The integration of knowledge graphs (KG) into recommender systems has been proven to effectively alleviate data sparsity. A recent technical trend in knowledge-aware studies is to develop multi-task learning models founded on contrastive learning (CL). However, the current CL-based methods have not taken into account the issue of the long-tail distributions of the user-item graph and the KG, leading to sub-optimal performances. This paper proposes an approach called self-augmented contrastive learning (SACL) that addresses long-tail problems by learning to supplement missing information in their neighborhood. A view generator is first designed to transform head nodes into pseudo-tail nodes that simulate the distribution of real tail nodes, and a knowledge transfer function is then proposed to normalize the tail nodes and obtain pseudo-head nodes. Following this, we apply the principles of adversarial game theory to the real head and tail nodes as well as the pseudo-head and tail nodes to naturally enable model-agnostic contrastive learning through the self-augmented representations of graph embeddings of the user-item graph and the KG. The results of extensive comparative and ablation experiments on three public datasets demonstrated that our proposed SACL outperforms state-of-the-art approaches. Our code will be available on https://github.com/bangzuo/SACL .