Branch prediction techniques for data-dependent branches

Branch prediction is an essential performance feature as it avoids executing instructions on wrong-path. There are some branches difficult to predict with traditional dynamicprediction techniques (HTP). In particular, in this thesis, data-dependent branches have been considered and some techniques that are software/compiler assisted have been presented. Mainly the purpose of this internship is to improve the mispredictionrate of a SLAM vision algorithm through a novel predictor that will benefit from compiler assistance. The report is organized in four chapters. The first chapter concerns branch prediction in general and some techniques of predictors software and compiler assisted. The second chapter focus on the implementation of a new branch predictor, named Register Parsing Branch Predictor (RPBP) and its operating mechanism has been reported. The third chapter focus on the performance results in which the different changes have been evaluated. The final chapter contains remarks and considerations about furtherwork.