faster version of AllocSetFreeIndex for x86 architecture

Hi,

I made a faster version of AllocSetFreeIndex for x86 architecture.

Neat, I have a version for PowerPC too.

In order to prevent writing multiple copies of AllocSetFreeIndex, I
propose that we add a fls() function ("find last set"); this can be
defined in an architecture-independent manner (ie, shift mask & test in
a loop), and re-defined for arches that have faster ways of doing the
same (ie, cntlz instruction on powerpc).

We can then change AllocSetFreeIndex to use fls().

Patches coming...

Jeremy

Add a utility header for simple bit operatios - bitops.h.

At present, just contains the fls() (find last set bit) function.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>

---
src/include/utils/bitops.h | 52 +++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 52 insertions(+)

diff --git a/src/include/utils/bitops.h b/src/include/utils/bitops.h
new file mode 100644
index 0000000..de11624
--- /dev/null
+++ b/src/include/utils/bitops.h
@@ -0,0 +1,52 @@
+/*-------------------------------------------------------------------------
+ *
+ * bitops.h
+ *	  Simple bit operations.
+ *
+ * Portions Copyright (c) 2009, PostgreSQL Global Development Group
+ *
+ * $PostgreSQL$
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef BITOPS_H
+#define BITOPS_H
+
+#if defined(__ppc__) || defined(__powerpc__) || \
+	defined(__ppc64__) || defined (__powerpc64__)
+
+static inline int
+fls(unsigned int x)
+{
+	int lz;
+	asm("cntlz %0,%1" : "=r" (lz) : "r" (x));
+	return 32 - lz;
+}
+
+#else /* !powerpc */
+
+/* Architecture-independent implementations */
+
+/*
+ * fls: find last set bit.
+ *
+ * Returns the 1-based index of the most-significant bit in x. The MSB
+ * is bit number 32, the LSB is bit number 1. If x is zero, returns zero.
+ */
+static inline int
+fls(unsigned int x)
+{
+	int ls = 0;
+
+	while (x != 0)
+	{
+		ls++;
+		x >>= 1;
+	}
+
+	return ls;
+}
+
+#endif
+
+#endif /* BITOPS_H */

Results in a ~2% performance increase by using the powerpc fls()
implementation.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>

---
src/backend/utils/mmgr/aset.c | 8 ++------
1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/src/backend/utils/mmgr/aset.c b/src/backend/utils/mmgr/aset.c
index 0e2d4d5..762cf72 100644
--- a/src/backend/utils/mmgr/aset.c
+++ b/src/backend/utils/mmgr/aset.c
@@ -65,6 +65,7 @@
 #include "postgres.h"

#include "utils/memutils.h"
+#include "utils/bitops.h"

/* Define this to detail debug alloc information */
/* #define HAVE_ALLOCINFO */
@@ -270,12 +271,7 @@ AllocSetFreeIndex(Size size)

 	if (size > 0)
 	{
-		size = (size - 1) >> ALLOC_MINBITS;
-		while (size != 0)
-		{
-			idx++;
-			size >>= 1;
-		}
+		idx = fls((size - 1) >> ALLOC_MINBITS);
 		Assert(idx < ALLOCSET_NUM_FREELISTS);
 	}

Jeremy Kerr <jk@ozlabs.org> writes:

Add a utility header for simple bit operatios - bitops.h.

This will fail outright on any non-gcc compiler.

regards, tom lane

Add a utility header for simple bit operatios - bitops.h.

At present, just contains the fls() (find last set bit) function.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>

---
v2: only use inline asm with gcc

---
src/include/utils/bitops.h | 53 +++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 53 insertions(+)

diff --git a/src/include/utils/bitops.h b/src/include/utils/bitops.h
new file mode 100644
index 0000000..4f2bbc9
--- /dev/null
+++ b/src/include/utils/bitops.h
@@ -0,0 +1,53 @@
+/*-------------------------------------------------------------------------
+ *
+ * bitops.h
+ *	  Simple bit operations.
+ *
+ * Portions Copyright (c) 2009, PostgreSQL Global Development Group
+ *
+ * $PostgreSQL$
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef BITOPS_H
+#define BITOPS_H
+
+#if defined(__GNUC__) && \
+	(defined(__ppc__) || defined(__powerpc__) || \
+	 defined(__ppc64__) || defined (__powerpc64__))
+
+static inline int
+fls(unsigned int x)
+{
+	int lz;
+	asm("cntlz %0,%1" : "=r" (lz) : "r" (x));
+	return 32 - lz;
+}
+
+#else /* !(gcc && powerpc) */
+
+/* Architecture-independent implementations */
+
+/*
+ * fls: find last set bit.
+ *
+ * Returns the 1-based index of the most-significant bit in x. The MSB
+ * is bit number 32, the LSB is bit number 1. If x is zero, returns zero.
+ */
+static inline int
+fls(unsigned int x)
+{
+	int ls = 0;
+
+	while (x != 0)
+	{
+		ls++;
+		x >>= 1;
+	}
+
+	return ls;
+}
+
+#endif
+
+#endif /* BITOPS_H */

