[Mlir-commits] [mlir] [mlir][vector][xegpu] Accept uniform values in `getDistributedType` (PR #163887)

[Charitha Saumya]

================
@@ -214,3 +214,54 @@ gpu.module @xevm_module{
 
   }
 }
+
+// -----
+// CHECK-LABEL: gpu.func @warp_scf_for_unused_uniform_for_result(
+// CHECK:         %[[W:.*]]:2 = gpu.warp_execute_on_lane_0(%{{.*}})[16] args(%{{.*}} : index,
+// CHECK-SAME:      !xegpu.tensor_desc<16x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>,
+// CHECK-SAME:      memref<16x16xf32>) -> (vector<16x1xf32>, vector<16x1xf32>) {
+// CHECK:           gpu.yield %{{.*}}, {{.*}} : vector<16x16xf32>, vector<16x1xf32>
+// CHECK:         }
+// CHECK:         %{{.*}}:2 = scf.for {{.*}} to %{{.*}} step %{{.*}} iter_args
+// CHECK-SAME:      (%{{.*}} = %[[W]]#0, %{{.*}} = %[[W]]#1) -> (vector<16x1xf32>, vector<16x1xf32>) {
+// CHECK:           %[[W1:.*]]:2 = gpu.warp_execute_on_lane_0(%{{.*}})[16]
+// CHECK-SAME:        args(%{{.*}} : vector<16x1xf32>, vector<16x1xf32>) -> (vector<16x1xf32>, vector<16x1xf32>) {
+// CHECK:             gpu.yield %{{.*}}, %{{.*}} : vector<16x16xf32>, vector<16x1xf32>
+// CHECK:           }
+// CHECK:           scf.yield %[[W1]]#0, %[[W1]]#1 : vector<16x1xf32>, vector<16x1xf32>
+// CHECK:         }
+gpu.module @xevm_module{
----------------
charithaintc wrote:

this would not trigger the bug. if the vector of 2 is extracted, then it means it has a user. 

This bug triggers when "before sinking a region op we don't know the distributed type of all the operands of this region op". This won't ever be triggered for a non-region op that has at least one user. 

https://github.com/llvm/llvm-project/pull/163887