How do you write (portably) reverse network byte order?

Here's a template based version:

#include <iostream>
#include <iomanip>

enum endianness_t {
   BIG,         // 0x44332211  => 0x44 0x33 0x22 0x11
   LITTLE,      // 0x44332211  => 0x11 0x22 0x33 0x44
  UNKNOWN
};

const uint32_t test_value    = 0x44332211;
const bool is_little_endian  = (((char *)&test_value)[0] == 0x11) && (((char *)&test_value)[1] == 0x22);
const bool is_big_endian     = (((char *)&test_value)[0] == 0x44) && (((char *)&test_value)[1] == 0x33);

const endianness_t endianness = 
   is_big_endian ? BIG: 
  (is_little_endian ? LITTLE : UNKNOWN);


template <typename T>
T identity(T v){
  return v;
}

// 16 bits values ------

uint16_t swap_(uint16_t v){
  return ((v & 0xFF) << 8) | ((v & 0xFF00) >> 8);
}

// 32 bits values ------

uint32_t swap_(uint32_t v){
  return ((v & 0xFF) << 24) | ((v & 0xFF00) << 8) | ((v & 0xFF0000) >> 8) | ((v & 0xFF000000) >> 24);
}

template <typename T, endianness_t HOST, endianness_t REMOTE>
 struct en_swap{
  static T conv(T v){
    return swap_(v);
  }
};

template <typename T>
struct en_swap<T, BIG, BIG>{
  static T conv(T v){
    return v;
  }
};

template <typename T>
struct en_swap<T, LITTLE, LITTLE> {
  static T conv(T v){
    return v;
  }
};

template <typename T>
T to_big(T v) {

  switch (endianness){
  case LITTLE :
    return en_swap<T,LITTLE,BIG>::conv(v);
  case BIG :
    return en_swap<T,BIG,BIG>::conv(v);
  }
}

template <typename T>
T to_little(T v) {
   switch (endianness){
   case LITTLE :
     return en_swap<T,LITTLE,LITTLE>::conv(v);
   case BIG :
     return en_swap<T,BIG,LITTLE>::conv(v);
  }
}


int main(){

  using namespace std;

  uint32_t x = 0x0ABCDEF0;
  uint32_t y = to_big(x);
  uint32_t z = to_little(x);

  cout << hex << setw(8) << setfill('0') << x << " " << y << " " << setw(8) << setfill('0') << z << endl;

}

In fact, the MSDN functions ntohl() and htonl() are the inverse of eachother:

The htonl function converts a u_long from host to TCP/IP network byte order (which is big-endian).

The ntohl function converts a u_long from TCP/IP network order to host byte order (which is little-endian on Intel processors).

Yes, runtime detecting endianness is a very sane thing to do, and basically what any ready-to-use macro/function would do at some point anyway.

And if you want to do little-big endian conversions yourself, see answer by @R-Martinho-Fernandes.

Warning: This only works on unsigned integers, because signed right shift is implementation defined and can lead to vulnerabilities (https://stackoverflow.com/a/7522498/395029)

C already provides an abstraction over the host's endianness: the number† or int†.

Producing output in a given endianness can be done portably by not trying to be clever: simply interpret the numbers as numbers and use bit shifts to extract each byte:

uint32_t value;
uint8_t lolo = (value >> 0) & 0xFF;
uint8_t lohi = (value >> 8) & 0xFF;
uint8_t hilo = (value >> 16) & 0xFF;
uint8_t hihi = (value >> 24) & 0xFF;

Then you just write the bytes in whatever order you desire.

When you are taking byte sequences with some endianness as input, you can reconstruct them in the host's endianness by again constructing numbers with bit operations:

uint32_t value = (hihi << 24)
               | (hilo << 16)
               | (lohi << 8)
               | (lolo << 0);

---

† Only the representations of numbers as byte sequences have endianness; numbers (i.e. quantities) don't.