#include <stdlib.h>

size_t f(void)
{

#if 1
	__asm__ __volatile__(
		".arch	iwmmxt\n"
		"mov	r0, #0\n"
		"mov	r1, #1\n"
		"tmcrr	wr1, r1, r0\n"	/* wr1 = 1ull */

		/* 'n = 0' */
		"mov	r3, #0\n"

		/* 'masked = 0x12' */
		"mov	r6, #0x12\n"
		"mov	r7, #0\n"

		/* while (masked) { ... */
		"0:\n"
		"orrs	r0, r6, r7\n"
		"beq	4f\n"

		/* BUG_ON(n > 4) */
		"cmp	r3, #4\n"
		"bls	1f\n"
		"tmrc	r5,  wcgr0\n"
		"tmrrc	r8,  r9,  wr0\n"
		"tmrrc	r10, r11, wr1\n"

		"strb	r7, [r7]\n"	/* *(char *)0 = 0 */

		"1:\n"
		/* 'masked >> n' */
		"tmcr	wcgr0, r3\n"	/* wcgr0 = n */
		"tmcrr	wr0, r6, r7\n"	/* wr0 = masked */
		"wsrldg	wr0, wr0, wcgr0\n" /* masked >> n */
		"tmrrc	r0, r1, wr0\n"	   /* (r0,r1) = masked >> n */

		/* if ((masked >> n) & 0x01) { ... */
		"and	r0, r0, #1\n"
		"cmp	r0, #0\n"
		"wslldg	wr0, wr1, wcgr0\n" /* wr0 = 1ull << n */
		"beq	3f\n"

		/* masked ^= 1ull << n */
		"tmrrc	r0, r1, wr0\n"	/* (r0,r1) = 1ull << n */
		"eor	r6, r6, r0\n"
		"eor	r7, r7, r1\n"

		"3:\n"
		"add	r3, r3, #1\n"	/* ++n */
		"b	0b\n"

		"4:\n" : : :
		"r0", "r1", "r2", "r3", "r4", "r6", "r7");

	return 2;

#else
	uint64_t	masked = a;
	size_t		n;
	size_t		m = 0;

	for (n = 0; masked != 0; ++n) {
		if ((masked & (1ULL << n)) != 0)
		{
			masked ^= 1ULL << n;
			++m;
		}

		if (n > 4)
			__asm__ __volatile__(
				".arch	iwmmxt\n"
				"tmrrc	r0, r1, wr10\n"
				"tmrrc	r2, r3, wr0\n"
				"tmrrc	r4, r5, wr1\n"
				"mov	r6, #0\n"
				"strb	r6, [r6]");
	}

	return m;
#endif
}


int main(int argc, char *argv[])
{
	size_t		i;
	pid_t		p = getpid();
	int		tmp;

	__asm__ __volatile__(
		".arch	iwmmxt\n"
		"tmcrr	wr0, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr1, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr2, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr3, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr4, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr5, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr6, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr7, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr8, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr9, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr10, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr11, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr12, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr13, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr14, %0, %2\n"
		"add	%0, %0, #1\n"
		"tmcrr	wr15, %0, %2\n"
		"add	%0, %0, #1\n"
		: "=r"(tmp) : "0"(0), "r"(p));

	return f() != 2;
}