Analysing value substitution and confidence estimation for value prediction

Abstract

Value Prediction is one of the newest techniques used to break down ILP limits. Despite being under continuous study during the last few years, a few aspects related to this emerging technique remain unanalysed in depth. Exhaustively investigated in the context of control speculation, confidence estimation has usually played a secondary role on value prediction and speculation. Closely linked to confidence estimation, value substitution also represents a relegated subject of research. This paper is focussed on analysing, in an isolated way, the respective impact on predictor performance of both confidence estimation and value substitution mechanisms. By using detailed pipeline-level simulations, we prove that improvements in these mechanisms are as important as reducing the predictor aliasing or even improving the prediction model.