Looking at history to filter allocations in prediction tables

Dependencies between instructions impose an execution order that must be preserved to guarantee the semantic correctness of programs. Recent works propose the use of prediction techniques to speculatively execute dependent operations, showing a significant increment in IPC. We propose a mechanism that reduces the area cost of a typical address predictor: the last-address predictor. Our proposal classifies load instructions at run-time and records the classifications in a table with more entries than the prediction table. Moreover, it uses this information to initialize its confidence information and to filter the allocation of the load instructions in the prediction table. Using direct mapped tables, our proposal captures a similar predictability and increases the accuracy of the typical address predictor and represents around a 40% area-cost saving