%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,

    }
  }
  return new_arg_type;
}

// Combination of value producer and port of value produced (e.g.,
//   <value result output>:<value in output tensor>,
// so for tf.Const -> tensor<10x20xf32>, [0,2,18] would point to a unique output
// scalar value).
struct ValuePort {
  PointerUnion<Operation*, BlockArgument> producer;
  SmallVector<unsigned int, 2> port;

mlir_module = '''python
func @main(%arg0 : tensor<10xf32>, %arg1 : tensor<10xf32>) -> tensor<10x10xf32> {
   %add = "magic.op"(%arg0, %arg1) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10x10xf32>
   return %ret : tensor<10x10xf32>
}
'''

@tf.function
def foo(x, y):
  return mlir_passthrough_op([x, y], mlir_module, Toutputs=[tf.float32])