A Framework for Exploring Optimization Properties

Important challenges for compiler optimization include determining what optimizations to apply, where to apply them and what is a good sequence in which to apply them. To address these challenges, an understanding of optimization properties is needed. We present a model-based framework, FOP, to determine how optimizations enable and disable one another. We combine the interaction and profitability properties to determine a "best" sequence for applying optimizations. FOP has three components: (1) a code model for the program, (2) optimization models that capture when optimizations are applicable and their actions, and (3) a resource model that expresses the hardware resources affected by optimizations. FOP determines interactions by comparing the create- and destroy-conditions of each optimization with the post conditions of other optimizations. We develop a technique with FOP to construct code-specific optimization sequences. Experimentally, we demonstrate that our approach achieves similarly good code quality as empirical techniques with less compile-time.