[PATCH v2] libstdc++: simd: x86: accept 64-bit long double as double [PR124657]
Jonathan Wakely
jwakely@redhat.com
Tue May 5 18:18:27 GMT 2026
On Wed, 22 Apr 2026 at 07:20, Alexandre Oliva <oliva@adacore.com> wrote:
>
> On Apr 21, 2026, Matthias Kretz <MatthiasKretz@gmx.net> wrote:
>
> >> This would be ever so slightly more efficient as:
>
> >> because it wouldn't instantiate the is_same_v variable templates
> >> unless required, and wouldn't care about long double for the common
> >> case of 80-bit long double.
> >>
> >> But the patch seems reasonable to me. Matthias?
>
> > Yes. It looks like a clean solution to me. I like Jonathan's suggestion.
> > std::experimental::simd is too expensive to compile already.
>
> Here's a patch that implements the suggested change, and that, along
> with https://gcc.gnu.org/pipermail/gcc-patches/2026-April/712616.html
> (Ping?) enables pr109261_constexpr_simd.cc to succeed with
> -mlong-double-64 on x86_64 -msse. Regstrapped on x86_64-linux-gnu and
> powerpc64le-linux-gnu. Ok for stage1?
>
> Jonathan, should I credit you for the preprocessor micro-optimization?
> I didn't quite follow it to the letter, because I thought it looked
> cleaner this way, and IIUC it would not bring any notable inefficiency.
>
>
> Various simd_x86 functions that handle double need to be adjusted to
> match 64-bit long double as well.
>
> Introduce __is_x86_ps<_Tp>() and __is_x86_pd<_Tp>() and use them
> instead of is_same_v<_Tp, float> and is_same_v<_Tp, double>,
> respectively.
>
>
> for libstdc++-v3/ChangeLog
>
> PR libstdc++/124657
> * include/experimental/bits/simd_x86.h
> (__is_x86_ps<_Tp>): New. Replace is_same_v<_Tp, float> with it.
> (__is_x86_pd<_Tp>): New. Replace is_same_v<_Tp, double> with it.
> ---
> libstdc++-v3/include/experimental/bits/simd_x86.h | 82 +++++++++++++--------
> 1 file changed, 50 insertions(+), 32 deletions(-)
>
> diff --git a/libstdc++-v3/include/experimental/bits/simd_x86.h b/libstdc++-v3/include/experimental/bits/simd_x86.h
> index 74c7a61998e0d..5d4ed5ad58c8e 100644
> --- a/libstdc++-v3/include/experimental/bits/simd_x86.h
> +++ b/libstdc++-v3/include/experimental/bits/simd_x86.h
> @@ -415,13 +415,31 @@ template <size_t _Np, typename _Tp, typename _Kp>
> #endif
>
> // ISA & type detection {{{
> +template <typename _Tp>
> + constexpr bool
> + __is_x86_ps()
> + {
> + return is_same_v<_Tp, float>;
> + }
> +
> +template <typename _Tp>
> + constexpr bool
> + __is_x86_pd()
> + {
> +#if __LDBL_MANT_DIG == __DBL_MANT_DIG
Missing __ on the macros, so this is 0 == 0, and we're instantiating
is_same_v<T, long double> for every call to this function, even in the
usual x86-linux case where long double is not 64-bit.
I'll push a fix after the tests finish.
> + if constexpr (is_same_v<_Tp, long double>)
> + return true;
> +#endif
> + return is_same_v<_Tp, double>;
> + }
> +
> template <typename _Tp, size_t _Np>
> constexpr bool
> __is_sse_ps()
> {
> return __have_sse
> - && is_same_v<_Tp,
> - float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
> + && __is_x86_ps<_Tp>()
> + && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
> }
>
> template <typename _Tp, size_t _Np>
> @@ -429,8 +447,8 @@ template <typename _Tp, size_t _Np>
> __is_sse_pd()
> {
> return __have_sse2
> - && is_same_v<_Tp,
> - double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
> + && __is_x86_pd<_Tp>()
> + && sizeof(__intrinsic_type_t<_Tp, _Np>) == 16;
> }
>
> template <typename _Tp, size_t _Np>
> @@ -438,8 +456,8 @@ template <typename _Tp, size_t _Np>
> __is_avx_ps()
> {
> return __have_avx
> - && is_same_v<_Tp,
> - float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
> + && __is_x86_ps<_Tp>()
> + && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
> }
>
> template <typename _Tp, size_t _Np>
> @@ -447,8 +465,8 @@ template <typename _Tp, size_t _Np>
> __is_avx_pd()
> {
> return __have_avx
> - && is_same_v<_Tp,
> - double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
> + && __is_x86_pd<_Tp>()
> + && sizeof(__intrinsic_type_t<_Tp, _Np>) == 32;
> }
>
> template <typename _Tp, size_t _Np>
> @@ -456,8 +474,8 @@ template <typename _Tp, size_t _Np>
> __is_avx512_ps()
> {
> return __have_avx512f
> - && is_same_v<_Tp,
> - float> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
> + && __is_x86_ps<_Tp>()
> + && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
> }
>
> template <typename _Tp, size_t _Np>
