Profile-Driven Compile-Time Optimizations Targeting Commodity Desktop Environments

点击翻译按钮获取中文摘要

While the commodity desktop environment is growing in popularity, researchers continue to study architecture-based research questions using standard benchmark programs. Benchmarks possess behavior that is fundamentally different from the behavior of desktop applications. For instance, the reliance of desktop applications on the hosting operating system and dynamically linked libraries warrants careful consideration when pursuing performance-driven architectural tradeoffs targeting a desktop environment.In the future these applications will be developed in object-oriented languages. Again, this introduces a different program model that has not been rigorously addressed in the architectural research community. Object-oriented programs place significant pressure on instruction caches and branch prediction mechanisms. The focus of this research is to provide compile-time optimizations that take advantage of the profile information currently available in the desktop application environment. The goal is to develop new profile-based compiler optimization algorithms that can significantly accelerate the performance of desktop applications.This project pursues the following research areas: Parameter Value Prediction, Indirect Branch Prediction, Instruction Memory Coloring, Weighted Completion Time Instruction Scheduling. Graph-based program analysis using program profiles is used to guide these optimizations. Improved optimization techniques are employed to produce tuned executable images. These optimizations are implemented in existing compilation environments to quantify their performance gains.