Simple and Effective Static Analysis to Find Bugs

Abstract

Much research in recent years has focused on using static analysis to find bugs in software. Many new approaches employing sophisticated program analysis techniques---inter-procedural, context-sensitive, or path-sensitive---have been developed. However, comparatively little work has been done on determining what bugs can be found using <i>simple</i> analysis techniques. We have found that simple static analysis techniques are effective at finding hundreds or thousands of serious software defects in several large commercial software applications.

In our research, we have attempted to characterize the bugs that can be found in production software using simple analysis techniques. Examples of simple analysis techniques include inspection of class hierarchies and method signatures, sequential scanning of program instructions using a state machine recognizer, intra-procedural dataflow analysis, and flow-insensitive whole program analysis. To determine what bugs may be found using these techniques, we performed <i>bug-driven</i> research. Starting from examples of real bugs, we tried to develop simple analysis techniques to find similar bugs. Using this approach, we found a large number of serious bugs in production applications and libraries with a relatively low percentage of false positives. The types of bugs our analysis uncovered in production code include null pointer dereferences, infinite recursive loops, data races, deadlocks, and missed thread notifications. One product of this work is a static analysis tool called FindBugs, which analyzes Java programs at the bytecode level. We have distributed FindBugs under an open-source license, and it has been widely adopted by commercial organizations and open-source projects. FindBugs has been downloaded more than 110,000 times since its initial release.

Our work makes several contributions to the field. First, we have cataloged many commonly-occurring bug patterns, described effective ways of finding occurrences of those patterns automatically, and classified common reasons why these bugs occur. Second, we have measured the accuracy of our bug detectors on production software and student programming projects, validating the effectiveness of the underlying static analysis techniques. Finally, we have described techniques for determining when static analysis warnings are added or removed as software evolves.