https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70154
Bug ID: 70154
Summary: [openacc] failing alias analysis in oacc kernels
region
Product: gcc
Version: 6.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
Target Milestone: ---
Created attachment 37908
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37908&action=edit
kernels-alias-ipa-pta-fail.c, to be run in gcc/testsuite/c-c++-common/goacc/
parloops fails to parallelize this testcase.
After ipa-pta, we have:
...
# VUSE <.MEM_2>
# PT = nonlocal unit-escaped { D.1869 } (nonlocal, escaped)
_5 = MEM[(const struct .omp_data_t.2D.1887 &).omp_data_i_4(D) clique 1 base
1].aD.1893;
# .MEM_6 = VDEF <.MEM_2>
*_5[i_1] = i_1;
# VUSE <.MEM_6>
# PT = nonlocal unit-escaped { D.1870 } (nonlocal, escaped)
_7 = MEM[(const struct .omp_data_t.2D.1887 &).omp_data_i_4(D) clique 1 base
1].bD.1891;
# .MEM_8 = VDEF <.MEM_6>
*_7[i_1] = i_1;
i_9 = i_1 + 1;
goto <bb 3>;
...
If I remove the parallel region, I get this instead (so no more 'nonlocal
unit-escaped'), and parallelization succeeds:
...
# PT = { D.1869 } (nonlocal, escaped)
...
# PT = { D.1870 } (nonlocal, escaped)
...
The same happens if I instead remove the data enter/exit bits.
My first suspicion was that the oacc internal fns were missing certain
annotations that should help ipa-tpa. But the fact that the failure doesn't
seem to be tied to either the parallel construct, or the data enter/exit
construct makes me suspect that I'm running into some ipa-pta limitation.
