Auto-Vectorization in HotSpot

Explore the ongoing development and improvements in HotSpot C2's auto vectoriser in this 43-minute conference talk from the 2025 JVM Language Summit. Begin with a brief introduction to the SuperWord algorithm before diving into significant enhancements already achieved and plans for future advancements, illustrated through real-world examples and benchmarks. Learn about dependency graph management for instruction reordering while preserving essential data and memory dependencies, and discover aliasing analysis techniques for safely packing loads and stores, with emphasis on dynamic aliasing checks for MemorySegment loops. Examine the profitability of vectorization by addressing challenges like reductions and lane-crossing operations with advanced cost modeling, and understand strategies to mitigate performance regressions including unaligned memory access, store-to-load forwarding failures, and cache line boundary splits. Discover future goals including ambitions for if-conversion, efficient scans such as prefix-sum and hash-code, more robust and interpretable vectorization algorithms, improved compilation times, and handling larger loop bodies. Gain technical insights and practical examples from current industry advancements in compiler optimization, vectorization, and JVM performance, presented by Emanuel Peter, Senior Member of Technical Staff from Oracle's Java Platform Group.