Reducing Dynamic Branch Predictor Lookups by Dynamically Collecting Non-branch Instructions

Abstract

Deeply pipelined architecture uses the dynamic branch predictor to enhance the performance. The dynamic branch predictor of the traditional scheme is exercised every cycle and then causes many power consumptions. For this reason, we intended to eliminate unnecessary lookups for dynamic branch predictor. In this paper, we proposed a scheme to dynamically collecting non-branch instructions between adjacent branch instructions during execution stage. We save the power by using the number we collected to filter the unnecessary lookups. Simulation results show that the dynamic branch predictor power is reduced by 63.79% and 76.58% in average for SPECint2000 and SPECfp2000 respectively with negligible performance loss.