* Jeremy Kerr:

+#if defined(__GNUC__) && \
+	(defined(__ppc__) || defined(__powerpc__) || \
+	 defined(__ppc64__) || defined (__powerpc64__))

If you require GCC anyway, you can use __builtin_clz instead.
(It's been available since GCC 4.1 at least.)

--
Florian Weimer <fweimer@bfk.de>
BFK edv-consulting GmbH http://www.bfk.de/
Kriegsstraße 100 tel: +49-721-96201-1
D-76133 Karlsruhe fax: +49-721-96201-99

Florian,

+#if defined(__GNUC__) && \
+	(defined(__ppc__) || defined(__powerpc__) || \
+	 defined(__ppc64__) || defined (__powerpc64__))

If you require GCC anyway, you can use __builtin_clz instead.
(It's been available since GCC 4.1 at least.)

Because now we have to test the compiler *and* the version as well?

But I do agree that using the builtins makes for much better code; I'm
looking at a future change that does this.

Cheers,

Jeremy

* Jeremy Kerr:

Florian,

+#if defined(__GNUC__) && \
+	(defined(__ppc__) || defined(__powerpc__) || \
+	 defined(__ppc64__) || defined (__powerpc64__))

If you require GCC anyway, you can use __builtin_clz instead.
(It's been available since GCC 4.1 at least.)

Because now we have to test the compiler *and* the version as well?

This builtin is not architecture-specific, so you'd save the
architecture check.

--
Florian Weimer <fweimer@bfk.de>
BFK edv-consulting GmbH http://www.bfk.de/
Kriegsstraße 100 tel: +49-721-96201-1
D-76133 Karlsruhe fax: +49-721-96201-99

Florian Weimer <fweimer@bfk.de> writes:

* Jeremy Kerr:

Because now we have to test the compiler *and* the version as well?

This builtin is not architecture-specific, so you'd save the
architecture check.

The appropriate way to handle it would be a configure probe to see if
the function is available, thus avoiding any wired-in knowledge about
compiler or compiler version *or* architecture.

The other thing I didn't like about the patch was the assumption that
it's okay to have a "static inline" function in a header. You can
get away with that in gcc but *not* in other compilers. Look at the
existing coding patterns for, eg, list_head; then go thou and do
likewise. Or, since there's currently no need for the code outside
aset.c, forget about putting it in a header and just plop it into
aset.c.

regards, tom lane

Hi Tom,

The other thing I didn't like about the patch was the assumption that
it's okay to have a "static inline" function in a header. You can
get away with that in gcc but *not* in other compilers.

Gee, you user-space guys have it tough! :D

Point taken, will rework.

Look at the existing coding patterns for, eg, list_head; then go thou
and do likewise. Or, since there's currently no need for the code
outside aset.c, forget about putting it in a header and just plop it
into aset.c.

OK, I'll add a configure check and conditionally use the builtin if it's
available. I have some other patches that could be improved by using
other builtins, so it would be a good opportunity to figure out a nice
pattern for doing this.

Cheers,

Jeremy

Move the shift-and-test login into a separate fls() function, which
can use __builtin_clz() if it's available.

This requires a new check for __builtin_clz in the configure script.

Results in a ~2% performance increase on PowerPC.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>

---
configure.in | 13 +++++++++++++
src/backend/utils/mmgr/aset.c | 32 ++++++++++++++++++++++++++------
2 files changed, 39 insertions(+), 6 deletions(-)

diff --git a/configure.in b/configure.in
index b8d2685..6a317b0 100644
--- a/configure.in
+++ b/configure.in
@@ -1361,6 +1361,19 @@ case $host_os in
 		AC_FUNC_FSEEKO;;
 esac

+# GCC builtins
+#
+# We need AC_TRY_LINK here, as the prototype generated by AC_CHECK_FUNC
+# will cause gcc to try to reference a non-builtin symbol.
+
+AC_MSG_CHECKING([for __builtin_clz])
+AC_TRY_LINK([],
+	    [__builtin_clz(0);],
+	    [AC_DEFINE(HAVE_BUILTIN_CLZ, 1,
+				[Define to 1 if you have __builtin_clz().])
+			AC_MSG_RESULT(yes)],
+		[AC_MSG_RESULT(no)])
+

