Return Value Prediction meets Information Theory
Jeremy Singer and Gavin Brown
Abstract
Accurate return value prediction is a key tool for enabling effective speculative method-level parallelism, which will be a standard feature in the next generation of chip-multiprocessor architectures. In this paper we give some information theoretic measures that indicate intrinsic predictability of method return values. This is in stark contrast to the current ad-hoc heuristic measures imposed by specific prediction techniques. Our hope is that the application of information theoretic principles to the field of return value prediction should result in new kinds of predictors, and better deployment of existing prediction techniques. The two main contributions of this work are: (i) to show that there is some correlation between information theoretic measures and return value predictor performance; (ii) to highlight some major issues that need to be resolved before information theory can be adopted practically by the return value prediction community.