A Feasibility Study on Hyperblock-based Aggressive Speculative Execution Model

Speculation execution model which executes sequential programs in parallel through speculation is an effective technique for making better use of growing on-chip resources and exploiting more instruction-level parallelism of applications. However, accompanied high communication overheads and roll-back penalties can not be neglected. This paper focuses on analyzing the feasibility of aggressive speculation execution model and finding an appropriate degree of "aggressiveness" under hyperblock-based execution model. We analyze the characteristic of control dependences and data dependences between adjacent hyperblocks, and propose a quantitative analysis method to detect data dependences on hyperblock-based execution model, and then evaluate the feasibility of aggressive speculative execution model on 8 applications from SPEC2K. Our experiments show most applications can get high prediction accuracy on control-flow from hyperblock-based prediction mechanisms, especially SPECFP. Furthermore, we analyze factors which impact expected prediction depth and find depth depends more on application than predictors.