Date of Award
Doctor of Philosophy
Electrical and Computer Engineering
Tien N. Nguyen
Software systems inevitably contain a large amount of repeated artifacts at different level of abstraction---from ideas, requirements, designs, algorithms to implementation. This dissertation focuses on analyzing software repetitiveness at implementation code level and leveraging the derived knowledge for easing tasks in software maintenance and evolution such as program comprehension, API use, change understanding, API adaptation and bug fixing. The guiding philosophy of this work is that, in a large corpus, code that conforms to specifications appears more frequently than code that does not, and similar code is changed similarly and similar code could have similar bugs that can be fixed similarly.
We have developed different representations for software artifacts at source code level, and the corresponding algorithms for measuring code similarity and mining repeated code. Our mining techniques bases on the key insight that code that conforms to programming patterns and specifications appears more frequently than code that does not. Thus, correct patterns and specifications can be mined from large code corpus. We also have built program differencing techniques for analyzing changes in software evolution. Our key insight is that similar code is likely changed in similar ways and similar code likely has similar bug(s) which can be fixed similarly. Therefore, learning changes and fixes from the past can help automatically detect and suggest changes/fixes to the repeated code in software development.
Our empirical evaluation shows that our techniques can accurately and efficiently detect repeated code, mine useful programming patterns and API specifications, and recommend changes. It can also detect bugs and suggest fixes, and provide actionable insights to ease maintenance tasks. Specifically, our code clone detection tool detects more meaningful clones than other tools. Our mining tools recover high quality programming patterns and API preconditions. The mined results have been used to successfully detect many bugs violating patterns and specifications in mature open-source systems. The mined API preconditions are shown to help API specification writer identify missing preconditions in already-specified APIs and start building preconditions for the not-yet-specified ones. The tools are scalable which analyze large systems in reasonable times. Our study on repeated changes give useful insights for program auto-repair tools. Our automated change suggestion approach achieves top-1 accuracy of 45%-51% which relatively improves more than 200% over the base approach. For a special type of change suggestion, API adaptation, our tool is highly correct and useful.
Hoan Anh Nguyen
Nguyen, Hoan Anh, "Analyzing repetitiveness in big code to support software maintenance and evolution" (2015). Graduate Theses and Dissertations. 14591.