TG-share Branch Predictor Designed For Frequent Context Switching

Accurate branch prediction is very important to improve computers'performance. Because of superscalar and superpipelining techniques, which are used in most of the popular computers, a branch misprediction may cause several to dozens of cycles'stall. Thus, how to get the satisfying predictions has become an important field in computer architecture research.SMT (Simultaneous Multi-threading Processor) is a promising way in the future high performance microprocessor development. But there are frequent context switching in SMT processors. Common branch predictors will not work as efficiently as they used to do.Though, to solve the problem, some researchers have given some branch predictors of multi-threaded processors, they do not work because of unpractical implementation. So a new branch predictor is brought forward in the thesis. It is named as TG-share, which means Threaded Global History with Index Sharing. TG-share is devised for a condition of frequent context switching. It has higher performance than common predictors. However, it does not cause much complexity in hardware, because it only uses the thread global branch history which can be implemented easily. Thus, TG-share has overcome the shortcoming that common predictors do not suit for multi-threading environment and some ideal multi-threaded predictors are too complex to be implement.