[gfortran,patch] FPE options, once again

[Richard Guenther]

On 9/8/05, Paul Brook <paul@codesourcery.com> wrote:
> On Thursday 01 September 2005 21:46, FX Coudert wrote:
> > This is the new version of the FPE patch for gfortran, using
> > glibc-specific code on systems where it is available, and runtime
> > detection of SSE units.
> >
> > I think the code itself is now in pretty good shape, although i'm a bit
> > worried of having heard no word from other gfortran developpers about
> > things such as the option name and the need/convenience of having
> > environment variables.
> 
> I thought the last time this was discussed (for a different option), the
> conclusion was that environment variables were more trouble that they were
> worth. Applications are liable to mysteriously break because a user didn't
> have the exact same combination of magic environment variables set.
> 
> My preference is for a commandline switch (like you implemented), and maybe a
> new intrinsic procedure so an application can control the setting at runtime.
> 
> > Regtested on i686-linux. Testing in progress on i686-freebsd,
> > x86_64-linux and i386-openbsd (no glibc, no SSE).
> >
> > OK for mainline? Do we want this on 4.0?
> 
> > + int invalid = exceptions & 0x01;
> > + int denormal = exceptions & 0x02;
> > + int zero = exceptions & 0x04;
> > + int overflow = exceptions & 0x08;
> > + int underflow = exceptions & 0x10;
> > + int precision = exceptions & 0x20;
> >...
> > + gfc_option.fpe_invalid = 0;
> > + gfc_option.fpe_denormal = 0;
> > + gfc_option.fpe_zero = 0;
> > + gfc_option.fpe_overflow = 0;
> > + gfc_option.fpe_underflow = 0;
> > + gfc_option.fpe_precision = 0;
> >[etc]
> 
> I agree with RTHs comment that there's no point having all there as separate
> variables. We convert to/from a bitmap anyway, so we may as well just use
> that bitmap all the time. Also make the bitmasks into #defines so we don't
> have hardcoded values in multiple places.

What about my earlier comment to make this available to C/C++, too, by
putting it in either a different crt*.o like HJL did with the denormals 
switching, or in libgcc.  rth will probably know the best place to put this.

Richard.