Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
tjh.dev
kernel
os
linux
1/*
2 * String handling functions for PowerPC.
3 *
4 * Copyright (C) 1996 Paul Mackerras.
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 */
11#include <asm/processor.h>
12#include <asm/cache.h>
13#include <asm/errno.h>
14#include <asm/ppc_asm.h>
15
16#define COPY_16_BYTES \
17 lwz r7,4(r4); \
18 lwz r8,8(r4); \
19 lwz r9,12(r4); \
20 lwzu r10,16(r4); \
21 stw r7,4(r6); \
22 stw r8,8(r6); \
23 stw r9,12(r6); \
24 stwu r10,16(r6)
25
26#define COPY_16_BYTES_WITHEX(n) \
278 ## n ## 0: \
28 lwz r7,4(r4); \
298 ## n ## 1: \
30 lwz r8,8(r4); \
318 ## n ## 2: \
32 lwz r9,12(r4); \
338 ## n ## 3: \
34 lwzu r10,16(r4); \
358 ## n ## 4: \
36 stw r7,4(r6); \
378 ## n ## 5: \
38 stw r8,8(r6); \
398 ## n ## 6: \
40 stw r9,12(r6); \
418 ## n ## 7: \
42 stwu r10,16(r6)
43
44#define COPY_16_BYTES_EXCODE(n) \
459 ## n ## 0: \
46 addi r5,r5,-(16 * n); \
47 b 104f; \
489 ## n ## 1: \
49 addi r5,r5,-(16 * n); \
50 b 105f; \
51.section __ex_table,"a"; \
52 .align 2; \
53 .long 8 ## n ## 0b,9 ## n ## 0b; \
54 .long 8 ## n ## 1b,9 ## n ## 0b; \
55 .long 8 ## n ## 2b,9 ## n ## 0b; \
56 .long 8 ## n ## 3b,9 ## n ## 0b; \
57 .long 8 ## n ## 4b,9 ## n ## 1b; \
58 .long 8 ## n ## 5b,9 ## n ## 1b; \
59 .long 8 ## n ## 6b,9 ## n ## 1b; \
60 .long 8 ## n ## 7b,9 ## n ## 1b; \
61 .text
62
63 .text
64 .stabs "arch/ppc/lib/",N_SO,0,0,0f
65 .stabs "string.S",N_SO,0,0,0f
66
67CACHELINE_BYTES = L1_CACHE_BYTES
68LG_CACHELINE_BYTES = L1_CACHE_SHIFT
69CACHELINE_MASK = (L1_CACHE_BYTES-1)
70
71_GLOBAL(strcpy)
72 addi r5,r3,-1
73 addi r4,r4,-1
741: lbzu r0,1(r4)
75 cmpwi 0,r0,0
76 stbu r0,1(r5)
77 bne 1b
78 blr
79
80/* This clears out any unused part of the destination buffer,
81 just as the libc version does. -- paulus */
82_GLOBAL(strncpy)
83 cmpwi 0,r5,0
84 beqlr
85 mtctr r5
86 addi r6,r3,-1
87 addi r4,r4,-1
881: lbzu r0,1(r4)
89 cmpwi 0,r0,0
90 stbu r0,1(r6)
91 bdnzf 2,1b /* dec ctr, branch if ctr != 0 && !cr0.eq */
92 bnelr /* if we didn't hit a null char, we're done */
93 mfctr r5
94 cmpwi 0,r5,0 /* any space left in destination buffer? */
95 beqlr /* we know r0 == 0 here */
962: stbu r0,1(r6) /* clear it out if so */
97 bdnz 2b
98 blr
99
100_GLOBAL(strcat)
101 addi r5,r3,-1
102 addi r4,r4,-1
1031: lbzu r0,1(r5)
104 cmpwi 0,r0,0
105 bne 1b
106 addi r5,r5,-1
1071: lbzu r0,1(r4)
108 cmpwi 0,r0,0
109 stbu r0,1(r5)
