joker-eph wrote: > This change does not add the new pattern to the tensor.pad canonicalizers, so it is opt-in and could be parameterized by a cost function. Yeah it is fine! https://github.com/llvm/llvm-project/pull/92691