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.
CITATION STYLE
Sato, T. (1999). Profile-based selection of load value and address predictors. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 1615, pp. 17–28). Springer Verlag. https://doi.org/10.1007/BFb0094908
Mendeley helps you to discover research relevant for your work.