[libcxx] Implement <simd> ABI for Clang/GCC vector extension, constructors, copy_from and copy_to.

[timshen]

Summary

This patch adds a new macro _LIBCPP_HAS_NO_VECTOR_EXTENSION for detecting
whether a vector extension (\_\_attribute\_\_((vector_size(num_bytes)))) is
available.

On the top of that, this patch implements the following API:

• all constructors
• operator[]
• copy_from
• copy_to

It also defines simd_abi::native to use vector extension, if available.
In GCC and Clang, certain values with vector extension are passed by registers,
instead of memory.

Based on D41148.

[timshen]

Update description.

[timshen]

Fix reference's access control.

[timshen]

Fix return type of reference::opreator--.

[timshen]

s/_LIBCPP_HAS_VECTOR_EXTENSION/_LIBCPP_HAS_NO_VECTOR_EXTENSION/

[timshen]

Rebase. Also modify the simd reference implementation to separate the stored type and value_type (which could be bool later).

[timshen]

Rebased.

[timshen]

Rebase.

[timshen]

Chatted with Marshall a bit, we thought that it's bad for toolchain portability to support a C++17 library in C++11, especially with modifications w.r.t std::plus<>.

Remove the backport of std::integer_sequence to C++11.

[mclow.lists]

Looks good; a bunch of minor things.
Remember to target c++17.
You need to indulge your OCD when writing the tests.

libcxx/include/__config

1278

We ususally organize these into "chunks", one per compiler.
The "Clang chunk" starts on 159, and goes to 483.
The "GCC chunk" starts on 484 and goes to 575.

It would match the organization better if you put this line:

#define _LIBCPP_HAS_NO_VECTOR_EXTENSION

in the file twice - once in the Clang chunk and once in the GCC chunk.

See _LIBCPP_HAS_UNIQUE_OBJECT_REPRESENTATIONS for an example of this.

libcxx/include/experimental/simd

660

When we define user-visible macros in libc++, we prepend them with _LIBCPP_ to avoid collisions.
Accordingly, _SPECIALIZE_VEC_EXT should be named _LIBCPP_SPECIALIZE_VEC_EXT

828

We got rid of this for N4755 - but I think you're fixing this in D44663.

857–863

I have a weak preference for putting entire statements inside #ifdef blocks where possible.

1236

Is this TODO still necessary?

libcxx/test/std/experimental/simd/simd.cons/broadcast.pass.cpp

76

Again, do we care about other types than native_simd here?

libcxx/test/std/experimental/simd/simd.cons/default.pass.cpp

22

Do we need any other ctors tested here? fixed_size_simd<int32_t, 4> for example?
Are there any post-conditions on the object created?

calling size() for example?

libcxx/test/std/experimental/simd/simd.mem/load.pass.cpp

23

I am unsure about the desirability of testing this by hoisting it all into the global namespace.

for the pmr stuff, we did:

namespace ex = std::experimental::pmr;

and then referred to things as ex::XXX. That way, you don't have to do all the typing, but we don't hide any ADL issues.

Obviously, this applies to all the tests.

libcxx/test/std/experimental/simd/simd.mem/store.pass.cpp

11

more unsupported :-)

52

Don't need this anymore.

[timshen]

Update based on the comments.

libcxx/include/experimental/simd

660

Well _SPECIALIZE_VEC_EXT and _SPECIALIZE_VEC_EXT_32 are not user-visible, as they are (1) '#undef'ed after all uses, and (2) they start with underscore, so don't collide with existing user-defined macros.

1236

I think so, as some operations are still not implemented, for example operator++().

libcxx/test/std/experimental/simd/simd.cons/default.pass.cpp

22

Yes. Added more.

The test space is large and it's hard to test everything. I went "every line of the implementation should be covered" standard with myself, and only had native_simd's default ctor. fixed_size_simd<>'s ctor is exactly implemented by the same line, so I didn't bother.

But now the attention is already raised, it doesn't hurt to add more tests.

22

Yes, size() is the post-condition. There is not much beyond that - the elements are not even value initialized.

[timshen]

I'm not going to rebase all the succeeding patches immediately onto this one, as it is painful and spamming emails. Rather, I'll only rebase the next patch in the line. So if you review more than one patch ahead (as you already did), you may see some stale patch context.

[timshen]

libcxx/include/experimental/simd

1236

For test coverage I turned many of the tests into templates, in the hope not to duplicate the text and increase the coverage. I'm not sure if you like the idea.

[mclow.lists]

Getting really close.

libcxx/include/experimental/simd

637

Can these (__get and __set) be noexcept? Obviously, it depends on _Tp.

660

I agree, but starting them with _LIBCPP means that people (like me) know that they come from libc++, and not (say) clang, or whatever C library we're sitting on top of.

745

If _Vp and _Tp have to name the same type, why have two of them?
What is the difference, and when would you use each one internally?

__simd_reference(__simd_storage<_Tp, _Abi>* __ptr, size_t __index);

vs.

__simd_reference operator=(_Vp __value) &&;

[timshen]

Update based on comments.

[timshen]

libcxx/include/experimental/simd

637

Currently _Tp is always a subset of arithmetic type, so it's safe to say noexcept. Added them, as they may help with -O0 performance.

745

If _Vp and _Tp have to name the same type, why have two of them?

Currently there is no difference. It's going to be different when simd_mask is implemented (_Vp can be bool, but _Tp may not be).

What is the difference, and when would you use each one internally?

This is not an internal class. This is part of the interface. IIRC since P0214R9 opreator=() is changed to something. In a later patch operator=() should be implemented as R9 specified.

