Compiler-driven value speculation scheduling

Modern microprocessors utilize several techniques for extracting instruction-level parallelism (ILP) to improve the performance. Current techniques employed in the microprocessor include register renaming to eliminate register anti- and output (false) dependences, branch prediction to overcome control dependences, and data disambiguation to resolve memory dependences. The techniques for value prediction and value speculation have been proposed to break register flow (true) dependences among operations, so that dependent operations can be speculatively executed without waiting for producer operations to finish. This thesis presents a new combined hardware and compiler synergy, value speculation scheduling (VSS), to exploit the predictability of operations to improve the performance of microprocessors. The VSS scheme can be applied to dynamically-scheduled machines and statically-scheduled machines. To improve the techniques for value speculation, a value speculation model is proposed as solving an optimal edge selection problem in a data dependence graph. Based on three properties observed from the optimal edge selection problem, an efficient algorithm is designed and serves as a new compilation phase of benefit analysis to know which dependences should be broken to obtain maximal benefits from value speculation. A pure software technique is also proposed, so that existing microprocessors can employ software-only value speculation scheduling (SVSS) without adding new value prediction hardware and modifying processor pipelines. Hardware-based value profiling is investigated to collect highly predictable operations at run-time for reducing the overhead of program profiling and eliminating the need of profile training inputs.