Branch-Target Correlation to Resolve Sub-Trace Aliasing in Dead Block Predictor

The traditional dead-block predictors have the problem of sub-trace aliasing, leading to unbounded wrong predictions. We believe that by including the information of execution footprints after the last touch, we can better deal with the sub-trace aliasing problem. This paper presents a new method, namely branch-target correlated dead-block predictor (BT-DBP), which adds the correlation of branch targets after the last touch to a memory address as a suffix to a DBP signature to unravel the sub-trace aliasing in a global prediction table. The results show that BT-DBP can reduce the misses incurred in using a global prediction table by 51% on average without scarifying the correct prediction coverage.