Delay-Sensitive Branch Predictors for Future Technologies DRAFT

Accurate branch prediction is an essential component of a modern, deeply pipelined microprocessors. Because the branch predictor is on the critical path for fetching instructions, it must deliver a prediction in a single cycle. However, as feature sizes shrink and clock rates increase, access delay will significantly decrease the size and accuracy of branch predictors that can be accessed in a single cycle. Thus, there is a tradeoff between branch prediction accuracy and latency. Deeper pipelines improve overall performance by allowing more aggressive clock rates, but some performance is lost due to increased branch misprediction penalties. Ironically, with shorter clock periods, the branch predictor has less time to make a prediction and might have to be scaled back to make it faster, which decreases accuracy and reduces the advantage of higher clock rates. We propose several methods for breaking the tradeoff between accuracy and latency in branch predictors. Our methods fall into two broad categories: hierarchical predictors using purely hardware implementations, and cooperative predictors that off-load some prediction work to the compiler. We describe hierarchical organizations that extend traditional predictors. We then describe a highly accurate branch predictor based on a neural learning technique. Using a hierarchical organization, this complex multi-cycle predictor can be used as a component of a fast delay-sensitive predictor. We introduce a novel cooperative branch predictor that off-loads most of the prediction work to the compiler with profiling. The compiler communicates profiled information to the microprocessor using extensions to the instruction set. This Boolean formula predictor has a small and fast hardware implementation, and will work in less than one cycle in even the smallest technologies with the most aggressive projected clock rates. Finally, we present another cooperative technique, branch path re-aliasing, that moves complexity off of the critical path for making a prediction and into the compiler; this technique increases accuracy by reducing destructive aliasing during the less critical update stage.