Combining GIN and PMD for Code Improvements
Amidst the recent excitement about using large language models to write software, it's easy to overlook the other ways that researchers are getting computers to write or improve code. This paper is an early look at combining two ideas: static analysis to detect problems in code found on Stack Overflow and genetic algorithms to improve those code snippets. The results are intriguing, but unsurprisingly, the authors found that the tools they used (PMD for code analysis and GIN for code mutation) may need some tweaking in order to play well together in this new way.
On a personal note, work like this has convinced me that the next major advance in programming languages won't come from research on classical concerns like type systems or concurrency mechanisms, but rather from asking, "How can we design a language so that automated tools can find and generate what we need more quickly and more accurately?" As we learn more about why tools like the ones described above can't (yet) do our programming for us, I believe we will see rapid evolution of both better tools and better languages for them to work on—better, that is, for other programs.
Sherlock A. Licorish and Markus Wagner. Combining GIN and PMD for code improvements. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, Jul 2022. doi:10.1145/3520304.3528772.
Software developers are increasingly dependent on question and answer portals and blogs for coding solutions. While such interfaces provide useful information, there are concerns that code hosted here is often incorrect, insecure or incomplete. Previous work indeed detected a range of faults in code provided on Stack Overflow through the use of static analysis. Static analysis may go a far way towards quickly establishing the health of software code available online. In addition, mechanisms that enable rapid automated program improvement may then enhance such code. Accordingly, we present this proof of concept. We use the PMD static analysis tool to detect performance faults for a sample of Stack Overflow Java code snippets, before performing mutations on these snippets using GIN. We then re-analyse the performance faults in these snippets after the GIN mutations. GIN's RandomSampler was used to perform 17,986 unique line and statement patches on 3,034 snippets where PMD violations were removed from 770 patched versions. Our outcomes indicate that static analysis techniques may be combined with automated program improvement methods to enhance publicly available code with very little resource requirements. We discuss our planned research agenda in this regard.