[Openmp-commits] [llvm] [openmp] [openmp] Add support for arm64ec/arm64x to libomp (PR #175988)

[Martin Storsjö via Openmp-commits]

mstorsjo wrote:

> I suppose it could be two separate PRs that I don't merge until both are accepted but I'm not sure what the utility of that is. But I don't mind splitting it if you think it'd be easier to review.

Yes, I still think it would be significantly easier to review.

> Edit: sorry I missed this; just to check I haven't missed something, I see you said
> 
> > At least for the mingw arm64ec support I don't see an issue with it...
> 
> Does the GNU toolchain support arm64ec at all? Last time I checked it didn't even support normal arm64 on Windows but that was a while ago. I haven't made any attempt to test this with the GNU toolchain if it does; I did make the assumption that arm64ec will always be arm64ec-pc-windows-msvc so I'd be (pleasantly!) surprised if it works with GNU.

No, the GNU toolchains don’t support arm64ec, at least yet. There is support under development for regular aarch64 windows though, but it’s not quite complete yet, as far as I know.

But I didn’t mean using the GNU toolstoolchains; clang and LLVM also can operate in mingw mode, with similar installation layouts, tool interfaces to what you get with GNU tools. https://github.com/mstorsjo/llvm-mingw/releases provides this, and other toolchains providers also build similar things. Since LLVM 21, arm64ec works fine in this setup. It’s so far treated entirely like a separate, fifth architecture, so you invoke e.g. `arm64ec-w64-mingw32-clang` instead of `aarch64-w64-mingw32-clang` etc. (Just calling clang and setting the `-target` parameter also works.)

In that setup, just plain arm64ec builds of openmp should work just fine. We haven’t really figured out the best way to do arm64x in this toolchain setup though, so for now it’s just like a separate architecture.

https://github.com/llvm/llvm-project/pull/175988