 #
 # Pthreads
diff --git a/src/backend/utils/mmgr/aset.c b/src/backend/utils/mmgr/aset.c
index 0e2d4d5..af352b8 100644
--- a/src/backend/utils/mmgr/aset.c
+++ b/src/backend/utils/mmgr/aset.c
@@ -255,6 +255,31 @@ static MemoryContextMethods AllocSetMethods = {
 #define AllocAllocInfo(_cxt, _chunk)
 #endif

+/*
+ * fls: find last set bit.
+ *
+ * Returns the 1-based index of the most-significant bit in x. The MSB
+ * is bit number 32, the LSB is bit number 1. If x is zero, the result is
+ * undefined.
+ */
+static inline int
+fls(unsigned int x)
+{
+#ifdef HAVE_BUILTIN_CLZ
+	return 32 - __builtin_clz(x);
+#else
+	int ls = 0;
+
+	while (x != 0)
+	{
+		ls++;
+		x >>= 1;
+	}
+
+	return ls;
+#endif
+}
+
 /* ----------
  * AllocSetFreeIndex -
  *
@@ -270,12 +295,7 @@ AllocSetFreeIndex(Size size)

 	if (size > 0)
 	{
-		size = (size - 1) >> ALLOC_MINBITS;
-		while (size != 0)
-		{
-			idx++;
-			size >>= 1;
-		}
+		idx = fls((size - 1) >> ALLOC_MINBITS);
 		Assert(idx < ALLOCSET_NUM_FREELISTS);
 	}

+/*
+ * fls: find last set bit.
+ *
+ * Returns the 1-based index of the most-significant bit in x. The MSB
+ * is bit number 32, the LSB is bit number 1. If x is zero, the result is
+ * undefined.
+ */
+static inline int
+fls(unsigned int x)

...

+ idx = fls((size - 1) >> ALLOC_MINBITS);

If size <= 8, fls((size - 1) >> ALLOC_MINBITS) is fls(0).
The result of fls(0) is undefined.

I think we have to never call fls(0) from AllocSetFreeIndex().
My proposal code:

if (size > (1 << ALLOC_MINBITS))
{
idx = fls((size - 1) >> ALLOC_MINBITS);
Assert(idx < ALLOCSET_NUM_FREELISTS);
}

Best regards,

---
Atsushi Ogawa

On Tue, 2009-06-02 at 23:53 +0900, Atsushi Ogawa wrote:

I made a faster version of AllocSetFreeIndex for x86 architecture.

Results of benchmark script:
Xeon(Core architecture), RedHat EL4, gcc 3.4.6
bytes : 4 8 16 32 64 128 256 512 1024 mix
original: 0.780 0.780 0.820 0.870 0.930 0.970 1.030 1.080 1.130 0.950
patched : 0.380 0.170 0.170 0.170 0.170 0.180 0.170 0.180 0.180 0.280

The effect of the patch that I measured by oprofile is:
- test program: pgbench -c 1 -t 50000 (fsync=off)

original:
CPU: P4 / Xeon with 2 hyper-threads, speed 2793.55 MHz (estimated)
Counted GLOBAL_POWER_EVENTS events
with a unit mask of 0x01 (mandatory) count 100000
samples % symbol name
66854 6.6725 AllocSetAlloc

patched:
CPU: P4 / Xeon with 2 hyper-threads, speed 2793.55 MHz (estimated)
Counted GLOBAL_POWER_EVENTS events
with a unit mask of 0x01 (mandatory) count 100000
samples % symbol name
47610 4.9333 AllocSetAlloc

I think this patch improves AllocSetAlloc/AllocSetFree performance.

Looks like very good work. Much appreciated.

--
Simon Riggs www.2ndQuadrant.com
PostgreSQL Training, Services and Support

Hi Atsushi,

If size <= 8, fls((size - 1) >> ALLOC_MINBITS) is fls(0).
The result of fls(0) is undefined.

Yep, got caught out by this because my previous fls() supported zero.

I think we have to never call fls(0) from AllocSetFreeIndex().
My proposal code:

if (size > (1 << ALLOC_MINBITS))
{
idx = fls((size - 1) >> ALLOC_MINBITS);
Assert(idx < ALLOCSET_NUM_FREELISTS);
}

Looks good, I'll send an updated patch.

Also, are you still seeing the same improvement with the __builtin_clz
as your inline asm implementation?

Cheers,

Jeremy

Move the shift-and-test login into a separate fls() function, which
can use __builtin_clz() if it's available.

This requires a new check for __builtin_clz in the configure script.

