]> git.ipfire.org Git - thirdparty/glibc.git/blob - sysdeps/x86_64/multiarch/strcat-ssse3.S
Revert " * sysdeps/x86_64/strlen.S: Replace with new SSE2 based implementation"
[thirdparty/glibc.git] / sysdeps / x86_64 / multiarch / strcat-ssse3.S
1 /* strcat with SSSE3
2 Copyright (C) 2011-2013 Free Software Foundation, Inc.
3 Contributed by Intel Corporation.
4 This file is part of the GNU C Library.
5
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
10
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
15
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, see
18 <http://www.gnu.org/licenses/>. */
19
20 #ifndef NOT_IN_libc
21
22 # include <sysdep.h>
23
24 # ifndef STRCAT
25 # define STRCAT __strcat_ssse3
26 # endif
27
28 # define USE_AS_STRCAT
29
30 .text
31 ENTRY (STRCAT)
32 # ifdef USE_AS_STRNCAT
33 mov %rdx, %r8
34 # endif
35
36 # define RETURN jmp L(StartStrcpyPart)
37 # include "strlen-sse2-no-bsf.S"
38
39 # undef RETURN
40
41 L(StartStrcpyPart):
42 mov %rsi, %rcx
43 lea (%rdi, %rax), %rdx
44 # ifdef USE_AS_STRNCAT
45 test %r8, %r8
46 jz L(StrncatExit0)
47 cmp $8, %r8
48 jbe L(StrncatExit8Bytes)
49 # endif
50 cmpb $0, (%rcx)
51 jz L(Exit1)
52 cmpb $0, 1(%rcx)
53 jz L(Exit2)
54 cmpb $0, 2(%rcx)
55 jz L(Exit3)
56 cmpb $0, 3(%rcx)
57 jz L(Exit4)
58 cmpb $0, 4(%rcx)
59 jz L(Exit5)
60 cmpb $0, 5(%rcx)
61 jz L(Exit6)
62 cmpb $0, 6(%rcx)
63 jz L(Exit7)
64 cmpb $0, 7(%rcx)
65 jz L(Exit8)
66 cmpb $0, 8(%rcx)
67 jz L(Exit9)
68 # ifdef USE_AS_STRNCAT
69 cmp $16, %r8
70 jb L(StrncatExit15Bytes)
71 # endif
72 cmpb $0, 9(%rcx)
73 jz L(Exit10)
74 cmpb $0, 10(%rcx)
75 jz L(Exit11)
76 cmpb $0, 11(%rcx)
77 jz L(Exit12)
78 cmpb $0, 12(%rcx)
79 jz L(Exit13)
80 cmpb $0, 13(%rcx)
81 jz L(Exit14)
82 cmpb $0, 14(%rcx)
83 jz L(Exit15)
84 cmpb $0, 15(%rcx)
85 jz L(Exit16)
86 # ifdef USE_AS_STRNCAT
87 cmp $16, %r8
88 je L(StrncatExit16)
89 # define USE_AS_STRNCPY
90 # endif
91
92 # include "strcpy-ssse3.S"
93
94 .p2align 4
95 L(CopyFrom1To16Bytes):
96 add %rsi, %rdx
97 add %rsi, %rcx
98
99 test %al, %al
100 jz L(ExitHigh)
101 test $0x01, %al
102 jnz L(Exit1)
103 test $0x02, %al
104 jnz L(Exit2)
105 test $0x04, %al
106 jnz L(Exit3)
107 test $0x08, %al
108 jnz L(Exit4)
109 test $0x10, %al
110 jnz L(Exit5)
111 test $0x20, %al
112 jnz L(Exit6)
113 test $0x40, %al
114 jnz L(Exit7)
115 movlpd (%rcx), %xmm0
116 movlpd %xmm0, (%rdx)
117 mov %rdi, %rax
118 ret
119
120 .p2align 4
121 L(ExitHigh):
122 test $0x01, %ah
123 jnz L(Exit9)
124 test $0x02, %ah
125 jnz L(Exit10)
126 test $0x04, %ah
127 jnz L(Exit11)
128 test $0x08, %ah
129 jnz L(Exit12)
130 test $0x10, %ah
131 jnz L(Exit13)
