Context

In large NVIDIA C++ codebases such as CUDA runtime components or TensorRT execution paths, concurrency changes are risky because correctness depends on ordering, shared state, and synchronization semantics.

Core question

Explain how you would structure a code review for a complex C++ concurrency change. Your answer should cover:

1. How to break the change into reviewable units and identify critical shared-state interactions.
2. How to reason about ordering, lock dependencies, deadlock risk, races, and memory visibility.
3. How to validate the change with targeted tests, tooling, and rollout safeguards.

Scope guidance

The interviewer expects a coding-focused, systems-oriented explanation rather than people-process advice. Discuss concrete techniques such as dependency graphs, lock-order analysis, invariants, state transitions, and use of NVIDIA-relevant tooling where appropriate. You do not need to implement a full static analyzer, but you should explain the algorithmic structure of a strong review approach and the failure modes it is designed to catch.