c++20无分支版本，现在std::endian已经存在，但在c++23之前增加了std::byteswap

#include <bit>
#include <type_traits>
#include <concepts>
#include <array>
#include <cstring>
#include <iostream>
#include <bitset>

template <int LEN, int OFF=LEN/2>
class do_swap
{
    // FOR 8 bytes:
    // LEN=8 (LEN/2==4)       <H><G><F><E><D><C><B><A>
    // OFF=4: FROM=0, TO=7 => [A]<G><F><E><D><C><B>[H]
    // OFF=3: FROM=1, TO=6 => [A][B]<F><E><D><C>[G][H]
    // OFF=2: FROM=2, TO=5 => [A][B][C]<E><D>[F][G][H]
    // OFF=1: FROM=3, TO=4 => [A][B][C][D][E][F][G][H]
    // OFF=0: FROM=4, TO=3 => DONE
public:
    enum consts {FROM=LEN/2-OFF, TO=(LEN-1)-FROM};
    using NXT=do_swap<LEN, OFF-1>;
// flip the first and last for the current iteration's range
    static void flip(std::array<std::byte, LEN>& b)
    {
        std::byte tmp=b[FROM];
        b[FROM]=b[TO];
        b[TO]=tmp;
        NXT::flip(b);
    }
};
template <int LEN>
class do_swap<LEN, 0> // STOP the template recursion
{
public:
    static void flip(std::array<std::byte, LEN>&)
    {
    }
};

template<std::integral T, std::endian TO, std::endian FROM=std::endian::native>
        requires ((TO==std::endian::big) || (TO==std::endian::little))
              && ((FROM==std::endian::big) || (FROM==std::endian::little))
class endian_swap
{
public:
    enum consts {BYTE_COUNT=sizeof(T)};
    static T cvt(const T integral)
    {
    // if FROM and TO are the same -- nothing to do
        if (TO==FROM)
        {
                return integral;
        }

    // endian::big --> endian::little is the same as endian::little --> endian::big
    // the bytes have to be reversed
    // memcpy seems to be the most supported way to do byte swaps in a defined way
        std::array<std::byte, BYTE_COUNT> bytes;
        std::memcpy(&bytes, &integral, BYTE_COUNT);
        do_swap<BYTE_COUNT>::flip(bytes);
        T ret;
        std::memcpy(&ret, &bytes, BYTE_COUNT);
        return ret;
    }
};

std::endian big()
{
    return std::endian::big;
}

std::endian little()
{
    return std::endian::little;
}

std::endian native()
{
    return std::endian::native;
}

long long swap_to_big(long long x)
{
    return endian_swap<long long, std::endian::big>::cvt(x);
}

long long swap_to_little(long long x)
{
    return endian_swap<long long, std::endian::little>::cvt(x);
}

void show(std::string label, long long x)
{
    std::cout << label << "\t: " << std::bitset<64>(x) << " (" << x << ")" << std::endl;
}

int main(int argv, char ** argc)
{
    long long init=0xF8FCFEFF7F3F1F0;
    long long to_big=swap_to_big(init);
    long long to_little=swap_to_little(init);
    show("Init", init);
    show(">big", to_big);
    show(">little", to_little);
}

摘自Rob Pike的《字节顺序谬误》:

假设数据流有一个小端编码的32位整数。下面是如何提取它(假设无符号字节):

i = (data[0]<<0) | (data[1]<<8) | (data[2]<<16) | ((unsigned)data[3]<<24);

如果它是big-endian，下面是如何提取它:

i = (data[3]<<0) | (data[2]<<8) | (data[1]<<16) | ((unsigned)data[0]<<24);

TL;DR:不要担心你的平台原生顺序，重要的是你从中读取的流的字节顺序，你最好希望它是定义良好的。

注1:这里int和unsigned int是32位，否则类型可能需要调整。

注2:最后一个字节必须在移位前显式转换为unsigned，因为默认情况下它被提升为int，移位24位意味着操作符号位，这是未定义行为。

请注意，至少对于Windows, htonl()比它们的内在对应_byteswap_ulong()慢得多。前者是对ws2_32.dll的一个DLL库调用，后者是一条BSWAP汇编指令。因此，如果你正在编写一些依赖于平台的代码，为了提高速度，最好使用intrinsic:

#define htonl(x) _byteswap_ulong(x)

这对于。png图像处理尤其重要，其中所有整数都保存在大端格式中，并说明“One can use htonl()…”{用来降低典型Windows程序的速度，如果你没有准备好}。

如果您这样做是为了在不同平台之间传输数据，请查看ntoh和hton函数。

从大端序到小端序的过程与从小端序到大端序的过程是一样的。

下面是一些示例代码:

void swapByteOrder(unsigned short& us)
{
    us = (us >> 8) |
         (us << 8);
}

void swapByteOrder(unsigned int& ui)
{
    ui = (ui >> 24) |
         ((ui<<8) & 0x00FF0000) |
         ((ui>>8) & 0x0000FF00) |
         (ui << 24);
}

void swapByteOrder(unsigned long long& ull)
{
    ull = (ull >> 56) |
          ((ull<<40) & 0x00FF000000000000) |
          ((ull<<24) & 0x0000FF0000000000) |
          ((ull<<8) & 0x000000FF00000000) |
          ((ull>>8) & 0x00000000FF000000) |
          ((ull>>24) & 0x0000000000FF0000) |
          ((ull>>40) & 0x000000000000FF00) |
          (ull << 56);
}

我有这个代码，允许我从HOST_ENDIAN_ORDER(无论它是什么)转换为LITTLE_ENDIAN_ORDER或BIG_ENDIAN_ORDER。我使用一个模板，所以如果我试图从HOST_ENDIAN_ORDER转换为LITTLE_ENDIAN_ORDER，他们恰好是相同的机器为我编译，不会生成任何代码。

下面是带有注释的代码:

// We define some constant for little, big and host endianess. Here I use 
// BOOST_LITTLE_ENDIAN/BOOST_BIG_ENDIAN to check the host indianess. If you
// don't want to use boost you will have to modify this part a bit.
enum EEndian
{
  LITTLE_ENDIAN_ORDER,
  BIG_ENDIAN_ORDER,
#if defined(BOOST_LITTLE_ENDIAN)
  HOST_ENDIAN_ORDER = LITTLE_ENDIAN_ORDER
#elif defined(BOOST_BIG_ENDIAN)
  HOST_ENDIAN_ORDER = BIG_ENDIAN_ORDER
#else
#error "Impossible de determiner l'indianness du systeme cible."
#endif
};

// this function swap the bytes of values given it's size as a template
// parameter (could sizeof be used?).
template <class T, unsigned int size>
inline T SwapBytes(T value)
{
  union
  {
     T value;
     char bytes[size];
  } in, out;

  in.value = value;

  for (unsigned int i = 0; i < size / 2; ++i)
  {
     out.bytes[i] = in.bytes[size - 1 - i];
     out.bytes[size - 1 - i] = in.bytes[i];
  }

  return out.value;
}

// Here is the function you will use. Again there is two compile-time assertion
// that use the boost librarie. You could probably comment them out, but if you
// do be cautious not to use this function for anything else than integers
// types. This function need to be calles like this :
//
//     int x = someValue;
//     int i = EndianSwapBytes<HOST_ENDIAN_ORDER, BIG_ENDIAN_ORDER>(x);
//
template<EEndian from, EEndian to, class T>
inline T EndianSwapBytes(T value)
{
  // A : La donnée à swapper à une taille de 2, 4 ou 8 octets
  BOOST_STATIC_ASSERT(sizeof(T) == 2 || sizeof(T) == 4 || sizeof(T) == 8);

  // A : La donnée à swapper est d'un type arithmetic
  BOOST_STATIC_ASSERT(boost::is_arithmetic<T>::value);

  // Si from et to sont du même type on ne swap pas.
  if (from == to)
     return value;

  return SwapBytes<T, sizeof(T)>(value);
}