PMPM: Prediction by Combining Multiple Partial Matches

The prediction by partial matching (PPM) algorithm has been well known for its high prediction accuracy. Recent proposals of PPM-like predictors confirm its effectiveness on branch prediction. In this paper, we introduce a new branch prediction algorithm, named Prediction by combining Multiple Partial Matches (PMPM). The PMPM algorithm selectively combines multiple matches instead of using the longest match as in PPM. We analyze the PPM and PMPM algorithms and show why PMPM is capable of making more accurate predictions than PPM. Based on PMPM, we propose both an idealistic predictor to push the limit of branch prediction accuracy and a realistic predictor for practical implementation. The simulation results show that the proposed PMPM predictors achieve higher prediction accuracy than the existing PPM-like branch predictors such as the TAGE predictor. In addition, we model the effect of ahead pipelining on our implementation and the results show that the accuracy loss is relatively small.