https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79720
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |RESOLVED
Resolution|--- |INVALID
--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Thomas Koenig from comment #12)
> (In reply to Jakub Jelinek from comment #11)
>
> > Yeah, exactly. But then this is really NOTABUG. The compile time
> > evaluation needs to be .5ulp precise, not emulate whatever precision issues
> > the library has.
>
> When -fcx-fortran-rules is active(by default in Fortran,
> or when explicitly specified by the user) we generate inline
> code. The issue is the same then.
But the bug is likely still in the generated inline code, not in the constant
folding which uses mpfr/libmpc and should be 0.5ulp precise. And both the
generated inline code and library is a compromise between speed and precision,
I think for most of the complex math functions glibc doesn't guarantee 0.5ulp
precision even for double (unlike for many scalar math functions), for 0.5ulp
we'd most likely need a library implementation that would deal with the
problematic cases.
Closing as invalid.
