SVP: Safe and Efficient Speculative Execution Mechanism Through Value Prediction.

Kaixuan Wang,Xinyu Qin,Zhuoyuan Yang,Weiliang He,Yifan Liu,Jun Han
DOI: https://doi.org/10.1145/3583781.3590308
2023-01-01
Abstract:Speculative execution attacks such as Spectre and Meltdown exploit the wrong execution patch to leak private data. In current state-of-the-art defense strategies, executions of all memory accesses that use speculatively-loaded addresses are blocked, resulting in high overhead. Our key observation is that these blocked memory accesses can be executed without operand-dependent hardware resource usage through value prediction. Therefore, we propose a novel hardware defense framework, named Speculative Value Prediction (SVP), to safely and efficiently execute the potentially unsafe memory accesses earlier. We build SVP on the cycle-accurate Gem5 simulator and its performance improvement is positively correlated with the coverage of value predictors. Experiments show that when using the value predictor with 30%/60%/100% coverage, SVP outperforms the state-of-the-art defense mechanism STT in the Spectre model by 21.5%/50.3%/107.7% respectively, and in the Futuristic model by 28.7%/55.4%/105.7% respectively.
What problem does this paper attempt to address?