Compiler Techniques for Improving Data Locality in Complex Program Constructs

随着处理器速度和内存延迟之间的差距不断加大，内存访问已经成为许多应用程序的关键性能瓶颈。这个项目开发新的编译器技术，通过提高内存层次结构的利用率来提高内存性能。研究对象包括单处理器工作站和并行计算机。设计新的程序转换是为了最大限度地重用存储在存储器层次结构中的快速存储器单元（如缓存）中的数据。这些新的转换处理了在实践中常见但以前无法处理的困难的程序结构。为了支持新的转换，开发了一个过程间分析来分析整个程序的内存访问行为。提出的程序分析和转换是在一个研究编译器中实现的。这些技术的成功开发可以显著提高各种应用程序的执行速度。

Zhiyuan LiCCR-9975309As the gap between processor speed and memory latency continues to increase, memory accesses have become the key performance bottleneck in many application programs. This project develops new compiler techniques to enhance memory performance by improving the utilization of the memory hierarchy. The research targets both uniprocessor workstations and parallel computers. New program transformations are devised to maximize the reuse of data which are stored in the fast memory units, such as caches, in the memory hierarchy. These new transformations handle difficult program constructs which are commonly found in practice but could not be handled previously. To support the new transformations, an interprocedural analysis is developed to analyze the memory access behavior throughout the program. The proposed program analysis and transformations are implemented within a research compiler. Successful development of the are techniques can lead to a significant improvement of the execution speed of a wide range of applications.