https://github.com/Harry-Ramsey approved this pull request. LGTM. I think someone with more expertise should just confirm that the latency with add is faster than two madds. https://github.com/llvm/llvm-project/pull/214796