Re: [PATCH] x86: Add optimized popcnt variants

From: Borislav Petkov
Date: Tue Feb 23 2010 - 10:58:49 EST

From: "H. Peter Anvin" <hpa@xxxxxxxxx>
Date: Mon, Feb 22, 2010 at 11:55:48AM -0800

> On 02/22/2010 10:49 AM, Borislav Petkov wrote:
> >>
> >> Just a note: this still means rdi is clobbered on x86-64, which is
> >> probably fine, but needs to be recorded as such. Since gcc doesn't
> >> support clobbers for registers used as operands (sigh), you have to
> >> create a dummy output and assign it a "=D" constraint.
> >>
> >> I don't know if gcc would handle -fcall-saved-rdi here... and if so, how
> >> reliably.

Hmm, we cannot do that with the current design since __arch_hweight64
is being inlined into every callsite and AFAICT we would have to build
every callsite with "-fcall-saved-rdi" which is clearly too much. The
explicit "=D" dummy constraint is straightforward, instead.

> > Ok, from looking at kernel/sched.s output it looks like it saves rdi
> > content over the alternative where needed. I'll do some more testing
> > just to make sure.
> >
>
> No, you can't rely on behavioral observation. A different version of
> gcc could behave differently. We need to make sure we tell gcc what the
> requirements actually are, as opposed to thinking we can just fix them.

Just to make clear: those observations were referring to the version
_with_ the dummy "=D" constraint - I was simply verifying the asm
output and wasn't relying on behavioral observation.

> +#define POPCNT ".byte 0xf3\n\t.byte 0x48\n\t.byte 0x0f\n\t.byte
> 0xb8\n\t.byte 0xc7"
>
> BTW, this can be written:
>
> #define POPCNT ".byte 0xf3,0x48,0x0f,0xb8,0xc7"

Done, here's the latest version, it boots fine and testing is ongoing:

--
From: Borislav Petkov <borislav.petkov@xxxxxxx>
Date: Thu, 11 Feb 2010 00:48:31 +0100
Subject: [PATCH] x86: Add optimized popcnt variants

Add support for the hardware version of the Hamming weight function,
popcnt, present in CPUs which advertize it under CPUID, Function
0x0000_0001_ECX[23]. On CPUs which don't support it, we fallback to the
default lib/hweight.c sw versions.

A synthetic benchmark comparing popcnt with __sw_hweight64 showed almost
a 3x speedup on a F10h machine.

Signed-off-by: Borislav Petkov <borislav.petkov@xxxxxxx>
---
arch/x86/Kconfig | 5 ++
arch/x86/include/asm/alternative.h | 9 +++-
arch/x86/include/asm/arch_hweight.h | 61 +++++++++++++++++++++++++++++
arch/x86/include/asm/bitops.h | 4 +-
include/asm-generic/bitops/arch_hweight.h | 22 ++++++++--
lib/Makefile | 3 +
lib/hweight.c | 20 +++++-----
scripts/Makefile.lib | 4 ++
8 files changed, 110 insertions(+), 18 deletions(-)
create mode 100644 arch/x86/include/asm/arch_hweight.h

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index eb40925..176950e 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -230,6 +230,11 @@ config X86_32_LAZY_GS
def_bool y
depends on X86_32 && !CC_STACKPROTECTOR

+config ARCH_HWEIGHT_CFLAGS
+ string
+ default "-fcall-saved-ecx -fcall-saved-edx" if X86_32
+ default "-fcall-saved-rsi -fcall-saved-rdx -fcall-saved-rcx -fcall-saved-r8 -fcall-saved-r9 -fcall-saved-r10 -fcall-saved-r11" if X86_64
+
config KTIME_SCALAR
def_bool X86_32
source "init/Kconfig"
diff --git a/arch/x86/include/asm/alternative.h b/arch/x86/include/asm/alternative.h
index 69b74a7..0720c96 100644
--- a/arch/x86/include/asm/alternative.h
+++ b/arch/x86/include/asm/alternative.h
@@ -39,9 +39,6 @@
#define LOCK_PREFIX ""
#endif

-/* This must be included *after* the definition of LOCK_PREFIX */
-#include <asm/cpufeature.h>
-
struct alt_instr {
u8 *instr; /* original instruction */
u8 *replacement;
@@ -91,6 +88,12 @@ static inline void alternatives_smp_switch(int smp) {}
".previous"

