]>
Commit | Line | Data |
---|---|---|
83ffe9cd | 1 | /* Copyright (C) 2012-2023 Free Software Foundation, Inc. |
0969ec7d EB |
2 | |
3 | This file is part of GCC. | |
4 | ||
5 | GCC is free software; you can redistribute it and/or modify it | |
6 | under the terms of the GNU General Public License as published by | |
7 | the Free Software Foundation; either version 3, or (at your option) | |
8 | any later version. | |
9 | ||
10 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
11 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
12 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
13 | License for more details. | |
14 | ||
15 | Under Section 7 of GPL version 3, you are granted additional | |
16 | permissions described in the GCC Runtime Library Exception, version | |
17 | 3.1, as published by the Free Software Foundation. | |
18 | ||
19 | You should have received a copy of the GNU General Public License and | |
20 | a copy of the GCC Runtime Library Exception along with this program; | |
21 | see the files COPYING3 and COPYING.RUNTIME respectively. If not, see | |
22 | <http://www.gnu.org/licenses/>. */ | |
23 | ||
24 | /* This file must be kept in sync with newlib/libc/machine/visium/memcpy.c */ | |
25 | ||
26 | #include <stddef.h> | |
27 | #include "memcpy.h" | |
28 | ||
29 | #define INST_BARRIER __asm__ __volatile__ ("":::"memory"); | |
30 | ||
31 | #define MOVE_32_OBJECTS(in,out) \ | |
32 | do { \ | |
33 | INST_BARRIER \ | |
34 | m0 = in [0]; \ | |
35 | m1 = in [1]; \ | |
36 | m2 = in [2]; \ | |
37 | m3 = in [3]; \ | |
38 | out [0] = m0; \ | |
39 | out [1] = m1; \ | |
40 | out [2] = m2; \ | |
41 | out [3] = m3; \ | |
42 | INST_BARRIER \ | |
43 | m0 = in [4]; \ | |
44 | m1 = in [5]; \ | |
45 | m2 = in [6]; \ | |
46 | m3 = in [7]; \ | |
47 | out [4] = m0; \ | |
48 | out [5] = m1; \ | |
49 | out [6] = m2; \ | |
50 | out [7] = m3; \ | |
51 | INST_BARRIER \ | |
52 | m0 = in [8]; \ | |
53 | m1 = in [9]; \ | |
54 | m2 = in [10]; \ | |
55 | m3 = in [11]; \ | |
56 | out [8] = m0; \ | |
57 | out [9] = m1; \ | |
58 | out [10] = m2; \ | |
59 | out [11] = m3; \ | |
60 | INST_BARRIER \ | |
61 | m0 = in [12]; \ | |
62 | m1 = in [13]; \ | |
63 | m2 = in [14]; \ | |
64 | m3 = in [15]; \ | |
65 | out [12] = m0; \ | |
66 | out [13] = m1; \ | |
67 | out [14] = m2; \ | |
68 | out [15] = m3; \ | |
69 | INST_BARRIER \ | |
70 | m0 = in [16]; \ | |
71 | m1 = in [17]; \ | |
72 | m2 = in [18]; \ | |
73 | m3 = in [19]; \ | |
74 | out [16] = m0; \ | |
75 | out [17] = m1; \ | |
76 | out [18] = m2; \ | |
77 | out [19] = m3; \ | |
78 | INST_BARRIER \ | |
79 | m0 = in [20]; \ | |
80 | m1 = in [21]; \ | |
81 | m2 = in [22]; \ | |
82 | m3 = in [23]; \ | |
83 | out [20] = m0; \ | |
84 | out [21] = m1; \ | |
85 | out [22] = m2; \ | |
86 | out [23] = m3; \ | |
87 | INST_BARRIER \ | |
88 | m0 = in [24]; \ | |
89 | m1 = in [25]; \ | |
90 | m2 = in [26]; \ | |
