Long-tail recommendation has emerged as a promising solution to mitigate the popularity bias inherent in graph-based recommender systems. Recently, contrastive learning-based models have shown their effectiveness in this field. However, typical data augmentation strategies relying on random edge dropout or noise injection are often ill-suited to the sparsity and imbalance inherent in long-tail data. In this paper, we propose two contrastive learning strategies, combined in the framework GCLLT, that are specifically designed for long-tail recommendation. The popularity-aware contrastive learning strategy enhances tail-item connectivity by linking tail nodes to nearby head nodes in the embedding space and pruning less informative edges from head items. This results in cleaner and more informative graph views for tail items. In neighborhood-aware contrastive learning, head and tail items are clustered separately, and each tail item is assigned the centroid of its nearest head cluster as a prototype, enabling semantic supervision without relying on structural augmentation. Finally, these contrastive learnings are applied to tail items, allowing their representations to be refined without being dominated by head items during optimization. Experimental results on real-world datasets demonstrate that GCLLT outperforms baseline models, achieving up to 10.25% improvement in Recall@20.