Prediction Outcome History-Based Confidence Estimation for Load Value Prediction

Load instructions occasionally incur very long latencies that can significantly affect system performance. Load value prediction alleviates this problem by allowing the CPU to speculatively continue processing without having to wait for the slow memory access to complete. Current load value predictors can only correctly predict about forty to seventy percent of the fetched load values. To avoid the cycle-penalty for mispredictions in the remaining cases, confidence estimators are employed. They inhibit all predictions that are not likely to be correct. In this paper we present a novel confidence estimator that is based on prediction outcome histories. Profiles are used to identify the high-confidence history patterns. Our confidence estimator is able to trade off coverage for accuracy and vice-versa with great flexibility and reaches an average prediction accuracy over SPECint95 of as high as 99.3%. Cycle-accurate pipeline-level simulations show that a simple last value predictor combined with our confidence estimator outperforms other predictors, sometimes by over 100%. Furthermore, this predictor is one of two predictors that yield a genuine speedup for all eight SPECint95 programs.