91 | m3 = in [27]; \ | |
92 | out [24] = m0; \ | |
93 | out [25] = m1; \ | |
94 | out [26] = m2; \ | |
95 | out [27] = m3; \ | |
96 | INST_BARRIER \ | |
97 | m0 = in [28]; \ | |
98 | m1 = in [29]; \ | |
99 | m2 = in [30]; \ | |
100 | m3 = in [31]; \ | |
101 | out [28] = m0; \ | |
102 | out [29] = m1; \ | |
103 | out [30] = m2; \ | |
104 | out [31] = m3; \ | |
105 | INST_BARRIER \ | |
106 | in += 32; \ | |
107 | out += 32; \ | |
108 | } while(0) | |
109 | ||
110 | #define MOVE_16_OBJECTS(in,out) \ | |
111 | do { \ | |
112 | INST_BARRIER \ | |
113 | m0 = in [0]; \ | |
114 | m1 = in [1]; \ | |
115 | m2 = in [2]; \ | |
116 | m3 = in [3]; \ | |
117 | out [0] = m0; \ | |
118 | out [1] = m1; \ | |
119 | out [2] = m2; \ | |
120 | out [3] = m3; \ | |
121 | INST_BARRIER \ | |
122 | m0 = in [4]; \ | |
123 | m1 = in [5]; \ | |
124 | m2 = in [6]; \ | |
125 | m3 = in [7]; \ | |
126 | out [4] = m0; \ | |
127 | out [5] = m1; \ | |
128 | out [6] = m2; \ | |
129 | out [7] = m3; \ | |
130 | INST_BARRIER \ | |
131 | m0 = in [8]; \ | |
132 | m1 = in [9]; \ | |
133 | m2 = in [10]; \ | |
134 | m3 = in [11]; \ | |
135 | out [8] = m0; \ | |
136 | out [9] = m1; \ | |
137 | out [10] = m2; \ | |
138 | out [11] = m3; \ | |
139 | INST_BARRIER \ | |
140 | m0 = in [12]; \ | |
141 | m1 = in [13]; \ | |
142 | m2 = in [14]; \ | |
143 | m3 = in [15]; \ | |
144 | out [12] = m0; \ | |
145 | out [13] = m1; \ | |
146 | out [14] = m2; \ | |
147 | out [15] = m3; \ | |
148 | INST_BARRIER \ | |
149 | in += 16; \ | |
150 | out += 16; \ | |
151 | } while(0) | |
152 | ||
153 | #define MOVE_12_OBJECTS(in,out) \ | |
154 | do { \ | |
155 | INST_BARRIER \ | |
156 | m0 = in [0]; \ | |
157 | m1 = in [1]; \ | |
158 | m2 = in [2]; \ | |
159 | m3 = in [3]; \ | |
160 | out [0] = m0; \ | |
161 | out [1] = m1; \ | |
162 | out [2] = m2; \ | |
163 | out [3] = m3; \ | |
164 | INST_BARRIER \ | |
165 | m0 = in [4]; \ | |
166 | m1 = in [5]; \ | |
167 | m2 = in [6]; \ | |
168 | m3 = in [7]; \ | |
169 | out [4] = m0; \ | |
170 | out [5] = m1; \ | |
171 | out [6] = m2; \ | |
172 | out [7] = m3; \ | |
173 | INST_BARRIER \ | |
174 | m0 = in [8]; \ | |
175 | m1 = in [9]; \ | |
176 | m2 = in [10]; \ | |
177 | m3 = in [11]; \ | |
178 | out [8] = m0; \ | |
179 | out [9] = m1; \ | |
180 | out [10] = m2; \ | |
181 | out [11] = m3; \ | |
182 | INST_BARRIER \ | |
183 | in += 12; \ | |
184 | out += 12; \ | |
185 | } while(0) | |
186 | ||
187 | #define MOVE_11_OBJECTS(in,out) \ | |
188 | do { \ | |
189 | INST_BARRIER \ | |
190 | m0 = in [0]; \ | |
191 | m1 = in [1]; \ | |
192 | m2 = in [2]; \ | |
193 | m3 = in [3]; \ | |
194 | out [0] = m0; \ | |
195 | out [1] = m1; \ | |
196 | out [2] = m2; \ | |
197 | out [3] = m3; \ | |
198 | INST_BARRIER \ | |
199 | m0 = in [4]; \ | |
200 | m1 = in [5]; \ | |
201 | m2 = in [6]; \ | |
202 | m3 = in [7]; \ | |
203 | out [4] = m0; \ | |
204 | out [5] = m1; \ | |
