https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87834
Tom de Vries <vries at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |RESOLVED
Resolution|--- |FIXED
Target Milestone|--- |9.0
--- Comment #1 from Tom de Vries <vries at gcc dot gnu.org> ---
This test fails in this mode until r265929. At r265938, this test-case starts
passing. Inbetween, build problems.
The main failure is "libgomp: cuCtxSynchronize error: an illegal memory access
was encountered", and is as mentioned in PR81691 comment 2 not specific to the
target.
Presumably, the root cause for the main failure is fixed by the OpenMP 5.0
changes in r265930.
Consequently, the npvtx libgomp plugin errors following the main failure, be it
either the hang in this PR or the PR81691 "libgomp: cuMemFreeHost
error/libgomp: device finalization failed" errors don't trigger anymore.
Going back to the failing r265929, and applying:
- [nvptx, libgomp] Fix cuMemAlloc with size zero
- [nvptx, libgomp] Fix assert (!s->map->active) in map_fini
- [nvptx, libgomp] Fix map_push
gets us back the main failure:
...
libgomp: cuCtxSynchronize error: an illegal memory access was encountered
FAIL: libgomp.fortran/target2.f90 -O0 execution test
libgomp: cuCtxSynchronize error: an illegal memory access was encountered
FAIL: libgomp.fortran/target2.f90 -O1 execution test
...
So it seems this is a duplicate of one or a combination of the PRs fixed by
those three commits. Marking it resolved-fixed, since I'm not sure which one to
mark it a duplicate of.
