Currently, the hazard analysis complains about having a UniquePtr<T> live across a GC call if T contains any unrooted pointers. The idea is that during a GC, we need to trace the T to update its pointers, and that's very unlikely to happen. (Note that Rooted<UniquePtr<T>> works and is understood.)

Nika pointed out that RefPtr and nsCOMPtr are similarly unsafe, though they are commonly used for this purpose and the rooting needs are normally handled via HoldJSObjects, which will require special detection.