https://github.com/Artem-B approved this pull request. LGTM. It would be good to summarize and incorporate the reason why 32-bit CAS done by LLVM directly results in better SASS. https://github.com/llvm/llvm-project/pull/120220