132 test $0x20, %ah
133 jnz L(Exit14)
134 test $0x40, %ah
135 jnz L(Exit15)
136 movlpd (%rcx), %xmm0
137 movlpd 8(%rcx), %xmm1
138 movlpd %xmm0, (%rdx)
139 movlpd %xmm1, 8(%rdx)
140 mov %rdi, %rax
141 ret
142
143 .p2align 4
144 L(StrncatExit1):
145 xor %ah, %ah
146 movb %ah, 1(%rdx)
147 L(Exit1):
148 movb (%rcx), %al
149 movb %al, (%rdx)
150 mov %rdi, %rax
151 ret
152
153 .p2align 4
154 L(StrncatExit2):
155 xor %ah, %ah
156 movb %ah, 2(%rdx)
157 L(Exit2):
158 movw (%rcx), %ax
159 movw %ax, (%rdx)
160 mov %rdi, %rax
161 ret
162
163 .p2align 4
164 L(StrncatExit3):
165 xor %ah, %ah
166 movb %ah, 3(%rdx)
167 L(Exit3):
168 movw (%rcx), %ax
169 movw %ax, (%rdx)
170 movb 2(%rcx), %al
171 movb %al, 2(%rdx)
172 mov %rdi, %rax
173 ret
174
175 .p2align 4
176 L(StrncatExit4):
177 xor %ah, %ah
178 movb %ah, 4(%rdx)
179 L(Exit4):
180 mov (%rcx), %eax
181 mov %eax, (%rdx)
182 mov %rdi, %rax
183 ret
184
185 .p2align 4
186 L(StrncatExit5):
187 xor %ah, %ah
188 movb %ah, 5(%rdx)
189 L(Exit5):
190 mov (%rcx), %eax
191 mov %eax, (%rdx)
192 movb 4(%rcx), %al
193 movb %al, 4(%rdx)
194 mov %rdi, %rax
195 ret
196
197 .p2align 4
198 L(StrncatExit6):
199 xor %ah, %ah
200 movb %ah, 6(%rdx)
201 L(Exit6):
202 mov (%rcx), %eax
203 mov %eax, (%rdx)
204 movw 4(%rcx), %ax
205 movw %ax, 4(%rdx)
206 mov %rdi, %rax
207 ret
208
209 .p2align 4
210 L(StrncatExit7):
211 xor %ah, %ah
212 movb %ah, 7(%rdx)
213 L(Exit7):
214 mov (%rcx), %eax
215 mov %eax, (%rdx)
216 mov 3(%rcx), %eax
217 mov %eax, 3(%rdx)
218 mov %rdi, %rax
219 ret
220
221 .p2align 4
222 L(StrncatExit8):
223 xor %ah, %ah
224 movb %ah, 8(%rdx)
225 L(Exit8):
226 movlpd (%rcx), %xmm0
227 movlpd %xmm0, (%rdx)
228 mov %rdi, %rax
229 ret
230
231 .p2align 4
232 L(StrncatExit9):
233 xor %ah, %ah
234 movb %ah, 9(%rdx)
235 L(Exit9):
236 movlpd (%rcx), %xmm0
237 movlpd %xmm0, (%rdx)
238 movb 8(%rcx), %al
239 movb %al, 8(%rdx)
240 mov %rdi, %rax
241 ret
242
243 .p2align 4
244 L(StrncatExit10):
245 xor %ah, %ah
246 movb %ah, 10(%rdx)
247 L(Exit10):
248 movlpd (%rcx), %xmm0
249 movlpd %xmm0, (%rdx)
250 movw 8(%rcx), %ax
251 movw %ax, 8(%rdx)
252 mov %rdi, %rax
253 ret
254
255 .p2align 4
256 L(StrncatExit11):
257 xor %ah, %ah
258 movb %ah, 11(%rdx)
259 L(Exit11):
260 movlpd (%rcx), %xmm0
261 movlpd %xmm0, (%rdx)
262 mov 7(%rcx), %eax
263 mov %eax, 7(%rdx)
264 mov %rdi, %rax
265 ret
266
267 .p2align 4
268 L(StrncatExit12):
269 xor %ah, %ah
270 movb %ah, 12(%rdx)
271 L(Exit12):
272 movlpd (%rcx), %xmm0
273 movlpd %xmm0, (%rdx)
274 mov 8(%rcx), %eax
275 mov %eax, 8(%rdx)
276 mov %rdi, %rax
277 ret
278
279 .p2align 4
280 L(StrncatExit13):
281 xor %ah, %ah
282 movb %ah, 13(%rdx)