[mclow.lists]

LGTM. Thanks!

libcxx/include/__config

 #if defined(_LIBCPP_ABI_MICROSOFT) && !defined(_LIBCPP_BUILDING_LIBRARY)
 # if defined(_DLL)
 #   pragma(lib, "c++.lib")
 # else
 #   pragma(lib, "libc++.lib")
 # endif
 #endif // defined(_LIBCPP_ABI_MICROSOFT) && !defined(_LIBCPP_BUILDING_LIBRARY)
 
+#if defined(_LIBCPP_COMPILER_CLANG) || defined(_LIBCPP_COMPILER_GCC)
+#define _LIBCPP_HAS_VECTOR_EXTENSION
+#endif // defined(_LIBCPP_COMPILER_CLANG) || defined(_LIBCPP_COMPILER_GCC)
+
 #endif // __cplusplus
 
 #endif // _LIBCPP_CONFIG

libcxx/include/experimental/__config

#define _LIBCPP_END_NAMESPACE_EXPERIMENTAL_SIMD \
     } _LIBCPP_END_NAMESPACE_EXPERIMENTAL
 
 #define _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL_SIMD_ABI \
     _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL_SIMD namespace simd_abi {
 
 #define _LIBCPP_END_NAMESPACE_EXPERIMENTAL_SIMD_ABI \
     } _LIBCPP_END_NAMESPACE_EXPERIMENTAL_SIMD
 
+// TODO: support more targets
+#if defined(__AVX__)
+#define _LIBCPP_NATIVE_SIMD_WIDTH_IN_BYTES 32
+#else
+#define _LIBCPP_NATIVE_SIMD_WIDTH_IN_BYTES 16
+#endif
+
 #endif

libcxx/include/experimental/simd

 };
 
 } // parallelism_v2
 } // std::experimental
 
 */
 
 #include <experimental/__config>
+#include <algorithm>
 #include <array>
 #include <cstddef>
 #include <functional>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #pragma GCC system_header
 #endif
 
 _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL_SIMD
 
 enum class _StorageKind {
   _Scalar,
   _Array,
+  _VecExt,
 };
 
 template <_StorageKind __kind, int _Np>
 struct __simd_abi {};
 
 template <class _Tp, class _Abi>
-struct __simd_storage_traits {};
+class __simd_storage {};
 
 template <class _Tp, int __num_element>
-struct __simd_storage_traits<_Tp,
-                             __simd_abi<_StorageKind::_Array, __num_element>> {
-  using type = std::array<_Tp, __num_element>;
+class __simd_storage<_Tp, __simd_abi<_StorageKind::_Array, __num_element>> {
+  std::array<_Tp, __num_element> __storage_;
+
+public:
+  using reference = _Tp&;
+
+  _Tp __get(size_t __index) const { return __storage_[__index]; };
+  reference __ref(size_t __index) { return __storage_[__index]; };
 };
 
