Load Value Prediction Using Prediction Outcome Histories

Due to their occasional very long latency, load instructions are among the slowest instructions of current highperformance microprocessors. Unfortunately, their long latency also delays the execution of all the dependent instructions, which 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 memory access to complete. Today’s load value predictors can only correctly predict about 40 to 70 percent of the load instructions. Confidence estimators are employed to estimate how likely a prediction is to be correct and to keep the predictor from making a (probably incorrect) prediction if the confidence is below a preset threshold. Despite its simplicity, the adaptive prediction outcome history-based confidence estimator we present in this paper outperforms other proposed mechanisms and reaches average prediction accuracies over SPECint95 in excess of 99%, even with small predictor sizes. A detailed pipeline-level simulation shows that a load value predictor equipped with our confidence estimator not only outperforms other predictors by more than 65% when a re-fetch misprediction recovery policy is used, but is also the only predictor that yields a genuine speedup for all eight SPECint95 programs.