https://gcc.gnu.org/bugzilla/show_bug.cgi?id=29776
--- Comment #25 from Adam Warner <adam.warner.nz at gmail dot com> ---
Documenting a workaround I've found for the unnecessary sign extension. I'm
still perplexed at the improbability of this appearing to work!
workaround_bsr_sign_extension.c:
#include <stdint.h>
uint64_t bsr_u64(uint64_t a) {
if (sizeof(unsigned long) == 8) return 63 - __builtin_clzl(a);
if (sizeof(unsigned long long) == 8) return 63 - __builtin_clzll(a);
}
uint64_t bsr_u64_alt(uint64_t a) {
if (sizeof(unsigned long) == 8) return UINT64_C(63) - __builtin_clzl(a);
if (sizeof(unsigned long long) == 8) return UINT64_C(63) -
__builtin_clzll(a);
}
int main(void) {return 0;}
$ gcc -O3 workaround_bsr_sign_extension.c && objdump -d -m i386:x86-64:intel
a.out|less
Relevant output:
0000000000001140 <bsr_u64>:
1140: 48 0f bd c7 bsr rax,rdi
1144: 48 98 cdqe
1146: c3 ret
1147: 66 0f 1f 84 00 00 00 nop WORD PTR [rax+rax*1+0x0]
114e: 00 00
0000000000001150 <bsr_u64_alt>:
1150: 48 0f bd c7 bsr rax,rdi
1154: c3 ret
In the alternative implementation the superfluous 32- to 64-bit sign extension
instruction CDQE is no longer generated.
