[Mlir-commits] [mlir] [mlir][xegpu] Fix operand layouts when rematerializing on layout conflict (PR #224730)
Charitha Saumya
llvmlistbot at llvm.org
Tue Oct 6 08:48:22 PDT 2026
================
@@ -1652,6 +1652,36 @@ LogicalResult ResolveLayoutConflicts::assignResultLayout(OpResult &result) {
return success();
}
+// Recursively rematerialize a trivially-rematerializable producer subtree so
+// that the clone yields `layout`. Cloning only the producer would leave its
+// operands pointing at values that carry the original (conflicting) layout,
----------------
charithaintc wrote:
I get that. but this new recursive traversal seems unmercenary. I am concerned about the compile time as well and maintainability.
Instead I would prefer a while loop over the original traversal and check for a fixed point (i.e. no more new conflicts). This is the standard way to such things.
https://github.com/llvm/llvm-project/pull/224730
More information about the Mlir-commits
mailing list