Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* checksum.S: Sparc V9 optimized checksum code. |
2 | * | |
3 | * Copyright(C) 1995 Linus Torvalds | |
4 | * Copyright(C) 1995 Miguel de Icaza | |
5 | * Copyright(C) 1996, 2000 David S. Miller | |
6 | * Copyright(C) 1997 Jakub Jelinek | |
7 | * | |
8 | * derived from: | |
9 | * Linux/Alpha checksum c-code | |
10 | * Linux/ix86 inline checksum assembly | |
11 | * RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code) | |
12 | * David Mosberger-Tang for optimized reference c-code | |
13 | * BSD4.4 portable checksum routine | |
14 | */ | |
15 | ||
16 | .text | |
17 | ||
18 | csum_partial_fix_alignment: | |
19 | /* We checked for zero length already, so there must be | |
20 | * at least one byte. | |
21 | */ | |
22 | be,pt %icc, 1f | |
23 | nop | |
24 | ldub [%o0 + 0x00], %o4 | |
25 | add %o0, 1, %o0 | |
26 | sub %o1, 1, %o1 | |
27 | 1: andcc %o0, 0x2, %g0 | |
28 | be,pn %icc, csum_partial_post_align | |
29 | cmp %o1, 2 | |
30 | blu,pn %icc, csum_partial_end_cruft | |
31 | nop | |
32 | lduh [%o0 + 0x00], %o5 | |
33 | add %o0, 2, %o0 | |
34 | sub %o1, 2, %o1 | |
35 | ba,pt %xcc, csum_partial_post_align | |
36 | add %o5, %o4, %o4 | |
37 | ||
38 | .align 32 | |
39 | .globl csum_partial | |
40 | csum_partial: /* %o0=buff, %o1=len, %o2=sum */ | |
41 | prefetch [%o0 + 0x000], #n_reads | |
42 | clr %o4 | |
43 | prefetch [%o0 + 0x040], #n_reads | |
44 | brz,pn %o1, csum_partial_finish | |
45 | andcc %o0, 0x3, %g0 | |
46 | ||
47 | /* We "remember" whether the lowest bit in the address | |
48 | * was set in %g7. Because if it is, we have to swap | |
49 | * upper and lower 8 bit fields of the sum we calculate. | |
50 | */ | |
51 | bne,pn %icc, csum_partial_fix_alignment | |
52 | andcc %o0, 0x1, %g7 | |
53 | ||
54 | csum_partial_post_align: | |
55 | prefetch [%o0 + 0x080], #n_reads | |
56 | andncc %o1, 0x3f, %o3 | |
57 | ||
58 | prefetch [%o0 + 0x0c0], #n_reads | |
59 | sub %o1, %o3, %o1 | |
60 | brz,pn %o3, 2f | |
61 | prefetch [%o0 + 0x100], #n_reads | |
62 | ||
63 | /* So that we don't need to use the non-pairing | |
64 | * add-with-carry instructions we accumulate 32-bit | |
65 | * values into a 64-bit register. At the end of the | |
66 | * loop we fold it down to 32-bits and so on. | |
67 | */ | |
68 | prefetch [%o0 + 0x140], #n_reads | |
69 | 1: lduw [%o0 + 0x00], %o5 | |
70 | lduw [%o0 + 0x04], %g1 | |
71 | lduw [%o0 + 0x08], %g2 | |
72 | add %o4, %o5, %o4 | |
73 | lduw [%o0 + 0x0c], %g3 | |
74 | add %o4, %g1, %o4 | |
75 | lduw [%o0 + 0x10], %o5 | |
76 | add %o4, %g2, %o4 | |
