.global _strlen
.type _strlen, @function

_strlen:
	mov	r4, r0
	mov	#0, r2

	/* Check 3 bytes to make sure we don't skip any when aligning */
	mov.b	@r0+, r1
	tst	r1, r1
	bt	.end
	mov.b	@r0+, r1
	tst	r1, r1
	bt	.end
	mov.b	@r0+, r1
	tst	r1, r1
	bt	.end

	/* Align to a 4-byte boundary */
	or	#3, r0
	xor	#3, r0

	/* Read bytes by groups of 4 */
1:	mov.l	@r0+, r1
	cmp/str	r1, r2
	bf	1b

	/* Go back to find out which of the last 4 bytes is the NUL */
	add	#-4, r0
2:	mov.b	@r0+, r1
	tst	r1, r1
	bf	2b

.end:	add	#-1, r0
	rts
	sub	r4, r0