https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127849
Bug ID: 127849
Summary: different vectorization of AND reduction for a
single-bit set initial value
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: manu at gcc dot gnu.org
Target Milestone: ---
Split from PR127824.
These two functions should produce the same vectorization, but the version with
unsigned seems worse?
__attribute__((optimize("no-trapping-math")))
bool
weakly_dominates_2(const double * restrict a, const double * restrict b, const
unsigned dim)
{
unsigned a_leq_b = 1;
for (unsigned d = 0; d < dim; d++)
a_leq_b &= (a[d] <= b[d]);
return (bool) a_leq_b;
}
__attribute__((optimize("no-trapping-math")))
bool
weakly_dominates_1(const double * restrict a, const double * restrict b, const
unsigned dim)
{
bool a_leq_b = true;
for (unsigned d = 0; d < dim; d++)
a_leq_b &= (a[d] <= b[d]);
return a_leq_b;
}
With -O3 -march=x86-64-v3 -fopt-info-vec,
weakly_dominates_2 produces:
vmovupd ymm0, YMMWORD PTR [rdi+rax]
vmovupd ymm1, YMMWORD PTR [rdi+32+rax]
vcmplepd ymm0, ymm0, YMMWORD PTR [rsi+rax]
vcmplepd ymm1, ymm1, YMMWORD PTR [rsi+32+rax]
add rax, 64
vpand ymm1, ymm4, ymm1
vpand ymm0, ymm4, ymm0
vperm2i128 ymm2, ymm0, ymm1, 32
vperm2i128 ymm0, ymm0, ymm1, 49
vpshufd ymm0, ymm0, 216
vpshufd ymm1, ymm2, 216
vpunpcklqdq ymm0, ymm1, ymm0
vpand ymm3, ymm3, ymm0
while weakly_dominates_1 produces:
vmovupd ymm0, YMMWORD PTR [rdi+rax]
vcmplepd ymm0, ymm0, YMMWORD PTR [rsi+rax]
add rax, 32
vpand ymm1, ymm1, ymm0
See https://godbolt.org/z/rWccjdMvv
