Commit | Line | Data |
---|---|---|
f5e706ad | 1 | /* |
a439fe51 | 2 | * include/asm/xor.h |
f5e706ad SR |
3 | * |
4 | * Optimized RAID-5 checksumming functions for 32-bit Sparc. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License as published by | |
8 | * the Free Software Foundation; either version 2, or (at your option) | |
9 | * any later version. | |
10 | * | |
11 | * You should have received a copy of the GNU General Public License | |
12 | * (for example /usr/src/linux/COPYING); if not, write to the Free | |
13 | * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
14 | */ | |
15 | ||
16 | /* | |
17 | * High speed xor_block operation for RAID4/5 utilizing the | |
18 | * ldd/std SPARC instructions. | |
19 | * | |
20 | * Copyright (C) 1999 Jakub Jelinek (jj@ultra.linux.cz) | |
21 | */ | |
22 | ||
23 | static void | |
24 | sparc_2(unsigned long bytes, unsigned long *p1, unsigned long *p2) | |
25 | { | |
26 | int lines = bytes / (sizeof (long)) / 8; | |
27 | ||
28 | do { | |
29 | __asm__ __volatile__( | |
30 | "ldd [%0 + 0x00], %%g2\n\t" | |
31 | "ldd [%0 + 0x08], %%g4\n\t" | |
32 | "ldd [%0 + 0x10], %%o0\n\t" | |
33 | "ldd [%0 + 0x18], %%o2\n\t" | |
34 | "ldd [%1 + 0x00], %%o4\n\t" | |
35 | "ldd [%1 + 0x08], %%l0\n\t" | |
36 | "ldd [%1 + 0x10], %%l2\n\t" | |
37 | "ldd [%1 + 0x18], %%l4\n\t" | |
38 | "xor %%g2, %%o4, %%g2\n\t" | |
39 | "xor %%g3, %%o5, %%g3\n\t" | |
40 | "xor %%g4, %%l0, %%g4\n\t" | |
41 | "xor %%g5, %%l1, %%g5\n\t" | |
42 | "xor %%o0, %%l2, %%o0\n\t" | |
43 | "xor %%o1, %%l3, %%o1\n\t" | |
44 | "xor %%o2, %%l4, %%o2\n\t" | |
45 | "xor %%o3, %%l5, %%o3\n\t" | |
46 | "std %%g2, [%0 + 0x00]\n\t" | |
47 | "std %%g4, [%0 + 0x08]\n\t" | |
48 | "std %%o0, [%0 + 0x10]\n\t" | |
49 | "std %%o2, [%0 + 0x18]\n" | |
50 | : | |
51 | : "r" (p1), "r" (p2) | |
52 | : "g2", "g3", "g4", "g5", | |
53 | "o0", "o1", "o2", "o3", "o4", "o5", | |
54 | "l0", "l1", "l2", "l3", "l4", "l5"); | |
55 | p1 += 8; | |
56 | p2 += 8; | |
57 | } while (--lines > 0); | |
58 | } | |
59 | ||
60 | static void | |
61 | sparc_3(unsigned long bytes, unsigned long *p1, unsigned long *p2, | |
62 | unsigned long *p3) | |
63 | { | |
64 | int lines = bytes / (sizeof (long)) / 8; | |
65 | ||
66 | do { | |
67 | __asm__ __volatile__( | |
68 | "ldd [%0 + 0x00], %%g2\n\t" | |
