]>
Commit | Line | Data |
---|---|---|
0b1cbaae | 1 | /* strcpy with SSE2 and unaligned load |
04277e02 | 2 | Copyright (C) 2011-2019 Free Software Foundation, Inc. |
0b1cbaae L |
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 | |
59ba27a6 | 17 | License along with the GNU C Library; if not, see |
5a82c748 | 18 | <https://www.gnu.org/licenses/>. */ |
0b1cbaae L |
19 | |
20 | ||
4f41c682 | 21 | #if IS_IN (libc) |
0b1cbaae L |
22 | |
23 | # include <sysdep.h> | |
24 | ||
25 | ||
26 | # define CFI_PUSH(REG) \ | |
27 | cfi_adjust_cfa_offset (4); \ | |
28 | cfi_rel_offset (REG, 0) | |
29 | ||
30 | # define CFI_POP(REG) \ | |
31 | cfi_adjust_cfa_offset (-4); \ | |
32 | cfi_restore (REG) | |
33 | ||
34 | # define PUSH(REG) pushl REG; CFI_PUSH (REG) | |
35 | # define POP(REG) popl REG; CFI_POP (REG) | |
36 | ||
37 | # ifndef STRCPY | |
38 | # define STRCPY __strcpy_sse2 | |
39 | # endif | |
40 | ||
41 | # define STR1 PARMS | |
42 | # define STR2 STR1+4 | |
43 | # define LEN STR2+4 | |
44 | ||
45 | # ifdef USE_AS_STRNCPY | |
46 | # define PARMS 16 | |
47 | # define ENTRANCE PUSH(%ebx); PUSH(%esi); PUSH(%edi) | |
48 | # define RETURN POP(%edi); POP(%esi); POP(%ebx); ret; \ | |
acb0d739 | 49 | CFI_PUSH(%ebx); CFI_PUSH(%esi); CFI_PUSH(%edi); |
0b1cbaae | 50 | |
dfc93c41 | 51 | # ifdef PIC |
0b1cbaae L |
52 | # define JMPTBL(I, B) I - B |
53 | ||
54 | /* Load an entry in a jump table into ECX and branch to it. TABLE is a | |
55 | jump table with relative offsets. | |
56 | INDEX is a register contains the index into the jump table. | |
57 | SCALE is the scale of INDEX. */ | |
58 | ||
59 | # define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \ | |
60 | /* We first load PC into ECX. */ \ | |
9a1d9254 | 61 | SETUP_PIC_REG(cx); \ |
0b1cbaae L |
62 | /* Get the address of the jump table. */ \ |
63 | addl $(TABLE - .), %ecx; \ | |
64 | /* Get the entry and convert the relative offset to the \ | |
65 | absolute address. */ \ | |
66 | addl (%ecx,INDEX,SCALE), %ecx; \ | |
c0c3f78a | 67 | /* We loaded the jump table and adjusted ECX. Go. */ \ |
f1574581 | 68 | _CET_NOTRACK jmp *%ecx |
0b1cbaae L |
69 | # else |
70 | # define JMPTBL(I, B) I | |
71 | ||
72 | /* Branch to an entry in a jump table. TABLE is a jump table with | |
73 | absolute offsets. INDEX is a register contains the index into the | |
74 | jump table. SCALE is the scale of INDEX. */ | |
75 | ||
76 | # define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \ | |
f1574581 | 77 | _CET_NOTRACK jmp *TABLE(,INDEX,SCALE) |
0b1cbaae L |
78 | # endif |
79 | ||
80 | .text | |
81 | ENTRY (STRCPY) | |
82 | ENTRANCE | |
83 | mov STR1(%esp), %edi | |
84 | mov STR2(%esp), %esi | |
85 | movl LEN(%esp), %ebx | |
86 | test %ebx, %ebx | |
87 | jz L(ExitZero) | |
88 | ||
89 | mov %esi, %ecx | |
90 | # ifndef USE_AS_STPCPY | |
91 | mov %edi, %eax /* save result */ | |
92 | # endif | |
93 | and $15, %ecx | |
94 | jz L(SourceStringAlignmentZero) | |
95 | ||
96 | and $-16, %esi | |
97 | pxor %xmm0, %xmm0 | |
98 | pxor %xmm1, %xmm1 | |
99 | ||
100 | pcmpeqb (%esi), %xmm1 | |
101 | add %ecx, %ebx | |
102 | pmovmskb %xmm1, %edx | |
103 | shr %cl, %edx | |
104 | # ifdef USE_AS_STPCPY | |
105 | cmp $16, %ebx | |
106 | jbe L(CopyFrom1To16BytesTailCase2OrCase3) | |
107 | # else | |
108 | cmp $17, %ebx | |
109 | jbe L(CopyFrom1To16BytesTailCase2OrCase3) | |
110 | # endif | |
111 | test %edx, %edx | |
112 | jnz L(CopyFrom1To16BytesTail) | |
113 | ||
114 | pcmpeqb 16(%esi), %xmm0 | |
115 | pmovmskb %xmm0, %edx | |
116 | # ifdef USE_AS_STPCPY | |
117 | cmp $32, %ebx | |
118 | jbe L(CopyFrom1To32BytesCase2OrCase3) | |
119 | # else | |
120 | cmp $33, %ebx | |
121 | jbe L(CopyFrom1To32BytesCase2OrCase3) | |
122 | # endif | |
123 | test %edx, %edx | |
124 | jnz L(CopyFrom1To32Bytes) | |
125 | ||
126 | movdqu (%esi, %ecx), %xmm1 /* copy 16 bytes */ | |
127 | movdqu %xmm1, (%edi) | |
128 | ||
129 | sub %ecx, %edi | |
130 | ||
c0c3f78a | 131 | /* If source address alignment != destination address alignment */ |
0b1cbaae L |
132 | .p2align 4 |
133 | L(Unalign16Both): | |
134 | mov $16, %ecx | |
135 | movdqa (%esi, %ecx), %xmm1 | |
136 | movaps 16(%esi, %ecx), %xmm2 | |
137 | movdqu %xmm1, (%edi, %ecx) | |
138 | pcmpeqb %xmm2, %xmm0 | |
139 | pmovmskb %xmm0, %edx | |
140 | add $16, %ecx | |
141 | sub $48, %ebx | |
142 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
143 | test %edx, %edx | |
144 | jnz L(CopyFrom1To16BytesUnalignedXmm2) | |
145 | ||
146 | movaps 16(%esi, %ecx), %xmm3 | |
147 | movdqu %xmm2, (%edi, %ecx) | |
148 | pcmpeqb %xmm3, %xmm0 | |
149 | pmovmskb %xmm0, %edx | |
150 | add $16, %ecx | |
151 | sub $16, %ebx | |
152 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
153 | test %edx, %edx | |
154 | jnz L(CopyFrom1To16BytesUnalignedXmm3) | |
155 | ||
156 | movaps 16(%esi, %ecx), %xmm4 | |
157 | movdqu %xmm3, (%edi, %ecx) | |
158 | pcmpeqb %xmm4, %xmm0 | |
159 | pmovmskb %xmm0, %edx | |
160 | add $16, %ecx | |
161 | sub $16, %ebx | |
162 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
163 | test %edx, %edx | |
164 | jnz L(CopyFrom1To16BytesUnalignedXmm4) | |
165 | ||
166 | movaps 16(%esi, %ecx), %xmm1 | |
167 | movdqu %xmm4, (%edi, %ecx) | |
168 | pcmpeqb %xmm1, %xmm0 | |
169 | pmovmskb %xmm0, %edx | |
170 | add $16, %ecx | |
171 | sub $16, %ebx | |
172 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
173 | test %edx, %edx | |
174 | jnz L(CopyFrom1To16BytesUnalignedXmm1) | |
175 | ||
176 | movaps 16(%esi, %ecx), %xmm2 | |
177 | movdqu %xmm1, (%edi, %ecx) | |
178 | pcmpeqb %xmm2, %xmm0 | |
179 | pmovmskb %xmm0, %edx | |
180 | add $16, %ecx | |
181 | sub $16, %ebx | |
182 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
183 | test %edx, %edx | |
184 | jnz L(CopyFrom1To16BytesUnalignedXmm2) | |
185 | ||
186 | movaps 16(%esi, %ecx), %xmm3 | |
187 | movdqu %xmm2, (%edi, %ecx) | |
188 | pcmpeqb %xmm3, %xmm0 | |
189 | pmovmskb %xmm0, %edx | |
190 | add $16, %ecx | |
191 | sub $16, %ebx | |
192 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
193 | test %edx, %edx | |
194 | jnz L(CopyFrom1To16BytesUnalignedXmm3) | |
195 | ||
196 | movdqu %xmm3, (%edi, %ecx) | |
197 | mov %esi, %edx | |
198 | lea 16(%esi, %ecx), %esi | |
199 | and $-0x40, %esi | |
200 | sub %esi, %edx | |
201 | sub %edx, %edi | |
202 | lea 128(%ebx, %edx), %ebx | |
203 | ||
204 | L(Unaligned64Loop): | |
205 | movaps (%esi), %xmm2 | |
206 | movaps %xmm2, %xmm4 | |
207 | movaps 16(%esi), %xmm5 | |
208 | movaps 32(%esi), %xmm3 | |
209 | movaps %xmm3, %xmm6 | |
210 | movaps 48(%esi), %xmm7 | |
211 | pminub %xmm5, %xmm2 | |
212 | pminub %xmm7, %xmm3 | |
213 | pminub %xmm2, %xmm3 | |
214 | pcmpeqb %xmm0, %xmm3 | |
215 | pmovmskb %xmm3, %edx | |
216 | sub $64, %ebx | |
217 | jbe L(UnalignedLeaveCase2OrCase3) | |
218 | test %edx, %edx | |
219 | jnz L(Unaligned64Leave) | |
220 | L(Unaligned64Loop_start): | |
221 | add $64, %edi | |
222 | add $64, %esi | |
223 | movdqu %xmm4, -64(%edi) | |
224 | movaps (%esi), %xmm2 | |
225 | movdqa %xmm2, %xmm4 | |
226 | movdqu %xmm5, -48(%edi) | |
227 | movaps 16(%esi), %xmm5 | |
228 | pminub %xmm5, %xmm2 | |
229 | movaps 32(%esi), %xmm3 | |
230 | movdqu %xmm6, -32(%edi) | |
231 | movaps %xmm3, %xmm6 | |
232 | movdqu %xmm7, -16(%edi) | |
233 | movaps 48(%esi), %xmm7 | |
234 | pminub %xmm7, %xmm3 | |
235 | pminub %xmm2, %xmm3 | |
236 | pcmpeqb %xmm0, %xmm3 | |
237 | pmovmskb %xmm3, %edx | |
