Solução integrada para os problemas de seleção e ordenação de fase

Abstract

Modern compilers ofIer several optimizations to be applied to the source code of a program to order to increase its performance. Due to the complex relationship which transformations have for each difIerent program, to find out the best optimizations set and the best order to apply them with no human interference are two hard problems which it is faced by compilers designers. Iterative compilation techniques try to mitigate these two problems evaluating compiled program performance with several sequences and choosing the best generated version. Good results are obtained, but in most cases requirement of a extremely big number of application executions is a limitative factor. Approaches use machine learning restrict quantity of evaluations, but usually they do not overcome iterative strategies. Besides, it is scarse research of strategies to try to integrate two or more from these techniques to mitigate problem of optimizations selecting and ordering at the same time. In this context, this work presents a strategy to mitigate phase selection problem and another one to phase ordering problem, which uses a prior knowledge structure to select and order sequences. Two approaches were implemented in a programs optimization framework and, despite they employ very difIerent strategies, they are integrated to provide a well-ordered high performance sequence. Experiments results lead to programs from SPEC2006 and cBench demonstrated integrated approach can optimize that applications consistently.