Non-Viable Path Branch Prediction

For modern superscalar processors which implement deeper and wider pipelines, accurate branch prediction is crucial for feeding sufficient number of correct instructions into the superscalar’s highly-parallelized execution core. In this paper, we show that what the branch predictor is learning has significant implications for its ability to make effective use of branch correlation and its ability to exploit longer history lengths. Contrary to the commonly employed approach of training each case using both successes and failures of previous branch behavior, a NonViable Path branch predictor learns those paths that the program will never follow, given prior history; such a predictor is referred to as Non-Viable Path Predictor (NVPP). An NVPP can learn any boolean function, given enough space. In this paper, we present an analysis of the design space and arrive at a practically feasible implementation of the concept which uses simple traditional one or two-bit tables and traditional table indexing mechanisms. Because of its unique properties, NVPP can also exploit very long global histories with only a linear increase in the space requirement and provide rapid training times. Although NVPP exploits large histories, its access time remains relatively constant and it is comparable to small Gshare predictors, permitting a single cycle implementation. The performance of NVPP is compared with most recently published highly accurate predictors including piecewise linear and the O-GEometric History Length predictors. Our results indicate that the NVPP accuracy compares well with these predictors while maintaining the advantages of simplicity and the speed of a traditional style predictor design.