Soluções para o problema da seleção de otimizações

Abstract

Compiler generated codes may not have the best quality possible because it is difficult to obtain the optimal sequence of instructions as it has endless possibilities. Compilers developers tried to improve code quality implementing some optimizations. However, when it is not used correctly, the application of optimizations may impair code quality. Among dozens of optimizations usually provided by a compiler, it is a challenge to know which ones will generate a better target code for a specific source code, even for the most experienced programmer. In this context, development of automated optimizations selectors is a challenge today. Approaches to the implementation of these selectors are found in the literature and include use of random, exhaustive and heuristics searches, genetic algorithms and machine learning. In view of the optimizations automatic selection problematic, this work presents four new approaches for selection of good optimizations sets. The experimental evaluation of this approaches showed that they make it possible a significant performance gain compared to approaches found in literature and suggests a wide applicability in various contexts.