With this config:
http://busybox.net/~vda/kernel_config_OPTIMIZE_INLINING_and_Os
gcc-4.7.2 generates many copies of these tiny functions:
__arch_hweight32 (35 copies):
55 push %rbp
e8 66 9b 4a 00 callq __sw_hweight32
48 89 e5 mov %rsp,%rbp
5d pop %rbp
c3 retq
__arch_hweight64 (8 copies):
55 push %rbp
e8 5e c2 8a 00 callq __sw_hweight64
48 89 e5 mov %rsp,%rbp
5d pop %rbp
c3 retq
See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66122
This patch fixes this via s/inline/__always_inline/
To avoid touching 32-bit case where such change was not tested
to be a win, reformat __arch_hweight64() to have completely
disjoint 64-bit and 32-bit implementations. IOW: made #ifdef /
32 bits and 64 bits instead of having #ifdef / #else / #endif
inside a single function body. Only 64-bit __arch_hweight64() is
__always_inline'd.
text data bss dec filename
86971120 17195912 36659200 140826232 vmlinux.before
86970954 17195912 36659200 140826066 vmlinux
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Thomas Graf <tgraf@suug.ch>
Cc: linux-kernel@vger.kernel.org
Link: http://lkml.kernel.org/r/1438697716-28121-2-git-send-email-dvlasenk@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Use a 32-bit popcnt instruction for __arch_hweight32(), even on
x86-64. Even though the input register will *usually* be
zero-extended due to the standard operation of the hardware, it isn't
necessarily so if the input value was the result of truncating a
64-bit operation.
Note: the POPCNT32 variant used on x86-64 has a technically
unnecessary REX prefix to make it five bytes long, the same as a CALL
instruction, therefore avoiding an unnecessary NOP.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Cc: Borislav Petkov <borislav.petkov@amd.com>
LKML-Reference: <alpine.LFD.2.00.1005171443060.4195@i5.linux-foundation.org>
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@amd.com>
LKML-Reference: <20100318112015.GC11152@aftab>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Denys Vlasenko