func.return %0 : tensor<1x112x112x32xf32>
}

func.func @fully_connected(%arg0: tensor<1x37xf32>, %arg1: tensor<40x37xf32>, %arg2: tensor<40xf32>) -> tensor<1x40xf32> {
  // CHECK: _arithmetic_count = 3000 : i64
  %0 = "tfl.fully_connected"(%arg0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> tensor<1x40xf32>
  func.return %0 : tensor<1x40xf32>
}

  %one = "glue.constant"() { value = 1: i32 } : () -> i32
  %done = "glue.compare" (%one, %one) { predicate = #glue<"compare LTE"> } : (i32, i32) -> i1
  %2 = mhlo.constant dense<[[1.1]]> : tensor<1x1xf32>
  %3 = mhlo.multiply %2, %2 : tensor<1x1xf32>
  %cst = "tf.Const"() {value = dense<0.0> : tensor<f32>} : () -> tensor<f32>
  %0 = "tf.AddV2"(%arg0, %cst) {device = "/device:CPU:0"} : (tensor<f32>, tensor<f32>) -> tensor<f32>

  %9 = "mhlo.compare"(%0, %8) {comparison_direction = #mhlo<comparison_direction NE>} : (tensor<10x8xf32>, tensor<10x8xf32>) -> tensor<10x8xi1>
  %10 = mhlo.and %7, %9 : tensor<10x8xi1>
  %11 = mhlo.subtract %arg0, %6 : tensor<10x8xf32>
  %12 = mhlo.divide %11, %5 : tensor<10x8xf32>
  %13 = mhlo.add %12, %3 : tensor<10x8xf32>

    ```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

  %0:3 = "tfl.unpack"(%arg0) {axis = 1 : i32, num = 3 : i32} : (tensor<1x3x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>)
  %1 = "tfl.concatenation"(%0#0, %0#1, %0#2) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x6xf32>
  func.return %1 : tensor<1x6xf32>

func.func private @batched_function(%arg0: tensor<1x3xf32>, %arg1: tensor<*x!tf_type.resource>) -> tensor<1x3xf32> {
  %0 = "tf.Identity"(%arg0) : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %0 : tensor<1x3xf32>
}

// -----

func.func @test_batch_function_with_invalid_symbol(%arg0: tensor<1x3xf32>, %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> () {

  // CHECK: %[[OLD_SLICE1:.*]] = "tf.Slice"(%[[READ1]],
  // CHECK: %[[RESHAPE1:.*]] = "tf.Reshape"(%[[VALUE]],
  // CHECK: %[[ADD1:.*]] = "tf.AddV2"(%[[RESHAPE1]], %[[OLD_SLICE1]]) : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  // CHECK: %[[UPDATE1:.*]] = "tf.XlaDynamicUpdateSlice"(%[[READ1]], %[[ADD1]],
  // CHECK: "tf.AssignVariableOp"(%[[GVAR1]], %[[UPDATE1]])

  %cst = arith.constant dense<[-1, 7]> : tensor<2xi32>
  %0 = "tf.TensorListReserve"(%cst, %arg0) : (tensor<2xi32>, tensor<i32>) -> tensor<!tf_type.variant<tensor<?x7xf32>>>
  %1 = "tf.TensorListGetItem"(%0, %arg1, %cst) : (tensor<!tf_type.variant<tensor<?x7xf32>>>, tensor<i32>, tensor<2xi32>) -> tensor<?x7xf32>
  func.return %1 : tensor<?x7xf32>

      // 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(),

  // CHECK-DAG: %[[ITEMS0_0:.*]] = "tf.ExpandDims"(%[[ITEMS0]], %[[AXIS]])
  // CHECK-DAG: "tf.ConcatV2"(%[[ITEMS1_3]], %[[ITEMS1_2]], %[[ITEMS1_1]], %[[ITEMS1_0]], %[[ITEMS0_0]], %[[AXIS]]) : (tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>, tensor<i64>) -> tensor<5x2xf32>

  %indices0 = "tf.Const"() {value = dense<4> : tensor<i32>} : () -> tensor<i32>