agora inbox for [email protected]help / color / mirror / Atom feed
[PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms 261+ messages / 2 participants [nested] [flat]
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms @ 2023-01-17 04:38 Andres Freund <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw) Until now we used struct timespec on all platforms but windows. Using struct timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime overhead (much more complicated additions). Instead we can convert the time to nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations cheaper. Representing time as int64 nanoseconds provides sufficient range, ~292 years relative to a starting point (depending on clock source, relative to the unix epoch or the system's boot time). That'd not be sufficient for calendar time stored on disk, but is plenty for runtime interval time measurement. On windows instr_time already is represented as cycles. It might make sense to represent time as cycles on other platforms as well, as using cycle acquisition instructions like rdtsc directly can reduce the overhead of time acquisition substantially. This could be done in a fairly localized manner as the code stands after this commit. Because the windows and non-windows paths are now more similar, use a common set of macros. To make that possible, most of the use of LARGE_INTEGER had to be removed, which looks nicer anyway. To avoid users of the API relying on the integer representation, we wrap the 64bit integer inside struct struct instr_time. Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Author: David Geier <[email protected]> Reviewed-by: Tom Lane <[email protected]> Discussion: https://postgr.es/m/[email protected] --- src/include/portability/instr_time.h | 162 ++++++++++++++------------- 1 file changed, 86 insertions(+), 76 deletions(-) diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index 9ea1a68bd94..c0ed491395d 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -34,6 +34,8 @@ * * INSTR_TIME_GET_MICROSEC(t) convert t to uint64 (in microseconds) * + * INSTR_TIME_GET_NANOSEC(t) convert t to uint64 (in nanoseconds) + * * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert * absolute times to intervals. The INSTR_TIME_GET_xxx operations are * only useful on intervals. @@ -54,8 +56,32 @@ #ifndef INSTR_TIME_H #define INSTR_TIME_H + +/* + * We store interval times as an int64 integer on all platforms, as int64 is + * cheap to add/subtract, the most common operation for instr_time. The + * acquisition of time and converting to specific units of time is platform + * specific. + * + * To avoid users of the API relying on the integer representation, we wrap + * the 64bit integer in a struct. + */ +typedef struct instr_time +{ + int64 ticks; /* in platforms specific unit */ +} instr_time; + + +/* helpers macros used in platform specific code below */ + +#define NS_PER_S INT64CONST(1000000000) +#define NS_PER_MS INT64CONST(1000000) +#define NS_PER_US INT64CONST(1000) + + #ifndef WIN32 + /* Use clock_gettime() */ #include <time.h> @@ -80,93 +106,43 @@ #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -typedef struct timespec instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_clock_gettime_ns(void) +{ + instr_time now; + struct timespec tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).tv_nsec == 0 && (t).tv_sec == 0) + clock_gettime(PG_INSTR_CLOCK, &tmp); + now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec; -#define INSTR_TIME_SET_ZERO(t) ((t).tv_sec = 0, (t).tv_nsec = 0) + return now; +} -#define INSTR_TIME_SET_CURRENT(t) ((void) clock_gettime(PG_INSTR_CLOCK, &(t))) +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_clock_gettime_ns()) -#define INSTR_TIME_ADD(x,y) \ - do { \ - (x).tv_sec += (y).tv_sec; \ - (x).tv_nsec += (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (t).ticks) -#define INSTR_TIME_SUBTRACT(x,y) \ - do { \ - (x).tv_sec -= (y).tv_sec; \ - (x).tv_nsec -= (y).tv_nsec; \ - /* Normalize */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - } while (0) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - do { \ - (x).tv_sec += (y).tv_sec - (z).tv_sec; \ - (x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \ - /* Normalize after each add to avoid overflow/underflow of tv_nsec */ \ - while ((x).tv_nsec < 0) \ - { \ - (x).tv_nsec += 1000000000; \ - (x).tv_sec--; \ - } \ - while ((x).tv_nsec >= 1000000000) \ - { \ - (x).tv_nsec -= 1000000000; \ - (x).tv_sec++; \ - } \ - } while (0) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0) - -#define INSTR_TIME_GET_MICROSEC(t) \ - (((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000)) #else /* WIN32 */ + /* Use QueryPerformanceCounter() */ -typedef LARGE_INTEGER instr_time; +/* helper for INSTR_TIME_SET_CURRENT */ +static inline instr_time +pg_query_performance_counter(void) +{ + instr_time now; + LARGE_INTEGER tmp; -#define INSTR_TIME_IS_ZERO(t) ((t).QuadPart == 0) + QueryPerformanceCounter(&tmp); + now.ticks = tmp.QuadPart; -#define INSTR_TIME_SET_ZERO(t) ((t).QuadPart = 0) - -#define INSTR_TIME_SET_CURRENT(t) QueryPerformanceCounter(&(t)) - -#define INSTR_TIME_ADD(x,y) \ - ((x).QuadPart += (y).QuadPart) - -#define INSTR_TIME_SUBTRACT(x,y) \ - ((x).QuadPart -= (y).QuadPart) - -#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ - ((x).QuadPart += (y).QuadPart - (z).QuadPart) - -#define INSTR_TIME_GET_DOUBLE(t) \ - (((double) (t).QuadPart) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MILLISEC(t) \ - (((double) (t).QuadPart * 1000.0) / GetTimerFrequency()) - -#define INSTR_TIME_GET_MICROSEC(t) \ - ((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency())) + return now; +} static inline double GetTimerFrequency(void) @@ -177,11 +153,45 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_SET_CURRENT(t) \ + ((t) = pg_query_performance_counter()) + +#define INSTR_TIME_GET_NANOSEC(t) \ + ((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency())) + #endif /* WIN32 */ -/* same macro on all platforms */ + +/* + * Common macros + */ + +#define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) + + +#define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) #define INSTR_TIME_SET_CURRENT_LAZY(t) \ (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) + +#define INSTR_TIME_ADD(x,y) \ + ((x).ticks += (y).ticks) + +#define INSTR_TIME_SUBTRACT(x,y) \ + ((x).ticks -= (y).ticks) + +#define INSTR_TIME_ACCUM_DIFF(x,y,z) \ + ((x).ticks += (y).ticks - (z).ticks) + + +#define INSTR_TIME_GET_DOUBLE(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S) + +#define INSTR_TIME_GET_MILLISEC(t) \ + ((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS) + +#define INSTR_TIME_GET_MICROSEC(t) \ + (INSTR_TIME_GET_NANOSEC(t) / NS_PER_US) + #endif /* INSTR_TIME_H */ -- 2.38.0 --vt42kbp2j7rjcapp Content-Type: text/x-diff; charset=us-ascii Content-Disposition: attachment; filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
* [PATCH v25 2/9] Row pattern recognition patch (parse/analysis). @ 2024-12-21 06:19 Tatsuo Ishii <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Tatsuo Ishii @ 2024-12-21 06:19 UTC (permalink / raw) --- src/backend/parser/parse_agg.c | 7 + src/backend/parser/parse_clause.c | 297 +++++++++++++++++++++++++++++- src/backend/parser/parse_expr.c | 6 + src/backend/parser/parse_func.c | 3 + 4 files changed, 312 insertions(+), 1 deletion(-) diff --git a/src/backend/parser/parse_agg.c b/src/backend/parser/parse_agg.c index 04b4596a65..6af3bcb375 100644 --- a/src/backend/parser/parse_agg.c +++ b/src/backend/parser/parse_agg.c @@ -580,6 +580,10 @@ check_agglevels_and_constraints(ParseState *pstate, Node *expr) errkind = true; break; + case EXPR_KIND_RPR_DEFINE: + errkind = true; + break; + /* * There is intentionally no default: case here, so that the * compiler will warn if we add a new ParseExprKind without @@ -970,6 +974,9 @@ transformWindowFuncCall(ParseState *pstate, WindowFunc *wfunc, case EXPR_KIND_CYCLE_MARK: errkind = true; break; + case EXPR_KIND_RPR_DEFINE: + errkind = true; + break; /* * There is intentionally no default: case here, so that the diff --git a/src/backend/parser/parse_clause.c b/src/backend/parser/parse_clause.c index 979926b605..5a44c68b08 100644 --- a/src/backend/parser/parse_clause.c +++ b/src/backend/parser/parse_clause.c @@ -96,7 +96,14 @@ static WindowClause *findWindowClause(List *wclist, const char *name); static Node *transformFrameOffset(ParseState *pstate, int frameOptions, Oid rangeopfamily, Oid rangeopcintype, Oid *inRangeFunc, Node *clause); - +static void transformRPR(ParseState *pstate, WindowClause *wc, WindowDef *windef, + List **targetlist); +static List *transformDefineClause(ParseState *pstate, WindowClause *wc, WindowDef *windef, + List **targetlist); +static void transformPatternClause(ParseState *pstate, WindowClause *wc, + WindowDef *windef); +static List *transformMeasureClause(ParseState *pstate, WindowClause *wc, + WindowDef *windef); /* * transformFromClause - @@ -2954,6 +2961,10 @@ transformWindowDefinitions(ParseState *pstate, rangeopfamily, rangeopcintype, &wc->endInRangeFunc, windef->endOffset); + + /* Process Row Pattern Recognition related clauses */ + transformRPR(pstate, wc, windef, targetlist); + wc->winref = winref; result = lappend(result, wc); @@ -3821,3 +3832,287 @@ transformFrameOffset(ParseState *pstate, int frameOptions, return node; } + +/* + * transformRPR + * Process Row Pattern Recognition related clauses + */ +static void +transformRPR(ParseState *pstate, WindowClause *wc, WindowDef *windef, + List **targetlist) +{ + /* + * Window definition exists? + */ + if (windef == NULL) + return; + + /* + * Row Pattern Common Syntax clause exists? + */ + if (windef->rpCommonSyntax == NULL) + return; + + /* Check Frame option. Frame must start at current row */ + if ((wc->frameOptions & FRAMEOPTION_START_CURRENT_ROW) == 0) + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("FRAME must start at current row when row patttern recognition is used"))); + + /* Transform AFTER MACH SKIP TO clause */ + wc->rpSkipTo = windef->rpCommonSyntax->rpSkipTo; + + /* Transform AFTER MACH SKIP TO variable */ + wc->rpSkipVariable = windef->rpCommonSyntax->rpSkipVariable; + + /* Transform SEEK or INITIAL clause */ + wc->initial = windef->rpCommonSyntax->initial; + + /* Transform DEFINE clause into list of TargetEntry's */ + wc->defineClause = transformDefineClause(pstate, wc, windef, targetlist); + + /* Check PATTERN clause and copy to patternClause */ + transformPatternClause(pstate, wc, windef); + + /* Transform MEASURE clause */ + transformMeasureClause(pstate, wc, windef); +} + +/* + * transformDefineClause + * Process DEFINE clause and transform ResTarget into list of + * TargetEntry. + * + * XXX we only support column reference in row pattern definition search + * condition, e.g. "price". <row pattern definition variable name>.