arsenm wrote: I found a way to hack in the old assembler behavior with both wave sizes. We're really not handling assembler predicates correctly. Essentially none of the manual validation code is correct, and shouldn't be necessary https://github.com/llvm/llvm-project/pull/159234