Recent advances in machine learning have significantly enhanced the capabilities of static profilers, enabling them to capture complex program behaviors beyond traditional hand-crafted heuristics. While static profilers offer advantages such as easy integration and high resource efficiency compared with dynamic profilers, their accuracy remains limited because they rely on static program features and lack runtime information. Static profilers produce lower-quality profiles that lead to less optimized programs. However, the field still lacks a detailed breakdown across profile types and a systematic estimate of the potential of static profilers.

This paper provides the first comprehensive evaluation of the potential and practical upper bounds of ML-driven static profiling techniques across a diverse range of program profiles, including branch execution probabilities, method call counts, and profiles of virtual calls and monitors. We propose an idealized framework for static profiling and identify new avenues for static profiling beyond conventional branch prediction. Our results clarify the inherent trade-offs between static and dynamic approaches and reveal opportunities to expand the role of static profiling within modern optimizing compilers.