typedef unsigned long int li;
// Reversing byte order in a 64 bit word.
li rev(li x){li a = x<<32 | x>>32,
 m = 0xffff0000ffffL, b = (a&m)<<16 | (a&~m)>>16,
 n = 0x00ff00ff00ff00ffL; return (b&n)<<8 | (b&~n)>>8;}
// gcc thinks that instruction bswapq does the whole job.
// Now (2017 Sept) clang thinks so too.
// Indeed it seems to!

/* demo & test */
#include <stdlib.h>
li ran64(){return (li)random() ^ ((li)random()<<16) ^ ((li)random()<<33);}
#include <stdio.h>
int main(){printf("%016lX\n", rev(0x2201020304050607L));
{int j=100000000; while(j--){li o = ran64();
  if(rev(o) != __builtin_bswap64(o)) printf("%016lX\n", o);}}
return 0;}
// 2.4 sec per 10^8 with clang -O3
// 1.9 sec per 10^8 with gcc -O3
// -> 0706050403020122