State Machine Replication (SMR) is a cornerstone for building highly available and consistent services. Modern systems, such as blockchains, push SMR scalability demands to tens or hundreds of nodes. Gossip-based communication has been proposed in this context to ensure scalable and resilient message dissemination. However, since gossip and consensus protocols are designed to tolerate failures and message loss, their combination introduces redundant overhead. This paper addresses this inefficiency by proposing Semantic Gossip, a cross-layer optimization that leverages consensus semantics to reduce gossip overhead without compromising safety or liveness. Semantic Gossip introduces semantic filtering to suppress redundant messages and semantic aggregation to combine similar messages, maintaining modularity and resilience. We implement these ideas atop Tendermint and demonstrate through experiments with 32 and 128 nodes that Semantic Gossip increases throughput by up to 1.24× and 3.42×, respectively, and reduces latency, while preserving consensus properties even under Byzantine failures and message loss. Our results show that semantic awareness in gossip communication can significantly enhance the scalability and efficiency of blockchain consensus.