<column + * reference> is not supported, e.g. "A.price". + */ +static List * +transformDefineClause(ParseState *pstate, WindowClause *wc, WindowDef *windef, + List **targetlist) +{ + /* DEFINE variable name initials */ + static char *defineVariableInitials = "abcdefghijklmnopqrstuvwxyz"; + + ListCell *lc, + *l; + ResTarget *restarget, + *r; + List *restargets; + List *defineClause; + char *name; + int initialLen; + int i; + + /* + * If Row Definition Common Syntax exists, DEFINE clause must exist. (the + * raw parser should have already checked it.) + */ + Assert(windef->rpCommonSyntax->rpDefs != NULL); + + /* + * Check and add "A AS A IS TRUE" if pattern variable is missing in DEFINE + * per the SQL standard. + */ + restargets = NIL; + foreach(lc, windef->rpCommonSyntax->rpPatterns) + { + A_Expr *a; + bool found = false; + + if (!IsA(lfirst(lc), A_Expr)) + ereport(ERROR, + errmsg("node type is not A_Expr")); + + a = (A_Expr *) lfirst(lc); + name = strVal(a->lexpr); + + foreach(l, windef->rpCommonSyntax->rpDefs) + { + restarget = (ResTarget *) lfirst(l); + + if (!strcmp(restarget->name, name)) + { + found = true; + break; + } + } + + if (!found) + { + /* + * "name" is missing. So create "name AS name IS TRUE" ResTarget + * node and add it to the temporary list. + */ + A_Const *n; + + restarget = makeNode(ResTarget); + n = makeNode(A_Const); + n->val.boolval.type = T_Boolean; + n->val.boolval.boolval = true; + n->location = -1; + restarget->name = pstrdup(name); + restarget->indirection = NIL; + restarget->val = (Node *) n; + restarget->location = -1; + restargets = lappend((List *) restargets, restarget); + } + } + + if (list_length(restargets) >= 1) + { + /* add missing DEFINEs */ + windef->rpCommonSyntax->rpDefs = + list_concat(windef->rpCommonSyntax->rpDefs, restargets); + list_free(restargets); + } + + /* + * Check for duplicate row pattern definition variables. The standard + * requires that no two row pattern definition variable names shall be + * equivalent. + */ + restargets = NIL; + foreach(lc, windef->rpCommonSyntax->rpDefs) + { + restarget = (ResTarget *) lfirst(lc); + name = restarget->name; + + /* + * Add DEFINE expression (Restarget->val) to the targetlist as a + * TargetEntry if it does not exist yet. Planner will add the column + * ref var node to the outer plan's target list later on. This makes + * DEFINE expression could access the outer tuple while evaluating + * PATTERN. + * + * XXX: adding whole expressions of DEFINE to the plan.targetlist is + * not so good, because it's not necessary to evalute the expression + * in the target list while running the plan. We should extract the + * var nodes only then add them to the plan.targetlist. + */ + findTargetlistEntrySQL99(pstate, (Node *) restarget->val, + targetlist, EXPR_KIND_RPR_DEFINE); + + /* + * Make sure that the row pattern definition search condition is a + * boolean expression. + */ + transformWhereClause(pstate, restarget->val, + EXPR_KIND_RPR_DEFINE, "DEFINE"); + + foreach(l, restargets) + { + char *n; + + r = (ResTarget *) lfirst(l); + n = r->name; + + if (!strcmp(n, name)) + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("row pattern definition variable name \"%s\" appears more than once in DEFINE clause", + name), + parser_errposition(pstate, exprLocation((Node *) r)))); + } + restargets = lappend(restargets, restarget); + } + list_free(restargets); + + /* + * Create list of row pattern DEFINE variable name's initial. We assign + * [a-z] to them (up to 26 variable names are allowed). + */ + restargets = NIL; + i = 0; + initialLen = strlen(defineVariableInitials); + + foreach(lc, windef->rpCommonSyntax->rpDefs) + { + char initial[2]; + + restarget = (ResTarget *) lfirst(lc); + name = restarget->name; + + if (i >= initialLen) + { + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("number of row pattern definition variable names exceeds %d", + initialLen), + parser_errposition(pstate, + exprLocation((Node *) restarget)))); + } + initial[0] = defineVariableInitials[i++]; + initial[1] = '\0'; + wc->defineInitial = lappend(wc->defineInitial, + makeString(pstrdup(initial))); + } + + defineClause = transformTargetList(pstate, windef->rpCommonSyntax->rpDefs, + EXPR_KIND_RPR_DEFINE); + + /* mark column origins */ + markTargetListOrigins(pstate, defineClause); + + /* mark all nodes in the DEFINE clause tree with collation information */ + assign_expr_collations(pstate, (Node *) defineClause); + + return defineClause; +} + +/* + * transformPatternClause + * Process PATTERN clause and return PATTERN clause in the raw parse tree + */ +static void +transformPatternClause(ParseState *pstate, WindowClause *wc, + WindowDef *windef) +{ + ListCell *lc; + + /* + * Row Pattern Common Syntax clause exists? + */ + if (windef->rpCommonSyntax == NULL) + return; + + wc->patternVariable = NIL; + wc->patternRegexp = NIL; + foreach(lc, windef->rpCommonSyntax->rpPatterns) + { + A_Expr *a; + char *name; + char *regexp; + + if (!IsA(lfirst(lc), A_Expr)) + ereport(ERROR, + errmsg("node type is not A_Expr")); + + a = (A_Expr *) lfirst(lc); + name = strVal(a->lexpr); + + wc->patternVariable = lappend(wc->patternVariable, makeString(pstrdup(name))); + regexp = strVal(lfirst(list_head(a->name))); + + wc->patternRegexp = lappend(wc->patternRegexp, makeString(pstrdup(regexp))); + } +} + +/* + * transformMeasureClause + * Process MEASURE clause + * XXX MEASURE clause is not supported yet + */ +static List * +transformMeasureClause(ParseState *pstate, WindowClause *wc, + WindowDef *windef) +{ + if (windef->rowPatternMeasures == NIL) + return NIL; + + ereport(ERROR, + (errcode(ERRCODE_SYNTAX_ERROR), + errmsg("%s", "MEASURE clause is not supported yet"), + parser_errposition(pstate, exprLocation((Node *) windef->rowPatternMeasures)))); + return NIL; +} diff --git a/src/backend/parser/parse_expr.c b/src/backend/parser/parse_expr.c index c2806297aa..e827c59fd4 100644 --- a/src/backend/parser/parse_expr.c +++ b/src/backend/parser/parse_expr.c @@ -575,6 +575,7 @@ transformColumnRef(ParseState *pstate, ColumnRef *cref) case EXPR_KIND_COPY_WHERE: case EXPR_KIND_GENERATED_COLUMN: case EXPR_KIND_CYCLE_MARK: + case EXPR_KIND_RPR_DEFINE: /* okay */ break; @@ -1858,6 +1859,9 @@ transformSubLink(ParseState *pstate, SubLink *sublink) case EXPR_KIND_GENERATED_COLUMN: err = _("cannot use subquery in column generation expression"); break; + case EXPR_KIND_RPR_DEFINE: + err = _("cannot use subquery in DEFINE expression"); + break; /* * There is intentionally no default: case here, so that the @@ -3197,6 +3201,8 @@ ParseExprKindName(ParseExprKind exprKind) return "GENERATED AS"; case EXPR_KIND_CYCLE_MARK: return "CYCLE"; + case EXPR_KIND_RPR_DEFINE: + return "DEFINE"; /* * There is intentionally no default: case here, so that the diff --git a/src/backend/parser/parse_func.c b/src/backend/parser/parse_func.c index 9b23344a3b..4c482abb30 100644 --- a/src/backend/parser/parse_func.c +++ b/src/backend/parser/parse_func.c @@ -2658,6 +2658,9 @@ check_srf_call_placement(ParseState *pstate, Node *last_srf, int location) case EXPR_KIND_CYCLE_MARK: errkind = true; break; + case EXPR_KIND_RPR_DEFINE: + errkind = true; + break; /* * There is intentionally no default: case here, so that the -- 2.25.1 ----Next_Part(Sat_Dec_21_18_20_04_2024_526)-- Content-Type: Text/X-Patch; charset=us-ascii Content-Transfer-Encoding: 7bit Content-Disposition: inline; filename="v25-0003-Row-pattern-recognition-patch-rewriter.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
end of thread, other threads:[~2024-12-21 06:19 UTC | newest] Thread overview: 261+ messages (download: mbox mbox.gz follow: Atom feed) -- links below jump to the message on this page -- 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2023-01-17 04:38 [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms Andres Freund <[email protected]> 2024-12-21 06:19 [PATCH v25 2/9] Row pattern recognition patch (parse/analysis). Tatsuo Ishii <[email protected]>
This inbox is served by agora; see mirroring instructions for how to clone and mirror all data and code used for this inbox