1 /* SPDX-License-Identifier: GPL-2.0-or-later */
2 #ifndef _ASM_X86_XOR_32_H
3 #define _ASM_X86_XOR_32_H
4
5 /*
6 * Optimized RAID-5 checksumming functions for MMX.
7 */
8
9 /*
10 * High-speed RAID5 checksumming functions utilizing MMX instructions.
11 * Copyright (C) 1998 Ingo Molnar.
12 */
13
14 #define LD(x, y) " movq 8*("#x")(%1), %%mm"#y" ;\n"
15 #define ST(x, y) " movq %%mm"#y", 8*("#x")(%1) ;\n"
16 #define XO1(x, y) " pxor 8*("#x")(%2), %%mm"#y" ;\n"
17 #define XO2(x, y) " pxor 8*("#x")(%3), %%mm"#y" ;\n"
18 #define XO3(x, y) " pxor 8*("#x")(%4), %%mm"#y" ;\n"
19 #define XO4(x, y) " pxor 8*("#x")(%5), %%mm"#y" ;\n"
20
21 #include <asm/fpu/api.h>
22
23 static void
xor_pII_mmx_2(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2)24 xor_pII_mmx_2(unsigned long bytes, unsigned long * __restrict p1,
25 const unsigned long * __restrict p2)
26 {
27 unsigned long lines = bytes >> 7;
28
29 kernel_fpu_begin();
30
31 asm volatile(
32 #undef BLOCK
33 #define BLOCK(i) \
34 LD(i, 0) \
35 LD(i + 1, 1) \
36 LD(i + 2, 2) \
37 LD(i + 3, 3) \
38 XO1(i, 0) \
39 ST(i, 0) \
40 XO1(i+1, 1) \
41 ST(i+1, 1) \
42 XO1(i + 2, 2) \
43 ST(i + 2, 2) \
44 XO1(i + 3, 3) \
45 ST(i + 3, 3)
46
47 " .align 32 ;\n"
48 " 1: ;\n"
49
50 BLOCK(0)
51 BLOCK(4)
52 BLOCK(8)
53 BLOCK(12)
54
55 " addl $128, %1 ;\n"
56 " addl $128, %2 ;\n"
57 " decl %0 ;\n"
58 " jnz 1b ;\n"
59 : "+r" (lines),
60 "+r" (p1), "+r" (p2)
61 :
62 : "memory");
63
64 kernel_fpu_end();
65 }
66
67 static void
xor_pII_mmx_3(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2,const unsigned long * __restrict p3)68 xor_pII_mmx_3(unsigned long bytes, unsigned long * __restrict p1,
69 const unsigned long * __restrict p2,
70 const unsigned long * __restrict p3)
71 {
72 unsigned long lines = bytes >> 7;
73
74 kernel_fpu_begin();
75
76 asm volatile(
77 #undef BLOCK
78 #define BLOCK(i) \
79 LD(i, 0) \
80 LD(i + 1, 1) \
81 LD(i + 2, 2) \
82 LD(i + 3, 3) \
83 XO1(i, 0) \
84 XO1(i + 1, 1) \
85 XO1(i + 2, 2) \
86 XO1(i + 3, 3) \
87 XO2(i, 0) \
88 ST(i, 0) \
89 XO2(i + 1, 1) \
90 ST(i + 1, 1) \
91 XO2(i + 2, 2) \
92 ST(i + 2, 2) \
93 XO2(i + 3, 3) \
94 ST(i + 3, 3)
95
96 " .align 32 ;\n"
97 " 1: ;\n"
98
99 BLOCK(0)
100 BLOCK(4)
101 BLOCK(8)
102 BLOCK(12)
103
104 " addl $128, %1 ;\n"
105 " addl $128, %2 ;\n"
106 " addl $128, %3 ;\n"
107 " decl %0 ;\n"
108 " jnz 1b ;\n"
109 : "+r" (lines),
110 "+r" (p1), "+r" (p2), "+r" (p3)
111 :
112 : "memory");
113
114 kernel_fpu_end();
115 }
116
117 static void
xor_pII_mmx_4(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2,const unsigned long * __restrict p3,const unsigned long * __restrict p4)118 xor_pII_mmx_4(unsigned long bytes, unsigned long * __restrict p1,
119 const unsigned long * __restrict p2,
