基于同时多线程的TBHBP分支预测器研究

针对传统处理器分支预测器存在分支预测信息混乱、分支指令别名冲突和容量冲突率高的缺点,提出基于同时多线程处理器的分支预测器TBHBP。该分支预测器采取线程历史信息与基于地址索引的局部历史信息相结合的综合历史信息作为模式匹配表PHT的索引,并采取线程独立拥有线程历史寄存器和分支历史寄存器的方式,通过新增分支结果输出表来提高指令的分支预测执行速度。研究结果表明,TBHBP分支预测器有效解决了分支信息过时、分支指令别名和容量冲突的问题。与Gshare分支预测器相比,其指令吞吐率提升了12.5%,分支误预测率和误预测路径取指率分别下降了0.5%和2.1%。

Branch Predictor with TBHBP Based on Simultaneous Multithreaded Processors

Aiming at the shortages of high conflict of alias and capacity as well as prediction information disorder exisled in the branch instruction, a new branch predictor TI3HI3P which is based on simultaneous multi-threading processor was proposed. In this branch predictor, the comprehensive historical information of the historical thread information and an index based on address information of the local history are taken as the pattern-matching table index of PHT.Though the way of history register is owned by separate thread and branch history register, the instruction execution speed of branch prediction is improved by adding the branch result output table. The researches results show that branch predictor of I}BHBP can effectively improve the problems of out-date branch information, branch instruction aliases and capacity conflicts. Comparing with the Gshare, instruction throughput rises up to 12.5% , and branch prediclion error rate and error forecast path rate can separately drop down by 0. 5% and 2.1%.