https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100573
Bug ID: 100573
Summary: [OpenMP] 'omp target teams' fails with nvptx and GCN
offloading: FAIL libgomp.c-c++-common/for-3.c +
for-9.c
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Keywords: openmp, wrong-code
Severity: normal
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: jakub at gcc dot gnu.org
Target Milestone: ---
The testcase
libgomp.c-c++-common/for-9.c
is a thin wrapper around libgomp.c-c++-common/for-3.c with slightly different
settings. Hence, it is affected as well.
libgomp.c-c++-common/for-3.c has:
int err = 0;
#pragma omp target teams reduction(|:err)
{
err |= test_d_normal ();
...
}
if (err)
abort ();
This will currently never fail because 'err' is mapped as firstprivate to the
target – at least until PR99928 is fixed (→ patch there).
Alternatively, manually adding a 'map(tofrom: err)' also works.
However, with that 'map added or the PR99928 patch applied:
* it still PASSES on x86-64-gnu-linux (-m64, -m32) with HOST FALLBACK
* BUT it FAILS both with AMD GCN and with NVPTX offloading.
But even without reduction and by checking each of the twenty-odd testcase
separately, it FAILS. Namely, I used the following:
int err = 0, i = 1;
#pragma omp target teams map(tofrom: err)
{
err = test_d_normal ();
}
__builtin_printf("%d -> %d\n", i++, err);
err = 0;
And the result is '-> 1' for all of them.
However, when I use 'omp target' instead of 'target teams', it passes.
