When would you choose ReentrantLock instead of synchronized?

I do not choose ReentrantLock as a reflexive performance optimization; modern intrinsic locks are highly optimized and workload behavior decides. An explicit lock that crosses methods or has several exit paths is easy to leak and hard to review. Lock scope, ordering, timeouts, and failure behavior need tests and documentation. If coordination becomes elaborate, a queue, semaphore, actor-style owner, or higher-level concurrent component may remove the shared-lock protocol entirely.