Generating, Locating, and Applying Systematic Edits by Learning from Example(s) Ph.D. Proposal

Programmers make systematic edits—similar, but not identical changes to multiple places during software development and maintenance. Finding all the correct locations and making correct edits is a tedious and error-prone process. Existing tools for automating systematic edits are limited because they do not support edit generation, edit location suggestion, or edit application at the same time, except for specialized or trivial edits. However, since many similar changes are needed in locations that contain similar context (the surrounding dependent code and syntactic structures), there is an opportunity to automate the systematic editing process by inferring edit scripts and characterizing their context from code that developers changed already. The challenge is we need to abstract and generalize from example(s) in order to create an edit script that is correct in many different contexts. This thesis seeks to substantially improve the efficiency and correctness of automatic systematic program transformation. (1) We design and implement Sydit to generate an abstract, context-aware edit script from a single changed method, and apply it to user-selected target method(s). This approach correctly performs many edits, but we show that the edit scripts from one example are not well suited to finding new locations. (2) We thus design and implement Lase to generate a partially abstract, context-aware edit script from multiple changed methods and show how to use the same script to automatically find edit locations and apply edits. Our experiments compare the systematic edits generated by Lase and those created by developers and find that Lase is effective at finding correct locations and making correct edits. (3) We propose to explore a new approach to generate high-level, collaborative edit scripts that span more than one method to handle sophisticated program transformations, such as refactoring. (4) Finally, we propose an approach for checking the correctness of the transformations. Our work aims at relieving developers from tedious and error-prone code changes by performing systematic transformations in a consistent way, and by guiding automated program repair, feature addition, and refactoring.