ML-SOCO: Machine Learning-Based Self-Optimizing Compiler Optimizations

Compiler optimizations often involve hand-crafted heuris- tics to guide the optimization process. These heuristics are designed to benefit the average program and are otherwise static or only customized by profiling information. We pro- pose machine learning-based self optimizing compiler op- timizations (ML-SOCO), a novel approach for fitting opti- mizations in a dynamic compiler to a specific environment. ML-SOCO explores—at run time—the impact of optimization decisions and uses this data to train or update a machine learning model. Related work which has primarily targeted static compilers has already shown that machine learning can outperform human-crafted heuristics. Our approach is specifically tailored to dynamic compilation and uses con- cepts like deoptimization for transparently switching be- tween generating data and performing machine learning decisions during compilation. We implemented ML-SOCO in the GraalVM compiler which is one of the most highly optimizing Java compilers on the market. When evaluat- ing ML-SOCO by replacing a loop peeling heuristics with a learned model we encountered multiple speedups larger than 30% in established benchmarks. Apart from improving the performance, ML-SOCO can also be used to assist compiler engineers when improving heuristics for specific domains.