Mail Archives: cygwin-developers/1999/01/18/19:54:21
Geoffrey Noer <noer AT cygnus DOT com> writes:
> On Tue, Jan 12, 1999 at 10:51:40AM -0500, DJ Delorie wrote:
> >
> > > Why is gcc not consistent with the "HOST"
> > > compiler in this area?
> >
> > I don't know, but remember that the x86 support for gcc isn't limited
> > to cygwin - it supports many other x86 targets like djgpp, linux, SCO,
> > etc.
>
> I would think that gcc should be consistent with VC++ in the way that
> was suggested. Couldn't this be set in a Cygwin- and Mingw-target
> specific x86 config file?
Actually, it affects all the x86 platforms, it's just a question of to
what degree.
There's been quite a bit of discussion regarding this in egcs group. My
current fix (local, haven't submitted yet) is to change the rounding mode
to MSVC/Intel-VTune compatible mode; if you want to go to extended fp
mode, you specify -mextended-fp (or a similarly named switch, can't
remember offhand) that simply links in a .o file that changes the mode
via GCC's __attribute__((constructor)) mechanism.
Interix on the other has solved in a much more general way that applies
to *all* x86 ports, and I'm currently working with Donn Terry to get these
changes into egcs.
For those of who follow egcs list, you've already seen all the discussion.
I'm dismayed at the fact that people keep on telling me that it's wrong
to expect proper numerical results from x86 and so on. I get expected
results from a host of other compilers on this very platform.
For those interested, please check out following threads:
"x86 regression: optimizer bug in 1.1 and dev"
http://www.cygnus.com/ml/egcs-bugs/1998-Nov/0269.html
and,
"Floating-point Consistency, -ffloat-store, and x86 (mostly)"
http://www.cygnus.com/ml/egcs/1998-Dec/0049.html
Regards,
Mumit
- Raw text -