 template <class _Tp>
-struct __simd_storage_traits<_Tp, __simd_abi<_StorageKind::_Scalar, 1>> {
-  using type = _Tp;
+class __simd_storage<_Tp, __simd_abi<_StorageKind::_Scalar, 1>> {
+  _Tp __storage_;
+
+public:
+  using reference = _Tp&;
+
+  _Tp __get(size_t __index) const { return (&__storage_)[__index]; };
+  reference __ref(size_t __index) { return (&__storage_)[__index]; };
+};
+
+#ifdef _LIBCPP_HAS_VECTOR_EXTENSION
+
+constexpr size_t __floor_pow_of_2(size_t __val) {
+  return ((__val - 1) & __val) == 0 ? __val
+                                    : __floor_pow_of_2((__val - 1) & __val);
+}
+
+constexpr size_t __ceil_pow_of_2(size_t __val) {
+  return __val == 1 ? 1 : __floor_pow_of_2(__val - 1) << 1;
+}
+
+template <class _Tp, size_t __bytes>
+struct __vec_ext_traits {};
+
+#define _SPECIALIZE_VEC_EXT(_TYPE, _NUM_ELEMENT)                               \
+  template <>                                                                  \
+  struct __vec_ext_traits<_TYPE, sizeof(_TYPE) * _NUM_ELEMENT> {               \
+    using type =                                                               \
+        _TYPE __attribute__((vector_size(sizeof(_TYPE) * _NUM_ELEMENT)));      \
+  }
+
+#define _SPECIALIZE_VEC_EXT_32(_TYPE)                                          \
+  _SPECIALIZE_VEC_EXT(_TYPE, 1);                                               \
+  _SPECIALIZE_VEC_EXT(_TYPE, 2);                                               \
+  _SPECIALIZE_VEC_EXT(_TYPE, 4);                                               \
+  _SPECIALIZE_VEC_EXT(_TYPE, 8);                                               \
+  _SPECIALIZE_VEC_EXT(_TYPE, 16);                                              \
+  _SPECIALIZE_VEC_EXT(_TYPE, 32)
+
+_SPECIALIZE_VEC_EXT_32(char);
+_SPECIALIZE_VEC_EXT_32(char16_t);
+_SPECIALIZE_VEC_EXT_32(char32_t);
+_SPECIALIZE_VEC_EXT_32(wchar_t);
+_SPECIALIZE_VEC_EXT_32(signed char);
+_SPECIALIZE_VEC_EXT_32(signed short);
+_SPECIALIZE_VEC_EXT_32(signed int);
+_SPECIALIZE_VEC_EXT_32(signed long);
+_SPECIALIZE_VEC_EXT_32(signed long long);
+_SPECIALIZE_VEC_EXT_32(unsigned char);
+_SPECIALIZE_VEC_EXT_32(unsigned short);
+_SPECIALIZE_VEC_EXT_32(unsigned int);
+_SPECIALIZE_VEC_EXT_32(unsigned long);
+_SPECIALIZE_VEC_EXT_32(unsigned long long);
+_SPECIALIZE_VEC_EXT_32(float);
+_SPECIALIZE_VEC_EXT_32(double);
+_SPECIALIZE_VEC_EXT_32(long double);
+
+#undef _SPECIALIZE_VEC_EXT_32
+#undef _SPECIALIZE_VEC_EXT
+
+template <class _Tp, int __num_element>
+class __simd_storage<_Tp, __simd_abi<_StorageKind::_VecExt, __num_element>> {
+  using _StorageType =
+      typename __vec_ext_traits<_Tp, __ceil_pow_of_2(sizeof(_Tp) *
+                                                     __num_element)>::type;
+
+  _StorageType __storage_;
+
+public:
+  class reference {
+    template <class, class>
+    friend class __simd_storage;
+    template <class, class>
+    friend class simd;
+
+    __simd_storage* __ptr_;
+    size_t __index_;
+
+    reference(__simd_storage* __ptr, size_t __index)
+        : __ptr_(__ptr), __index_(__index) {}
+    reference(const reference&) = default;
+
+  public:
+    reference() = delete;
+    reference& operator=(const reference&) = delete;
+
+    operator _Tp() const { return __ptr_->__storage_[__index_]; }
+
+    reference operator=(_Tp __value) && {
+      __ptr_->__storage_[__index_] = __value;
+      return *this;
+    }
+
+    reference operator++() && {
+      ++__ptr_->__storage_[__index_];
+      return *this;
+    }
+
+    _Tp operator++(int)&& { return __ptr_->__storage_[__index_]++; }
+
+    reference operator--() && {
+      --__ptr_->__storage_[__index_];
+      return *this;
+    }
+
+    _Tp operator--(int)&& { return __ptr_->__storage_[__index_]--; }
+
+    reference operator+=(_Tp __value) && {
+      __ptr_->__storage_[__index_] += __value;
+      return *this;
+    }
+
+    reference operator-=(_Tp __value) && {
+      __ptr_->__storage_[__index_] -= __value;
+      return *this;
+    }
+
+    reference operator*=(_Tp __value) && {
+      __ptr_->__storage_[__index_] *= __value;
+      return *this;
+    }
+
+    reference operator/=(_Tp __value) && {
+      __ptr_->__storage_[__index_] /= __value;
+      return *this;
+    }
+
+    reference operator%=(_Tp __value) && {
+      __ptr_->__storage_[__index_] %= __value;
+      return *this;
+    }
+
+    reference operator>>=(_Tp __value) && {
+      __ptr_->__storage_[__index_] >>= __value;
+      return *this;
+    }
+
+    reference operator<<=(_Tp __value) && {
+      __ptr_->__storage_[__index_] <<= __value;
+      return *this;
+    }
+
+    reference operator&=(_Tp __value) && {
+      __ptr_->__storage_[__index_] &= __value;
+      return *this;
+    }
+
+    reference operator|=(_Tp __value) && {
+      __ptr_->__storage_[__index_] |= __value;
+      return *this;
+    }
+
+    reference operator^=(_Tp __value) && {
+      __ptr_->__storage_[__index_] ^= __value;
+      return *this;
+    }
+  };
+
+  _Tp __get(size_t __index) const { return __storage_[__index]; };
+  reference __ref(size_t __index) { return reference(this, __index); }
 };
 
+#endif // _LIBCPP_HAS_VECTOR_EXTENSION
+
 template <class _To, class _From>
 constexpr auto __is_non_narrowing_convertible_impl(_From a[[gnu::unused]])
     -> decltype(_To{a}, true) {
   return true;
 }
 
 template <class _To>
 constexpr bool __is_non_narrowing_convertible_impl(...) {
constexpr _Tp __variadic_sum() {
   return {};
 }
 
 template <class _Tp, class _Up, class... _Args>
 constexpr _Tp __variadic_sum(_Up __first, _Args... __rest) {
   return static_cast<_Tp>(__first) + __variadic_sum<_Tp>(__rest...);
 }
 
+template <class _Tp>
+struct __nodeduce {
+  using type = _Tp;
+};
+
+template <class _Tp>
+constexpr bool __vectorizable() {
+  return std::is_arithmetic<_Tp>::value && !std::is_const<_Tp>::value &&
+         !std::is_volatile<_Tp>::value && !std::is_same<_Tp, bool>::value;
+}
+
 _LIBCPP_END_NAMESPACE_EXPERIMENTAL_SIMD
 _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL_SIMD_ABI
 
 using scalar = __simd_abi<_StorageKind::_Scalar, 1>;
 
 template <int _Np>
 using fixed_size = __simd_abi<_StorageKind::_Array, _Np>;
 
 #if _LIBCPP_STD_VER > 14 && !defined(_LIBCPP_HAS_NO_VARIABLE_TEMPLATES) &&     \
     !defined(_LIBCPP_HAS_NO_INLINE_VARIABLES)
 template <class _Tp>
 inline constexpr int max_fixed_size = 32;
 #endif
+
 template <class _Tp>
 using compatible = fixed_size<16 / sizeof(_Tp)>;
+
 template <class _Tp>
-using native = compatible<_Tp>;
+using native =
+#ifdef _LIBCPP_HAS_VECTOR_EXTENSION
+    __simd_abi<_StorageKind::_VecExt,
+               _LIBCPP_NATIVE_SIMD_WIDTH_IN_BYTES / sizeof(_Tp)>;
+#else
+    fixed_size<_Tp, _LIBCPP_NATIVE_SIMD_WIDTH_IN_BYTES / sizeof(_Tp)>;
+#endif
 
 _LIBCPP_END_NAMESPACE_EXPERIMENTAL_SIMD_ABI
 _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL_SIMD
 
 template <class _Tp, class _Abi = simd_abi::compatible<_Tp>>
 class simd;
 template <class _Tp, class _Abi = simd_abi::compatible<_Tp>>
 class simd_mask;

 struct simd_size<_Tp, __simd_abi<__kind, _Np>>
     : std::integral_constant<size_t, _Np> {
   static_assert(
       std::is_arithmetic<_Tp>::value &&
           !std::is_same<typename std::remove_const<_Tp>::type, bool>::value,
       "Element type should be vectorizable");
 };
 
