8 July 2022 in Empirical Software Engineering
Compiler fuzzing techniques require a means of generating programs that are free from undefined behaviour (UB) to reliably reveal miscompilation bugs. Existing program generators such as Csmith achieve UB-freedom by heavily restricting the form of generated programs. The idiomatic nature of the resulting programs risks limiting the test coverage they can offer, and thus the compiler bugs they can discover. We investigate the idea of adapting existing fuzzers to be less restrictive concerning UB, in the practical setting of C compiler testing via a new tool, CsmithEdge, which extends Csmith. CsmithEdge probabilistically weakens the constraints used to enforce UB-freedom, thus generated programs are no longer guaranteed to be UB-free. It then employs several off-the-shelf UB detection tools and a novel dynamic analysis to (a) detect cases where the generated program exhibits UB and (b) determine where Csmith has been too conservative in its use of safe math wrappers that guarantee UB-freedom for arithmetic operations, removing the use of redundant ones. The resulting UB-free programs can be used to test for miscompilation bugs via differential testing. The non-UB-free programs can still be used to check that the compiler under test does not crash or hang. Our experiments on recent versions of GCC, LLVM and the Microsoft Visual Studio Compiler show that CsmithEdge was able to discover 7 previously unknown miscompilation bugs (5 already fixed in response to our reports) that could not be found via intensive testing using Csmith, and 2 compiler-hang bugs that were fixed independently shortly before we considered reporting them.
10 July 2025 in Automated Software Engineering
Despite the wide availability of automated testing techniques such as fuzzing, little attention has been devoted to testing computer architecture simulators. We propose a fully automated approach for this task. Our approach uses large language models (LLM) to generate input programs, including information about their parameters and types, as test cases for the simulators. The LLM’s output becomes the initial seed for an existing fuzzer, , which has been enhanced with three mutation operators, targeting both the input binary program and its parameters. We implement our approach in a tool called . We use it to test the system simulator. discovered 21 new bugs in , 14 where ’s software prediction differs from the real behaviour on actual hardware, and 7 where it crashed. New defects were uncovered with each of the 6 LLMs used.
21 January 2025 in Automated Software Engineering
Ever since the first large language models (LLMs) have become available, both academics and practitioners have used them to aid software engineering tasks. However, little research as yet has been done in combining search-based software engineering (SBSE) and LLMs. In this paper, we evaluate the use of LLMs as mutation operators for genetic improvement (GI), an SBSE approach, to improve the GI search process. In a preliminary work, we explored the feasibility of combining the Gin Java GI toolkit with OpenAI LLMs in order to generate an edit for the tool. Here we extend this investigation involving three LLMs and three types of prompt, and five real-world software projects. We sample the edits at random, as well as using local search. We also conducted a qualitative analysis to understand why LLM-generated code edits break as part of our evaluation. Our results show that, compared with conventional statement GI edits, LLMs produce fewer unique edits, but these compile and pass tests more often, with the model finding test-passing edits 77% of the time. The and LLMs are roughly equal in finding the best run-time improvements. Simpler prompts are more successful than those providing more context and examples. The qualitative analysis reveals a wide variety of areas where LLMs typically fail to produce valid edits commonly including inconsistent formatting, generating non-Java syntax, or refusing to provide a solution.