[Mlir-commits] [mlir] [mlir][gpu] Allow WarpOpDeadResult, WarpOpForwardOperand patterns to be used in isolation. (PR #132860)

Charitha Saumya llvmlistbot at llvm.org
Wed Mar 26 12:11:28 PDT 2025


charithaintc wrote:

> I think you could move the pattern to the canonicalization. There's nothing specific to vector there iirc.

I considered this approach now, but the issue is then at op level we need to rely on `WarpDistributionPattern` or otherwise reimplement `moveRegionToNewWarpOpAndReplaceReturns`. Does that sound ok to you?

I think moving these simplification patterns to a separate function is minimally invasive and gets the job done. 


https://github.com/llvm/llvm-project/pull/132860


More information about the Mlir-commits mailing list