简单地说:

#include <climits>

template <typename T>
T swap_endian(T u)
{
    static_assert (CHAR_BIT == 8, "CHAR_BIT != 8");

    union
    {
        T u;
        unsigned char u8[sizeof(T)];
    } source, dest;

    source.u = u;

    for (size_t k = 0; k < sizeof(T); k++)
        dest.u8[k] = source.u8[sizeof(T) - k - 1];

    return dest.u;
}

用法:swap_endian < uint32_t >(42)。

如果您采用反转单词中位序的常见模式，并剔除每个字节中反转位的部分，那么您将只剩下反转单词中的字节的部分。对于64位:

x = ((x & 0x00000000ffffffff) << 32) ^ ((x >> 32) & 0x00000000ffffffff);
x = ((x & 0x0000ffff0000ffff) << 16) ^ ((x >> 16) & 0x0000ffff0000ffff);
x = ((x & 0x00ff00ff00ff00ff) <<  8) ^ ((x >>  8) & 0x00ff00ff00ff00ff);

编译器应该清除多余的位屏蔽操作(我把它们留在了突出显示模式)，但如果它没有，你可以这样重写第一行:

x = ( x                       << 32) ^  (x >> 32);

在大多数架构上，这通常应该简化为一条旋转指令(忽略整个操作可能是一条指令)。

在RISC处理器上，大而复杂的常量可能会导致编译困难。不过，您可以简单地计算前一个的每个常数。像这样:

uint64_t k = 0x00000000ffffffff; /* compiler should know a trick for this */
x = ((x & k) << 32) ^ ((x >> 32) & k);
k ^= k << 16;
x = ((x & k) << 16) ^ ((x >> 16) & k);
k ^= k << 8;
x = ((x & k) <<  8) ^ ((x >>  8) & k);

如果你愿意，你可以把它写成一个循环。这样做效率不高，只是为了好玩:

int i = sizeof(x) * CHAR_BIT / 2;
uintmax_t k = (1 << i) - 1;
while (i >= 8)
{
    x = ((x & k) << i) ^ ((x >> i) & k);
    i >>= 1;
    k ^= k << i;
}

为了完整起见，这里是第一种形式的简化32位版本:

x = ( x               << 16) ^  (x >> 16);
x = ((x & 0x00ff00ff) <<  8) ^ ((x >>  8) & 0x00ff00ff);

和在C中一样:

short big = 0xdead;
short little = (((big & 0xff)<<8) | ((big & 0xff00)>>8));

您还可以声明一个无符号字符的向量，将输入值memcpy放入其中，将字节反向转换为另一个向量，然后将字节memcpy取出，但这将花费比旋转位长几个数量级的时间，特别是对于64位值。

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);
}

我从这篇文章中得到了一些建议，并把它们放在一起形成了这个:

#include <boost/type_traits.hpp>
#include <boost/static_assert.hpp>
#include <boost/detail/endian.hpp>
#include <stdexcept>
#include <cstdint>

enum endianness
{
    little_endian,
    big_endian,
    network_endian = big_endian,
    
    #if defined(BOOST_LITTLE_ENDIAN)
        host_endian = little_endian
    #elif defined(BOOST_BIG_ENDIAN)
        host_endian = big_endian
    #else
        #error "unable to determine system endianness"
    #endif
};

namespace detail {

template<typename T, size_t sz>
struct swap_bytes
{
    inline T operator()(T val)
    {
        throw std::out_of_range("data size");
    }
};

template<typename T>
struct swap_bytes<T, 1>
{
    inline T operator()(T val)
    {
        return val;
    }
};

template<typename T>
struct swap_bytes<T, 2>
{
    inline T operator()(T val)
    {
        return ((((val) >> 8) & 0xff) | (((val) & 0xff) << 8));
    }
};

template<typename T>
struct swap_bytes<T, 4>
{
    inline T operator()(T val)
    {
        return ((((val) & 0xff000000) >> 24) |
                (((val) & 0x00ff0000) >>  8) |
                (((val) & 0x0000ff00) <<  8) |
                (((val) & 0x000000ff) << 24));
    }
};

template<>
struct swap_bytes<float, 4>
{
    inline float operator()(float val)
    {
        uint32_t mem =swap_bytes<uint32_t, sizeof(uint32_t)>()(*(uint32_t*)&val);
        return *(float*)&mem;
    }
};

template<typename T>
struct swap_bytes<T, 8>
{
    inline T operator()(T val)
    {
        return ((((val) & 0xff00000000000000ull) >> 56) |
                (((val) & 0x00ff000000000000ull) >> 40) |
                (((val) & 0x0000ff0000000000ull) >> 24) |
                (((val) & 0x000000ff00000000ull) >> 8 ) |
                (((val) & 0x00000000ff000000ull) << 8 ) |
                (((val) & 0x0000000000ff0000ull) << 24) |
                (((val) & 0x000000000000ff00ull) << 40) |
                (((val) & 0x00000000000000ffull) << 56));
    }
};

template<>
struct swap_bytes<double, 8>
{
    inline double operator()(double val)
    {
        uint64_t mem =swap_bytes<uint64_t, sizeof(uint64_t)>()(*(uint64_t*)&val);
        return *(double*)&mem;
    }
};

template<endianness from, endianness to, class T>
struct do_byte_swap
{
    inline T operator()(T value)
    {
        return swap_bytes<T, sizeof(T)>()(value);
    }
};
// specialisations when attempting to swap to the same endianess
template<class T> struct do_byte_swap<little_endian, little_endian, T> { inline T operator()(T value) { return value; } };
template<class T> struct do_byte_swap<big_endian,    big_endian,    T> { inline T operator()(T value) { return value; } };

} // namespace detail

template<endianness from, endianness to, class T>
inline T byte_swap(T value)
{
    // ensure the data is only 1, 2, 4 or 8 bytes
    BOOST_STATIC_ASSERT(sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4 || sizeof(T) == 8);
    // ensure we're only swapping arithmetic types
    BOOST_STATIC_ASSERT(boost::is_arithmetic<T>::value);

    return detail::do_byte_swap<from, to, T>()(value);
}

然后你可以这样使用它:

// swaps val from host-byte-order to network-byte-order
auto swapped = byte_swap<host_endian, network_endian>(val);

反之亦然

// swap a value received from the network into host-byte-order
auto val = byte_swap<network_endian, host_endian>(val_from_network);

如果你这样做是为了网络/主机兼容性，你应该使用:

ntohl() //Network to Host byte order (Long)
htonl() //Host to Network byte order (Long)

ntohs() //Network to Host byte order (Short)
htons() //Host to Network byte order (Short)

如果是出于其他原因，这里提供的byte_swap解决方案之一也可以很好地工作。