[libcxx-commits] [libcxx] [libc++][atomic_ref] Use __atomic_fetch_{add, sub} builtins on floating-points whenever possible (PR #135685)

[via libcxx-commits]

https://github.com/huixie90 requested changes to this pull request.

I would make atomic_ref and atomic share the similar logic (or just share the implantation) for these rmw operations. The reason why atomic has special branch for fp80 long double is that , the clang builtin that emit the llvm ir, has serious bug that affects all types whose data size is not power of 2. ( I think I have a patch but I never got a chance to finalise the tests. )

https://github.com/llvm/llvm-project/pull/135685