摘要
底层虚拟机(LLVM)是一个广泛使用的编译框架,其中间表示(IR)中包含有丰富的程序分析信息,众多以LLVM为平台的相关工作均以IR为基础开展。数据依赖关系在错误检测、定位及程序调试等领域有着重要应用,基于IR的数据依赖关系计算多采用串行迭代方式,但在应对较大规模IR文件时可扩展性不够理想。对此进行了数据依赖关系计算中指令读写的可并行性挖掘,结合图形处理器并行计算优势,提出一种基于LLVM IR的数据依赖关系并行计算方法DRPC。以IR为输入,采用CPU-GPU双端协同方式实现程序数据依赖关系的高效计算。实验结果表明,针对基准程序集SPEC,DRPC分别在直接及传递数据依赖关系计算上最高获得了3.48x和4.91x的加速比。
The low level virtual machine(LLVM)is a widely used compiler framework.Its intermediate representation(IR)contains abundant program analysis information,and many related work based on program analysis takes a basis of IR.Data dependence plays an important role in fault detection,fault location,and program debugging.IR-based data-dependence computing often adopts serial iterative approaches,such approaches usually suffer scalability issues when dealing with larger-scale IR contents.To address above issues,this paper carried out the parallelism mining of instructions reading and writing in data dependence calculation.Combined with the parallel computing advantages of graphics processing unit,it proposed a parallel computing method DRPC based on LLVM IR.This method accepted IR as inputs,and achieved efficient program data dependence computation through a CPU-GPU cooperation way.The experimental results show that DRPC’s highest speedup is 3.48 x and 4.9 x respectively in the direct and transitive data dependence computation with benchmark SPEC.
作者
朱燕
衷璐洁
Zhu Yan;Zhong Lujie(College of Information Engineering,Capital Normal University,Beijing 100048,China)
出处
《计算机应用研究》
CSCD
北大核心
2020年第2期437-442,共6页
Application Research of Computers
基金
国家自然科学基金资助项目(61872253,61402303).
关键词
底层虚拟机
中间表示
数据依赖
图形处理器
可扩展性
LLVM
intermediate representation
data dependence
graphics processing unit
scalability
