movp:
	push cx
	push ds
	push si
	push es
	push di

	cmp ax, dx
	ja .up
	je .return
	push ax
	add ax, cx
	cmp ax, dx
	pop ax
	ja .down



.up:
	push ax
	push dx
.uploop:
	mov ds, ax
	mov es, dx
	xor di, di
	xor si, si
	sub cx, 1000h
	jbe .uplast
	push cx
	mov cx, 10000h /2
	rep movsw
	pop cx
	add ax, 1000h
	add dx, 1000h
	jmp short .uploop
.uplast:
	add cx, 1000h
	shl cx, 1
	shl cx, 1
	shl cx, 1
	rep movsw
	pop dx
	pop ax
	jmp short .return

.down:
	std
.dnloop:
	sub cx, 1000h
	jbe .dnlast
	push ax
	push dx
	add ax, cx
	add dx, cx
	mov ds, ax
	mov es, dx
	pop dx
	pop ax
	mov di, -2
	mov si, di
	push cx
	mov cx, 10000h /2
	rep movsw
	pop cx
	jmp short .dnloop
.dnlast:
	add cx, 1000h
	shl cx, 1
	shl cx, 1
	shl cx, 1
	mov di, cx
	dec di
	shl di, 1
	mov si, di
	mov ds, ax
	mov es, dx


	numdef AMD_ERRATUM_109_WORKAROUND, 1


%if _AMD_ERRATUM_109_WORKAROUND
	jcxz @FF
	cmp cx, 20
	ja @FF
@@:
	movsw
	loop @B
@@:
%endif
	rep movsw
	cld
.return:
	pop di
	pop es
	pop si
	pop ds
	pop cx
	retn