https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91824
Bug ID: 91824
Summary: unnecessary sign-extension after _mm_movemask_epi8 or
__builtin_popcount
Product: gcc
Version: 9.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: vanyacpp at gmail dot com
Target Milestone: ---
gcc -O2 -mpopcnt leaves unnecessary cdqe:
#include <cstdint>
#include <emmintrin.h>
void f(uint64_t& val, __m128i mask)
{
val += __builtin_popcount(_mm_movemask_epi8(mask));
}
void g(uint64_t& val, __m128i mask)
{
val += __builtin_popcountll(_mm_movemask_epi8(mask));
}
f:
pmovmskb eax, xmm0
popcnt eax, eax
cdqe
add QWORD PTR [rdi], rax
ret
g:
pmovmskb eax, xmm0
cdqe
popcnt rax, rax
add QWORD PTR [rdi], rax
ret
Both cdqe are unnecessary, because the results of both pmovmskb and
__builtin_popcount can not be negative.
Only lower 16 bits of pmovmskb can be non-zero. And the image of popcnt is
either [0..32] or [0..64] depending on the argument.
