Communications of the ACM
Neural Software Analysis
Credit: Getty Images
Software is increasingly dominating the world. The huge demand for more and better software is turning tools and techniques for software developers into an important factor toward a productive economy and strong society. Such tools aim at making developers more productive by supporting them through (partial) automation in various development tasks. For example, developer tools complete partially written code, warn about potential bugs and vulnerabilities, find code clones, or help developers search through huge code bases.
Key Insights
The conventional way of building developer tools is program analysis based on precise, logical reasoning. Such traditional program analysis is deployed in compilers and many other widely used tools. Despite its success, there are many problems that traditional program analysis can only partially address. The reason is that practically all interesting program analysis problems are undecidable, that is, giving answers guaranteed to be precise and correct is impossible for non-trivial programs. Instead, program analysis must approximate the behavior of the analyzed software, often with the help of carefully crafted heuristics.
No entries found