Eliminate Redundant Loop-Carried Flow Dependences for VLIW Architectures
消除VLIW结构上的循环体间冗余流相关

Data dependence is the basic guideline and constraint for parallel processing. The structural distinctiveness of VLIW (very long instruction word), i.e., the lockstep property, leads to remarkable uniqueness in its data dependence analysis, as is shown in this paper. Flow dependences with the same iteration distance comprise a linear ordered set, while characteristic flow dependences with different iteration distances may also have inclusion relationship between each other. From such discovery, a novel method is presented to eliminate redundant loop-carried flow dependences for VLIW architectures, so as to facilitate loop scheduling. The method is complete in that it gets rid of all redundancies. The sufficient and necessary preconditions to identify redundancy for single and multiple iteration distances are given, and a linear ordered algorithm is described. The method is generic, and can serve as a foundation for research on VLIW, such as software pipelining and multi-instruction stream scheduling.