摘要
在编译优化中,多面体模型可以对计算密集型程序中的耗时较多的循环代码进行并行性和数据局部性优化。但是,多面体建模过程中存在诸多限制,程序中只有少量代码可以被识别进而转换为多面体表示进行优化。基于LLVM编译框架提出一种分析方法,对多面体建模中的非规则因素进行了规范化处理,对非仿射因素提出一种定值扩展方法,消除了多面体建模的部分限制。实验表明,该方法在SPEC 2006测试集中静态控制块的识别数目增加了44%,同时提升了多面体优化效果。
In compilation optimization,the polyhedral model can optimize the parallelism and data locality of the time-consuming loop code in computation-intensive programs.However,there are many limitations in the process of polyhedron modeling,and only a few codes in the program can be identified and transformed into polyhedron representation for optimization.This paper proposes an analysis method based on the LLVM compilation framework,which normalizes the irregular factors in polyhedron modeling,and proposes a fixed value expansion method for non-affine factors,which eliminates some limitations of polyhedron modeling.The experiments show that the method increases the recognition number of static control parts in the SPEC 2006 test set by 44%,and improves the polyhedron optimization effect.
作者
夏文博
胡伟方
郭浩然
Xia Wenbo;Hu Weifang;Guo Haoran(School of Information Engineering,Zhengzhou University,Zhengzhou 450000,Henan,China;Beijing Space Information Relay Transmission Technology Research Center,Beijing 100094,China)
出处
《计算机应用与软件》
北大核心
2022年第3期19-24,共6页
Computer Applications and Software
基金
郑州大学2018年科研启动基金项目(38879938)。
