return
  }
  // CHECK-LABEL: func private @foo.impl
  func.func private @foo.impl(%arg0: tensor<1xf32>, %arg1: tensor<2xf32>) -> (tensor<2xf32>, tensor<1xf32>) {
    return %arg1, %arg0 : tensor<2xf32>, tensor<1xf32>
  }

  %fill = "tf.Fill"(%cst_0, %cst) : (tensor<1xi64>, tensor<f32>) -> tensor<1xf32>
  tf_device.replicate([%0, %fill] as %arg_r0: tensor<1xf32>) {n = 2 : i32} {
    %1 = "tf_device.launch"() <{device = "TPU_REPLICATED_HOST_0"}> ({
      %2 = "tf.Identity"(%arg_r0) : (tensor<1xf32>) -> tensor<1xf32>
      tf_device.return %2 : tensor<1xf32>
    }) : () -> tensor<1xf32>

func.func @add(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Add"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0: tensor<1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0, %arg1) <{custom_code = "FlexAdd", custom_option = #tfl<const_bytes : "0x03416464001412034164641A001A002A070A015412023001320000021B171414042801">}> : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %2 = "tfl.add"(%arg0, %arg3) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>

module{
  func.func @main(%arg0: tensor<10xf32>) -> tensor<10xf32> {
    return %arg0 : tensor<10xf32>
  }
  func.func @func1(%arg0: tensor<10xf32>, %arg1: tensor<10xf32>) -> tensor<10xf32> {
    %0 = stablehlo.add %arg0, %arg1 : tensor<10xf32>
    %1 = stablehlo.add %arg0, %arg1 : tensor<10xf32>
    return %0 : tensor<10xf32>
  }
})mlir";

module {
  // CHECK-LABEL: main
  func.func @main(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
    %0 = "tfl.squared_difference"(%arg0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
    func.return %0 : tensor<4xf32>
    // CHECK:  [[VAL_0:%.*]] = tfl.sub %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<4xf32>

}

// -----

func.func @tensorlistPushBack(%arg0: tensor<3x10xf32>, %arg1: tensor<1xi32>, %arg2: tensor<10xf32>) -> tensor<?x10xf32> {
  %0 = "tf.TensorListFromTensor"(%arg0, %arg1) : (tensor<3x10xf32>, tensor<1xi32>) -> tensor<!tf_type.variant<tensor<10xf32>>>
  %1 = "tf.TensorListPushBack"(%0, %arg2) : (tensor<!tf_type.variant<tensor<10xf32>>>, tensor<10xf32>) -> tensor<!tf_type.variant<tensor<10xf32>>>

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<6xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

  %cst_0 = "tf.Const"() {value = dense<0.0235294122> : tensor<f32>} : () -> tensor<f32>
  %cst_1 = "tf.Const"() {value = dense<0.00708661414> : tensor<1xf32>} : () -> tensor<1xf32>
  %cst_2 = "tf.Const"() {value = dense<1.799000e+03> : tensor<1xf32>} : () -> tensor<1xf32>