自适应存储相关性预测器

Self-adaptive Memory Dependence Predictor

访存指令的乱序执行会导致存储相关性冲突。存储相关性预测技术能够减少相关性冲突,提升处理器性能。已有学术研究工作普遍存在硬件开销大、实现复杂度高的不足;商业处理器中的存储相关性预测技术虽然实现简单,但又存在不具有自适应性或不利于利用指令并行性等问题。设计了一种简单、高效的存储相关性预测器SMDP,它具有自适应、易实现、充分利用指令并行性等优点。实验表明,SMDP能有效提高处理器性能,在与实际处理器接近的较小指令窗口配置下,与盲预测机制相比,平均性能提高0.7991%,最高可达4.9225%。

The out of order execution of store and load instructions always results in memory dependence hazards.MDP（Memory Dependence Prediction） can reduce these hazards,and improve the processor performance.Most of academic research works are complex and spend high hardware cost.Although MDP is implemented simply in commercial processors,it still has some shortages,such as,no self-adaptive ability or blocking ILP（Instruction Level Parallelism）.This paper proposed a simple and effective memory dependence predictor SMDP,which possesses self-adaptive ability,low hardware cost,and high ILP.The simulation shows that SMDP improves the processor performance effectively,by 0.7991% on average over blind prediction,and by 4.9225% at best.