238 | sub $64, %ebx | |
239 | jbe L(UnalignedLeaveCase2OrCase3) | |
240 | test %edx, %edx | |
241 | jz L(Unaligned64Loop_start) | |
242 | L(Unaligned64Leave): | |
243 | pxor %xmm1, %xmm1 | |
244 | ||
245 | pcmpeqb %xmm4, %xmm0 | |
246 | pcmpeqb %xmm5, %xmm1 | |
247 | pmovmskb %xmm0, %edx | |
248 | pmovmskb %xmm1, %ecx | |
249 | test %edx, %edx | |
250 | jnz L(CopyFrom1To16BytesUnaligned_0) | |
251 | test %ecx, %ecx | |
252 | jnz L(CopyFrom1To16BytesUnaligned_16) | |
253 | ||
254 | pcmpeqb %xmm6, %xmm0 | |
255 | pcmpeqb %xmm7, %xmm1 | |
256 | pmovmskb %xmm0, %edx | |
257 | pmovmskb %xmm1, %ecx | |
258 | test %edx, %edx | |
259 | jnz L(CopyFrom1To16BytesUnaligned_32) | |
260 | ||
261 | bsf %ecx, %edx | |
262 | movdqu %xmm4, (%edi) | |
263 | movdqu %xmm5, 16(%edi) | |
264 | movdqu %xmm6, 32(%edi) | |
265 | # ifdef USE_AS_STPCPY | |
266 | lea 48(%edi, %edx), %eax | |
267 | # endif | |
268 | movdqu %xmm7, 48(%edi) | |
269 | add $15, %ebx | |
270 | sub %edx, %ebx | |
271 | lea 49(%edi, %edx), %edi | |
272 | jmp L(StrncpyFillTailWithZero) | |
273 | ||
c0c3f78a | 274 | /* If source address alignment == destination address alignment */ |
0b1cbaae L |
275 | |
276 | L(SourceStringAlignmentZero): | |
277 | pxor %xmm0, %xmm0 | |
278 | movdqa (%esi), %xmm1 | |
279 | pcmpeqb %xmm1, %xmm0 | |
280 | pmovmskb %xmm0, %edx | |
281 | # ifdef USE_AS_STPCPY | |
282 | cmp $16, %ebx | |
283 | jbe L(CopyFrom1To16BytesTail1Case2OrCase3) | |
284 | # else | |
285 | cmp $17, %ebx | |
286 | jbe L(CopyFrom1To16BytesTail1Case2OrCase3) | |
287 | # endif | |
288 | test %edx, %edx | |
289 | jnz L(CopyFrom1To16BytesTail1) | |
290 | ||
291 | pcmpeqb 16(%esi), %xmm0 | |
292 | movdqu %xmm1, (%edi) | |
293 | pmovmskb %xmm0, %edx | |
294 | # ifdef USE_AS_STPCPY | |
295 | cmp $32, %ebx | |
296 | jbe L(CopyFrom1To32Bytes1Case2OrCase3) | |
297 | # else | |
298 | cmp $33, %ebx | |
299 | jbe L(CopyFrom1To32Bytes1Case2OrCase3) | |
300 | # endif | |
301 | test %edx, %edx | |
302 | jnz L(CopyFrom1To32Bytes1) | |
303 | ||
304 | jmp L(Unalign16Both) | |
305 | ||
306 | /*-----------------End of main part---------------------------*/ | |
307 | ||
308 | /* Case1 */ | |
309 | .p2align 4 | |
310 | L(CopyFrom1To16BytesTail): | |
311 | sub %ecx, %ebx | |
312 | add %ecx, %esi | |
313 | bsf %edx, %edx | |
314 | BRANCH_TO_JMPTBL_ENTRY (L(ExitTable), %edx, 4) | |
315 | ||
316 | .p2align 4 | |
317 | L(CopyFrom1To32Bytes1): | |
318 | add $16, %esi | |
319 | add $16, %edi | |
320 | sub $16, %ebx | |
321 | L(CopyFrom1To16BytesTail1): | |
322 | bsf %edx, %edx | |
323 | BRANCH_TO_JMPTBL_ENTRY (L(ExitTable), %edx, 4) | |
324 | ||
325 | .p2align 4 | |
326 | L(CopyFrom1To32Bytes): | |
327 | sub %ecx, %ebx | |
328 | bsf %edx, %edx | |
329 | add %ecx, %esi | |
330 | add $16, %edx | |
331 | sub %ecx, %edx | |
332 | BRANCH_TO_JMPTBL_ENTRY (L(ExitTable), %edx, 4) | |
333 | ||
334 | .p2align 4 | |
335 | L(CopyFrom1To16BytesUnaligned_0): | |
336 | bsf %edx, %edx | |
337 | # ifdef USE_AS_STPCPY | |
338 | lea (%edi, %edx), %eax | |
339 | # endif | |
340 | movdqu %xmm4, (%edi) | |
341 | add $63, %ebx | |
342 | sub %edx, %ebx | |
343 | lea 1(%edi, %edx), %edi | |
344 | jmp L(StrncpyFillTailWithZero) | |
345 | ||
346 | .p2align 4 | |
347 | L(CopyFrom1To16BytesUnaligned_16): | |
348 | bsf %ecx, %edx | |
349 | movdqu %xmm4, (%edi) | |
350 | # ifdef USE_AS_STPCPY | |
351 | lea 16(%edi, %edx), %eax | |
352 | # endif | |
353 | movdqu %xmm5, 16(%edi) | |
354 | add $47, %ebx | |
355 | sub %edx, %ebx | |
356 | lea 17(%edi, %edx), %edi | |
357 | jmp L(StrncpyFillTailWithZero) | |
358 | ||
359 | .p2align 4 | |
360 | L(CopyFrom1To16BytesUnaligned_32): | |
361 | bsf %edx, %edx | |
362 | movdqu %xmm4, (%edi) | |
363 | movdqu %xmm5, 16(%edi) | |
364 | # ifdef USE_AS_STPCPY | |
365 | lea 32(%edi, %edx), %eax | |
366 | # endif | |
367 | movdqu %xmm6, 32(%edi) | |
368 | add $31, %ebx | |
369 | sub %edx, %ebx | |
370 | lea 33(%edi, %edx), %edi | |
371 | jmp L(StrncpyFillTailWithZero) | |
372 | ||
373 | .p2align 4 | |
374 | L(CopyFrom1To16BytesUnalignedXmm6): | |
375 | movdqu %xmm6, (%edi, %ecx) | |
376 | jmp L(CopyFrom1To16BytesXmmExit) | |
377 | ||
378 | .p2align 4 | |
379 | L(CopyFrom1To16BytesUnalignedXmm5): | |
380 | movdqu %xmm5, (%edi, %ecx) | |
381 | jmp L(CopyFrom1To16BytesXmmExit) | |
382 | ||
383 | .p2align 4 | |
384 | L(CopyFrom1To16BytesUnalignedXmm4): | |
385 | movdqu %xmm4, (%edi, %ecx) | |
386 | jmp L(CopyFrom1To16BytesXmmExit) | |
387 | ||
388 | .p2align 4 | |
389 | L(CopyFrom1To16BytesUnalignedXmm3): | |
390 | movdqu %xmm3, (%edi, %ecx) | |
391 | jmp L(CopyFrom1To16BytesXmmExit) | |
392 | ||
393 | .p2align 4 | |
394 | L(CopyFrom1To16BytesUnalignedXmm1): | |
395 | movdqu %xmm1, (%edi, %ecx) | |
396 | jmp L(CopyFrom1To16BytesXmmExit) | |
397 | ||
398 | .p2align 4 | |
399 | L(CopyFrom1To16BytesExit): | |
400 | BRANCH_TO_JMPTBL_ENTRY (L(ExitTable), %edx, 4) | |
401 | ||
402 | /* Case2 */ | |
403 | ||
404 | .p2align 4 | |
405 | L(CopyFrom1To16BytesCase2): | |
406 | add $16, %ebx | |
407 | add %ecx, %edi | |
408 | add %ecx, %esi | |
409 | bsf %edx, %edx | |
410 | cmp %ebx, %edx | |
411 | jb L(CopyFrom1To16BytesExit) | |
412 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
413 | ||
414 | .p2align 4 | |
415 | L(CopyFrom1To32BytesCase2): | |
416 | sub %ecx, %ebx | |
417 | add %ecx, %esi | |
418 | bsf %edx, %edx | |
419 | add $16, %edx | |
420 | sub %ecx, %edx | |
421 | cmp %ebx, %edx | |
422 | jb L(CopyFrom1To16BytesExit) | |
423 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
424 | ||
425 | L(CopyFrom1To16BytesTailCase2): | |
426 | sub %ecx, %ebx | |
427 | add %ecx, %esi | |
428 | bsf %edx, %edx | |
429 | cmp %ebx, %edx | |
430 | jb L(CopyFrom1To16BytesExit) | |
431 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
432 | ||
433 | L(CopyFrom1To16BytesTail1Case2): | |
434 | bsf %edx, %edx | |
435 | cmp %ebx, %edx | |
436 | jb L(CopyFrom1To16BytesExit) | |
437 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
438 | ||
439 | /* Case2 or Case3, Case3 */ | |
440 | ||
441 | .p2align 4 | |
442 | L(CopyFrom1To16BytesCase2OrCase3): | |
443 | test %edx, %edx | |
444 | jnz L(CopyFrom1To16BytesCase2) | |
445 | L(CopyFrom1To16BytesCase3): | |
446 | add $16, %ebx | |
447 | add %ecx, %edi | |
448 | add %ecx, %esi | |
449 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
450 | ||
451 | .p2align 4 | |
452 | L(CopyFrom1To32BytesCase2OrCase3): | |
453 | test %edx, %edx | |
454 | jnz L(CopyFrom1To32BytesCase2) | |
455 | sub %ecx, %ebx | |
456 | add %ecx, %esi | |
457 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
458 | ||
459 | .p2align 4 | |
460 | L(CopyFrom1To16BytesTailCase2OrCase3): | |
461 | test %edx, %edx | |
462 | jnz L(CopyFrom1To16BytesTailCase2) | |
463 | sub %ecx, %ebx | |
464 | add %ecx, %esi | |
465 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
466 | ||
467 | .p2align 4 | |
468 | L(CopyFrom1To32Bytes1Case2OrCase3): | |
469 | add $16, %edi | |
470 | add $16, %esi | |
471 | sub $16, %ebx | |
472 | L(CopyFrom1To16BytesTail1Case2OrCase3): | |
473 | test %edx, %edx | |
474 | jnz L(CopyFrom1To16BytesTail1Case2) | |
475 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
476 | ||
477 | .p2align 4 | |
478 | L(Exit0): | |
479 | # ifdef USE_AS_STPCPY | |
480 | mov %edi, %eax | |
481 | # endif | |
482 | RETURN | |
483 | ||
484 | .p2align 4 | |
485 | L(Exit1): | |
486 | movb %dh, (%edi) | |
487 | # ifdef USE_AS_STPCPY | |
