This project addresses new techniques to break pure dependencies for horizontal (EPIC, VLIW) microarchitectures. The ultimate limit to schedule length is the longest chain through the computation. The research presents techniques for dependence chain splitting, which attack pure dependence chains using methods presented earlier for superscalar microarchitectures. The techniques for superscalar processors are not directly applicable to VLIW or EPIC architectures, since such architectures have no or little dynamic speculation mechanisms. Essentially, they execute code that is statically speculated by the compiler. The core idea of this work is to break a chain at a beneficial split point, and then speculatively execute the second half of the chain. This project develops techniques to find the best split point(s) of a region of code, develop hardware-based very low overhead value profiling, study VLIW/EPIC-specific value predictor designs, investigate heuristics to address the penalty branches to patch-up code, look into methods for adding confidence prediction for VLIW/EPIC value speculation, and study the potential of software-only value speculation. Issues such as register pressure impact and branch insertion for value speculation across control flow are also being explored.