> @@ -465,8 +483,8 @@ template <typename _Tp, size_t _Np>
> __is_avx512_pd()
> {
> return __have_avx512f
> - && is_same_v<_Tp,
> - double> && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
> + && __is_x86_pd<_Tp>()
> + && sizeof(__intrinsic_type_t<_Tp, _Np>) == 64;
> }
>
> // }}}
> @@ -2397,9 +2415,9 @@ template <typename _Abi, typename>
> [[maybe_unused]] const auto __yi = __to_intrin(__y);
> if constexpr (sizeof(__xi) == 64)
> {
> - if constexpr (is_same_v<_Tp, float>)
> + if constexpr (__is_x86_ps<_Tp> ())
> return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
> else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
> return _mm512_mask_cmplt_epi8_mask(__k1, __xi, __yi);
> @@ -2422,9 +2440,9 @@ template <typename _Abi, typename>
> }
> else if constexpr (sizeof(__xi) == 32)
> {
> - if constexpr (is_same_v<_Tp, float>)
> + if constexpr (__is_x86_ps<_Tp> ())
> return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
> else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
> return _mm256_mask_cmplt_epi8_mask(__k1, __xi, __yi);
> @@ -2447,9 +2465,9 @@ template <typename _Abi, typename>
> }
> else if constexpr (sizeof(__xi) == 16)
> {
> - if constexpr (is_same_v<_Tp, float>)
> + if constexpr (__is_x86_ps<_Tp> ())
> return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
> else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
> return _mm_mask_cmplt_epi8_mask(__k1, __xi, __yi);
> @@ -2505,9 +2523,9 @@ template <typename _Abi, typename>
> [[maybe_unused]] const auto __yi = __to_intrin(__y);
> if constexpr (sizeof(__xi) == 64)
> {
> - if constexpr (is_same_v<_Tp, float>)
> + if constexpr (__is_x86_ps<_Tp> ())
> return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
> else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
> return _mm512_mask_cmple_epi8_mask(__k1, __xi, __yi);
> @@ -2530,9 +2548,9 @@ template <typename _Abi, typename>
> }
> else if constexpr (sizeof(__xi) == 32)
> {
> - if constexpr (is_same_v<_Tp, float>)
> + if constexpr (__is_x86_ps<_Tp> ())
> return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
> else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
> return _mm256_mask_cmple_epi8_mask(__k1, __xi, __yi);
> @@ -2555,9 +2573,9 @@ template <typename _Abi, typename>
> }
> else if constexpr (sizeof(__xi) == 16)
> {
> - if constexpr (is_same_v<_Tp, float>)
> + if constexpr (__is_x86_ps<_Tp> ())
> return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
> else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) == 1)
> return _mm_mask_cmple_epi8_mask(__k1, __xi, __yi);
> @@ -5021,10 +5039,10 @@ template <typename _Abi, typename>
> = _Abi::template _S_implicit_mask_intrin<_Tp>();
> return 0 != __testc(__a, __b);
> }
> - else if constexpr (is_same_v<_Tp, float>)
> + else if constexpr (__is_x86_ps<_Tp> ())
> return (_mm_movemask_ps(__a) & ((1 << _Np) - 1))
> == (1 << _Np) - 1;
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return (_mm_movemask_pd(__a) & ((1 << _Np) - 1))
> == (1 << _Np) - 1;
> else
> @@ -5084,9 +5102,9 @@ template <typename _Abi, typename>
> else
> return 0 == __testz(__a, __a);
> }
> - else if constexpr (is_same_v<_Tp, float>)
> + else if constexpr (__is_x86_ps<_Tp> ())
> return (_mm_movemask_ps(__a) & ((1 << _Np) - 1)) != 0;
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return (_mm_movemask_pd(__a) & ((1 << _Np) - 1)) != 0;
> else
> return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
> @@ -5120,9 +5138,9 @@ template <typename _Abi, typename>
> else
> return 0 != __testz(__a, __a);
> }
> - else if constexpr (is_same_v<_Tp, float>)
> + else if constexpr (__is_x86_ps<_Tp> ())
> return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
> - else if constexpr (is_same_v<_Tp, double>)
> + else if constexpr (__is_x86_pd<_Tp> ())
> return (__movemask(__a) & ((1 << _Np) - 1)) == 0;
> else
> return (__movemask(__a) & int((1ull << (_Np * sizeof(_Tp))) - 1))
> @@ -5150,13 +5168,13 @@ template <typename _Abi, typename>
> = _Abi::template _S_implicit_mask_intrin<_Tp>();
> return 0 != __testnzc(__a, __b);
> }
> - else if constexpr (is_same_v<_Tp, float>)
> + else if constexpr (__is_x86_ps<_Tp> ())
> {
> constexpr int __allbits = (1 << _Np) - 1;
> 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_x86_pd<_Tp> ())
> {
> 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!
>
More information about the Libstdc++
mailing list