Target encoding for efficient indirect jump prediction

Abstract

Accurate indirect jump prediction is critical for some applications. Proposed methods are not efficient in terms of chip area. Our proposal evaluates a mechanism called target encoding that provides a better ratio between prediction accuracy and the amount of bits devoted to the predictor. The idea is to encode full target addresses into shorter target identifiers, so that more entries can be stored with a fixed memory budget, and longer branch histories can be used to increase prediction accuracy. With a fixed area budget, the increase in accuracy for the proposed scheme ranges from 10% to up to 90%. On the other hand, the new scheme provides the same accuracy while reducing predictor size by between 35% and 70%.