Identifying Behavioural Implications of Source Code Changes

Abstract

The dynamic behaviour of a software system changes as a consequence of developer’s static source code modifications. In this thesis, we improve upon a previous approach that combines static and dynamic analyses to categorize behavioural changes by greatly improving its accuracy through polymorphic mapping. We further refine the previous model by introducing a change-centric state transition model that captures the flow of call pairs among different partitions based on static and dynamic call graphs. We also extend the approach by incorporating complete dynamic call stacks into the analysis. Finally, we perform a longitudinal analysis of three software systems to categorize how they have dynamically evolved across 100 program versions. In our evaluation, the polymorphic mapping algorithm decreased mismatches between the static and dynamic analyses by 53%. In particular, we decreased the mismatch by 71% in the most important category of changes from the developer’s point of view. We found that developers introduce new behaviour more often than eliminating old behaviour. Our results show that developers are more likely to remove unexecuted/dead code than code that is executed dynamically. In terms of change types, we found that changes made to fix defects encountered the least inconsistent and unexpected behaviour, while changes made to add new functionality experienced the highest unexecuted behaviour. Finally, we argue that augmenting the dynamic analyses with call stacks provides useful information that helps developers analyze the implications of the call pairs highlighted by our analyses.