Adaptive Software Return Value Prediction

Return value prediction (RVP) is a technique for guessing the return value from a function before it actually completes, enabling a number of program optimizations and analyses. However, despite the apparent usefulness, RVP and value prediction in general have seen limited uptake in practice. Hardware proposals have been successful in terms of speed and prediction accuracy, but the cost of dedicated circuitry is high, the available memory for prediction is low, and the flexibility is negligible. Software solutions are inherently much more flexible, but can only achieve high accuracies in exchange for reduced speed and increased memory consumption. In this work we first express many different existing prediction strategies in a unification framework, using it as the basis for a software implementation. We then explore an adaptive software RVP design that relies on simple object-orientation in a hybrid predictor. It allocates predictors on a per-callsite basis instead of globally, and frees the resources associated with unused hybrid sub-predictors after an initial warmup period. We find that these techniques dramatically improve speed and reduce memory consumption while maintaining high prediction accuracy.