顺序程序的并行性识别——一种基于变量活动域的方法

PARALLELISM DETECTION IN SEQUENTIAL PROGRAMS——A METHOD BASED ON THE ACTIVE REGIONSOF VARIABLES

本文提出了一种新的顺序程序并行性的识别方法——基于变量活动域的识别方法,这种方法在程序转换的基础上,系统地削弱了语句之间的数据相关性,提高了程序的并行性;并在变量活动域的基础上将转向控制相关转化为数据相关,最终生成一有条件的数据流图,而不带来任何额外副作用;另外,本文对此识别方法的正确性给出了严格证明。

This paper presents a new method to detect parallelism in sequential programs——a method based on the Active Regions (AR) of variables. With program transformation, this method alleviates the data-dependencies among the statements and enhances parallelism-in sequential programs. With the help of AR, this method transfers the control-dependency into data-dependency and creates a conditional dataflow graph without side-effects. The correctness of this method is proved.