[PATCH] libstdc++: simd: use preprocessor to test for 64-bit long double [PR124657]

[Jonathan Wakely]

On Wed, 22 Apr 2026 at 07:28, Alexandre Oliva <oliva@adacore.com> wrote:
>
> On Apr 21, 2026, Jonathan Wakely <jwakely@redhat.com> wrote:
>
> > #if  __LDBL_MANT_DIG__ == __DBL_MANT_DIG__
>
> I tried to apply the same technique to simd.h, on top of
> https://gcc.gnu.org/pipermail/gcc-patches/2026-April/712616.html, but it
> doesn't save any instantiations, neither on x86 nor on ppc.  It seems to
> make things generally uglier, but it might offer a little compile-time
> benefit on both, so I left it as a separate patch...  WDYT?
>
> Also regstrapped on x86_64-linux-gnu and powerpc64le-linux-gnu.  I also
> tested pr109261_constexpr_simd.cc with -mlong-double-64 on both.  Ok for
> stage1?
>
>
> When long double is as wide as double on x86 or powerpc
> (-mlong-double-64), we remap it to double in __intrinsic_type.
>
> Use the preprocessor to compare __L?DBL_MANT_DIG instead of
> floating-type type sizes on x86 SSE and PowerPC AltiVec.
>
>
> for  libstdc++-v3/ChangeLog
>
>         PR libstdc++/124657
>         * include/experimental/bits/simd.h [_GLIBCXX_SIMD_HAVE_SSE]
>         (__intrinsic_type::_S_is_ldouble): New.  Use all over.
>         (__intrinsic_type::type): Use preprocessor tests.
>         [__ALTIVEC__] (__intrinsic_type): Likewise.
> ---
>  libstdc++-v3/include/experimental/bits/simd.h |   25 ++++++++++++++++++++-----
>  1 file changed, 20 insertions(+), 5 deletions(-)
>
> diff --git a/libstdc++-v3/include/experimental/bits/simd.h b/libstdc++-v3/include/experimental/bits/simd.h
> index 8c30c8af26881..177d3598b54b5 100644
> --- a/libstdc++-v3/include/experimental/bits/simd.h
> +++ b/libstdc++-v3/include/experimental/bits/simd.h
> @@ -2469,16 +2469,24 @@ template <>
>  template <typename _Tp, size_t _Bytes>
>    struct __intrinsic_type<_Tp, _Bytes, enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <= 64>>
>    {
> +    static constexpr bool _S_is_ldouble = is_same_v<_Tp, long double>;
> +
>      // allow _Tp == long double with -mlong-double-64
> -    static_assert(!(is_same_v<_Tp, long double>
> -                   && sizeof(long double) > sizeof(double)),
> +#if __LDBL_MANT_DIG != __DBL_MANT_DIG

I think both of these macros need to have __ at the end, so this
condition is always expanded to 0 != 0, so you're effectively removing
the static_assert that was previously present.

> +    static_assert(!_S_is_ldouble,
>                   "no __intrinsic_type support for long double on x86");
> +#endif
>
>      static constexpr size_t _S_VBytes = _Bytes <= 16 ? 16 : _Bytes <= 32 ? 32 : 64;
>
>      using type [[__gnu__::__vector_size__(_S_VBytes)]]
>        = conditional_t<is_integral_v<_Tp>, long long int,
> -                     conditional_t<is_same_v<_Tp, long double>, double, _Tp> >;
> +#if __LDBL_MANT_DIG != __DBL_MANT_DIG

Ditto

> +                     _Tp
> +#else
> +                     conditional_t<_S_is_ldouble, double, _Tp>
> +#endif
> +                     >;
>    };
>  #endif // _GLIBCXX_SIMD_HAVE_SSE
>
> @@ -2582,12 +2590,17 @@ template <typename _Tp, size_t _Bytes>
>      static constexpr bool _S_is_ldouble = is_same_v<_Tp, long double>;
>
>      // allow _Tp == long double with -mlong-double-64
> -    static_assert(!(_S_is_ldouble && sizeof(long double) > sizeof(double)),
> +#if __LDBL_MANT_DIG != __DBL_MANT_DIG
> +    static_assert(!_S_is_ldouble,
>                   "no __intrinsic_type support for 128-bit floating point on PowerPC");
> +#endif
>
>  #ifndef __VSX__
>      static_assert(!(is_same_v<_Tp, double>
> -                   || (_S_is_ldouble && sizeof(long double) == sizeof(double))),
> +#if __LDBL_MANT_DIG == __DBL_MANT_DIG
> +                   || _S_is_ldouble
> +#endif
> +                   ),
>                   "no __intrinsic_type support for 64-bit floating point on PowerPC w/o VSX");
>  #endif
>
> @@ -2595,9 +2608,11 @@ template <typename _Tp, size_t _Bytes>
>      {
>        if constexpr (is_floating_point_v<_Tp>)
>         {
> +#if __LDBL_MANT_DIG == __DBL_MANT_DIG
>           if constexpr (_S_is_ldouble)
>             return double {};
>           else
> +#endif
>             return _Tp {};
>         }
>        else if constexpr (is_signed_v<_Tp>)
>
>
> --
> Alexandre Oliva, happy hacker            https://blog.lx.oliva.nom.br/
> Free Software Activist     FSFLA co-founder     GNU Toolchain Engineer
> More tolerance and less prejudice are key for inclusion and diversity.
> Excluding neuro-others for not behaving ""normal"" is *not* inclusive!
>