On Wed, 17 Dec 2025 at 08:56, Tomasz Kamiński <[email protected]> wrote:
>
> The r16-6177-g866bc8a9214b1d introduced type-constraint on _Urbg template
> parameter in __glibcxx_concepts, with was inconsistent with declaration in
> bits/random.h and definition in bits/random.tcc causing the missing symbol
> errors in tests.
>
> Furthermore, this made the mangled name of generate_canonical in C++20
> mode different from older standard and previous versions of GCC.
Which isn't really a problem since it's a function template, so old
code has its own instantiation of the old definition.
In fact, maybe it's better to have a different mangled name, so that
the new definition is a distinct symbol and the linker won't choose a
symbol generated from the old definition, giving surprising results in
a program where some TUs are compiled by GCC 16 and some TUs are
compiled by GCC 15. If all calls to generate_canonical use the old
definition, even for TUs compiled with GCC 16, that would be
surprising.
So I agree that we need the declaration and definition to match
(obviously!) and that we should not use a type constraint on the new
definition (the standard doesn't say it has any Constraints, so
there's no reason to add that, the static_assert is better). But I
think we should put the new definition in an inline namespace.
In <chrono> we do something similar for system_clock and steady_clock:
_GLIBCXX_BEGIN_INLINE_ABI_NAMESPACE(_V2)
struct system_clock
...
_GLIBCXX_END_INLINE_ABI_NAMESPACE(_V2)
also in <condition_variable> and <bits/algo.h>.
We should do that for the new generate_canonical, ensuring that the
declaration in <bits/random.h> matches it, like so:
--- a/libstdc++-v3/include/bits/random.h
+++ b/libstdc++-v3/include/bits/random.h
@@ -51,6 +51,9 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
// std::uniform_random_bit_generator is defined in <bits/uniform_int_dist.h>
+#ifndef _GLIBCXX_USE_OLD_GENERATE_CANONICAL
+_GLIBCXX_BEGIN_INLINE_ABI_NAMESPACE(_V2)
+#endif
/**
* @brief A function template for converting the output of a (integral)
* uniform random number generator to a floatng point result in the range
@@ -60,6 +63,9 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
typename _UniformRandomNumberGenerator>
_RealType
generate_canonical(_UniformRandomNumberGenerator& __g);
+#ifndef _GLIBCXX_USE_OLD_GENERATE_CANONICAL
+_GLIBCXX_END_INLINE_ABI_NAMESPACE(_V2)
+#endif
/// @cond undocumented
// Implementation-space details.
>
> libstdc++-v3/ChangeLog:
>
> * include/bits/random.tcc (generate_canonical)
> [!_GLIBCXX_USE_OLD_GENERATE_CANONICAL]: Use static_assert
> instead of type-constraint on template parameter.
> ---
> The missing error was much more prosaic in their root cause, than I
> have originally suspected.
>
> Testing on x86_64-linux. OK for trunk?
>
> libstdc++-v3/include/bits/random.tcc | 11 ++++-------
> 1 file changed, 4 insertions(+), 7 deletions(-)
>
> diff --git a/libstdc++-v3/include/bits/random.tcc
> b/libstdc++-v3/include/bits/random.tcc
> index 38e8645c88c..d0aed028ed6 100644
> --- a/libstdc++-v3/include/bits/random.tcc
> +++ b/libstdc++-v3/include/bits/random.tcc
> @@ -3619,12 +3619,6 @@ namespace __detail
> template <> const bool __is_rand_dist_float_v<long double> = true;
> #endif
>
> -#ifdef __glibcxx_concepts
> -# define _Uniform_random_bit_generator uniform_random_bit_generator
> -#else
> -# define _Uniform_random_bit_generator typename
> -#endif
> -
> // Note, this works even when (__range + 1) overflows:
> template <typename _Rng>
> constexpr bool __is_power_of_2_less_1(_Rng __range)
> @@ -3646,10 +3640,13 @@ namespace __detail
> * @since C++11
> */
> template<typename _RealT, size_t __digits,
> - _Uniform_random_bit_generator _Urbg>
> + typename _Urbg>
> _RealT
> generate_canonical(_Urbg& __urng)
> {
> +#ifdef __glibcxx_concepts
> + static_assert(uniform_random_bit_generator<_Urbg>);
> +#endif
> static_assert(__is_rand_dist_float_v<_RealT>,
> "template argument must be floating point");
> static_assert(__digits != 0 && _Urbg::max() > _Urbg::min(),
> --
> 2.52.0
>
