A case study: optimizing GCC on ARM for performance of libevas rasterization library

This paper reports on the work for optimizing GCC on ARM to improve performance of libevas rasterization library. We used manual profiling and analysis as well as ACOVEA [3] compiler options tuning tool to identify weak places and tune GCC optimization parameters. We identified a number of deficiencies in GCC optimizations with libevas on ARM, including GCSE, register allocation, autovectorization and loop prefetching, and proposed solutions to them, that altogether brought 15.78% average performance increase, and with up to 119% increase on certain tests, as measured with customized expedite benchmark. These results show that tuning existing GCC optimizations for specific platform and application may provide significant performance boost, comparable to that of developing a new compiler optimization.