[Mlir-commits] [mlir] [mlir][ArmSME] Remove `ConvertIllegalShapeCastOpsToTransposes` (PR #139706)

[Andrzej Warzyński]

banach-space wrote:

Note: I’ve rebased this on top of main to keep it up to date — please let me know if that’s problematic, and I’ll avoid doing it in the future.

@MacDue, I’ve restored the ability to lower the following (thanks for flagging this!):
```mlir
  %illegalRead = vector.transfer_read %memref[%a, %b], %pad: memref<?x?xf32>, vector<[4]x1xf32>
```

This is now handled in `LowerColumnTransferReadToLoops`. There are two follow-up points worth discussing:
* I might have been better off using [TransferOpConversion](https://github.com/llvm/llvm-project/blob/d04d335a2ef48a433cb661665edd3e1ef47e7a04/mlir/lib/Conversion/VectorToSCF/VectorToSCF.cpp#L873-L875) - I’ll take another look.
* `in_bounds` isn’t handled properly yet.

Personally, I’d suggest leaving these as TODOs for now, and focus this PR on removing `ConvertIllegalShapeCastOpsToTransposes`, since we now have a clear forward path. That said, I don’t feel strongly - just trying to prioritise, especially considering that `ConvertIllegalShapeCastOpsToTransposes` currently isn’t exercised at all.

WDYT?

https://github.com/llvm/llvm-project/pull/139706