An Effective Local Search Algorithm for an Adaptive Compiler

Most algorithms currently used to find good compilation sequences in an iterative adaptive compiler, such as genetic algorithms and hill climbing, search in the space of sequences of fixed length. In this paper, we argue that restricting the search to fixed-length sequences limits the ability of search algorithms to find good sequences for some benchmarks. We propose a new local search algorithm that uses greedy construction and cleanup to effectively explore the neighborhood of a start sequence by randomized insertion and deletion of transformations. Preliminary experimental results show that the quality of the local minima found by our local search algorithm are superior to those sequences found by GAs and HCs, and are close to the best sequence we know. Such local minima are found with significantly lower search effort than GAs and HCs working with fixed-length sequences.