diff --git a/include/boost/endian/buffers.hpp b/include/boost/endian/buffers.hpp
index d8be041..51fe4f8 100644
--- a/include/boost/endian/buffers.hpp
+++ b/include/boost/endian/buffers.hpp
@@ -37,13 +37,9 @@
 #include <boost/config.hpp>
 #include <boost/config/workaround.hpp>
 #include <boost/predef/other/endian.h>
+#include <boost/endian/detail/endian_store.hpp>
 #include <boost/endian/detail/endian_load.hpp>
 #include <boost/endian/conversion.hpp>
-#include <boost/type_traits/is_signed.hpp>
-#include <boost/type_traits/make_unsigned.hpp>
-#include <boost/type_traits/conditional.hpp>
-#include <boost/type_traits/is_integral.hpp>
-#include <boost/type_traits/type_identity.hpp>
 #include <boost/cstdint.hpp>
 #include <boost/static_assert.hpp>
 #include <boost/core/scoped_enum.hpp>
@@ -221,65 +217,6 @@ namespace endian
   namespace detail
   {
 
-    // Unrolled loops for loading and storing streams of bytes.
-
-    template <typename T, std::size_t n_bytes,
-      bool sign=boost::is_signed<T>::value >
-    struct unrolled_byte_loops
-    {
-      typedef unrolled_byte_loops<T, n_bytes - 1, sign> next;
-
-      // shifting a negative number is flagged by -fsanitize=undefined
-      // so use the corresponding unsigned type for the shifts
-
-      typedef typename boost::conditional<
-        boost::is_integral<T>::value,
-        boost::make_unsigned<T>, boost::type_identity<T> >::type::type U;
-
-      static T load_big(const unsigned char* bytes) BOOST_NOEXCEPT
-        { return static_cast<T>(*(bytes - 1) | (static_cast<U>(next::load_big(bytes - 1)) << 8)); }
-      static T load_little(const unsigned char* bytes) BOOST_NOEXCEPT
-        { return static_cast<T>(*bytes | (static_cast<U>(next::load_little(bytes + 1)) << 8)); }
-
-      static void store_big(unsigned char* bytes, T value) BOOST_NOEXCEPT
-        {
-          *(bytes - 1) = static_cast<unsigned char>(value);
-          next::store_big(bytes - 1, static_cast<T>(static_cast<U>(value) >> 8));
-        }
-      static void store_little(unsigned char* bytes, T value) BOOST_NOEXCEPT
-        {
-          *bytes = static_cast<unsigned char>(value);
-          next::store_little(bytes + 1, static_cast<T>(static_cast<U>(value) >> 8));
-        }
-    };
-
-    template <typename T>
-    struct unrolled_byte_loops<T, 1, false>
-    {
-      static T load_big(const unsigned char* bytes) BOOST_NOEXCEPT
-        { return *(bytes - 1); }
-      static T load_little(const unsigned char* bytes) BOOST_NOEXCEPT
-        { return *bytes; }
-      static void store_big(unsigned char* bytes, T value) BOOST_NOEXCEPT
-        { *(bytes - 1) = static_cast<unsigned char>(value); }
-      static void store_little(unsigned char* bytes, T value) BOOST_NOEXCEPT
-        { *bytes = static_cast<unsigned char>(value); }
-
-    };
-
-    template <typename T>
-    struct unrolled_byte_loops<T, 1, true>
-    {
-      static T load_big(const unsigned char* bytes) BOOST_NOEXCEPT
-        { return *reinterpret_cast<const signed char*>(bytes - 1); }
-      static T load_little(const unsigned char* bytes) BOOST_NOEXCEPT
-        { return *reinterpret_cast<const signed char*>(bytes); }
-      static void store_big(unsigned char* bytes, T value)  BOOST_NOEXCEPT
-        { *(bytes - 1) = static_cast<unsigned char>(value); }
-      static void store_little(unsigned char* bytes, T value) BOOST_NOEXCEPT
-        { *bytes = static_cast<unsigned char>(value); }
-    };
-
     template <typename T, std::size_t n_bytes>
     inline
     T load_big_endian(const void* bytes) BOOST_NOEXCEPT
@@ -298,37 +235,14 @@ namespace endian
     inline
     void store_big_endian(void* bytes, T value) BOOST_NOEXCEPT
     {
-#     if defined(__x86_64__) || defined(_M_X64) || defined(__i386) || defined(_M_IX86)
-      // All major x86 compilers elide this test and optimize out memcpy
-      // (the x86 architecture allows unaligned loads, but -fsanitize=undefined does not)
-      if (sizeof(T) == n_bytes)
-      {
-        endian::native_to_big_inplace(value);
-        std::memcpy( bytes, &value, sizeof(T) );
-        return;
-      }
-#     endif
-      unrolled_byte_loops<T, n_bytes>::store_big
-        (static_cast<unsigned char*>(bytes) + n_bytes, value);
+        endian::endian_store<T, n_bytes, order::big>( value, static_cast<unsigned char*>( bytes ) );
     }
 