120 const unsigned long * __restrict p3,
121 const unsigned long * __restrict p4)
122 {
123 unsigned long lines = bytes >> 7;
124
125 kernel_fpu_begin();
126
127 asm volatile(
128 #undef BLOCK
129 #define BLOCK(i) \
130 LD(i, 0) \
131 LD(i + 1, 1) \
132 LD(i + 2, 2) \
133 LD(i + 3, 3) \
134 XO1(i, 0) \
135 XO1(i + 1, 1) \
136 XO1(i + 2, 2) \
137 XO1(i + 3, 3) \
138 XO2(i, 0) \
139 XO2(i + 1, 1) \
140 XO2(i + 2, 2) \
141 XO2(i + 3, 3) \
142 XO3(i, 0) \
143 ST(i, 0) \
144 XO3(i + 1, 1) \
145 ST(i + 1, 1) \
146 XO3(i + 2, 2) \
147 ST(i + 2, 2) \
148 XO3(i + 3, 3) \
149 ST(i + 3, 3)
150
151 " .align 32 ;\n"
152 " 1: ;\n"
153
154 BLOCK(0)
155 BLOCK(4)
156 BLOCK(8)
157 BLOCK(12)
158
159 " addl $128, %1 ;\n"
160 " addl $128, %2 ;\n"
161 " addl $128, %3 ;\n"
162 " addl $128, %4 ;\n"
163 " decl %0 ;\n"
164 " jnz 1b ;\n"
165 : "+r" (lines),
166 "+r" (p1), "+r" (p2), "+r" (p3), "+r" (p4)
167 :
168 : "memory");
169
170 kernel_fpu_end();
171 }
172
173
174 static void
xor_pII_mmx_5(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2,const unsigned long * __restrict p3,const unsigned long * __restrict p4,const unsigned long * __restrict p5)175 xor_pII_mmx_5(unsigned long bytes, unsigned long * __restrict p1,
176 const unsigned long * __restrict p2,
177 const unsigned long * __restrict p3,
178 const unsigned long * __restrict p4,
179 const unsigned long * __restrict p5)
180 {
181 unsigned long lines = bytes >> 7;
182
183 kernel_fpu_begin();
184
185 /* Make sure GCC forgets anything it knows about p4 or p5,
186 such that it won't pass to the asm volatile below a
187 register that is shared with any other variable. That's
188 because we modify p4 and p5 there, but we can't mark them
189 as read/write, otherwise we'd overflow the 10-asm-operands
190 limit of GCC < 3.1. */
191 asm("" : "+r" (p4), "+r" (p5));
192
193 asm volatile(
194 #undef BLOCK
195 #define BLOCK(i) \
196 LD(i, 0) \
197 LD(i + 1, 1) \
198 LD(i + 2, 2) \
199 LD(i + 3, 3) \
200 XO1(i, 0) \
201 XO1(i + 1, 1) \
202 XO1(i + 2, 2) \
203 XO1(i + 3, 3) \
204 XO2(i, 0) \
205 XO2(i + 1, 1) \
206 XO2(i + 2, 2) \
207 XO2(i + 3, 3) \
208 XO3(i, 0) \
209 XO3(i + 1, 1) \
210 XO3(i + 2, 2) \
211 XO3(i + 3, 3) \
212 XO4(i, 0) \
213 ST(i, 0) \
214 XO4(i + 1, 1) \
215 ST(i + 1, 1) \
216 XO4(i + 2, 2) \
217 ST(i + 2, 2) \
218 XO4(i + 3, 3) \
219 ST(i + 3, 3)
220
221 " .align 32 ;\n"
222 " 1: ;\n"
223
224 BLOCK(0)
225 BLOCK(4)
226 BLOCK(8)
227 BLOCK(12)
228
229 " addl $128, %1 ;\n"
230 " addl $128, %2 ;\n"
231 " addl $128, %3 ;\n"
232 " addl $128, %4 ;\n"
233 " addl $128, %5 ;\n"
234 " decl %0 ;\n"
235 " jnz 1b ;\n"
236 : "+r" (lines),
237 "+r" (p1), "+r" (p2), "+r" (p3)
238 : "r" (p4), "r" (p5)
239 : "memory");
240
241 /* p4 and p5 were modified, and now the variables are dead.
