哇，我简直不敢相信我在这里读到的一些答案。实际上汇编中有一条指令比其他任何程序都快。bswap。你可以简单地写一个这样的函数…

__declspec(naked) uint32_t EndianSwap(uint32 value)
{
    __asm
    {
        mov eax, dword ptr[esp + 4]
        bswap eax
        ret
    }
}

它比之前提到的内在函数要快得多。我把它们拆开看了看。上面的函数没有序言/尾声，因此实际上没有任何开销。

unsigned long _byteswap_ulong(unsigned long value);

做16位同样容易，除了你会使用xchg al，啊。Bswap仅适用于32位寄存器。

64位有点棘手，但也不过分。比上面所有带有循环和模板的例子都要好得多。

这里有一些注意事项……首先，bswap只能在80x486以上的CPU上使用。有人打算在386上运行吗?!?如果是这样，你仍然可以用…替换bswap。

mov ebx, eax
shr ebx, 16
xchg al, ah
xchg bl, bh
shl eax, 16
or eax, ebx

内联汇编也只能在Visual Studio的x86代码中使用。裸函数不能内衬，而且在x64版本中也不可用。对于那个实例，你必须使用编译器的内在函数。

大多数平台都有一个系统头文件，提供了有效的byteswap函数。在Linux上是<end .h>。你可以用c++很好地包装它:

#include <iostream>

#include <endian.h>

template<size_t N> struct SizeT {};

#define BYTESWAPS(bits) \
template<class T> inline T htobe(T t, SizeT<bits / 8>) { return htobe ## bits(t); } \
template<class T> inline T htole(T t, SizeT<bits / 8>) { return htole ## bits(t); } \
template<class T> inline T betoh(T t, SizeT<bits / 8>) { return be ## bits ## toh(t); } \
template<class T> inline T letoh(T t, SizeT<bits / 8>) { return le ## bits ## toh(t); }

BYTESWAPS(16)
BYTESWAPS(32)
BYTESWAPS(64)

#undef BYTESWAPS

template<class T> inline T htobe(T t) { return htobe(t, SizeT<sizeof t>()); }
template<class T> inline T htole(T t) { return htole(t, SizeT<sizeof t>()); }
template<class T> inline T betoh(T t) { return betoh(t, SizeT<sizeof t>()); }
template<class T> inline T letoh(T t) { return letoh(t, SizeT<sizeof t>()); }

int main()
{
    std::cout << std::hex;
    std::cout << htobe(static_cast<unsigned short>(0xfeca)) << '\n';
    std::cout << htobe(0xafbeadde) << '\n';

    // Use ULL suffix to specify integer constant as unsigned long long 
    std::cout << htobe(0xfecaefbeafdeedfeULL) << '\n';
}

输出:

cafe
deadbeaf
feeddeafbeefcafe

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);

实现优化器友好的未对齐非就地末端访问器的可移植技术。它们处理每个编译器、每个边界对齐和每个字节排序。这些未对齐的例程被补充或讨论，取决于本机的端序和对齐方式。部分列出，但你懂的。BO*是基于本机字节排序的常数值。

uint32_t sw_get_uint32_1234(pu32)
uint32_1234 *pu32;
{
  union {
    uint32_1234 u32_1234;
    uint32_t u32;
  } bou32;
  bou32.u32_1234[0] = (*pu32)[BO32_0];
  bou32.u32_1234[1] = (*pu32)[BO32_1];
  bou32.u32_1234[2] = (*pu32)[BO32_2];
  bou32.u32_1234[3] = (*pu32)[BO32_3];
  return(bou32.u32);
}

void sw_set_uint32_1234(pu32, u32)
uint32_1234 *pu32;
uint32_t u32;
{
  union {
    uint32_1234 u32_1234;
    uint32_t u32;
  } bou32;
  bou32.u32 = u32;
  (*pu32)[BO32_0] = bou32.u32_1234[0];
  (*pu32)[BO32_1] = bou32.u32_1234[1];
  (*pu32)[BO32_2] = bou32.u32_1234[2];
  (*pu32)[BO32_3] = bou32.u32_1234[3];
}

#if HAS_SW_INT64
int64 sw_get_int64_12345678(pi64)
int64_12345678 *pi64;
{
  union {
    int64_12345678 i64_12345678;
    int64 i64;
  } boi64;
  boi64.i64_12345678[0] = (*pi64)[BO64_0];
  boi64.i64_12345678[1] = (*pi64)[BO64_1];
  boi64.i64_12345678[2] = (*pi64)[BO64_2];
  boi64.i64_12345678[3] = (*pi64)[BO64_3];
  boi64.i64_12345678[4] = (*pi64)[BO64_4];
  boi64.i64_12345678[5] = (*pi64)[BO64_5];
  boi64.i64_12345678[6] = (*pi64)[BO64_6];
  boi64.i64_12345678[7] = (*pi64)[BO64_7];
  return(boi64.i64);
}
#endif

int32_t sw_get_int32_3412(pi32)
int32_3412 *pi32;
{
  union {
    int32_3412 i32_3412;
    int32_t i32;
  } boi32;
  boi32.i32_3412[2] = (*pi32)[BO32_0];
  boi32.i32_3412[3] = (*pi32)[BO32_1];
  boi32.i32_3412[0] = (*pi32)[BO32_2];
  boi32.i32_3412[1] = (*pi32)[BO32_3];
  return(boi32.i32);
}

