Correlation and Aliasing in Dynamic Branch Predictors: Full Simulation Results

Previous branch prediction studies have relied primarily upon the SPECint89 and SPECint92 benchmarks for evaluation. Most of these benchrnarb exercise a very small amount of code. As a con-sequence, the resources required by these schemes for accurate predictions of larger programs have not been clear Moreove~ many of these studies have simulated a very limited number of con-figurations. Here we report on simulations of a variety of branch prediction schemes using a set of relatively large benchmark pro-grams that we believe to be more representative of likely system workloads. We have examined the sensitivity of these prediction schemes tovariation in workload, in resources, andin design and configuration. We show that for predictors with small available resources, aliasing between distinct branches can have the domi-nant injluence on prediction accuracy. For global history based schemes, such as GAs andgshare, aliasing in the predictor table can eliminate any advantage gained through inter branch correla-tion. For self-history based prediction scheme, such as PAs, it is aliasing inthebuffer recording branch history, rather than thepre-dictor table, that poses problems. Past studies have sometimes confused these effects and allocated resources incorrectly.