I Fuzzed, and Vibe Fixed, the Vibed C Compiler

10 hours ago

John Regehr initially didn't take much interest in CCC (Claude's C Compiler) but became curious after fuzzing it with Csmith and YARPGen.
CCC showed significant miscompilation rates: 14 out of 101 Csmith programs and 5 out of 101 YARPGen programs.
Csmith and YARPGen are randomized compiler testing tools that have detected many compiler defects, including miscompilations.
Regehr used YARPGen and C-Vise to reduce miscompilation bugs in CCC, then fixed them with the help of Codex (GPT-5.3-codex).
After 11 bug fixes, CCC no longer miscompiled programs in overnight fuzzing runs with both YARPGen and Csmith.
The bugs fixed were mostly surface-level, related to incorrect handling of C language semantics, such as sign-extension and narrowing optimizations.
Regehr noted that production-grade compilers like GCC and Clang rarely have such surface-level bugs; most bugs are found in optimizers.
CCC is impressive for a non-production compiler but lacks optimization and still has fundamental C interpretation issues.
Regehr's fork of CCC, with fixes, is available on GitHub for further testing.
The experiment showed that Codex can effectively fix compiler bugs without guidance, producing working solutions.

Hasty Briefsbeta