Decline infeasible full CUDA rewrite of Fuzzilli

[Copilot]

Two successive requests asked for a complete port of the Fuzzilli Swift codebase to CUDA, positioning the GPU as the primary engine for program generation, mutation, type propagation, SSA validation, FuzzIL→JS lifting, and testcase minimization.

Why the request was declined

• Wrong execution model: CUDA targets dense, regular, data-parallel workloads. Every component requested for GPU offload (recursive IR construction, constraint-based type propagation, use-def chain validation, serial string lifting, minimization round-trips) relies on pointer-heavy, variable-length, control-flow-divergent operations that map poorly to warp execution and produce severe divergence penalties.
• Bottleneck misidentified: Fuzzilli throughput is bounded by V8 REPRL execution (~1–5 ms/exec). Program generation takes microseconds. GPU-accelerating generation produces no measurable throughput gain.
• Scope incompatible with task constraints: The request is a new project (~50,000+ lines of Swift reimplemented in CUDA C++), not a repository change. Task rules require the smallest possible change to address a concrete issue.

Suggested alternatives for GPU-accelerated fuzzing research

• Offload coverage bitmap processing (popcount, novelty checks) to GPU via Thrust/CUB — this is the only stage with natural data parallelism
• Scale horizontally using Fuzzilli's existing distributed corpus-sharing across many CPU instances
• Use a GPU-resident language model (via ONNX Runtime or TensorRT) to score or rank mutation candidates, keeping the Swift fuzzing loop intact

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.