77 | lduw [%o0 + 0x14], %g1 | |
78 | add %o4, %g3, %o4 | |
79 | lduw [%o0 + 0x18], %g2 | |
80 | add %o4, %o5, %o4 | |
81 | lduw [%o0 + 0x1c], %g3 | |
82 | add %o4, %g1, %o4 | |
83 | lduw [%o0 + 0x20], %o5 | |
84 | add %o4, %g2, %o4 | |
85 | lduw [%o0 + 0x24], %g1 | |
86 | add %o4, %g3, %o4 | |
87 | lduw [%o0 + 0x28], %g2 | |
88 | add %o4, %o5, %o4 | |
89 | lduw [%o0 + 0x2c], %g3 | |
90 | add %o4, %g1, %o4 | |
91 | lduw [%o0 + 0x30], %o5 | |
92 | add %o4, %g2, %o4 | |
93 | lduw [%o0 + 0x34], %g1 | |
94 | add %o4, %g3, %o4 | |
95 | lduw [%o0 + 0x38], %g2 | |
96 | add %o4, %o5, %o4 | |
97 | lduw [%o0 + 0x3c], %g3 | |
98 | add %o4, %g1, %o4 | |
99 | prefetch [%o0 + 0x180], #n_reads | |
100 | add %o4, %g2, %o4 | |
101 | subcc %o3, 0x40, %o3 | |
102 | add %o0, 0x40, %o0 | |
103 | bne,pt %icc, 1b | |
104 | add %o4, %g3, %o4 | |
105 | ||
106 | 2: and %o1, 0x3c, %o3 | |
107 | brz,pn %o3, 2f | |
108 | sub %o1, %o3, %o1 | |
109 | 1: lduw [%o0 + 0x00], %o5 | |
110 | subcc %o3, 0x4, %o3 | |
111 | add %o0, 0x4, %o0 | |
112 | bne,pt %icc, 1b | |
113 | add %o4, %o5, %o4 | |
114 | ||
115 | 2: | |
116 | /* fold 64-->32 */ | |
117 | srlx %o4, 32, %o5 | |
118 | srl %o4, 0, %o4 | |
119 | add %o4, %o5, %o4 | |
120 | srlx %o4, 32, %o5 | |
121 | srl %o4, 0, %o4 | |
122 | add %o4, %o5, %o4 | |
123 | ||
124 | /* fold 32-->16 */ | |
125 | sethi %hi(0xffff0000), %g1 | |
126 | srl %o4, 16, %o5 | |
127 | andn %o4, %g1, %g2 | |
128 | add %o5, %g2, %o4 | |
129 | srl %o4, 16, %o5 | |
130 | andn %o4, %g1, %g2 | |
131 | add %o5, %g2, %o4 | |
132 | ||
133 | csum_partial_end_cruft: | |
134 | /* %o4 has the 16-bit sum we have calculated so-far. */ | |
135 | cmp %o1, 2 | |
136 | blu,pt %icc, 1f | |
137 | nop | |
138 | lduh [%o0 + 0x00], %o5 | |
139 | sub %o1, 2, %o1 | |
140 | add %o0, 2, %o0 | |
141 | add %o4, %o5, %o4 | |
142 | 1: brz,pt %o1, 1f | |
143 | nop | |
144 | ldub [%o0 + 0x00], %o5 | |
145 | sub %o1, 1, %o1 | |
146 | add %o0, 1, %o0 | |
147 | sllx %o5, 8, %o5 | |
148 | add %o4, %o5, %o4 | |
149 | 1: | |
150 | /* fold 32-->16 */ | |
151 | sethi %hi(0xffff0000), %g1 | |
152 | srl %o4, 16, %o5 | |
153 | andn %o4, %g1, %g2 | |
154 | add %o5, %g2, %o4 | |
155 | srl %o4, 16, %o5 | |
156 | andn %o4, %g1, %g2 | |
157 | add %o5, %g2, %o4 | |
158 | ||
159 | 1: brz,pt %g7, 1f | |
160 | nop | |
161 | ||
162 | /* We started with an odd byte, byte-swap the result. */ | |
163 | srl %o4, 8, %o5 | |
164 | and %o4, 0xff, %g1 | |
165 | sll %g1, 8, %g1 | |
166 | or %o5, %g1, %o4 | |
167 | ||
168 | 1: add %o2, %o4, %o2 | |
169 | ||
170 | csum_partial_finish: | |
171 | retl | |
172 | mov %o2, %o0 |