agora inbox for [email protected]  

help / color / mirror / Atom feed

Possible typo in nodeAgg.c
261+ messages / 2 participants
[nested] [flat]

* Possible typo in nodeAgg.c
@ 2020-10-16 09:03  Hou, Zhijie <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Hou, Zhijie @ 2020-10-16 09:03 UTC (permalink / raw)
  To: PostgreSQL Developers <[email protected]>

Hi

In /src/backend/executor/nodeAgg.c

I found the following comment still use work mem,
Since hash_mem has been introduced, Is it more accurate to use hash_mem here ?

@@ -1827,7 +1827,7 @@ hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
 	/*
 	 * Don't set the limit below 3/4 of hash_mem. In that case, we are at the
 	 * minimum number of partitions, so we aren't going to dramatically exceed
-	 * work mem anyway.
+	 * hash_mem anyway.

Best regards,
houzj




Attachments:

  [application/octet-stream] 0001-fix-typo-in-nodeAgg.c.patch (876B, ../../7065ae4343494632bdd3ac7226d52ae6@G08CNEXMBPEKD05.g08.fujitsu.local/2-0001-fix-typo-in-nodeAgg.c.patch)
  download | inline diff:
From f2dc0c06824f4361f81e5d866fda4d67f6340da8 Mon Sep 17 00:00:00 2001
From: root <[email protected]>
Date: Fri, 16 Oct 2020 04:45:06 -0400
Subject: [PATCH] fix typo in nodeAgg.c

---
 src/backend/executor/nodeAgg.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/backend/executor/nodeAgg.c b/src/backend/executor/nodeAgg.c
index 75e5bbf..255d014 100644
--- a/src/backend/executor/nodeAgg.c
+++ b/src/backend/executor/nodeAgg.c
@@ -1827,7 +1827,7 @@ hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
 	/*
 	 * Don't set the limit below 3/4 of hash_mem. In that case, we are at the
 	 * minimum number of partitions, so we aren't going to dramatically exceed
-	 * work mem anyway.
+	 * hash_mem anyway.
 	 */
 	if (hash_mem * 1024L > 4 * partition_mem)
 		*mem_limit = hash_mem * 1024L - partition_mem;
-- 
1.8.3.1



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread

* [PATCH v8 2/5] instr_time: Represent time as an int64 on all platforms
@ 2023-01-17 04:38  Andres Freund <[email protected]>
  0 siblings, 0 replies; 261+ messages in thread

From: Andres Freund @ 2023-01-17 04:38 UTC (permalink / raw)

Until now we used struct timespec on all platforms but windows. Using struct
timespe causes a fair bit of memory (struct timeval is 16 bytes) and runtime
overhead (much more complicated additions). Instead we can convert the time to
nanoseconds in INSTR_TIME_SET_CURRENT(), making the remaining operations
cheaper.

Representing time as int64 nanoseconds provides sufficient range, ~292 years
relative to a starting point (depending on clock source, relative to the unix
epoch or the system's boot time). That'd not be sufficient for calendar time
stored on disk, but is plenty for runtime interval time measurement.

On windows instr_time already is represented as cycles. It might make sense to
represent time as cycles on other platforms as well, as using cycle
acquisition instructions like rdtsc directly can reduce the overhead of time
acquisition substantially. This could be done in a fairly localized manner as
the code stands after this commit.

Because the windows and non-windows paths are now more similar, use a common
set of macros. To make that possible, most of the use of LARGE_INTEGER had to
be removed, which looks nicer anyway.

To avoid users of the API relying on the integer representation, we wrap the
64bit integer inside struct struct instr_time.

Author: Andres Freund <[email protected]>
Author: Lukas Fittl <[email protected]>
Author: David Geier <[email protected]>
Reviewed-by: Tom Lane <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/include/portability/instr_time.h | 162 ++++++++++++++-------------
 1 file changed, 86 insertions(+), 76 deletions(-)

diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h
index 9ea1a68bd94..c0ed491395d 100644
--- a/src/include/portability/instr_time.h
+++ b/src/include/portability/instr_time.h
@@ -34,6 +34,8 @@
  *
  * INSTR_TIME_GET_MICROSEC(t)		convert t to uint64 (in microseconds)
  *
+ * INSTR_TIME_GET_NANOSEC(t)		convert t to uint64 (in nanoseconds)
+ *
  * Note that INSTR_TIME_SUBTRACT and INSTR_TIME_ACCUM_DIFF convert
  * absolute times to intervals.  The INSTR_TIME_GET_xxx operations are
  * only useful on intervals.
@@ -54,8 +56,32 @@
 #ifndef INSTR_TIME_H
 #define INSTR_TIME_H
 
+
+/*
+ * We store interval times as an int64 integer on all platforms, as int64 is
+ * cheap to add/subtract, the most common operation for instr_time. The
+ * acquisition of time and converting to specific units of time is platform
+ * specific.
+ *
+ * To avoid users of the API relying on the integer representation, we wrap
+ * the 64bit integer in a struct.
+ */
+typedef struct instr_time
+{
+	int64		ticks;			/* in platforms specific unit */
+} instr_time;
+
+
+/* helpers macros used in platform specific code below */
+
+#define NS_PER_S	INT64CONST(1000000000)
+#define NS_PER_MS	INT64CONST(1000000)
+#define NS_PER_US	INT64CONST(1000)
+
+
 #ifndef WIN32
 
