[llvm] [LangRef] No target-specific size limit for atomics (PR #136864)

[Nikita Popov via llvm-commits]

nikic wrote:

I think as far as IR and the backend is concerned, it's fine to say that atomics of arbitrary width are supported. If this results in a libatomic call and libatomic is not linked, that's a linker error and no longer LLVM's problem. The frontend will not be able to improve on this situation by producing some kind of better lowering than LLVM.

Worth pointing out that amdgpu and nvptx disable all libcalls *apart from* the atomic libcalls. They'll still emit the usual atomic_* libcalls and not crash / produce a backend error.

https://github.com/llvm/llvm-project/pull/136864