Evaluation of Alternative Data Speculation Approaches for Superscalar Processors

Data dependences are one of the main limits to the amount of ILP that current processors can exploit. Data speculation are becoming a promising mechanism to avoid the ordering imposed by data dependences. This work presents two different approaches for improving the performance of current superscalar processors through data speculation. First, the impact of data speculation is studied for an ideal machine with infinite resources. Results show that arithmetic prediction is the best approach for boosting the performance of future processors. Then, load speculation and arithmetic speculation are evaluated for a current superscalar processor with a realistic configuration. Results show that, unlike an ideal machine, load speculation is the best approach to improve the performance of superscalar processors through data speculation. In addition, both mechanisms are combined. The performance gain for this configuration is about 12% in average, and it is up to 45% for some benchmarks.