What is acceptable for -ffast-math? A numerical viewpoint

[dewar@gnat.com]

<<You'll have to show me a _realistic_ example before I'll believe that. I
don't consider anything involving the likes of DBL_MAX or 10**1000 to be
realistic.
>>

OK, but there are plenty of casual programs that use large numbers, that's
why we expand the exponent when we go to double as well as expanding the
mantissa. Poorly designed fpt systems in which this is not done (e.g.
mainframe 360) proved a real nuisance, not to experts, who of course
carefully prove that their values are in range and avoid overflows, but
to casual programmers. It's often surprising how large intermediate
values can get.