A Branch Prediction Mechanism Using Adaptive Branch History Length

Processor pipelines have been growing deeper and issue widths wider over the years. If this trend continues, the branch misprediction penalty will become very high. Branch misprediction is the single most significant performance limiter for improving processor performance using deeper pipelining. Therefore, more accurate branch predictor becomes an essential part of modern processors. Several branch predictors combine a part of the branch address with a fixed amount of global branch history to make a prediction. These predictors cannot perform uniformly well across all programs because the best amount of branch history to be used depends on the program and branches in the program. Therefore, predictors that use a fixed history length are unable to perform up to their potential performance. In this paper, we propose a branch prediction mechanism, using variable length history, which predicts using a bank having higher prediction accuracy among predictions from five banks. Bank 0 is a bimodal predictor which is indexed with the 12 least significant bits of the branch address. Banks 1, 2, 3 and 4 are predictors which are indexed with different global history bits and the branch PC. In simulation results, the proposed mechanism outperforms gshare predictors using fixed history length of 12 and 13 , up to 6.34% in prediction accuracy. Furthermore, the proposed mechanism outperforms gshare predictors using best history lengths for benchmarks, up to 2.3% in prediction accuracy.