/*
+ * This must be included *after* the definition of ALTERNATIVE due to
+ * <asm/arch_hweight.h>
+ */
+#include <asm/cpufeature.h>
+
+/*
* Alternative instructions for different CPU types or capabilities.
*
* This allows to use optimized instructions even on generic binary
diff --git a/arch/x86/include/asm/arch_hweight.h b/arch/x86/include/asm/arch_hweight.h
new file mode 100644
index 0000000..eaefadc
--- /dev/null
+++ b/arch/x86/include/asm/arch_hweight.h
@@ -0,0 +1,61 @@
+#ifndef _ASM_X86_HWEIGHT_H
+#define _ASM_X86_HWEIGHT_H
+
+#ifdef CONFIG_64BIT
+/* popcnt %rdi, %rax */
+#define POPCNT ".byte 0xf3,0x48,0x0f,0xb8,0xc7"
+#define REG_IN "D"
+#define REG_OUT "a"
+#else
+/* popcnt %eax, %eax */
+#define POPCNT ".byte 0xf3,0x0f,0xb8,0xc0"
+#define REG_IN "a"
+#define REG_OUT "a"
+#endif
+
+/*
+ * __sw_hweightXX are called from within the alternatives below
+ * and callee-clobbered registers need to be taken care of. See
+ * ARCH_HWEIGHT_CFLAGS in <arch/x86/Kconfig> for the respective
+ * compiler switches.
+ */
+static inline unsigned int __arch_hweight32(unsigned int w)
+{
+ unsigned int res = 0;
+
+ asm (ALTERNATIVE("call __sw_hweight32", POPCNT, X86_FEATURE_POPCNT)
+ : "="REG_OUT (res)
+ : REG_IN (w));
+
+ return res;
+}
+
+static inline unsigned int __arch_hweight16(unsigned int w)
+{
+ return __arch_hweight32(w & 0xffff);
+}
+
+static inline unsigned int __arch_hweight8(unsigned int w)
+{
+ return __arch_hweight32(w & 0xff);
+}
+
+static inline unsigned long __arch_hweight64(__u64 w)
+{
+ unsigned long res = 0;
+ /* tell gcc that %rdi is clobbered as an input operand */
+ unsigned long dummy;
+
+#ifdef CONFIG_X86_32
+ return __arch_hweight32((u32)w) +
+ __arch_hweight32((u32)(w >> 32));
+#else
+ asm (ALTERNATIVE("call __sw_hweight64", POPCNT, X86_FEATURE_POPCNT)
+ : "="REG_OUT (res), "="REG_IN (dummy)
+ : REG_IN (w));
+#endif /* CONFIG_X86_32 */
+
+ return res;
+}
+
+#endif
diff --git a/arch/x86/include/asm/bitops.h b/arch/x86/include/asm/bitops.h
index 02b47a6..545776e 100644
--- a/arch/x86/include/asm/bitops.h
+++ b/arch/x86/include/asm/bitops.h
@@ -444,7 +444,9 @@ static inline int fls(int x)

#define ARCH_HAS_FAST_MULTIPLIER 1

-#include <asm-generic/bitops/hweight.h>
+#include <asm/arch_hweight.h>
+
+#include <asm-generic/bitops/const_hweight.h>

#endif /* __KERNEL__ */

diff --git a/include/asm-generic/bitops/arch_hweight.h b/include/asm-generic/bitops/arch_hweight.h
index 3a7be84..9a81c1e 100644
--- a/include/asm-generic/bitops/arch_hweight.h
+++ b/include/asm-generic/bitops/arch_hweight.h
@@ -3,9 +3,23 @@

#include <asm/types.h>

-extern unsigned int __arch_hweight32(unsigned int w);
-extern unsigned int __arch_hweight16(unsigned int w);
-extern unsigned int __arch_hweight8(unsigned int w);
-extern unsigned long __arch_hweight64(__u64 w);
+inline unsigned int __arch_hweight32(unsigned int w)
+{
+ return __sw_hweight32(w);
+}

