diff --git a/libcxx/include/CMakeLists.txt b/libcxx/include/CMakeLists.txt
--- a/libcxx/include/CMakeLists.txt
+++ b/libcxx/include/CMakeLists.txt
@@ -907,6 +907,8 @@
   experimental/__simd/abi_tag.h
   experimental/__simd/aligned_tag.h
   experimental/__simd/declaration.h
+  experimental/__simd/internal_declaration.h
+  experimental/__simd/reference.h
   experimental/__simd/scalar.h
   experimental/__simd/simd.h
   experimental/__simd/simd_mask.h
diff --git a/libcxx/include/experimental/__simd/scalar.h b/libcxx/include/experimental/__simd/internal_declaration.h
copy from libcxx/include/experimental/__simd/scalar.h
copy to libcxx/include/experimental/__simd/internal_declaration.h
--- a/libcxx/include/experimental/__simd/scalar.h
+++ b/libcxx/include/experimental/__simd/internal_declaration.h
@@ -7,23 +7,28 @@
 //
 //===----------------------------------------------------------------------===//
 
-#ifndef _LIBCPP_EXPERIMENTAL___SIMD_SCALAR_H
-#define _LIBCPP_EXPERIMENTAL___SIMD_SCALAR_H
-
-#include <cstddef>
+#ifndef _LIBCPP_EXPERIMENTAL___SIMD_INTERNAL_DECLARATION_H
+#define _LIBCPP_EXPERIMENTAL___SIMD_INTERNAL_DECLARATION_H
 
 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
 
 _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL
 inline namespace parallelism_v2 {
-namespace simd_abi {
-struct __scalar {
-  static constexpr size_t __simd_size = 1;
-};
 
-} // namespace simd_abi
+template <class _Tp, class _Abi>
+struct __simd_storage;
+
+template <class _Tp, class _Abi>
+struct __mask_storage;
+
+template <class _Tp, class _Abi>
+struct __simd_operations;
+
+template <class _Tp, class _Abi>
+struct __mask_operations;
+
 } // namespace parallelism_v2
 _LIBCPP_END_NAMESPACE_EXPERIMENTAL
 
 #endif // _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
-#endif // _LIBCPP_EXPERIMENTAL___SIMD_SCALAR_H
+#endif // _LIBCPP_EXPERIMENTAL___SIMD_INTERNAL_DECLARATION_H
diff --git a/libcxx/include/experimental/__simd/reference.h b/libcxx/include/experimental/__simd/reference.h
new file mode 100644
--- /dev/null
+++ b/libcxx/include/experimental/__simd/reference.h
@@ -0,0 +1,53 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP_EXPERIMENTAL___SIMD_REFERENCE_H
+#define _LIBCPP_EXPERIMENTAL___SIMD_REFERENCE_H
+
+#include <experimental/__simd/utility.h>
+
+#if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
+
+_LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL
+inline namespace parallelism_v2 {
+template <class _Tp, class _Storage, class _Vp>
+class __reference { // exposition only
+  template <class, class>
+  friend class simd;
+  template <class, class>
+  friend class simd_mask;
+
+  _Storage& __s_;
+  size_t __idx_;
+
+  __reference(_Storage& __s, size_t __idx) : __s_(__s), __idx_(__idx) {}
+
+  _Vp __get() const { return __s_.__get(__idx_); }
+
+  void __set(_Vp __v) {
+    if constexpr (is_same_v<_Vp, bool>)
+      __s_.__set(__idx_, __set_all_bits<_Tp>(__v));
+    else
+      __s_.__set(__idx_, __v);
+  }
+
+public:
+  using value_type = _Vp;
+
+  __reference()                   = delete;
+  __reference(const __reference&) = delete;
+
+  operator value_type() const noexcept { return __get(); }
+};
+
+} // namespace parallelism_v2
+_LIBCPP_END_NAMESPACE_EXPERIMENTAL
+
+#endif // _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
+#endif // _LIBCPP_EXPERIMENTAL___SIMD_REFERENCE_H
diff --git a/libcxx/include/experimental/__simd/scalar.h b/libcxx/include/experimental/__simd/scalar.h
--- a/libcxx/include/experimental/__simd/scalar.h
+++ b/libcxx/include/experimental/__simd/scalar.h
@@ -11,6 +11,7 @@
 #define _LIBCPP_EXPERIMENTAL___SIMD_SCALAR_H
 
