https://gcc.gnu.org/g:09561afe3ed2335d3ff6ba7c8ceb927e9223fff5
commit r17-5501-g09561afe3ed2335d3ff6ba7c8ceb927e9223fff5 Author: Thomas Schwinge <[email protected]> Date: Tue Jul 18 16:58:15 2023 +0200 libgomp: Handle OpenMP's reverse offloads: For NVIDIA GPU code offloading, shorten 'libgomp.fortran/reverse-offload-5.f90' timeout to 10 s Ever since 'libgomp.fortran/reverse-offload-5.f90' had gotten added in commit ea4b23d9c82d9be3b982c3519fe5e8e9d833a6a8 "libgomp: Handle OpenMP's reverse offloads", we ran into a deadlock for NVIDIA GPU code offloading (likely to be qualified as pre-existing implementation deficiency in NVIDIA GPU code offloading), and that stalled libgomp testing until DejaGnu's timeout handler eventually kicked in. Shorten that timeout to 10 s. libgomp/ * testsuite/libgomp.fortran/reverse-offload-5.f90: For NVIDIA GPU code offloading, shorten timeout to 10 s. Co-authored-by: Thomas Schwinge <[email protected]> Diff: --- libgomp/testsuite/libgomp.fortran/reverse-offload-5.f90 | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/libgomp/testsuite/libgomp.fortran/reverse-offload-5.f90 b/libgomp/testsuite/libgomp.fortran/reverse-offload-5.f90 index 16810eb47deb..d44600b53a1e 100644 --- a/libgomp/testsuite/libgomp.fortran/reverse-offload-5.f90 +++ b/libgomp/testsuite/libgomp.fortran/reverse-offload-5.f90 @@ -1,5 +1,9 @@ ! { dg-additional-options "-foffload-options=nvptx-none=-misa=sm_35" { target { offload_target_nvptx } } } ! { dg-xfail-run-if "Copying on-device allocated memory fails with cuMemcpyDtoHAsync error: invalid argument" { offload_device_nvptx } } +! ..., and upon process termination, we then deadlock during device shutdown, +! as the device kernel is still running, waiting for the reverse offload to +! complete. Until that's addressed properly: +! { dg-timeout 10 { target offload_device_nvptx } } ... to make sure that happens quickly. ! Because of the nvptx fail, a non-device alloc version has been ! created: reverse-offload-5a.f90
