[Mlir-commits] [mlir] [MLIR][SCF] Add canonicalization pattern to fold away iter args of scf.forall (PR #90189)

[Abhishek Varma]

Abhishek-Varma wrote:

Hi @matthias-springer @MaheshRavishankar 

Thank you for your review comments! They were really helpful and helped clear out certain concepts pertaining to `scf.forall`.

A few notes to help understand the current stance as per the discussion thread above :-
1. `scf.forall.in_parallel` implements `ParallelCombiningOpInterface` through which we can get hold of all the yield ops - @matthias-springer thank you for this suggestion (I was able to directly use the interface method without having to `cast`)
2. From step 1 I'm basing the search on just `tensor.parallel_insert_slice` and thus not going through the loop body unnecessarily.
3. The above has been added as a method in `scf.forall` - @MaheshRavishankar thank you for this suggestion.

Please re-review/approve whenever you both get a chance. Thanks!

https://github.com/llvm/llvm-project/pull/90189