205 | out [6] = m2; \ | |
206 | out [7] = m3; \ | |
207 | INST_BARRIER \ | |
208 | m0 = in [8]; \ | |
209 | m1 = in [9]; \ | |
210 | m2 = in [10]; \ | |
211 | out [8] = m0; \ | |
212 | out [9] = m1; \ | |
213 | out [10] = m2; \ | |
214 | INST_BARRIER \ | |
215 | in += 11; \ | |
216 | out += 11; \ | |
217 | } while(0) | |
218 | ||
219 | #define MOVE_10_OBJECTS(in,out) \ | |
220 | do { \ | |
221 | INST_BARRIER \ | |
222 | m0 = in [0]; \ | |
223 | m1 = in [1]; \ | |
224 | m2 = in [2]; \ | |
225 | m3 = in [3]; \ | |
226 | out [0] = m0; \ | |
227 | out [1] = m1; \ | |
228 | out [2] = m2; \ | |
229 | out [3] = m3; \ | |
230 | INST_BARRIER \ | |
231 | m0 = in [4]; \ | |
232 | m1 = in [5]; \ | |
233 | m2 = in [6]; \ | |
234 | m3 = in [7]; \ | |
235 | out [4] = m0; \ | |
236 | m0 = in [8]; \ | |
237 | out [5] = m1; \ | |
238 | m1 = in [9]; \ | |
239 | out [6] = m2; \ | |
240 | out [7] = m3; \ | |
241 | out [8] = m0; \ | |
242 | out [9] = m1; \ | |
243 | INST_BARRIER \ | |
244 | in += 10; \ | |
245 | out += 10; \ | |
246 | } while(0) | |
247 | ||
248 | #define MOVE_9_OBJECTS(in,out) \ | |
249 | do { \ | |
250 | INST_BARRIER \ | |
251 | m0 = in [0]; \ | |
252 | m1 = in [1]; \ | |
253 | m2 = in [2]; \ | |
254 | m3 = in [3]; \ | |
255 | out [0] = m0; \ | |
256 | out [1] = m1; \ | |
257 | out [2] = m2; \ | |
258 | out [3] = m3; \ | |
259 | INST_BARRIER \ | |
260 | m0 = in [4]; \ | |
261 | m1 = in [5]; \ | |
262 | m2 = in [6]; \ | |
263 | m3 = in [7]; \ | |
264 | out [4] = m0; \ | |
265 | out [5] = m1; \ | |
266 | out [6] = m2; \ | |
267 | out [7] = m3; \ | |
268 | INST_BARRIER \ | |
269 | m0 = in [8]; \ | |
270 | out [8] = m0; \ | |
271 | in += 9; \ | |
272 | out += 9; \ | |
273 | } while(0) | |
274 | ||
275 | #define MOVE_8_OBJECTS(in,out) \ | |
276 | do { \ | |
277 | INST_BARRIER \ | |
278 | m0 = in [0]; \ | |
279 | m1 = in [1]; \ | |
280 | m2 = in [2]; \ | |
281 | m3 = in [3]; \ | |
282 | out [0] = m0; \ | |
283 | out [1] = m1; \ | |
284 | out [2] = m2; \ | |
285 | out [3] = m3; \ | |
286 | INST_BARRIER \ | |
287 | m0 = in [4]; \ | |
288 | m1 = in [5]; \ | |
289 | m2 = in [6]; \ | |
290 | m3 = in [7]; \ | |
291 | out [4] = m0; \ | |
292 | out [5] = m1; \ | |
293 | out [6] = m2; \ | |
294 | out [7] = m3; \ | |
295 | INST_BARRIER \ | |
296 | in += 8; \ | |
297 | out += 8; \ | |
298 | } while(0) | |
299 | ||
300 | #define MOVE_7_OBJECTS(in,out) \ | |
301 | do { \ | |
302 | INST_BARRIER \ | |
303 | m0 = in [0]; \ | |
304 | m1 = in [1]; \ | |
305 | m2 = in [2]; \ | |
306 | m3 = in [3]; \ | |
307 | out [0] = m0; \ | |
308 | out [1] = m1; \ | |
309 | out [2] = m2; \ | |
310 | out [3] = m3; \ | |
311 | INST_BARRIER \ | |
312 | m0 = in [4]; \ | |
313 | m1 = in [5]; \ | |
314 | m2 = in [6]; \ | |
315 | out [4] = m0; \ | |
316 | out [5] = m1; \ | |
317 | out [6] = m2; \ | |
318 | INST_BARRIER \ | |
319 | in += 7; \ | |
320 | out += 7; \ | |
321 | } while(0) | |
