[gfortran] Fix NEAREST constant folder

[Tobias Schlüter]

Steve Kargl wrote:
> MPFR does not support subnormal numbers, so there isn't much you 
> can do.  If you don't range_check, then underflows elsewhere will
> be treated as if you set the result to zero.  It's concievable that
> someone may use nearest(0,1) to avoid division by zero.  So, I
> think our only option is to return +-tiny().  I suggest we have
> a comment with this info in the source and perhaps issue a warning
> that gfortran does not currently support subnormal numbers.

IOW we can't get compile-time arithmetic right with mpfr?  This is very
unfortunate.  (Another possible application for NEAREST(0.,1.) is to check for
the existence of denormals: NEAREST(0.,1.) < TINY(0.) should be true iff the
machine representation supports subnormals)

> OK to commit to mainline.  OK for 4.0 branch when Mark re-opens the
> branch.

After I had committed this, I realized that my testcase doesn't test double
precision numbers, so I tried adding a few tests for double precision to the
testcase but they all failed.  After a close look the constant folder appears
to do the right thing, and the library implementation simply calls nexafter()
from the C library so I believe the culprit is gfc_conv_mpfr_to_tree (maybe
we're losing precision because we go via a decimal string?), I'll investigate
this further tomorrow.

- Tobi