Improving Emciency of Dynamic Speculation via Data Address Prediction Using Instruction Reissue Mechanism

摘要

In this paper, we evaluate resolving ambiguous memory aliasing speculatively. Utilizing data address prediction, an ambiguous memory dependence is speculatively resolved and a load instruction is speculatively executed. In order to realize the data speculation mechanism, a mechanism which recovers processor state when a misprediction occurs is indispensable. We introduce an instruction reissue mechanism. Since instructions independent of misspeculated instructions are not squashed, the effect of data dependence speculation is enhanced. From the experimental evaluation, we have found that the instruction reissue mechanism improves the efficiency of the dynamic data dependence speculation and instruction level parallelism is increased by up to 26.5/100.