https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91539
Bug ID: 91539
Summary: #pragma omp simd disables -ffp-contract=fast
Product: gcc
Version: 9.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: dmatthews at utexas dot edu
Target Milestone: ---
When compiling this program:
#include <math.h>
void foo(int n,
const double* restrict a,
const double* restrict b,
const double* restrict c,
double* restrict d)
{
#pragma omp simd
for (int i = 0;i < n;i++)
d[i] = a[i]*b[i]+c[i];
}
void bar(int n,
const double* restrict a,
const double* restrict b,
const double* restrict c,
double* restrict d)
{
#pragma omp simd
for (int i = 0;i < n;i++)
d[i] = fma(a[i],b[i],c[i]);
}
with gcc 9.2.0 and the options "-std=c99 -fopenmp-simd -O3 -march=core-avx2",
the "bar" function generates FMA instructions but the "foo" function does not.
Explicitly adding "-ffp-contract=fast" restores FMA in "foo". According to the
documentation, "-ffp-contract=fast" should be the default, and indeed
autovectorization (not using the #pragma) will generate FMA instructions
without explicitly specifying this option.
This issue is the same as the optimization issue reported in bug 91511, but is
*not* reflected in bug 61727 of which the former is marked a duplicate.
