%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>

///      : (tensor<f32>, tensor<f32>, tensor<f32>) -> tensor<5xf32>
///
/// Output would be:
///   %iota = "mhlo.iota"() {iota_dimension = 0 : i64} : () -> tensor<5xf32>
///   %scaled = "mhlo.multiply"(%iota, %delta)
///       {broadcast_dimensions = dense<[]> : tensor<0xi64>} :
///       (tensor<5xf32>, tensor<f32>) -> tensor<5xf32>
///   %result = "mhlo.add"(%scaled, %offset)

          %1 = call @then_branch_func(%arg1) : (tensor<*xf32>) -> tensor<*xf32>
          "tf.Yield"(%1) : (tensor<*xf32>) -> ()
        },  {
          %1 = call @else_branch_func(%arg1) : (tensor<*xf32>) -> tensor<*xf32>
          "tf.Yield"(%1) : (tensor<*xf32>) -> ()
        }) {is_stateless = false} : (tensor<i1>) -> tensor<*xf32>
    ```

    will be transformed into this functional operation

  // Look for resource or variant element type and ensure we refine the subtype.
  // We only support a single subtype at the moment, we won't handle something
  // like:
  //   tensor<!tf_type.variant<tensor<10xf32>, tensor<8xf32>>
  if (rhs_element_type_with_subtype &&
      rhs_element_type_with_subtype.GetSubtypes().size() == 1) {
    auto lhs_element_type_with_subtype =
        mlir::dyn_cast<TF::TensorFlowTypeWithSubtype>(lhs_element_type);