[Mlir-commits] [mlir] [mlir][vector] Implement IndexedAccessOpInterface for load, store, etc. (PR #196216)

[Erick Ochoa Lopez]

================
@@ -642,6 +642,25 @@ func.func @fold_vector_load_subview(%src : memref<24x64xf32>,
 
 // -----
 
+// This should fold, but implementing IndexedAccessOpInterface on vector.load
+// in a way that would allow the fold added complexity (emitting
+// `vector.shape_cast`s) that people wanted to keep out of the initial
+// implementation during previous discussions, so we leave it negative for now.
+func.func @no_fold_scalar_equivalent_vector_load_subview(
----------------
amd-eochoalo wrote:

Can you use negative prefix for the negative tests?

Is this the only fold that is no longer possible now? Was it possible before and just untested? If it was possible before and untested, I would just add a comment to the PR description stating this. I think it would also be good to state in the PR description that drop-leading-unit-dims should take care of the other patterns as well.

https://github.com/llvm/llvm-project/pull/196216