高性能代价比的两层关联间接转移预测器设计

随着面向对象语言程序、动态链接库(DLL)等的普遍应用,间接转移指令的使用越来越频繁.两层关联间接转移预测器预测准确度高,但实现硬件代价较高,因此并不实用.文中深入分析了两层关联间接转移预测器中产生误预测的原因,通过改进索引方法、压缩存储等实用方法减小硬件实现代价.实验结果表明,通过这些方法的改进,在133K比特硬件存储代价下,使用一组SPEC CPU2000测试程序进行评估,间接转移误预测率为9.6%,仅比两层关联预测器理想误预测率高2.3%,而4路组相联BTB预测器的误预测率为31%.

Design of Cost-Effective 2-Level Correlation Indirect-Branch Predictor

Indirect branches occur frequently in Object-Oriented Language(OOL),and in Dynamically-Linked Libraries(DLL),two programming environments rapidly increasing in popularity.2-level correlation indirect-branch predictor tracks branch behavior well but costs unaccepted hardware budget.Through analyzing principle of mis-prediction in 2-level correlation indirect-branch predictor,the authors propose several cost-effective methods like improving index function and storage compressing to reduce hardware cost of 2-level correlation indirect-branch predictor.With 133K-bit hardware storage budget,for a set of benchmarks comes from SPEC CPU2000,miss-prediction rate drops from 31% on 4-way set-associative BTB to 9.6% on proposed predictor,which is 2.3% larger than the optimal mis-prediction rate.