1/*
2 * A fast checksum routine using movem
3 * Copyright (c) 1998-2001 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	movem	$r8,[$sp]
30
31	;; do a movem checksum
32
33	subq	10*4,$r11	; update length for the first loop
34
35_mloop:	movem	[$r10+],$r9	; read 10 longwords
36
37	;; perform dword checksumming on the 10 longwords
38
39	add.d	$r0,$r12
40	ax
41	add.d	$r1,$r12
42	ax
43	add.d	$r2,$r12
44	ax
45	add.d	$r3,$r12
46	ax
47	add.d	$r4,$r12
48	ax
49	add.d	$r5,$r12
50	ax
51	add.d	$r6,$r12
52	ax
53	add.d	$r7,$r12
54	ax
55	add.d	$r8,$r12
56	ax
57	add.d	$r9,$r12
58
59	;; fold the carry into the checksum, to avoid having to loop the carry
60	;; back into the top
61
62	ax
63	addq	0,$r12
64
65	subq	10*4,$r11
66	bge	_mloop
67	nop
68
69	addq	10*4,$r11	; compensate for last loop underflowing length
70
71	movem	[$sp+],$r8	; restore regs
72
73_word_loop:
74	;; only fold if there is anything to fold.
75
76	cmpq	0,$r12
77	beq	_no_fold
78
79	;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below.
80	;; r9 and r13 can be used as temporaries.
81
82	moveq	-1,$r9		; put 0xffff in r9, faster than move.d 0xffff,r9
83	lsrq	16,$r9
84
85	move.d	$r12,$r13
86	lsrq	16,$r13		; r13 = checksum >> 16
87	and.d	$r9,$r12		; checksum = checksum & 0xffff
88	add.d	$r13,$r12		; checksum += r13
89
90_no_fold:
91	cmpq	2,$r11
92	blt	_no_words
93	nop
94
95	;; checksum the rest of the words
96
97	subq	2,$r11
98
99_wloop:	subq	2,$r11
100	bge	_wloop
101	addu.w	[$r10+],$r12
102
103	addq	2,$r11
104
105_no_words:
106	;; see if we have one odd byte more
107	cmpq	1,$r11
108	beq	_do_byte
109	nop
110	ret
111	move.d	$r12, $r10
112
113_do_byte:
114	;; copy and checksum the last byte
115	addu.b	[$r10],$r12
116	ret
117	move.d	$r12, $r10
118
119