I looked at the source now and found nothing specific.

The optimisation seems to be wrong; clang returns a constant 1, which assumes that the function is never called with n=0. Which is true for the Collatz conjecture (it starts at 1), but surely clang has no right to assume that n≥1?