Evaluation and Analysis of Effects of Auto-vectorization in Typical Compilers

SIMD(Single-Instruction-Multiple-Data) architecture plays an important role in the architecture of modern processors.Also,it is supported in multiple types of homemade high performance general purpose processors.For exploiting the data-parallel performance potentials of the short vector processing ability presented by the SIMD architecture,the auto-vectorization of compilers is one of the main means for improving the performance of applications.Evaluating the effects of auto-vectorization in typical compilers using mature commercial general purpose processor with SIMD support,is beneficial both to processor architecture design and to compiler analysis and design.We evaluated the effects of auto-vectorization in the typical compilers(including Intel compiler,PGI compiler and GCC compiler) with the stan-dard benchmarks SPECCPU2006 and SPECOMPM2001.Then taking the open source product level GCC compiler as the target,we thoroughly evaluated the effects of auto-vectori-zation in GCC with hand-coded program segments(mainly multiple types of loops),and we analyzed the ability and limitations of the current implementations of the auto-vectori-zation in GCC.Our work provides a valuable contribution for the research and development of auto-vectorization in compilers.