[Openmp-commits] [PATCH] D23115: __kmp_free_task: Fix for serial explicit tasks producing proxy tasks

[Hal Finkel via Openmp-commits]

hfinkel added a comment.

In https://reviews.llvm.org/D23115#504656, @Hahnfeld wrote:

> In https://reviews.llvm.org/D23115#504652, @hfinkel wrote:
>
> > In https://reviews.llvm.org/D23115#504641, @Hahnfeld wrote:
> >
> > > In https://reviews.llvm.org/D23115#504638, @hfinkel wrote:
> > >
> > > > Please add your example as a regression test case.
> > >
> > >
> > > Hmm, that will be tricky: Currently only the Intel Compiler will generate proxy tasks for `target nowait`...
> >
> >
> > In general, I'd still add the test (just with a comment that it had failed using the Intel compiler). I presume that the test suite runs with the Intel compiler too. Since this uses the target directive, however, I'm not sure it will run with Clang currently, so we might need to wait.
>
>
> Hmm, I don't quite see much sense in a test that would not fail with the compilers the tests are most commonly (and automatically) run with.


We probably should have a buildbot for libomp that uses the Intel compiler so long as that's a configuration we don't want to break. Regardless, this is already an issue for lots of different reasons. We have tests that only fail with asserts enabled, we have tests for the JIT that only represent bugs on certain host platforms, etc. A test that everyone can use to verify correct behavior is best, but even if not, there is still value in it.

> Maybe we could reuse the idea of `tasking/kmp_taskloop.c` where the compiler generated code is manually emitted...


That works too.


https://reviews.llvm.org/D23115