On Fri, 10 Apr 2026 at 00:13, Alexandre Oliva <[email protected]> wrote:
>
>
> Various simd_x86 functions that handle double need to be adjusted to
> match 64-bit long double as well.
I don't think this is a good use of time.
Matthias has suggested dropping long double support from
experimental::simd (which would reduce maintenance costs and testsuite
time). Complicating the code for -mlong-double-64 seems to be going in
the opposite direction.
I would prefer to just add a static_assert to reject -mlong-double-64
and say "don't do that, this isn't supported" (and XFAIL the relevant
tests). It's an experimental implementation which has been superseded
by std::simd in C++26, we have no obligation to make it work on all
targets for all configurations. If it's not compatible with
-mlong-double-64 targets, then so be it.
>
> This patch deals with _S_*_of, to test the concept.
>
> There are many occurrences of double in simd_x86.h file, and we
> probably have to adjust them all. But this is more than enough to get
> pr109261_constexpr_simd.cc to compile with -mlong-double-64.
>
> We probably need a better abstraction.
>
>
> Regstrapped on x86_64-linux-gnu. With this change on top of the
> previous one, experimental/simd/pr109261_constexpr_simd.cc compiles
> successfully with -mlong-double-64. This works, but I realize it's ugly
> and incomplete, so take this as a conversation starter about preferences
> on how to spell out the long-double-64 equivalence to double all over
> simd_x86.h. Unless you like it, in which case, ok to install? :-)
>
>
> for libstdc++-v3/ChangeLog
>
> PR libstdc++/124657
> * include/experimental/bits/simd_x86.h (_S_all_of): Handle
> long double like double.
> (_S_any_of, _S_none_of, _S_some_of): Likewise.
> ---
> libstdc++-v3/include/experimental/bits/simd_x86.h | 16 ++++++++++++----
> 1 file changed, 12 insertions(+), 4 deletions(-)
>
> diff --git a/libstdc++-v3/include/experimental/bits/simd_x86.h
> b/libstdc++-v3/include/experimental/bits/simd_x86.h
> index 74c7a61998e0d..400b56e40b33f 100644
> --- a/libstdc++-v3/include/experimental/bits/simd_x86.h
> +++ b/libstdc++-v3/include/experimental/bits/simd_x86.h
> @@ -5024,7 +5024,9 @@ template <typename _Abi, typename>
> else if constexpr (is_same_v<_Tp, float>)
> return (_mm_movemask_ps(__a) & ((1 << _Np) - 1))
> == (1 << _Np) - 1;
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (is_same_v<_Tp, double>
> + || (sizeof (long double) == sizeof (double)
> + && is_same_v<_Tp, long double>))
> return (_mm_movemask_pd(__a) & ((1 << _Np) - 1))
> == (1 << _Np) - 1;
> else
> @@ -5086,7 +5088,9 @@ template <typename _Abi, typename>
> }
> else if constexpr (is_same_v<_Tp, float>)
> return (_mm_movemask_ps(__a) & ((1 << _Np) - 1)) != 0;
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (is_same_v<_Tp, double>
> + || (sizeof (long double) == sizeof (double)
> + && is_same_v<_Tp, long double>))
> return (_mm_movemask_pd(__a) & ((1 << _Np) - 1)) != 0;
> else
> return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) -
> 1))
> @@ -5122,7 +5126,9 @@ template <typename _Abi, typename>
> }
> else if constexpr (is_same_v<_Tp, float>)
> return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (is_same_v<_Tp, double>
> + || (sizeof (long double) == sizeof (double)
> + && is_same_v<_Tp, long double>))
> return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
> else
> return (__movemask(__a) & int((1ull << (_Np * sizeof(_Tp))) -
> 1))
> @@ -5156,7 +5162,9 @@ template <typename _Abi, typename>
> const auto __tmp = _mm_movemask_ps(__a) & __allbits;
> return __tmp > 0 && __tmp < __allbits;
> }
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (is_same_v<_Tp, double>
> + || (sizeof (long double) == sizeof (double)
> + && is_same_v<_Tp, long double>))
> {
> constexpr int __allbits = (1 << _Np) - 1;
> const auto __tmp = _mm_movemask_pd(__a) & __allbits;
>
> --
> 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!
>