110 bne 1b
111 blr
112
113_GLOBAL(strcmp)
114 addi r5,r3,-1
115 addi r4,r4,-1
1161: lbzu r3,1(r5)
117 cmpwi 1,r3,0
118 lbzu r0,1(r4)
119 subf. r3,r0,r3
120 beqlr 1
121 beq 1b
122 blr
123
124_GLOBAL(strncmp)
125 PPC_LCMPI r5,0
126 beqlr
127 mtctr r5
128 addi r5,r3,-1
129 addi r4,r4,-1
1301: lbzu r3,1(r5)
131 cmpwi 1,r3,0
132 lbzu r0,1(r4)
133 subf. r3,r0,r3
134 beqlr 1
135 bdnzt eq,1b
136 blr
137
138_GLOBAL(strlen)
139 addi r4,r3,-1
1401: lbzu r0,1(r4)
141 cmpwi 0,r0,0
142 bne 1b
143 subf r3,r3,r4
144 blr
145
146/*
147 * Use dcbz on the complete cache lines in the destination
148 * to set them to zero. This requires that the destination
149 * area is cacheable. -- paulus
150 */
151_GLOBAL(cacheable_memzero)
152 mr r5,r4
153 li r4,0
154 addi r6,r3,-4
155 cmplwi 0,r5,4
156 blt 7f
157 stwu r4,4(r6)
158 beqlr
159 andi. r0,r6,3
160 add r5,r0,r5
161 subf r6,r0,r6
162 clrlwi r7,r6,32-LG_CACHELINE_BYTES
163 add r8,r7,r5
164 srwi r9,r8,LG_CACHELINE_BYTES
165 addic. r9,r9,-1 /* total number of complete cachelines */
166 ble 2f
167 xori r0,r7,CACHELINE_MASK & ~3
168 srwi. r0,r0,2
169 beq 3f
170 mtctr r0
1714: stwu r4,4(r6)
172 bdnz 4b
1733: mtctr r9
174 li r7,4
175#if !defined(CONFIG_8xx)
17610: dcbz r7,r6
177#else
17810: stw r4, 4(r6)
179 stw r4, 8(r6)
180 stw r4, 12(r6)
181 stw r4, 16(r6)
182#if CACHE_LINE_SIZE >= 32
183 stw r4, 20(r6)
184 stw r4, 24(r6)
185 stw r4, 28(r6)
186 stw r4, 32(r6)
187#endif /* CACHE_LINE_SIZE */
188#endif
189 addi r6,r6,CACHELINE_BYTES
190 bdnz 10b
191 clrlwi r5,r8,32-LG_CACHELINE_BYTES
192 addi r5,r5,4
1932: srwi r0,r5,2
194 mtctr r0
195 bdz 6f
1961: stwu r4,4(r6)
197 bdnz 1b
1986: andi. r5,r5,3
1997: cmpwi 0,r5,0
200 beqlr
201 mtctr r5
202 addi r6,r6,3
2038: stbu r4,1(r6)
204 bdnz 8b
205 blr
206
207_GLOBAL(memset)
208 rlwimi r4,r4,8,16,23
209 rlwimi r4,r4,16,0,15
210 addi r6,r3,-4
211 cmplwi 0,r5,4
212 blt 7f
213 stwu r4,4(r6)
214 beqlr
215 andi. r0,r6,3
216 add r5,r0,r5
217 subf r6,r0,r6
218 srwi r0,r5,2
219 mtctr r0
220 bdz 6f
2211: stwu r4,4(r6)
222 bdnz 1b
2236: andi. r5,r5,3
2247: cmpwi 0,r5,0
225 beqlr
226 mtctr r5
227 addi r6,r6,3
2288: stbu r4,1(r6)
229 bdnz 8b
230 blr
231
232/*
233 * This version uses dcbz on the complete cache lines in the
234 * destination area to reduce memory traffic. This requires that
235 * the destination area is cacheable.
236 * We only use this version if the source and dest don't overlap.
