[PATCH] libstdc++: Adjust generate_canonical (P0952, LWG2524) [PR119739]

[Tomasz Kaminski]

On Sat, Aug 9, 2025 at 2:00 PM Nathan Myers <ncm@cantrip.org> wrote:

> Changes in V3:
>  * Implement exactly as specified in the WP, for C++26 and up.
>  * Eliminate massive overhead seen in present implementation.
>
> Implement P0952R2 "A new specification for std::generate_canonical".
> It has us start over if the naive generation of a float in 0..1
> comes up exactly equal to 1, which occurs too rarely for the change
> to measurably affect performance.
>
> A test for the case already appears in 64351.cc. This change amounts
> to a different resolution for PR64351 and LWG 2524.
>
> libstdc++-v3/ChangeLog:
>         PR libstdc++/119739
>         * include/bits/random.tcc
> ---
>  libstdc++-v3/include/bits/random.tcc | 50 ++++++++++++++++++++++++++++
>  1 file changed, 50 insertions(+)
>
> diff --git a/libstdc++-v3/include/bits/random.tcc
> b/libstdc++-v3/include/bits/random.tcc
> index 53ccacb2e38..941495654c0 100644
> --- a/libstdc++-v3/include/bits/random.tcc
> +++ b/libstdc++-v3/include/bits/random.tcc
> @@ -3349,6 +3349,7 @@ namespace __detail
>         }
>      }
>
> +#if __cplusplus < 202300
>    template<typename _RealType, size_t __bits,
>            typename _UniformRandomNumberGenerator>
>      _RealType
> @@ -3385,6 +3386,55 @@ namespace __detail
>         }
>        return __ret;
>      }
> +#else
> +  template<typename _RealT, size_t __digits, typename _URBG>
> +    _RealT
> +    generate_canonical(_URBG& __urng)
> +    {
> +      static_assert(std::is_floating_point<_RealT>::value,
> +       "template argument must be a floating point type");
> +      static_assert(
> +       requires { [](uniform_random_bit_generator auto) {}(__urng); },
> +       "template argument must be an std::uniform_random_bit_generator");
> +
> +      const size_t __r = std::numeric_limits<_RealT>::radix;
> +      const auto __rdiff = [](auto __u) constexpr -> _RealT
> +       { return _RealT(__u - _URBG::min()); };
> +      constexpr static _RealT __R = __rdiff(_URBG::max()) + _RealT(1.0);
>
You do not need static here I believe, constexpr should be sufficient. Same
for
all other variables.

> +      const size_t __reg_digits = std::numeric_limits<_RealT>::digits;
> +      const size_t __d = __reg_digits < __digits ? __reg_digits :
> __digits;
> +      const auto __pow = [](_RealT __r, auto __d) consteval -> _RealT
>
I would use __builtin_pow here, it is constexpr in more modes than
corresponding
standard function, similarly for floor.

> +       {
> +         _RealT __res = __r;
> +         while (--__d) __res *= __r;
> +         return __res;
> +       };
> +      constexpr static _RealT __rd = __pow(_RealT(__r), __d);

+      const int __k = [](_RealT __rd, _RealT __R) consteval
>
I will use if constexpr for __R being a power of two, and use __builtin_log
in that case. You may need to add #pragma GCC diagnostic ignored
"-Wc++17-extensions"
sections around it to avoid issues.

> +       {
> +         unsigned __k = 1;
> +         _RealT __Ri = __R;
> +         while (__Ri < __rd)
> +           { __Ri *= __R, ++__k; }
> +         return __k;
> +       } (__rd, __R);
>

+      constexpr static _RealT __Rk = __pow(__R, __k);
> +      constexpr static  _RealT __x = [](_RealT __Rk, _RealT __rd)
> consteval
> +       { return std::floor(__Rk/__rd); }(__Rk, __rd);
>
I think this will not compile in clang, even built-ins are not constexpr
there.

> +      constexpr static _RealT __xrd = __x * __rd;
> +
> +      _RealT __sum;
> +      do
> +       {
> +         _RealT __Ri = _RealT(1.0);
> +         __sum = __rdiff(__urng());
> +         for (int __i = 1; __i != __k; ++__i)
> +           __Ri *= __R, __sum += __rdiff(__urng()) * __Ri;
> +       } while (__builtin_expect(__sum >= __xrd, false));
> +      _RealT __ret = std::floor(__sum / __x) / __rd;
> +      return __ret;
> +    }
> +#endif
>
>  _GLIBCXX_END_NAMESPACE_VERSION
>  } // namespace
> --
> 2.50.0
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://gcc.gnu.org/pipermail/libstdc++/attachments/20250811/478745d0/attachment-0001.htm>