Efficient and Accurate Value Prediction Using Dynamic Classification

this paper, we propose combining three prediction mechanisms into a hybrid predictor. Each predictor has different hardware cost and prediction capability. The choice of a predictor for each instruction is guided by a dynamic classification mechanism. This mechanism partitions the instruction stream to maximize the prediction rate for each predictor. It also utilizes the predictors more efficiently by allocating an entry for each static instruction in at most one predictor. We achieve prediction rates of 72% and 43% with two possible realistic classification mechanisms for the SPECint95 benchmark set