Using Dataflow Based Context for Accurate Value Prediction

We explore the reasons behind the rather low prediction accuracy of existing data value predictors. Our studies show that contexts formed only from the outcomes of the last several instances of a static instruction do not always encapsulate all of the information required for correct prediction. Complex interactions between data flow and control flow change the context in ways that result in predictability loss for a significant number of dynamic instructions. For improving the prediction accuracy, we propose the concept of using contexts derived from the predictable portions of the data flow graph. That is, the predictability of hardto -predict instructions can be improved by taking advantage of the predictability of the easy-to-predict instructions that precede it in the data flow graph. We propose and investigate a run-time scheme for producing such an improved context from the predicted values of previous instructions. We also propose a novel predictor called dynamic dataflow-inherited speculative context (DDISC) based predictor for specifically predicting hard-to-predict instructions. Simulation results verify that the use of dataflow-based contexts yields significant improvements in prediction accuracies, ranging from 35% to 99%. This translates to an overall prediction accuracy of 68% to 99.9%