Automatic Reproducible Crash Detection

Abstract

Crash reproduction, which spends much time of developers in reading and understanding source code, is a crucial yet time-consuming task in program debugging. To reduce the time and resource cost, automatic techniques of test generation have been proposed. These techniques aim to automatically generate test cases to reproduce the scenario of a crashed project. Unfortunately, due to the lack of a detailed comprehension of the source code, a generated test case may fail in reproducing an expected crash. In this paper, we propose an automatic approach to reproducible bug detection. This approach predicts whether a crash is difficult to reproduce or not via training a classifier based on historical reproducible crash data. If a crash is difficult to reproduce, it is better to assign the crash to a developer, instead of using an automatic technique of test generation. Our work can help to prioritize crashes and to save the cost of developers. Preliminary experiments show that our approach effectively detects reproducible crashes via evaluating 45 crashes.