Exploring last n value prediction

Most load value predictors retain a large number of previously loaded values for making future predictions. In this paper we evaluate the trade-off between tall and slim versus short and wide predictors of the same total size, i.e., between retaining a few values for a large number of load instructions and many values for a proportionately, smaller number of loads. But results show, for example, that even modest predictors holding sixteen kilobytes of values benefit from retaining four values per load instruction when running SPECint95. A detailed comparison of eight load value predictors on a cycle-accurate simulator of a superscalar out-of-order microprocessor shows that our implementation of a last four value predictor outperforms other predictors from the literature, often significantly. With 21 kB of state, it yields a harmonic mean speedup of 12.5% with existing re-fetch misprediction recovery hardware and 13.7% with a not yet realized re-execution recovery mechanism.