Prediction and speculation techniques in ILP

In this article we review the concepts of branch, value and memory prediction used in conjunction with control and data speculative execution in superscalar processors. Since the amount of available instruction level parallelism within a basic block is relatively small, control speculation techniques increase the number of candidate instructions for execution. Moreover, the integration of value, and memory prediction in superscalar processors introduces a new kind of speculative execution. Data speculation techniques allow the processor to execute instructions beyond the limit of true data RAW dependencies. We identify data speculation as a natural extension of control speculation and capture their similarities and differences. We also raise a new question: is the analytical modelling approach really infeasible?.