Uma solução híbrida para mitigação do problema de seleção de otimizações

Abstract

Compiler optimizations are transformations, usually in intermediate representation of the code, that tries to improve the performance of the executable program. Choosing one sequence of these transformations, that change the code to its best possible state, is a complex task, and generally inviable. So, it is usually tried to mitigate the problem. In the context of Optimization Selection Problem, are applied two known approaches: iterative compilation and machine learning. This work shows an approach that mixes those two, applying iterative compilation in machine learning results. The obtained results with this approach, using support vector machine with genetic algorithm, show better results than a pure iterative compilation using a genetic algorithm, reaching speedups of 2,115x over program without optimizations, while a pure iterative compilation reached 2,074x.