efriedma-quic wrote: It was implemented downstream in a completely separate vectorization framework. I'm only mentioning it because we don't want to block off the possibility of adding such support in the future. https://github.com/llvm/llvm-project/pull/139893