+// TODO: implement it.
 template <class _Tp, class _Up = typename _Tp::value_type>
 struct memory_alignment;
 
 #if _LIBCPP_STD_VER > 14 && !defined(_LIBCPP_HAS_NO_VARIABLE_TEMPLATES) &&     \
     !defined(_LIBCPP_HAS_NO_INLINE_VARIABLES)
 template <class _Tp, class _Abi = simd_abi::compatible<_Tp>>
 inline constexpr size_t simd_size_v = simd_size<_Tp, _Abi>::value;
 

 
 // masked assignment [simd.whereexpr]
 template <class _MaskType, class _Tp>
 class const_where_expression;
 template <class _MaskType, class _Tp>
 class where_expression;
 
 // masked assignment [simd.mask.where]
-template <class _Tp>
-struct __nodeduce {
-  using type = _Tp;
-};
-
 template <class _Tp, class _Abi>
 where_expression<simd_mask<_Tp, _Abi>, simd<_Tp, _Abi>>
 where(const typename simd<_Tp, _Abi>::mask_type&, simd<_Tp, _Abi>&) noexcept;
 
 template <class _Tp, class _Abi>
 const_where_expression<simd_mask<_Tp, _Abi>, const simd<_Tp, _Abi>>
 where(const typename simd<_Tp, _Abi>::mask_type&,
       const simd<_Tp, _Abi>&) noexcept;
public:
   void operator++(int);
   void operator--();
   void operator--(int);
   template <class _Up, class _Flags>
   void copy_from(const _Up*, _Flags);
 };
 
 // [simd.class]
 // TODO: implement simd
 template <class _Tp, class _Abi>
 class simd {
+public:
+  using value_type = _Tp;
+  using reference = typename __simd_storage<_Tp, _Abi>::reference;
+  using mask_type = simd_mask<_Tp, _Abi>;
+  using abi_type = _Abi;
+
+  simd() = default;
+  simd(const simd&) = default;
+  simd& operator=(const simd&) = default;
+
+  static constexpr size_t size() noexcept {
+    return simd_size<_Tp, _Abi>::value;
+  }
+
 private:
+  __simd_storage<_Tp, _Abi> __s_;
+
   template <class _Up>
   static constexpr bool __can_broadcast() {
     return (std::is_arithmetic<_Up>::value &&
             __is_non_narrowing_arithmetic_convertible<_Up, _Tp>()) ||
            (!std::is_arithmetic<_Up>::value &&
             std::is_convertible<_Up, _Tp>::value) ||
            std::is_same<typename std::remove_const<_Up>::type, int>::value ||
            (std::is_same<typename std::remove_const<_Up>::type,
                          unsigned int>::value &&
             std::is_unsigned<_Tp>::value);
   }
 
-public:
-  using value_type = _Tp;
-  // TODO: this is strawman implementation. Turn it into a proxy type.
-  using reference = _Tp&;
-  using mask_type = simd_mask<_Tp, _Abi>;
-
-  using abi_type = _Abi;
+  template <class _Generator, size_t... __indicies>
+  static constexpr decltype(
+      std::forward_as_tuple(std::declval<_Generator>()(
+          std::integral_constant<size_t, __indicies>())...),
+      bool())
+  __can_generate(std::index_sequence<__indicies...>) {
+    return !__variadic_sum<bool>(
+        !__can_broadcast<decltype(std::declval<_Generator>()(
+            std::integral_constant<size_t, __indicies>()))>()...);
+  }
 
-  static constexpr size_t size() noexcept {
-    return simd_size<_Tp, _Abi>::value;
+  template <class _Generator>
+  static bool __can_generate(...) {
+    return false;
   }
 
-  simd() = default;
+  template <class _Generator, size_t... __indicies>
+  void __generator_init(_Generator&& __g, std::index_sequence<__indicies...>) {
+    int __unused[] = {((*this)[__indicies] =
+                           __g(std::integral_constant<size_t, __indicies>()),
+                       0)...};
+    (void)__unused;
+  }
 
+public:
   // implicit type conversion constructor
   template <class _Up,
             class = typename std::enable_if<
                 std::is_same<_Abi, simd_abi::fixed_size<size()>>::value &&
                 __is_non_narrowing_arithmetic_convertible<_Up, _Tp>()>::type>
-  simd(const simd<_Up, simd_abi::fixed_size<size()>>&) {}
+  simd(const simd<_Up, simd_abi::fixed_size<size()>>& __v) {
+    for (size_t __i = 0; __i < size(); __i++) {
+      (*this)[__i] = static_cast<_Tp>(__v[__i]);
+    }
+  }
 
   // implicit broadcast constructor
   template <class _Up,
             class = typename std::enable_if<__can_broadcast<_Up>()>::type>
-  simd(_Up&&);
+  simd(_Up&& __rv) {
+    auto __v = static_cast<_Tp>(__rv);
+    for (size_t __i = 0; __i < size(); __i++) {
+      (*this)[__i] = __v;
+    }
+  }
 
   // generator constructor
-  // TODO: for now only check for the index 0. This is because C++11 doesn't
-  // have index_sequence, and it's hard to check for all indicies without using
-  // index_sequence.
   template <class _Generator,
-            int = decltype(simd(std::declval<_Generator>()(
-                               std::integral_constant<size_t, 0>())),
-                           int())()>
-  explicit simd(_Generator&&);
+            int = typename std::enable_if<
+                __can_generate<_Generator>(std::make_index_sequence<size()>()),
+                int>::type()>
+  explicit simd(_Generator&& __g) {
+    __generator_init(std::forward<_Generator>(__g),
+                     std::make_index_sequence<size()>());
+  }
 
