1/*
2 * A fast checksum routine using movem
3 * Copyright (c) 1998-2001, 2003 Axis Communications AB
4 *
5 * csum_partial(const unsigned char * buff, int len, unsigned int sum)
6 */
7
8	.globl	csum_partial
9csum_partial:
10
11	;; r10 - src
12	;; r11 - length
13	;; r12 - checksum
14
15	;; check for breakeven length between movem and normal word looping versions
16	;; we also do _NOT_ want to compute a checksum over more than the
17	;; actual length when length < 40
18
19	cmpu.w	80,$r11
20	blo	_word_loop
21	nop
22
23	;; need to save the registers we use below in the movem loop
24	;; this overhead is why we have a check above for breakeven length
25	;; only r0 - r8 have to be saved, the other ones are clobber-able
26	;; according to the ABI
27
28	subq	9*4,$sp
29	subq	10*4,$r11	; update length for the first loop
30	movem	$r8,[$sp]
31
32	;; do a movem checksum
33
34_mloop:	movem	[$r10+],$r9	; read 10 longwords
35
36	;; perform dword checksumming on the 10 longwords
37
38	add.d	$r0,$r12
39	addc	$r1,$r12
40	addc	$r2,$r12
41	addc	$r3,$r12
42	addc	$r4,$r12
43	addc	$r5,$r12
44	addc	$r6,$r12
45	addc	$r7,$r12
46	addc	$r8,$r12
47	addc	$r9,$r12
48
49	;; fold the carry into the checksum, to avoid having to loop the carry
50	;; back into the top
51
52	addc	0,$r12
53	addc	0,$r12		; do it again, since we might have generated a carry
54
55	subq	10*4,$r11
56	bge	_mloop
57	nop
58
59	addq	10*4,$r11	; compensate for last loop underflowing length
60
61	movem	[$sp+],$r8	; restore regs
62
63_word_loop:
64	;; only fold if there is anything to fold.
65
66	cmpq	0,$r12
67	beq	_no_fold
68
69	;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below.
70	;; r9 and r13 can be used as temporaries.
71
72	moveq	-1,$r9		; put 0xffff in r9, faster than move.d 0xffff,r9
73	lsrq	16,$r9
74
75	move.d	$r12,$r13
76	lsrq	16,$r13		; r13 = checksum >> 16
77	and.d	$r9,$r12		; checksum = checksum & 0xffff
78	add.d	$r13,$r12		; checksum += r13
79	move.d	$r12,$r13		; do the same again, maybe we got a carry last add
80	lsrq	16,$r13
81	and.d	$r9,$r12
82	add.d	$r13,$r12
83
84_no_fold:
85	cmpq	2,$r11
86	blt	_no_words
87	nop
88
89	;; checksum the rest of the words
90
91	subq	2,$r11
92
93_wloop:	subq	2,$r11
94	bge	_wloop
95	addu.w	[$r10+],$r12
96
97	addq	2,$r11
98
99_no_words:
100	;; see if we have one odd byte more
101	cmpq	1,$r11
102	beq	_do_byte
103	nop
104	ret
105	move.d	$r12,$r10
106
107_do_byte:
108	;; copy and checksum the last byte
109	addu.b	[$r10],$r12
110	ret
111	move.d	$r12,$r10
112