cmp() on integers - is there guarantee of returning only +-1 or 0?

[Dmitry Anikin]

doc says that it must be > 0, or < 0, but it seems that
it returns +1 or -1. Can it be reliably used to get the sign of x:
cmp(x, 0) like pascal Sign() function does? I mean, I'm
pretty sure that it can be used, but is it mentioned somewhere
in language spec, or it may be implementation defined?
If so, any other simple means of _reliably_ getting the sign?