   // load constructor
-  template <class _Up, class _Flags>
-  simd(const _Up*, _Flags);
+  template <
+      class _Up, class _Flags,
+      class = typename std::enable_if<__vectorizable<_Up>()>::type,
+      class = typename std::enable_if<is_simd_flag_type<_Flags>::value>::type>
+  simd(const _Up* __buffer, _Flags) {
+    // TODO: optimize for overaligned flags
+    for (size_t __i = 0; __i < size(); __i++) {
+      (*this)[__i] = static_cast<_Tp>(__buffer[__i]);
+    }
+  }
 
   // loads [simd.load]
   template <class _Up, class _Flags>
-  void copy_from(const _Up*, _Flags);
+  typename std::enable_if<__vectorizable<_Up>() &&
+                          is_simd_flag_type<_Flags>::value>::type
+  copy_from(const _Up* __buffer, _Flags) {
+    *this = simd(__buffer, _Flags());
+  }
 
   // stores [simd.store]
   template <class _Up, class _Flags>
-  void copy_to(_Up*, _Flags) const;
+  typename std::enable_if<__vectorizable<_Up>() &&
+                          is_simd_flag_type<_Flags>::value>::type
+  copy_to(_Up* __buffer, _Flags) const {
+    // TODO: optimize for overaligned flags
+    for (size_t __i = 0; __i < size(); __i++) {
+      __buffer[__i] = static_cast<_Up>((*this)[__i]);
+    }
+  }
 
   // scalar access [simd.subscr]
-  reference operator[](size_t);
-  value_type operator[](size_t) const;
+  reference operator[](size_t __i) { return __s_.__ref(__i); }
+
+  value_type operator[](size_t __i) const { return __s_.__get(__i); }
 
   // unary operators [simd.unary]
   simd& operator++();
   simd operator++(int);
   simd& operator--();
   simd operator--(int);
   mask_type operator!() const;
   simd operator~() const;

libcxx/include/utility

 inline _LIBCPP_INLINE_VISIBILITY
 constexpr _T1 const && get(pair<_T2, _T1> const&& __p) _NOEXCEPT
 {
     return __get_pair<1>::get(_VSTD::move(__p));
 }
 
 #endif
 
-#if _LIBCPP_STD_VER > 11
+#if _LIBCPP_STD_VER >= 11
 
 template<class _Tp, _Tp... _Ip>
 struct _LIBCPP_TEMPLATE_VIS integer_sequence
 {
     typedef _Tp value_type;
     static_assert( is_integral<_Tp>::value,
                   "std::integer_sequence can only be instantiated with an integral type" );
     static
template<class _Tp, _Tp _Np>
     using make_integer_sequence = __make_integer_sequence<_Tp, _Np>;
 
 template<size_t _Np>
     using make_index_sequence = make_integer_sequence<size_t, _Np>;
 
 template<class... _Tp>
     using index_sequence_for = make_index_sequence<sizeof...(_Tp)>;
 
-#endif  // _LIBCPP_STD_VER > 11
+#endif  // _LIBCPP_STD_VER >= 11
 
 #if _LIBCPP_STD_VER > 11
 template<class _T1, class _T2 = _T1>
 inline _LIBCPP_INLINE_VISIBILITY
 _T1 exchange(_T1& __obj, _T2 && __new_value)
 {
     _T1 __old_value = _VSTD::move(__obj);
     __obj = _VSTD::forward<_T2>(__new_value);

libcxx/test/std/experimental/simd/simd.abi/vector_extension.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++98, c++03
+
+// <experimental/simd>
+//
+// [simd.abi]
+
+#include <experimental/simd>
+#include <cstdint>
+
+using namespace std::experimental::parallelism_v2;
+
+constexpr inline int reg_width() {
+#if defined(__AVX__)
+  return 32;
+#else
+  return 16;
+#endif
+}
+
+#ifdef _LIBCPP_HAS_VECTOR_EXTENSION
+
+static_assert(sizeof(simd<char, __simd_abi<_StorageKind::_VecExt, 1>>) == 1,
+              "");
+static_assert(sizeof(simd<char, __simd_abi<_StorageKind::_VecExt, 2>>) == 2,
+              "");
+static_assert(sizeof(simd<char, __simd_abi<_StorageKind::_VecExt, 3>>) == 4,
+              "");
+static_assert(sizeof(simd<char, __simd_abi<_StorageKind::_VecExt, 12>>) == 16,
+              "");
+static_assert(sizeof(simd<int32_t, __simd_abi<_StorageKind::_VecExt, 3>>) == 16,
+              "");
+static_assert(sizeof(simd<int32_t, __simd_abi<_StorageKind::_VecExt, 5>>) == 32,
+              "");
+static_assert(
+    std::is_same<simd_abi::native<int8_t>,
+                 __simd_abi<_StorageKind::_VecExt, reg_width()>>::value,
+    "");
+#else
+static_assert(
+    std::is_same<simd_abi::native<int8_t>,
+                 __simd_abi<_StorageKind::_Array, reg_width()>>::value,
+    "");
+
+#endif
+
+static_assert(std::is_same<simd_abi::compatible<int8_t>,
+                           __simd_abi<_StorageKind::_Array, 16>>::value,
+              "");
+
+int main() {}

