Semantic Conflicts in Version Control Systems

Tese de mestrado, Engenharia Informática (Engenharia de Software),Universidade de Lisboa, Faculdade de Ciências, 2021Merging parallel changes is a common occurrence for developers working in collaborative software projects. Nowadays, however, developers still rely on tools that perform textual merge to detect possible merge conflicts. While practical and efficient, this merging approach fails to detect semantic conflicts, that is, concurrent changes that cause the merged result to misbehave. It is well known that software testing is a powerful method to check if a software system matches its requirements and build the developer’s confidence that it is defect-free. As such, tests can be used to detect wrong behaviour in the merged result. Despite their capabilities, however, tests might not cover the changed parts and, hence, do not reveal the bugs introduced during the merge. As such, the effectiveness of testing for conflict detection is directly dependant on the quality of the test suite being used. So far, the detection of semantic conflicts has been mainly approached through techniques that rely on the static analysis of programs and target specific types of concurrent changes. This work explores a different approach and proposes a solution to detect semantic conflicts that relies on automatic test generation and targets a variety of causes of such conflicts. The main idea behind the solution is to detect semantic conflicts as soon as possible, overcoming the dependency on the quality of the test suite by using automatically generated tests. The information about the concurrent changes that happened in the merge is used to guide the test generation. This led to the development of UNSETTLE, a tool capable of comparing the changes in different versions of a code base to a set of change patterns capturing known and common causes of semantic conflicts and automatically search for tests that reveal a semantic conflict. We performed a preliminary evaluation that showed us that UNSETTLE has the potential to be useful in detecting semantic conflicts. UNSETTLE was capable of generating a test that detected the emergent behaviour in 6 out of 19 cases of real semantic conflicts