Model-driven Code Optimization

Although code optimizations have been applied by compilers for over 40 years, much of the research has been devoted to the development of particular optimizations. Certain problems with the application of optimizations have yet to be addressed, including when, where and in what order to apply optimizations to get the most benefit. A number of occurring events demand these problems to be considered. For example, cost-sensitive embedded systems are widely used, where any performance improvement from applying optimizations can help reduce cost. Although several approaches have been proposed for handling some of these issues, there is no systematic way to address the problems.This dissertation presents a novel model-based framework for effectively applying optimizations. The goal of the framework is to determine optimization properties and use these properties to drive the application of optimizations. This dissertation describes three framework instances: FPSO for predicting the profitability of scalar optimizations; FPLO for predicting the profitability of loop optimizations; and FIO for determining the interaction property. Based on profitability and the interaction properties, compilers will selectively apply only beneficial optimizations and determine code-specific optimization sequences to get the most benefit. We implemented the framework instances and performed the experiments to demonstrate their effectiveness and efficiency. On average, FPSO and FPLO can accurately predict profitability 90% of the time. Compared with a heuristic approach for selectively applying optimizations, our model-driven approach can achieve similar or better performance improvement without tuning the parameters necessary in the heuristic approach. Compared with an empirical approach that experimentally chooses a good order to apply optimizations, our model-driven approach can find similarly good sequences with up to 43 times compile-time savings.This dissertation demonstrates that analytic models can be used to address the effective application of optimizations. Our model-driven approach is practical and scalable. With model-driven optimizations, compilers can produce higher quality code in less time than what is possible with current approaches