This PR adds a safer wrapper for setjmp called call_with_setjmp that helps to prevent LLVM miscompilations such as happened here where LLVM emitted a store immediate 1 instead of a load-add-store sequence for this code snippet:

let mut c = 0;
let mut buf = JumpBuf::new();

let r = setjmp(ptr::from_mut(&mut buf));
c += 1; // on opt-lvl >=1 this becomes a store immediate 1 instead of a load-add-store
        // because LLVM rightly assumes that this point will only be reached once per frame
if r == 0 {
    assert_eq!(c, 1);
    longjmp(ptr::from_mut(&mut buf), 1234567);
}
assert_eq!(c, 2);
assert_eq!(r, 1234567);

(see on godbolt here)

If you force c into a different stack frame or into a static things are fine btw if you were curious (see the setjmp_longjmp_simple test for more).

The proper solution though - documented here and adapted into a crate https://github.com/pnkfelix/cee-scape - is to use inline assembly and mark all registers as clobbered along with preventing accidental misuse of the JmpBuf in invalid contexts.