;
; bcopy(p1, p2, nbytes)
;
	.globl	_bcopy
_bcopy:
	jsr	popax		; get nbytes
	sta	tmp1
	stx	tmp1+1
	jsr	popax		; p2
	sta	ptr2
	stx	ptr2+1
	jsr	popax		; p1
	sta	ptr1
	stx	ptr1+1
bc1:
	lda	tmp1+1		; nbytes^
	beq	bc3
	ldy	#0		; set up to move 256
bc2:
	lda	(ptr1),y	; get a byte
	sta	(ptr2),y	; store it
	iny
	bne	bc2
	inc	ptr1+1		; bump ptrs
	inc	ptr2+1
	dec	tmp1+1
	jmp	bc1		; do another block
bc3:
	ldy	#0		; set up for last section of < 256
bc4:
	cpy	tmp1		; done yet?
	beq	bc5		; yup, exit
	lda	(ptr1),y	; get a byte
	sta	(ptr2),y	; store it
	iny
	bne	bc4
bc5:
	rts			; done!
;