A variety of branch predictors was previously proposed, with the analysis of branch behavior. To further improve the branch prediction accuracy, it is necessary to analyze branch behavior more deeply and feed it back to the function of branch predictors. This work, however, is very difficult. Thus, automating this work is helpful to reduce this difficulty. Although there exists an previous work for this automation using genetic algorithm, it was not successful. We attempted to generate a logic circuit of a practical branch predictor with high prediction accuracy by exploiting the existing framework of two-level branch predictors. Our results show that the branch predictor generated by genetic algorithm can achieve 0.5% higher prediction accuracy than the gshare.