 #include <cstddef>
+#include <experimental/__simd/internal_declaration.h>
 
 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
 
@@ -20,8 +21,30 @@
 struct __scalar {
   static constexpr size_t __simd_size = 1;
 };
-
 } // namespace simd_abi
+
+template <class _Tp>
+struct __simd_storage<_Tp, simd_abi::__scalar> {
+  _Tp __data;
+
+  _Tp __get(size_t __idx) const noexcept { return (&__data)[__idx]; }
+  void __set(size_t __idx, _Tp __v) noexcept { (&__data)[__idx] = __v; }
+};
+
+template <class _Tp>
+struct __mask_storage<_Tp, simd_abi::__scalar> : __simd_storage<bool, simd_abi::__scalar> {};
+
+template <class _Tp>
+struct __simd_operations<_Tp, simd_abi::__scalar> {
+  using _Simd = __simd_storage<_Tp, simd_abi::__scalar>;
+  using _Mask = __mask_storage<_Tp, simd_abi::__scalar>;
+};
+
+template <class _Tp>
+struct __mask_operations<_Tp, simd_abi::__scalar> {
+  using _Mask = __mask_storage<_Tp, simd_abi::__scalar>;
+};
+
 } // namespace parallelism_v2
 _LIBCPP_END_NAMESPACE_EXPERIMENTAL
 
diff --git a/libcxx/include/experimental/__simd/simd.h b/libcxx/include/experimental/__simd/simd.h
--- a/libcxx/include/experimental/__simd/simd.h
+++ b/libcxx/include/experimental/__simd/simd.h
@@ -12,7 +12,10 @@
 
 #include <experimental/__simd/abi_tag.h>
 #include <experimental/__simd/declaration.h>
+#include <experimental/__simd/reference.h>
+#include <experimental/__simd/scalar.h>
 #include <experimental/__simd/traits.h>
+#include <experimental/__simd/vec_ext.h>
 
 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
 
@@ -23,8 +26,14 @@
 // TODO: implement simd class
 template <class _Tp, class _Abi>
 class simd {
+  using _Impl    = __simd_operations<_Tp, _Abi>;
+  using _Storage = typename _Impl::_Simd;
+
+  _Storage __s_;
+
 public:
   using value_type = _Tp;
+  using reference  = __reference<_Tp, _Storage, value_type>;
   using mask_type  = simd_mask<_Tp, _Abi>;
   using abi_type   = _Abi;
 
diff --git a/libcxx/include/experimental/__simd/simd_mask.h b/libcxx/include/experimental/__simd/simd_mask.h
--- a/libcxx/include/experimental/__simd/simd_mask.h
+++ b/libcxx/include/experimental/__simd/simd_mask.h
@@ -12,6 +12,9 @@
 
 #include <experimental/__simd/abi_tag.h>
 #include <experimental/__simd/declaration.h>
+#include <experimental/__simd/reference.h>
+#include <experimental/__simd/scalar.h>
+#include <experimental/__simd/vec_ext.h>
 
 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
 
@@ -22,8 +25,14 @@
 // TODO: implement simd_mask class
 template <class _Tp, class _Abi>
 class simd_mask {
+  using _Impl    = __mask_operations<_Tp, _Abi>;
+  using _Storage = typename _Impl::_Mask;
+
+  _Storage __s_;
+
 public:
   using value_type = bool;
+  using reference  = __reference<_Tp, _Storage, value_type>;
   using simd_type  = simd<_Tp, _Abi>;
   using abi_type   = _Abi;
 
diff --git a/libcxx/include/experimental/__simd/utility.h b/libcxx/include/experimental/__simd/utility.h
--- a/libcxx/include/experimental/__simd/utility.h
+++ b/libcxx/include/experimental/__simd/utility.h
@@ -15,6 +15,7 @@
 #include <__type_traits/is_constant_evaluated.h>
 #include <__type_traits/is_same.h>
 #include <__type_traits/is_volatile.h>
+#include <cstdint>
 
 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
 
@@ -23,6 +24,36 @@
 template <class _Tp>
 constexpr bool __is_vectorizable_v =
     is_arithmetic_v<_Tp> && !is_const_v<_Tp> && !is_volatile_v<_Tp> && !is_same_v<_Tp, bool>;
+
+template <class _Tp>
+_LIBCPP_HIDE_FROM_ABI auto __choose_mask_type() {
+  if constexpr (sizeof(_Tp) == 1) {
+    return uint8_t{};
+  } else if constexpr (sizeof(_Tp) == 2) {
+    return uint16_t{};
+  } else if constexpr (sizeof(_Tp) == 4) {
+    return uint32_t{};
+  } else if constexpr (sizeof(_Tp) == 8) {
+    return uint64_t{};
+  }
+#  ifndef _LIBCPP_HAS_NO_INT128
+  else if constexpr (sizeof(_Tp) == 16) {
+    return __uint128_t{};
+  }
+#  endif
+  else
+    static_assert(sizeof(_Tp) == 0);
+}
+
+template <class _Tp>
+_LIBCPP_HIDE_FROM_ABI auto constexpr __set_all_bits(bool __v) {
+  using _Up = decltype(__choose_mask_type<_Tp>());
+  _Up __res = 0;
+  for (unsigned long __i = 0; __i < __CHAR_BIT__ * sizeof(_Tp); __i++)
+    __res |= static_cast<_Up>(__v) << __i;
+  return __res;
+}
+
 } // namespace parallelism_v2
 _LIBCPP_END_NAMESPACE_EXPERIMENTAL
 
diff --git a/libcxx/include/experimental/__simd/vec_ext.h b/libcxx/include/experimental/__simd/vec_ext.h
--- a/libcxx/include/experimental/__simd/vec_ext.h
+++ b/libcxx/include/experimental/__simd/vec_ext.h
@@ -11,6 +11,8 @@
 #define _LIBCPP_EXPERIMENTAL___SIMD_VEC_EXT_H
 