242 Clobber them just to be sure nobody does something stupid
243 like assuming they have some legal value. */
244 asm("" : "=r" (p4), "=r" (p5));
245
246 kernel_fpu_end();
247 }
248
249 #undef LD
250 #undef XO1
251 #undef XO2
252 #undef XO3
253 #undef XO4
254 #undef ST
255 #undef BLOCK
256
257 static void
xor_p5_mmx_2(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2)258 xor_p5_mmx_2(unsigned long bytes, unsigned long * __restrict p1,
259 const unsigned long * __restrict p2)
260 {
261 unsigned long lines = bytes >> 6;
262
263 kernel_fpu_begin();
264
265 asm volatile(
266 " .align 32 ;\n"
267 " 1: ;\n"
268 " movq (%1), %%mm0 ;\n"
269 " movq 8(%1), %%mm1 ;\n"
270 " pxor (%2), %%mm0 ;\n"
271 " movq 16(%1), %%mm2 ;\n"
272 " movq %%mm0, (%1) ;\n"
273 " pxor 8(%2), %%mm1 ;\n"
274 " movq 24(%1), %%mm3 ;\n"
275 " movq %%mm1, 8(%1) ;\n"
276 " pxor 16(%2), %%mm2 ;\n"
277 " movq 32(%1), %%mm4 ;\n"
278 " movq %%mm2, 16(%1) ;\n"
279 " pxor 24(%2), %%mm3 ;\n"
280 " movq 40(%1), %%mm5 ;\n"
281 " movq %%mm3, 24(%1) ;\n"
282 " pxor 32(%2), %%mm4 ;\n"
283 " movq 48(%1), %%mm6 ;\n"
284 " movq %%mm4, 32(%1) ;\n"
285 " pxor 40(%2), %%mm5 ;\n"
286 " movq 56(%1), %%mm7 ;\n"
287 " movq %%mm5, 40(%1) ;\n"
288 " pxor 48(%2), %%mm6 ;\n"
289 " pxor 56(%2), %%mm7 ;\n"
290 " movq %%mm6, 48(%1) ;\n"
291 " movq %%mm7, 56(%1) ;\n"
292
293 " addl $64, %1 ;\n"
294 " addl $64, %2 ;\n"
295 " decl %0 ;\n"
296 " jnz 1b ;\n"
297 : "+r" (lines),
298 "+r" (p1), "+r" (p2)
299 :
300 : "memory");
301
302 kernel_fpu_end();
303 }
304
305 static void
xor_p5_mmx_3(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2,const unsigned long * __restrict p3)306 xor_p5_mmx_3(unsigned long bytes, unsigned long * __restrict p1,
307 const unsigned long * __restrict p2,
308 const unsigned long * __restrict p3)
309 {
310 unsigned long lines = bytes >> 6;
311
312 kernel_fpu_begin();
313
314 asm volatile(
315 " .align 32,0x90 ;\n"
316 " 1: ;\n"
317 " movq (%1), %%mm0 ;\n"
318 " movq 8(%1), %%mm1 ;\n"
319 " pxor (%2), %%mm0 ;\n"
320 " movq 16(%1), %%mm2 ;\n"
321 " pxor 8(%2), %%mm1 ;\n"
322 " pxor (%3), %%mm0 ;\n"
323 " pxor 16(%2), %%mm2 ;\n"
324 " movq %%mm0, (%1) ;\n"
325 " pxor 8(%3), %%mm1 ;\n"
326 " pxor 16(%3), %%mm2 ;\n"
327 " movq 24(%1), %%mm3 ;\n"
328 " movq %%mm1, 8(%1) ;\n"
329 " movq 32(%1), %%mm4 ;\n"
330 " movq 40(%1), %%mm5 ;\n"
331 " pxor 24(%2), %%mm3 ;\n"
