Output arg1 = ops::_Arg(s.WithOpName("arg1"), DT_RESOURCE, 1);
    Output arg2 = ops::_Arg(s.WithOpName("arg2"), DT_INT32, 2);
    auto ret0 = ops::_Retval(s.WithOpName("ret0"), arg1, 0);
    auto ret1 = ops::_Retval(s.WithOpName("ret1"), arg0, 1);
    auto ret2 = ops::_Retval(s.WithOpName("ret2"), arg2, 2);
    std::unique_ptr<Graph> g(new Graph(OpRegistry::Global()));
    TF_CHECK_OK(s.ToGraph(g.get()));

    TF_RETURN_IF_ERROR(status);
    graph->AddEdge(node, 0, main_node, i);
  }

  // Similarly with return values, create dummy _Retval nodes fed by `node`.
  for (int64_t i = 0, end = result_types.size(); i < end; ++i) {
    Node* node;
    string retval_name = absl::StrCat("_retval", i);
    Status status = NodeBuilder(retval_name, FunctionLibraryDefinition::kRetOp)
                        .Input(main_node, i)

  might become the name of Retval nodes as well (with an index suffix if there
  are multiple output tensors from one node). Since Retval nodes are not used in
  SavedModel, this function removes them and restore the names to the actual
  output tensors.

  Args:
    graph_def: the converted GraphDef.

  Returns:
    The GraphDef with Retval nodes removed and output tensor names restored.

  // Computes the permutation to produce the correct retval order, and update
  // the argument indices.
  for (int i = 0; i < num_retvals; ++i) {
    int index;
    TF_RETURN_IF_ERROR(GetIndexAttr(*retvals[i], num_retvals, &index));
    (*output_permutation)[index] = i;
    retvals[i]->AddAttr("index", i);
  }

  AddNodeAttr(kXlaClusterIdAttr, call_def->name(), call_def);

// Builds XlaSendFromHost node, and replaces all _Retval nodes with it.
absl::StatusOr<Node*> ReplaceRetNodesWithSendFromHostNode(
    Graph* g, const string& oc_cluster_name,
    std::vector<DataType>* send_from_host_dtypes, Node* key_placeholder) {
  // TODO(b/77601805): use in nodes for sink node, instead of traversing all
  // nodes.
  std::vector<Node*> ret_nodes = GatherNodesWithType(*g, "_Retval");

        const Edge* edge,
        const absl::flat_hash_map<const Node*, Node*>& node_images);

    // Creates a _Retval node for the src node of edge, and add it to results_,
    // if none exists yet. If a new _Retval node is created, also adds the edge
    // within the subgraph from the src to the _Retval node.
    Status RecordResult(
        const Edge* edge,
        const absl::flat_hash_map<const Node*, Node*>& node_images);

      count: 4
    }
  }
  unclustered_op_histogram {
    op: "NoOp"
    count: 2
  }
  unclustered_op_histogram {
    op: "_Arg"
    count: 1
  }
  unclustered_op_histogram {
    op: "_Retval"
    count: 1
  }
}
)",
      &expected_auto_clustering_activity);
  EXPECT_EQ(listener()->auto_clustering_activity().DebugString(),
            expected_auto_clustering_activity.DebugString());

      StrEq("input_tensor:0"));

  // Match the `_Retval` node that corresponds to the return value of @main.
  const auto retval_node_itr =
      llvm::find_if(exported_model->graph_def().node(),
                    [](const NodeDef& node) { return node.op() == "_Retval"; });
  ASSERT_NE(retval_node_itr, exported_model->graph_def().node().end());

    return false;
  }

  // _Arg nodes in a top-level function represent feeds and _Retval nodes in a
  // top-level function represent fetches.
  if (stack_depth == 1 &&
      (node.type_string() == "_Arg" || node.type_string() == "_Retval")) {
    absl::string_view uncompilable_reason = "top level _Arg or _Retval";
    MaybeMarkUncompilableNode(uncompilable_reason, *stack_trace,

    if (mul) {
      auto c = ops::Mul(scope.WithOpName("C"), a, b);
      auto d = ops::_Retval(scope.WithOpName("D"), c, 0);
      TF_RETURN_IF_ERROR(scope.ToGraph(graph.get()));
    } else {
      auto c = ops::Add(scope.WithOpName("C"), a, b);
      auto d = ops::_Retval(scope.WithOpName("D"), c, 0);
      TF_RETURN_IF_ERROR(scope.ToGraph(graph.get()));
    }