[Mlir-commits] [mlir] [mlir][ArmSME] Remove `ConvertIllegalShapeCastOpsToTransposes` (PR #139706)

[Benjamin Maxwell]

================
@@ -491,51 +491,6 @@ func.func @illegal_transpose_no_defining_source_op(%vec: vector<[4]x1xf32>) -> v
 
 // -----
 
-// CHECK-LABEL: @illegal_shape_cast_to_transpose_2d(
-// CHECK-SAME:                                      %[[VEC:.*]]: vector<[4]x1xf32>)
-func.func @illegal_shape_cast_to_transpose_2d(%vec: vector<[4]x1xf32>) -> vector<1x[4]xf32> {
-  // CHECK: vector.transpose %[[VEC]], [1, 0] : vector<[4]x1xf32> to vector<1x[4]xf32>
-  %0 = vector.shape_cast %vec : vector<[4]x1xf32> to vector<1x[4]xf32>
-  return %0 : vector<1x[4]xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @illegal_shape_cast_to_transpose_1d(
-// CHECK-SAME:                                      %[[VEC:.*]]: vector<[4]x1xf32>)
-func.func @illegal_shape_cast_to_transpose_1d(%vec: vector<[4]x1xf32>) -> vector<[4]xf32> {
-  // CHECK: %[[TRANSPOSE:.*]] = vector.transpose %[[VEC]], [1, 0] : vector<[4]x1xf32> to vector<1x[4]xf32>
-  // CHECK: vector.shape_cast %[[TRANSPOSE]] : vector<1x[4]xf32> to vector<[4]xf32>
-  %0 = vector.shape_cast %vec : vector<[4]x1xf32> to vector<[4]xf32>
-  return %0 : vector<[4]xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @lift_illegal_2d_shape_cast_to_memory
-func.func @lift_illegal_2d_shape_cast_to_memory(%a: index, %b: index, %memref: memref<?x?xf32>) -> vector<1x[4]xf32> {
-  // CHECK: vector.transfer_read {{.*}} : memref<?x?xf32, {{.*}}>, vector<1x[4]xf32>
-  // CHECK-NOT: vector.shape_cast
-  %pad = arith.constant 0.0 : f32
-  %illegalRead = vector.transfer_read %memref[%a, %b], %pad {in_bounds = [false, true]}: memref<?x?xf32>, vector<[4]x1xf32>
-  %cast = vector.shape_cast %illegalRead : vector<[4]x1xf32> to vector<1x[4]xf32>
-  return %cast : vector<1x[4]xf32>
----------------
MacDue wrote:

> I just tried it and it does not lower. And, AFAICT, it shouldn't, as the last dimension of the memerf

This lowering does not depend on the memref/transpose being contiguous. It lowers the `transfer_read` to a `memref.transpose` + `transfer_read`, which lowers to a loop in the case of a non-contiguous read (such as in this test case): https://godbolt.org/z/b96E7aYq4

https://github.com/llvm/llvm-project/pull/139706