Techniques for reducing the misprediction recovery latency in out-of-order processors.

摘要

Most high performance processors speculatively execute instructions following conditional branches. When speculation fails, performance is lost while the processor recovers from the mispredicted instructions. This thesis proposes "Turbo-ROB" and "BranchTap", two techniques that improve performance by speeding up this process. BranchTap is an adaptive speculation control strategy that temporarily throttles speculation when speculating more is likely to deteriorate performance by increasing the recovery latency. Turbo-ROB reduces the amount of work that is necessary to recover from most mispredictions. Our techniques work with existing conventional recovery mechanisms. We demonstrate their utility with just a reorder buffer or with global checkpoints. Turbo-ROB and BranchTap improve average performance of a 512-entry window processor with a reorder buffer by 11.8% and 7.9% respectively. Combined, the two techniques improve average performance by 14.3%.