Optimization-Directed Compiler Fuzzing for Continuous Translation Validation
summary
Optimuzz combines directed grey-box fuzzing with translation validation to find miscompilations; it targets optimization trigger locations, slices optimizer CFGs for selective coverage feedback, prioritizes seeds by a distance metric, and applies targeted mutation that focuses on def‑use neighbors. Instantiated for LLVM (Alive2) and TurboFan (TurboTV), it reproduces more benchmark miscompilations than prior fuzzers (e.g., 27 reproductions vs. none in some cases) and uncovers 55 new LLVM miscompilation bugs in the latest release.
AI-written summary · about
chapters (11)
- 0:00Compiler optimization correctness
- 2:51Directed fuzzing motivation
- 3:54Optimus framework overview
- 4:58Target identification in optimizations
- 5:21Guided search with coverage
- 7:49Targeted mutation strategy
- 10:14Evaluation results
- 11:41Ablation study effectiveness
- 13:06Bug discovery impact
- 14:35Summary and contributions
- 14:57Q&A
related talks
31m“Fuzz or lose..."
52mStructure aware fuzzing
14mFuzzing C++ Compilers via Type-Driven Mutation
14mFinding Compiler Bugs through Cross-Language Code Generator and Differential Testing
53mThree Colours of Fuzzing: Reflections and Open Challenges
20mAFL++ : Combining Incremental Steps of Fuzzing Research
20mKFuzzTest: Targeted Fuzzing of Internal Kernel Functions
53mFuzzing 101