[Mlir-commits] [mlir] [mlir][vector] Add a negative test for `FoldTransposeBroadcast` (PR #217915)

[Andrzej Warzyński]

================
@@ -67,18 +67,31 @@ func.func @broadcast_transpose_partial_ones_to_broadcast(%arg0 : vector<1xi8>) -
 
 // -----
 
-// CHECK-LABEL: broadcast_transpose_mixed_example
+// CHECK-LABEL: broadcast_transpose_mixed_order_preserving
 //  CHECK-SAME:  %[[ARG:.*]]: vector<4x1x1x7xi8>) -> vector<3x2x4x5x6x7xi8> {
 //       CHECK:  %[[RES:.*]] = vector.broadcast %[[ARG]] : vector<4x1x1x7xi8> to vector<3x2x4x5x6x7xi8>
 //       CHECK:  return %[[RES]] : vector<3x2x4x5x6x7xi8>
-func.func @broadcast_transpose_mixed_example(%arg0 : vector<4x1x1x7xi8>) -> vector<3x2x4x5x6x7xi8> {
+func.func @broadcast_transpose_mixed_order_preserving(%arg0 : vector<4x1x1x7xi8>) -> vector<3x2x4x5x6x7xi8> {
   %0 = vector.broadcast %arg0 : vector<4x1x1x7xi8> to vector<2x3x4x5x6x7xi8>
   %1 = vector.transpose %0, [1, 0, 2, 3, 4, 5] : vector<2x3x4x5x6x7xi8> to vector<3x2x4x5x6x7xi8>
   return %1 : vector<3x2x4x5x6x7xi8>
 }
 
 // -----
 
+// CHECK-LABEL: negative_broadcast_transpose_mixed_not_order_preserving
+//  CHECK-SAME:  %[[ARG:.*]]: vector<4x1x1x7xi8>) -> vector<6x2x4x5x3x7xi8> {
+//       CHECK:  %[[BC:.*]] = vector.broadcast
+//       CHECK:  %[[TR:.*]] = vector.transpose %[[BC]]
+//       CHECK:  return %[[TR]] : vector<6x2x4x5x3x7xi8>
+func.func @negative_broadcast_transpose_mixed_not_order_preserving(%arg0 : vector<4x1x1x7xi8>) -> vector<6x2x4x5x3x7xi8> {
+  %0 = vector.broadcast %arg0 : vector<4x1x1x7xi8> to vector<2x3x4x5x6x7xi8>
+  %1 = vector.transpose %0, [4, 0, 2, 3, 1, 5] : vector<2x3x4x5x6x7xi8> to vector<6x2x4x5x3x7xi8>
----------------
banach-space wrote:

> I don't know the historical reason behind this implementation choice.

Since it's a canonicalisation, it must have been related to preserving information that would otherwise be lost.

In any case, we should always take extra care when modifying the semantics of a folder. This PR is the first step. Let's land it and then return to #215940 . I am hoping that @dcaballe will be able to chime in too.

https://github.com/llvm/llvm-project/pull/217915