| Concurrency Errors, such as atomicity violations, are notorious to detect and confirm. Although various (static, dynamic, and hybrid) approaches have been proposed for detecting atomicity violations, they are difficult to be applied on real-world applications because they are inaccurate and can not distinguish false and benign alarms from real bugs.;This thesis presents a novel method that accurately classifies the above three kinds of alarms using machine learning and program analysis. In the program analysis phase, we generate a comprehensive scenario for each warning by analyzing program data flows and control flows. In the machine learning phase, we encode each scenario into a fixed-size sequence by analyzing its various properties. Finally, we generate a decision tree for classification through learning given training sequences.;We have implemented the approach and tested it on a few benchmarks to demonstrate its effectiveness. |
