Semantic Impact and Faults in Source Code Changes: An Empirical Study

Changes to source code have become a critical factor in fault predictions. Text or syntactic approaches have been widely used. Textual analysis focuses on changed text fragments while syntactic analysis focuses on changed syntactic entities. Although both of them have demonstrated their advantages in experimental results, they only study code fragments modified during changes. Because of semantic dependencies within programs, we believe that code fragments impacted by changes are also helpful. Given a source code change, we identify its impact by program slicing along the variable def-use chains. To evaluate the effectiveness of change impacts in fault detection and prediction, we compare impacted code with changed code according to size and fault density. Our experiment on the change history of a successful industrial project shows that: fault density in changed and impacted fragments are higher than other areas; for large changes, their impacts have higher fault density than changes themselves; interferences within change impact contribute to the high fault density in large changes. Our study suggests that, like change itself, change impact is also a high priority indicator in fault prediction, especially for changes of large scales.