ABSTRACT
Speculative execution has proven to be enormously valuable for increasing execution-time performance in recent and current processors. The use of speculative execution provides a powerful latency-hiding mechanism for those microarchitectural operations that would otherwise cause unacceptable stalls within the processor, such as waiting for conditional branches to resolve or for memory reads to be fulfilled from the memory hierarchy. Further, in order to extract ever larger amounts of instruction level parallelism from existing programs (generally quite sequential in nature) over many basic blocks, much more speculative execution is usually required. However, the complexity of implementing speculative execution is substantial and has been a limiting factor in its evolution to more aggressive forms beyond control-flow speculation.
