Combining Branch Predictors：（结合分支预测）.pdfVIP

Combining Branch Predictors Scott McFarling June 1993 d i g i t a l Western Research Laboratory 250 University Avenue Palo Alto, California 94301 USA Abstract One of the key factors determining computer performance is the degree to which the implementation can take advantage of instruction-level paral- lelism. Perhaps the most critical limit to this parallelism is the presence of conditional branches that determine which instructions need to be executed next. To increase parallelism, several authors have suggested ways of predicting the direction of conditional branches with hardware that uses the history of previous branches. The different proposed predictors take advan- tage of different observed patterns in branch behavior. This paper presents a method of combining the advantages of these different types of predictors. The new method uses a history mechanism to keep track of which predictor is most accurate for each branch so that the most accurate predictor can be used. In addition, this paper describes a method of increasing the usefulness of branch history by hashing it together with the branch address. Together, these new techniques are shown to outperform previously known approaches both in terms of maximum prediction accuracy and the prediction accuracy for a given predictor size. Specifically, prediction accuracy reaches 98.1% correct versus 97.1% correct for the most accurate previously known ap- proach. Also, this new approach is typically at least a factor of two smaller than other schemes for a given prediction accuracy. Finally, this new ap- proach allows predictors with a single level of history array access to outper- form schemes with multiple level