[Bug libfortran/59774] [Regression] Inconsistent rounding between -m32 and -m64

["jvdelisle at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59774

--- Comment #10 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> ---
Very interesting and good sleuthing!

The way this is suppose to work:

For G formatting, we compute the equivalent F or E formatting, as defined in
the standard, and pass this new format to output_float which then uses the
regular existing formatting processes to do the work.

This line: (on or about line 1105 in write_float.def

  newf.u.real.d = m == 0.0 ? d - 1 : -(mid - d - 1) ;\

is suppose to compute the new 'd' from mid and the given "d" and pass that into
output_float using the newf fnode structure.  In the failing case the new 'd'
is being set to zero and being passed on.

As far as your 'kludge'.  I don't think of it that way, but we should see if
there is a way to correctly compute the d_o value where you are using it in
output_float. There should be a way. Since the standard defines all we do in
terms of F and E formatting, I think we are mishandling something there in
output_float.  You are very close to the solution here. (of course I could be
wrong). Someone on the list once said, if it fixes the bug, its probably good
enough.  The computer has no feelings about "correctness" of approach.