A new method to evaluate instruction cache performance was designed and implemented. The cache performance of programs for various cache configurations can be obtained without recompiling the analyzed program. No special operating system support or dedicated registers are required. The new method outperforms conventional trace-driven cache simulation by almost an order of a magnitude without any loss of accuracy of the measurements. By making extensive use of static cache simulation and reducing code instrumentation to simple frequency counting in many places, this method reduces the execution overhead of analyzed programs to a factor of 2 on average. In addition, different cache sizes and resulting hit ratios have little influence on the overhead. Therefore, one can conclude that instruction cache analysis via static cache simulation is a general method to quickly obtain accurate measurements outperforming any other published methods.
-20pt