https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81920
Bug ID: 81920
Summary: Improve timing-sensitive libgomp.oacc testcases
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: testsuite
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
Target Milestone: ---
There is a number of libgomp.oacc testcases that tests timing behaviour, and
fails on and off. In the .sum file they show up as execution failure, which
normally is associated with a wrong-code bug:
...
FAIL: libgomp.oacc-c++/../libgomp.oacc-c-c++-common/lib-81.c
-DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0
execution test
...
In the .log file we find evidence that in fact it's a timing issue, but the
message doesn't tell us what is the actual and expected time:
...
'./lib-81.exe'; echo XYZ$?ZYX^M
actual time too long^M
Aborted (core dumped)^M
XYZ134ZYX^M
OK>
Elapsed time: 1031 ms
FAIL: libgomp.oacc-c++/../libgomp.oacc-c-c++-common/lib-81.c
-DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -fofflo\
ad=nvptx-none -O0 execution test
...
1. These should fail in output testing, not in execution, making clear in the
.sum file we're dealing with a timing issue.
2. The error message in the .log file should give details about actual and
expected time.
3. Having more detailed reporting may give us a better idea whether we need to
recalibrate, or whether there's an underlying issue that really needs fixing.
