RESEARCH AND DEVELOPMENT OF A PARALLELIZING COMPILER FOR ARBITRARY GRAIN SIZES AND GENERAL PARARLLEL ARCHITECTURES

在京都大学津田实验室开发了十多年的向量化编译器V-Pascal的基础上，添加了自动并行化功能以及各种功能。(1)强大的并行性检测能力，（2）设施，以找到最佳的粒度，（3）提高数据的局部性，通过数据分区和循环转换，和（3）最小化的机器依赖性，在自动并行化的并行结构抽象仍然有待解决的主要困难。在这个项目中，开发了新的别名分析算法，以便在分析指针引用引起的依赖项时具有更高的精度。在存在指向堆分配的内存区域的指针变量的情况下，内存分析特别困难，但该项目的结果为问题提供了重要线索。这部分回答了上面的问题（1）。关于问题（2），已经考虑了一种特殊的视图机制，使得依赖性分析的结果可以在任意层次级别上显示，即，粗粒度级别或细粒度级别。这使得检查和选择最佳并行化变得更容易。

On top of the vectorizing compiler V-Pascal that had been developed by the Tsuda laboratory of Kyoto University over ten years, automatic parallelizing functions were added, together with various features. (1) Powerful ability to detect parallelism, (2) facility to find optimal grain sizes, (3) enhancing the locality of data references by means of data partitioning and loop transformation, and (3) minimizing machine-dependences in automatic parallelization by abstraction of parallel architectures are the main difficulties that remain to be resolved. In this project, new algorithms were developed for alias analyzes to have more precision in the analysis of dependeces due to pointer references. Alias analyzes in the presence of pointer variables that point to memory area that has been assigned by heap are especially difficult, but the result of this project gives an important clue to the problem. This partially answers the question (1) above. Concerning the question (2), a special view mechanism has been considered so that the results of dependence analyzes can be displayd at an arbitrary hierarchical level, namely, either coarse-grain level or fine-grain level. This makes it easier to check on, and choose, optimal parallelization.