The awkward thing about trigonometry is that (as far as I've heard, at least) it's not yet known whether it's feasible to implement it with tolerable performance to the ½ULP accuracy that's needed in order to get consistent results. It's doable for things like +-*/, which are why those are portable and might plausibly be constable one day (if we can figure out how to deal with NAN), and maybe even could happen for relatively-simple well-behaved monotonic functions like pow2(x).
But something like sin(10300) is just fundamentally hard to get right. Just think how many decimal places of accuracy you need on π for the range reduction...