A Stochastic Approach to Instruction Cache Optimization

The memory alignment of executable code can have significant yet hard to predict effects on the run-time performance of programs, even in the presence of existing aggressive optimization. We present an investigation of two different approaches to instruction cache optimization—the first is an implementation of a well-known and established compile-time heuristic, and the second is our own stochastic, genetic search algorithm based on active profiling. Both algorithms were tested on 7 real-world programs and speed gains of up to 10% were observed. Our results show that, although more time consuming, our approach yields higher average performance gains than established heuristic solutions. This suggests there is room for improvement, and moreover that stochastic and learning-based optimization approaches like our own may be worth investigating further as a means of better exploiting hardware features.