237 * -- paulus.
238 */
239_GLOBAL(cacheable_memcpy)
240 add r7,r3,r5 /* test if the src & dst overlap */
241 add r8,r4,r5
242 cmplw 0,r4,r7
243 cmplw 1,r3,r8
244 crand 0,0,4 /* cr0.lt &= cr1.lt */
245 blt memcpy /* if regions overlap */
246
247 addi r4,r4,-4
248 addi r6,r3,-4
249 neg r0,r3
250 andi. r0,r0,CACHELINE_MASK /* # bytes to start of cache line */
251 beq 58f
252
253 cmplw 0,r5,r0 /* is this more than total to do? */
254 blt 63f /* if not much to do */
255 andi. r8,r0,3 /* get it word-aligned first */
256 subf r5,r0,r5
257 mtctr r8
258 beq+ 61f
25970: lbz r9,4(r4) /* do some bytes */
260 stb r9,4(r6)
261 addi r4,r4,1
262 addi r6,r6,1
263 bdnz 70b
26461: srwi. r0,r0,2
265 mtctr r0
266 beq 58f
26772: lwzu r9,4(r4) /* do some words */
268 stwu r9,4(r6)
269 bdnz 72b
270
27158: srwi. r0,r5,LG_CACHELINE_BYTES /* # complete cachelines */
272 clrlwi r5,r5,32-LG_CACHELINE_BYTES
273 li r11,4
274 mtctr r0
275 beq 63f
27653:
277#if !defined(CONFIG_8xx)
278 dcbz r11,r6
279#endif
280 COPY_16_BYTES
281#if L1_CACHE_BYTES >= 32
282 COPY_16_BYTES
283#if L1_CACHE_BYTES >= 64
284 COPY_16_BYTES
285 COPY_16_BYTES
286#if L1_CACHE_BYTES >= 128
287 COPY_16_BYTES
288 COPY_16_BYTES
289 COPY_16_BYTES
290 COPY_16_BYTES
291#endif
292#endif
293#endif
294 bdnz 53b
295
29663: srwi. r0,r5,2
297 mtctr r0
298 beq 64f
29930: lwzu r0,4(r4)
300 stwu r0,4(r6)
301 bdnz 30b
302
30364: andi. r0,r5,3
304 mtctr r0
305 beq+ 65f
30640: lbz r0,4(r4)
307 stb r0,4(r6)
308 addi r4,r4,1
309 addi r6,r6,1
310 bdnz 40b
31165: blr
312
313_GLOBAL(memmove)
314 cmplw 0,r3,r4
315 bgt backwards_memcpy
316 /* fall through */
317
318_GLOBAL(memcpy)
319 srwi. r7,r5,3
320 addi r6,r3,-4
321 addi r4,r4,-4
322 beq 2f /* if less than 8 bytes to do */
323 andi. r0,r6,3 /* get dest word aligned */
324 mtctr r7
325 bne 5f
3261: lwz r7,4(r4)
327 lwzu r8,8(r4)
328 stw r7,4(r6)
329 stwu r8,8(r6)
330 bdnz 1b
331 andi. r5,r5,7
3322: cmplwi 0,r5,4
333 blt 3f
334 lwzu r0,4(r4)
335 addi r5,r5,-4
336 stwu r0,4(r6)
3373: cmpwi 0,r5,0
338 beqlr
339 mtctr r5
340 addi r4,r4,3
341 addi r6,r6,3
3424: lbzu r0,1(r4)
343 stbu r0,1(r6)
344 bdnz 4b
345 blr
3465: subfic r0,r0,4
347 mtctr r0
3486: lbz r7,4(r4)
349 addi r4,r4,1
350 stb r7,4(r6)
351 addi r6,r6,1
352 bdnz 6b
353 subf r5,r0,r5
354 rlwinm. r7,r5,32-3,3,31
355 beq 2b
356 mtctr r7
357 b 1b
358
359_GLOBAL(backwards_memcpy)
360 rlwinm. r7,r5,32-3,3,31 /* r0 = r5 >> 3 */
361 add r6,r3,r5
362 add r4,r4,r5
