Exploiting Computational Locality in Global Value Histories.

Abstract

Value prediction is a speculative technique to break true data dependencies by predicting uncomputed values based on history. Previous research focused on exploiting two types of value locality (computation-based and context-based) in the local value history, which is the value sequence produced by the same instruction that is being predicted. Besides local value history, value locality also exists in global value history, which is the value sequence produced by all dynamic instructions according to their execution order. In this thesis, a new type of value locality, computational locality in global value history is studied. A prediction scheme, called gDiff, is designed to exploit one special and most common case of this computational model, the stride-based computation, in global value history. Experiments show that there exists very strong stride type of locality in global value sequences and ideally the gDiff predictor can achieve 73% prediction accuracy for all value producing instructions without any hybrid scheme, much higher than local stride and local context prediction schemes. However, the ability to realistically exploit locality in global value history is greatly challenged by the value delay issue, i.e., the correlated value may not be available when the prediction is being made. The value delay issue is studied in an out-of-order (OOO) execution pipeline model and the gDiff predictor is improved by maintaining an order in the value queue and utilizing local stride predictions when global values are unavailable to avoid the value delay problem. This improved predictor, called hgDiff, demonstrates 88% accuracy and 69% prediction coverage on average, outperforming a local stride predictor by 2% higher accuracy and 13% higher coverage.