JavART - A Lightweight Rule-Based JIT Compiler Using Translation Rules Extracted from a Learning Approach

Explore a conference presentation introducing a novel lightweight rule-based Just-In-Time (JIT) compiler designed to optimize Java compilation performance on mobile and client-side devices. Learn about JavART, an innovative approach that addresses the critical balance between compilation time and code quality by using translation rules automatically extracted through machine learning techniques from existing JIT compilers. Discover how this system bypasses traditional interpretive phases and first-level optimizations at runtime, instead generating high-quality native code directly through offline-learned translation rules. Examine the implementation details of a prototype built on Android 14 and understand how it achieves significant performance improvements, including 1.23× average speedup compared to default interpreter mode with two-tier JIT compilation, 5.5× faster native code generation than existing first-tier JIT compilers, and 6% faster execution of generated code. Analyze the research methodology that demonstrates effectiveness across real-world applications and explore the extension to client-side systems running HotSpot JVM with 1.20× average speedup. Gain insights into the technical challenges of JIT compilation for short-running, frequently restarted applications typical in Android environments, and understand how this learning-based approach represents a paradigm shift from traditional JIT compilation strategies by leveraging offline learning to inform runtime compilation decisions.