Records Branch Predictor

Modern processors get great performance improvement by branch prediction, and two levels pattern branch predictor is widely used because of its simple structure and high performance. But this kind of predictor has 3 limitations—Branch Alias, History Noise and Multi-way Access. Alias and noise have been researched by many previous works, but those 2 limitations cannot be solved absolutely and have space to improve more. Meanwhile previous works do not notice multi-way access.This paper summarizes alias and noise, poses multi-way access, pictures miss predictions from multi-way access, and suggests a Stack Branch Structure according to loops and function calls. Based on stack structure, this paper suggests another structure—Records Branch Structure (RBS). RBS is a structure adding to predictor to maintain and choice GHR when predicting. It does not change primary predictor's update and architecture. It can be used by any pattern-based predictor. Predictor with RBS is called Records Branch Predictor. RBS has 3 rules to guarantee predictor accesses the PHT entry, modified by current predicting branch, to predict. Those 3 rules reduce bad effect from noise at last. RBS is useless for alias.This paper implements Gshare, Records-Gshare, 2Bc-gskew and Records-2Bc-gskew on Simplescalar to compare Records Branch Predictor with primary predictor without RBS. The results show that predictors with RBS can get less miss predictions than primary predictor. Besides, longer PC-Records and larger PHT capacity can reduce more miss prediction percentage for primary predictor in general. As an example, Records-Gshare can reduce 7.6% miss prediction for Gshare and Records-2Bc-gskew can reduce 14.5% miss prediction for 2Bc-gskew on average when number of PC-Records entries is 20 and number of PHT entries is 4096.