https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99928
Bug ID: 99928
Summary: [OpenMP] reduction variable in combined target
construct wrongly mapped as firstprivate
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Keywords: openmp, wrong-code
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: anlauf at gcc dot gnu.org, jakub at gcc dot gnu.org
Target Milestone: ---
Reported by Harald at
https://gcc.gnu.org/pipermail/fortran/2021-March/055896.html
with a Fortran program. But it also occurs for C as shown below.
OpenMP 5.1 has:
"If a list item appears in a reduction, lastprivate or linear clause
on a combined target construct then it is treated as if it also appears
in a map clause with a map-type of tofrom." (2.21.7)
Likewise OpenMP 5.0 in 2.19.7.
The FE chops the target-teams into target + teams and for target;
's' is implicitly mapped as 'firstprivate' instead of as 'map(tofrom:'.
It seems as if this is best done on the C/C++/Fortran FE which still sees that
it is a combined target construct.
int a[10];
int s2()
{
int s = 0;
#pragma omp target data map(a,s)
{
#pragma omp target teams reduction(+:s)
{
for (int i=0; i < 10; i++)
s += a[i];
}
}
return s;
}
Original dump:
#pragma omp target data map(tofrom:s) map(tofrom:a)
#pragma omp target
#pragma omp teams reduction(+:s)
omplower dump:
#pragma omp target data map(tofrom:s [len: 4]) map(tofrom:a [len: 40])
#pragma omp target num_teams(0) thread_limit(0) firstprivate(s) map(tofrom:a
[len: 40])
#pragma omp teams reduction(+:s) shared(a)