// CHECK-DAG: %[[VAR_HANDLE_1:.*]] = "tf.VarHandleOp"()
// CHECK-DAG: "tf.AssignVariableOp"(%[[VAR_HANDLE_1]], %[[CST_1]])

  func.func @serving_default(%arg0: tensor<8xf32> {tf_saved_model.index_path = ["input"]}) -> (tensor<8xf32> {tf_saved_model.index_path = ["output"]})
    attributes {tf.entry_function = {control_outputs = "", inputs = "", outputs = "StatefulPartitionedCall:0"}, tf_saved_model.exported_names = ["serving_default"]} {

  func.func @main(%arg0: tensor<?x384xf32> {tf_saved_model.index_path = ["input2"]}, %arg1: tensor<?x384xf32> {tf_saved_model.index_path = ["input1"]}) -> (tensor<?x5xf32> {tf_saved_model.index_path = ["start_logits"]}, tensor<?x5xf32> {tf_saved_model.index_path = ["end_logits"]}) attributes {tf.entry_function = {control_outputs = "", inputs = "serving_default_input2:0,serving_default_input1:0", outputs = "S...

# CHECK:      func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME:   [[ARG0:%.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]},
# CHECK-SAME:   [[ARG1:%.*]]: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.bound_input = @[[VAR]]})
# CHECK-SAME:             -> (tensor<3x3xf32> {tf_saved_model.index_path = ["r"]})
# CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["key"]

          func.removeArgAttr(i, bound_input);
        }
        func.removeArgAttr(i, index_path);
      }
      for (unsigned i = 0, e = func.getNumResults(); i != e; ++i) {
        func.removeResultAttr(i, bound_input);
        func.removeResultAttr(i, index_path);
      }
    });

    // Clean up modules from tf_saved_model attributes.
    module->removeAttr("tf_saved_model.semantics");
  }

  }

  return nullptr;
}

// Adds the file prefix argument to `main_func_op`. The file prefix argument
// is the argument whose "tf_saved_model.index_path" attribute has
// "__tf_file_prefix". Its type is `tensor<!tf_type.string>`. Also, the value
// "__tf_file_prefix:0" is appended to the "tf.entry_function" attribute's
// "inputs" key.

# The function for the signature "key".
# CHECK: func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME: %[[ARG_1:.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]}
# CHECK-SAME: %[[ARG_2:.*]]: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.bound_input = @y}
# CHECK-SAME: -> (tensor<3x3xf32> {tf_saved_model.index_path = ["r"]})
# CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["key"]

// CHECK-SAME: -> (!tfrt.chain, !tfrt_fallback.tf_tensor)
func.func @func_basic(
    %arg0: tensor<3x1xf32> {tf_saved_model.index_path = [0]},
    %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.index_path = [1]})
      -> (tensor<3x3xf32> {tf_saved_model.index_path = []})
  attributes {tf_saved_model.exported_names = ["basic"]} {

  # has a corresponding tf.LegacyCall op.
  # CHECK:      func {{@[a-zA-Z_0-9]+}}(
  # CHECK-SAME: [[ARG0:%.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["y"]},
  # CHECK-SAME: [[ARG1:%.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]}
  #
  # CHECK-NEXT: [[R0:%.*]] = "tf.LegacyCall"([[ARG1]], [[ARG0]])
  z = plus(x, y)

  tensor_info_x = tf.compat.v1.saved_model.utils.build_tensor_info(x)

  func.func @serving_default(%arg0: tensor<!tf_type.resource<tensor<100x50xf32>>> {tf.resource_name = "dense/kernel"}, %arg1: tensor<!tf_type.resource<tensor<50xf32>>> {tf.resource_name = "dense/bias"}) -> (tensor<100x50xf32> {tf_saved_model.index_path = ["dense_2"]})
  attributes {tf.entry_function = {control_outputs = "", inputs = "", outputs = "dense_2/Add:0"}, tf_saved_model.exported_names = ["serving_default"]} {

  func.func @gather_with_float_output(%arg0: tensor<1x3x2x1024xf32> {tf_saved_model.index_path = ["input1"]}, %arg1: tensor<1xi32> {tf_saved_model.index_path = ["input2"]}) -> (tensor<1x3x1x1xf32> {tf_saved_model.index_path = ["output"]}) attributes {tf.entry_function = {control_outputs = "", inputs = "serving_default_input:0", outputs = "PartitionedCall:0"}, tf_saved_model.exported_names = ["serving_default"]}...