#include <stdio.h>

#define rdtscl(low) \
     __asm__ __volatile__ ("rdtsc" : "=a" (low) : : "edx")

#define MAD_F_SCALEBITS 24

#define shrd(__lo_,__hi_,__result) \
   asm volatile("shrdl %3,%2,%1"  \
        : "=rm" (__result)  \
        : "0" (__lo_), "r" (__hi_), "I" (MAD_F_SCALEBITS)  \
        : "cc");  \

#define shrd_new(__lo_,__hi_,__result) \
   asm volatile("shrl %3,%1\n\t"  \
        "shll %4,%2\n\t"  \
        "orl %2,%1\n\t"  \
        : "=rm" (__result)  \
        : "0" (__lo_), "r" (__hi_), "I" (MAD_F_SCALEBITS), \
          "I" (32-MAD_F_SCALEBITS)  \
        : "cc");  \


void main() {
  long a,b,c;
  size_t i;
  int x,y,r;
  x=0x12345678;
  y=0x87654321;
  rdtscl(a);
  for (i=0; i<100; ++i)
    shrd(x,y,r);
  rdtscl(b);
  printf("normal: %ld (%lx)\n",b-a,r);
  rdtscl(a);
  for (i=0; i<100; ++i)
    shrd_new(x,y,r);
  rdtscl(b);
  printf("pointer arithmetic: %ld (%lx)\n",b-a,r);
}