[Mlir-commits] [mlir] [mlir][func] Move return-type verification from ReturnOp to FuncOp (PR #184153)

[Ingo Müller]

================
@@ -284,23 +284,37 @@ FuncOp FuncOp::clone() {
 // ReturnOp
 //===----------------------------------------------------------------------===//
 
-LogicalResult ReturnOp::verify() {
-  auto function = cast<FuncOp>((*this)->getParentOp());
-
-  // The operand number and types must match the function signature.
-  const auto &results = function.getFunctionType().getResults();
-  if (getNumOperands() != results.size())
-    return emitOpError("has ")
-           << getNumOperands() << " operands, but enclosing function (@"
-           << function.getName() << ") returns " << results.size();
-
-  for (unsigned i = 0, e = results.size(); i != e; ++i)
-    if (getOperand(i).getType() != results[i])
-      return emitError() << "type of return operand " << i << " ("
-                         << getOperand(i).getType()
-                         << ") doesn't match function result type ("
-                         << results[i] << ")"
-                         << " in function @" << function.getName();
+LogicalResult FuncOp::verify() {
+  // External declarations have no body to check.
+  if (isDeclaration())
+    return success();
+  // Hoist the result types once; they are the same for every return site.
+  auto resultTypes = getFunctionType().getResults();
+  for (Block &block : getBody()) {
+    if (block.empty())
+      continue;
+    // Check func.return or other return-like terminators ops (e.g.
+    // llvm.return, test.return).
+    auto returnOp = dyn_cast<RegionBranchTerminatorOpInterface>(&block.back());
----------------
ingomueller-net wrote:

I think that this isn't nice to have -- it's crucial for the verification to be correct. The current version breaks several downstream users with false negatives. One [example](https://github.com/tensorflow/tensorflow/blob/e9b73649/tensorflow/compiler/mlir/tensorflow/tests/tf-reduce-identity.mlir#L4) from Tensorflow:

```mlir
func.func @target_function() -> tensor<i32> {
  %0 = "tf_device.cluster"() ({
    // ...
    tf_device.return %4 : tensor<i32>
  }) {...} : () -> tensor<i32>
  func.return %0 : tensor<i32>
}
```

The `tf_device.return` is casted to `returnOp` successfully, such that the remaining code verifies properties that don't need to hold.

https://github.com/llvm/llvm-project/pull/184153