69 | "ldd [%0 + 0x08], %%g4\n\t" | |
70 | "ldd [%0 + 0x10], %%o0\n\t" | |
71 | "ldd [%0 + 0x18], %%o2\n\t" | |
72 | "ldd [%1 + 0x00], %%o4\n\t" | |
73 | "ldd [%1 + 0x08], %%l0\n\t" | |
74 | "ldd [%1 + 0x10], %%l2\n\t" | |
75 | "ldd [%1 + 0x18], %%l4\n\t" | |
76 | "xor %%g2, %%o4, %%g2\n\t" | |
77 | "xor %%g3, %%o5, %%g3\n\t" | |
78 | "ldd [%2 + 0x00], %%o4\n\t" | |
79 | "xor %%g4, %%l0, %%g4\n\t" | |
80 | "xor %%g5, %%l1, %%g5\n\t" | |
81 | "ldd [%2 + 0x08], %%l0\n\t" | |
82 | "xor %%o0, %%l2, %%o0\n\t" | |
83 | "xor %%o1, %%l3, %%o1\n\t" | |
84 | "ldd [%2 + 0x10], %%l2\n\t" | |
85 | "xor %%o2, %%l4, %%o2\n\t" | |
86 | "xor %%o3, %%l5, %%o3\n\t" | |
87 | "ldd [%2 + 0x18], %%l4\n\t" | |
88 | "xor %%g2, %%o4, %%g2\n\t" | |
89 | "xor %%g3, %%o5, %%g3\n\t" | |
90 | "xor %%g4, %%l0, %%g4\n\t" | |
91 | "xor %%g5, %%l1, %%g5\n\t" | |
92 | "xor %%o0, %%l2, %%o0\n\t" | |
93 | "xor %%o1, %%l3, %%o1\n\t" | |
94 | "xor %%o2, %%l4, %%o2\n\t" | |
95 | "xor %%o3, %%l5, %%o3\n\t" | |
96 | "std %%g2, [%0 + 0x00]\n\t" | |
97 | "std %%g4, [%0 + 0x08]\n\t" | |
98 | "std %%o0, [%0 + 0x10]\n\t" | |
99 | "std %%o2, [%0 + 0x18]\n" | |
100 | : | |
101 | : "r" (p1), "r" (p2), "r" (p3) | |
102 | : "g2", "g3", "g4", "g5", | |
103 | "o0", "o1", "o2", "o3", "o4", "o5", | |
104 | "l0", "l1", "l2", "l3", "l4", "l5"); | |
105 | p1 += 8; | |
106 | p2 += 8; | |
107 | p3 += 8; | |
108 | } while (--lines > 0); | |
109 | } | |
110 | ||
111 | static void | |
112 | sparc_4(unsigned long bytes, unsigned long *p1, unsigned long *p2, | |
113 | unsigned long *p3, unsigned long *p4) | |
114 | { | |
115 | int lines = bytes / (sizeof (long)) / 8; | |
116 | ||
117 | do { | |
118 | __asm__ __volatile__( | |
119 | "ldd [%0 + 0x00], %%g2\n\t" | |
120 | "ldd [%0 + 0x08], %%g4\n\t" | |
121 | "ldd [%0 + 0x10], %%o0\n\t" | |
122 | "ldd [%0 + 0x18], %%o2\n\t" | |
123 | "ldd [%1 + 0x00], %%o4\n\t" | |
124 | "ldd [%1 + 0x08], %%l0\n\t" | |
125 | "ldd [%1 + 0x10], %%l2\n\t" | |
126 | "ldd [%1 + 0x18], %%l4\n\t" | |
127 | "xor %%g2, %%o4, %%g2\n\t" | |
128 | "xor %%g3, %%o5, %%g3\n\t" | |
129 | "ldd [%2 + 0x00], %%o4\n\t" | |
130 | "xor %%g4, %%l0, %%g4\n\t" | |
131 | "xor %%g5, %%l1, %%g5\n\t" | |
132 | "ldd [%2 + 0x08], %%l0\n\t" | |
133 | "xor %%o0, %%l2, %%o0\n\t" | |
134 | "xor %%o1, %%l3, %%o1\n\t" | |
135 | "ldd [%2 + 0x10], %%l2\n\t" | |