Results in a ~2% performance increase on PowerPC.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>

---
v2: prevent fls(0)

---
configure.in | 13 +++++++++++++
src/backend/utils/mmgr/aset.c | 34 +++++++++++++++++++++++++++-------
2 files changed, 40 insertions(+), 7 deletions(-)

diff --git a/configure.in b/configure.in
index b8d2685..6a317b0 100644
--- a/configure.in
+++ b/configure.in
@@ -1361,6 +1361,19 @@ case $host_os in
 		AC_FUNC_FSEEKO;;
 esac

+# GCC builtins
+#
+# We need AC_TRY_LINK here, as the prototype generated by AC_CHECK_FUNC
+# will cause gcc to try to reference a non-builtin symbol.
+
+AC_MSG_CHECKING([for __builtin_clz])
+AC_TRY_LINK([],
+	    [__builtin_clz(0);],
+	    [AC_DEFINE(HAVE_BUILTIN_CLZ, 1,
+				[Define to 1 if you have __builtin_clz().])
+			AC_MSG_RESULT(yes)],
+		[AC_MSG_RESULT(no)])
+

 #
 # Pthreads
diff --git a/src/backend/utils/mmgr/aset.c b/src/backend/utils/mmgr/aset.c
index 0e2d4d5..9eb3117 100644
--- a/src/backend/utils/mmgr/aset.c
+++ b/src/backend/utils/mmgr/aset.c
@@ -255,6 +255,31 @@ static MemoryContextMethods AllocSetMethods = {
 #define AllocAllocInfo(_cxt, _chunk)
 #endif

+/*
+ * fls: find last set bit.
+ *
+ * Returns the 1-based index of the most-significant bit in x. The MSB
+ * is bit number 32, the LSB is bit number 1. If x is zero, the result is
+ * undefined.
+ */
+static inline int
+fls(unsigned int x)
+{
+#ifdef HAVE_BUILTIN_CLZ
+	return 32 - __builtin_clz(x);
+#else
+	int ls = 0;
+
+	while (x != 0)
+	{
+		ls++;
+		x >>= 1;
+	}
+
+	return ls;
+#endif
+}
+
 /* ----------
  * AllocSetFreeIndex -
  *
@@ -268,14 +293,9 @@ AllocSetFreeIndex(Size size)
 {
 	int			idx = 0;

-	if (size > 0)
+	if (size > (1 << ALLOC_MINBITS))
 	{
-		size = (size - 1) >> ALLOC_MINBITS;
-		while (size != 0)
-		{
-			idx++;
-			size >>= 1;
-		}
+		idx = fls((size - 1) >> ALLOC_MINBITS);
 		Assert(idx < ALLOCSET_NUM_FREELISTS);
 	}

Hi,

Also, are you still seeing the same improvement with the __builtin_clz
as your inline asm implementation?

In my benchmark program, it is a little different performance
in fls implementation and inline asm implementation.
However, the result of a pgbench is almost the same improvement.

Here is the result of my benchmark.

Xeon(Core architecture)
bytes : 4 8 16 32 64 128 256 512 1024 mix
original : 0.780 0.790 0.820 0.870 0.930 0.980 1.040 1.080 1.140 0.910
inline asm: 0.320 0.180 0.190 0.180 0.190 0.180 0.190 0.180 0.190 0.170
fls : 0.270 0.260 0.290 0.290 0.290 0.290 0.290 0.300 0.290 0.380

Xeon(P4 architecrure)
bytes : 4 8 16 32 64 128 256 512 1024 mix
original : 0.520 0.520 0.670 0.780 0.950 1.000 1.060 1.190 1.250 0.940
inline asm: 0.610 0.530 0.530 0.520 0.520 0.540 0.540 0.580 0.540 0.600
fls : 0.390 0.370 0.780 0.780 0.780 0.790 0.780 0.780 0.780 0.520

pgbench result (measured by oprofile)

CPU: Xeon(P4 architecrure)
test program: pgbench -c 1 -t 50000 (fsync=off)

original
samples % symbol name
66854 6.6725 AllocSetAlloc
11817 1.1794 AllocSetFree

inline asm
samples % symbol name
47610 4.9333 AllocSetAlloc
6248 0.6474 AllocSetFree

fls
samples % symbol name
48779 4.9954 AllocSetAlloc
7648 0.7832 AllocSetFree

Best regards,

---
Atsushi Ogawa

Hi Atsushi,

In my benchmark program, it is a little different performance
in fls implementation and inline asm implementation.
However, the result of a pgbench is almost the same improvement.

Here is the result of my benchmark.

Excellent, thank you for getting this extra set of numbers.

Cheers,

Jeremy