%2 = "tf.AddV2"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    // CHECK: %[[SUB:.*]] = "tf.Sub"
    %3 = "tf.Sub"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    // CHECK: tf_device.return %arg0, %[[ADD]], %[[MUL]], %[[SUB]]
    tf_device.return %arg0, %2, %0, %3 : tensor<i32>, tensor<i32>, tensor<i32>, tensor<i32>
  }) : () -> (tensor<i32>, tensor<i32>, tensor<i32>, tensor<i32>)

    %Enter, %ctl_0 = Enter(%Const) [%arg.ctl] name("while/Enter") {T = i32, frame_name = "while/while_context", is_constant = false, parallel_iterations = 10 : i64} : (tensor<i32>) -> (tensor<*xi32>)
    %NextIteration, %ctl_1 = NextIteration(%Const) [%arg.ctl] name("while/NextIteration") {T = i32} : (tensor<i32>) -> (tensor<*xi32>)

                      %n: i32 {tfr.name="N"},
                      %axis: i32 {tfr.name="axis"}) -> !tfr.tensor {
  %index = arith.constant 0 : index
  %cst = arith.constant 1 : i32
  %eq = arith.cmpi eq, %n, %cst : i32
  %v1 = tfr.get_element %values[%index] : (!tfr.tensor_list, index) -> !tfr.tensor
  %temp = tfr.call @tf__expand_dims(%v1, %axis) : (!tfr.tensor, i32) -> !tfr.tensor
  %res = scf.if %eq -> !tfr.tensor {

  %1 = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
  %2 = "tf.Add"(%0, %1) : (tensor<i32>, tensor<i32>) -> (tensor<i32>)
  func.return %2 : tensor<i32>
}

func.func @stateful_pcall_func(%arg0: tensor<i32>) -> tensor<i32> {
  func.return %arg0 : tensor<i32>
}

// -----

  "func.func"() ({
  ^bb0(%arg0: tensor<i32>):
    "func.return"(%arg0) : (tensor<i32>) -> ()
  }) {sym_name = "main_while_body_4225150", sym_visibility = "private", tf._input_shapes = [#tf_type.shape<>], tf.signature.is_stateful, function_type = (tensor<i32>) -> (tensor<i32>)} : () -> ()
  "func.func"() ({
  ^bb0(%arg0: tensor<i32>):

     %2 = builtin.unrealized_conversion_cast to tensor<i32>
     %3 = builtin.unrealized_conversion_cast to tensor<i32>
     // CHECK: TPUCopyWithDynamicShape{{.*}}device = "foobar"
     %4, %5 = "tf.TPUCopyWithDynamicShape"(%0, %1, %2, %3) {operandSegmentSizes = array<i32: 2, 2>} : (tensor<2048xi32>, tensor<2048xi32>, tensor<i32>, tensor<i32>) -> (tensor<2048xi32>, tensor<2048xi32>)

func.func @while_cond(%arg0: tensor<i32>) -> tensor<i1> {
  %0 = "tf.Const"() {value = dense<9> : tensor<i32>} : () -> tensor<i32>
  %1 = "tf.Less"(%arg0, %0) {} : (tensor<i32>, tensor<i32>) -> tensor<i1>
  func.return %1 : tensor<i1>
}

func.func @while_body(%arg0: tensor<i32>) -> tensor<i32> {
  %1 = "tf.AddV2"(%arg0, %arg0) {} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  func.return %1 : tensor<i32>
}

  func.return %0 : tensor<i32>
}

// -----

func.func @Callee(%arg0: tensor<i32>) -> tensor<i32> {
  "tf.XlaSendToHost"(%arg0) {key = "send_key_call3"} : (tensor<i32>) -> ()
  %1 = "tf.XlaRecvFromHost"() {key = "recv_key_call3", shape = #tf_type.shape<>} : () -> tensor<i32>
  func.return %1 : tensor<i32>
}

func.func @target_function() -> tensor<i32> {
  %0 = "tf_device.cluster"() ({
    // CHECK: tf.UnsupportedOp
    %1 = "tf.UnsupportedOp"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
    // CHECK: tf.Identity
    %2 = "tf.Identity"(%1) : (tensor<i32>) -> tensor<i32>
    // CHECK-NOT: tf.Identity
    %3 = "tf.Identity"(%2) : (tensor<i32>) -> tensor<i32>
    %4 = "tf.Identity"(%3) : (tensor<i32>) -> tensor<i32>
    // CHECK: tf_device.return

module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    // CHECK-LABEL: func.func @main
    // CHECK-DAG: %[[CST_0:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<f32>}> : () -> tensor<f32>
    // CHECK-DAG: %[[CST_1:.*]] = "tf.Const"() <{value = dense<3> : tensor<i32>}> : () -> tensor<i32>