// CHECK-NEXT:    return [[RES]], [[MATRES]] : tensor<1x1xf32>, tensor<1x1xf32>
//
module {
  func.func @serving_default(%arg0: tensor<1x3xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>) {
    %0 = "tf.VarHandleOp"() <{container = "", shared_name = "y"}> : () -> tensor<!tf_type.resource<tensor<3x1xf32>>>
    %2 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<3x1xf32>>>) -> tensor<3x1xf32>

}

// CHECK-LABEL: QuantizeConcat
func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

// CHECK:           return %[[VAL_3]] : tensor<3x6xf32>
// CHECK:         }
func.func @concat_v2_1d_axis(%arg0: tensor<3x3xf32>, %arg1: tensor<3x3xf32>) -> tensor<3x6xf32> {
  %2 = "mhlo.concatenate"(%arg0, %arg1) <{dimension = 1 : i64}> : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x6xf32>
  func.return %2 : tensor<3x6xf32>
}

}

func.func @atan2(%arg0: tensor<8xf32>, %arg1: tensor<8xf32>) -> tensor<8xf32> {
  %0 = "tf.Atan2"(%arg0, %arg1) : (tensor<8xf32>, tensor<8xf32>) -> tensor<8xf32>
  func.return %0 : tensor<8xf32>

// CHECK-LABEL: atan2
// CHECK: %[[RES0:.*]] = "tfl.atan2"(%arg0, %arg1) : (tensor<8xf32>, tensor<8xf32>) -> tensor<8xf32>
// CHECK:  return %[[RES0]] : tensor<8xf32>
}

  %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>
  func.return %3 : tensor<2x1xf32>
})";
  const std::string kExpectedFB = CreateRuntimeMetadata();
  mlir::DialectRegistry registry;
  registry.insert<mlir::TFL::TensorFlowLiteDialect, mlir::arith::ArithDialect,

  func.return %24 : tensor<1x4xf32>
}

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir

    %7 = mhlo.add %arg1, %arg2 : tensor<f32>
    mhlo.return %7 : tensor<f32>
  }) {window_dimensions = dense<3> : tensor<2xi64>} : (tensor<8x8xf32>, tensor<f32>) -> tensor<6x6xf32>
  %5 = "mhlo.broadcast_in_dim"(%4) {broadcast_dimensions = dense<[2, 3]> : tensor<2xi64>} : (tensor<6x6xf32>) -> tensor<1x3x6x6xf32>
  %6 = mhlo.divide %3, %5 : tensor<1x3x6x6xf32>
  return %6 : tensor<1x3x6x6xf32>
}

// CHECK-LABEL:   func.func @avg_pool2d_2(

    } else {
      // Recurse on the subtypes in the variant/resource. Basically if the input
      // were:
      //   tensor<!tf_type.variant<tensor<?x8xf32>>>
      // and:
      //   tensor<!tf_type.variant<tensor<10x8xf32>>>
      // we'll try here to refine tensor<?x8xf32> with tensor<10x8xf32>.
      auto refined_subtype = mlir::cast<TensorType>(
          TypeMeet(lhs_element_type_with_subtype.GetSubtypes().front(),

//                (tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>)
//
// will be converted into:
//
//   %0 = "mhlo.slice"(%input) {
//             limit_indices = dense<[4, 2]> : tensor<2xi64>,
//             start_indices = dense<0> : tensor<2xi64>,
//             strides = dense<1> : tensor<2xi64>} :
//        (tensor<4x6xf32>) -> tensor<4x2xf32>