From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 7857) id E39313861823; Fri, 25 Jun 2021 15:35:15 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org E39313861823 MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Content-Type: text/plain; charset="utf-8" From: Matthias Kretz To: gcc-cvs@gcc.gnu.org, libstdc++-cvs@gcc.gnu.org Subject: [gcc r12-1812] libstdc++: Make use of __builtin_bit_cast for simd X-Act-Checkin: gcc X-Git-Author: Matthias Kretz X-Git-Refname: refs/heads/master X-Git-Oldrev: 3c3474679419707710a899a03151cbe79a91c360 X-Git-Newrev: 74ebd1297e9cfa9f7d05bfcac5510d4968cc6ba8 Message-Id: <20210625153515.E39313861823@sourceware.org> Date: Fri, 25 Jun 2021 15:35:15 +0000 (GMT) X-BeenThere: libstdc++-cvs@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Libstdc++-cvs mailing list List-Unsubscribe: , List-Archive: List-Help: List-Subscribe: , X-List-Received-Date: Fri, 25 Jun 2021 15:35:16 -0000 https://gcc.gnu.org/g:74ebd1297e9cfa9f7d05bfcac5510d4968cc6ba8 commit r12-1812-g74ebd1297e9cfa9f7d05bfcac5510d4968cc6ba8 Author: Matthias Kretz Date: Thu Jan 28 21:04:03 2021 +0100 libstdc++: Make use of __builtin_bit_cast for simd The __bit_cast function was a hack to achieve what __builtin_bit_cast can do, therefore use __builtin_bit_cast if possible. However, __builtin_bit_cast cannot be used to cast from/to fixed_size_simd, since it isn't trivially copyable (in the language sense — in principle it is). Therefore add __proposed::simd_bit_cast to enable the use case required in the test framework. Signed-off-by: Matthias Kretz libstdc++-v3/ChangeLog: * include/experimental/bits/simd.h (__bit_cast): Implement via __builtin_bit_cast #if available. (__proposed::simd_bit_cast): Add overloads for simd and simd_mask, which use __builtin_bit_cast (or __bit_cast #if not available), which return an object of the requested type with the same bits as the argument. * include/experimental/bits/simd_math.h: Use simd_bit_cast instead of __bit_cast to allow casts to fixed_size_simd. (copysign): Remove branch that was only required if __bit_cast cannot be constexpr. * testsuite/experimental/simd/tests/bits/test_values.h: Switch from __bit_cast to __proposed::simd_bit_cast since the former will not cast fixed_size objects anymore. Diff: --- libstdc++-v3/include/experimental/bits/simd.h | 57 +++++++++++++++++++++- libstdc++-v3/include/experimental/bits/simd_math.h | 37 ++++++-------- .../experimental/simd/tests/bits/test_values.h | 8 +-- 3 files changed, 76 insertions(+), 26 deletions(-) diff --git a/libstdc++-v3/include/experimental/bits/simd.h b/libstdc++-v3/include/experimental/bits/simd.h index c396ebd3ee3..ed2645bebc8 100644 --- a/libstdc++-v3/include/experimental/bits/simd.h +++ b/libstdc++-v3/include/experimental/bits/simd.h @@ -1602,7 +1602,9 @@ template _GLIBCXX_SIMD_INTRINSIC constexpr _To __bit_cast(const _From __x) { - // TODO: implement with / replace by __builtin_bit_cast ASAP +#if __has_builtin(__builtin_bit_cast) + return __builtin_bit_cast(_To, __x); +#else static_assert(sizeof(_To) == sizeof(_From)); constexpr bool __to_is_vectorizable = is_arithmetic_v<_To> || is_enum_v<_To>; @@ -1633,6 +1635,7 @@ template reinterpret_cast(&__x), sizeof(_To)); return __r; } +#endif } // }}} @@ -2904,6 +2907,58 @@ template (__x)}; } + +template + _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR + _To + simd_bit_cast(const