136 | "xor %%o2, %%l4, %%o2\n\t" | |
137 | "xor %%o3, %%l5, %%o3\n\t" | |
138 | "ldd [%2 + 0x18], %%l4\n\t" | |
139 | "xor %%g2, %%o4, %%g2\n\t" | |
140 | "xor %%g3, %%o5, %%g3\n\t" | |
141 | "ldd [%3 + 0x00], %%o4\n\t" | |
142 | "xor %%g4, %%l0, %%g4\n\t" | |
143 | "xor %%g5, %%l1, %%g5\n\t" | |
144 | "ldd [%3 + 0x08], %%l0\n\t" | |
145 | "xor %%o0, %%l2, %%o0\n\t" | |
146 | "xor %%o1, %%l3, %%o1\n\t" | |
147 | "ldd [%3 + 0x10], %%l2\n\t" | |
148 | "xor %%o2, %%l4, %%o2\n\t" | |
149 | "xor %%o3, %%l5, %%o3\n\t" | |
150 | "ldd [%3 + 0x18], %%l4\n\t" | |
151 | "xor %%g2, %%o4, %%g2\n\t" | |
152 | "xor %%g3, %%o5, %%g3\n\t" | |
153 | "xor %%g4, %%l0, %%g4\n\t" | |
154 | "xor %%g5, %%l1, %%g5\n\t" | |
155 | "xor %%o0, %%l2, %%o0\n\t" | |
156 | "xor %%o1, %%l3, %%o1\n\t" | |
157 | "xor %%o2, %%l4, %%o2\n\t" | |
158 | "xor %%o3, %%l5, %%o3\n\t" | |
159 | "std %%g2, [%0 + 0x00]\n\t" | |
160 | "std %%g4, [%0 + 0x08]\n\t" | |
161 | "std %%o0, [%0 + 0x10]\n\t" | |
162 | "std %%o2, [%0 + 0x18]\n" | |
163 | : | |
164 | : "r" (p1), "r" (p2), "r" (p3), "r" (p4) | |
165 | : "g2", "g3", "g4", "g5", | |
166 | "o0", "o1", "o2", "o3", "o4", "o5", | |
167 | "l0", "l1", "l2", "l3", "l4", "l5"); | |
168 | p1 += 8; | |
169 | p2 += 8; | |
170 | p3 += 8; | |
171 | p4 += 8; | |
172 | } while (--lines > 0); | |
173 | } | |
174 | ||
175 | static void | |
176 | sparc_5(unsigned long bytes, unsigned long *p1, unsigned long *p2, | |
177 | unsigned long *p3, unsigned long *p4, unsigned long *p5) | |
178 | { | |
179 | int lines = bytes / (sizeof (long)) / 8; | |
180 | ||
181 | do { | |
182 | __asm__ __volatile__( | |
183 | "ldd [%0 + 0x00], %%g2\n\t" | |
184 | "ldd [%0 + 0x08], %%g4\n\t" | |
185 | "ldd [%0 + 0x10], %%o0\n\t" | |
186 | "ldd [%0 + 0x18], %%o2\n\t" | |
187 | "ldd [%1 + 0x00], %%o4\n\t" | |
188 | "ldd [%1 + 0x08], %%l0\n\t" | |
189 | "ldd [%1 + 0x10], %%l2\n\t" | |
190 | "ldd [%1 + 0x18], %%l4\n\t" | |
191 | "xor %%g2, %%o4, %%g2\n\t" | |
192 | "xor %%g3, %%o5, %%g3\n\t" | |
193 | "ldd [%2 + 0x00], %%o4\n\t" | |
194 | "xor %%g4, %%l0, %%g4\n\t" | |
195 | "xor %%g5, %%l1, %%g5\n\t" | |
196 | "ldd [%2 + 0x08], %%l0\n\t" | |
197 | "xor %%o0, %%l2, %%o0\n\t" | |
198 | "xor %%o1, %%l3, %%o1\n\t" | |
199 | "ldd [%2 + 0x10], %%l2\n\t" | |
200 | "xor %%o2, %%l4, %%o2\n\t" | |