322 | ||
323 | #define MOVE_6_OBJECTS(in,out) \ | |
324 | do { \ | |
325 | INST_BARRIER \ | |
326 | m0 = in [0]; \ | |
327 | m1 = in [1]; \ | |
328 | m2 = in [2]; \ | |
329 | m3 = in [3]; \ | |
330 | out [0] = m0; \ | |
331 | INST_BARRIER \ | |
332 | m0 = in [4]; \ | |
333 | out [1] = m1; \ | |
334 | INST_BARRIER \ | |
335 | m1 = in [5]; \ | |
336 | out [2] = m2; \ | |
337 | out [3] = m3; \ | |
338 | out [4] = m0; \ | |
339 | out [5] = m1; \ | |
340 | INST_BARRIER \ | |
341 | in += 6; \ | |
342 | out += 6; \ | |
343 | } while(0) | |
344 | ||
345 | #define MOVE_5_OBJECTS(in,out) \ | |
346 | do { \ | |
347 | INST_BARRIER \ | |
348 | m0 = in [0]; \ | |
349 | m1 = in [1]; \ | |
350 | m2 = in [2]; \ | |
351 | m3 = in [3]; \ | |
352 | INST_BARRIER \ | |
353 | out [0] = m0; \ | |
354 | m0 = in [4]; \ | |
355 | INST_BARRIER \ | |
356 | out [1] = m1; \ | |
357 | out [2] = m2; \ | |
358 | out [3] = m3; \ | |
359 | out [4] = m0; \ | |
360 | INST_BARRIER \ | |
361 | in += 5; \ | |
362 | out += 5; \ | |
363 | } while(0) | |
364 | ||
365 | #define MOVE_4_OBJECTS(in,out) \ | |
366 | do { \ | |
367 | INST_BARRIER \ | |
368 | m0 = in [0]; \ | |
369 | m1 = in [1]; \ | |
370 | m2 = in [2]; \ | |
371 | m3 = in [3]; \ | |
372 | out [0] = m0; \ | |
373 | out [1] = m1; \ | |
374 | out [2] = m2; \ | |
375 | out [3] = m3; \ | |
376 | INST_BARRIER \ | |
377 | in += 4; \ | |
378 | out += 4; \ | |
379 | } while(0) | |
380 | ||
381 | #define MOVE_3_OBJECTS(in,out) \ | |
382 | do { \ | |
383 | INST_BARRIER \ | |
384 | m0 = in [0]; \ | |
385 | m1 = in [1]; \ | |
386 | m2 = in [2]; \ | |
387 | out [0] = m0; \ | |
388 | out [1] = m1; \ | |
389 | out [2] = m2; \ | |
390 | INST_BARRIER \ | |
391 | in += 3; \ | |
392 | out += 3; \ | |
393 | } while(0) | |
394 | ||
395 | #define MOVE_2_OBJECTS(in,out) \ | |
396 | do { \ | |
397 | INST_BARRIER \ | |
398 | m0 = in [0]; \ | |
399 | m1 = in [1]; \ | |
400 | out [0] = m0; \ | |
401 | out [1] = m1; \ | |
402 | INST_BARRIER \ | |
403 | in += 2; \ | |
404 | out += 2; \ | |
405 | } while(0) | |
406 | ||
407 | #define MOVE_1_OBJECT(in,out) \ | |
408 | do { \ | |
409 | INST_BARRIER \ | |
410 | m0 = in [0]; \ | |
411 | out [0] = m0; \ | |
412 | INST_BARRIER \ | |
413 | in += 1; \ | |
414 | out += 1; \ | |
415 | } while(0) | |
416 | ||
417 | ||
418 | static inline void | |
419 | __int_memcpy (void *__restrict s1, const void *__restrict s2, size_t n) | |
420 | { | |
421 | int value = n; | |
422 | int loop_var; | |
423 | const int *in = s2; | |
424 | int *out = s1; | |
425 | int count; | |
426 | int m0,m1,m2,m3; | |
427 | ||
428 | /* This code currently give a stall for any value with a 1->2 in the low 5 | |
429 | bits, i.e. 1,2, 33,34 ? not acceptable! */ | |
430 | switch (value & 0x1f) | |
431 | { | |
432 | case 0: | |
433 | break; | |
434 | case 1: | |
435 | MOVE_1_OBJECT (in, out); | |
436 | break; | |
437 | case 2: | |
