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 v19 4/6] Move conversion of a "historic" to MVCC snapshot to a separate function. @ 2025-08-11 13:23 Antonin Houska <[email protected]> 0 siblings, 0 replies; 261+ messages in thread From: Antonin Houska @ 2025-08-11 13:23 UTC (permalink / raw) The conversion is now handled by SnapBuildMVCCFromHistoric(). REPACK CONCURRENTLY will also need it. --- src/backend/replication/logical/snapbuild.c | 51 +++++++++++++++++---- src/backend/utils/time/snapmgr.c | 3 +- src/include/replication/snapbuild.h | 1 + src/include/utils/snapmgr.h | 1 + 4 files changed, 45 insertions(+), 11 deletions(-) diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c index 98ddee20929..a2f1803622c 100644 --- a/src/backend/replication/logical/snapbuild.c +++ b/src/backend/replication/logical/snapbuild.c @@ -440,10 +440,7 @@ Snapshot SnapBuildInitialSnapshot(SnapBuild *builder) { Snapshot snap; - TransactionId xid; TransactionId safeXid; - TransactionId *newxip; - int newxcnt = 0; Assert(XactIsoLevel == XACT_REPEATABLE_READ); Assert(builder->building_full_snapshot); @@ -485,6 +482,31 @@ SnapBuildInitialSnapshot(SnapBuild *builder) MyProc->xmin = snap->xmin; + /* Convert the historic snapshot to MVCC snapshot. */ + return SnapBuildMVCCFromHistoric(snap, true); +} + +/* + * Turn a historic MVCC snapshot into an ordinary MVCC snapshot. + * + * Unlike a regular (non-historic) MVCC snapshot, the xip array of this + * snapshot contains not only running main transactions, but also their + * subtransactions. This difference does has no impact on XidInMVCCSnapshot(). + * + * Pass true for 'in_place' if you don't care about modifying the source + * snapshot. If you need a new instance, and one that was allocated as a + * single chunk of memory, pass false. + */ +Snapshot +SnapBuildMVCCFromHistoric(Snapshot snapshot, bool in_place) +{ + TransactionId xid; + TransactionId *oldxip = snapshot->xip; + uint32 oldxcnt = snapshot->xcnt; + TransactionId *newxip; + int newxcnt = 0; + Snapshot result; + /* allocate in transaction context */ newxip = (TransactionId *) palloc(sizeof(TransactionId) * GetMaxSnapshotXidCount()); @@ -495,7 +517,7 @@ SnapBuildInitialSnapshot(SnapBuild *builder) * classical snapshot by marking all non-committed transactions as * in-progress. This can be expensive. */ - for (xid = snap->xmin; NormalTransactionIdPrecedes(xid, snap->xmax);) + for (xid = snapshot->xmin; NormalTransactionIdPrecedes(xid, snapshot->xmax);) { void *test; @@ -503,7 +525,7 @@ SnapBuildInitialSnapshot(SnapBuild *builder) * Check whether transaction committed using the decoding snapshot * meaning of ->xip. */ - test = bsearch(&xid, snap->xip, snap->xcnt, + test = bsearch(&xid, snapshot->xip, snapshot->xcnt, sizeof(TransactionId), xidComparator); if (test == NULL) @@ -520,11 +542,22 @@ SnapBuildInitialSnapshot(SnapBuild *builder) } /* adjust remaining snapshot fields as needed */ - snap->snapshot_type = SNAPSHOT_MVCC; - snap->xcnt = newxcnt; - snap->xip = newxip; + snapshot->xcnt = newxcnt; + snapshot->xip = newxip; - return snap; + if (in_place) + result = snapshot; + else + { + result = CopySnapshot(snapshot); + + /* Restore the original values so the source is intact. */ + snapshot->xip = oldxip; + snapshot->xcnt = oldxcnt; + } + result->snapshot_type = SNAPSHOT_MVCC; + + return result; } /* diff --git a/src/backend/utils/time/snapmgr.c b/src/backend/utils/time/snapmgr.c index 65561cc6bc3..bc7840052fe 100644 --- a/src/backend/utils/time/snapmgr.c +++ b/src/backend/utils/time/snapmgr.c @@ -212,7 +212,6 @@ typedef struct ExportedSnapshot static List *exportedSnapshots = NIL; /* Prototypes for local functions */ -static Snapshot CopySnapshot(Snapshot snapshot); static void UnregisterSnapshotNoOwner(Snapshot snapshot); static void FreeSnapshot(Snapshot snapshot); static void SnapshotResetXmin(void); @@ -602,7 +601,7 @@ SetTransactionSnapshot(Snapshot sourcesnap, VirtualTransactionId *sourcevxid, * The copy is palloc'd in TopTransactionContext and has initial refcounts set * to 0. The returned snapshot has the copied flag set. */ -static Snapshot +Snapshot CopySnapshot(Snapshot snapshot) { Snapshot newsnap; diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h index 44031dcf6e3..6d4d2d1814c 100644 --- a/src/include/replication/snapbuild.h +++ b/src/include/replication/snapbuild.h @@ -73,6 +73,7 @@ extern void FreeSnapshotBuilder(SnapBuild *builder); extern void SnapBuildSnapDecRefcount(Snapshot snap); extern Snapshot SnapBuildInitialSnapshot(SnapBuild *builder); +extern Snapshot SnapBuildMVCCFromHistoric(Snapshot snapshot, bool in_place); extern const char *SnapBuildExportSnapshot(SnapBuild *builder); extern void SnapBuildClearExportedSnapshot(void); extern void SnapBuildResetExportedSnapshotState(void); diff --git a/src/include/utils/snapmgr.h b/src/include/utils/snapmgr.h index 604c1f90216..f65f83c85cd 100644 --- a/src/include/utils/snapmgr.h +++ b/src/include/utils/snapmgr.h @@ -63,6 +63,7 @@ extern Snapshot GetTransactionSnapshot(void); extern Snapshot GetLatestSnapshot(void); extern void SnapshotSetCommandId(CommandId curcid); +extern Snapshot CopySnapshot(Snapshot snapshot); extern Snapshot GetCatalogSnapshot(Oid relid); extern Snapshot GetNonHistoricCatalogSnapshot(Oid relid); extern void InvalidateCatalogSnapshot(void); -- 2.39.5 --ipmcfwv2cm3fpxaf Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v19-0005-Add-CONCURRENTLY-option-to-REPACK-command.patch" ^ permalink raw reply [nested|flat] 261+ messages in thread
end of thread, other threads:[~2025-08-11 13:23 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]> 2025-08-11 13:23 [PATCH v19 4/6] Move conversion of a "historic" to MVCC snapshot to a separate function. Antonin Houska <[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