Optimizing compilers often sacrifice binary size in pursuit of higher run-time performance. In the absence of method execution profiles, they uniformly apply performance-oriented optimizations, typically various forms of code duplication. Duplications in methods that are rarely or never executed only increase binary size without improving performance. Modern static profiler use ML to predict branch profiles, yet they do not identify which methods will be frequently executed at run time. Doing so would enable more selective optimizations, reducing binary size while preserving or only minimally affecting run-time performance.

We present GraalMHC, a machine–learning–based static profiler that predicts method hotness. GraalMHC uses the XGBoost ensemble to classify methods as cold and warm. For cold methods, GraalMHC enables code-size-reducing optimizations, and for warm methods, it enables performance-improving optimizations. In this way, GraalMHC enables binary-size reductions with no or minimal impact on run-time performance. In addition, GraalMHC allows users to choose between three different size-optimization levels: (S1) 9–13% binary-size reduction with 1–2% performance loss, (S2) 15–25% reduction with 3–5% performance loss, and (S3) 17–35% reduction with 5–7% performance loss. We integrate GraalMHC into the Oracle GraalVM Native Image compiler, delivering a complete end-to-end solution.