On Mon, 20 Aug 2007, Nicholas Clark wrote:
> On Mon, Aug 20, 2007 at 10:18:33AM -0400, Andy Dougherty wrote:
> > Your trick (extended, if needed, by clever Configure.pl probes) will
> > cleverly suppress the warning for the f == 0 case, while still retaining
> > the exact comparison. (Hmm. Not sure about negative zero offhand. I'd
> > have to test.)
>
> I'm not convinced that it's great. It assumes
>
> 1: that the floating point representation is all bits zero [Alphas can run in
> non IEEE, IIRC]
> 2: that there is an integer type exactly the same size as the floating point
> type. (Larger and smaller will both have problems)
Yes. This sort of thing is what I meant by "clever Configure.pl probes".
Configure.pl could figure out what the bit pattern is for a FLOATVAL
of 0.0. And check for -0.0, and then do special testing for the
10-byte variety, and perhaps even more special testing for the extra
precision that math co-processors used to keep (but throw away
when it moved values into registers) (Insert painful flashback of
post-processing 68000 assembler output to use a math
coprocessor). But it's doable, I think, with
persistence and a reasonable amount of useful feedback.
> I can't think of a clean, portable, efficient way to test that a floating
> point variable is zero other than == 0.0;
Nor can I. Thus you either use lots of platform-specific code (determined
by Configure.pl) or turn off gcc's -Wfloat-equal warning flag and just
write f == 0.0.
Personally, I'd just turn off the -Wfloat-equal flag, but I have no
objection if someone else wants to do all the work to do it the other way.
--
Andy Dougherty [EMAIL PROTECTED]
