Reducing Branch Misprediction Penalty through Confidence Estimation

Abstract

The goal of this Thesis is reducing the global penalty associated to branch mispredictions, in terms of both performance degradation and energy consumption, through the use of confidence estimation. The reduction of this global penalty has been achieved, firstly, by increasing the accuracy of branch predictors, next, by reducing the time necessary to restore the processor from a mispredicted branch, and finally, by reducing the energy consumption due to the execution of incorrect instructions. All these proposals rely on the use of confidence estimation, a mechanism that assesses the quality of branch predictions by means of estimating the probability of a dynamic branch prediction to be correct or incorrect.