488 | lea (%edi), %eax | |
489 | # endif | |
490 | sub $1, %ebx | |
491 | lea 1(%edi), %edi | |
492 | jnz L(StrncpyFillTailWithZero) | |
493 | RETURN | |
494 | ||
495 | .p2align 4 | |
496 | L(Exit2): | |
497 | movw (%esi), %dx | |
498 | movw %dx, (%edi) | |
499 | # ifdef USE_AS_STPCPY | |
500 | lea 1(%edi), %eax | |
501 | # endif | |
502 | sub $2, %ebx | |
503 | lea 2(%edi), %edi | |
504 | jnz L(StrncpyFillTailWithZero) | |
505 | RETURN | |
506 | ||
507 | .p2align 4 | |
508 | L(Exit3): | |
509 | movw (%esi), %cx | |
510 | movw %cx, (%edi) | |
511 | movb %dh, 2(%edi) | |
512 | # ifdef USE_AS_STPCPY | |
513 | lea 2(%edi), %eax | |
514 | # endif | |
515 | sub $3, %ebx | |
516 | lea 3(%edi), %edi | |
517 | jnz L(StrncpyFillTailWithZero) | |
518 | RETURN | |
519 | ||
520 | .p2align 4 | |
521 | L(Exit4): | |
522 | movl (%esi), %edx | |
523 | movl %edx, (%edi) | |
524 | # ifdef USE_AS_STPCPY | |
525 | lea 3(%edi), %eax | |
526 | # endif | |
527 | sub $4, %ebx | |
528 | lea 4(%edi), %edi | |
529 | jnz L(StrncpyFillTailWithZero) | |
530 | RETURN | |
531 | ||
532 | .p2align 4 | |
533 | L(Exit5): | |
534 | movl (%esi), %ecx | |
535 | movb %dh, 4(%edi) | |
536 | movl %ecx, (%edi) | |
537 | # ifdef USE_AS_STPCPY | |
538 | lea 4(%edi), %eax | |
539 | # endif | |
540 | sub $5, %ebx | |
541 | lea 5(%edi), %edi | |
542 | jnz L(StrncpyFillTailWithZero) | |
543 | RETURN | |
544 | ||
545 | .p2align 4 | |
546 | L(Exit6): | |
547 | movl (%esi), %ecx | |
548 | movw 4(%esi), %dx | |
549 | movl %ecx, (%edi) | |
550 | movw %dx, 4(%edi) | |
551 | # ifdef USE_AS_STPCPY | |
552 | lea 5(%edi), %eax | |
553 | # endif | |
554 | sub $6, %ebx | |
555 | lea 6(%edi), %edi | |
556 | jnz L(StrncpyFillTailWithZero) | |
557 | RETURN | |
558 | ||
559 | .p2align 4 | |
560 | L(Exit7): | |
561 | movl (%esi), %ecx | |
562 | movl 3(%esi), %edx | |
563 | movl %ecx, (%edi) | |
564 | movl %edx, 3(%edi) | |
565 | # ifdef USE_AS_STPCPY | |
566 | lea 6(%edi), %eax | |
567 | # endif | |
568 | sub $7, %ebx | |
569 | lea 7(%edi), %edi | |
570 | jnz L(StrncpyFillTailWithZero) | |
571 | RETURN | |
572 | ||
573 | .p2align 4 | |
574 | L(Exit8): | |
575 | movlpd (%esi), %xmm0 | |
576 | movlpd %xmm0, (%edi) | |
577 | # ifdef USE_AS_STPCPY | |
578 | lea 7(%edi), %eax | |
579 | # endif | |
580 | sub $8, %ebx | |
581 | lea 8(%edi), %edi | |
582 | jnz L(StrncpyFillTailWithZero) | |
583 | RETURN | |
584 | ||
585 | .p2align 4 | |
586 | L(Exit9): | |
587 | movlpd (%esi), %xmm0 | |
588 | movb %dh, 8(%edi) | |
589 | movlpd %xmm0, (%edi) | |
590 | # ifdef USE_AS_STPCPY | |
591 | lea 8(%edi), %eax | |
592 | # endif | |
593 | sub $9, %ebx | |
594 | lea 9(%edi), %edi | |
595 | jnz L(StrncpyFillTailWithZero) | |
596 | RETURN | |
597 | ||
598 | .p2align 4 | |
599 | L(Exit10): | |
600 | movlpd (%esi), %xmm0 | |
601 | movw 8(%esi), %dx | |
602 | movlpd %xmm0, (%edi) | |
603 | movw %dx, 8(%edi) | |
604 | # ifdef USE_AS_STPCPY | |
605 | lea 9(%edi), %eax | |
606 | # endif | |
607 | sub $10, %ebx | |
608 | lea 10(%edi), %edi | |
609 | jnz L(StrncpyFillTailWithZero) | |
610 | RETURN | |
611 | ||
612 | .p2align 4 | |
613 | L(Exit11): | |
614 | movlpd (%esi), %xmm0 | |
615 | movl 7(%esi), %edx | |
616 | movlpd %xmm0, (%edi) | |
617 | movl %edx, 7(%edi) | |
618 | # ifdef USE_AS_STPCPY | |
619 | lea 10(%edi), %eax | |
620 | # endif | |
621 | sub $11, %ebx | |
622 | lea 11(%edi), %edi | |
623 | jnz L(StrncpyFillTailWithZero) | |
624 | RETURN | |
625 | ||
626 | .p2align 4 | |
627 | L(Exit12): | |
628 | movlpd (%esi), %xmm0 | |
629 | movl 8(%esi), %edx | |
630 | movlpd %xmm0, (%edi) | |
631 | movl %edx, 8(%edi) | |
632 | # ifdef USE_AS_STPCPY | |
633 | lea 11(%edi), %eax | |
634 | # endif | |
635 | sub $12, %ebx | |
636 | lea 12(%edi), %edi | |
637 | jnz L(StrncpyFillTailWithZero) | |
638 | RETURN | |
639 | ||
640 | .p2align 4 | |
641 | L(Exit13): | |
642 | movlpd (%esi), %xmm0 | |
643 | movlpd 5(%esi), %xmm1 | |
644 | movlpd %xmm0, (%edi) | |
645 | movlpd %xmm1, 5(%edi) | |
646 | # ifdef USE_AS_STPCPY | |
647 | lea 12(%edi), %eax | |
648 | # endif | |
649 | sub $13, %ebx | |
650 | lea 13(%edi), %edi | |
651 | jnz L(StrncpyFillTailWithZero) | |
652 | RETURN | |
653 | ||
654 | .p2align 4 | |
655 | L(Exit14): | |
656 | movlpd (%esi), %xmm0 | |
657 | movlpd 6(%esi), %xmm1 | |
658 | movlpd %xmm0, (%edi) | |
659 | movlpd %xmm1, 6(%edi) | |
660 | # ifdef USE_AS_STPCPY | |
661 | lea 13(%edi), %eax | |
662 | # endif | |
663 | sub $14, %ebx | |
664 | lea 14(%edi), %edi | |
665 | jnz L(StrncpyFillTailWithZero) | |
666 | RETURN | |
667 | ||
668 | .p2align 4 | |
669 | L(Exit15): | |
670 | movlpd (%esi), %xmm0 | |
671 | movlpd 7(%esi), %xmm1 | |
672 | movlpd %xmm0, (%edi) | |
673 | movlpd %xmm1, 7(%edi) | |
674 | # ifdef USE_AS_STPCPY | |
675 | lea 14(%edi), %eax | |
676 | # endif | |
677 | sub $15, %ebx | |
678 | lea 15(%edi), %edi | |
679 | jnz L(StrncpyFillTailWithZero) | |
680 | RETURN | |
681 | ||
682 | .p2align 4 | |
683 | L(Exit16): | |
684 | movdqu (%esi), %xmm0 | |
685 | movdqu %xmm0, (%edi) | |
686 | # ifdef USE_AS_STPCPY | |
687 | lea 15(%edi), %eax | |
688 | # endif | |
689 | sub $16, %ebx | |
690 | lea 16(%edi), %edi | |
691 | jnz L(StrncpyFillTailWithZero) | |
692 | RETURN | |
693 | ||
694 | .p2align 4 | |
695 | L(Exit17): | |
696 | movdqu (%esi), %xmm0 | |
697 | movdqu %xmm0, (%edi) | |
698 | movb %dh, 16(%edi) | |
699 | # ifdef USE_AS_STPCPY | |
700 | lea 16(%edi), %eax | |
701 | # endif | |
702 | sub $17, %ebx | |
703 | lea 17(%edi), %edi | |
704 | jnz L(StrncpyFillTailWithZero) | |
705 | RETURN | |
706 | ||
707 | .p2align 4 | |
708 | L(Exit18): | |
709 | movdqu (%esi), %xmm0 | |
710 | movw 16(%esi), %cx | |
711 | movdqu %xmm0, (%edi) | |
712 | movw %cx, 16(%edi) | |
713 | # ifdef USE_AS_STPCPY | |
714 | lea 17(%edi), %eax | |
715 | # endif | |
716 | sub $18, %ebx | |
717 | lea 18(%edi), %edi | |
718 | jnz L(StrncpyFillTailWithZero) | |
719 | RETURN | |
720 | ||
721 | .p2align 4 | |
722 | L(Exit19): | |
723 | movdqu (%esi), %xmm0 | |
724 | movl 15(%esi), %ecx | |
725 | movdqu %xmm0, (%edi) | |
726 | movl %ecx, 15(%edi) | |
727 | # ifdef USE_AS_STPCPY | |
728 | lea 18(%edi), %eax | |
729 | # endif | |
730 | sub $19, %ebx | |
731 | lea 19(%edi), %edi | |
732 | jnz L(StrncpyFillTailWithZero) | |
733 | RETURN | |
734 | ||
735 | .p2align 4 | |
736 | L(Exit20): | |
737 | movdqu (%esi), %xmm0 | |
738 | movl 16(%esi), %ecx | |
739 | movdqu %xmm0, (%edi) | |
740 | movl %ecx, 16(%edi) | |
741 | # ifdef USE_AS_STPCPY | |
742 | lea 19(%edi), %eax | |
743 | # endif | |
744 | sub $20, %ebx | |
745 | lea 20(%edi), %edi | |
746 | jnz L(StrncpyFillTailWithZero) | |
747 | RETURN | |
748 | ||
749 | .p2align 4 | |
750 | L(Exit21): | |
751 | movdqu (%esi), %xmm0 | |
752 | movl 16(%esi), %ecx | |
753 | movdqu %xmm0, (%edi) | |
754 | movl %ecx, 16(%edi) | |
755 | movb %dh, 20(%edi) | |
756 | # ifdef USE_AS_STPCPY | |
757 | lea 20(%edi), %eax | |
758 | # endif | |
759 | sub $21, %ebx | |
760 | lea 21(%edi), %edi | |
761 | jnz L(StrncpyFillTailWithZero) | |
762 | RETURN | |
763 | ||
764 | .p2align 4 | |
765 | L(Exit22): | |
766 | movdqu (%esi), %xmm0 | |
767 | movlpd 14(%esi), %xmm3 | |
768 | movdqu %xmm0, (%edi) | |
769 | movlpd %xmm3, 14(%edi) | |
770 | # ifdef USE_AS_STPCPY | |
771 | lea 21(%edi), %eax | |
772 | # endif | |
773 | sub $22, %ebx | |
774 | lea 22(%edi), %edi | |
775 | jnz L(StrncpyFillTailWithZero) | |
776 | RETURN | |
