@mstorsjo, sorry, didn't ask you earlier: what is your environment? Our combination: everything is on Windows; we are using MSVC compiler to compile OpenMP code (for x64, x86, or arm64), and using libomp140 for runtime. LLVM libomp runtime is compiled by MSVC (for x64, x86, arm64) with additional definition 'OPENMP_MSVC_NAME_SCHEME' for cmake. We also have LLVM_ENABLE_PDB=ON for cmake for release, AFAIK it's different from usual.

I've finally managed to somehow cobble things together so I can run the lit tests for this configuration: it mostly seems to work, but there are a couple failures that I probably should look into (although I don't know if they're related to what's being changed here or whether it's a later issue.

Anyway, with x86_64 clang-cl/msvc mode, there's 4 legitimate test failures among the tests; with x86_64 mingw there's those 4 plus yet another one failing. With aarch64 mingw there's 9 more failing tests (but still 180 tests passing).

We never tested this configuration. I'd suggest to look at the additional failures, and if they are related to atomics, to investigate them (otherwise consider them to be a different issue). If you can share all logs, I'd be curious to look at the results too (it won't save you any time though)

In D137168#3900133, @mstorsjo wrote:

FWIW I tested this patch for a mingw/aarch64 environment, and at least a very small testcase does seem to run as expected.

@jdoerfert : yes, it's a great idea, I can join next Wednesday. From the agenda I didn't know you'd like to discuss this today. Hopefully Chris will be able to come too, but I'll let him tell for himself