Static prediction of recursion frequency using machine learning to enable hot spot optimizations

Recursion poses a severe problem for static optimizations because its execution frequency usually depends upon runtime values, hence being rarely predictable at compile time. As a consequence, optimization potential of programs is sacrificed since possible hot paths where most of the execution time is spent and where optimization would be beneficial might be undiscovered. In this paper, we propose a sophisticated machine learning based approach to statically predict the recursion frequency of functions for programs in real-world application domains, which can be used to guide various hot spot optimizations. Our experiments with 369 programs of 25 benchmark suites from different domains demonstrate that our approach is applicable to a wide range of programs with different behavior and yields more precise heuristics than those generated by pure static analyses. Moreover, our results provide valuable insights into recursive structures in general, when they appear and how deep they are.