777 | ||
778 | .p2align 4 | |
779 | L(Exit23): | |
780 | movdqu (%esi), %xmm0 | |
781 | movlpd 15(%esi), %xmm3 | |
782 | movdqu %xmm0, (%edi) | |
783 | movlpd %xmm3, 15(%edi) | |
784 | # ifdef USE_AS_STPCPY | |
785 | lea 22(%edi), %eax | |
786 | # endif | |
787 | sub $23, %ebx | |
788 | lea 23(%edi), %edi | |
789 | jnz L(StrncpyFillTailWithZero) | |
790 | RETURN | |
791 | ||
792 | .p2align 4 | |
793 | L(Exit24): | |
794 | movdqu (%esi), %xmm0 | |
795 | movlpd 16(%esi), %xmm2 | |
796 | movdqu %xmm0, (%edi) | |
797 | movlpd %xmm2, 16(%edi) | |
798 | # ifdef USE_AS_STPCPY | |
799 | lea 23(%edi), %eax | |
800 | # endif | |
801 | sub $24, %ebx | |
802 | lea 24(%edi), %edi | |
803 | jnz L(StrncpyFillTailWithZero) | |
804 | RETURN | |
805 | ||
806 | .p2align 4 | |
807 | L(Exit25): | |
808 | movdqu (%esi), %xmm0 | |
809 | movlpd 16(%esi), %xmm2 | |
810 | movdqu %xmm0, (%edi) | |
811 | movlpd %xmm2, 16(%edi) | |
812 | movb %dh, 24(%edi) | |
813 | # ifdef USE_AS_STPCPY | |
814 | lea 24(%edi), %eax | |
815 | # endif | |
816 | sub $25, %ebx | |
817 | lea 25(%edi), %edi | |
818 | jnz L(StrncpyFillTailWithZero) | |
819 | RETURN | |
820 | ||
821 | .p2align 4 | |
822 | L(Exit26): | |
823 | movdqu (%esi), %xmm0 | |
824 | movlpd 16(%esi), %xmm2 | |
825 | movw 24(%esi), %cx | |
826 | movdqu %xmm0, (%edi) | |
827 | movlpd %xmm2, 16(%edi) | |
828 | movw %cx, 24(%edi) | |
829 | # ifdef USE_AS_STPCPY | |
830 | lea 25(%edi), %eax | |
831 | # endif | |
832 | sub $26, %ebx | |
833 | lea 26(%edi), %edi | |
834 | jnz L(StrncpyFillTailWithZero) | |
835 | RETURN | |
836 | ||
837 | .p2align 4 | |
838 | L(Exit27): | |
839 | movdqu (%esi), %xmm0 | |
840 | movlpd 16(%esi), %xmm2 | |
841 | movl 23(%esi), %ecx | |
842 | movdqu %xmm0, (%edi) | |
843 | movlpd %xmm2, 16(%edi) | |
844 | movl %ecx, 23(%edi) | |
845 | # ifdef USE_AS_STPCPY | |
846 | lea 26(%edi), %eax | |
847 | # endif | |
848 | sub $27, %ebx | |
849 | lea 27(%edi), %edi | |
850 | jnz L(StrncpyFillTailWithZero) | |
851 | RETURN | |
852 | ||
853 | .p2align 4 | |
854 | L(Exit28): | |
855 | movdqu (%esi), %xmm0 | |
856 | movlpd 16(%esi), %xmm2 | |
857 | movl 24(%esi), %ecx | |
858 | movdqu %xmm0, (%edi) | |
859 | movlpd %xmm2, 16(%edi) | |
860 | movl %ecx, 24(%edi) | |
861 | # ifdef USE_AS_STPCPY | |
862 | lea 27(%edi), %eax | |
863 | # endif | |
864 | sub $28, %ebx | |
865 | lea 28(%edi), %edi | |
866 | jnz L(StrncpyFillTailWithZero) | |
867 | RETURN | |
868 | ||
869 | .p2align 4 | |
870 | L(Exit29): | |
871 | movdqu (%esi), %xmm0 | |
872 | movdqu 13(%esi), %xmm2 | |
873 | movdqu %xmm0, (%edi) | |
874 | movdqu %xmm2, 13(%edi) | |
875 | # ifdef USE_AS_STPCPY | |
876 | lea 28(%edi), %eax | |
877 | # endif | |
878 | sub $29, %ebx | |
879 | lea 29(%edi), %edi | |
880 | jnz L(StrncpyFillTailWithZero) | |
881 | RETURN | |
882 | ||
883 | .p2align 4 | |
884 | L(Exit30): | |
885 | movdqu (%esi), %xmm0 | |
886 | movdqu 14(%esi), %xmm2 | |
887 | movdqu %xmm0, (%edi) | |
888 | movdqu %xmm2, 14(%edi) | |
889 | # ifdef USE_AS_STPCPY | |
890 | lea 29(%edi), %eax | |
891 | # endif | |
892 | sub $30, %ebx | |
893 | lea 30(%edi), %edi | |
894 | jnz L(StrncpyFillTailWithZero) | |
895 | RETURN | |
896 | ||
897 | ||
898 | .p2align 4 | |
899 | L(Exit31): | |
900 | movdqu (%esi), %xmm0 | |
901 | movdqu 15(%esi), %xmm2 | |
902 | movdqu %xmm0, (%edi) | |
903 | movdqu %xmm2, 15(%edi) | |
904 | # ifdef USE_AS_STPCPY | |
905 | lea 30(%edi), %eax | |
906 | # endif | |
907 | sub $31, %ebx | |
908 | lea 31(%edi), %edi | |
909 | jnz L(StrncpyFillTailWithZero) | |
910 | RETURN | |
911 | ||
912 | .p2align 4 | |
913 | L(Exit32): | |
914 | movdqu (%esi), %xmm0 | |
915 | movdqu 16(%esi), %xmm2 | |
916 | movdqu %xmm0, (%edi) | |
917 | movdqu %xmm2, 16(%edi) | |
918 | # ifdef USE_AS_STPCPY | |
919 | lea 31(%edi), %eax | |
920 | # endif | |
921 | sub $32, %ebx | |
922 | lea 32(%edi), %edi | |
923 | jnz L(StrncpyFillTailWithZero) | |
924 | RETURN | |
925 | ||
926 | .p2align 4 | |
927 | L(StrncpyExit1): | |
928 | movb (%esi), %dl | |
929 | movb %dl, (%edi) | |
930 | # ifdef USE_AS_STPCPY | |
931 | lea 1(%edi), %eax | |
932 | # endif | |
933 | RETURN | |
934 | ||
935 | .p2align 4 | |
936 | L(StrncpyExit2): | |
937 | movw (%esi), %dx | |
938 | movw %dx, (%edi) | |
939 | # ifdef USE_AS_STPCPY | |
940 | lea 2(%edi), %eax | |
941 | # endif | |
942 | RETURN | |
943 | ||
944 | .p2align 4 | |
945 | L(StrncpyExit3): | |
946 | movw (%esi), %cx | |
947 | movb 2(%esi), %dl | |
948 | movw %cx, (%edi) | |
949 | movb %dl, 2(%edi) | |
950 | # ifdef USE_AS_STPCPY | |
951 | lea 3(%edi), %eax | |
952 | # endif | |
953 | RETURN | |
954 | ||
955 | .p2align 4 | |
956 | L(StrncpyExit4): | |
957 | movl (%esi), %edx | |
958 | movl %edx, (%edi) | |
959 | # ifdef USE_AS_STPCPY | |
960 | lea 4(%edi), %eax | |
961 | # endif | |
962 | RETURN | |
963 | ||
964 | .p2align 4 | |
965 | L(StrncpyExit5): | |
966 | movl (%esi), %ecx | |
967 | movb 4(%esi), %dl | |
968 | movl %ecx, (%edi) | |
969 | movb %dl, 4(%edi) | |
970 | # ifdef USE_AS_STPCPY | |
971 | lea 5(%edi), %eax | |
972 | # endif | |
973 | RETURN | |
974 | ||
975 | .p2align 4 | |
976 | L(StrncpyExit6): | |
977 | movl (%esi), %ecx | |
978 | movw 4(%esi), %dx | |
979 | movl %ecx, (%edi) | |
980 | movw %dx, 4(%edi) | |
981 | # ifdef USE_AS_STPCPY | |
982 | lea 6(%edi), %eax | |
983 | # endif | |
984 | RETURN | |
985 | ||
986 | .p2align 4 | |
987 | L(StrncpyExit7): | |
988 | movl (%esi), %ecx | |
989 | movl 3(%esi), %edx | |
990 | movl %ecx, (%edi) | |
991 | movl %edx, 3(%edi) | |
992 | # ifdef USE_AS_STPCPY | |
993 | lea 7(%edi), %eax | |
994 | # endif | |
995 | RETURN | |
996 | ||
997 | .p2align 4 | |
998 | L(StrncpyExit8): | |
999 | movlpd (%esi), %xmm0 | |
1000 | movlpd %xmm0, (%edi) | |
1001 | # ifdef USE_AS_STPCPY | |
1002 | lea 8(%edi), %eax | |
1003 | # endif | |
1004 | RETURN | |
1005 | ||
1006 | .p2align 4 | |
1007 | L(StrncpyExit9): | |
1008 | movlpd (%esi), %xmm0 | |
1009 | movb 8(%esi), %dl | |
1010 | movlpd %xmm0, (%edi) | |
1011 | movb %dl, 8(%edi) | |
1012 | # ifdef USE_AS_STPCPY | |
1013 | lea 9(%edi), %eax | |
1014 | # endif | |
1015 | RETURN | |
1016 | ||
1017 | .p2align 4 | |
1018 | L(StrncpyExit10): | |
1019 | movlpd (%esi), %xmm0 | |
1020 | movw 8(%esi), %dx | |
1021 | movlpd %xmm0, (%edi) | |
1022 | movw %dx, 8(%edi) | |
1023 | # ifdef USE_AS_STPCPY | |
1024 | lea 10(%edi), %eax | |
1025 | # endif | |
1026 | RETURN | |
1027 | ||
1028 | .p2align 4 | |
1029 | L(StrncpyExit11): | |
1030 | movlpd (%esi), %xmm0 | |
1031 | movl 7(%esi), %edx | |
1032 | movlpd %xmm0, (%edi) | |
1033 | movl %edx, 7(%edi) | |
1034 | # ifdef USE_AS_STPCPY | |
1035 | lea 11(%edi), %eax | |
1036 | # endif | |
1037 | RETURN | |
1038 | ||
1039 | .p2align 4 | |
1040 | L(StrncpyExit12): | |
1041 | movlpd (%esi), %xmm0 | |
1042 | movl 8(%esi), %edx | |
1043 | movlpd %xmm0, (%edi) | |
1044 | movl %edx, 8(%edi) | |
1045 | # ifdef USE_AS_STPCPY | |
1046 | lea 12(%edi), %eax | |
1047 | # endif | |
1048 | RETURN | |
1049 | ||
1050 | .p2align 4 | |
1051 | L(StrncpyExit13): | |
1052 | movlpd (%esi), %xmm0 | |
1053 | movlpd 5(%esi), %xmm1 | |
1054 | movlpd %xmm0, (%edi) | |
1055 | movlpd %xmm1, 5(%edi) | |
1056 | # ifdef USE_AS_STPCPY | |
1057 | lea 13(%edi), %eax | |
1058 | # endif | |
1059 | RETURN | |
1060 | ||
1061 | .p2align 4 | |
1062 | L(StrncpyExit14): | |
1063 | movlpd (%esi), %xmm0 | |
1064 | movlpd 6(%esi), %xmm1 | |
