Predicting Instruction Cache Behavior

It has been claimed that the execution time of a program can often be predicted more accurately on an uncached system than on a system with cache memory 5, 20]. Thus, caches are often disabled for critical real-time tasks to ensure the predictability required for scheduling analysis. This work shows that instruction caching can be exploited to gain execution speed without sacriicing predictability. A new method called Static Cache Simulation is introduced which uses control-ow information provided by the back-end of a compiler. This simulator statically predicts the caching behavior of a large portion of the instruction cache references of a program. In addition, a fetch-from-memory bit is added to the instruction encoding which indicates whether an instruction shall be fetched from the instruction cache or from main memory. This bit-encoding approach provides a signiicant speedup in execution time (factor 3-8) over systems with a disabled instruction cache without any sacriice in the predictability of worst-case execution time. Even without bit-encoding, the ability to predict the caching behavior of a large percentage of the instruction references is very useful for providing tight worst-case execution time predictions of large code segments on machines with instruction caches.