simd<_Up, _Abi>& __x) + { + using _Tp = typename _To::value_type; + using _ToMember = typename _SimdTraits<_Tp, typename _To::abi_type>::_SimdMember; + using _From = simd<_Up, _Abi>; + using _FromMember = typename _SimdTraits<_Up, _Abi>::_SimdMember; + // with concepts, the following should be constraints + static_assert(sizeof(_To) == sizeof(_From)); + static_assert(is_trivially_copyable_v<_Tp> && is_trivially_copyable_v<_Up>); + static_assert(is_trivially_copyable_v<_ToMember> && is_trivially_copyable_v<_FromMember>); +#if __has_builtin(__builtin_bit_cast) + return {__private_init, __builtin_bit_cast(_ToMember, __data(__x))}; +#else + return {__private_init, __bit_cast<_ToMember>(__data(__x))}; +#endif + } + +template + _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR + _To + simd_bit_cast(const simd_mask<_Up, _Abi>& __x) + { + using _From = simd_mask<_Up, _Abi>; + static_assert(sizeof(_To) == sizeof(_From)); + static_assert(is_trivially_copyable_v<_From>); + // _To can be simd, specifically simd> in which case _To is not trivially + // copyable. + if constexpr (is_simd_v<_To>) + { + using _Tp = typename _To::value_type; + using _ToMember = typename _SimdTraits<_Tp, typename _To::abi_type>::_SimdMember; + static_assert(is_trivially_copyable_v<_ToMember>); +#if __has_builtin(__builtin_bit_cast) + return {__private_init, __builtin_bit_cast(_ToMember, __x)}; +#else + return {__private_init, __bit_cast<_ToMember>(__x)}; +#endif + } + else + { + static_assert(is_trivially_copyable_v<_To>); +#if __has_builtin(__builtin_bit_cast) + return __builtin_bit_cast(_To, __x); +#else + return __bit_cast<_To>(__x); +#endif + } + } } // namespace __proposed // simd_cast {{{2 diff --git a/libstdc++-v3/include/experimental/bits/simd_math.h b/libstdc++-v3/include/experimental/bits/simd_math.h index dd9aaa35967..c81ad7cf08c 100644 --- a/libstdc++-v3/include/experimental/bits/simd_math.h +++ b/libstdc++-v3/include/experimental/bits/simd_math.h @@ -405,10 +405,11 @@ template using _Vp = simd<_Tp, _Abi>; using _Up = make_unsigned_t<__int_for_sizeof_t<_Tp>>; using namespace std::experimental::__float_bitwise_operators; + using namespace std::experimental::__proposed; const _Vp __exponent_mask = __infinity_v<_Tp>; // 0x7f800000 or 0x7ff0000000000000 return static_simd_cast>( - __bit_cast>(__v & __exponent_mask) + simd_bit_cast>(__v & __exponent_mask) >> (__digits_v<_Tp> - 1)); } @@ -697,11 +698,9 @@ template // (inf and NaN are excluded by -ffinite-math-only) const auto __iszero_inf_nan = __x == 0; #else - const auto __as_int - = __bit_cast, _V>>(abs(__x)); - const auto __inf - = __bit_cast, _V>>( - _V(__infinity_v<_Tp>)); + using _Ip = __int_for_sizeof_t<_Tp>; + const auto __as_int = simd_bit_cast>(abs(__x)); + const auto __inf = simd_bit_cast>(_V(__infinity_v<_Tp>)); const auto __iszero_inf_nan = static_simd_cast( __as_int == 0 || __as_int >= __inf); #endif @@ -719,10 +718,10 @@ template where(__value_isnormal.