332 " movq %%mm2, 16(%1) ;\n"
333 " pxor 32(%2), %%mm4 ;\n"
334 " pxor 24(%3), %%mm3 ;\n"
335 " pxor 40(%2), %%mm5 ;\n"
336 " movq %%mm3, 24(%1) ;\n"
337 " pxor 32(%3), %%mm4 ;\n"
338 " pxor 40(%3), %%mm5 ;\n"
339 " movq 48(%1), %%mm6 ;\n"
340 " movq %%mm4, 32(%1) ;\n"
341 " movq 56(%1), %%mm7 ;\n"
342 " pxor 48(%2), %%mm6 ;\n"
343 " movq %%mm5, 40(%1) ;\n"
344 " pxor 56(%2), %%mm7 ;\n"
345 " pxor 48(%3), %%mm6 ;\n"
346 " pxor 56(%3), %%mm7 ;\n"
347 " movq %%mm6, 48(%1) ;\n"
348 " movq %%mm7, 56(%1) ;\n"
349
350 " addl $64, %1 ;\n"
351 " addl $64, %2 ;\n"
352 " addl $64, %3 ;\n"
353 " decl %0 ;\n"
354 " jnz 1b ;\n"
355 : "+r" (lines),
356 "+r" (p1), "+r" (p2), "+r" (p3)
357 :
358 : "memory" );
359
360 kernel_fpu_end();
361 }
362
363 static void
xor_p5_mmx_4(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2,const unsigned long * __restrict p3,const unsigned long * __restrict p4)364 xor_p5_mmx_4(unsigned long bytes, unsigned long * __restrict p1,
365 const unsigned long * __restrict p2,
366 const unsigned long * __restrict p3,
367 const unsigned long * __restrict p4)
368 {
369 unsigned long lines = bytes >> 6;
370
371 kernel_fpu_begin();
372
373 asm volatile(
374 " .align 32,0x90 ;\n"
375 " 1: ;\n"
376 " movq (%1), %%mm0 ;\n"
377 " movq 8(%1), %%mm1 ;\n"
378 " pxor (%2), %%mm0 ;\n"
379 " movq 16(%1), %%mm2 ;\n"
380 " pxor 8(%2), %%mm1 ;\n"
381 " pxor (%3), %%mm0 ;\n"
382 " pxor 16(%2), %%mm2 ;\n"
383 " pxor 8(%3), %%mm1 ;\n"
384 " pxor (%4), %%mm0 ;\n"
385 " movq 24(%1), %%mm3 ;\n"
386 " pxor 16(%3), %%mm2 ;\n"
387 " pxor 8(%4), %%mm1 ;\n"
388 " movq %%mm0, (%1) ;\n"
389 " movq 32(%1), %%mm4 ;\n"
390 " pxor 24(%2), %%mm3 ;\n"
391 " pxor 16(%4), %%mm2 ;\n"
392 " movq %%mm1, 8(%1) ;\n"
393 " movq 40(%1), %%mm5 ;\n"
394 " pxor 32(%2), %%mm4 ;\n"
395 " pxor 24(%3), %%mm3 ;\n"
396 " movq %%mm2, 16(%1) ;\n"
397 " pxor 40(%2), %%mm5 ;\n"
398 " pxor 32(%3), %%mm4 ;\n"
399 " pxor 24(%4), %%mm3 ;\n"
400 " movq %%mm3, 24(%1) ;\n"
401 " movq 56(%1), %%mm7 ;\n"
402 " movq 48(%1), %%mm6 ;\n"
403 " pxor 40(%3), %%mm5 ;\n"
404 " pxor 32(%4), %%mm4 ;\n"
405 " pxor 48(%2), %%mm6 ;\n"
406 " movq %%mm4, 32(%1) ;\n"
407 " pxor 56(%2), %%mm7 ;\n"
408 " pxor 40(%4), %%mm5 ;\n"
409 " pxor 48(%3), %%mm6 ;\n"
410 " pxor 56(%3), %%mm7 ;\n"
411 " movq %%mm5, 40(%1) ;\n"
412 " pxor 48(%4), %%mm6 ;\n"
413 " pxor 56(%4), %%mm7 ;\n"
