Surrogate-Assisted Multi-objective Optimization for Compiler Optimization Sequence Selection

Abstract

Compiler developers typically design various optimization options to produce optimized programs. Generally, it is a challenging task to identify a reasonable set of optimization options (i.e., compiler optimization sequence) in modern compilers. Optimization objectives, in addition to the target architecture and source code of the program, influence the selection of optimization sequences. Current applications are often required to optimize two or more conflicting objectives simultaneously, such as execution time and code size. Existing approaches employ evolutionary algorithms to find appropriate optimization sequences to trade off the above two objectives. However, since program compilation and execution are time-consuming, and the two objectives are inherently conflicting, applying evolutionary algorithms faces the diverse objectives influence and computationally expensive problem. In this study, we present a surrogate-assisted multi-objective optimization approach. To speed up the convergence, it employs a fast global search based on non-dominated sorting. The approach then uses two surrogate models for each objective to generate approximate fitness evaluations rather than using actual expensive evaluations. Extensive experiments on the benchmark suite cBench show that our approach outperforms the baseline NSGA-II on hypervolume by an average of 11.7%. Furthermore, experiments verify that the surrogate model contributes to solving the computationally expensive problem and taking fewer actual fitness evaluations.