.global _memchr
.type _memchr, @function

_memchr:
	mov	r4, r0
	exts.b	r5, r5

	/* For small inputs, simply check bytes individually */
	mov	#64, r2
	cmp/hi	r6, r2
	bt	.last

.large:	/* Make a 4-byte version of r5 for cmp/str */
	extu.b	r5, r3
	swap.b	r3, r2
	or	r3, r2
	swap.w	r2, r3
	or	r3, r2

	/* First check 3 bytes to ensure we don't skip bytes when aligning */
	mov.b	@r0+, r1
	cmp/eq	r1, r5
	bt	.end
	mov.b	@r0+, r1
	cmp/eq	r1, r5
	bt	.end
	mov.b	@r0+, r1
	cmp/eq	r1, r5
	bt	.end

	/* Align to a 4-byte boundary */
	shlr2	r0
	shll2	r0
	add	r4, r6
	sub	r0, r6

	mov	r6, r7
	shlr2	r7
	mov	#3, r3
	and	r3, r6

	/* Read longwords */
1:	mov.l	@r0+, r1
	cmp/str	r1, r2
	bt	.found
	dt	r7
	bf	1b

.last:	/* Don't read if there are no bytes left */
	tst	r6, r6
	bt	.none

2:	mov.b	@r0+, r1
	cmp/eq	r1, r5
	bt	.end
	dt	r6
	bf	2b

.none:	rts
	mov	#0, r0

.found:	/* Go back to find out which of the last 4 bytes is r5 */
	add	#-4, r0
	bra	2b
	mov	#4, r6

.end:	rts
	add	#-1, r0