Reframing of Classification and Regression Tasks for Predicting the Effects of Compiler Settings on Multiple Embedded Systems

Compiler settings can have a significant impact on the performance of software but the task of finding effective configurations is time-consuming due to the large number of optimizations available and complex interactions between them. Furthermore, effective configurations are dependent on the target program and architecture. Previous work used prior knowledge about the performance of similar training programs in order to predict good configurations but these methods required retraining for each new architecture. In this paper we identify interesting classification and regression tasks for evaluating the performance of compiler configurations on two embedded system architectures. We show how a model learned for one architecture is not directly applicable to the other architecture and we discuss potential ideas for modeling the shift between architectures which in future could allow for the reuse of a single model on many platforms rather than retraining for each new system.