llvm-project

4b12ba82 - [mlir][vector] Fix mask rank extension for TransferWriteNonPermutationLowering (#229852)

[mlir][vector] Fix mask rank extension for TransferWriteNonPermutationLowering (#229852) When a `vector.transfer_write`'s permutation map skips a dim between the dims that are written to, e.g. `(d0, d1, d2, d3) -> (d1, d3)`, `TransferWriteNonPermutationLowering` adds a unit dim for it. The vector got that unit dim in the right place, but the mask always got it at the end. The new write then failed to verify: ``` 'vector.transfer_write' op inferred mask type ('vector<7x1x8xi1>') and mask operand type ('vector<7x8x1xi1>') don't match ``` `-convert-vector-to-scf` runs the same pattern and fails on such a write too. This puts the mask's unit dims at the positions of the missing dims. This is #144893 by @Groverkss, moved to the current builder API. The test uses a map that also skips the leading dim, so the offset from the first written dim is covered as well. Part of #227821. Assisted-by: Claude --------- Co-authored-by: Kunwar Grover <groverkss@gmail.com>