1065 | movlpd %xmm0, (%edi) | |
1066 | movlpd %xmm1, 6(%edi) | |
1067 | # ifdef USE_AS_STPCPY | |
1068 | lea 14(%edi), %eax | |
1069 | # endif | |
1070 | RETURN | |
1071 | ||
1072 | .p2align 4 | |
1073 | L(StrncpyExit15): | |
1074 | movlpd (%esi), %xmm0 | |
1075 | movlpd 7(%esi), %xmm1 | |
1076 | movlpd %xmm0, (%edi) | |
1077 | movlpd %xmm1, 7(%edi) | |
1078 | # ifdef USE_AS_STPCPY | |
1079 | lea 15(%edi), %eax | |
1080 | # endif | |
1081 | RETURN | |
1082 | ||
1083 | .p2align 4 | |
1084 | L(StrncpyExit16): | |
1085 | movdqu (%esi), %xmm0 | |
1086 | movdqu %xmm0, (%edi) | |
1087 | # ifdef USE_AS_STPCPY | |
1088 | lea 16(%edi), %eax | |
1089 | # endif | |
1090 | RETURN | |
1091 | ||
1092 | .p2align 4 | |
1093 | L(StrncpyExit17): | |
1094 | movdqu (%esi), %xmm0 | |
1095 | movb 16(%esi), %cl | |
1096 | movdqu %xmm0, (%edi) | |
1097 | movb %cl, 16(%edi) | |
1098 | # ifdef USE_AS_STPCPY | |
1099 | lea 17(%edi), %eax | |
1100 | # endif | |
1101 | RETURN | |
1102 | ||
1103 | .p2align 4 | |
1104 | L(StrncpyExit18): | |
1105 | movdqu (%esi), %xmm0 | |
1106 | movw 16(%esi), %cx | |
1107 | movdqu %xmm0, (%edi) | |
1108 | movw %cx, 16(%edi) | |
1109 | # ifdef USE_AS_STPCPY | |
1110 | lea 18(%edi), %eax | |
1111 | # endif | |
1112 | RETURN | |
1113 | ||
1114 | .p2align 4 | |
1115 | L(StrncpyExit19): | |
1116 | movdqu (%esi), %xmm0 | |
1117 | movl 15(%esi), %ecx | |
1118 | movdqu %xmm0, (%edi) | |
1119 | movl %ecx, 15(%edi) | |
1120 | # ifdef USE_AS_STPCPY | |
1121 | lea 19(%edi), %eax | |
1122 | # endif | |
1123 | RETURN | |
1124 | ||
1125 | .p2align 4 | |
1126 | L(StrncpyExit20): | |
1127 | movdqu (%esi), %xmm0 | |
1128 | movl 16(%esi), %ecx | |
1129 | movdqu %xmm0, (%edi) | |
1130 | movl %ecx, 16(%edi) | |
1131 | # ifdef USE_AS_STPCPY | |
1132 | lea 20(%edi), %eax | |
1133 | # endif | |
1134 | RETURN | |
1135 | ||
1136 | .p2align 4 | |
1137 | L(StrncpyExit21): | |
1138 | movdqu (%esi), %xmm0 | |
1139 | movl 16(%esi), %ecx | |
1140 | movb 20(%esi), %dl | |
1141 | movdqu %xmm0, (%edi) | |
1142 | movl %ecx, 16(%edi) | |
1143 | movb %dl, 20(%edi) | |
1144 | # ifdef USE_AS_STPCPY | |
1145 | lea 21(%edi), %eax | |
1146 | # endif | |
1147 | RETURN | |
1148 | ||
1149 | .p2align 4 | |
1150 | L(StrncpyExit22): | |
1151 | movdqu (%esi), %xmm0 | |
1152 | movlpd 14(%esi), %xmm3 | |
1153 | movdqu %xmm0, (%edi) | |
1154 | movlpd %xmm3, 14(%edi) | |
1155 | # ifdef USE_AS_STPCPY | |
1156 | lea 22(%edi), %eax | |
1157 | # endif | |
1158 | RETURN | |
1159 | ||
1160 | .p2align 4 | |
1161 | L(StrncpyExit23): | |
1162 | movdqu (%esi), %xmm0 | |
1163 | movlpd 15(%esi), %xmm3 | |
1164 | movdqu %xmm0, (%edi) | |
1165 | movlpd %xmm3, 15(%edi) | |
1166 | # ifdef USE_AS_STPCPY | |
1167 | lea 23(%edi), %eax | |
1168 | # endif | |
1169 | RETURN | |
1170 | ||
1171 | .p2align 4 | |
1172 | L(StrncpyExit24): | |
1173 | movdqu (%esi), %xmm0 | |
1174 | movlpd 16(%esi), %xmm2 | |
1175 | movdqu %xmm0, (%edi) | |
1176 | movlpd %xmm2, 16(%edi) | |
1177 | # ifdef USE_AS_STPCPY | |
1178 | lea 24(%edi), %eax | |
1179 | # endif | |
1180 | RETURN | |
1181 | ||
1182 | .p2align 4 | |
1183 | L(StrncpyExit25): | |
1184 | movdqu (%esi), %xmm0 | |
1185 | movlpd 16(%esi), %xmm2 | |
1186 | movb 24(%esi), %cl | |
1187 | movdqu %xmm0, (%edi) | |
1188 | movlpd %xmm2, 16(%edi) | |
1189 | movb %cl, 24(%edi) | |
1190 | # ifdef USE_AS_STPCPY | |
1191 | lea 25(%edi), %eax | |
1192 | # endif | |
1193 | RETURN | |
1194 | ||
1195 | .p2align 4 | |
1196 | L(StrncpyExit26): | |
1197 | movdqu (%esi), %xmm0 | |
1198 | movlpd 16(%esi), %xmm2 | |
1199 | movw 24(%esi), %cx | |
1200 | movdqu %xmm0, (%edi) | |
1201 | movlpd %xmm2, 16(%edi) | |
1202 | movw %cx, 24(%edi) | |
1203 | # ifdef USE_AS_STPCPY | |
1204 | lea 26(%edi), %eax | |
1205 | # endif | |
1206 | RETURN | |
1207 | ||
1208 | .p2align 4 | |
1209 | L(StrncpyExit27): | |
1210 | movdqu (%esi), %xmm0 | |
1211 | movlpd 16(%esi), %xmm2 | |
1212 | movl 23(%esi), %ecx | |
1213 | movdqu %xmm0, (%edi) | |
1214 | movlpd %xmm2, 16(%edi) | |
1215 | movl %ecx, 23(%edi) | |
1216 | # ifdef USE_AS_STPCPY | |
1217 | lea 27(%edi), %eax | |
1218 | # endif | |
1219 | RETURN | |
1220 | ||
1221 | .p2align 4 | |
1222 | L(StrncpyExit28): | |
1223 | movdqu (%esi), %xmm0 | |
1224 | movlpd 16(%esi), %xmm2 | |
1225 | movl 24(%esi), %ecx | |
1226 | movdqu %xmm0, (%edi) | |
1227 | movlpd %xmm2, 16(%edi) | |
1228 | movl %ecx, 24(%edi) | |
1229 | # ifdef USE_AS_STPCPY | |
1230 | lea 28(%edi), %eax | |
1231 | # endif | |
1232 | RETURN | |
1233 | ||
1234 | .p2align 4 | |
1235 | L(StrncpyExit29): | |
1236 | movdqu (%esi), %xmm0 | |
1237 | movdqu 13(%esi), %xmm2 | |
1238 | movdqu %xmm0, (%edi) | |
1239 | movdqu %xmm2, 13(%edi) | |
1240 | # ifdef USE_AS_STPCPY | |
1241 | lea 29(%edi), %eax | |
1242 | # endif | |
1243 | RETURN | |
1244 | ||
1245 | .p2align 4 | |
1246 | L(StrncpyExit30): | |
1247 | movdqu (%esi), %xmm0 | |
1248 | movdqu 14(%esi), %xmm2 | |
1249 | movdqu %xmm0, (%edi) | |
1250 | movdqu %xmm2, 14(%edi) | |
1251 | # ifdef USE_AS_STPCPY | |
1252 | lea 30(%edi), %eax | |
1253 | # endif | |
1254 | RETURN | |
1255 | ||
1256 | .p2align 4 | |
1257 | L(StrncpyExit31): | |
1258 | movdqu (%esi), %xmm0 | |
1259 | movdqu 15(%esi), %xmm2 | |
1260 | movdqu %xmm0, (%edi) | |
1261 | movdqu %xmm2, 15(%edi) | |
1262 | # ifdef USE_AS_STPCPY | |
1263 | lea 31(%edi), %eax | |
1264 | # endif | |
1265 | RETURN | |
1266 | ||
1267 | .p2align 4 | |
1268 | L(StrncpyExit32): | |
1269 | movdqu (%esi), %xmm0 | |
1270 | movdqu 16(%esi), %xmm2 | |
1271 | movdqu %xmm0, (%edi) | |
1272 | movdqu %xmm2, 16(%edi) | |
1273 | # ifdef USE_AS_STPCPY | |
1274 | lea 32(%edi), %eax | |
1275 | # endif | |
1276 | RETURN | |
1277 | ||
1278 | .p2align 4 | |
1279 | L(StrncpyExit33): | |
1280 | movdqu (%esi), %xmm0 | |
1281 | movdqu 16(%esi), %xmm2 | |
1282 | movb 32(%esi), %cl | |
1283 | movdqu %xmm0, (%edi) | |
1284 | movdqu %xmm2, 16(%edi) | |
1285 | movb %cl, 32(%edi) | |
1286 | RETURN | |
1287 | ||
1288 | .p2align 4 | |
1289 | L(Fill0): | |
1290 | RETURN | |
1291 | ||
1292 | .p2align 4 | |
1293 | L(Fill1): | |
1294 | movb %dl, (%edi) | |
1295 | RETURN | |
1296 | ||
1297 | .p2align 4 | |
1298 | L(Fill2): | |
1299 | movw %dx, (%edi) | |
1300 | RETURN | |
1301 | ||
1302 | .p2align 4 | |
1303 | L(Fill3): | |
1304 | movl %edx, -1(%edi) | |
1305 | RETURN | |
1306 | ||
1307 | .p2align 4 | |
1308 | L(Fill4): | |
1309 | movl %edx, (%edi) | |
1310 | RETURN | |
1311 | ||
1312 | .p2align 4 | |
1313 | L(Fill5): | |
1314 | movl %edx, (%edi) | |
1315 | movb %dl, 4(%edi) | |
1316 | RETURN | |
1317 | ||
1318 | .p2align 4 | |
1319 | L(Fill6): | |
1320 | movl %edx, (%edi) | |
1321 | movw %dx, 4(%edi) | |
1322 | RETURN | |
1323 | ||
1324 | .p2align 4 | |
1325 | L(Fill7): | |
1326 | movlpd %xmm0, -1(%edi) | |
1327 | RETURN | |
1328 | ||
1329 | .p2align 4 | |
1330 | L(Fill8): | |
1331 | movlpd %xmm0, (%edi) | |
1332 | RETURN | |
1333 | ||
1334 | .p2align 4 | |
1335 | L(Fill9): | |
1336 | movlpd %xmm0, (%edi) | |
1337 | movb %dl, 8(%edi) | |
1338 | RETURN | |
1339 | ||
1340 | .p2align 4 | |
1341 | L(Fill10): | |
1342 | movlpd %xmm0, (%edi) | |
1343 | movw %dx, 8(%edi) | |
1344 | RETURN | |
1345 | ||
1346 | .p2align 4 | |
1347 | L(Fill11): | |
1348 | movlpd %xmm0, (%edi) | |
1349 | movl %edx, 7(%edi) | |
1350 | RETURN | |
1351 | ||
1352 | .p2align 4 | |
1353 | L(Fill12): | |
1354 | movlpd %xmm0, (%edi) | |
1355 | movl %edx, 8(%edi) | |
1356 | RETURN | |
1357 | ||