438 | MOVE_2_OBJECTS (in, out); | |
439 | break; | |
440 | case 3: | |
441 | MOVE_3_OBJECTS (in, out); | |
442 | break; | |
443 | case 4: | |
444 | MOVE_4_OBJECTS (in, out); | |
445 | break; | |
446 | case 5: | |
447 | MOVE_5_OBJECTS (in, out); | |
448 | break; | |
449 | case 6: | |
450 | MOVE_6_OBJECTS (in, out); | |
451 | break; | |
452 | case 7: | |
453 | MOVE_7_OBJECTS (in, out); | |
454 | break; | |
455 | case 8: | |
456 | MOVE_8_OBJECTS (in, out); | |
457 | break; | |
458 | case 9: | |
459 | MOVE_9_OBJECTS (in, out); | |
460 | break; | |
461 | case 10: | |
462 | MOVE_10_OBJECTS (in, out); | |
463 | break; | |
464 | case 11: | |
465 | MOVE_11_OBJECTS (in, out); | |
466 | break; | |
467 | case 12: | |
468 | MOVE_12_OBJECTS (in, out); | |
469 | break; | |
470 | case 13: | |
471 | MOVE_9_OBJECTS (in, out); | |
472 | MOVE_4_OBJECTS (in, out); | |
473 | break; | |
474 | case 14: | |
475 | MOVE_12_OBJECTS (in, out); | |
476 | MOVE_2_OBJECTS (in, out); | |
477 | break; | |
478 | case 15: | |
479 | MOVE_11_OBJECTS (in, out); | |
480 | MOVE_4_OBJECTS (in, out); | |
481 | break; | |
482 | case 16: | |
483 | MOVE_16_OBJECTS (in, out); | |
484 | break; | |
485 | case 17: | |
486 | MOVE_11_OBJECTS (in, out); | |
487 | MOVE_6_OBJECTS (in, out); | |
488 | break; | |
489 | case 18: | |
490 | MOVE_9_OBJECTS (in, out); | |
491 | MOVE_9_OBJECTS (in, out); | |
492 | break; | |
493 | case 19: | |
494 | MOVE_16_OBJECTS (in, out); | |
495 | MOVE_3_OBJECTS (in, out); | |
496 | break; | |
497 | case 20: | |
498 | MOVE_16_OBJECTS (in, out); | |
499 | MOVE_4_OBJECTS (in, out); | |
500 | break; | |
501 | case 21: | |
502 | MOVE_16_OBJECTS (in, out); | |
503 | MOVE_5_OBJECTS (in, out); | |
504 | break; | |
505 | case 22: | |
506 | MOVE_16_OBJECTS (in, out); | |
507 | MOVE_6_OBJECTS (in, out); | |
508 | break; | |
509 | case 23: | |
510 | MOVE_16_OBJECTS (in, out); | |
511 | MOVE_7_OBJECTS (in, out); | |
512 | break; | |
513 | case 24: | |
514 | MOVE_16_OBJECTS (in, out); | |
515 | MOVE_8_OBJECTS (in, out); | |
516 | break; | |
517 | case 25: | |
518 | MOVE_16_OBJECTS (in, out); | |
519 | MOVE_9_OBJECTS (in, out); | |
520 | break; | |
521 | case 26: | |
522 | MOVE_16_OBJECTS (in, out); | |
523 | MOVE_10_OBJECTS (in, out); | |
524 | break; | |
525 | case 27: | |
526 | MOVE_16_OBJECTS (in, out); | |
527 | MOVE_11_OBJECTS (in, out); | |
528 | break; | |
529 | case 28: | |
530 | MOVE_16_OBJECTS (in, out); | |
531 | MOVE_8_OBJECTS (in, out); | |
532 | MOVE_4_OBJECTS (in, out); | |
533 | break; | |
534 | case 29: | |
535 | MOVE_16_OBJECTS (in, out); | |
536 | MOVE_9_OBJECTS (in, out); | |
537 | MOVE_4_OBJECTS (in, out); | |
538 | break; | |
539 | case 30: | |
540 | MOVE_16_OBJECTS (in, out); | |
541 | MOVE_12_OBJECTS (in, out); | |
542 | MOVE_2_OBJECTS (in, out); | |
543 | break; | |
544 | case 31: | |
545 | MOVE_16_OBJECTS (in, out); | |
546 | MOVE_11_OBJECTS (in, out); | |
