Hybrid Indirect Branch Predictors
As discussed in the previous section, predictors with short path lengths adapt more quickly when the program goes through a phase change because it doesn’t take much time for a short history to fill up. Longer path length predictors are capable of detecting longer-term correlations but take longer to adapt and suffer more from table size limitations because a larger pattern set is mapped to the same number of targets. In this section we combine basic predictors into a hybrid predictor in order to obtain the advantages of both.
Unable to display preview. Download preview PDF.