1358 | .p2align 4 | |
1359 | L(Fill13): | |
1360 | movlpd %xmm0, (%edi) | |
1361 | movlpd %xmm0, 5(%edi) | |
1362 | RETURN | |
1363 | ||
1364 | .p2align 4 | |
1365 | L(Fill14): | |
1366 | movlpd %xmm0, (%edi) | |
1367 | movlpd %xmm0, 6(%edi) | |
1368 | RETURN | |
1369 | ||
1370 | .p2align 4 | |
1371 | L(Fill15): | |
1372 | movdqu %xmm0, -1(%edi) | |
1373 | RETURN | |
1374 | ||
1375 | .p2align 4 | |
1376 | L(Fill16): | |
1377 | movdqu %xmm0, (%edi) | |
1378 | RETURN | |
1379 | ||
1380 | .p2align 4 | |
1381 | L(CopyFrom1To16BytesUnalignedXmm2): | |
1382 | movdqu %xmm2, (%edi, %ecx) | |
1383 | ||
1384 | .p2align 4 | |
1385 | L(CopyFrom1To16BytesXmmExit): | |
1386 | bsf %edx, %edx | |
1387 | add $15, %ebx | |
1388 | add %ecx, %edi | |
1389 | # ifdef USE_AS_STPCPY | |
1390 | lea (%edi, %edx), %eax | |
1391 | # endif | |
1392 | sub %edx, %ebx | |
1393 | lea 1(%edi, %edx), %edi | |
1394 | ||
1395 | .p2align 4 | |
1396 | L(StrncpyFillTailWithZero): | |
1397 | pxor %xmm0, %xmm0 | |
1398 | xor %edx, %edx | |
1399 | sub $16, %ebx | |
1400 | jbe L(StrncpyFillExit) | |
1401 | ||
1402 | movdqu %xmm0, (%edi) | |
1403 | add $16, %edi | |
1404 | ||
1405 | mov %edi, %esi | |
1406 | and $0xf, %esi | |
1407 | sub %esi, %edi | |
1408 | add %esi, %ebx | |
1409 | sub $64, %ebx | |
1410 | jb L(StrncpyFillLess64) | |
1411 | ||
1412 | L(StrncpyFillLoopMovdqa): | |
1413 | movdqa %xmm0, (%edi) | |
1414 | movdqa %xmm0, 16(%edi) | |
1415 | movdqa %xmm0, 32(%edi) | |
1416 | movdqa %xmm0, 48(%edi) | |
1417 | add $64, %edi | |
1418 | sub $64, %ebx | |
1419 | jae L(StrncpyFillLoopMovdqa) | |
1420 | ||
1421 | L(StrncpyFillLess64): | |
1422 | add $32, %ebx | |
1423 | jl L(StrncpyFillLess32) | |
1424 | movdqa %xmm0, (%edi) | |
1425 | movdqa %xmm0, 16(%edi) | |
1426 | add $32, %edi | |
1427 | sub $16, %ebx | |
1428 | jl L(StrncpyFillExit) | |
1429 | movdqa %xmm0, (%edi) | |
1430 | add $16, %edi | |
1431 | BRANCH_TO_JMPTBL_ENTRY (L(FillTable), %ebx, 4) | |
1432 | ||
1433 | L(StrncpyFillLess32): | |
1434 | add $16, %ebx | |
1435 | jl L(StrncpyFillExit) | |
1436 | movdqa %xmm0, (%edi) | |
1437 | add $16, %edi | |
1438 | BRANCH_TO_JMPTBL_ENTRY (L(FillTable), %ebx, 4) | |
1439 | ||
1440 | L(StrncpyFillExit): | |
1441 | add $16, %ebx | |
1442 | BRANCH_TO_JMPTBL_ENTRY (L(FillTable), %ebx, 4) | |
1443 | ||
1444 | .p2align 4 | |
1445 | L(UnalignedLeaveCase2OrCase3): | |
1446 | test %edx, %edx | |
1447 | jnz L(Unaligned64LeaveCase2) | |
1448 | L(Unaligned64LeaveCase3): | |
1449 | lea 64(%ebx), %ecx | |
1450 | and $-16, %ecx | |
1451 | add $48, %ebx | |
1452 | jl L(CopyFrom1To16BytesCase3) | |
1453 | movdqu %xmm4, (%edi) | |
1454 | sub $16, %ebx | |
1455 | jb L(CopyFrom1To16BytesCase3) | |
1456 | movdqu %xmm5, 16(%edi) | |
1457 | sub $16, %ebx | |
1458 | jb L(CopyFrom1To16BytesCase3) | |
1459 | movdqu %xmm6, 32(%edi) | |
1460 | sub $16, %ebx | |
1461 | jb L(CopyFrom1To16BytesCase3) | |
1462 | movdqu %xmm7, 48(%edi) | |
1463 | # ifdef USE_AS_STPCPY | |
1464 | lea 64(%edi), %eax | |
1465 | # endif | |
1466 | RETURN | |
1467 | ||
1468 | .p2align 4 | |
1469 | L(Unaligned64LeaveCase2): | |
1470 | xor %ecx, %ecx | |
1471 | pcmpeqb %xmm4, %xmm0 | |
1472 | pmovmskb %xmm0, %edx | |
1473 | add $48, %ebx | |
1474 | jle L(CopyFrom1To16BytesCase2OrCase3) | |
1475 | test %edx, %edx | |
1476 | jnz L(CopyFrom1To16BytesUnalignedXmm4) | |
1477 | ||
1478 | pcmpeqb %xmm5, %xmm0 | |
1479 | pmovmskb %xmm0, %edx | |
1480 | movdqu %xmm4, (%edi) | |
1481 | add $16, %ecx | |
1482 | sub $16, %ebx | |
1483 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
1484 | test %edx, %edx | |
1485 | jnz L(CopyFrom1To16BytesUnalignedXmm5) | |
1486 | ||
1487 | pcmpeqb %xmm6, %xmm0 | |
1488 | pmovmskb %xmm0, %edx | |
1489 | movdqu %xmm5, 16(%edi) | |
1490 | add $16, %ecx | |
1491 | sub $16, %ebx | |
1492 | jbe L(CopyFrom1To16BytesCase2OrCase3) | |
1493 | test %edx, %edx | |
1494 | jnz L(CopyFrom1To16BytesUnalignedXmm6) | |
1495 | ||
1496 | pcmpeqb %xmm7, %xmm0 | |
1497 | pmovmskb %xmm0, %edx | |
1498 | movdqu %xmm6, 32(%edi) | |
1499 | lea 16(%edi, %ecx), %edi | |
1500 | lea 16(%esi, %ecx), %esi | |
1501 | bsf %edx, %edx | |
1502 | cmp %ebx, %edx | |
1503 | jb L(CopyFrom1To16BytesExit) | |
1504 | BRANCH_TO_JMPTBL_ENTRY (L(ExitStrncpyTable), %ebx, 4) | |
1505 | ||
1506 | .p2align 4 | |
1507 | L(ExitZero): | |
1508 | movl %edi, %eax | |
1509 | RETURN | |
1510 | ||
1511 | END (STRCPY) | |
1512 | ||
1513 | .p2align 4 | |
1514 | .section .rodata | |
1515 | L(ExitTable): | |
1516 | .int JMPTBL(L(Exit1), L(ExitTable)) | |
1517 | .int JMPTBL(L(Exit2), L(ExitTable)) | |
1518 | .int JMPTBL(L(Exit3), L(ExitTable)) | |
1519 | .int JMPTBL(L(Exit4), L(ExitTable)) | |
1520 | .int JMPTBL(L(Exit5), L(ExitTable)) | |
1521 | .int JMPTBL(L(Exit6), L(ExitTable)) | |
1522 | .int JMPTBL(L(Exit7), L(ExitTable)) | |
1523 | .int JMPTBL(L(Exit8), L(ExitTable)) | |
1524 | .int JMPTBL(L(Exit9), L(ExitTable)) | |
1525 | .int JMPTBL(L(Exit10), L(ExitTable)) | |
1526 | .int JMPTBL(L(Exit11), L(ExitTable)) | |
1527 | .int JMPTBL(L(Exit12), L(ExitTable)) | |
1528 | .int JMPTBL(L(Exit13), L(ExitTable)) | |
1529 | .int JMPTBL(L(Exit14), L(ExitTable)) | |
1530 | .int JMPTBL(L(Exit15), L(ExitTable)) | |
1531 | .int JMPTBL(L(Exit16), L(ExitTable)) | |
1532 | .int JMPTBL(L(Exit17), L(ExitTable)) | |
1533 | .int JMPTBL(L(Exit18), L(ExitTable)) | |
1534 | .int JMPTBL(L(Exit19), L(ExitTable)) | |
1535 | .int JMPTBL(L(Exit20), L(ExitTable)) | |
1536 | .int JMPTBL(L(Exit21), L(ExitTable)) | |
1537 | .int JMPTBL(L(Exit22), L(ExitTable)) | |
1538 | .int JMPTBL(L(Exit23), L(ExitTable)) | |
1539 | .int JMPTBL(L(Exit24), L(ExitTable)) | |
1540 | .int JMPTBL(L(Exit25), L(ExitTable)) | |
1541 | .int JMPTBL(L(Exit26), L(ExitTable)) | |
1542 | .int JMPTBL(L(Exit27), L(ExitTable)) | |
1543 | .int JMPTBL(L(Exit28), L(ExitTable)) | |
1544 | .int JMPTBL(L(Exit29), L(ExitTable)) | |
1545 | .int JMPTBL(L(Exit30), L(ExitTable)) | |
1546 | .int JMPTBL(L(Exit31), L(ExitTable)) | |
1547 | .int JMPTBL(L(Exit32), L(ExitTable)) | |
1548 | ||
1549 | L(ExitStrncpyTable): | |
1550 | .int JMPTBL(L(Exit0), L(ExitStrncpyTable)) | |
1551 | .int JMPTBL(L(StrncpyExit1), L(ExitStrncpyTable)) | |
1552 | .int JMPTBL(L(StrncpyExit2), L(ExitStrncpyTable)) | |
1553 | .int JMPTBL(L(StrncpyExit3), L(ExitStrncpyTable)) | |
1554 | .int JMPTBL(L(StrncpyExit4), L(ExitStrncpyTable)) | |
1555 | .int JMPTBL(L(StrncpyExit5), L(ExitStrncpyTable)) | |
1556 | .int JMPTBL(L(StrncpyExit6), L(ExitStrncpyTable)) | |
1557 | .int JMPTBL(L(StrncpyExit7), L(ExitStrncpyTable)) | |
1558 | .int JMPTBL(L(StrncpyExit8), L(ExitStrncpyTable)) | |
1559 | .int JMPTBL(L(StrncpyExit9), L(ExitStrncpyTable)) | |
1560 | .int JMPTBL(L(StrncpyExit10), L(ExitStrncpyTable)) | |
1561 | .int JMPTBL(L(StrncpyExit11), L(ExitStrncpyTable)) | |
1562 | .int JMPTBL(L(StrncpyExit12), L(ExitStrncpyTable)) | |
1563 | .int JMPTBL(L(StrncpyExit13), L(ExitStrncpyTable)) | |
1564 | .int JMPTBL(L(StrncpyExit14), L(ExitStrncpyTable)) | |
1565 | .int JMPTBL(L(StrncpyExit15), L(ExitStrncpyTable)) | |
1566 | .int JMPTBL(L(StrncpyExit16), L(ExitStrncpyTable)) | |
1567 | .int JMPTBL(L(StrncpyExit17), L(ExitStrncpyTable)) | |
1568 | .int JMPTBL(L(StrncpyExit18), L(ExitStrncpyTable)) | |
1569 | .int JMPTBL(L(StrncpyExit19), L(ExitStrncpyTable)) | |
1570 | .int JMPTBL(L(StrncpyExit20), L(ExitStrncpyTable)) | |
1571 | .int JMPTBL(L(StrncpyExit21), L(ExitStrncpyTable)) | |
