https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91732
Bug ID: 91732
Summary: Adding omp simd pragma prevents vectorization
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: jed at 59A2 dot org
Target Milestone: ---
omp-simd.c:
void poisson(int Q, const double *restrict gsym, const double *restrict du,
double *restrict dv) {
#pragma omp simd
for (int i=0; i<Q; i++) {
const double g[2][2] = {{gsym[Q*0+i], gsym[Q*2+i]},
{gsym[Q*2+i], gsym[Q*1+i]}};
for (int j=0; j<2; j++)
dv[Q*j+i] = g[j][0] * du[Q*0+i] + g[j][1] * du[Q*1+i];
}
}
The above fails to vectorize despite unrolling the inner loop.
$ gcc -Ofast -march=skylake-avx512 -fopenmp -fopt-info -fopt-info-missed -c
omp-simd.c
omp-simd.c:6:5: optimized: loop with 2 iterations completely unrolled (header
execution count 357878152)
omp-simd.c:4:38: missed: couldn't vectorize loop
omp-simd.c:4:18: missed: not vectorized: not suitable for scatter store
D.4095[_37][0][0] = _4;
If I remove the "#pragma omp simd", it vectorizes:
$ gcc -Ofast -march=skylake-avx512 -fopenmp -fopt-info -fopt-info-missed -c
omp-simd.c
omp-simd.c:5:5: optimized: loop with 2 iterations completely unrolled (header
execution count 357878152)
omp-simd.c:2:3: optimized: loop vectorized using 32 byte vectors
omp-simd.c:2:3: optimized: loop versioned for vectorization because of
possible aliasing
omp-simd.c:2:3: optimized: loop with 2 iterations completely unrolled (header
execution count 18709371)
If instead, I replace "#pragma omp simd" with "#pragma GCC ivdep", it
vectorizes without possible aliasing.
$ gcc -Ofast -march=skylake-avx512 -fopenmp -fopt-info -fopt-info-missed -c
omp-simd.c
omp-simd.c:6:5: optimized: loop with 2 iterations completely unrolled (header
execution count 357878152)
omp-simd.c:3:3: optimized: loop vectorized using 32 byte vectors
omp-simd.c:3:3: optimized: loop with 2 iterations completely unrolled (header
execution count 24166268)
I think aliasing should not be a concern due to use of restrict. Also, if I
manually unroll the inner loop (which the compiler is unrolling for me), the
original "omp simd" version vectorizes nicely.
Reproduced on trunk: https://gcc.godbolt.org/z/wKdHg0
