一种混合型值关联间接跳转预测机制

A Hybrid Value Correlation Based Indirect Jump Prediction

准确的间接跳转预测对现代处理器的性能和能耗有效性都具有重要意义.本文提出了一种混合型值关联间接跳转预测机制,通过混合使用多种关联信息以降低间接跳转误预测率.该机制一方面依赖于编译器根据高层次数据流信息识别间接跳转指令所对应的初始关联数据值.另一方面,该机制针对间接跳转预测的不同场景分别设计了两类关联信息:单一数据值和值历史,并实现了一种低开销的硬件结构,该硬件结构在运行时刻根据不同应用场景动态选择最佳关联信息引导间接跳转预测.实验结果表明,相对于传统的BTB预测器和最新的VBBI预测器,本文机制能够有效降低误预测率,提高程序性能并降低系统能耗.

Accurate indirect jump prediction is critical for the performance and energy efficiency of modern high-performance processors.This paper proposes the Hybrid Value Correlation（HVC） based indirect jump prediction,which combines various types of correlated information to reduce indirect jump mispredictions.First of all,our mechanism relies on the compiler to identify the correlated data values based on high-level dataflow information.Second,our mechanism maintains two kinds of correlated information：the single data value and the value history.Our mechanism makes use of a low-cost hardware structure,which dynamically chooses the best correlated information for indirect jump prediction according to different processor states.Experimental results show that HVC prediction can significantly reduce the misprediction rate over the baseline-BTB prediction and the state-of-the-art VBBI prediction.The low misprediction rate of HVC prediction leads to better performance and lower power consumption over previous predictors.