llvm-project

6b617d4a - [mlir][vector] Drop leading unit dims in memref op folding (#197278)

[mlir][vector] Drop leading unit dims in memref op folding (#197278) Make the IndexedAccessOpInterface implementations for vector dialect operations a bit smarter by letting them drop redundant unit dimensions and shape_cast operations back. This allows folding in, for example, an expand_shape of a 1-D memref to a load of a `vector<1x1x8xf32>` by turning it into a load of a `vector<8xf32>` instead. This functionality is somewhat reduntant with dropleadunitdims, but is included for completeness and to better match the interface contract of IndexedAccessOpInterface. While we're here, fix a bug in the verifier of vector.expandshape/collapsestore - as near as I can tell, they were meant to check for the mask being the same shape is the input vector, but they just checked the first dimension. AI: Co-authored by Codex 5.5 (which also caught a bunch of my old typos)