[Mlir-commits] [mlir] [mlir][vector] Generalize the canonicalization of transpose(broadcast(x)) (PR #153056)

[James Newling]

================
@@ -369,9 +369,8 @@ func.func @transfer_write_broadcast_unit_dim_tensor(
   %c0 = arith.constant 0 : index
 
   %res = vector.transfer_write %vec_0, %dst_0[%c0, %c0, %c0, %c0] {in_bounds = [false, false, true], permutation_map = affine_map<(d0, d1, d2, d3) -> (d0, d1, d3)>} : vector<14x8x16xf32>, tensor<?x?x?x?xf32>
-  // CHECK: %[[NEW_VEC0:.*]] = vector.broadcast %{{.*}} : vector<14x8x16xf32> to vector<1x14x8x16xf32>
-  // CHECK: %[[NEW_VEC1:.*]] = vector.transpose %[[NEW_VEC0]], [1, 2, 0, 3] : vector<1x14x8x16xf32> to vector<14x8x1x16xf32>
-  // CHECK: %[[NEW_RES0:.*]] = vector.transfer_write %[[NEW_VEC1]], %[[DST0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {in_bounds = [false, false, true, true]} : vector<14x8x1x16xf32>, tensor<?x?x?x?xf32>
+  // CHECK: %[[NEW_VEC0:.*]] = vector.shape_cast %{{.*}} : vector<14x8x16xf32> to vector<14x8x1x16xf32>
----------------
newling wrote:

Interesting. I have a slow moving PR where this same update is made: https://github.com/llvm/llvm-project/pull/140583/files

https://github.com/llvm/llvm-project/pull/153056