Exploring Dynamic Branch Prediction Methods

How to resolve the control flow breaking caused by the branch instructions is a major issue in modern deep pipeline processor design. Our project is based on the paper of J. Stark et. al. [1], a variable length path branch predictor. It uses the branch path information for prediction, and change the length of the path dynamically based on the profiling of the application. It shows that a “clever” extraction of branch path information is a critical direction for branch prediction. This paper investigates the branch predictors which explore different information from either control or data flow, including history prediction outcome, branch path and data value. We classified the predictors into three categories: <1>Adding more information to increase the accuracy of branch predictor; <2> Changing the amount of path and history information that will be used for branch prediction; <3> Using different weight for different path and history prediction outcome. By analyzing some typical predictors of each category, the paper attempts to find the most important factors to improve the performance of dynamic branch predictors.