]>
Commit | Line | Data |
---|---|---|
ab59284e | 1 | /* |
1da177e4 LT |
2 | * A fast checksum+copy routine using movem |
3 | * Copyright (c) 1998, 2001 Axis Communications AB | |
4 | * | |
5 | * Authors: Bjorn Wesen | |
6 | * | |
7 | * csum_partial_copy_nocheck(const char *src, char *dst, | |
8 | * int len, unsigned int sum) | |
9 | */ | |
10 | ||
11 | .globl csum_partial_copy_nocheck | |
12 | csum_partial_copy_nocheck: | |
13 | ||
14 | ;; r10 - src | |
15 | ;; r11 - dst | |
16 | ;; r12 - length | |
17 | ;; r13 - checksum | |
18 | ||
19 | ;; check for breakeven length between movem and normal word looping versions | |
20 | ;; we also do _NOT_ want to compute a checksum over more than the | |
21 | ;; actual length when length < 40 | |
22 | ||
23 | cmpu.w 80, $r12 | |
24 | blo _word_loop | |
25 | nop | |
26 | ||
27 | ;; need to save the registers we use below in the movem loop | |
28 | ;; this overhead is why we have a check above for breakeven length | |
29 | ;; only r0 - r8 have to be saved, the other ones are clobber-able | |
30 | ;; according to the ABI | |
31 | ||
32 | subq 9*4, $sp | |
33 | movem $r8, [$sp] | |
34 | ||
35 | ;; do a movem copy and checksum | |
36 | ||
37 | subq 10*4, $r12 ; update length for the first loop | |
38 | ||
39 | _mloop: movem [$r10+],$r9 ; read 10 longwords | |
40 | 1: ;; A failing userspace access will have this as PC. | |
41 | movem $r9,[$r11+] ; write 10 longwords | |
42 | ||
43 | ;; perform dword checksumming on the 10 longwords | |
44 | ||
45 | add.d $r0,$r13 | |
46 | ax | |
47 | add.d $r1,$r13 | |
48 | ax | |
49 | add.d $r2,$r13 | |
50 | ax | |
51 | add.d $r3,$r13 | |
52 | ax | |
53 | add.d $r4,$r13 | |
54 | ax | |
55 | add.d $r5,$r13 | |
56 | ax | |
57 | add.d $r6,$r13 | |
58 | ax | |
59 | add.d $r7,$r13 | |
60 | ax | |
61 | add.d $r8,$r13 | |
62 | ax | |
63 | add.d $r9,$r13 | |
64 | ||
65 | ;; fold the carry into the checksum, to avoid having to loop the carry | |
66 | ;; back into the top | |
67 | ||
68 | ax | |
69 | addq 0,$r13 | |
1da177e4 LT |
70 | |
71 | subq 10*4,$r12 | |
72 | bge _mloop | |
73 | nop | |
74 | ||
75 | addq 10*4,$r12 ; compensate for last loop underflowing length | |
76 | ||
77 | movem [$sp+],$r8 ; restore regs | |
78 | ||
79 | _word_loop: | |
80 | ;; only fold if there is anything to fold. | |
81 | ||
82 | cmpq 0,$r13 | |
83 | beq _no_fold | |
84 | ||
85 | ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below | |
86 | ;; r9 can be used as temporary. | |
87 | ||
88 | move.d $r13,$r9 | |
89 | lsrq 16,$r9 ; r0 = checksum >> 16 | |
90 | and.d 0xffff,$r13 ; checksum = checksum & 0xffff | |
91 | add.d $r9,$r13 ; checksum += r0 | |
1da177e4 LT |
92 | |
93 | _no_fold: | |
94 | cmpq 2,$r12 | |
95 | blt _no_words | |
96 | nop | |
97 | ||
98 | ;; copy and checksum the rest of the words | |
99 | ||
100 | subq 2,$r12 | |
101 | ||
102 | _wloop: move.w [$r10+],$r9 | |
103 | 2: ;; A failing userspace access will have this as PC. | |
104 | addu.w $r9,$r13 | |
105 | subq 2,$r12 | |
106 | bge _wloop | |
107 | move.w $r9,[$r11+] | |
108 | ||
109 | addq 2,$r12 | |
110 | ||
111 | _no_words: | |
112 | ;; see if we have one odd byte more | |
113 | cmpq 1,$r12 | |
114 | beq _do_byte | |
115 | nop | |
116 | ret | |
117 | move.d $r13, $r10 | |
118 | ||
119 | _do_byte: | |
120 | ;; copy and checksum the last byte | |
121 | move.b [$r10],$r9 | |
122 | 3: ;; A failing userspace access will have this as PC. | |
123 | addu.b $r9,$r13 | |
124 | move.b $r9,[$r11] | |
125 | ret | |
126 | move.d $r13, $r10 |