Interesting, I've had to use it a number of times over the years despite never really doing much unsafe. At least to me, it seems pretty well-scoped as a workaround from the requirements that the compiler has around needing to use generic type parameters in type definitions, which certainly isn't something you need to be writing unsafe code to run into. I wouldn't be shocked if it used unsafe under the hood, but then again, so does Vec.

The original reason to design it (instead of the previously inferred bivariance) was so that unsafe code that really does not want bivariance, and will be unsound if it will be used, will remember to consider that.

It doesn't use unsafe under the hood, rather it's compiler magic.

> At least to me, it seems pretty well-scoped as a workaround from the requirements that the compiler has around needing to use generic type parameters in type definitions

The reason those requirements exist is (primarily) to do with unsafe code. Specifically it’s about deciding the variance of the type (which doesn’t matter for a truely unused type parameter).