Split Point Selection and Recovery for Value Speculation Scheduling

This paper extends previous work in value speculation scheduling [14] with an expanded ISA, new recovery code generation methods, and better instruction selection techniques. Our results show that critical instructions are not as predictable as non-critical instructions, instructions at the top of local critical paths have higher predictability than instructions in the middle, and that long-latency instructions have lower predictability than one-cycle instructions. Results from various split point selection heuristics show that breaking true data dependencies of long-latency instructions on local critical paths provides larger cycle savings than only predicting top or middle instructions. This paper also provides empirical results showing the relationship between the predictability of instructions with their latency, location within a dependence chain, and criticality. We also analyze potential critical path reduction when a compiler is allowed to choose the best predictor on a per-instruction basis.