547 | MOVE_4_OBJECTS (in, out); | |
548 | break; | |
549 | } | |
550 | ||
551 | /* This loop governs the asmptoptic behaviour of this algorithm, for long | |
552 | word copies. */ | |
553 | count = value >> 5; | |
554 | for (loop_var = 0; loop_var < count; loop_var++) | |
555 | MOVE_32_OBJECTS (in, out); | |
556 | } | |
557 | ||
558 | static inline void | |
559 | __shrt_int_memcpy (void *__restrict s1, const void *__restrict s2, size_t n) | |
560 | { | |
561 | int value = n; | |
562 | int loop_var; | |
563 | const short int *in = s2; | |
564 | int short *out = s1; | |
565 | int count; | |
566 | int m0,m1,m2,m3; | |
567 | ||
568 | /* This code currently give a stall for any value with a 1->2 in the low 5 | |
569 | bits, i.e. 1,2, 33,34 ? not acceptable! */ | |
570 | switch (value & 0x1f) | |
571 | { | |
572 | case 0: | |
573 | break; | |
574 | case 1: | |
575 | MOVE_1_OBJECT (in, out); | |
576 | break; | |
577 | case 2: | |
578 | MOVE_2_OBJECTS (in, out); | |
579 | break; | |
580 | case 3: | |
581 | MOVE_3_OBJECTS (in, out); | |
582 | break; | |
583 | case 4: | |
584 | MOVE_4_OBJECTS (in, out); | |
585 | break; | |
586 | case 5: | |
587 | MOVE_5_OBJECTS (in, out); | |
588 | break; | |
589 | case 6: | |
590 | MOVE_6_OBJECTS (in, out); | |
591 | break; | |
592 | case 7: | |
593 | MOVE_7_OBJECTS (in, out); | |
594 | break; | |
595 | case 8: | |
596 | MOVE_8_OBJECTS (in, out); | |
597 | break; | |
598 | case 9: | |
599 | MOVE_9_OBJECTS (in, out); | |
600 | break; | |
601 | case 10: | |
602 | MOVE_10_OBJECTS (in, out); | |
603 | break; | |
604 | case 11: | |
605 | MOVE_11_OBJECTS (in, out); | |
606 | break; | |
607 | case 12: | |
608 | MOVE_12_OBJECTS (in, out); | |
609 | break; | |
610 | case 13: | |
611 | MOVE_9_OBJECTS (in, out); | |
612 | MOVE_4_OBJECTS (in, out); | |
613 | break; | |
614 | case 14: | |
615 | MOVE_12_OBJECTS (in, out); | |
616 | MOVE_2_OBJECTS (in, out); | |
617 | break; | |
618 | case 15: | |
619 | MOVE_11_OBJECTS (in, out); | |
620 | MOVE_4_OBJECTS (in, out); | |
621 | break; | |
622 | case 16: | |
623 | MOVE_16_OBJECTS (in, out); | |
624 | break; | |
625 | case 17: | |
626 | MOVE_11_OBJECTS (in, out); | |
627 | MOVE_6_OBJECTS (in, out); | |
628 | break; | |
629 | case 18: | |
630 | MOVE_9_OBJECTS (in, out); | |
631 | MOVE_9_OBJECTS (in, out); | |
632 | break; | |
633 | case 19: | |
634 | MOVE_16_OBJECTS (in, out); | |
635 | MOVE_3_OBJECTS (in, out); | |
636 | break; | |
637 | case 20: | |
638 | MOVE_16_OBJECTS (in, out); | |
639 | MOVE_4_OBJECTS (in, out); | |
640 | break; | |
641 | case 21: | |
642 | MOVE_16_OBJECTS (in, out); | |
643 | MOVE_5_OBJECTS (in, out); | |
644 | break; | |
645 | case 22: | |
646 | MOVE_16_OBJECTS (in, out); | |
647 | MOVE_6_OBJECTS (in, out); | |
648 | break; | |
649 | case 23: | |
650 | MOVE_16_OBJECTS (in, out); | |
651 | MOVE_7_OBJECTS (in, out); | |
652 | break; | |
653 | case 24: | |
