Trace Substitution

Trace caches deliver a high number of instructions per cycle to wide-issue superscalar processors. To overcome complex control flow, multiple branch predictors have to predict up to 3 conditional branches per cycle. These multiple branch predictors sometimes predict completely wrong paths of execution, degrading the average fetch bandwidth. This paper shows that such mispredictions can be detected by monitoring trace cache misses. Based on this observation, a new technique called trace substitution is introduced. On a trace cache miss, trace substitution overrides the predicted trace with a cached trace. If the substitution is correct, the fetch bandwidth increases. We show that trace substitution consistently improves the fetch bandwidth with 0.2 instructions per access. For inaccurate predictors, trace substitution can increase the fetch bandwidth with up to 2 instructions per access.