Social recommendation leverages friends’ preferences to improve user recommendation quality. However, recent studies reveal an inherent misalignment between the social and interaction graphs. Directly using these raw graphs leads to suboptimal performance. Recently, contrastive learning (CL) has been introduced to mitigate this issue. It treats the same user’s representations from social and interaction views as a positive pair, while representations of different users across the two views as negative pairs, aligning views by maximizing the similarity of positive pairs and minimizing that of negative pairs. Nevertheless, existing approaches treat all different users uniformly when constructing sample pairs, regardless of their behavioral similarity. As a result, highly similar users are pushed away like dissimilar ones, weakening alignment effectiveness. In this paper, we propose Friend-aware Contrastive Learning with Aligned Social Graph for Social Recommendation (FACAR). Specially, to alleviate the graph-level misalignment and facilitate subsequent processing, FACAR first performs preference-aware social relation mining by integrating user preferences into social relations to refine the social graph. Then, to avoid uniformly sampling pairs in CL so as to better align the two views in representations-level, we construct a Trust Friend Set for each user and design a friend-aware contrastive learning framework based on relations characterized above. Experiments on three benchmarks demonstrate that FACAR outperforms state-of-the-art baselines. Our implementation is available at: https://github.com/ZeasonJuan/FACAR .