libcxx/test/std/experimental/simd/simd.access/default.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++98, c++03
+
+// <experimental/simd>
+//
+// scalar access [simd.subscr]
+// reference operator[](size_t);
+// value_type operator[](size_t) const;
+
+#include <experimental/simd>
+#include <cassert>
+#include <cstdint>
+
+using namespace std::experimental::parallelism_v2;
+
+void test_access() {
+  {
+    native_simd<int> a(42), b(4);
+    static_assert(std::is_convertible<decltype(a[0]), int8_t>::value, "");
+
+    assert(a[0] == 42);
+    assert(!a[0] == !42);
+    assert(~a[0] == ~42);
+    assert(+a[0] == +42);
+    assert(-a[0] == -42);
+    assert(a[0] + b[0] == 42 + 4);
+    assert(a[0] - b[0] == 42 - 4);
+    assert(a[0] * b[0] == 42 * 4);
+    assert(a[0] / b[0] == 42 / 4);
+    assert(a[0] % b[0] == 42 % 4);
+    assert(a[0] << b[0] == (42 << 4));
+    assert(a[0] >> b[0] == (42 >> 4));
+    assert(a[0] < b[0] == false);
+    assert(a[0] <= b[0] == false);
+    assert(a[0] > b[0] == true);
+    assert(a[0] >= b[0] == true);
+    assert(a[0] == b[0] == false);
+    assert(a[0] != b[0] == true);
+    assert((a[0] & b[0]) == (42 & 4));
+    assert((a[0] | b[0]) == (42 | 4));
+    assert((a[0] ^ b[0]) == (42 ^ 4));
+    assert((a[0] && b[0]) == true);
+    assert((a[0] || b[0]) == true);
+
+    {
+      auto c = a;
+      ++c[0];
+      assert(c[0] == 42 + 1);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      auto ret = c[0]++;
+      assert(ret == 42);
+      assert(c[0] == 42 + 1);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      --c[0];
+      assert(c[0] == 42 - 1);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      auto ret = c[0]--;
+      assert(ret == 42);
+      assert(c[0] == 42 - 1);
+      assert(c[1] == 42);
+    }
+
+    {
+      auto c = a;
+      c[0] += b[0];
+      assert(c[0] == 42 + 4);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] -= b[0];
+      assert(c[0] == 42 - 4);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] *= b[0];
+      assert(c[0] == 42 * 4);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] /= b[0];
+      assert(c[0] == 42 / 4);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] %= b[0];
+      assert(c[0] == 42 % 4);
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] >>= b[0];
+      assert(c[0] == (42 >> 4));
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] <<= b[0];
+      assert(c[0] == (42 << 4));
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] &= b[0];
+      assert(c[0] == (42 & 4));
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] |= b[0];
+      assert(c[0] == (42 | 4));
+      assert(c[1] == 42);
+    }
+    {
+      auto c = a;
+      c[0] ^= b[0];
+      assert(c[0] == (42 ^ 4));
+      assert(c[1] == 42);
+    }
+
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] += a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] -= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] *= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] /= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] %= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] >>= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] <<= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] &= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] |= a[0]));
+    }
+    {
+      auto c = a;
+      (void)(a[0] + (c[0] ^= a[0]));
+    }
+  }
+  {
+    const native_simd<int> a(42);
+    const native_simd<int> b(4);
+    static_assert(std::is_same<decltype(a[0]), int>::value, "");
+
+    assert(a[0] == 42);
+    assert(!a[0] == !42);
+    assert(~a[0] == ~42);
+    assert(+a[0] == +42);
+    assert(-a[0] == -42);
+    assert(a[0] + b[0] == 42 + 4);
+    assert(a[0] - b[0] == 42 - 4);
+    assert(a[0] * b[0] == 42 * 4);
+    assert(a[0] / b[0] == 42 / 4);
+    assert(a[0] % b[0] == 42 % 4);
+    assert(a[0] << b[0] == (42 << 4));
+    assert(a[0] >> b[0] == (42 >> 4));
+    assert(a[0] < b[0] == false);
+    assert(a[0] <= b[0] == false);
+    assert(a[0] > b[0] == true);
+    assert(a[0] >= b[0] == true);
+    assert(a[0] == b[0] == false);
+    assert(a[0] != b[0] == true);
+    assert((a[0] & b[0]) == (42 & 4));
+    assert((a[0] | b[0]) == (42 | 4));
+    assert((a[0] ^ b[0]) == (42 ^ 4));
+    assert((a[0] && b[0]) == true);
+    assert((a[0] || b[0]) == true);
+  }
+}
+
+int main() { test_access(); }

libcxx/test/std/experimental/simd/simd.casts/simd_cast.pass.cpp

 //===----------------------------------------------------------------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is dual licensed under the MIT and the University of Illinois Open
 // Source Licenses. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 
 // UNSUPPORTED: c++98, c++03
 
 // <experimental/simd>
 //
 // [simd.casts]
 // template <class T, class U, class Abi> see below simd_cast(const simd<U, Abi>&);
+
 #include <experimental/simd>
 #include <cstdint>
 
 using namespace std::experimental::parallelism_v2;
 
 static_assert(std::is_same<decltype(simd_cast<int32_t>(native_simd<int32_t>())),
                            native_simd<int32_t>>::value,
               "");

libcxx/test/std/experimental/simd/simd.cons/broadcast.pass.cpp

void compile_narrowing_conversion() {
   supported_native_simd_ctor<long double>(3.);
 
   not_supported_native_simd_ctor<float>(3.);
   not_supported_native_simd_ctor<int8_t>(long(3));
   not_supported_native_simd_ctor<float>(long(3));
   not_supported_native_simd_ctor<int>(3.);
 }
 