654 | MOVE_16_OBJECTS (in, out); | |
655 | MOVE_8_OBJECTS (in, out); | |
656 | break; | |
657 | case 25: | |
658 | MOVE_16_OBJECTS (in, out); | |
659 | MOVE_9_OBJECTS (in, out); | |
660 | break; | |
661 | case 26: | |
662 | MOVE_16_OBJECTS (in, out); | |
663 | MOVE_10_OBJECTS (in, out); | |
664 | break; | |
665 | case 27: | |
666 | MOVE_16_OBJECTS (in, out); | |
667 | MOVE_11_OBJECTS (in, out); | |
668 | break; | |
669 | case 28: | |
670 | MOVE_16_OBJECTS (in, out); | |
671 | MOVE_8_OBJECTS (in, out); | |
672 | MOVE_4_OBJECTS (in, out); | |
673 | break; | |
674 | case 29: | |
675 | MOVE_16_OBJECTS (in, out); | |
676 | MOVE_9_OBJECTS (in, out); | |
677 | MOVE_4_OBJECTS (in, out); | |
678 | break; | |
679 | case 30: | |
680 | MOVE_16_OBJECTS (in, out); | |
681 | MOVE_12_OBJECTS (in, out); | |
682 | MOVE_2_OBJECTS (in, out); | |
683 | break; | |
684 | case 31: | |
685 | MOVE_16_OBJECTS (in, out); | |
686 | MOVE_11_OBJECTS (in, out); | |
687 | MOVE_4_OBJECTS (in, out); | |
688 | break; | |
689 | } | |
690 | ||
691 | /* This loop governs the asmptoptic behaviour of this algorithm, for long | |
692 | word copies. */ | |
693 | count = value >> 5; | |
694 | for (loop_var = 0; loop_var < count; loop_var++) | |
695 | MOVE_32_OBJECTS (in, out); | |
696 | } | |
697 | ||
698 | ||
699 | static inline void | |
700 | __byte_memcpy (void *__restrict s1, const void *__restrict s2, size_t n) | |
701 | { | |
702 | int value = n; | |
703 | int loop_var; | |
704 | const char *in = s2; | |
705 | char *out = s1; | |
706 | int count; | |
707 | int m0,m1,m2,m3; | |
708 | ||
709 | /* This code currently give a stall for any value with a 1->2 in the low 5 | |
710 | bits, i.e. 1,2, 33,34 ? not acceptable! */ | |
711 | switch (value & 0x1f) | |
712 | { | |
713 | case 0: | |
714 | break; | |
715 | case 1: | |
716 | MOVE_1_OBJECT (in, out); | |
717 | break; | |
718 | case 2: | |
719 | MOVE_2_OBJECTS (in, out); | |
720 | break; | |
721 | case 3: | |
722 | MOVE_3_OBJECTS (in, out); | |
723 | break; | |
724 | case 4: | |
725 | MOVE_4_OBJECTS (in, out); | |
726 | break; | |
727 | case 5: | |
728 | MOVE_5_OBJECTS (in, out); | |
729 | break; | |
730 | case 6: | |
731 | MOVE_6_OBJECTS (in, out); | |
732 | break; | |
733 | case 7: | |
734 | MOVE_7_OBJECTS (in, out); | |
735 | break; | |
736 | case 8: | |
737 | MOVE_8_OBJECTS (in, out); | |
738 | break; | |
739 | case 9: | |
740 | MOVE_9_OBJECTS (in, out); | |
741 | break; | |
742 | case 10: | |
743 | MOVE_10_OBJECTS (in, out); | |
744 | break; | |
745 | case 11: | |
746 | MOVE_11_OBJECTS (in, out); | |
747 | break; | |
748 | case 12: | |
749 | MOVE_12_OBJECTS (in, out); | |
750 | break; | |
751 | case 13: | |
752 | MOVE_9_OBJECTS (in, out); | |
753 | MOVE_4_OBJECTS (in, out); | |
754 | break; | |
755 | case 14: | |
756 | MOVE_12_OBJECTS (in, out); | |
757 | MOVE_2_OBJECTS (in, out); | |
758 | break; | |
