Automatic WCET Reduction by Machine Learning Based Heuristics for Function Inlining

The application of machine learning techniques in compiler frameworks has become a challenging research area. Learning algorithms are exploited for an automatic generation of optimization heuristics which often outperform hand-crafted models. Moreover, these automatic approaches can effectively tune the compilers’ heuristics after larger changes in the optimization sequence or they can be leveraged to tailor heuristics towards a particular architectural model. Previous works focussed on a reduction of the average-case performance. In this paper, learning approaches are studied in the context of an automatic minimization of the worst-case execution time (WCET) which is the upper bound of the program’s maximum execution time. We show that explicitly taking the new timing model into account allows the construction of compiler heuristics that effectively reduce the WCET. This is demonstrated for the well-known optimization function inlining. Our WCET-driven inlining heuristics based on a fast classifier called random forests outperform standard heuristics by up to 9.1% on average in terms of the WCET reduction. Moreover, we point out that our classifier is highly accurate with a prediction rate for inlining candidates of 84.0%.