Research On Profile-Guided SIMD Vectorization Identification And Optimization

The rise of multimedia extension in the microprocessor architecture, which is caused by the rapid development of multimedia applications, makes the SIMD extension components integrate into popular processors widely. It forces the auto-vectorization technology develop efficiently. However, the appearance of non-aligned and non-continuous reference problems in programs, which is caused by frequently using of complex structures, is impacting negatively the identification and optimization capability of the traditional compilation technology.Based on the development and research of the auto-vectorization compiler SW-VEC, this thesis focuses on some issues about identification and optimization capability of the profile-guided SIMD vectorization compilation technology. The following are the main contents and contributions of the work:1. The profile-guided compilation technology was applied in the SIMD vectorization identification. Aiming at the problem that traditional auto-vectorization technology can't identify complex structures, a feedback architecture using the profile-guided compilation technology in SIMD vectorization was proposed. It can improve the capability of the traditional compiler's vectorization identification effectively.2. Based on the primary performance analysis technology on the Linux system, the thesis implemented a loop-level performance analysis technology. According to this technology, the compiler can implement selectively hot-spot loop auto-vectorization. It can decrease the cost of profile-guided compilation and improve the efficiency of compiler indirectly.3. The thesis used the profile information so that the continuous analysis problem, alignment analysis problem and dependence analysis problem inside the loop can be implemented. Continuous analysis and optimization algorithm, alignment analysis and optimization algorithm and dependence analysis algorithm were proposed in order to solve non-continuous reference problem, non-aligned reference problem and implicit data dependence problem respectively. All of these algorithms can improve the vectorization capability of compiler.All of the algorithms presented above have been implemented on the SW-VEC system. It ensured the validity of all the algorithms efficiently.