【作者】 钟明； 郭振宇； 汤志忠；

【Author】 Zhong Ming Guo Zhenyu Tang Zhizhong Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China

【机构】 清华大学计算机科学与技术系；

【摘要】 软件流水是一种重要的指令调度技术,通过重叠地执行不同的循环体来提高指令级并行性。模调度是一类重要的软件流水调度算法。保守的相关性分析可能会引入较多的模糊相关,这阻碍了模调度生成高效的调度结果。数据猜测能克服保守的相关性分析带来的调度限制,开发潜在的并行性。本文提出了模调度中的一种数据猜测方法,在开放源代码编译器ORC上实现了该方法,并用SPEC2000基准程序进行了测试,实验结果表明,该方法收到了较好的效果。更多还原

【Abstract】 Software pipelining is an important scheduling technique which exploits instruction level parallelism by overlapping several consecutive iterations. Modulo scheduling is a state-of-art scheme in the literature of software pipelining. Conservative alias analysis may introduce many non-definite dependencies, which prevents modulo scheduling from finding a more efficient scheduling result. Data speculation can be applied to ignore those constraints and exploit potential parallelism. This paper proposes a data speculation approach in modulo scheduling. We implement the approach in ORC and test it on SPEC2000 benchmarks. Experimental results show that the approach boosts performance effectively.更多还原