Re: [PATCH] msvcrt: Fix strtof() error reporting for values out of float range

[Martin Storsjö]

On Wed, 28 Jul 2021, Piotr Caban wrote:
> Hi Martin,
>
> On 7/28/21 12:44 AM, Martin Storsjo wrote:
> > float CDECL _strtof_l( const char *str, char **end, _locale_t locale )
> >   {
> > - return _strtod_l(str, end, locale);
> > - double ret = _strtod_l(str, end, locale);
> > - if (isfinite(ret)) {
> > -
> >    /* Check for cases that aren't out of range for doubles, but that 
> >
> > are
> > -
> >     * for floats. */
> > -
> >    if (ret > FLT_MAX)
> > -
> >        *_errno() = ERANGE;
> > -
> >    else if (ret < -FLT_MAX)
> > -
> >        *_errno() = ERANGE;
> > -
> >    else if (ret > 0 && ret < FLT_MIN)
> > -
> >        *_errno() = ERANGE;
> > -
> >    else if (ret < 0 && ret > -FLT_MIN)
> > -
> >        *_errno() = ERANGE;
> > - }
> > - return ret;
>
> It doesn't work for denormals (e.g. for "1.4e-45" input string).

Hmm, right... When I studied this aspect for mingw-w64, I noted that 
different CRTs seem to differ in what they do for denormals in 
strtod/strtof; some set ERANGE, some don't. But you're right that ucrtbase 
doesn't seem to set it for these denormals, so we shouldn't either.
> Maybe something along this lines will work better:
> if (ret && isfinite(ret)) {
>    float f = ret;
>    if (!f || !isfinite(f))
>        *_errno() = ERANGE;
> }

Thanks! That does seem to work. I'll amend the patch to include testcases 
for 0 too, in addition to denormals.

// Martin