[Mlir-commits] [mlir] [mlir][linalg] Add splat transpose canonicalization patterns (PR #195991)

[Hocky Yudhiono]

================
@@ -2124,6 +2134,14 @@ LogicalResult TransposeOp::fold(FoldAdaptor adaptor,
     return success();
   }
 
+  if (getInput().getType() == getInit().getType()) {
----------------
hockyy wrote:

```mlir
  func.func @transpose_splat_constant_same_shape_permutations(%arg0: tensor<3x3x3xf32>, %arg1: tensor<3x3x3xf32>) -> (tensor<3x3x3xf32>, tensor<3x3x3xf32>) {
    %cst = arith.constant dense<1.000000e+00> : tensor<3x3x3xf32>
    %transposed = linalg.transpose ins(%cst : tensor<3x3x3xf32>) outs(%arg0 : tensor<3x3x3xf32>) permutation = [0, 1, 2] 
    %transposed_0 = linalg.transpose ins(%cst : tensor<3x3x3xf32>) outs(%arg1 : tensor<3x3x3xf32>) permutation = [2, 0, 1] 
    return %transposed, %transposed_0 : tensor<3x3x3xf32>, tensor<3x3x3xf32>
  }
```

I wanted to handle the case above in the fold, so, regardless of permutation maps, both can be folded. But now that we support the general canonicalization, I think the fold is redundant. I will just yeet out this `::fold` addition.

```mlir
  func.func @transpose_splat_constant_same_shape_permutations(%arg0: tensor<3x3x3xf32>, %arg1: tensor<3x3x3xf32>) -> (tensor<3x3x3xf32>, tensor<3x3x3xf32>) {
    %cst = arith.constant dense<1.000000e+00> : tensor<3x3x3xf32>
    return %cst, %cst : tensor<3x3x3xf32>, tensor<3x3x3xf32>
  }
```

https://github.com/llvm/llvm-project/pull/195991