https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93554
Tobias Burnus <burnus at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |burnus at gcc dot gnu.org
--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> ---
The analogous OpenMP variant [omp target / omp do private(x)] works.
The problem seems to be that "private(x)" implies that the data is device
private and, hence, it is finalized after the loop and before returning to the
host. – But the OpenACC code expects that directly after the loop the "omp
return" (internal) pragma comes.
* * *
The code flow is:
expand_omp_for
-> t (for OpenACC)
-> expand_omp_for_static_nochunk (for OpenMP)
i.e. the input in those functions is the same.
* * *
One has:
debug_bb(region->entry):
<bb 6> :
#pragma omp for private(i) private(x)
for (i = 0; i <= 31; i = i + 1)
debug_bb(region->cont):
<bb 7> :
x.a = 1;
#pragma omp continue (i, i)
debug_bb(region->exit):
<bb 10> :
#pragma omp return
HOWEVER:
BRANCH_EDGE (entry_bb)->dest
is the code after the region->exit (autodeallocation of the allocatable
component x%b); namely "debug_bb((*region->entry->succs)[1]->dest)"
[with …[0]->flag == 1 == EDGE_FALLTHRU]:
D.3967 = x.b.data;
if (D.3967 != 0B)
goto <bb 9>; [INV]
else
goto <bb 10>; [INV]
Hence, the second condition of the assert fails for OpenACC:
gcc_assert (EDGE_COUNT (entry_bb->succs) == 2
&& BRANCH_EDGE (entry_bb)->dest == exit_bb);
and likewise the later
gcc_assert (FALLTHRU_EDGE (cont_bb)->dest == exit_bb);
In OpenMP, the result is the same but there one operates checks:
BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest
[OpenMP: "bool broken_loop = region->cont == NULL", which is similar to the
handling of cont == NULL in the OpenACC code.]
