HERCULES: Strong Patterns Towards More Intelligent Predictive Modeling

Abstract

Recent work has showed that program analysis techniques to select meaningful features of programs are important in predictive modeling on compiler optimization selection problems. Although there are many successful state-of-the-art program analysis techniques, they often do not provide enough loop information which can be helpful, especially when a target program is computation intensive with many complex loops. In this paper, we introduce a novel technique to characterize loops using a pattern-driven system named HERCULES and automatically select compiler optimizations. Using this system allows us to derive meaningful and expressive loop features which is not trivial to collect using other static techniques in state-of-the-art. We evaluate our loop features by building prediction models and compare the models based on three successful state-of-the-art program characterization techniques found in the literature. We show that our models outperform three state-of-the-art program characterization techniques introduced in this paper in terms of buildling a prediction model. We show that our models with loop features achieved up to 63% of the best speedup we can reach in our given optimization space for one of platforms we conducted experiments. In addition, we also show how prediction can be affected when we use our loop features and one of dynamic features together.