CanhottracesimproveValueprediction_final-Abstract：可以提高价值prediction_final-abstract热的痕迹.pdfVIP

Can Hot Traces Improve Value Prediction? Assad Khamaisee(1), Avinoam Kolodny(1) and Avi Mendelson(2) (1) EE department, Technion, (2) Intel Corporation, Israel Abstract This paper focuses on the interrelations between two well-known techniques for processor performance acceleration: trace cache (in particular hot traces) and value prediction. Value prediction was proven to have potential for performance improvement, but it suffers from the need to use relatively large tables and it needs an accurate classification mechanism to indicate which instructions are most likely to be predicted correctly in order to minimize the amount of mispredictions. Trace cache was proposed as a mechanism to improve the effective fetch bandwidth. Recently, several papers proposed to use “hot traces” (traces that have been used repeatedly), as a basis for different trace cache optimizations such as reducing the power consumption of the core and gaining better performance given a limited-size cache. This paper proposes to use hot traces as a filter to the value pr ediction, so that the value prediction will be done only for instructions belonging to hot traces. We will show that the new proposed technique improves the prediction accuracy for most of the applications, and reduces the miss ratio of value prediction. 1. Introduction In order to extract more performance out of single threaded applications, various techniques were proposed, including the use of trace cache and value prediction. Trace cache was introduced as a mechanism for increasing the fetch bandwidth of out-of-order machines [rot96], while Value prediction was proposed for overco