http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47558
--- Comment #65 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-02-07 20:55:31 UTC --- > /* The system ___divdc3 routine in libSystem on darwin10 is not > accurate to 1ulp, ours is, so we avoid ever using the system name > for this routine and instead install a non-conflicting name that > is accurate. I think this comment in gcc/config/darwin.c and gcc/config/i386/darwin.h is at best misleading. The problem is not about accuracy, but the range of validity of the chosen algorithm. In darwin10, it is tiny<arg(complex_denominator)**2<huge, while in the FSF lib it is tiny<arg(complex_denominator)<huge. Note that codes relying on this extended range are likely to generate infinities elsewhere.