283 L(Exit13):
284 movlpd (%rcx), %xmm0
285 movlpd %xmm0, (%rdx)
286 movlpd 5(%rcx), %xmm1
287 movlpd %xmm1, 5(%rdx)
288 mov %rdi, %rax
289 ret
290
291 .p2align 4
292 L(StrncatExit14):
293 xor %ah, %ah
294 movb %ah, 14(%rdx)
295 L(Exit14):
296 movlpd (%rcx), %xmm0
297 movlpd %xmm0, (%rdx)
298 movlpd 6(%rcx), %xmm1
299 movlpd %xmm1, 6(%rdx)
300 mov %rdi, %rax
301 ret
302
303 .p2align 4
304 L(StrncatExit15):
305 xor %ah, %ah
306 movb %ah, 15(%rdx)
307 L(Exit15):
308 movlpd (%rcx), %xmm0
309 movlpd %xmm0, (%rdx)
310 movlpd 7(%rcx), %xmm1
311 movlpd %xmm1, 7(%rdx)
312 mov %rdi, %rax
313 ret
314
315 .p2align 4
316 L(StrncatExit16):
317 xor %ah, %ah
318 movb %ah, 16(%rdx)
319 L(Exit16):
320 movlpd (%rcx), %xmm0
321 movlpd 8(%rcx), %xmm1
322 movlpd %xmm0, (%rdx)
323 movlpd %xmm1, 8(%rdx)
324 mov %rdi, %rax
325 ret
326
327 # ifdef USE_AS_STRNCPY
328
329 .p2align 4
330 L(CopyFrom1To16BytesCase2):
331 add $16, %r8
332 add %rsi, %rcx
333 lea (%rsi, %rdx), %rsi
334 lea -9(%r8), %rdx
335 and $1<<7, %dh
336 or %al, %dh
337 test %dh, %dh
338 lea (%rsi), %rdx
339 jz L(ExitHighCase2)
340
341 test $0x01, %al
342 jnz L(Exit1)
343 cmp $1, %r8
344 je L(StrncatExit1)
345 test $0x02, %al
346 jnz L(Exit2)
347 cmp $2, %r8
348 je L(StrncatExit2)
349 test $0x04, %al
350 jnz L(Exit3)
351 cmp $3, %r8
352 je L(StrncatExit3)
353 test $0x08, %al
354 jnz L(Exit4)
355 cmp $4, %r8
356 je L(StrncatExit4)
357 test $0x10, %al
358 jnz L(Exit5)
359 cmp $5, %r8
360 je L(StrncatExit5)
361 test $0x20, %al
362 jnz L(Exit6)
363 cmp $6, %r8
364 je L(StrncatExit6)
365 test $0x40, %al
366 jnz L(Exit7)
367 cmp $7, %r8
368 je L(StrncatExit7)
369 movlpd (%rcx), %xmm0
370 movlpd %xmm0, (%rdx)
371 lea 7(%rdx), %rax
372 cmpb $1, (%rax)
373 sbb $-1, %rax
374 xor %cl, %cl
375 movb %cl, (%rax)
376 mov %rdi, %rax
377 ret
378
379 .p2align 4
380 L(ExitHighCase2):
381 test $0x01, %ah
382 jnz L(Exit9)
383 cmp $9, %r8
384 je L(StrncatExit9)
385 test $0x02, %ah
386 jnz L(Exit10)
387 cmp $10, %r8
388 je L(StrncatExit10)
389 test $0x04, %ah
390 jnz L(Exit11)
391 cmp $11, %r8
392 je L(StrncatExit11)
393 test $0x8, %ah
394 jnz L(Exit12)
395 cmp $12, %r8
396 je L(StrncatExit12)
397 test $0x10, %ah
398 jnz L(Exit13)
399 cmp $13, %r8
400 je L(StrncatExit13)
401 test $0x20, %ah
402 jnz L(Exit14)
403 cmp $14, %r8
404 je L(StrncatExit14)
405 test $0x40, %ah
406 jnz L(Exit15)
407 cmp $15, %r8
408 je L(StrncatExit15)
409 movlpd (%rcx), %xmm0
410 movlpd %xmm0, (%rdx)
411 movlpd 8(%rcx), %xmm1
412 movlpd %xmm1, 8(%rdx)
413 mov %rdi, %rax
414 ret
415
416 L(CopyFrom1To16BytesCase2OrCase3):
417 test %rax, %rax
418 jnz L(CopyFrom1To16BytesCase2)
