[Mlir-commits] [mlir] [mlir] Expose linalg vectorization without replacement (PR #144158)

[Diego Caballero]

dcaballe wrote:

Nice discussion! I haven't had the time to dive deep into the details but I'm generally +1 to making the vectorizer more modular and reusable. We talked about introducing interfaces to decouple the information of the iteration space to be vectorized from the actual ops, which is something we could explore. My only concern would be if we head towards implementing an ad-hoc "conversion rewriter" where IR modifications have to live in data structures on the side, or we set the expectations that the vectorizer can't materialize changes directly in the IR. That would be difficult to justify and maintain. Happy to provide feedback on a specific example, if you have one you can share.

https://github.com/llvm/llvm-project/pull/144158