Re: [PATCH] Fix for bug libstdc++/110860

[Jonathan Wakely <jwakely@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 3499 bytes --]

On Mon, 14 Aug 2023 at 10:58, Paul Dreik via Libstdc++ <
libstdc++@gcc.gnu.org> wrote:

> The patch below fixes an issue with the fix already committed for
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110860 which unfortunately
> was not sufficient for small magnitude floating point values.
>
> With the patch in place, the code now survives the fuzzing I used to
> find the problem in the first place. Tested on amd64.
>
> I prepared the patch using git show, which should include the signoff as
> instructed per the DCO.
>

I couldn't apply the patch directly from the email, but I'm not sure where
it got mangled. I just applied it by hand instead.

Pushed to trunk, thanks for the patch!

I'll push it to gcc-13 shortly too.




>
> Thanks, Paul
>
> ------------------------------------------------------------------------
> commit 848b8d948787495e64ed9c55d681eccf730b74fb
> Author: Paul Dreik <gccpatches@pauldreik.se>
> Date:   Mon Aug 14 11:52:30 2023 +0200
>
>      libstdc++: Avoid problematic use of log10 in std::format [PR110860]
>
>      If abs(__v) is smaller than one, the result will be on the
>      form 0.xxxxx. It is only if the magnitude is large that more digits
>      are needed before the decimal dot.
>
>      This uses frexp instead of log10 which should be less expensive
>      and have sufficient precision for the desired purpose.
>
>      It removes the problematic cases where log10 will be negative or not
>      fit in an int.
>
>      Signed-off-by: Paul Dreik <gccpatches@pauldreik.se>
>
> diff --git a/libstdc++-v3/include/std/format
> b/libstdc++-v3/include/std/format
> index f4520ff3f..729e3d4b9 100644
> --- a/libstdc++-v3/include/std/format
> +++ b/libstdc++-v3/include/std/format
> @@ -1490,14 +1490,22 @@ namespace __format
>               // If the buffer is too small it's probably because of a
> large
>               // precision, or a very large value in fixed format.
>               size_t __guess = 8 + __prec;
> -             if (__fmt == chars_format::fixed && __v != 0) // +ddd.prec
> +             if (__fmt == chars_format::fixed) // +ddd.prec
>                 {
> -                 if constexpr (is_same_v<_Fp, float>)
> -                   __guess += __builtin_log10f(__v < 0.0f ? -__v : __v);
> -                 else if constexpr (is_same_v<_Fp, double>)
> -                   __guess += __builtin_log10(__v < 0.0 ? -__v : __v);
> -                 else if constexpr (is_same_v<_Fp, long double>)
> -                   __guess += __builtin_log10l(__v < 0.0l ? -__v : __v);
> +                 if constexpr (is_same_v<_Fp, float> || is_same_v<_Fp,
> double> ||
> is_same_v<_Fp, long double>)
> +                   {
> +                     // the number of digits to the left of the decimal
> point
> +                     // is floor(log10(max(abs(__v),1)))+1
> +                     int __exp{};
> +                     if constexpr (is_same_v<_Fp, float>)
> +                       __builtin_frexpf(__v, &__exp);
> +                     else if constexpr (is_same_v<_Fp, double>)
> +                       __builtin_frexp(__v, &__exp);
> +                     else if constexpr (is_same_v<_Fp, long double>)
> +                       __builtin_frexpl(__v, &__exp);
> +                     if (__exp>0)
> +                       __guess += 1U + __exp * 4004U / 13301U; //
> log10(2) approx.
> +                   }
>                   else
>                     __guess += numeric_limits<_Fp>::max_exponent10;
>                 }
>