Exploring design space of scalable per-address branch predictors

While a scalable per-address (SPA) branch predictor reduces the cost of implementing per-address two-level branch predictors by about 50% through the exploitation of value locality in the history of branch outcomes, it suffers from internal conflict within the pattern history table (PHT), called pattern aliasing or interference. In this paper, we propose and evaluate alternative designs of SPA predictor to reduce aliasing or interference rate in its PHT as well as to improve its prediction accuracy. Several PHT designs are evaluated against a direct-mapped PHT used in the original SPA branch predictor design. Our experimental results on eight SPEC2000 benchmark programs reveal that a SPA predictor with 4-way set-associative PHT using 7-tag bit yields the highest prediction accuracy.