void sw_set_int32_3412(pi32, i32)
int32_3412 *pi32;
int32_t i32;
{
  union {
    int32_3412 i32_3412;
    int32_t i32;
  } boi32;
  boi32.i32 = i32;
  (*pi32)[BO32_0] = boi32.i32_3412[2];
  (*pi32)[BO32_1] = boi32.i32_3412[3];
  (*pi32)[BO32_2] = boi32.i32_3412[0];
  (*pi32)[BO32_3] = boi32.i32_3412[1];
}

uint32_t sw_get_uint32_3412(pu32)
uint32_3412 *pu32;
{
  union {
    uint32_3412 u32_3412;
    uint32_t u32;
  } bou32;
  bou32.u32_3412[2] = (*pu32)[BO32_0];
  bou32.u32_3412[3] = (*pu32)[BO32_1];
  bou32.u32_3412[0] = (*pu32)[BO32_2];
  bou32.u32_3412[1] = (*pu32)[BO32_3];
  return(bou32.u32);
}

void sw_set_uint32_3412(pu32, u32)
uint32_3412 *pu32;
uint32_t u32;
{
  union {
    uint32_3412 u32_3412;
    uint32_t u32;
  } bou32;
  bou32.u32 = u32;
  (*pu32)[BO32_0] = bou32.u32_3412[2];
  (*pu32)[BO32_1] = bou32.u32_3412[3];
  (*pu32)[BO32_2] = bou32.u32_3412[0];
  (*pu32)[BO32_3] = bou32.u32_3412[1];
}

float sw_get_float_1234(pf)
float_1234 *pf;
{
  union {
    float_1234 f_1234;
    float f;
  } bof;
  bof.f_1234[0] = (*pf)[BO32_0];
  bof.f_1234[1] = (*pf)[BO32_1];
  bof.f_1234[2] = (*pf)[BO32_2];
  bof.f_1234[3] = (*pf)[BO32_3];
  return(bof.f);
}

void sw_set_float_1234(pf, f)
float_1234 *pf;
float f;
{
  union {
    float_1234 f_1234;
    float f;
  } bof;
  bof.f = (float)f;
  (*pf)[BO32_0] = bof.f_1234[0];
  (*pf)[BO32_1] = bof.f_1234[1];
  (*pf)[BO32_2] = bof.f_1234[2];
  (*pf)[BO32_3] = bof.f_1234[3];
}

double sw_get_double_12345678(pd)
double_12345678 *pd;
{
  union {
    double_12345678 d_12345678;
    double d;
  } bod;
  bod.d_12345678[0] = (*pd)[BO64_0];
  bod.d_12345678[1] = (*pd)[BO64_1];
  bod.d_12345678[2] = (*pd)[BO64_2];
  bod.d_12345678[3] = (*pd)[BO64_3];
  bod.d_12345678[4] = (*pd)[BO64_4];
  bod.d_12345678[5] = (*pd)[BO64_5];
  bod.d_12345678[6] = (*pd)[BO64_6];
  bod.d_12345678[7] = (*pd)[BO64_7];
  return(bod.d);
}

void sw_set_double_12345678(pd, d)
double_12345678 *pd;
double d;
{
  union {
    double_12345678 d_12345678;
    double d;
  } bod;
  bod.d = d;
  (*pd)[BO64_0] = bod.d_12345678[0];
  (*pd)[BO64_1] = bod.d_12345678[1];
  (*pd)[BO64_2] = bod.d_12345678[2];
  (*pd)[BO64_3] = bod.d_12345678[3];
  (*pd)[BO64_4] = bod.d_12345678[4];
  (*pd)[BO64_5] = bod.d_12345678[5];
  (*pd)[BO64_6] = bod.d_12345678[6];
  (*pd)[BO64_7] = bod.d_12345678[7];
}

如果不与访问器一起使用，这些类型def的好处是会引发编译器错误，从而减少被遗忘的访问器错误。

typedef char int8_1[1], uint8_1[1];

typedef char int16_12[2], uint16_12[2]; /* little endian */
typedef char int16_21[2], uint16_21[2]; /* big endian */

typedef char int24_321[3], uint24_321[3]; /* Alpha Micro, PDP-11 */

typedef char int32_1234[4], uint32_1234[4]; /* little endian */
typedef char int32_3412[4], uint32_3412[4]; /* Alpha Micro, PDP-11 */
typedef char int32_4321[4], uint32_4321[4]; /* big endian */

typedef char int64_12345678[8], uint64_12345678[8]; /* little endian */
typedef char int64_34128756[8], uint64_34128756[8]; /* Alpha Micro, PDP-11 */
typedef char int64_87654321[8], uint64_87654321[8]; /* big endian */

typedef char float_1234[4]; /* little endian */
typedef char float_3412[4]; /* Alpha Micro, PDP-11 */
typedef char float_4321[4]; /* big endian */

typedef char double_12345678[8]; /* little endian */
typedef char double_78563412[8]; /* Alpha Micro? */
typedef char double_87654321[8]; /* big endian */

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。