#include <stdio.h>
#include <string.h>
#include <time.h>


#define mmxcpy(dest, src, len) \
	__asm__ ( \
	"movl %2, %%ecx\n\t" \
	"shrl $5, %%ecx\n\t" \
	"cplp:\n\t" \
	"movq (%1), %%mm0\n\t" \
	"movq 8(%1), %%mm1\n\t" \
	"movq 16(%1), %%mm2\n\t" \
	"movq 24(%1), %%mm3\n\t" \
	"movq (%0), %%mm4\n\t" \
	"movq %%mm0, (%0)\n\t" \
	"movq %%mm1, 8(%0)\n\t" \
	"movq %%mm2, 16(%0)\n\t" \
	"movq %%mm3, 24(%0)\n\t" \
	"addl $32, %1\n\t" \
	"addl $32, %0\n\t" \
	"decl %%ecx\n\t" \
	"jnz cplp\n\t" \
	: : "D" (dest), "S" (src), "g" (len) : "%ecx", "%esi", "%edi" )



static unsigned char a[4000] __attribute__ ((aligned(32)));
static unsigned char b[4000] __attribute__ ((aligned(32)));


main()
{
	clock_t t1, t2;
	int i;

	t1 = clock();
	while(t1 == clock());
	t1 = clock();

	for(i = 0; i < 200000; i++)
		mmxcpy(a, b, 4000);

	t2 = clock();
	printf("mmxcpy: %d\n", t2 - t1);


	t1 = clock();
	while(t1 == clock());
	t1 = clock();

	for(i = 0; i < 200000; i++)
		memcpy(a, b, 4000);

	t2 = clock();
	printf("memcpy: %d\n", t2 - t1);
}
.