Design of a High Performance Branch Predictor Based on Global History Considering Hardware Cost

Branch prediction is an indispensable technology in modern processor design and a key method to improve processor performance. When the CPU is executing programs, branch instructions account for 20% to 30% of all instructions, which affect the processor performance. Although dynamic branch predictor based on global history has a simple structure and occupies less memory, the existence of index aliasing problem influences branch prediction accuracy. In this paper, the traditional Gselect branch predictor was improved by the XOR operation between the index generated by Gselect branch predictor and the last 8 bits of instruction's PC address to ease the problem of index aliasing. Finally, the simulation was carried out by using SimpleScalar 3.0 simulator and benchmark programs. The experimental results show that when the length of GHR register is 12, the prediction rate of the newly improved branch predictor called XOR-Gselect is best and 1.93% higher than Gselect's on average.