1572 | .int JMPTBL(L(StrncpyExit22), L(ExitStrncpyTable)) | |
1573 | .int JMPTBL(L(StrncpyExit23), L(ExitStrncpyTable)) | |
1574 | .int JMPTBL(L(StrncpyExit24), L(ExitStrncpyTable)) | |
1575 | .int JMPTBL(L(StrncpyExit25), L(ExitStrncpyTable)) | |
1576 | .int JMPTBL(L(StrncpyExit26), L(ExitStrncpyTable)) | |
1577 | .int JMPTBL(L(StrncpyExit27), L(ExitStrncpyTable)) | |
1578 | .int JMPTBL(L(StrncpyExit28), L(ExitStrncpyTable)) | |
1579 | .int JMPTBL(L(StrncpyExit29), L(ExitStrncpyTable)) | |
1580 | .int JMPTBL(L(StrncpyExit30), L(ExitStrncpyTable)) | |
1581 | .int JMPTBL(L(StrncpyExit31), L(ExitStrncpyTable)) | |
1582 | .int JMPTBL(L(StrncpyExit32), L(ExitStrncpyTable)) | |
1583 | .int JMPTBL(L(StrncpyExit33), L(ExitStrncpyTable)) | |
1584 | ||
1585 | .p2align 4 | |
1586 | L(FillTable): | |
1587 | .int JMPTBL(L(Fill0), L(FillTable)) | |
1588 | .int JMPTBL(L(Fill1), L(FillTable)) | |
1589 | .int JMPTBL(L(Fill2), L(FillTable)) | |
1590 | .int JMPTBL(L(Fill3), L(FillTable)) | |
1591 | .int JMPTBL(L(Fill4), L(FillTable)) | |
1592 | .int JMPTBL(L(Fill5), L(FillTable)) | |
1593 | .int JMPTBL(L(Fill6), L(FillTable)) | |
1594 | .int JMPTBL(L(Fill7), L(FillTable)) | |
1595 | .int JMPTBL(L(Fill8), L(FillTable)) | |
1596 | .int JMPTBL(L(Fill9), L(FillTable)) | |
1597 | .int JMPTBL(L(Fill10), L(FillTable)) | |
1598 | .int JMPTBL(L(Fill11), L(FillTable)) | |
1599 | .int JMPTBL(L(Fill12), L(FillTable)) | |
1600 | .int JMPTBL(L(Fill13), L(FillTable)) | |
1601 | .int JMPTBL(L(Fill14), L(FillTable)) | |
1602 | .int JMPTBL(L(Fill15), L(FillTable)) | |
1603 | .int JMPTBL(L(Fill16), L(FillTable)) | |
1604 | # else | |
1605 | # define PARMS 4 | |
1606 | # define ENTRANCE | |
1607 | # define RETURN POP (%edi); ret; CFI_PUSH (%edi) | |
1608 | # define RETURN1 ret | |
1609 | ||
1610 | .text | |
1611 | ENTRY (STRCPY) | |
1612 | ENTRANCE | |
1613 | mov STR1(%esp), %edx | |
1614 | mov STR2(%esp), %ecx | |
1615 | ||
1616 | cmpb $0, (%ecx) | |
1617 | jz L(ExitTail1) | |
1618 | cmpb $0, 1(%ecx) | |
1619 | jz L(ExitTail2) | |
1620 | cmpb $0, 2(%ecx) | |
1621 | jz L(ExitTail3) | |
1622 | cmpb $0, 3(%ecx) | |
1623 | jz L(ExitTail4) | |
1624 | cmpb $0, 4(%ecx) | |
1625 | jz L(ExitTail5) | |
1626 | cmpb $0, 5(%ecx) | |
1627 | jz L(ExitTail6) | |
1628 | cmpb $0, 6(%ecx) | |
1629 | jz L(ExitTail7) | |
1630 | cmpb $0, 7(%ecx) | |
1631 | jz L(ExitTail8) | |
1632 | cmpb $0, 8(%ecx) | |
1633 | jz L(ExitTail9) | |
1634 | cmpb $0, 9(%ecx) | |
1635 | jz L(ExitTail10) | |
1636 | cmpb $0, 10(%ecx) | |
1637 | jz L(ExitTail11) | |
1638 | cmpb $0, 11(%ecx) | |
1639 | jz L(ExitTail12) | |
1640 | cmpb $0, 12(%ecx) | |
1641 | jz L(ExitTail13) | |
1642 | cmpb $0, 13(%ecx) | |
1643 | jz L(ExitTail14) | |
1644 | cmpb $0, 14(%ecx) | |
1645 | jz L(ExitTail15) | |
1646 | cmpb $0, 15(%ecx) | |
1647 | jz L(ExitTail16) | |
1648 | ||
1649 | PUSH (%edi) | |
1650 | PUSH (%ebx) | |
1651 | ||
1652 | mov %edx, %edi | |
1653 | lea 16(%ecx), %ebx | |
1654 | and $-16, %ebx | |
1655 | pxor %xmm0, %xmm0 | |
1656 | movdqu (%ecx), %xmm1 | |
1657 | movdqu %xmm1, (%edx) | |
1658 | pcmpeqb (%ebx), %xmm0 | |
1659 | pmovmskb %xmm0, %eax | |
1660 | sub %ecx, %ebx | |
1661 | test %eax, %eax | |
1662 | jnz L(CopyFrom1To16Bytes) | |
1663 | ||
1664 | mov %ecx, %eax | |
1665 | lea 16(%ecx), %ecx | |
1666 | and $-16, %ecx | |
1667 | sub %ecx, %eax | |
1668 | sub %eax, %edx | |
1669 | xor %ebx, %ebx | |
1670 | ||
1671 | .p2align 4 | |
1672 | movdqa (%ecx), %xmm1 | |
1673 | movaps 16(%ecx), %xmm2 | |
1674 | movdqu %xmm1, (%edx) | |
1675 | pcmpeqb %xmm2, %xmm0 | |
1676 | pmovmskb %xmm0, %eax | |
1677 | add $16, %ebx | |
1678 | test %eax, %eax | |
1679 | jnz L(CopyFrom1To16Bytes) | |
1680 | ||
1681 | movaps 16(%ecx, %ebx), %xmm3 | |
1682 | movdqu %xmm2, (%edx, %ebx) | |
1683 | pcmpeqb %xmm3, %xmm0 | |
1684 | pmovmskb %xmm0, %eax | |
1685 | add $16, %ebx | |
1686 | test %eax, %eax | |
1687 | jnz L(CopyFrom1To16Bytes) | |
1688 | ||
1689 | movaps 16(%ecx, %ebx), %xmm4 | |
1690 | movdqu %xmm3, (%edx, %ebx) | |
1691 | pcmpeqb %xmm4, %xmm0 | |
1692 | pmovmskb %xmm0, %eax | |
1693 | add $16, %ebx | |
1694 | test %eax, %eax | |
1695 | jnz L(CopyFrom1To16Bytes) | |
1696 | ||
1697 | movaps 16(%ecx, %ebx), %xmm1 | |
1698 | movdqu %xmm4, (%edx, %ebx) | |
1699 | pcmpeqb %xmm1, %xmm0 | |
1700 | pmovmskb %xmm0, %eax | |
1701 | add $16, %ebx | |
1702 | test %eax, %eax | |
1703 | jnz L(CopyFrom1To16Bytes) | |
1704 | ||
1705 | movaps 16(%ecx, %ebx), %xmm2 | |
1706 | movdqu %xmm1, (%edx, %ebx) | |
1707 | pcmpeqb %xmm2, %xmm0 | |
1708 | pmovmskb %xmm0, %eax | |
1709 | add $16, %ebx | |
1710 | test %eax, %eax | |
1711 | jnz L(CopyFrom1To16Bytes) | |
1712 | ||
1713 | movaps 16(%ecx, %ebx), %xmm3 | |
1714 | movdqu %xmm2, (%edx, %ebx) | |
1715 | pcmpeqb %xmm3, %xmm0 | |
1716 | pmovmskb %xmm0, %eax | |
1717 | add $16, %ebx | |
1718 | test %eax, %eax | |
1719 | jnz L(CopyFrom1To16Bytes) | |
1720 | ||
1721 | movdqu %xmm3, (%edx, %ebx) | |
1722 | mov %ecx, %eax | |
1723 | lea 16(%ecx, %ebx), %ecx | |
1724 | and $-0x40, %ecx | |
1725 | sub %ecx, %eax | |
1726 | sub %eax, %edx | |
1727 | ||
1728 | L(Aligned64Loop): | |
1729 | movaps (%ecx), %xmm2 | |
1730 | movaps %xmm2, %xmm4 | |
1731 | movaps 16(%ecx), %xmm5 | |
1732 | movaps 32(%ecx), %xmm3 | |
1733 | movaps %xmm3, %xmm6 | |
1734 | movaps 48(%ecx), %xmm7 | |
1735 | pminub %xmm5, %xmm2 | |
1736 | add $64, %ecx | |
1737 | pminub %xmm7, %xmm3 | |
1738 | add $64, %edx | |
1739 | pminub %xmm2, %xmm3 | |
1740 | pcmpeqb %xmm0, %xmm3 | |
1741 | pmovmskb %xmm3, %eax | |
1742 | test %eax, %eax | |
1743 | jnz L(Aligned64Leave) | |
1744 | L(Aligned64Loop_start): | |
1745 | movdqu %xmm4, -64(%edx) | |
1746 | movaps (%ecx), %xmm2 | |
1747 | movdqa %xmm2, %xmm4 | |
1748 | movdqu %xmm5, -48(%edx) | |
1749 | movaps 16(%ecx), %xmm5 | |
1750 | pminub %xmm5, %xmm2 | |
1751 | movaps 32(%ecx), %xmm3 | |
1752 | movdqu %xmm6, -32(%edx) | |
1753 | movaps %xmm3, %xmm6 | |
1754 | movdqu %xmm7, -16(%edx) | |
1755 | movaps 48(%ecx), %xmm7 | |
1756 | pminub %xmm7, %xmm3 | |
1757 | pminub %xmm2, %xmm3 | |
1758 | pcmpeqb %xmm3, %xmm0 | |
1759 | pmovmskb %xmm0, %eax | |
1760 | add $64, %edx | |
1761 | add $64, %ecx | |
1762 | test %eax, %eax | |
1763 | jz L(Aligned64Loop_start) | |
1764 | L(Aligned64Leave): | |
1765 | sub $0xa0, %ebx | |
1766 | pxor %xmm0, %xmm0 | |
1767 | pcmpeqb %xmm4, %xmm0 | |
1768 | pmovmskb %xmm0, %eax | |
1769 | test %eax, %eax | |
1770 | jnz L(CopyFrom1To16Bytes) | |
1771 | ||
1772 | pcmpeqb %xmm5, %xmm0 | |
1773 | pmovmskb %xmm0, %eax | |
1774 | movdqu %xmm4, -64(%edx) | |
1775 | test %eax, %eax | |
1776 | lea 16(%ebx), %ebx | |
1777 | jnz L(CopyFrom1To16Bytes) | |
1778 | ||
1779 | pcmpeqb %xmm6, %xmm0 | |
1780 | pmovmskb %xmm0, %eax | |
1781 | movdqu %xmm5, -48(%edx) | |
1782 | test %eax, %eax | |
1783 | lea 16(%ebx), %ebx | |
1784 | jnz L(CopyFrom1To16Bytes) | |
1785 | ||
1786 | movdqu %xmm6, -32(%edx) | |
1787 | pcmpeqb %xmm7, %xmm0 | |
1788 | pmovmskb %xmm0, %eax | |
1789 | lea 16(%ebx), %ebx | |
1790 | ||
1791 | /*-----------------End of main part---------------------------*/ | |
1792 | ||
1793 | .p2align 4 | |
1794 | L(CopyFrom1To16Bytes): | |
1795 | add %ebx, %edx | |
1796 | add %ebx, %ecx | |
1797 | ||
1798 | POP (%ebx) | |
1799 | test %al, %al | |
1800 | jz L(ExitHigh) | |
1801 | test $0x01, %al | |
1802 | jnz L(Exit1) | |
1803 | test $0x02, %al | |
1804 | jnz L(Exit2) | |
1805 | test $0x04, %al | |
