Skip to Main content Skip to Navigation

De-aliased Hybrid Branch Predictors

André Seznec 1 Pierre Michaud 1
1 CAPS - Compilation, parallel architectures and system
IRISA - Institut de Recherche en Informatique et Systèmes Aléatoires, Inria Rennes – Bretagne Atlantique
Abstract : Fixed-size branch predictors tables suffer from a loss of prediction accuracy due to aliasing or interference. This is particularly true for predictors using a global history vector such as gshare. «De-aliased» global history predictors -the skewed branch predictor, the bimode predictor and the agree predictor- were recently proposed. «De-aliased» predictors consistently achieve the same prediction accuracy level as gshare or gselect using less than half the transistor budget. However different branches do not require the use of the same vector of information to be accurately predicted. Hybrid predictors combining several branch prediction schemes -- may deliver higher branch prediction accuracy than a branch predictor using a single branch prediction scheme. Then «de-aliased» branch are natural candidates as hybrid predictors components. In this paper, we show how cost-effective hybrid branch predictors can be derived from the enhanced skewed branch predictor e-gskew. 2Bc-gskew combines e-gskew and a bimodal branch predictor. It consists in four identical predic tor-table banks, i.e., the three banks from the e-gskew -including a bimodal bank- plus a meta-predictor. 2Bc-gskew-pskew combines a bimodal component, a global history register component and a per-address history component. These hybrid predictors are shown to achieve high prediction accuracy at a low hardware cost.
Document type :
Complete list of metadata
Contributor : Rapport de Recherche Inria <>
Submitted on : Wednesday, May 24, 2006 - 11:43:26 AM
Last modification on : Thursday, January 7, 2021 - 4:36:18 PM
Long-term archiving on: : Sunday, April 4, 2010 - 9:38:44 PM


  • HAL Id : inria-00073060, version 1


André Seznec, Pierre Michaud. De-aliased Hybrid Branch Predictors. [Research Report] RR-3618, INRIA. 1999. ⟨inria-00073060⟩



Record views


Files downloads