Thanks for that.  I'd call that pretty broken then:  the result of 
expm1(709.78) is effectively wrong by 24.4 trillion ulps (perhaps that 
should be 24.4 teraulps?).  I guess someone just coded a hard cutoff value 
somewhere between 709.7 and 709.78.

But since anyone getting this close to overflow is asking for trouble 
anyway, perhaps the easiest solution is just to change the test to test a 
value not quite as close to the overflow boundary.