     template <typename T, std::size_t n_bytes>
     inline
     void store_little_endian(void* bytes, T value) BOOST_NOEXCEPT
     {
-#     if defined(__x86_64__) || defined(_M_X64) || defined(__i386) || defined(_M_IX86)
-      // All major x86 compilers elide this test and optimize out memcpy
-      // (the x86 architecture allows unaligned loads, but -fsanitize=undefined does not)
-      if (sizeof(T) == n_bytes)
-      {
-        // if we ever extend the #ifdef to non-x86:
-        //   endian::native_to_little_inplace(value);
-        std::memcpy( bytes, &value, sizeof(T) );
-        return;
-      }
-#     endif
-      unrolled_byte_loops<T, n_bytes>::store_little
-        (static_cast<unsigned char*>(bytes), value);
+        endian::endian_store<T, n_bytes, order::little>( value, static_cast<unsigned char*>( bytes ) );
     }
 
   } // namespace detail
diff --git a/include/boost/endian/detail/endian_load.hpp b/include/boost/endian/detail/endian_load.hpp
index 323ce5e..c292e77 100644
--- a/include/boost/endian/detail/endian_load.hpp
+++ b/include/boost/endian/detail/endian_load.hpp
@@ -6,6 +6,7 @@
 // Distributed under the Boost Software License, Version 1.0.
 // http://www.boost.org/LICENSE_1_0.txt
 
+#include <boost/endian/detail/integral_by_size.hpp>
 #include <boost/endian/conversion.hpp>
 #include <boost/type_traits/is_signed.hpp>
 #include <cstddef>
@@ -34,30 +35,6 @@ inline T endian_load( unsigned char const * p ) BOOST_NOEXCEPT
 namespace detail
 {
 
-template<std::size_t N> struct integral_by_size
-{
-};
-
-template<> struct integral_by_size<1>
-{
-    typedef uint8_t type;
-};
-
-template<> struct integral_by_size<2>
-{
-    typedef uint16_t type;
-};
-
-template<> struct integral_by_size<4>
-{
-    typedef uint32_t type;
-};
-
-template<> struct integral_by_size<8>
-{
-    typedef uint64_t type;
-};
-
 // same endianness, same size
 
