[Mlir-commits] [mlir] [mlir] Exclude masked ops in VectorDropLeadUnitDim (PR #76468)

[llvmlistbot at llvm.org]

tanmaysachan wrote:

> We can implement a pattern rewrite for MaskOp that inherits from a base class like the `VectorMaskOpConversionBase` here:
> 
> https://github.com/llvm/llvm-project/blob/ba81477e9cdb5384a10c700ee247562f0f6c03e6/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp#L850
> 
> This should automatically deal with the insertion point issue of the mask op.

@dcaballe Here, are you suggesting pattern rewrite to convert MaskOp directly to llvm before this optimization to prevent insertion of cast ops into maskop's region?

https://github.com/llvm/llvm-project/pull/76468