Recent years have witnessed fast developments of graph neural networks (GNNs) that have benefited myriads of graph analytic tasks and applications. In general, most GNNs depend on the homophily assumption that nodes belonging to the same class are more likely to be connected. However, as a ubiquitous graph property in numerous real-world scenarios, heterophily, i.e., nodes with different labels tend to be linked, significantly limits the performance of tailor-made homophilic GNNs. Hence, \textit{GNNs for heterophilic graphs} are gaining increasing attention in this community. To the best of our knowledge, in this paper, we provide a comprehensive review of GNNs for heterophilic graphs for the first time. Specifically, we propose a systematic taxonomy that essentially governs existing heterophilic GNN models, along with a general summary and detailed analysis. Furthermore, we summarize the mainstream heterophilic graph benchmarks to facilitate robust and fair evaluations. In the end, we point out the potential directions to advance and stimulate future research and applications on heterophilic graphs.