[Mlir-commits] [mlir] [mlir][linalg] Update createWriteOrMaskedWrite (PR #174810)

[Diego Caballero]

dcaballe wrote:

Thanks for addressing the comments.

> Also, we load 5x1 elements and have to make sure we don't store spurious elements that we masked out for the load, right?

I think this point hasn't been addressed/answered. Let's say we want to insert a `tensor<4xf32>` into `tensor<64xf32>`, both from and to index 0 (to keep it simple). We use vector size 8. The masked vector load should load the first 4 elements and mask out the last 4 elements. The masked store should also mask out the last 4 elements even though storing an 8-element vector would be within the boundaries of the destination tensor. Otherwise, we'd be storing spurious elements. That is, the load mask should be combined (and) with the create mask of the store. 

Am I missing anything? Is this handled properly?

https://github.com/llvm/llvm-project/pull/174810