[Mlir-commits] [mlir] [mlir][tosa] Use traits to check output type aligns with input type (PR #193961)

[Luke Hutton]

================
@@ -2349,8 +2349,12 @@ def Tosa_ReshapeBlockScaledOp
 // Operator: reverse
 //===----------------------------------------------------------------------===//
 def Tosa_ReverseOp: Tosa_Op<"reverse", [
-    DeclareOpInterfaceMethods<InferShapedTypeOpInterface,
-                              ["inferReturnTypeComponents"]>, Pure]> {
+      DeclareOpInterfaceMethods<InferShapedTypeOpInterface,
+                                ["inferReturnTypeComponents"]>,
+      SameOperandsAndResultRank,
----------------
lhutton1 wrote:

Unfortunately, this would prevent constructs such as:
```
func.func @test_reverse(%arg0: tensor<2x3xi32>) -> tensor<2x?xi32> {
  %0 = tosa.reverse %arg0 {axis = 1 : i32} : (tensor<2x3xi32>) -> tensor<2x?xi32>
  return %0 : tensor<2x?xi32>
}
```
which are required for shape inference.

https://github.com/llvm/llvm-project/pull/193961