Cached Two-Level Adaptive Branch Predictors with Multiple Stages

Abstract

In this paper, we quantify the performance of a novel family of multi-stage Two-Level Adaptive Branch Predictors. In each two- level predictor, the PHT of a conventional Two-level Adaptive Branch Predictor is replaced by a Prediction Cache. Unlike a PHT, a Prediction Cache saves only relevant branch prediction information. Furthermore, predictions are never based on uninitialised entries and interference between branches is eliminated. In the case of a Prediction Cache miss in the first stage, our two-stage predictors use a default two-bit prediction counter stored in a second stage. We demonstrate that a two- stage Cached Predictor is more accurate than a conventional two-level predictor and quantify the crucial contribution made by the second prediction stage in achieving this high accuracy. We then extend our Cached Predictor by adding a third stage and demonstrate that a Three- Stage Cached Predictor further improves the accuracy of cached predictors.