func.func @do_not_annotate_without_replicate(%arg0: tensor<?xi32>) -> tensor<?xi32> {
    %0 = "tf._A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
    %1 = "tf._B"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
    %2 = "tf_device.cluster_func"(%0, %1) {func = @_func, device = ""} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf._C"(%2) : (tensor<?xi32>) -> tensor<?xi32>
    func.return %3 : tensor<?xi32>
  }

  // CHECK-LABEL: func @_func

	chki8(int8(i16), ci16&0xff)
	chki8(int8(i32), ci32&0xff-1<<8)
	chki8(int8(i64), ci64&0xff)
	chki8(int8(u8), cu8&0xff-1<<8)
	chki8(int8(u16), cu16&0xff)
	chki8(int8(u32), cu32&0xff)
	chki8(int8(u64), cu64&0xff)
	//	chki8(int8(f32), 0)
	//	chki8(int8(f64), 0)

	chki16(int16(i8), ci8&0xffff-1<<16)
	chki16(int16(i16), ci16&0xffff-1<<16)
	chki16(int16(i32), ci32&0xffff-1<<16)
	chki16(int16(i64), ci64&0xffff-1<<16)

  let arguments = (ins
     TFL_TensorOf<[I8, I16, I32, I64, UI8, UI16, UI32, UI64, F32, F64]>:$input,
     TFL_TensorOf<[I32]>:$dilations,
     TFL_0DTensorOf<[I8, I16, I32, I64, UI8, UI16, UI32, UI64, F32, F64]>:$padding_value
  );

  let results = (outs TFL_TensorOf<[I8, I16, I32, I64, UI8, UI16, UI32, UI64, F32, F64]>:$output);
}

def TFL_AddOp : TFL_Op<"add", [

func.func @fold_device_index() -> tensor<i32> {
  // CHECK-NOT: tf.DeviceIndex
  // CHECK: tf.Const
  // CHECK-SAME: value = dense<1> : tensor<i32>
  %0 = "tf.DeviceIndex"() {device = "/device:CPU:0", device_names = ["GPU", "CPU"]} : () -> tensor<i32>
  func.return %0 : tensor<i32>
}

// -----

// CHECK-LABEL: @not_fold_device_index
func.func @not_fold_device_index() -> tensor<i32> {
  // CHECK-NOT: tf.Const

}

func.func @body(%arg0: tensor<*xi32>, %arg1: tensor<*xf32>) -> (tensor<*xi32>, tensor<*xf32>) {
  %0:2 = tf_executor.graph {
    %outputs, %control = tf_executor.island wraps "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
    %outputs_0, %control_1 = tf_executor.island wraps "tf.Sub"(%arg0, %outputs) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>

  %max_iter = "tf.Const"() {__op_key = 1, value = dense<2> : tensor<i32>} : () -> tensor<i32>
  // CHECK: tf_mlrt.map_fn
  %result = "tf_mlrt.tf_map_fn"(%max_iter, %input2, %2) { operandSegmentSizes = array<i32: 1, 1, 1>, body_fn = @NopMapFnBody, num_tensor_list_or_flow_in = 1 : i32} : (tensor<i32>, tensor<!tf_type.variant<tensor<*xf32>>>, tensor<i32>) -> tensor<i32>
  return %result : tensor<i32>
}

// CHECK-LABEL: @NopMapFnBody

    %i32_max_filled = "tf.Fill" (%input_shape, %i32_max) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
    %i32_act_max_f32_filled = "tf.Fill" (%input_shape, %i32_act_max_f32) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
    %clip_min = "tf.Maximum"(%i32_min_filled, %input_zp) : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>

    // CHECK: [[TENSOR:%.+]] = "tf.TensorScatterAdd"([[READ]], [[EXPAND]], [[UPDATE]]) : (tensor<*xi32>, tensor<?x1xi32>, tensor<?x?x?xi32>) -> tensor<*xi32>
    // CHECK: "tf.AssignVariableOp"([[VAR]], [[TENSOR]])
    "tf.ResourceScatterAdd"(%resource, %indices, %updates) : (tensor<*x!tf_type.resource<tensor<*xi32>>>, tensor<?xi32>, tensor<?x?x?xi32>) -> ()

    tf_device.return

module attributes {tf.versions = {producer = 888 : i32}} {
  func.func @main(%arg0: tensor<?xi32>) -> tensor<?xi32> {
    %0 = "tf.Identity"(%arg0) : (tensor<?xi32>) -> tensor<?xi32> loc(unknown)
    func.return %0 : tensor<?xi32> loc(unknown)
  } loc(unknown)
} loc(unknown)

    "tf.AssignVariableOp"(%1, %3) {dtype = i32} : (tensor<*x!tf_type.resource<tensor<*xi32>>>, tensor<*xi32>) -> ()
    %4 = "tf.ReadVariableOp"(%1) {dtype = i32} : (tensor<*x!tf_type.resource<tensor<*xi32>>>) -> tensor<*xi32>
    tf_device.return %4 : tensor<*xi32>
  }) {cluster_attr = "cluster_attr"} : () -> tensor<*xi32>

  // CHECK: return %[[CLUSTER_RES]]
  return %1 : tensor<*xi32>
}