#include <stdint.h>
#include <stdio.h>

static uint64_t mul31by31to62(uint32_t a, uint32_t b)
{
	uint64_t i;
	__asm__(
		"fildl %1\n\t"
		"fildl %2\n\t"
		"fmulp\n\t"
		"fistpll %0\n\t"
		: "=m" (i)
		: "m" (a), "m" (b)
	);
	return i;
}

static uint64_t mul63by63to63(uint64_t a, uint64_t b)
{
	uint64_t i;
	__asm__(
		"fildll %1\n\t"
		"fildll %2\n\t"
		"fmulp\n\t"
		"fistpll %0\n\t"
		: "=m" (i)
		: "m" (a), "m" (b)
	);
	return i;
}

static void test(uint32_t a, uint32_t b)
{
	printf("%08x * %08x = %016llx %016llx %016llx\n", a, b,
	    (unsigned long long)((uint64_t)a * (uint64_t)b),
	    mul31by31to62(a, b),
	    mul63by63to63(a, b));
}

int main(void)
{
	test(0, 0);
	test(1, 0);
	test(0, 1);
	test(1, 1);
	test(0x10, 0x10);
	test(1000, 1000);
	test(0x1000, 0x1000);
	test(0x10000, 0xffff);
	test(0x10000, 0x10000);
	test(0x100000, 0x100000);
	test(1000000, 1000000);
	test(0x12345678, 0x78765432);
	test(0x12345678, 0x7fffffff);
	test(0x7fffffff, 0x7fffffff);
	test(0x80000000, 0x80000000);
	test(0x80000000, 0x80000001);
	test(0x80000000, 0xffffffff);
	test(1, 0x80000000);
	test(1, 0x80000001);
	test(0x12345678, 0x98765432);
	test(0x7fffffff, 0xffffffff);
	test(0xffffffff, 0xffffffff);

	return 0;
}