[Mlir-commits] [mlir] [mlir][xegpu] Fix lane-local classification for packed-lane-data reductions (PR #223077)

[Artem Kroviakov]

================
@@ -740,6 +740,29 @@ gpu.func @vector_multi_reduction_3d_leading_unit_dim_cross_lane() {
   gpu.return
 }
 
+// lane_data packs all 16 reduced elements of dim1 into a single lane and
+// lane_layout is 1 everywhere, so the reduction is lane-local and must lower to
+// a plain vector.reduction. Classifying it by result type would route it
----------------
akroviakov wrote:

No need to specify what happens in a suboptimal implementation not present after the PR merge.

https://github.com/llvm/llvm-project/pull/223077