A Fuzzer for the Toy Optimizer

3 days ago

Optimizers are difficult to implement correctly, often missing corner cases.
A fuzzer was developed to automatically catch bugs in optimizers, requiring a correctness oracle to detect meaningful issues.
Random programs are generated using APIs, with operations like load, store, and escape to simulate real-world usage.
The verifier checks invariants by comparing heap states between optimized and un-optimized programs under different aliasing conditions.
An interpreter tracks heap states and escaped values to ensure program correctness post-optimization.
Testing revealed that the optimizer correctly handles previously identified bug scenarios, demonstrating its robustness.
Commenting out key optimization logic quickly exposes bugs, proving the fuzzer's effectiveness in identifying issues.
Feedback suggests the equivalence definition may need adjustment for optimizations like allocation removal.

Hasty Briefsbeta