Collaborative Research: Next Generation Compilers for Emerging Multicore Systems

多核作为标准机器设计的出现给软件工业带来了计算历史上最重要的挑战之一。为了利用每一代新机器的额外计算能力，程序必须能够从多核的最重要特征中获利：多个处理器的存在。换句话说，程序必须能够并行执行。此外，为了高效执行，这种并行性必须采用与程序要执行的多核机器的内部组织一致的形式。如果这些程序是手工设计的，那么考虑机器特性的需要将大大增加程序开发的成本。此外，由于较新的多核设计可能包含新颖的体系结构特性，因此将程序从一代移植到下一代的过程也将涉及大量成本。换句话说，如果什么都不做，软件成本的显著增加将是广泛接受多核的必要条件。在这个项目中，我们的目标是开发能够自动生成高效并行程序的技术。为此，我们将扩展Pivot，一个在Texas A&amp；M开发的c++编译器原型，它具有能够自动检测大多数传统c++程序中隐含的并行性并将其映射到广泛的多核系统的技术。也就是说，我们将使用自动并行化技术扩展Pivot。我们将建立在伊利诺伊州和德克萨斯州农工大学为数值计算开发的静态和混合（静态和动态）分析技术的基础上，并将其扩展到处理c++程序中经常使用的不规则数据结构。

The emergence of multicores as the standard machine design has created one of the most important challenges to the software industry in the history of computing. To take advantage of the additional computational power of each new generation of machines, programs must be able to profit from the most important characteristic of multicores: the presence of multiple processors. In other words, programs must be able to execute in parallel. Furthermore, for efficient execution, this parallelism must take a form that is consistent with the internal organization of the multicore machine where the program is to execute. If these programs were to be manually designed, the need to take into account machine characteristics would increase the cost of program development significantly. Also, since newer multicore designs are likely to include novel architectural features, the process of porting programs from one generation to the next will also involve significant costs. In other words, if nothing is done, the significant increases in the cost of software will be necessary for widespread acceptance of multicores. Our objective in this project is to develop techniques for automating the process of generating efficient parallel programs. To this end, we will extend Pivot, a prototype C++ compiler, under development at Texas A&M, with techniques capable of automatically detecting the parallelism implicit in most conventional C++ programs and mapping it onto a wide range of multicore systems. That is, we will extend Pivot with automatic parallelization techniques. We will build on static and hybrid (static and dynamic) analysis techniques developed at Illinois and Texas A&M for numerical computations and extend them to handle the irregular data structures that are often used in C++ programs.