https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111661
--- Comment #3 from Patrick Bégou <patrick.be...@univ-grenoble-alpes.fr> --- Hi Tobias, thanks for this information. - yes removing the "finalize" make this small test case working. In my mind it should only remove the allocated attribute from the GPU with setting the count to zero. Is it because the attribute is an allocatable and not a pointer ? Is it the same behaviour with a pointer as attribute ? - unfortunately this modification doesn't make some significant progress with porting my large code (things are much more complex) but with Gnu compilers Gdb is working. So it is a big step for investigating. I have isolated the low level data management to test this module independently. It works with ftn and Nvidia but not with Gnu at this time. I have to investigate. Patrick Le 13/10/2023 à 10:45, burnus at gcc dot gnu.org a écrit : > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111661 > > --- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> --- > @Patrick: It seems to work fine without "finalize". > > Can you check whether the big program then works for you? > Usually, one should be able to do proper ref counting such that a > 'finalize' -> force setting refcounts to zero > shouldn't be needed. > > * * * > > Looking at the code more closely, the problem is: > > #pragma omp target oacc_exit_data map(delete:tab.val.data [len: 88]) > > this tries to 'delete' the array descriptor - but as tab.val.data is part of > 'tab', this deletes all of "tab". > > > Compare the C example: > > struct t { int *a; int n; }; > void f() { > struct t s; > #pragma acc enter data copyin(s.a[:s.n]) > #pragma acc exit data delete(s.a[:s.n]) > // for completeness, not relevant here: > #pragma acc exit data detach(s.a) > #pragma acc exit data delete(s.a) > } > > > GCC does: > > #pragma omp target oacc_enter_data map(struct:s [len: 1]) \ > map(alloc:s.a [len: 8]) map(to:*_4 [len: _3]) map(attach:s.a [bias: 0]) > > #pragma omp target oacc_exit_data map(release:s.a [len: 8]) \ > map(release:*_8 [len: _7]) map(detach:s.a [bias: 0]) > > #pragma omp target oacc_exit_data map(detach:s.a [bias: 0]) > #pragma omp target oacc_exit_data map(release:s.a [len: 8]) > > which seems to be at least consistent. Again, here a 'finalize' would force > the > reference counts to zero and, hence, also delete 's' and not only the > pointee/pointer target *s.a / s.a[0:.n] but also the pointer 's.a' itself. > > (BTW: Same result since GCC 10; GCC 9 rejects that code.) > > * * * > > QUESTION: Is the current code for C (and Fortran) correct according to the > OpenACC specification or not? > > FOLLOW UP QUESTION: If GCC's result is incorrect, what should the compiler do > instead? > And if it is correct, the question is: why do both ftn and nvfortran work in > the same way? >
