Focalising dynamic value prediction to CPU's context

Abstract

Value prediction (VP) is a relatively new technique that increases performance by eliminating true data dependency constraints. VP architectures allow data dependent instructions to issue and execute speculatively using the predicted value. This technique is built on the concept of value locality, which describes the likelihood of a previously seen value recurring within a storage location. The authors extend dynamic VP by introducing the concept of register-centric prediction instead of instruction-centric prediction. The value localities obtained on some registers of MIPS architecture were quite remarkable leading to the conclusion that VP might be successfully applied, at least on these favourable registers. The idea of attaching a value predictor for the processor's favourable registers is original and might involve new architectural techniques for improving performance and reducing the hardware cost of speculative micro-architectures. The register VP technique consists in predicting the registers' next values based on the previously seen values. It executes the subsequent data dependent instructions using the predicted values. The speculative execution will be validated when the correct values are known. If the value was correctly predicted the critical path is reduced, otherwise the instructions executed with wrong entries must be executed again. The authors examine different favourable register selections and different basic value predictors to capture certain type of value predictabilities from the SPEC benchmarks (1995 and 2000) to obtain higher prediction accuracies. Their results show that there is a time correlation between the names of the destination registers and the values stored in these registers. The simulations show that the hybrid predictor optimally exploits this correlation with an average prediction accuracy of 85.44%, which is quite remarkable (on some benchmarks the values are over 96%). Considering an eight-issue out-of-order superscalar processor it is shown that register-centric VP produces average speedups of 17.30% for the SPECint95 benchmarks, and of 13.58% for the SPECint2000 benchmarks.