363 beq 2f
364 andi. r0,r6,3
365 mtctr r7
366 bne 5f
3671: lwz r7,-4(r4)
368 lwzu r8,-8(r4)
369 stw r7,-4(r6)
370 stwu r8,-8(r6)
371 bdnz 1b
372 andi. r5,r5,7
3732: cmplwi 0,r5,4
374 blt 3f
375 lwzu r0,-4(r4)
376 subi r5,r5,4
377 stwu r0,-4(r6)
3783: cmpwi 0,r5,0
379 beqlr
380 mtctr r5
3814: lbzu r0,-1(r4)
382 stbu r0,-1(r6)
383 bdnz 4b
384 blr
3855: mtctr r0
3866: lbzu r7,-1(r4)
387 stbu r7,-1(r6)
388 bdnz 6b
389 subf r5,r0,r5
390 rlwinm. r7,r5,32-3,3,31
391 beq 2b
392 mtctr r7
393 b 1b
394
395_GLOBAL(memcmp)
396 cmpwi 0,r5,0
397 ble- 2f
398 mtctr r5
399 addi r6,r3,-1
400 addi r4,r4,-1
4011: lbzu r3,1(r6)
402 lbzu r0,1(r4)
403 subf. r3,r0,r3
404 bdnzt 2,1b
405 blr
4062: li r3,0
407 blr
408
409_GLOBAL(memchr)
410 cmpwi 0,r5,0
411 ble- 2f
412 mtctr r5
413 addi r3,r3,-1
4141: lbzu r0,1(r3)
415 cmpw 0,r0,r4
416 bdnzf 2,1b
417 beqlr
4182: li r3,0
419 blr
420
421_GLOBAL(__copy_tofrom_user)
422 addi r4,r4,-4
423 addi r6,r3,-4
424 neg r0,r3
425 andi. r0,r0,CACHELINE_MASK /* # bytes to start of cache line */
426 beq 58f
427
428 cmplw 0,r5,r0 /* is this more than total to do? */
429 blt 63f /* if not much to do */
430 andi. r8,r0,3 /* get it word-aligned first */
431 mtctr r8
432 beq+ 61f
43370: lbz r9,4(r4) /* do some bytes */
43471: stb r9,4(r6)
435 addi r4,r4,1
436 addi r6,r6,1
437 bdnz 70b
43861: subf r5,r0,r5
439 srwi. r0,r0,2
440 mtctr r0
441 beq 58f
44272: lwzu r9,4(r4) /* do some words */
44373: stwu r9,4(r6)
444 bdnz 72b
445
446 .section __ex_table,"a"
447 .align 2
448 .long 70b,100f
449 .long 71b,101f
450 .long 72b,102f
451 .long 73b,103f
452 .text
453
45458: srwi. r0,r5,LG_CACHELINE_BYTES /* # complete cachelines */
455 clrlwi r5,r5,32-LG_CACHELINE_BYTES
456 li r11,4
457 beq 63f
458
459#ifdef CONFIG_8xx
460 /* Don't use prefetch on 8xx */
461 mtctr r0
462 li r0,0
46353: COPY_16_BYTES_WITHEX(0)
464 bdnz 53b
465
466#else /* not CONFIG_8xx */
467 /* Here we decide how far ahead to prefetch the source */
468 li r3,4
469 cmpwi r0,1
470 li r7,0
471 ble 114f
472 li r7,1
473#if MAX_COPY_PREFETCH > 1
474 /* Heuristically, for large transfers we prefetch
475 MAX_COPY_PREFETCH cachelines ahead. For small transfers
476 we prefetch 1 cacheline ahead. */
477 cmpwi r0,MAX_COPY_PREFETCH
478 ble 112f
479 li r7,MAX_COPY_PREFETCH
480112: mtctr r7
481111: dcbt r3,r4
482 addi r3,r3,CACHELINE_BYTES
483 bdnz 111b
484#else
485 dcbt r3,r4
486 addi r3,r3,CACHELINE_BYTES
487#endif /* MAX_COPY_PREFETCH > 1 */
