Multi-objective Exploration for Compiler Optimizations and Parameters

Abstract

AbstractIdentifying the suitable set of optimization options and determining nearly optimal values for the compiler parameter set in modern day compilers is a combinatorial problem. These values not only depend on the underlying target architecture and application source, but also on the optimization objective. Standard optimization options provide inferior solutions and also often specific to a particular optimization objective. Most common requirement of the current day systems is to optimize with multiple objectives, especially among average execution time, size and power. In this paper we apply Genetic Algorithm using Weighted Cost Function to obtain the best set of optimization options and optimal parameter set values for the multi-objective optimization of average execution time and code size. The effectiveness of this approach is demonstrated with the benchmark programs from SPEC 2006 benchmark suite. It is observed that the results obtained with parameter tuning and optimization option selection are better or equal to the results obtained with ‘-Ofast’ option in terms of execution time and at the same time equal to ’-Os’ option in terms of code size.KeywordsCompiler OptimizationOptimization Option SelectionParameter TuningGenetic AlgorithmsMulti-Objective Optimization