| Feedback-Driven Optimization (FDO) is a technique that has been widely adopted by modern compilers. By allowing the compiler to gather profiles of a program's dynamic behaviors and to perform optimization based on the profiles, it often substantially enhances the quality of the generated executable. However, an important open question is how to advance FDO to adapt to the arising trends in modern computing, such as the rapid upgrading of software, the increased variety of program input data sets and the popular demands of mobile devices that have stringent memory and storage budget. This dissertation aims to answer the question through a systematic exploration from three key aspects: How to reduce profiling overhead in the context of frequent software enhancement and upgrading? How to improve the usefulness of FDO when the profile is gathered through sampling, which is a common technique to alleviate profiling overhead? How to space-efficiently address the input sensitivity problem, which refers to the case when applying optimization based on one profile leads to inferior performance on a different input? We developed several novel techniques, namely profile migration, profile rectification and space-efficient versioning to address those problems. Experiments demonstrate that these techniques are promising for advancing FDO to meet the arising needs of modern computing. |
