一种面向向量化的动态指针别名分析框架  

Dynamic Pointer Alias Analysis Framework for Vectorization

作　　者：刘鹏[1,2] 赵荣彩[1,2] 李朋远[1,2] 

机构地区：[1]信息工程大学,郑州450001 [2]数学工程与先进计算国家重点实验室,郑州450001

出　　处：《计算机科学》2015年第3期26-30,共5页Computer Science

基　　金："核高基"国家科技重大专项(2009ZX01036)资助

摘　　要：指针别名分析是数据流分析中的关键性技术,其分析结果是编译优化和程序变换的基础。在向量化方法和动态指针别名分析相关研究的基础上,设计了一种面向向量化的动态指针别名分析框架。该框架通过动态插桩和试运行提取指针别名信息,并反馈到向量化阶段指导向量化代码生成。从提取候选别名分析集、插桩及试运行和反馈优化3个方面对整体框架进行分析和研究。该框架基于Open64实现,并以通用测试集GCC-VECT和典型应用进行了实验评估,结果表明,该框架相比静态指针别名分析具有更精确的别名分析结果,该结果能够有效改进向量化程序的加速比。Pointer alias analysis is a key technology in data-flow analysis,whose results are the basis of compiler optimization and program transformation.Based on the related research of vectorization method and dynamic pointer alias analysis,a dynamic pointer alias analysis framework oriented to vectorization was designed.The dynamic pointer alias information is extracted by dynamic instrument and test run,and the vectorization code generation is guided by the feedback information.The whole framework was studied from three aspects,which are candidate alias analysis set extraction,instrument and test run.The framework was implemented on Open 64 and evaluated in benchmark GCC-VECT and typical applications.The experimental results show that the framework has the more precise alias analysis results compared with static pointer alias analysis,and can significantly improve the speedup of vectorization program.

关 键 词：指针别名分析 向量化 动态分析 依赖分析 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]