A Penalty-Sensitive Branch Predictor

Branch predictor design is typically focused only on minimizing the misprediction rate (MR), while ignores misprediction penalty.Because the misprediction penalty varies widely from branch to branch, performance might get improved by using a predictor that makes a greater effort to predict high-penalty branches, at the expense of the other, even if the total number of mispredictions doesn't change. A penalty-sensitive predictor was developed based on this idea. It includes a penalty predictor to predict whether a branch is high or low penalty. Then, a twoclass TAGE predictor is developed to favor high-penalty branches at the expense of low-penalty branches. Experiment shows although the overall performance improvement is limited, the penalty-sensitive mechanism successfully decreases the MR of the highpenalty branches while increasing the MR of the lowpenalty branches by a small amount.