Re: better float2str() C code, anyone?

[David Jones <drj@...>]

On Dec 01, 2004, at 14:54, Luiz Henrique de Figueiredo wrote:

> > One issue I've had from time to time is that the default
> >    #define LUA_NUMBER_FMT        "%.14g"
> > is not long enough to do reversible conversions on doubles. I believe
> > you need %.19g for the conversion to be reversible.
>
> I think it's %.17g for IEEE doubles. (Despite what DBL_DIG says!)
> Lua was like that once, but then you get ugly things like this:
>
> > =string.format("%.17g",0.34)
> 0.34000000000000002
>
> > =0.34
> 0.34
>
> > =34/100
> 0.34
>
> Users complaint that this is an error in Lua! And it does not help to 
> point
> them to docs like http://lua-users.org/wiki/FloatingPoint ...

Can you imagine a document that it would be helpful to point your users 
at?  If you can, then perhaps we can write it.

As I see it Lua used to do The Right Thing.  It used to use %.17g which 
would give reliable external and internal conversion.  As long as the C 
library was "good enough".  glibc and FreeBSD libc were good enough, 
but many others weren't.  It's a shame that Lua now masks poor quality 
C implementations at the expense of reliability.  It's not particularly 
a shame that you changed it, it's a shame that the state of C libraries 
meant you felt compelled to change it.

David Jones