488
489114: subf r8,r7,r0
490 mr r0,r7
491 mtctr r8
492
49353: dcbt r3,r4
49454: dcbz r11,r6
495 .section __ex_table,"a"
496 .align 2
497 .long 54b,105f
498 .text
499/* the main body of the cacheline loop */
500 COPY_16_BYTES_WITHEX(0)
501#if L1_CACHE_BYTES >= 32
502 COPY_16_BYTES_WITHEX(1)
503#if L1_CACHE_BYTES >= 64
504 COPY_16_BYTES_WITHEX(2)
505 COPY_16_BYTES_WITHEX(3)
506#if L1_CACHE_BYTES >= 128
507 COPY_16_BYTES_WITHEX(4)
508 COPY_16_BYTES_WITHEX(5)
509 COPY_16_BYTES_WITHEX(6)
510 COPY_16_BYTES_WITHEX(7)
511#endif
512#endif
513#endif
514 bdnz 53b
515 cmpwi r0,0
516 li r3,4
517 li r7,0
518 bne 114b
519#endif /* CONFIG_8xx */
520
52163: srwi. r0,r5,2
522 mtctr r0
523 beq 64f
52430: lwzu r0,4(r4)
52531: stwu r0,4(r6)
526 bdnz 30b
527
52864: andi. r0,r5,3
529 mtctr r0
530 beq+ 65f
53140: lbz r0,4(r4)
53241: stb r0,4(r6)
533 addi r4,r4,1
534 addi r6,r6,1
535 bdnz 40b
53665: li r3,0
537 blr
538
539/* read fault, initial single-byte copy */
540100: li r9,0
541 b 90f
542/* write fault, initial single-byte copy */
543101: li r9,1
54490: subf r5,r8,r5
545 li r3,0
546 b 99f
547/* read fault, initial word copy */
548102: li r9,0
549 b 91f
550/* write fault, initial word copy */
551103: li r9,1
55291: li r3,2
553 b 99f
554
555/*
556 * this stuff handles faults in the cacheline loop and branches to either
557 * 104f (if in read part) or 105f (if in write part), after updating r5
558 */
559 COPY_16_BYTES_EXCODE(0)
560#if L1_CACHE_BYTES >= 32
561 COPY_16_BYTES_EXCODE(1)
562#if L1_CACHE_BYTES >= 64
563 COPY_16_BYTES_EXCODE(2)
564 COPY_16_BYTES_EXCODE(3)
565#if L1_CACHE_BYTES >= 128
566 COPY_16_BYTES_EXCODE(4)
567 COPY_16_BYTES_EXCODE(5)
568 COPY_16_BYTES_EXCODE(6)
569 COPY_16_BYTES_EXCODE(7)
570#endif
571#endif
572#endif
573
574/* read fault in cacheline loop */
575104: li r9,0
576 b 92f
577/* fault on dcbz (effectively a write fault) */
578/* or write fault in cacheline loop */
579105: li r9,1
58092: li r3,LG_CACHELINE_BYTES
581 mfctr r8
582 add r0,r0,r8
583 b 106f
584/* read fault in final word loop */
585108: li r9,0
586 b 93f
587/* write fault in final word loop */
588109: li r9,1
58993: andi. r5,r5,3
590 li r3,2
591 b 99f
592/* read fault in final byte loop */
593110: li r9,0
594 b 94f
595/* write fault in final byte loop */
596111: li r9,1
59794: li r5,0
598 li r3,0
599/*
600 * At this stage the number of bytes not copied is
601 * r5 + (ctr << r3), and r9 is 0 for read or 1 for write.
