https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108916
Bug ID: 108916
Summary: Miss vectorization for masked gather w/o restrict
qualifier.
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: crazylht at gmail dot com
Target Milestone: ---
void
foo1 (double* __restrict a, long long* b, double* c, long long* d)
{
for (long long i = 0; i != 10000; i++)
{
if (b[i])
a[i] = 0;
else
a[i] = c[d[i]];
}
}
void
foo2 (double* __restrict a, long long* b, double* __restrict c, long long* d)
{
for (long long i = 0; i != 10000; i++)
{
if (b[i])
a[i] = 0;
else
a[i] = c[d[i]];
}
}
w/ -Ofast -march=skylake-avx512, foo1 can't be vectorized, but foo2 can, the
only difference between foo1 and foo2 is restrict qualifier for double* c.
But there's already a restrict qualifier for double* __restrict a in both
cases, why double* __restrict c matters.
