Approaches to improve performance for history-based branch predictors

Abstract

This paper investigates the aliasing problems in global-history-based and local-history-based branch predictors and presents two approaches to improve the performance of global-history-based branch predictors. Global-history-based predictors have more critical aliasing problems but show the better performance than local-history-based predictors. Therefore, our approaches mainly focus on alleviating the aliasing problems for global-history-based predictors. The performance of each approach is evaluated and compared by using the Simplescalar simulator with SPEC95CINT benchmark programs. Our experimental results show that the approaches outperform conventional global-history-based branch predictors.