Limited dual path execution

This work presents a hybrid branch predictor scheme that uses a limited form of dual path execution along with dynamic branch prediction to improve execution times. The ability to execute down both paths of a conditional branch enables the branch penalty to be minimized; however, relying exclusively on dual path execution is infeasible due because instruction fetch rates far exceed the capability of the pipeline to retire a single branch before others must be processed. By using confidence information, available in the dynamic branch prediction state tables, a limited form of dual path execution becomes feasible. This reduces the burden on the branch predictor by allowing predictions of low confidence to be avoided. In this study we present a new approach to gather branch prediction confidence with little or no overhead, and use this confidence mechanism to determine whether dual path execution or branch predic-tion should be used. Comparing this hybrid predictor model to the dynamic branch predictor shows a dra-matic decrease in misprediction rate, which translates to an reduction in runtime of over 20%. These results imply that dual path execution, which often is thought to be an excessively resource consuming method, may be a worthy approach if restricted with an appropriate predicting set. 1