+inline unsigned int __arch_hweight16(unsigned int w)
+{
+ return __sw_hweight16(w);
+}
+
+inline unsigned int __arch_hweight8(unsigned int w)
+{
+ return __sw_hweight8(w);
+}
+
+inline unsigned long __arch_hweight64(__u64 w)
+{
+ return __sw_hweight64(w);
+}
#endif /* _ASM_GENERIC_BITOPS_HWEIGHT_H_ */
diff --git a/lib/Makefile b/lib/Makefile
index 3b0b4a6..e2ad17c 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -39,7 +39,10 @@ lib-$(CONFIG_RWSEM_XCHGADD_ALGORITHM) += rwsem.o
lib-$(CONFIG_GENERIC_FIND_FIRST_BIT) += find_next_bit.o
lib-$(CONFIG_GENERIC_FIND_NEXT_BIT) += find_next_bit.o
obj-$(CONFIG_GENERIC_FIND_LAST_BIT) += find_last_bit.o
+
+CFLAGS_hweight.o = $(subst $(quote),,$(CONFIG_ARCH_HWEIGHT_CFLAGS))
obj-$(CONFIG_GENERIC_HWEIGHT) += hweight.o
+
obj-$(CONFIG_LOCK_KERNEL) += kernel_lock.o
obj-$(CONFIG_DEBUG_PREEMPT) += smp_processor_id.o
obj-$(CONFIG_DEBUG_LIST) += list_debug.o
diff --git a/lib/hweight.c b/lib/hweight.c
index 9ff86df..f9ce440 100644
--- a/lib/hweight.c
+++ b/lib/hweight.c
@@ -9,7 +9,7 @@
* The Hamming Weight of a number is the total number of bits set in it.
*/

-unsigned int __arch_hweight32(unsigned int w)
+unsigned int __sw_hweight32(unsigned int w)
{
unsigned int res = w - ((w >> 1) & 0x55555555);
res = (res & 0x33333333) + ((res >> 2) & 0x33333333);
@@ -17,30 +17,30 @@ unsigned int __arch_hweight32(unsigned int w)
res = res + (res >> 8);
return (res + (res >> 16)) & 0x000000FF;
}
-EXPORT_SYMBOL(__arch_hweight32);
+EXPORT_SYMBOL(__sw_hweight32);

-unsigned int __arch_hweight16(unsigned int w)
+unsigned int __sw_hweight16(unsigned int w)
{
unsigned int res = w - ((w >> 1) & 0x5555);
res = (res & 0x3333) + ((res >> 2) & 0x3333);
res = (res + (res >> 4)) & 0x0F0F;
return (res + (res >> 8)) & 0x00FF;
}
-EXPORT_SYMBOL(__arch_hweight16);
+EXPORT_SYMBOL(__sw_hweight16);

-unsigned int __arch_hweight8(unsigned int w)
+unsigned int __sw_hweight8(unsigned int w)
{
unsigned int res = w - ((w >> 1) & 0x55);
res = (res & 0x33) + ((res >> 2) & 0x33);
return (res + (res >> 4)) & 0x0F;
}
-EXPORT_SYMBOL(__arch_hweight8);
+EXPORT_SYMBOL(__sw_hweight8);

-unsigned long __arch_hweight64(__u64 w)
+unsigned long __sw_hweight64(__u64 w)
{
#if BITS_PER_LONG == 32
- return __arch_hweight32((unsigned int)(w >> 32)) +
- __arch_hweight32((unsigned int)w);
+ return __sw_hweight32((unsigned int)(w >> 32)) +
+ __sw_hweight32((unsigned int)w);
#elif BITS_PER_LONG == 64
#ifdef ARCH_HAS_FAST_MULTIPLIER
w -= (w >> 1) & 0x5555555555555555ul;
@@ -57,4 +57,4 @@ unsigned long __arch_hweight64(__u64 w)
#endif
#endif
}
-EXPORT_SYMBOL(__arch_hweight64);
+EXPORT_SYMBOL(__sw_hweight64);
diff --git a/scripts/Makefile.lib b/scripts/Makefile.lib
index f9bdf26..cbcd654 100644
--- a/scripts/Makefile.lib
+++ b/scripts/Makefile.lib
@@ -245,3 +245,7 @@ quiet_cmd_lzo = LZO $@
cmd_lzo = (cat $(filter-out FORCE,$^) | \
lzop -9 && $(call size_append, $(filter-out FORCE,$^))) > $@ || \
(rm -f $@ ; false)
+
+# misc stuff
+# ---------------------------------------------------------------------------
+quote:="
--
1.6.4.2


--
Regards/Gruss,
Boris.

--
Advanced Micro Devices, Inc.
Operating Systems Research Center
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/