 #include <cstddef>
+#include <experimental/__simd/internal_declaration.h>
+#include <experimental/__simd/utility.h>
 
 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
 
@@ -21,8 +23,31 @@
 struct __vec_ext {
   static constexpr size_t __simd_size = _Np;
 };
-
 } // namespace simd_abi
+
+template <class _Tp, int _Np>
+struct __simd_storage<_Tp, simd_abi::__vec_ext<_Np>> {
+  _Tp __data __attribute__((vector_size(sizeof(_Tp) * _Np)));
+
+  _Tp __get(size_t __idx) const noexcept { return __data[__idx]; }
+  void __set(size_t __idx, _Tp __v) noexcept { __data[__idx] = __v; }
+};
+
+template <class _Tp, int _Np>
+struct __mask_storage<_Tp, simd_abi::__vec_ext<_Np>>
+    : __simd_storage<decltype(__choose_mask_type<_Tp>()), simd_abi::__vec_ext<_Np>> {};
+
+template <class _Tp, int _Np>
+struct __simd_operations<_Tp, simd_abi::__vec_ext<_Np>> {
+  using _Simd = __simd_storage<_Tp, simd_abi::__vec_ext<_Np>>;
+  using _Mask = __mask_storage<_Tp, simd_abi::__vec_ext<_Np>>;
+};
+
+template <class _Tp, int _Np>
+struct __mask_operations<_Tp, simd_abi::__vec_ext<_Np>> {
+  using _Mask = __mask_storage<_Tp, simd_abi::__vec_ext<_Np>>;
+};
+
 } // namespace parallelism_v2
 _LIBCPP_END_NAMESPACE_EXPERIMENTAL
 
diff --git a/libcxx/test/std/experimental/simd/simd.class/simd_alias.pass.cpp b/libcxx/test/std/experimental/simd/simd.class/simd_alias.pass.cpp
--- a/libcxx/test/std/experimental/simd/simd.class/simd_alias.pass.cpp
+++ b/libcxx/test/std/experimental/simd/simd.class/simd_alias.pass.cpp
@@ -12,6 +12,7 @@
 // template <class T, class Abi> class simd {
 // public:
 //   using value_type = T;
+//   using reference = see below;
 //   using mask_type = simd_mask<T, Abi>;
 //   using abi_type = Abi;
 // };
@@ -26,6 +27,7 @@
   template <class SimdAbi>
   void operator()() {
     static_assert(std::is_same_v<typename ex::simd<T, SimdAbi>::value_type, T>);
+    static_assert(is_simd_reference<typename ex::simd<T, SimdAbi>::reference>::value);
     static_assert(std::is_same_v<typename ex::simd<T, SimdAbi>::mask_type, ex::simd_mask<T, SimdAbi>>);
     static_assert(std::is_same_v<typename ex::simd<T, SimdAbi>::abi_type, SimdAbi>);
   }
diff --git a/libcxx/test/std/experimental/simd/simd.mask.class/simd_mask_alias.pass.cpp b/libcxx/test/std/experimental/simd/simd.mask.class/simd_mask_alias.pass.cpp
--- a/libcxx/test/std/experimental/simd/simd.mask.class/simd_mask_alias.pass.cpp
+++ b/libcxx/test/std/experimental/simd/simd.mask.class/simd_mask_alias.pass.cpp
@@ -12,6 +12,7 @@
 // template <class T, class Abi> class simd_mask {
 // public:
 //   using value_type = bool;
+//   using reference = see below;
 //   using simd_type = simd<T, Abi>;
 //   using abi_type = Abi;
 // };
@@ -26,6 +27,7 @@
   template <class SimdAbi>
   void operator()() {
     static_assert(std::is_same_v<typename ex::simd_mask<T, SimdAbi>::value_type, bool>);
+    static_assert(is_simd_reference<typename ex::simd_mask<T, SimdAbi>::reference>::value);
     static_assert(std::is_same_v<typename ex::simd_mask<T, SimdAbi>::simd_type, ex::simd<T, SimdAbi>>);
     static_assert(std::is_same_v<typename ex::simd_mask<T, SimdAbi>::abi_type, SimdAbi>);
   }
diff --git a/libcxx/test/std/experimental/simd/test_utils.h b/libcxx/test/std/experimental/simd/test_utils.h
--- a/libcxx/test/std/experimental/simd/test_utils.h
+++ b/libcxx/test/std/experimental/simd/test_utils.h
@@ -50,4 +50,10 @@
   return pow;
 }
 
+template <class T>
+struct is_simd_reference : std::false_type {};
+
+template <class T, class S, class V>
+struct is_simd_reference<ex::__reference<T, S, V>> : std::true_type {};
+
 #endif // TEST_UTIL_H