Advanced Techniques for Improving Indirect Branch Prediction Accuracy

Deep pipelines and fast clock rates are necessitating the development of high accuracy branch predictors. From microarchitectural viewpoint, in the last decade the importance of indirect branch prediction increased even though, in the computing programs the indirect jumps remain less frequent than the more predictable conditional branches. One reason refers to predicative execution that implies decreasing of conditional branches number. The dimension took by the desktop, visual or object-oriented applications development (C++, Java – characterized by a large amount of indirect calls comparative to procedural programs), represents another reason which illustrates that indirect branch prediction misses start to dominate the overall misprediction cost. Since the maximum prediction accuracy obtained by a feasible PPM predictor and reported in literature is around 90% implies the necessity of implementing new efficient indirect branch prediction schemes. Thus, we developed a hybrid predictor with arity-based selection that improves indirect branch prediction accuracy reaching in average to 93.77%, comparable with a multi-stage cascaded predictor. We also showed that a modified Target Cache structure based on confidence mechanism and indexed with extended global correlation information represents a more simple and feasible solution that could replace the more complex PPM predictor.