[Mlir-commits] [mlir] [mlir][vector] Support warp distribution of `transfer_read` with dependencies (PR #77779)

[Nicolas Vasilache]

================
@@ -1264,10 +1264,11 @@ func.func @vector_insert_2d_broadcast(%laneid: index) -> (vector<4x96xf32>) {
 //       CHECK-PROP:     %[[EXTRACT:.*]] = vector.extract %[[GATHER]][0] : vector<64xi32> from vector<1x64xi32>
 //       CHECK-PROP:     %[[CAST:.*]] = arith.index_cast %[[EXTRACT]] : vector<64xi32> to vector<64xindex>
 //       CHECK-PROP:     %[[EXTRACTELT:.*]] = vector.extractelement %[[CAST]][{{.*}}: i32] : vector<64xindex>
-//       CHECK-PROP:     %[[TRANSFERREAD:.*]] = vector.transfer_read %[[AR2]][%[[C0]], %[[EXTRACTELT]], %[[C0]]],
-//       CHECK-PROP:     vector.yield %[[TRANSFERREAD]] : vector<64xf32>
-//       CHECK-PROP:   return %[[W]]
-func.func @transfer_read_no_prop(%in2: vector<1x2xindex>, %ar1 :  memref<1x4x2xi32>, %ar2 : memref<1x4x1024xf32>)-> vector<2xf32> {
+//       CHECK-PROP:     vector.yield %[[EXTRACTELT]] : index
+//       CHECK-PROP:     %[[APPLY:.*]] = affine.apply #[[$MAP]]()[%[[THREADID]]]
----------------
nicolasvasilache wrote:

nit: can we indent this better?
Right now it looks like it is in the region but after the terminator.

https://github.com/llvm/llvm-project/pull/77779