__cvt(), __e) = __exponent_bits; static_assert(sizeof(_IV) == sizeof(__value_isnormal)); const _IV __offset - = (__bit_cast<_IV>(__value_isnormal) & _IV(__exp_adjust)) - | (__bit_cast<_IV>(static_simd_cast<_MaskType>(__exponent_bits == 0) - & static_simd_cast<_MaskType>(__x != 0)) - & _IV(__exp_adjust + __exp_offset)); + = (simd_bit_cast<_IV>(__value_isnormal) & _IV(__exp_adjust)) + | (simd_bit_cast<_IV>(static_simd_cast<_MaskType>(__exponent_bits == 0) + & static_simd_cast<_MaskType>(__x != 0)) + & _IV(__exp_adjust + __exp_offset)); *__exp = simd_cast<_Samesize>(__e - __offset); return __mant; } @@ -786,7 +785,7 @@ template using namespace std::experimental::__proposed; using _IV = rebind_simd_t< conditional_t, _V>; - return (__bit_cast<_IV>(__v) >> (__digits_v<_Tp> - 1)) + return (simd_bit_cast<_IV>(__v) >> (__digits_v<_Tp> - 1)) - (__max_exponent_v<_Tp> - 1); }; _V __r = static_simd_cast<_V>(__exponent(abs_x)); @@ -953,6 +952,7 @@ template // Skylake-AVX512 (not even for SSE and AVX vectors, and really bad for // AVX-512). using namespace __float_bitwise_operators; + using namespace __proposed; _V __absx = abs(__x); // no error _V __absy = abs(__y); // no error _V __hi = max(__absx, __absy); // no error @@ -1000,9 +1000,9 @@ template #ifdef __FAST_MATH__ using _Ip = __int_for_sizeof_t<_Tp>; using _IV = rebind_simd_t<_Ip, _V>; - const auto __as_int = __bit_cast<_IV>(__hi_exp); + const auto __as_int = simd_bit_cast<_IV>(__hi_exp); const _V __scale - = __bit_cast<_V>(2 * __bit_cast<_Ip>(_Tp(1)) - __as_int); + = simd_bit_cast<_V>(2 * simd_bit_cast<_Ip>(_Tp(1)) - __as_int); #else const _V __scale = (__hi_exp ^ __inf) * _Tp(.5); #endif @@ -1090,6 +1090,7 @@ _GLIBCXX_SIMD_CVTING2(hypot) else { using namespace __float_bitwise_operators; + using namespace __proposed; const _V __absx = abs(__x); // no error const _V __absy = abs(__y); // no error const _V __absz = abs(__z); // no error @@ -1169,9 +1170,9 @@ _GLIBCXX_SIMD_CVTING2(hypot) #ifdef __FAST_MATH__ using _Ip = __int_for_sizeof_t<_Tp>; using _IV = rebind_simd_t<_Ip, _V>; - const auto __as_int = __bit_cast<_IV>(__hi_exp); + const auto __as_int = simd_bit_cast<_IV>(__hi_exp); const _V __scale - = __bit_cast<_V>(2 * __bit_cast<_Ip>(_Tp(1)) - __as_int); + = simd_bit_cast<_V>(2 * simd_bit_cast<_Ip>(_Tp(1)) - __as_int); #else const _V __scale = (__hi_exp ^ __inf) * _Tp(.5); #endif @@ -1278,12 +1279,6 @@ template return std::copysign(__x[0], __y[0]); else if constexpr (__is_fixed_size_abi_v<_Abi>) return {__private_init, _Abi::_SimdImpl::_S_copysign(__data(__x), __data(__y))}; - else if constexpr (is_same_v<_Tp, long double> && sizeof(_Tp) == 12) - // Remove this case once __bit_cast is implemented via __builtin_bit_cast. - // It is necessary, because __signmask below cannot be computed at compile - // time. - return simd<_Tp, _Abi>( - [&](auto __i) { return std::copysign(__x[__i], __y[__i]); }); else { using _V = simd<_Tp, _Abi>; diff --git a/libstdc++-v3/testsuite/experimental/simd/tests/bits/test_values.h b/libstdc++-v3/testsuite/experimental/simd/tests/bits/test_values.h index b69bd0b704d..67aa870659b 100644 --- a/libstdc++-v3/testsuite/experimental/simd/tests/bits/test_values.h +++ b/libstdc++-v3/testsuite/experimental/simd/tests/bits/test_values.h @@ -221,11 +221,11 @@ template if constexpr (sizeof(T) <= sizeof(double)) { using I = rebind_simd_t<__int_for_sizeof_t, V>; - const I abs_x = __bit_cast(abs(x)); - const I min = __bit_cast(V(std::__norm_min_v)); - const I max = __bit_cast(V(std::__finite_max_v)); + const I abs_x = simd_bit_cast(abs(x)); + const I min = simd_bit_cast(V(std::__norm_min_v)); + const I max = simd_bit_cast(V(std::__finite_max_v)); return static_simd_cast( - __bit_cast(x) == 0 || (abs_x >= min && abs_x <= max)); + simd_bit_cast(x) == 0 || (abs_x >= min && abs_x <= max)); } else {