[Mlir-commits] [mlir] [MLIR][Vector] Fix transferOps optimization inside maskOp (PR #90835)

[Hugo Trachino]

nujaa wrote:

> Is the plan to rebase this on top of #91987?

Yes.  As mentionned in your previous comment, there are lots of different patterns updated.

> Thanks for the updates, looks good!
> 
> I've noticed that the following patterns have been updated:
> 
> * `TransferReadOfExtractSliceOpFolder`
> * `TransferReadPermutationLowering`
> * `TransferWritePermutationLowering`
> * `TransferWriteNonPermutationLowering`
> * `TransferOpReduceRank`
> * `TransferReadToVectorLoadLowering`
> * `TransferWriteToVectorStoreLowering`
> 
> However, there's only 3 new tests :) And it's hard to tell which tests corresponds to which pattern. You could address that by creating multiple PRs and group similar patterns together (e.g. `TransferReadToVectorLoadLowering` and `TransferWriteToVectorStoreLowering`). Just a loose suggestion, not a requirement.

I have submitted : 
* https://github.com/llvm/llvm-project/pull/91987 for `TransferReadPermutationLowering`,  `TransferWritePermutationLowering` and `TransferWriteNonPermutationLowering`
* https://github.com/llvm/llvm-project/pull/91960  targetting  `TransferReadOfExtractSliceOpFolder`.
* I have 2 branches ready to be pushed on top of them to respectively address `TransferOpReduceRank` and `Transfer{Read|write}ToVector{Load|store}Lowering`


https://github.com/llvm/llvm-project/pull/90835