602 */
60399: mfctr r0
604106: slw r3,r0,r3
605 add. r3,r3,r5
606 beq 120f /* shouldn't happen */
607 cmpwi 0,r9,0
608 bne 120f
609/* for a read fault, first try to continue the copy one byte at a time */
610 mtctr r3
611130: lbz r0,4(r4)
612131: stb r0,4(r6)
613 addi r4,r4,1
614 addi r6,r6,1
615 bdnz 130b
616/* then clear out the destination: r3 bytes starting at 4(r6) */
617132: mfctr r3
618 srwi. r0,r3,2
619 li r9,0
620 mtctr r0
621 beq 113f
622112: stwu r9,4(r6)
623 bdnz 112b
624113: andi. r0,r3,3
625 mtctr r0
626 beq 120f
627114: stb r9,4(r6)
628 addi r6,r6,1
629 bdnz 114b
630120: blr
631
632 .section __ex_table,"a"
633 .align 2
634 .long 30b,108b
635 .long 31b,109b
636 .long 40b,110b
637 .long 41b,111b
638 .long 130b,132b
639 .long 131b,120b
640 .long 112b,120b
641 .long 114b,120b
642 .text
643
644_GLOBAL(__clear_user)
645 addi r6,r3,-4
646 li r3,0
647 li r5,0
648 cmplwi 0,r4,4
649 blt 7f
650 /* clear a single word */
65111: stwu r5,4(r6)
652 beqlr
653 /* clear word sized chunks */
654 andi. r0,r6,3
655 add r4,r0,r4
656 subf r6,r0,r6
657 srwi r0,r4,2
658 andi. r4,r4,3
659 mtctr r0
660 bdz 7f
6611: stwu r5,4(r6)
662 bdnz 1b
663 /* clear byte sized chunks */
6647: cmpwi 0,r4,0
665 beqlr
666 mtctr r4
667 addi r6,r6,3
6688: stbu r5,1(r6)
669 bdnz 8b
670 blr
67190: mr r3,r4
672 blr
67391: mfctr r3
674 slwi r3,r3,2
675 add r3,r3,r4
676 blr
67792: mfctr r3
678 blr
679
680 .section __ex_table,"a"
681 .align 2
682 .long 11b,90b
683 .long 1b,91b
684 .long 8b,92b
685 .text
686
687_GLOBAL(__strncpy_from_user)
688 addi r6,r3,-1
689 addi r4,r4,-1
690 cmpwi 0,r5,0
691 beq 2f
692 mtctr r5
6931: lbzu r0,1(r4)
694 cmpwi 0,r0,0
695 stbu r0,1(r6)
696 bdnzf 2,1b /* dec ctr, branch if ctr != 0 && !cr0.eq */
697 beq 3f
6982: addi r6,r6,1
6993: subf r3,r3,r6
700 blr
70199: li r3,-EFAULT
702 blr
703
704 .section __ex_table,"a"
705 .align 2
706 .long 1b,99b
707 .text
708
709/* r3 = str, r4 = len (> 0), r5 = top (highest addr) */
710_GLOBAL(__strnlen_user)
711 addi r7,r3,-1
712 subf r6,r7,r5 /* top+1 - str */
713 cmplw 0,r4,r6
714 bge 0f
715 mr r6,r4
7160: mtctr r6 /* ctr = min(len, top - str) */
7171: lbzu r0,1(r7) /* get next byte */
718 cmpwi 0,r0,0
719 bdnzf 2,1b /* loop if --ctr != 0 && byte != 0 */
720 addi r7,r7,1
721 subf r3,r3,r7 /* number of bytes we have looked at */
722 beqlr /* return if we found a 0 byte */
723 cmpw 0,r3,r4 /* did we look at all len bytes? */
724 blt 99f /* if not, must have hit top */
725 addi r3,r4,1 /* return len + 1 to indicate no null found */
726 blr
72799: li r3,0 /* bad address, return 0 */
728 blr
729
730 .section __ex_table,"a"
731 .align 2
732 .long 1b,99b