It might work great when you wrote it, but it might not even survive the next change to the codebase, and 10 years out, who is to know what programmers will come along and plug new stuff in without noticing the stack magic, etc

edit: another aspect of the same issue is that it can be difficult or impossible to write generic longjmp-safe code for many kinds of tasks. In C++ you have stack unwinding to e.g. delete large temporary heap allocations during a complex operation, but no such facility exists in C.

setjmp/longjmp will almost certainly require tracking any resource use carefully so that resources can be released at the setjmp point, when execution gets back there.

I mostly use it for terminating the recursive descent parsing on error. Otherwise it gets tedious to check return values everywhere in hand written code.

There is no reason libpng needed to use setjmp as part of its API, and that it does so is (I hope) widely regarded as a bad decision. It's "useful" only in so much as that is the API it defines, but there is no reason that this had to be the case, and simply returning an error code would have been very very much preferable.