Robert Dewar wrote:
Paolo Carlini wrote:
Actually, sorry, __builtin_cpow returns (nan, nan) (got sidetracked by a strange issue I'm seeing in the C++ library), even "worse", so to speak...
Paolo.
Well it certainly seems the right result in this case to me. Does the standard really require the wrong result here?
You mean (nan, nan) is ok? Why, exactly? Isn't (1,0) more consistent with the behavior of real power? (*) The standard is rather vague, in my reading, allowing "spurious exceptions" in "special cases", thus allowing a plain cexp(c*clog(z)).
Paolo.
(*) This "intuition" is confirmed by a range of independent "appliances", like high quality HP calculators or Maple...
