https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102409
Bug ID: 102409
Summary: _pragma ("omp ...") expansion issue - placed in the
wrong scope
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
Target Milestone: ---
Created attachment 51484
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51484&action=edit
Testcase. Compile with "g++ -fopenmp" or with "g++ -E" vs. "clang++ -E" and
inspect "omp ordered" placement
I have to admit that I do not know the fine print of _Pragma and nested macro
expansion. However, the result I get with clang matches what I naively would
expect – while I find the result for GCC odd. I also do not know whether this
use is an extension or matches the C (C++?) standard.
I think it is similar to PR 91669, PR 90400, PR91285, PR 82335
... except that here only the _Pragma that ends up at the wrong spot is the one
which is used in the macro call – while those which are in a #define are
handled properly.
If I read the other PRs correctly, there the issue is the _Pragma in the macro
#define. Nonetheless, they may still fail due to the very same reason.
* * *
In any case, the result which clang gives matches the expectation of the
testcase author
of the test at
https://github.com/clang-ykt/omptests/blob/master/t-for/test.c#L296
Due to GCC's macro/_Pragma expansion, that one (as well as the attached
stripped-down version) fails with:
error: ‘ordered’ region may not be closely nested inside of ‘critical’,
‘ordered’, explicit ‘task’ or ‘taskloop’ region
* * *
The code uses:
#define PARALLEL(X) TEST({ \
...
_Pragma("omp for ordered") \
X \
_Pragma("omp for schedule(auto) ordered") \
X \
})
...
PARALLEL(
for (int i = 0; i < N; i++) { \
_Pragma("omp ordered") \
EXPECTED: same result as with clang++-11 -E -fopenmp:
...
#pragma omp for ordered
for (int i = 0; i < (1024*3); i++) {
#pragma omp ordered
S[0] += C[i] + D[i]; }
#pragma omp for schedule(auto) ordered
for (int i = 0; i < (1024*3); i++) {
#pragma omp ordered
S[0] += C[i] + D[i]; } } }} } }
}
...
BUT: g++ -E -fopenmp gives the following, placing the 'omp ordered' before TEST
instead of inside before "S[0] = " together with the rest of X:
#pragma omp ordered
#pragma omp ordered
{ int fail = 0, trial; for (int trial = 0; trial < (1) && fail == 0;
trial++) {
...
#pragma omp for ordered
for (int i = 0; i < (1024*3); i++) { S[0] += C[i] + D[i]; }
#pragma omp for schedule(auto) ordered
for (int i = 0; i < (1024*3); i++) { S[0] += C[i] + D[i]; } } }} } }
}
