Profile-based selection of load value and address predictors

Abstract

In this paper, we investigate hybrid predictors for data speculation. In order to increase opportunities for data speculation as well as improve prediction accuracy, we propose to combine a load address predictor with a load value predictor. For each instruction, by choosing the more accurate predictor, we improve prediction accuracy. We investigate two types of hybrid predictors. One has an adaptive mechanism for choosing the more accurate one dynamically, and the other decides the selection statically using execution profiles. The latter one has the benefit that the hardware cost of the selecting mechanism is removed. We have evaluated the predictors using a cycle-by-cycle simulator and found that contribution of the static hybrid predictor to processor performance is comparable to that of the dynamic one.