Re: [PATCH 04/11 v3] libstdc++: Make use of __builtin_bit_cast

[Matthias Kretz <m.kretz@gsi.de>]

[-- Attachment #1: Type: text/plain, Size: 2303 bytes --]

For -ffast-math there was a missing using namespace __proposed left. The 
attached patch resolves the issue.

From: Matthias Kretz <m.kretz@gsi.de>

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 <m.kretz@gsi.de>

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.
---
 libstdc++-v3/include/experimental/bits/simd.h | 57 ++++++++++++++++++-
 .../include/experimental/bits/simd_math.h     | 37 ++++++------
 .../simd/tests/bits/test_values.h             |  8 +--
 3 files changed, 76 insertions(+), 26 deletions(-)


-- 
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Centre for Heavy Ion Research               https://gsi.de
 std::experimental::simd              https://github.com/VcDevel/std-simd
──────────────────────────────────────────────────────────────────────────

[-- Attachment #2: 0001-libstdc-Make-use-of-__builtin_bit_cast.patch --]
[-- Type: text/x-patch, Size: 9051 bytes --]

diff --git a/libstdc++-v3/include/experimental/bits/simd.h b/libstdc++-v3/include/experimental/bits/simd.h
index 163f1b574e2..852d0b62012 100644
--- a/libstdc++-v3/include/experimental/bits/simd.h
+++ b/libstdc++-v3/include/experimental/bits/simd.h
@@ -1598,7 +1598,9 @@ template <typename _To, typename _From>
   _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>;
@@ -1629,6 +1631,7 @@ template <typename _To, typename _From>
 			 reinterpret_cast<const char*>(&__x), sizeof(_To));
 	return __r;
       }
+#endif
   }
 
 // }}}
@@ -2900,6 +2903,58 @@ template <typename _Tp, typename _Up, typename _Ap,
     return {__private_init, _RM::abi_type::_MaskImpl::template _S_convert<
 			      typename _RM::simd_type::value_type>(__x)};
   }
+
+template <typename _To, typename _Up, typename _Abi>
+  _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 <typename _To, typename _Up, typename _Abi>
+  _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<T, A>, specifically simd<T, fixed_size<N>> 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 d954e761eee..ef2bdc641b8 100644
--- a/libstdc++-v3/include/experimental/bits/simd_math.h
+++ b/libstdc++-v3/include/experimental/bits/simd_math.h
@@ -405,10 +405,11 @@ template <typename _Tp, typename _Abi>
     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<rebind_simd_t<int, _Vp>>(
-      __bit_cast<rebind_simd_t<_Up, _Vp>>(__v & __exponent_mask)
+	     simd_bit_cast<rebind_simd_t<_Up, _Vp>>(__v & __exponent_mask)
       >> (__digits_v<_Tp> - 1));
   }
 
@@ -700,11 +701,9 @@ template <typename _Tp, typename _Abi>
 	// (inf and NaN are excluded by -ffinite-math-only)
 	const auto __iszero_inf_nan = __x == 0;
 #else
-	const auto __as_int
-	  = __bit_cast<rebind_simd_t<__int_for_sizeof_t<_Tp>, _V>>(abs(__x));
-	const auto __inf
-	  = __bit_cast<rebind_simd_t<__int_for_sizeof_t<_Tp>, _V>>(
-	    _V(__infinity_v<_Tp>));
+	using _Ip = __int_for_sizeof_t<_Tp>;
+	const auto __as_int = simd_bit_cast<rebind_simd_t<_Ip, _V>>(abs(__x));
+	const auto __inf = simd_bit_cast<rebind_simd_t<_Ip, _V>>(_V(__infinity_v<_Tp>));
 	const auto __iszero_inf_nan = static_simd_cast<typename _V::mask_type>(
 	  __as_int == 0 || __as_int >= __inf);
 #endif
@@ -722,10 +721,10 @@ template <typename _Tp, typename _Abi>
 	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<int, _V>>(__e - __offset);
 	return __mant;
       }
@@ -796,7 +795,7 @@ template <typename _Tp, typename _Abi>
 	  using namespace std::experimental::__proposed;
 	  using _IV = rebind_simd_t<
 	    conditional_t<sizeof(_Tp) == sizeof(_LLong), _LLong, int>, _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));
@@ -981,6 +980,7 @@ template <typename _VV>
 	// 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
@@ -1028,9 +1028,9 @@ template <typename _VV>
 #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
@@ -1118,6 +1118,7 @@ _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
@@ -1197,9 +1198,9 @@ _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
@@ -1306,12 +1307,6 @@ template <typename _Tp, typename _Abi>
       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 <class V>
     if constexpr (sizeof(T) <= sizeof(double))
       {
 	using I = rebind_simd_t<__int_for_sizeof_t<T>, V>;
-	const I abs_x = __bit_cast<I>(abs(x));
-	const I min = __bit_cast<I>(V(std::__norm_min_v<T>));
-	const I max = __bit_cast<I>(V(std::__finite_max_v<T>));
+	const I abs_x = simd_bit_cast<I>(abs(x));
+	const I min = simd_bit_cast<I>(V(std::__norm_min_v<T>));
+	const I max = simd_bit_cast<I>(V(std::__finite_max_v<T>));
 	return static_simd_cast<typename V::mask_type>(
-		 __bit_cast<I>(x) == 0 || (abs_x >= min && abs_x <= max));
+		 simd_bit_cast<I>(x) == 0 || (abs_x >= min && abs_x <= max));
       }
     else
       {