Generalized Worst Case Estimation of Misprediction Counts for Dynamic Branch Predictors

Abstract

Estimating the number of mispredictions is critically important for estimating the Worst-Case Execution Time for real-time systems. This paper generalizes and improves over previous attempts to provide a safe and tight mispredication count estimate for dynamic branch predictors. The paper gives closed formulas to compute mispredictions in case of simple and nested loops applicable to all variations of two-level adaptive branch predictors in addition to the gshare and gselect predictors. The given formulas are general enough to accommodate predictors with any counter size.