Suggesting Loop Unrolling Using a Heuristic-guided Approach

The development of embedded applications typically faces memory and/or execution time con-straints. In order to improve performance advanced compilers may use code transformations. There are cases where code transformations are not automatically applied and it is up to the devel-oper to manually apply them. One of the most relevant code transformations is Loop Unrolling. It is a widely studied transformation and it is able to improve the performance of many loops. It is easily implemented and its applicability is always legal. The main goal of this dissertation is to propose an approach to help developers decide about the application of Loop Unrolling and about the unroll factor to use. We propose an approach that uses a set of heuristics, based on char-acteristics extracted from the source code of the loops being analyzed, to suggest loops for Loop Unrolling. These heuristics are based on characteristics that likely lead to performance gains when Loop Unrolling is applied, but also on characteristics that can make a loop inadequate for Loop Unrolling. The approach consists in software extensions of an existent source-to-source tool to extract features and on a software tool implementing the heuristics. To validate and evaluate the proposed approach we developed a prototype that targets th