419
420 .p2align 4
421 L(CopyFrom1To16BytesCase3):
422 add $16, %r8
423 add %rsi, %rdx
424 add %rsi, %rcx
425
426 cmp $8, %r8
427 ja L(ExitHighCase3)
428 cmp $1, %r8
429 je L(StrncatExit1)
430 cmp $2, %r8
431 je L(StrncatExit2)
432 cmp $3, %r8
433 je L(StrncatExit3)
434 cmp $4, %r8
435 je L(StrncatExit4)
436 cmp $5, %r8
437 je L(StrncatExit5)
438 cmp $6, %r8
439 je L(StrncatExit6)
440 cmp $7, %r8
441 je L(StrncatExit7)
442 movlpd (%rcx), %xmm0
443 movlpd %xmm0, (%rdx)
444 xor %ah, %ah
445 movb %ah, 8(%rdx)
446 mov %rdi, %rax
447 ret
448
449 .p2align 4
450 L(ExitHighCase3):
451 cmp $9, %r8
452 je L(StrncatExit9)
453 cmp $10, %r8
454 je L(StrncatExit10)
455 cmp $11, %r8
456 je L(StrncatExit11)
457 cmp $12, %r8
458 je L(StrncatExit12)
459 cmp $13, %r8
460 je L(StrncatExit13)
461 cmp $14, %r8
462 je L(StrncatExit14)
463 cmp $15, %r8
464 je L(StrncatExit15)
465 movlpd (%rcx), %xmm0
466 movlpd %xmm0, (%rdx)
467 movlpd 8(%rcx), %xmm1
468 movlpd %xmm1, 8(%rdx)
469 xor %ah, %ah
470 movb %ah, 16(%rdx)
471 mov %rdi, %rax
472 ret
473
474 .p2align 4
475 L(StrncatExit0):
476 mov %rdi, %rax
477 ret
478
479 .p2align 4
480 L(StrncatExit15Bytes):
481 cmp $9, %r8
482 je L(StrncatExit9)
483 cmpb $0, 9(%rcx)
484 jz L(Exit10)
485 cmp $10, %r8
486 je L(StrncatExit10)
487 cmpb $0, 10(%rcx)
488 jz L(Exit11)
489 cmp $11, %r8
490 je L(StrncatExit11)
491 cmpb $0, 11(%rcx)
492 jz L(Exit12)
493 cmp $12, %r8
494 je L(StrncatExit12)
495 cmpb $0, 12(%rcx)
496 jz L(Exit13)
497 cmp $13, %r8
498 je L(StrncatExit13)
499 cmpb $0, 13(%rcx)
500 jz L(Exit14)
501 cmp $14, %r8
502 je L(StrncatExit14)
503 movlpd (%rcx), %xmm0
504 movlpd %xmm0, (%rdx)
505 movlpd 7(%rcx), %xmm1
506 movlpd %xmm1, 7(%rdx)
507 lea 14(%rdx), %rax
508 cmpb $1, (%rax)
509 sbb $-1, %rax
510 xor %cl, %cl
511 movb %cl, (%rax)
512 mov %rdi, %rax
513 ret
514
515 .p2align 4
516 L(StrncatExit8Bytes):
517 cmpb $0, (%rcx)
518 jz L(Exit1)
519 cmp $1, %r8
520 je L(StrncatExit1)
521 cmpb $0, 1(%rcx)
522 jz L(Exit2)
523 cmp $2, %r8
524 je L(StrncatExit2)
525 cmpb $0, 2(%rcx)
526 jz L(Exit3)
527 cmp $3, %r8
528 je L(StrncatExit3)
529 cmpb $0, 3(%rcx)
530 jz L(Exit4)
531 cmp $4, %r8
532 je L(StrncatExit4)
533 cmpb $0, 4(%rcx)
534 jz L(Exit5)
535 cmp $5, %r8
536 je L(StrncatExit5)
537 cmpb $0, 5(%rcx)
538 jz L(Exit6)
539 cmp $6, %r8
540 je L(StrncatExit6)
541 cmpb $0, 6(%rcx)
542 jz L(Exit7)
543 cmp $7, %r8
544 je L(StrncatExit7)
545 movlpd (%rcx), %xmm0
546 movlpd %xmm0, (%rdx)
547 lea 7(%rdx), %rax
548 cmpb $1, (%rax)
549 sbb $-1, %rax
550 xor %cl, %cl
551 movb %cl, (%rax)
552 mov %rdi, %rax
553 ret
554
555 # endif
556 END (STRCAT)
557 #endif