[SYCL][NFC] Use unique (or no at all) output files in LIT tests #7872
Conversation
This fixes sporadic issue with `check-sycl` on Windows. Apparently we had a number of tests, which didn't specify `-o %t.out` and it caused some tests to write to the same default `a.exe` and fail with error like: ``` LINK : fatal error LNK1104: cannot open file 'a.exe' ``` To resolve this, switch most of such tests to use `-fsyntax-only` or explicitly added `-o %t.out` to prevent possible races. "bad" tests were discovered using `grep -rnP 'clangxx(?!.*\-o)' sycl/test/`
| @@ -1,4 +1,4 @@ | |||
| // RUN: %clangxx -fsycl -c -fno-color-diagnostics -Xclang -fdump-record-layouts %s | FileCheck %s | |||
| // RUN: %clangxx -fsycl -c -fno-color-diagnostics -Xclang -fdump-record-layouts %s -o %t.out | FileCheck %s | |||
There was a problem hiding this comment.
Note for reviewers: most of abi tests started to fail with -fsyntax-only because output from clang was different. I haven't investigated it and simply decided to switch to -o %t.out instead.
| @@ -1,4 +1,4 @@ | |||
| // RUN: %clangxx -fsycl -c %s | |||
| // RUN: %clangxx -fsycl -fsyntax-only %s | |||
There was a problem hiding this comment.
Note for reviewers: -c still creates an output file, which is a fat object file in case of -fsycl compilation and that temporary file may still cause races. Therefore, I switched -c tests to -fsyntax-only: those tests do not check LLVM IR anyway and -fsyntax-only doesn't create output files, which should make them a tiny bit faster (less I/O)
In addition to that we need an automated check that all tests comply with these requirements.
This seems to be a good starting point, but we need to allow |
Yes, I will prepare some "Testing guidelines" document
Yeah, that would be a good idea. I will try to come up with some script for that. |
|
Failure in unit-tests is a known flaky issue, which is being addressed in #7855 |
|
Hi folks, I've finally prepared a PR to follow-up on testing guidelines, see #8145 |
Use no output file instead of default output file in LIT tests. This occasionally caused errors on windows when multiple files compile to the same output file. Using `-fsyntax-only` hjere also avoids running the full compiler pipeline See #7872 for more context
Use no output file instead of default output file in LIT tests. This occasionally caused errors on windows when multiple files compile to the same output file. Using `-fsyntax-only` hjere also avoids running the full compiler pipeline See intel#7872 for more context
Use no output file instead of default output file in LIT tests. This occasionally caused errors on windows when multiple files compile to the same output file. Using `-fsyntax-only` hjere also avoids running the full compiler pipeline See intel#7872 for more context
This is a joined cherry-pick of #17367, #17362 and #17746 intended to fix a few remaining issues we see in 6.1 nightly validation run. Descriptions of cherry-picked commits: --- [SYCL] Disable launch_policy_neg.cpp on Windows (#17367) Fails to compile on new MSVC. Patch-by: Sarnie, Nick <nick.sarnie@intel.com> --- [SYCL] Avoid testing race condition Use no output file instead of default output file in LIT tests. This occasionally caused errors on windows when multiple files compile to the same output file. Using `-fsyntax-only` hjere also avoids running the full compiler pipeline See #7872 for more context Patch-By: David Garcia Orozco <david.garcia.orozco@intel.com> --- [SYCL] Fix clang-linker-wrapper Windows test (#17746) Failing now because I recently added `ptxas` to the path on these runners. Patch-by: Sarnie, Nick <nick.sarnie@intel.com> ---
This fixes sporadic issue with
check-syclon Windows. Apparently we had a number of tests, which didn't specify-o %t.outand it caused some tests to write to the same defaulta.exeand fail with error like:To resolve this, switch most of such tests to use
-fsyntax-onlyor explicitly added-o %t.outto prevent possible races."bad" tests were discovered using
grep -rnP 'clangxx(?!.*\-o)' sycl/test/