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

typedef unsigned long long BITBOARD;

static BITBOARD bitboard __attribute__ ((aligned(32)));
static BITBOARD bitboard2 __attribute__ ((aligned(8)));
static BITBOARD Occupied __attribute__ ((aligned(8)));
static BITBOARD rank_mask __attribute__ ((aligned(8))); 
static BITBOARD all_mask __attribute__ ((aligned(8))); 

#define Old_And(a,b) ((a) & (b))
#define Old_Or(a,b) ((a) | (b))
#define Old_Xor(a,b) ((a) ^ (b))
#define Old_Compl(a) (~(a))
#define Old_Shiftl(a,b) ((a) << (b))
#define Old_Shiftr(a,b) ((a) >> (b))


#define And(a, b) \
({ \
	static BITBOARD __tmp __attribute__ ((aligned(8))); \
	__asm__ ( \
	"movq %1, %%mm0\n\t" \
	"pand %2, %%mm0\n\t" \
	"movq %%mm0, %0" \
	: "=m" (__tmp): "m" (a), "m" (b) : "memory", "%mm0" ); \
	__tmp; \
})


#define Or(a, b) \
({ \
	static BITBOARD __tmp __attribute__ ((aligned(8))); \
	__asm__ ( \
	"movq %1, %%mm0\n\t" \
	"por %2, %%mm0\n\t" \
	"movq %%mm0, %0" \
	: "=m" (__tmp) : "m" (a), "m" (b) : "memory" ); \
	__tmp; \
})


#define Xor(a, b) \
({ \
	static BITBOARD __tmp __attribute__ ((aligned(8))); \
	__asm__ ( \
	"movq %1, %%mm0\n\t" \
	"pxor %2, %%mm0\n\t" \
	"movq %%mm0, %0" \
	: "=m" (__tmp) : "m" (a), "m" (b) : "memory" ); \
	__tmp; \
})


#define Compl(a) \
({  \
	static BITBOARD __tmp __attribute__ ((aligned(8))); \
	__asm__ ( \
	"movq %1, %%mm0\n\t" \
	"pandn all_mask, %%mm0\n\t" \
	"movq %%mm0, %0" \
	: "=m" (__tmp) : "m" (a) : "memory" ); \
	__tmp; \
})


#define Shiftl(a, b) \
({ \
	static BITBOARD __tmp __attribute__ ((aligned(8))); \
	__asm__ ( \
	"movq %1, %%mm0\n\t" \
	"psllq %2, %%mm0\n\t" \
	"movq %%mm0, %0" \
	: "=m" (__tmp) : "m" (a), "im" (b) : "memory" ); \
	__tmp; \
})


#define Shiftr(a, b) \
({ \
	static BITBOARD __tmp __attribute__ ((aligned(8))); \
	__asm__ ( \
	"movq %1, %%mm0\n\t" \
	"psrlq %2, %%mm0\n\t" \
	"movq %%mm0, %0" \
	: "=m" (__tmp) : "m" (a), "im" (b) : "memory" ); \
	__tmp; \
})


#define Emms() __asm__("emms")


normal_run()
{
	clock_t t1,t2;
	unsigned int i;

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

	for(i = 0; i < 30000000; i++) {
		bitboard2 = Old_And(Old_And(bitboard, Old_Shiftl(Occupied,3)), Old_Compl(rank_mask));
		bitboard ++;
		rank_mask--;
	}

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


mmx_run()
{
	clock_t t1, t2; 
	unsigned int i;

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

	for(i = 0; i < 30000000; i++) {
		bitboard2 = And(And(bitboard, Shiftl(Occupied,3)), Compl(rank_mask));
		bitboard++;
		rank_mask--;
	}

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



main()
{
	int i;

	printf("%p %p %p %p\n", &bitboard, &bitboard2, &Occupied, &rank_mask);

	Occupied =   0x80FF442288445533LL;
	bitboard = 0x8080808080804423LL; 
	bitboard2 = 0x0110101001434400LL; 

	all_mask = 0xffffffffffffffffLL;

	for(i = 0; i < 8; i++) {
		normal_run();
		mmx_run();
	}
}
.