A systematic branch strategy and its evaluation

Control dependency introduced by branch (especially conditional branch) instructions, which is resolved only at run-time, can have a severe performance impact on pipelined machines. As programs usually contain a high percentage of branches, a variety of strategies are used to minimize overhead due to branches. However, each strategy has its own advantages and drawbacks. In this paper, a new systematic and integrated strategy that can achieve substantially higher performance than any other schemes reported in the literature is proposed. This strategy uses different mechanisms to deal with different branches based on different dynamic behavior. This is the first published paper that combines four methods to deal with branches. When given the same hardware resources, the proposed strategy can achieve a higher branch performance than the existing best method. For the same given branch prediction performance, it only requires 14% of the hardware when compared with a purely branch target buffer (BTB) approach