Automatic tuning of GCC optimization passes for ARM platform

Abstract

Modern compilers can work on many platforms and implement a lot of optimizations, which are not always tuned well for every target platform. In the paper we present a tool for automatic tuning of compilation parameters that was developed in ISP RAS. The tool uses genetic algorithm to evolve the best compiler optimization parameters. It supports cross-compilation, parallel build and execution on remote targets, and support for two-tier compilation mode for profile-guided optimizations. Also it provides the tools for automated analysis of the tuning results to help identify the most important compilation parameters.We used our tool to tune GCC parameters for ARM platform, and demonstrated how it can be used to improve performance of several popular applications. Using the tool’s automated analysis of the tuning results, we found several deficiencies in GCC compiler optimizations. The affected optimizations include if-conversion (conversion of conditional branches into predicated form), prefetching (adding of data prefetching instructions in loops), and autovectorization. We identified the reasons for suboptimal code generation, and improved affected optimizations accordingly. Also, we describe a few other improvements that we made to GCC based on the results of manual assembly code analysis. Described improvements to GCC yield significant speedup for chosen test applications on ARM platform.

Keywords

• генетические алгоритмы
• оптимизация программ
• векторизация
• предварительная загрузка данных
• архитектура arm