基于线性函数的值预测器设计

在微处理器中，为突破数据流限制以获取更高的指令级并行，指令值预测研究日益得到广泛重视，多种值预测器设计方案被提出。这些预测器可以获得很高的性能，但在性价比优化设计上还有很大的研究空间。本文提出的基于线性函数的值预测器在性能和硬件耗费两方面实现了较好的折衷。SPEC CINT95基准测试程序集模拟结果表明，与复杂的基于stride和2level的混和值预测器相比，基于线性函数的值预测器在性能上仅有很小损失。

Linear Function Based Data Value Predictor Design

In an effort to exploit instruction level parallelism in modern microprocessors, several schemes have been proposed using value prediction to exceed the dataflow limit. Some of these mechanisms achieve very high prediction hit rates, albeit at considerable hardware cost. Others with lower hardware cost but obtain mediocre prediction performance. This paper proposes a new simple value predictor based on linear functions. For the SPEC CINT95 benchmark suite, our scheme has little performance loss compared with that of the best but most complicated hybrid scheme.