michaelmaitland wrote: > Nice, is there a specific workload you're seeing that motivates this? We were experimenting with not using vp intrinsics and saw it in x264. https://github.com/llvm/llvm-project/pull/132236