[Mlir-commits] [mlir] [mlir][vector] Create `VectorToLLVMDialectInterface` (PR #121440)

[Andrzej Warzyński]

banach-space wrote:

Thanks for getting back to me.

> We cannot (easily) do this. These tests designed to work with `convert-vector-to-llvm` pass which includes some preparations patterns running through `applyPatternsGreedily`. 

This doesn't sound right. The tests in "Conversion/VectorToLLVM" should exercise the dialect conversion, i.e. https://github.com/llvm/llvm-project/blob/d839c06719128700bdd033361b20aa6899f6620a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp#L1906-L1929

That's exactly what the newly proposed `VectorToLLVMDialectInterface` is meant to run, right?

So, could you provide specific examples where things break (i.e. my assumption doesn't hold)? It's possible that "vector-to-llvm.mlir" contains tests that shouldn't be there. If that's the case, we should use this opportunity to fix that.

> Also, this pass includes armNeon/armSVE/amx/x86Vector specific transforms (which should probably be handled through their own dedicated `ToLLVMDialectInterface`s).

There's nothing architecture specific in "vector-to-llvm.mlir", so none of these should be required. If that's not the case, we should that file.

> 
> Motivation for this one is to unhardcode conversions from GPUtoROCDL/NVVM passes https://github.com/llvm/llvm-project/blob/main/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp#L296

Right now it's not clear to me how that's connected, but if it's going to help, then we should make it happen. In fact, it would be good to make sure that this change unblocks something useful. So - why not include the changes for "LowerGpuOpsToROCDLOps.cpp" in this PR?

https://github.com/llvm/llvm-project/pull/121440