[Mlir-commits] [mlir] [mlir][Vector][NFC] Move canonicalizers for DenseElementsAttr to folders (PR #127995)

[Kunwar Grover]

Groverkss wrote:

> > Folders are local, and it's always better to implement a folder than a canonicalizer.
> 
> In the past we tried to be cautious of IR size and bail out of canonicalizers if they produced large constants. I wonder if moving to folders raises the bar even higher.

I don't think folders have a higher bar than canonicalizers. A folder is just a canonicalizer that is local. I think what you are raising is valid point, but it's not related to this patch, as this patch is just implementing the same canonicalization in a more restricted fashion (local transformations only).

I think for InsertOp folders, we do have a fold limit. I think it would be good to set a limit for the ExtractOp too.

https://github.com/llvm/llvm-project/pull/127995