typedef unsigned long time;
typedef unsigned long long stamp64;

extern inline time stamp(void)
{
	time tsc;
	asm volatile("rdtsc" : "=a" (tsc) : : "edx");
	return tsc;
}

extern inline time measure(time t)
{
	time tsc;
	asm volatile("rdtsc" : "=a" (tsc) : : "edx");
	if (tsc>t)
		return tsc-t;
	else
		return t-tsc;
}

int c1 = 12345678;
int d1;

static long long var1 __attribute__ ((aligned(32)));
static long long var2;

#if 1
static long long dummy1; /* These two make var3 and var4 be outside  */
static long long dummy2; /* the cache line when reading var1         */
#endif

static long long var3;
static long long var4;

int x[64]; 

int main(void)
{
   time s;
   time si[10];
   int i;
   long double *p;

   p = (void *)&x;
   while((int)p & 7) p++;
   
   p++;
   *p=1234.567E123;
   
   /*
   printf("p=%08lx\n",p);

   printf("var1: %p\nvar2: %p\nvar3: %p\nvar4: %p\n", 
	&var1, &var2, &var3, &var4);
   
  */ 
   for(i=0; i < 10; i++) {
      int a,b;
      int c[10];
      for(a=0;a<10;a++)
        c[a]=-5;
      s=stamp();
      asm("
	not %%esi
	nop
	nop

	.align 16

	nop
	nop
	nop
	nop

	nop				# Comment these four nop's out, 
	nop				# and suddenly the loop	is 
	nop				# > 100%% slower! 
	nop

	movl  $30000, %%ecx		# Loop this many times

	.align 16
lp:					# Here is the loop (just some
	movl var1,%%edx			# nonsense code)	
	movl var1+4,%%eax
	movl var2,%%ebx
	andl %%edx,%%ebx
	movl var2+4,%%esi
	andl %%eax,%%esi
	movl %%ebx,var3
	movl %%esi,var3+4
	movl %%edx,%%eax
	addl $1,%%eax
	movl %%eax,var4
	movl %%ebx,%%edx
	adcl $0,%%eax
	movl %%eax,var4+4
	decl %%ecx
	jnz  lp

	nop
	nop
	not %%esi
      " : : : "eax", "ebx", "edx", "ecx", "esi", "edi" );
      si[i]=measure(s);
   }
   for(i=0; i<10; i++) {
      printf("%5d cycles  %5f ms\n",si[i], ((float)si[i] * (1000.0 / 166.0)) / 1000000.0);
   }
   return 0;
}


.