[Mlir-commits] [mlir] [mlir][xegpu] Fix operand layouts when rematerializing on layout conflict (PR #224730)

[Jianhui Li]

================
@@ -408,6 +408,31 @@ gpu.func @step_muli_clone_via_anchor(%arg0: i64, %arg1: memref<32x32xf32>) kerne
   gpu.return
 }
 
+// Rematerializing a producer must also rematerialize its operands so the clone
+// stays internally consistent. Here `arith.index_castui` (elementwise) is cloned
+// for the second consumer layout; its `vector.step` operand must be cloned with
+// the matching layout too, otherwise the clone would have an operand and result
----------------
Jianhui-Li wrote:

please remove the otherwise portion.  The comments could be shorten. 

https://github.com/llvm/llvm-project/pull/224730