Re: Default Floating Point Format Specifiers

[William Ahern <william@...>]

On Thu, Jul 03, 2014 at 01:36:06PM -0500, Paige DePol wrote:
> On Jul 3, 2014, at 6:09 AM, Luiz Henrique de Figueiredo <lhf@tecgraf.puc-rio.br> wrote:
> 
> >> The problem, according to the IEEE floating point entries on Wikipedia, is that there are not enough digits specified.
> > 
> > This is on purpose so that numbers appear right for the casual user:
> > 
> >> a=0.1
> >> =a
> >     0.1
> >> =string.format("%.17g",a)
> >     0.10000000000000001
> > 
> > We don't want needless noise when using print (and other automatic
> > conversions via tostring). If you need round-trip strings, then use
> > string.format.
> 
> Ah, well that is a great reason to not specify larger format specifiers, I
> didn't realise 0.1 would turn into 0.10000000000000001 with a larger
> specifier. When I was reading about specifier digits I did not find any
> information on this side-effect.
<snip>

I assumed you already understood that part of the issue considering the
examples you originally provided.

Round-tripping _within_ a single application is straight forward, at least
theoretically. (Presuming, as Luiz pointed out, that's what you actually
want to accomplish.) But then there are a bunch of other issues if the
strings you generate are released into the wild. It's kind of a headache for
things like JSON, as well. It's relatively safe to assume binary64 IEEE754
these days (especially since it's the standard in JavaScript), but you can't
always assume that the library routines used to parse the numbers do the
expected thing.

The latest version of Visual Studio (14 STP) actually supports
DBL_DECIMAL_DIG (from C11), but they also changed the behavior of the printf
and strtod, which suggests to me that their old versions were probably a
source of interoperability headaches, even regarding regular decimal
conversions.

See http://blogs.msdn.com/b/vcblog/archive/2014/06/18/crt-features-fixes-and-breaking-changes-in-visual-studio-14-ctp1.aspx

I think until hexidecimal floats are widely supported you just have to be
conservative about what kind of precision you can expect when round-tripping
floats.

Maybe it's not as much of an issue as I think it is. It would be a cool
project, though, to document all the various behaviors across platforms,
versions, and languages. You'd probably get serious linkage from Stack
Overflow and other technical sites.