Files
Justine Tunney 87d7010495 Improve performance of bitscanning intrinsics
This change helps spectre more intelligently plan execution, by working
around false output dependencies, impacting ops like popcnt bsr and bsf
2021-06-15 06:29:51 -07:00

25 lines
1.0 KiB
C

#ifndef COSMOPOLITAN_LIBC_BITS_POPCNT_H_
#define COSMOPOLITAN_LIBC_BITS_POPCNT_H_
#include "libc/nexgen32e/x86feature.h"
#if !(__ASSEMBLER__ + __LINKER__ + 0)
COSMOPOLITAN_C_START_
unsigned long popcnt(unsigned long) pureconst;
#if defined(__GNUC__) && !defined(__STRICT_ANSI__)
#define popcnt(X) \
(__builtin_constant_p(X) ? __builtin_popcountll(X) : ({ \
unsigned long PoP = (X); \
if (X86_HAVE(POPCNT)) { \
asm("popcnt\t%0,%0" : "+r"(PoP) : /* no inputs */ : "cc"); \
} else { \
PoP = (popcnt)(PoP); \
} \
PoP; \
}))
#endif /* GNUC && !ANSI */
COSMOPOLITAN_C_END_
#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */
#endif /* COSMOPOLITAN_LIBC_BITS_POPCNT_H_ */