基于历史长度自适应的分支预测方法

Branch Prediction Based on Branch History Length Adaptation

通过研究处理器动态分支预测器中预测效率与分支历史长度的关系,针对程序中各分支指令存在不同最优历史长度的规律,提出一种搜索各分支指令最佳历史长度的分支预测方法.该方法通过实时监测分支指令的预测准确率,在分支预测表硬件资源不变的情况下动态调整预测器的历史长度,以适应程序的动态运行特征.实验结果表明,在相同硬件资源下,文中方法相对于Gshare预测器错误率降低15.8%,相对于Bi-mode预测器预测错误率降低10.3%.

In this paper, we study the relationship between predict accuracy and branch history length in dynamic branch predictor, and found that different branch instruction will achieve optimal prediction accuracy with differ-ent history length. Thus, we propose a branch prediction strategy, which adapts the branch history length for each branch instruction to achieve high prediction accuracy in total. This approach monitors branch misprediction rate, and dynamically changes the branch history length when the misprediction rate higher than a threshold, this ap-proach can avoid some critical branch alias during the program’s execution, leading the predictor to achieve the optimal prediction accuracy. The experimental results show that, in comparison Gshare, this method achieved 15.8% reduction in misprediction rate, while in comparison with Bi-mode, 10.3% reduction in mispredition rate was achieved.