Abstract: Traditional vectorization method and superword level parallelism both rely on data dependence analysis to exploit the parallelism of the programs. However, the up-to-date compiler has less ability to analyze such statements as non-structured control flow statements, which the data dependence analysis cannot deal with. This paper presents a vectorization method of export branch statements for SIMD extension, which can effectively vectorize the export branch statements within a vector length. Test results show that this method can both fully exploit the parallelism of the data flow and ensure the semantic correctness of control flow.

Key words: SIMD extension, vectorization, export branch, vector factor, non-structured

摘要： 传统的向量化方法和超字并行方法依靠数据依赖关系分析确定程序中的并行性，而依赖关系分析无法处理非结构化控制流语句，现有的编译器对该类语句的向量化能力有限。为此，给出一种面向SIMD扩展体系结构的出口分支语句向量化方法，该方法针对一个向量因子内的出口分支语句，能够有效地进行自动向量化处理。测试结果表明，该方法既充分发掘了程序数据流中的并行性，又保证了控制流语义的正确性。

关键词: SIMD扩展, 向量化, 出口分支, 向量因子, 非结构化

CLC Number: 

• TP311