Evolutionary Optimization of Compiler Flag Selection by Learning and Exploiting Flags Interactions

Unai Garciarena, Roberto Santana

Abstract: Compiler flag selection can be an effective way to increase the quality of executable code according to different code quality criteria. Evolutionary algorithms have been successfully applied to this optimization problem. However, previous approaches have only partially addressed the question of capturing and exploiting the interactions between compilation options to improve the search. In this paper we deal with this question comparing estimation of distribution algorithms (EDAs) and a traditional genetic algorithm approach. We show that EDAs that learn bivariate interactions can improve the results of GAs for some of the programs considered. We also show that the probabilistic models generated as a result of the search for optimal flag combinations can be used to unveil the (problem-dependent) interactions between the flags, allowing the user a more informed choice of compilation options.