A Sample Profile-based Optimization Method with Better Precision

Conventional feedback-directed optimization is not widely adopted for the difficulties in generating representative training data sets. High runtime overhead and tedious re-compilation model obstruct its usability. Instruction-level hardware event sampling may overcome the drawbacks. There are still several challenges in creating accurate edge profiles, which is necessary to achieve competitive performance gains. This paper focuses on multiple hardware event profiles, supervised learning to discover patterns and generate heuristics to improve the precision of the instruction-level sample profile. We further enhance the efficacy of the smoothing algorithm used to construct the edge profiles from the instruction level and basic-block level samples. With these improvements, it is able to achieve about 70% of the performance obtained via instrumentation-based exact edge profiles for SPEC benchmarks, which also brings better performance of about 2.05%-13.81% improvement.