[all-commits] [llvm/llvm-project] 44da86: [MLIR][XeGPU] Clone trivial operations with multip...

[Nishant Patel via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 44da8601d71ea712b92589bc83f498d6e952c4f9
      https://github.com/llvm/llvm-project/commit/44da8601d71ea712b92589bc83f498d6e952c4f9
  Author: Nishant Patel <nishant.b.patel at intel.com>
  Date:   2026-05-26 (Tue, 26 May 2026)

  Changed paths:
    M mlir/include/mlir/Dialect/XeGPU/Transforms/XeGPULayoutImpl.h
    M mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp
    M mlir/lib/Dialect/XeGPU/Transforms/XeGPUPropagateLayout.cpp
    M mlir/test/Dialect/XeGPU/resolve-layout-conflicts.mlir

  Log Message:
  -----------
  [MLIR][XeGPU] Clone trivial operations with multiple consumers in layout propagation to avoid layout conflict (#197514)

In XeGPU layout propagation, when a trivially rematerializable producer
(e.g. a CSE-merged vector.step) has multiple uses demanding distinct
layouts, clone the producer once per required layout so each def retains
a single well-defined layout. This avoids inserting a bridging
xegpu.convert_layout whose WG-to-SG lowering would otherwise force a
costly SLM round-trip to redistribute data across subgroups.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications