ALPHAPROG: Reinforcement Generation of Valid Programs for Compiler Fuzzing

Fuzzing is a widely-used testing technique to assure software robustness. However, automatic generation of high-quality test suites challenging, especially for that takes in highly-structured inputs, such as the compilers. Compiler fuzzing remains difficult generating tons syntactically and semantically valid programs not trivial. Most previous methods either depend on human-crafted grammars or heuristics learn partial language patterns. They both suffer from completeness issue classic puzzle testing. To mitigate problem, we propose knowledge-guided reinforcement learning-based approach compiler fuzzing. We first design naive learning model which evolves with sequential mutation rewards provided by target test. By iterating training cycle, learns generate can improve efficacy well. implement proposed method into tool called ALPHAPROG. analyze framework four different reward functions our study reveal effectiveness ALPHAPROG also reported two important bugs production were confirmed addressed project owner, further demonstrates ALPHAPROG's applied value practice.