On Sun, Jan 03, 2016 at 05:56:41PM -0300, James Almer wrote:
[...]
> >> +static av_always_inline av_const int ff_parity(uint32_t v)
> >> +{
> >> +#if HAVE_PARITY
> >> + return __builtin_parity(v);
> >> +#else
> >> + return av_popcount(v) & 1;
> >> +#endif
> >
> > Do compilers really generate better code for the former?
>
> GCC does on x86 when the target cpu doesn't support the popcnt instruction,
> otherwise the end result would be the same (popcnt + and).
> av_popcount_c() is not optimal for this.
>
For the record, this is what it looks like here (GCC 5.3.0, clang 3.7.0,
i5-5250U):
[/tmp]☭ cat a.c
#include <stdint.h>
int parity0(uint32_t x) { return __builtin_popcount(x) & 1; }
int parity1(uint32_t x) { return __builtin_parity(x); }
[/tmp]☭ gcc -O2 -c a.c && objdump -r -d -Mintel a.o
a.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <parity0>:
0: 48 83 ec 08 sub rsp,0x8
4: 89 ff mov edi,edi
6: e8 00 00 00 00 call b <parity0+0xb>
7: R_X86_64_PC32 __popcountdi2-0x4
b: 48 83 c4 08 add rsp,0x8
f: 83 e0 01 and eax,0x1
12: c3 ret
13: 0f 1f 00 nop DWORD PTR [rax]
16: 66 2e 0f 1f 84 00 00 nop WORD PTR cs:[rax+rax*1+0x0]
1d: 00 00 00
0000000000000020 <parity1>:
20: 89 f8 mov eax,edi
22: c1 ef 10 shr edi,0x10
25: 31 f8 xor eax,edi
27: 30 e0 xor al,ah
29: 0f 9b c0 setnp al
2c: 0f b6 c0 movzx eax,al
2f: c3 ret
[/tmp]☭ clang -O2 -c a.c && objdump -r -d -Mintel a.o
a.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <parity0>:
0: 89 f8 mov eax,edi
2: d1 e8 shr eax,1
4: 25 55 55 55 55 and eax,0x55555555
9: 29 c7 sub edi,eax
b: 89 f8 mov eax,edi
d: 25 33 33 33 33 and eax,0x33333333
12: c1 ef 02 shr edi,0x2
15: 81 e7 33 33 33 33 and edi,0x33333333
1b: 01 c7 add edi,eax
1d: 89 f8 mov eax,edi
1f: c1 e8 04 shr eax,0x4
22: 01 f8 add eax,edi
24: 25 0f 0f 0f 01 and eax,0x10f0f0f
29: 69 c0 01 01 01 01 imul eax,eax,0x1010101
2f: c1 e8 18 shr eax,0x18
32: 83 e0 01 and eax,0x1
35: c3 ret
36: 66 2e 0f 1f 84 00 00 nop WORD PTR cs:[rax+rax*1+0x0]
3d: 00 00 00
0000000000000040 <parity1>:
40: 89 f8 mov eax,edi
42: d1 e8 shr eax,1
44: 25 55 55 55 55 and eax,0x55555555
49: 29 c7 sub edi,eax
4b: 89 f8 mov eax,edi
4d: 25 33 33 33 33 and eax,0x33333333
52: c1 ef 02 shr edi,0x2
55: 81 e7 33 33 33 33 and edi,0x33333333
5b: 01 c7 add edi,eax
5d: 89 f8 mov eax,edi
5f: c1 e8 04 shr eax,0x4
62: 01 f8 add eax,edi
64: 25 0f 0f 0f 01 and eax,0x10f0f0f
69: 69 c0 01 01 01 01 imul eax,eax,0x1010101
6f: c1 e8 18 shr eax,0x18
72: 83 e0 01 and eax,0x1
75: c3 ret
[/tmp]☭
Conclusion: with GCC it matters, not so much with Clang.
--
Clément B.
signature.asc
Description: PGP signature
_______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
