asraa wrote: I think that's fair - I'm happy to update our code to avoid using affine loops. In the end, this code is destined to lower to LLVM IR, and it really shouldn't matter if I use a lower level abstraction. https://github.com/llvm/llvm-project/pull/139069