414 " movq %%mm6, 48(%1) ;\n"
415 " movq %%mm7, 56(%1) ;\n"
416
417 " addl $64, %1 ;\n"
418 " addl $64, %2 ;\n"
419 " addl $64, %3 ;\n"
420 " addl $64, %4 ;\n"
421 " decl %0 ;\n"
422 " jnz 1b ;\n"
423 : "+r" (lines),
424 "+r" (p1), "+r" (p2), "+r" (p3), "+r" (p4)
425 :
426 : "memory");
427
428 kernel_fpu_end();
429 }
430
431 static void
xor_p5_mmx_5(unsigned long bytes,unsigned long * __restrict p1,const unsigned long * __restrict p2,const unsigned long * __restrict p3,const unsigned long * __restrict p4,const unsigned long * __restrict p5)432 xor_p5_mmx_5(unsigned long bytes, unsigned long * __restrict p1,
433 const unsigned long * __restrict p2,
434 const unsigned long * __restrict p3,
435 const unsigned long * __restrict p4,
436 const unsigned long * __restrict p5)
437 {
438 unsigned long lines = bytes >> 6;
439
440 kernel_fpu_begin();
441
442 /* Make sure GCC forgets anything it knows about p4 or p5,
443 such that it won't pass to the asm volatile below a
444 register that is shared with any other variable. That's
445 because we modify p4 and p5 there, but we can't mark them
446 as read/write, otherwise we'd overflow the 10-asm-operands
447 limit of GCC < 3.1. */
448 asm("" : "+r" (p4), "+r" (p5));
449
450 asm volatile(
451 " .align 32,0x90 ;\n"
452 " 1: ;\n"
453 " movq (%1), %%mm0 ;\n"
454 " movq 8(%1), %%mm1 ;\n"
455 " pxor (%2), %%mm0 ;\n"
456 " pxor 8(%2), %%mm1 ;\n"
457 " movq 16(%1), %%mm2 ;\n"
458 " pxor (%3), %%mm0 ;\n"
459 " pxor 8(%3), %%mm1 ;\n"
460 " pxor 16(%2), %%mm2 ;\n"
461 " pxor (%4), %%mm0 ;\n"
462 " pxor 8(%4), %%mm1 ;\n"
463 " pxor 16(%3), %%mm2 ;\n"
464 " movq 24(%1), %%mm3 ;\n"
465 " pxor (%5), %%mm0 ;\n"
466 " pxor 8(%5), %%mm1 ;\n"
467 " movq %%mm0, (%1) ;\n"
468 " pxor 16(%4), %%mm2 ;\n"
469 " pxor 24(%2), %%mm3 ;\n"
470 " movq %%mm1, 8(%1) ;\n"
471 " pxor 16(%5), %%mm2 ;\n"
472 " pxor 24(%3), %%mm3 ;\n"
473 " movq 32(%1), %%mm4 ;\n"
474 " movq %%mm2, 16(%1) ;\n"
475 " pxor 24(%4), %%mm3 ;\n"
476 " pxor 32(%2), %%mm4 ;\n"
477 " movq 40(%1), %%mm5 ;\n"
478 " pxor 24(%5), %%mm3 ;\n"
479 " pxor 32(%3), %%mm4 ;\n"
480 " pxor 40(%2), %%mm5 ;\n"
481 " movq %%mm3, 24(%1) ;\n"
482 " pxor 32(%4), %%mm4 ;\n"
483 " pxor 40(%3), %%mm5 ;\n"
484 " movq 48(%1), %%mm6 ;\n"
485 " movq 56(%1), %%mm7 ;\n"
486 " pxor 32(%5), %%mm4 ;\n"
487 " pxor 40(%4), %%mm5 ;\n"
488 " pxor 48(%2), %%mm6 ;\n"
489 " pxor 56(%2), %%mm7 ;\n"
