Research on Compilers that can Adaptively Adopt Optimizations

本文的研究目标是编译过程中的自适应优化。如今，计算平台正在分化。它们包括Pentium和RISC的单处理器计算机、小规模SMP、大规模并行系统和异构网格环境。我们强烈要求我们可以在统一的编程模型中编写程序，并且编译器可以负责获取。在每个平台上的最佳性能。因此，编译器中的优化器对良好的性能负有重要责任。然而，在当今计算机/网络/软件体系结构的复杂性中，需要大量的分析来提高性能。在本研究中，我们研究了（1）一个性能模型，具有准确的性能分析。具体来说，我们已经建立了展开的形状，一个象征性的和定量的性能模型循环展开，在顺序和顺序的处理器。该模型被证明是足够准确地捕捉CPU行为的循环展开，如软件流水的影响和内存延迟隐藏。(2)根据数值数据结构的特点自适应选择算法这种选择适用于安装时、编译时和运行时。我们观察到使用这种自适应算法选择的性能改进。作为优化的两个基本要求，我们可以列出它们不改变程序语义，并且它们真正提高性能。作为一个优化的供应商项目，我们试图保证这两个要求。事实上，我们在这项研究中研究了提高性能的适应性，这可以成为实现这一目标的垫脚石。

This research aims at Adaptive Optimization in compilation. Today, computing platforms are diverging. They include uniprocessor computer of Pentium and RISC' s, small scale SMP, large scale parallel systems, and heterogeneous GRID environment. It is strongly required that we can make programs in uniform programming model, and that compilers can take care of getting. best performance on each platform. As its result, optimizers in a compiler is heavily responsible for good performance. However, in today's complexity of computer/network/software architecture, heavy analysis is necessary to improve performance. In this research, we have studied(1) a performance model that has accuracy in performance analysis. Specifically, we have built unrolling shape, a symbolic and quantitative performance model for loop unrolling, for both in-order and out-of-order processors. This model is proved to be accurate enough to capture CPU behaviors in loop unrolling such as effect of software pipelining and memory latency hiding.(2) adaptive selection of algorithms in accordance with characteristics of numerical data structure This selection is applied in installation time, compile time, and runtime. We observed performance improvement using this adaptive algorithm selection.As two essential requirements to optimizations, we can list that they do not change program semantics and that they truly improve performance. As Optimization Verifying Compiler Project, we are trying to guarantee the two requirements. The fact that we have studied adaptation for improving performance in this research can be a stepping stone to this end.