201 | "xor %%o3, %%l5, %%o3\n\t" | |
202 | "ldd [%2 + 0x18], %%l4\n\t" | |
203 | "xor %%g2, %%o4, %%g2\n\t" | |
204 | "xor %%g3, %%o5, %%g3\n\t" | |
205 | "ldd [%3 + 0x00], %%o4\n\t" | |
206 | "xor %%g4, %%l0, %%g4\n\t" | |
207 | "xor %%g5, %%l1, %%g5\n\t" | |
208 | "ldd [%3 + 0x08], %%l0\n\t" | |
209 | "xor %%o0, %%l2, %%o0\n\t" | |
210 | "xor %%o1, %%l3, %%o1\n\t" | |
211 | "ldd [%3 + 0x10], %%l2\n\t" | |
212 | "xor %%o2, %%l4, %%o2\n\t" | |
213 | "xor %%o3, %%l5, %%o3\n\t" | |
214 | "ldd [%3 + 0x18], %%l4\n\t" | |
215 | "xor %%g2, %%o4, %%g2\n\t" | |
216 | "xor %%g3, %%o5, %%g3\n\t" | |
217 | "ldd [%4 + 0x00], %%o4\n\t" | |
218 | "xor %%g4, %%l0, %%g4\n\t" | |
219 | "xor %%g5, %%l1, %%g5\n\t" | |
220 | "ldd [%4 + 0x08], %%l0\n\t" | |
221 | "xor %%o0, %%l2, %%o0\n\t" | |
222 | "xor %%o1, %%l3, %%o1\n\t" | |
223 | "ldd [%4 + 0x10], %%l2\n\t" | |
224 | "xor %%o2, %%l4, %%o2\n\t" | |
225 | "xor %%o3, %%l5, %%o3\n\t" | |
226 | "ldd [%4 + 0x18], %%l4\n\t" | |
227 | "xor %%g2, %%o4, %%g2\n\t" | |
228 | "xor %%g3, %%o5, %%g3\n\t" | |
229 | "xor %%g4, %%l0, %%g4\n\t" | |
230 | "xor %%g5, %%l1, %%g5\n\t" | |
231 | "xor %%o0, %%l2, %%o0\n\t" | |
232 | "xor %%o1, %%l3, %%o1\n\t" | |
233 | "xor %%o2, %%l4, %%o2\n\t" | |
234 | "xor %%o3, %%l5, %%o3\n\t" | |
235 | "std %%g2, [%0 + 0x00]\n\t" | |
236 | "std %%g4, [%0 + 0x08]\n\t" | |
237 | "std %%o0, [%0 + 0x10]\n\t" | |
238 | "std %%o2, [%0 + 0x18]\n" | |
239 | : | |
240 | : "r" (p1), "r" (p2), "r" (p3), "r" (p4), "r" (p5) | |
241 | : "g2", "g3", "g4", "g5", | |
242 | "o0", "o1", "o2", "o3", "o4", "o5", | |
243 | "l0", "l1", "l2", "l3", "l4", "l5"); | |
244 | p1 += 8; | |
245 | p2 += 8; | |
246 | p3 += 8; | |
247 | p4 += 8; | |
248 | p5 += 8; | |
249 | } while (--lines > 0); | |
250 | } | |
251 | ||
252 | static struct xor_block_template xor_block_SPARC = { | |
253 | .name = "SPARC", | |
254 | .do_2 = sparc_2, | |
255 | .do_3 = sparc_3, | |
256 | .do_4 = sparc_4, | |
257 | .do_5 = sparc_5, | |
258 | }; | |
259 | ||
260 | /* For grins, also test the generic routines. */ | |
261 | #include <asm-generic/xor.h> | |
262 | ||
263 | #undef XOR_TRY_TEMPLATES | |
264 | #define XOR_TRY_TEMPLATES \ | |
265 | do { \ | |
266 | xor_speed(&xor_block_8regs); \ | |
267 | xor_speed(&xor_block_32regs); \ | |
268 | xor_speed(&xor_block_SPARC); \ | |
269 | } while (0) |