490 " movq %%mm4, 32(%1) ;\n"
491 " pxor 48(%3), %%mm6 ;\n"
492 " pxor 56(%3), %%mm7 ;\n"
493 " pxor 40(%5), %%mm5 ;\n"
494 " pxor 48(%4), %%mm6 ;\n"
495 " pxor 56(%4), %%mm7 ;\n"
496 " movq %%mm5, 40(%1) ;\n"
497 " pxor 48(%5), %%mm6 ;\n"
498 " pxor 56(%5), %%mm7 ;\n"
499 " movq %%mm6, 48(%1) ;\n"
500 " movq %%mm7, 56(%1) ;\n"
501
502 " addl $64, %1 ;\n"
503 " addl $64, %2 ;\n"
504 " addl $64, %3 ;\n"
505 " addl $64, %4 ;\n"
506 " addl $64, %5 ;\n"
507 " decl %0 ;\n"
508 " jnz 1b ;\n"
509 : "+r" (lines),
510 "+r" (p1), "+r" (p2), "+r" (p3)
511 : "r" (p4), "r" (p5)
512 : "memory");
513
514 /* p4 and p5 were modified, and now the variables are dead.
515 Clobber them just to be sure nobody does something stupid
516 like assuming they have some legal value. */
517 asm("" : "=r" (p4), "=r" (p5));
518
519 kernel_fpu_end();
520 }
521
522 static struct xor_block_template xor_block_pII_mmx = {
523 .name = "pII_mmx",
524 .do_2 = xor_pII_mmx_2,
525 .do_3 = xor_pII_mmx_3,
526 .do_4 = xor_pII_mmx_4,
527 .do_5 = xor_pII_mmx_5,
528 };
529
530 static struct xor_block_template xor_block_p5_mmx = {
531 .name = "p5_mmx",
532 .do_2 = xor_p5_mmx_2,
533 .do_3 = xor_p5_mmx_3,
534 .do_4 = xor_p5_mmx_4,
535 .do_5 = xor_p5_mmx_5,
536 };
537
538 static struct xor_block_template xor_block_pIII_sse = {
539 .name = "pIII_sse",
540 .do_2 = xor_sse_2,
541 .do_3 = xor_sse_3,
542 .do_4 = xor_sse_4,
543 .do_5 = xor_sse_5,
544 };
545
546 /* Also try the AVX routines */
547 #include <asm/xor_avx.h>
548
549 /* Also try the generic routines. */
550 #include <asm-generic/xor.h>
551
552 /* We force the use of the SSE xor block because it can write around L2.
553 We may also be able to load into the L1 only depending on how the cpu
554 deals with a load to a line that is being prefetched. */
555 #undef XOR_TRY_TEMPLATES
556 #define XOR_TRY_TEMPLATES \
557 do { \
558 AVX_XOR_SPEED; \
559 if (boot_cpu_has(X86_FEATURE_XMM)) { \
560 xor_speed(&xor_block_pIII_sse); \
561 xor_speed(&xor_block_sse_pf64); \
562 } else if (boot_cpu_has(X86_FEATURE_MMX)) { \
563 xor_speed(&xor_block_pII_mmx); \
564 xor_speed(&xor_block_p5_mmx); \
565 } else { \
566 xor_speed(&xor_block_8regs); \
567 xor_speed(&xor_block_8regs_p); \
568 xor_speed(&xor_block_32regs); \
569 xor_speed(&xor_block_32regs_p); \
570 } \
571 } while (0)
572
573 #endif /* _ASM_X86_XOR_32_H */
574