-int main() {}
+void compile_convertible() {
+  struct ConvertibleToInt {
+    operator int64_t() const;
+  };
+  supported_native_simd_ctor<int64_t>(ConvertibleToInt());
+
+  struct NotConvertibleToInt {};
+  not_supported_native_simd_ctor<int64_t>(NotConvertibleToInt());
+}
+
+void compile_unsigned() {
+  not_supported_native_simd_ctor<int>(3u);
+  supported_native_simd_ctor<uint16_t>(3u);
+}
+
+void test_broadcast() {
+  native_simd<int> a(3);
+  for (size_t i = 0; i < a.size(); i++) {
+    assert(a[i] == 3);
+  }
+}
+
+int main() { test_broadcast(); }

libcxx/test/std/experimental/simd/simd.cons/default.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++98, c++03
+
+// <experimental/simd>
+//
+// [simd.class]
+// template <class U> simd() = default;
+
+#include <cstdint>
+#include <experimental/simd>
+
+using namespace std::experimental::parallelism_v2;
+
+int main() { (void)native_simd<int32_t>(); }

libcxx/test/std/experimental/simd/simd.cons/geneartor.pass.cpp

 //===----------------------------------------------------------------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is dual licensed under the MIT and the University of Illinois Open
 // Source Licenses. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 
 // UNSUPPORTED: c++98, c++03
 
 // <experimental/simd>
 //
 // [simd.class]
 // template <class G> explicit simd(G&& gen);
 
-#include <cstdint>
 #include <experimental/simd>
+#include <cstdint>
 
 using namespace std::experimental::parallelism_v2;
 
 template <class T, class... Args>
-auto not_supported_native_simd_ctor(Args&&... args)
-    -> decltype(native_simd<T>(std::forward<Args>(args)...), void()) = delete;
+auto not_supported_simd128_ctor(Args&&... args)
+    -> decltype(fixed_size_simd<T, 16 / sizeof(T)>(std::forward<Args>(args)...),
+                void()) = delete;
 
 template <class T>
-void not_supported_native_simd_ctor(...) {}
+void not_supported_simd128_ctor(...) {}
 
 template <class T, class... Args>
-auto supported_native_simd_ctor(Args&&... args)
-    -> decltype(native_simd<T>(std::forward<Args>(args)...), void()) {}
+auto supported_simd128_ctor(Args&&... args)
+    -> decltype(fixed_size_simd<T, 16 / sizeof(T)>(std::forward<Args>(args)...),
+                void()) {}
 
 template <class T>
-void supported_native_simd_ctor(...) = delete;
+void supported_simd128_ctor(...) = delete;
+
+struct identity {
+  template <size_t value>
+  int operator()(std::integral_constant<size_t, value>) const {
+    return value;
+  }
+};
 
 void compile_generator() {
-  supported_native_simd_ctor<int>([](int i) { return i; });
-  not_supported_native_simd_ctor<int>([](int i) { return float(i); });
-  not_supported_native_simd_ctor<int>([](intptr_t i) { return (int*)(i); });
-  not_supported_native_simd_ctor<int>([](int* i) { return i; });
+  supported_simd128_ctor<int>(identity());
+  not_supported_simd128_ctor<int>([](int i) { return float(i); });
+  not_supported_simd128_ctor<int>([](intptr_t i) { return (int*)(i); });
+  not_supported_simd128_ctor<int>([](int* i) { return i; });
+}
+
+struct limited_identity {
+  template <size_t value>
+  typename std::conditional<value <= 2, int32_t, int64_t>::type
+  operator()(std::integral_constant<size_t, value>) const {
+    return value;
+  }
+};
+
+void compile_limited_identity() {
+  supported_simd128_ctor<int64_t>(limited_identity());
+  not_supported_simd128_ctor<int32_t>(limited_identity());
+}
+
+void test_generator() {
+  {
+    fixed_size_simd<int32_t, 4> a([](int i) { return i; });
+    assert(a[0] == 0);
+    assert(a[1] == 1);
+    assert(a[2] == 2);
+    assert(a[3] == 3);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a([](int i) { return 2 * i - 1; });
+    assert(a[0] == -1);
+    assert(a[1] == 1);
+    assert(a[2] == 3);
+    assert(a[3] == 5);
+  }
 }
 
-int main() {}
+int main() { test_generator(); }

libcxx/test/std/experimental/simd/simd.cons/load.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++98, c++03
+
+// <experimental/simd>
+//
+// [simd.class]
+// template <class U, class Flags> simd(const U* mem, Flags f);
+
+#include <cstdint>
+#include <experimental/simd>
+
+using namespace std::experimental::parallelism_v2;
+
+template <class T, class... Args>
+auto not_supported_native_simd_ctor(Args&&... args)
+    -> decltype(native_simd<T>(std::forward<Args>(args)...), void()) = delete;
+
+template <class T>
+void not_supported_native_simd_ctor(...) {}
+
+template <class T, class... Args>
+auto supported_native_simd_ctor(Args&&... args)
+    -> decltype(native_simd<T>(std::forward<Args>(args)...), void()) {}
+
+template <class T>
+void supported_native_simd_ctor(...) = delete;
+
+void compile_load_ctor() {
+  supported_native_simd_ctor<int>((int*)nullptr, element_aligned_tag());
+  supported_native_simd_ctor<uint32_t>((int*)nullptr, element_aligned_tag());
+  supported_native_simd_ctor<double>((float*)nullptr, element_aligned_tag());
+  supported_native_simd_ctor<uint16_t>((unsigned int*)nullptr,
+                                       element_aligned_tag());
+  supported_native_simd_ctor<uint32_t>((float*)nullptr, element_aligned_tag());
+
+  not_supported_native_simd_ctor<int>((int*)nullptr, int());
+}
+
+void test_load_ctor() {
+  alignas(32) int32_t buffer[] = {4, 3, 2, 1};
+  {
+    fixed_size_simd<int32_t, 4> a(buffer, element_aligned_tag());
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a(buffer, vector_aligned_tag());
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a(buffer, overaligned_tag<32>());
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+
+#if TEST_STD_VER > 14
+  {
+    fixed_size_simd<int32_t, 4> a(buffer, element_aligned);
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a(buffer, vector_aligned);
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a(buffer, overaligned);
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+#endif
+}
+
+void test_converting_load_ctor() {
+  float buffer[] = {1., 2., 4., 8.};
+  fixed_size_simd<int32_t, 4> a(buffer, element_aligned_tag());
+  assert(a[0] == 1);
+  assert(a[1] == 2);
+  assert(a[2] == 4);
+  assert(a[3] == 8);
+}
+
+int main() {
+  test_load_ctor();
+  test_converting_load_ctor();
+}

