Improving branch prediction by dynamic dataflow-based identification of correlated branches from a large global history

Deep pipelines and fast clock rates are necessitating the development of high accuracy, multi-stage branch predictors for future processors. Such a predictor uses a collection of predictors, each of which provides its predictions at a different stage of the pipeline front-end. A simple 1-cycle latency line predictor provides predictions in the first stage, followed in a couple of stages later by predictions from a more accurate global predictor. Finally, one or two stages later, a highly accurate corrector predictor selectively corrects the global predictor's prediction. As the corrector predictor has the final say, its accuracy must be very high. The focus of this paper is to propose and evaluate techniques to build high-accuracy corrector predictors.Our techniques rely on using a long global history, and identifying correlated branches in this history by using runtime dataflow information. In particular, we identify for each dynamic branch a set of branches called "affectors", which control the computation that affect that branch's outcome. We propose efficient hardware structures to track dataflow and to identify the affector branches for each dynamic branch; the hardware overhead for identifying affectors for all dynamic branches from a 64 branch global history is only 312 bytes. We then propose two prediction schemes that put to use the affector branch information. Experimental studies show that adding an 8KB corrector predictor (that uses affector information) to a 16KB perceptron predictor (total size 24.2KB) reduces the average misprediction rate for 12 benchmarks from 6.3% to 5.7%, an improvement achieved only by a 64KB perceptron predictor.