LLVM: Enable OpenMP on aarch64

[dim]

Summary

LLVM: Enable OpenMP on aarch64

It just works™

Feel free to upstream the patch to LLVM (probably with `LINUX||FREEBSD` instead of removing the OS check completely though).

PR: 248864
Differential_Revision: D21167
Reviewed_by: ??

[emaste]

contrib/openmp/runtime/src/kmp.h

3469

Came in from "aarch64 port sent by C. Bergstrom"
http://llvm.org/viewvc/llvm-project?view=revision&revision=225792
but this just added KMP_AARCH_64 to the existing ARM/X86_64 cases, KMP_OS_LINUX was already there

is there something FreeBSD-aarch64-specific here?

[val_packett.cool]

contrib/openmp/runtime/src/kmp.h

3469

The va_list usage in the else cases (which treats it as if it is already a pointer) does not compile on FreeBSD/aarch64, it needs to be in the same case as Linux/(aarch64|amd64|arm), which treats it as opaque.

This is just another "upstream only considered Linux, we need to be included in the same way as Linux" case.

[emaste]

This is just another "upstream only considered Linux, we need to be included in the same way as Linux" case.

But I mean this was already the case for the presumably-working amd64/i386 cases?

[val_packett.cool]

In D21167#463001, @emaste wrote:

This is just another "upstream only considered Linux, we need to be included in the same way as Linux" case.

But I mean this was already the case for the presumably-working amd64/i386 cases?

FreeBSD/amd64 works anyway because va_list *is* a pointer there

[val_packett.cool]

Rebased on top of the powerpc64 change

[brooks]

I've submitted this upstream https://reviews.llvm.org/D77333

[mikael]

ping?

[dim]

I'll have a look at this now. I was not aware of this review at all.

contrib/openmp/runtime/src/kmp.h

3469

This would typically not be acceptable for upstream. The condition should at least be rewritten as:

#if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && (KMP_OS_LINUX | KMP_OS_FREEBSD)

It should probably be defining something like KMP_HAVE_WEIRD_VA_LIST so the conditional expression does not have to be changed in many places.

[dim]

I've submitted an upstream review here: https://reviews.llvm.org/D86397

With this and the src.opts.mk part, a buildworld for aarch64 completed successfully, although I have no way to determine whether the produced libomp.so works at all.

[dim]

Thanks for this review, it's now effectively enabled via rS364732 for the change that went into upstream, and rS364733 to change src.opts.mk.