libcxx/test/std/experimental/simd/simd.mem/load.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++98, c++03
+
+// <experimental/simd>
+//
+// loads [simd.load]
+// template <class U, class Flags> void copy_from(const U* mem, Flags f);
+
+#include <experimental/simd>
+#include <cstdint>
+
+using namespace std::experimental::parallelism_v2;
+
+template <class T, class... Args>
+auto not_supported_load(Args&&... args) -> decltype(
+    std::declval<native_simd<T>>().copy_from(std::forward<Args>(args)...),
+    void()) = delete;
+
+template <class T>
+void not_supported_load(...) {}
+
+template <class T, class... Args>
+auto supported_load(Args&&... args) -> decltype(
+    std::declval<native_simd<T>>().copy_from(std::forward<Args>(args)...),
+    void()) {}
+
+template <class T>
+void supported_load(...) = delete;
+
+void compile_load() {
+  supported_load<int>((int*)nullptr, element_aligned_tag());
+  supported_load<uint32_t>((int*)nullptr, element_aligned_tag());
+  supported_load<double>((float*)nullptr, element_aligned_tag());
+  supported_load<uint16_t>((unsigned int*)nullptr, element_aligned_tag());
+  supported_load<uint32_t>((float*)nullptr, element_aligned_tag());
+
+  not_supported_load<int>((int*)nullptr, int());
+}
+
+void test_load() {
+  alignas(32) int32_t buffer[] = {4, 3, 2, 1};
+  {
+    fixed_size_simd<int32_t, 4> a;
+    a.copy_from(buffer, element_aligned_tag());
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a;
+    a.copy_from(buffer, vector_aligned_tag());
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a;
+    a.copy_from(buffer, overaligned_tag<32>());
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+
+#if TEST_STD_VER > 14
+  {
+    fixed_size_simd<int32_t, 4> a;
+    a.copy_from(buffer, element_aligned);
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a;
+    a.copy_from(buffer, vector_aligned);
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+  {
+    fixed_size_simd<int32_t, 4> a;
+    a.copy_from(buffer, overaligned);
+    assert(a[0] == 4);
+    assert(a[1] == 3);
+    assert(a[2] == 2);
+    assert(a[3] == 1);
+  }
+#endif
+}
+
+void test_converting_load() {
+  float buffer[] = {1., 2., 4., 8.};
+  fixed_size_simd<int32_t, 4> a;
+  a.copy_from(buffer, element_aligned_tag());
+  assert(a[0] == 1);
+  assert(a[1] == 2);
+  assert(a[2] == 4);
+  assert(a[3] == 8);
+}
+
+int main() {
+  test_load();
+  test_converting_load();
+}

libcxx/test/std/experimental/simd/simd.mem/store.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: c++98, c++03
+
+// <experimental/simd>
+//
+// // stores [simd.store]
+// template <class U, class Flags> void copy_to(U* mem, Flags f) const;
+
+#include <experimental/simd>
+#include <cstdint>
+
+using namespace std::experimental::parallelism_v2;
+
+void test_store() {
+  fixed_size_simd<int32_t, 4> a([](int i) { return 4 - i; });
+  {
+    alignas(32) int32_t buffer[4] = {0};
+    a.copy_to(buffer, element_aligned_tag());
+    assert(buffer[0] == 4);
+    assert(buffer[1] == 3);
+    assert(buffer[2] == 2);
+    assert(buffer[3] == 1);
+  }
+  {
+    alignas(32) int32_t buffer[4] = {0};
+    a.copy_to(buffer, vector_aligned_tag());
+    assert(buffer[0] == 4);
+    assert(buffer[1] == 3);
+    assert(buffer[2] == 2);
+    assert(buffer[3] == 1);
+  }
+  {
+    alignas(32) int32_t buffer[4] = {0};
+    a.copy_to(buffer, overaligned_tag<32>());
+    assert(buffer[0] == 4);
+    assert(buffer[1] == 3);
+    assert(buffer[2] == 2);
+    assert(buffer[3] == 1);
+  }
+
+#if TEST_STD_VER > 14
+  {
+    alignas(32) int32_t buffer[4] = {0};
+    a.copy_to(buffer, element_aligned);
+    assert(buffer[0] == 4);
+    assert(buffer[1] == 3);
+    assert(buffer[2] == 2);
+    assert(buffer[3] == 1);
+  }
+  {
+    alignas(32) int32_t buffer[4] = {0};
+    a.copy_to(buffer, vector_aligned);
+    assert(buffer[0] == 4);
+    assert(buffer[1] == 3);
+    assert(buffer[2] == 2);
+    assert(buffer[3] == 1);
+  }
+  {
+    alignas(32) int32_t buffer[4] = {0};
+    a.copy_to(buffer, overaligned);
+    assert(buffer[0] == 4);
+    assert(buffer[1] == 3);
+    assert(buffer[2] == 2);
+    assert(buffer[3] == 1);
+  }
+#endif
+}
+
+void test_converting_store() {
+  float buffer[4] = {0.};
+  fixed_size_simd<int32_t, 4> a([](int i) { return 1 << i; });
+  a.copy_to(buffer, element_aligned_tag());
+  assert(buffer[0] == 1.);
+  assert(buffer[1] == 2.);
+  assert(buffer[2] == 4.);
+  assert(buffer[3] == 8.);
+}
+
+int main() {
+  test_store();
+  test_converting_store();
+}