 template<class T, std::size_t N, BOOST_SCOPED_ENUM(order) O> struct endian_load_impl<T, N, O, N, O>
diff --git a/include/boost/endian/detail/endian_store.hpp b/include/boost/endian/detail/endian_store.hpp
new file mode 100644
index 0000000..33501b6
--- /dev/null
+++ b/include/boost/endian/detail/endian_store.hpp
@@ -0,0 +1,197 @@
+#ifndef BOOST_ENDIAN_DETAIL_ENDIAN_STORE_HPP_INCLUDED
+#define BOOST_ENDIAN_DETAIL_ENDIAN_STORE_HPP_INCLUDED
+
+// Copyright 2019 Peter Dimov
+//
+// Distributed under the Boost Software License, Version 1.0.
+// http://www.boost.org/LICENSE_1_0.txt
+
+#include <boost/endian/detail/integral_by_size.hpp>
+#include <boost/endian/conversion.hpp>
+#include <cstddef>
+#include <cstring>
+
+namespace boost
+{
+namespace endian
+{
+
+namespace detail
+{
+
+template<class T, std::size_t N1, BOOST_SCOPED_ENUM(order) O1, std::size_t N2, BOOST_SCOPED_ENUM(order) O2> struct endian_store_impl
+{
+};
+
+} // namespace detail
+
+template<class T, std::size_t N, BOOST_SCOPED_ENUM(order) Order>
+inline void endian_store( T const & v, unsigned char * p ) BOOST_NOEXCEPT
+{
+    return detail::endian_store_impl<T, sizeof(T), endian::order::native, N, Order>()( v, p );
+}
+
+namespace detail
+{
+
+// same endianness, same size
+
+template<class T, std::size_t N, BOOST_SCOPED_ENUM(order) O> struct endian_store_impl<T, N, O, N, O>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        std::memcpy( p, &v, N );
+    }
+};
+
+// same size, reverse endianness
+
+template<class T, std::size_t N, BOOST_SCOPED_ENUM(order) O1, BOOST_SCOPED_ENUM(order) O2> struct endian_store_impl<T, N, O1, N, O2>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        typename integral_by_size<N>::type tmp;
+        std::memcpy( &tmp, &v, N );
+
+        endian_reverse_inplace( tmp );
+
+        std::memcpy( p, &tmp, N );
+    }
+};
+
+// truncating store 4 -> 3
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 4, Order, 3, order::little>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 4 ];
+        endian::endian_store<T, 4, order::little>( v, tmp );
+
+        p[0] = tmp[0];
+        p[1] = tmp[1];
+        p[2] = tmp[2];
+    }
+};
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 4, Order, 3, order::big>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 4 ];
+        endian::endian_store<T, 4, order::big>( v, tmp );
+
+        p[0] = tmp[1];
+        p[1] = tmp[2];
+        p[2] = tmp[3];
+    }
+};
+
+// truncating store 8 -> 5
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 8, Order, 5, order::little>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 8 ];
+        endian::endian_store<T, 8, order::little>( v, tmp );
+
+        p[0] = tmp[0];
+        p[1] = tmp[1];
+        p[2] = tmp[2];
+        p[3] = tmp[3];
+        p[4] = tmp[4];
+    }
+};
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 8, Order, 5, order::big>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 8 ];
+        endian::endian_store<T, 8, order::big>( v, tmp );
+
+        p[0] = tmp[3];
+        p[1] = tmp[4];
+        p[2] = tmp[5];
+        p[3] = tmp[6];
+        p[4] = tmp[7];
+    }
+};
+
+// truncating store 8 -> 6
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 8, Order, 6, order::little>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 8 ];
+        endian::endian_store<T, 8, order::little>( v, tmp );
+
+        p[0] = tmp[0];
+        p[1] = tmp[1];
+        p[2] = tmp[2];
+        p[3] = tmp[3];
+        p[4] = tmp[4];
+        p[5] = tmp[5];
+    }
+};
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 8, Order, 6, order::big>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 8 ];
+        endian::endian_store<T, 8, order::big>( v, tmp );
+
+        p[0] = tmp[2];
+        p[1] = tmp[3];
+        p[2] = tmp[4];
+        p[3] = tmp[5];
+        p[4] = tmp[6];
+        p[5] = tmp[7];
+    }
+};
+
+// truncating store 8 -> 7
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 8, Order, 7, order::little>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 8 ];
+        endian::endian_store<T, 8, order::little>( v, tmp );
+
+        p[0] = tmp[0];
+        p[1] = tmp[1];
+        p[2] = tmp[2];
+        p[3] = tmp[3];
+        p[4] = tmp[4];
+        p[5] = tmp[5];
+        p[6] = tmp[6];
+    }
+};
+
+template<class T, BOOST_SCOPED_ENUM(order) Order> struct endian_store_impl<T, 8, Order, 7, order::big>
+{
+    inline void operator()( T const & v, unsigned char * p ) const BOOST_NOEXCEPT
+    {
+        unsigned char tmp[ 8 ];
+        endian::endian_store<T, 8, order::big>( v, tmp );
+
+        p[0] = tmp[1];
+        p[1] = tmp[2];
+        p[2] = tmp[3];
+        p[3] = tmp[4];
+        p[4] = tmp[5];
+        p[5] = tmp[6];
+        p[6] = tmp[7];
+    }
+};
+
+} // namespace detail
+
+} // namespace endian
+} // namespace boost
+
+#endif  // BOOST_ENDIAN_DETAIL_ENDIAN_STORE_HPP_INCLUDED
diff --git a/include/boost/endian/detail/integral_by_size.hpp b/include/boost/endian/detail/integral_by_size.hpp
new file mode 100644
index 0000000..b0fd885
--- /dev/null
+++ b/include/boost/endian/detail/integral_by_size.hpp
@@ -0,0 +1,47 @@
+#ifndef BOOST_ENDIAN_DETAIL_INTEGRAL_BY_SIZE_HPP_INCLUDED
+#define BOOST_ENDIAN_DETAIL_INTEGRAL_BY_SIZE_HPP_INCLUDED
+
+// Copyright 2019 Peter Dimov
+//
+// Distributed under the Boost Software License, Version 1.0.
+// http://www.boost.org/LICENSE_1_0.txt
+
+#include <boost/cstdint.hpp>
+#include <cstddef>
+
+namespace boost
+{
+namespace endian
+{
+namespace detail
+{
+
+template<std::size_t N> struct integral_by_size
+{
+};
+
+template<> struct integral_by_size<1>
+{
+    typedef uint8_t type;
+};
+
+template<> struct integral_by_size<2>
+{
+    typedef uint16_t type;
+};
+
+template<> struct integral_by_size<4>
+{
+    typedef uint32_t type;
+};
+
+template<> struct integral_by_size<8>
+{
+    typedef uint64_t type;
+};
+
+} // namespace detail
+} // namespace endian
+} // namespace boost
+
+#endif  // BOOST_ENDIAN_DETAIL_INTEGRAL_BY_SIZE_HPP_INCLUDED