Rethinking Structure Learning For Graph Neural Networks