// PartitionedCall are placed on the host, a producer that produces them on the
// device will incur a D2H copy, even if the PartitionedCall is not executed
// (i.e. even if we choose to execute the XLA compiled computation via _XlaRun).
// To prevent this, we add control dependencies to make the int32 input edges
// into the PartitionedCall dead.  With this change the D2H copy only happens if

// CHECK-NEXT: [[q_bias:%.+]] = "quantfork.qcast"([[bias]]) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>

    up creating nested XLA launch ops.

    For example, the `tf.PartitionedCall` operation in the following code

    ```mlir
    func.func @convert_partitioned_call_with_resources(%arg0: tensor<!tf_type.resource>, %arg1: tensor<i32>) -> tensor<i32> {

  *library.add_function() = identity_func;

  Output in = ops::Placeholder(root, DT_FLOAT);
  NameAttrList b_name_attr;
  b_name_attr.set_name("IdentityFunc");
  ops::PartitionedCall call(root.WithOpName("call"), {in}, {DT_FLOAT},
                            b_name_attr);

  GraphDef graph_def;
  TF_ASSERT_OK(root.graph()->AddFunctionLibrary(library));
  TF_ASSERT_OK(root.ToGraphDef(&graph_def));

    : (tensor<!tf_type.variant<tensor<f32>>>, tensor<i1>) -> tensor<!tf_type.variant<tensor<f32>>>
  // CHECK: %[[CALL2:.*]]:2 = "tf.PartitionedCall"(%[[INIT]],
  // CHECK-SAME: f = @callee_tensorlist_decomposed
  %call2 = "tf.PartitionedCall"(%tl, %arg0) {f = @callee, config = "", config_proto = "", executor_type = ""}
    : (tensor<!tf_type.variant<tensor<f32>>>, tensor<i1>) -> tensor<!tf_type.variant<tensor<f32>>>

}

void CreateSubgraphCallingFunctionWithRefVar(const Scope& s) {
  NameAttrList ref_float_function;
  ref_float_function.set_name("RefFloatFn");
  ops::PartitionedCall call(s.WithOpName("RefFloat"), {absl::Span<Input>{}},
                            {DT_FLOAT}, ref_float_function);
  Output constant =
      ops::Const(s.WithOpName("constant_ref_pco"), Input::Initializer(0.0));

// -----

// This test makes sure functions without tf._noinline=true is inlined.

module {
  func.func @partitioned_call(%arg0: tensor<1x2x2x3xf32>) -> (tensor<1x2x2x3xf32>) {
    %0 = "tf.PartitionedCall"(%arg0) <{
      config = "", config_proto = "", executor_type = "", f = @some_func
    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>

    %0 = tf_executor.graph {
      %out, %ctl = tf_executor.island wraps "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @serving_default} : (tensor<?x!tf_type.string>) -> tensor<*xi64>
      tf_executor.fetch %out : tensor<*xi64>
    }
    return %0 : tensor<*xi64>
  }

        tf_executor.graph {
          %control_0 = tf_executor.island wraps "tf.PartitionedCall"() <{config = "", config_proto = "", executor_type = "", f = @function_1}> : () -> ()
          %control_1 = tf_executor.island wraps "tf.PartitionedCall"() <{config = "", config_proto = "", executor_type = "", f = @function_2}> : () -> ()

  attributes {tf.entry_function = {control_outputs = "", inputs = "serving_default_input_tensor:0", outputs = "PartitionedCall:0"}, tf_saved_model.exported_names = ["serving_default"]} {
    %0 = "tf.PartitionedCall"(%arg0) {f = @__inference_main} : (tensor<1x5x5x1024xf32>) -> tensor<1x5x5x1024xf32>
    return %0 : tensor<1x5x5x1024xf32>
  }