Machine Learning Techniques in Compiler Design

The manual generation of heuristics for compiler optimizations is a tedious task which requires both a high amount of expertise and an extensive trial-and-error tuning. The consequence is that many optimizations operate inefficiently. In this chapter, an automatic generation of machine learning based (MLB) heuristics for WCET-aware compiler optimizations is presented. A feasibility study for the generation of WCET-aware heuristics for the source code optimization function inlining is presented. Moreover, this chapter shows that on the one hand MLB WCET-aware heuristics can also be easily constructed at other abstraction levels of the code, namely at assembly level as demonstrated for the optimization loop-invariant code motion. On the other hand, this chapter indicates that an automatic selection of appropriate machine learning algorithms and their parameters can significantly improve the performance of the generated heuristics. To demonstrate the effectiveness of the MLB heuristics, evaluations on real-life benchmarks were performed, showing that the novel heuristics outperform traditional, hand-crafted heuristics.