759 | case 15: | |
760 | MOVE_11_OBJECTS (in, out); | |
761 | MOVE_4_OBJECTS (in, out); | |
762 | break; | |
763 | case 16: | |
764 | MOVE_16_OBJECTS (in, out); | |
765 | break; | |
766 | case 17: | |
767 | MOVE_11_OBJECTS (in, out); | |
768 | MOVE_6_OBJECTS (in, out); | |
769 | break; | |
770 | case 18: | |
771 | MOVE_9_OBJECTS (in, out); | |
772 | MOVE_9_OBJECTS (in, out); | |
773 | break; | |
774 | case 19: | |
775 | MOVE_16_OBJECTS (in, out); | |
776 | MOVE_3_OBJECTS (in, out); | |
777 | break; | |
778 | case 20: | |
779 | MOVE_16_OBJECTS (in, out); | |
780 | MOVE_4_OBJECTS (in, out); | |
781 | break; | |
782 | case 21: | |
783 | MOVE_16_OBJECTS (in, out); | |
784 | MOVE_5_OBJECTS (in, out); | |
785 | break; | |
786 | case 22: | |
787 | MOVE_16_OBJECTS (in, out); | |
788 | MOVE_6_OBJECTS (in, out); | |
789 | break; | |
790 | case 23: | |
791 | MOVE_16_OBJECTS (in, out); | |
792 | MOVE_7_OBJECTS (in, out); | |
793 | break; | |
794 | case 24: | |
795 | MOVE_16_OBJECTS (in, out); | |
796 | MOVE_8_OBJECTS (in, out); | |
797 | break; | |
798 | case 25: | |
799 | MOVE_16_OBJECTS (in, out); | |
800 | MOVE_9_OBJECTS (in, out); | |
801 | break; | |
802 | case 26: | |
803 | MOVE_16_OBJECTS (in, out); | |
804 | MOVE_10_OBJECTS (in, out); | |
805 | break; | |
806 | case 27: | |
807 | MOVE_16_OBJECTS (in, out); | |
808 | MOVE_11_OBJECTS (in, out); | |
809 | break; | |
810 | case 28: | |
811 | MOVE_16_OBJECTS (in, out); | |
812 | MOVE_8_OBJECTS (in, out); | |
813 | MOVE_4_OBJECTS (in, out); | |
814 | break; | |
815 | case 29: | |
816 | MOVE_16_OBJECTS (in, out); | |
817 | MOVE_9_OBJECTS (in, out); | |
818 | MOVE_4_OBJECTS (in, out); | |
819 | break; | |
820 | case 30: | |
821 | MOVE_16_OBJECTS (in, out); | |
822 | MOVE_12_OBJECTS (in, out); | |
823 | MOVE_2_OBJECTS (in, out); | |
824 | break; | |
825 | case 31: | |
826 | MOVE_16_OBJECTS (in, out); | |
827 | MOVE_11_OBJECTS (in, out); | |
828 | MOVE_4_OBJECTS (in, out); | |
829 | break; | |
830 | } | |
831 | ||
832 | /* This loop governs the asmptoptic behaviour of this algorithm, for long | |
833 | word copies. */ | |
834 | count = value >> 5; | |
835 | for (loop_var = 0; loop_var < count; loop_var++) | |
836 | MOVE_32_OBJECTS (in, out); | |
837 | } | |
838 | ||
839 | ||
840 | /* Exposed interface. */ | |
841 | ||
842 | #ifndef __VISIUM_ARCH_BMI__ | |
843 | ||
844 | void | |
845 | __long_int_memcpy (void *__restrict s1, const void *__restrict s2, size_t n) | |
846 | { | |
847 | __int_memcpy (s1, s2, n); | |
848 | } | |
849 | ||
850 | #endif /* !__VISIUM_ARCH_BMI__ */ | |
851 | ||
852 | void | |
853 | __wrd_memcpy (void *__restrict s1, const void *__restrict s2, size_t n) | |
854 | { | |
855 | __shrt_int_memcpy (s1, s2, n); | |
856 | } | |
857 | ||
858 | void | |
859 | __byt_memcpy (void *__restrict s1, const void *__restrict s2, size_t n) | |
860 | { | |
861 | __byte_memcpy (s1, s2, n); | |
862 | } |