Revisiting the Message Passing in Heterophilous Graph Neural Networks