+
 /* Use clock_gettime() */
 
 #include <time.h>
@@ -80,93 +106,43 @@
 #define PG_INSTR_CLOCK	CLOCK_REALTIME
 #endif
 
-typedef struct timespec instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_clock_gettime_ns(void)
+{
+	instr_time	now;
+	struct timespec tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).tv_nsec == 0 && (t).tv_sec == 0)
+	clock_gettime(PG_INSTR_CLOCK, &tmp);
+	now.ticks = tmp.tv_sec * NS_PER_S + tmp.tv_nsec;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).tv_sec = 0, (t).tv_nsec = 0)
+	return now;
+}
 
-#define INSTR_TIME_SET_CURRENT(t)	((void) clock_gettime(PG_INSTR_CLOCK, &(t)))
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_clock_gettime_ns())
 
-#define INSTR_TIME_ADD(x,y) \
-	do { \
-		(x).tv_sec += (y).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (t).ticks)
 
-#define INSTR_TIME_SUBTRACT(x,y) \
-	do { \
-		(x).tv_sec -= (y).tv_sec; \
-		(x).tv_nsec -= (y).tv_nsec; \
-		/* Normalize */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	do { \
-		(x).tv_sec += (y).tv_sec - (z).tv_sec; \
-		(x).tv_nsec += (y).tv_nsec - (z).tv_nsec; \
-		/* Normalize after each add to avoid overflow/underflow of tv_nsec */ \
-		while ((x).tv_nsec < 0) \
-		{ \
-			(x).tv_nsec += 1000000000; \
-			(x).tv_sec--; \
-		} \
-		while ((x).tv_nsec >= 1000000000) \
-		{ \
-			(x).tv_nsec -= 1000000000; \
-			(x).tv_sec++; \
-		} \
-	} while (0)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).tv_sec) + ((double) (t).tv_nsec) / 1000000000.0)
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).tv_sec * 1000.0) + ((double) (t).tv_nsec) / 1000000.0)
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	(((uint64) (t).tv_sec * (uint64) 1000000) + (uint64) ((t).tv_nsec / 1000))
 
 #else							/* WIN32 */
 
+
 /* Use QueryPerformanceCounter() */
 
-typedef LARGE_INTEGER instr_time;
+/* helper for INSTR_TIME_SET_CURRENT */
+static inline instr_time
+pg_query_performance_counter(void)
+{
+	instr_time	now;
+	LARGE_INTEGER tmp;
 
-#define INSTR_TIME_IS_ZERO(t)	((t).QuadPart == 0)
+	QueryPerformanceCounter(&tmp);
+	now.ticks = tmp.QuadPart;
 
-#define INSTR_TIME_SET_ZERO(t)	((t).QuadPart = 0)
-
-#define INSTR_TIME_SET_CURRENT(t)	QueryPerformanceCounter(&(t))
-
-#define INSTR_TIME_ADD(x,y) \
-	((x).QuadPart += (y).QuadPart)
-
-#define INSTR_TIME_SUBTRACT(x,y) \
-	((x).QuadPart -= (y).QuadPart)
-
-#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
-	((x).QuadPart += (y).QuadPart - (z).QuadPart)
-
-#define INSTR_TIME_GET_DOUBLE(t) \
-	(((double) (t).QuadPart) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MILLISEC(t) \
-	(((double) (t).QuadPart * 1000.0) / GetTimerFrequency())
-
-#define INSTR_TIME_GET_MICROSEC(t) \
-	((uint64) (((double) (t).QuadPart * 1000000.0) / GetTimerFrequency()))
+	return now;
+}
 
 static inline double
 GetTimerFrequency(void)
@@ -177,11 +153,45 @@ GetTimerFrequency(void)
 	return (double) f.QuadPart;
 }
 
+#define INSTR_TIME_SET_CURRENT(t) \
+	((t) = pg_query_performance_counter())
+
+#define INSTR_TIME_GET_NANOSEC(t) \
+	((int64) (((double) (t).ticks * NS_PER_S) / GetTimerFrequency()))
+
 #endif							/* WIN32 */
 
-/* same macro on all platforms */
+
+/*
+ * Common macros
+ */
+
+#define INSTR_TIME_IS_ZERO(t)	((t).ticks == 0)
+
+
+#define INSTR_TIME_SET_ZERO(t)	((t).ticks = 0)
 
 #define INSTR_TIME_SET_CURRENT_LAZY(t) \
 	(INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false)
 
+
+#define INSTR_TIME_ADD(x,y) \
+	((x).ticks += (y).ticks)
+
+#define INSTR_TIME_SUBTRACT(x,y) \
+	((x).ticks -= (y).ticks)
+
+#define INSTR_TIME_ACCUM_DIFF(x,y,z) \
+	((x).ticks += (y).ticks - (z).ticks)
+
+
+#define INSTR_TIME_GET_DOUBLE(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_S)
+
+#define INSTR_TIME_GET_MILLISEC(t) \
+	((double) INSTR_TIME_GET_NANOSEC(t) / NS_PER_MS)
+
+#define INSTR_TIME_GET_MICROSEC(t) \
+	(INSTR_TIME_GET_NANOSEC(t) / NS_PER_US)
+
 #endif							/* INSTR_TIME_H */
-- 
2.38.0


--vt42kbp2j7rjcapp
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0003-instr_time-Add-INSTR_TIME_SET_SECONDS-INSTR_TIME_.patch"



^ permalink  raw  reply  [nested|flat] 261+ messages in thread