Dynamically fighting bugs : prevention, detection and elimination

Abstract

This dissertation presents three test-generation techniques that are used to improve software quality. Each of our techniques targets bugs that are found by different stake-holders: developers, testers, and maintainers. We implemented and evaluated our techniques on real code. We present the design of each tool and conduct experimental evaluation of the tools with available alternatives. Developers need to prevent regression errors when they create new functionality. This dissertation presents a technique that helps developers prevent regression errors in object-oriented programs by automatically generating unit-level regression tests. Our technique generates regressions tests by using models created dynamically from example executions. In our evaluation, our technique created effective regression tests, and achieved good coverage even for programs with constrained APIs. Testers need to detect bugs in programs. This dissertation presents a technique that helps testers detect and localize bugs in web applications. Our technique automatically creates tests that expose failures by combining dynamic test generation with explicit state model checking. In our evaluation, our technique discovered hundreds of faults in real applications. Maintainers have to reproduce failing executions in order to eliminate bugs found in deployed programs. This dissertation presents a technique that helps maintainers eliminate bugs by generating tests that reproduce failing executions. Our technique automatically generates tests that reproduce the failed executions by monitoring methods and storing optimized states of method arguments.

(cont.) In our evaluation, our technique reproduced failures with low overhead in real programs Analyses need to avoid unnecessary computations in order to scale. This dissertation presents a technique that helps our other techniques to scale by inferring the mutability classification of arguments. Our technique classifies mutability by combining both static analyses and a novel dynamic mutability analysis. In our evaluation, our technique efficiently and correctly classified most of the arguments for programs with more than hundred thousand lines of code.