1806 | jnz L(Exit3) | |
1807 | test $0x08, %al | |
1808 | jnz L(Exit4) | |
1809 | test $0x10, %al | |
1810 | jnz L(Exit5) | |
1811 | test $0x20, %al | |
1812 | jnz L(Exit6) | |
1813 | test $0x40, %al | |
1814 | jnz L(Exit7) | |
1815 | /* Exit 8 */ | |
1816 | movl (%ecx), %eax | |
1817 | movl %eax, (%edx) | |
1818 | movl 4(%ecx), %eax | |
1819 | movl %eax, 4(%edx) | |
1820 | # ifdef USE_AS_STPCPY | |
1821 | lea 7(%edx), %eax | |
1822 | # else | |
1823 | movl %edi, %eax | |
1824 | # endif | |
1825 | RETURN | |
1826 | ||
1827 | .p2align 4 | |
1828 | L(ExitHigh): | |
1829 | test $0x01, %ah | |
1830 | jnz L(Exit9) | |
1831 | test $0x02, %ah | |
1832 | jnz L(Exit10) | |
1833 | test $0x04, %ah | |
1834 | jnz L(Exit11) | |
1835 | test $0x08, %ah | |
1836 | jnz L(Exit12) | |
1837 | test $0x10, %ah | |
1838 | jnz L(Exit13) | |
1839 | test $0x20, %ah | |
1840 | jnz L(Exit14) | |
1841 | test $0x40, %ah | |
1842 | jnz L(Exit15) | |
1843 | /* Exit 16 */ | |
1844 | movlpd (%ecx), %xmm0 | |
1845 | movlpd %xmm0, (%edx) | |
1846 | movlpd 8(%ecx), %xmm0 | |
1847 | movlpd %xmm0, 8(%edx) | |
1848 | # ifdef USE_AS_STPCPY | |
1849 | lea 15(%edx), %eax | |
1850 | # else | |
1851 | movl %edi, %eax | |
1852 | # endif | |
1853 | RETURN | |
1854 | ||
1855 | .p2align 4 | |
1856 | L(Exit1): | |
1857 | movb (%ecx), %al | |
1858 | movb %al, (%edx) | |
1859 | # ifdef USE_AS_STPCPY | |
1860 | lea (%edx), %eax | |
1861 | # else | |
1862 | movl %edi, %eax | |
1863 | # endif | |
1864 | RETURN | |
1865 | ||
1866 | .p2align 4 | |
1867 | L(Exit2): | |
1868 | movw (%ecx), %ax | |
1869 | movw %ax, (%edx) | |
1870 | # ifdef USE_AS_STPCPY | |
1871 | lea 1(%edx), %eax | |
1872 | # else | |
1873 | movl %edi, %eax | |
1874 | # endif | |
1875 | RETURN | |
1876 | ||
1877 | .p2align 4 | |
1878 | L(Exit3): | |
1879 | movw (%ecx), %ax | |
1880 | movw %ax, (%edx) | |
1881 | movb 2(%ecx), %al | |
1882 | movb %al, 2(%edx) | |
1883 | # ifdef USE_AS_STPCPY | |
1884 | lea 2(%edx), %eax | |
1885 | # else | |
1886 | movl %edi, %eax | |
1887 | # endif | |
1888 | RETURN | |
1889 | ||
1890 | .p2align 4 | |
1891 | L(Exit4): | |
1892 | movl (%ecx), %eax | |
1893 | movl %eax, (%edx) | |
1894 | # ifdef USE_AS_STPCPY | |
1895 | lea 3(%edx), %eax | |
1896 | # else | |
1897 | movl %edi, %eax | |
1898 | # endif | |
1899 | RETURN | |
1900 | ||
1901 | .p2align 4 | |
1902 | L(Exit5): | |
1903 | movl (%ecx), %eax | |
1904 | movl %eax, (%edx) | |
1905 | movb 4(%ecx), %al | |
1906 | movb %al, 4(%edx) | |
1907 | # ifdef USE_AS_STPCPY | |
1908 | lea 4(%edx), %eax | |
1909 | # else | |
1910 | movl %edi, %eax | |
1911 | # endif | |
1912 | RETURN | |
1913 | ||
1914 | .p2align 4 | |
1915 | L(Exit6): | |
1916 | movl (%ecx), %eax | |
1917 | movl %eax, (%edx) | |
1918 | movw 4(%ecx), %ax | |
1919 | movw %ax, 4(%edx) | |
1920 | # ifdef USE_AS_STPCPY | |
1921 | lea 5(%edx), %eax | |
1922 | # else | |
1923 | movl %edi, %eax | |
1924 | # endif | |
1925 | RETURN | |
1926 | ||
1927 | .p2align 4 | |
1928 | L(Exit7): | |
1929 | movl (%ecx), %eax | |
1930 | movl %eax, (%edx) | |
1931 | movl 3(%ecx), %eax | |
1932 | movl %eax, 3(%edx) | |
1933 | # ifdef USE_AS_STPCPY | |
1934 | lea 6(%edx), %eax | |
1935 | # else | |
1936 | movl %edi, %eax | |
1937 | # endif | |
1938 | RETURN | |
1939 | ||
1940 | .p2align 4 | |
1941 | L(Exit9): | |
1942 | movl (%ecx), %eax | |
1943 | movl %eax, (%edx) | |
1944 | movl 4(%ecx), %eax | |
1945 | movl %eax, 4(%edx) | |
1946 | movb 8(%ecx), %al | |
1947 | movb %al, 8(%edx) | |
1948 | # ifdef USE_AS_STPCPY | |
1949 | lea 8(%edx), %eax | |
1950 | # else | |
1951 | movl %edi, %eax | |
1952 | # endif | |
1953 | RETURN | |
1954 | ||
1955 | .p2align 4 | |
1956 | L(Exit10): | |
1957 | movl (%ecx), %eax | |
1958 | movl %eax, (%edx) | |
1959 | movl 4(%ecx), %eax | |
1960 | movl %eax, 4(%edx) | |
1961 | movw 8(%ecx), %ax | |
1962 | movw %ax, 8(%edx) | |
1963 | # ifdef USE_AS_STPCPY | |
1964 | lea 9(%edx), %eax | |
1965 | # else | |
1966 | movl %edi, %eax | |
1967 | # endif | |
1968 | RETURN | |
1969 | ||
1970 | .p2align 4 | |
1971 | L(Exit11): | |
1972 | movl (%ecx), %eax | |
1973 | movl %eax, (%edx) | |
1974 | movl 4(%ecx), %eax | |
1975 | movl %eax, 4(%edx) | |
1976 | movl 7(%ecx), %eax | |
1977 | movl %eax, 7(%edx) | |
1978 | # ifdef USE_AS_STPCPY | |
1979 | lea 10(%edx), %eax | |
1980 | # else | |
1981 | movl %edi, %eax | |
1982 | # endif | |
1983 | RETURN | |
1984 | ||
1985 | .p2align 4 | |
1986 | L(Exit12): | |
1987 | movl (%ecx), %eax | |
1988 | movl %eax, (%edx) | |
1989 | movl 4(%ecx), %eax | |
1990 | movl %eax, 4(%edx) | |
1991 | movl 8(%ecx), %eax | |
1992 | movl %eax, 8(%edx) | |
1993 | # ifdef USE_AS_STPCPY | |
1994 | lea 11(%edx), %eax | |
1995 | # else | |
1996 | movl %edi, %eax | |
1997 | # endif | |
1998 | RETURN | |
1999 | ||
2000 | .p2align 4 | |
2001 | L(Exit13): | |
2002 | movlpd (%ecx), %xmm0 | |
2003 | movlpd %xmm0, (%edx) | |
2004 | movlpd 5(%ecx), %xmm0 | |
2005 | movlpd %xmm0, 5(%edx) | |
2006 | # ifdef USE_AS_STPCPY | |
2007 | lea 12(%edx), %eax | |
2008 | # else | |
2009 | movl %edi, %eax | |
2010 | # endif | |
2011 | RETURN | |
2012 | ||
2013 | .p2align 4 | |
2014 | L(Exit14): | |
2015 | movlpd (%ecx), %xmm0 | |
2016 | movlpd %xmm0, (%edx) | |
2017 | movlpd 6(%ecx), %xmm0 | |
2018 | movlpd %xmm0, 6(%edx) | |
2019 | # ifdef USE_AS_STPCPY | |
2020 | lea 13(%edx), %eax | |
2021 | # else | |
2022 | movl %edi, %eax | |
2023 | # endif | |
2024 | RETURN | |
2025 | ||
2026 | .p2align 4 | |
2027 | L(Exit15): | |
2028 | movlpd (%ecx), %xmm0 | |
2029 | movlpd %xmm0, (%edx) | |
2030 | movlpd 7(%ecx), %xmm0 | |
2031 | movlpd %xmm0, 7(%edx) | |
2032 | # ifdef USE_AS_STPCPY | |
2033 | lea 14(%edx), %eax | |
2034 | # else | |
2035 | movl %edi, %eax | |
2036 | # endif | |
2037 | RETURN | |
2038 | ||
2039 | CFI_POP (%edi) | |
2040 | ||
2041 | .p2align 4 | |
2042 | L(ExitTail1): | |
2043 | movb (%ecx), %al | |
2044 | movb %al, (%edx) | |
2045 | movl %edx, %eax | |
2046 | RETURN1 | |
2047 | ||
2048 | .p2align 4 | |
2049 | L(ExitTail2): | |
2050 | movw (%ecx), %ax | |
2051 | movw %ax, (%edx) | |
2052 | # ifdef USE_AS_STPCPY | |
2053 | lea 1(%edx), %eax | |
2054 | # else | |
2055 | movl %edx, %eax | |
2056 | # endif | |
2057 | RETURN1 | |
2058 | ||
2059 | .p2align 4 | |
2060 | L(ExitTail3): | |
2061 | movw (%ecx), %ax | |
2062 | movw %ax, (%edx) | |
2063 | movb 2(%ecx), %al | |
2064 | movb %al, 2(%edx) | |
2065 | # ifdef USE_AS_STPCPY | |
2066 | lea 2(%edx), %eax | |
2067 | # else | |
2068 | movl %edx, %eax | |
2069 | # endif | |
2070 | RETURN1 | |
2071 | ||
2072 | .p2align 4 | |
2073 | L(ExitTail4): | |
2074 | movl (%ecx), %eax | |
2075 | movl %eax, (%edx) | |
2076 | # ifdef USE_AS_STPCPY | |
2077 | lea 3(%edx), %eax | |
2078 | # else | |
2079 | movl %edx, %eax | |
2080 | # endif | |
2081 | RETURN1 | |
2082 | ||
2083 | .p2align 4 | |
2084 | L(ExitTail5): | |
2085 | movl (%ecx), %eax | |
2086 | movl %eax, (%edx) | |
2087 | movb 4(%ecx), %al | |
2088 | movb %al, 4(%edx) | |
2089 | # ifdef USE_AS_STPCPY | |
2090 | lea 4(%edx), %eax | |
2091 | # else | |
2092 | movl %edx, %eax | |
2093 | # endif | |
2094 | RETURN1 | |
2095 | ||
2096 | .p2align 4 | |
2097 | L(ExitTail6): | |
2098 | movl (%ecx), %eax | |
2099 | movl %eax, (%edx) | |
2100 | movw 4(%ecx), %ax | |
2101 | movw %ax, 4(%edx) | |
2102 | # ifdef USE_AS_STPCPY | |
2103 | lea 5(%edx), %eax | |
2104 | # else | |
2105 | movl %edx, %eax | |
2106 | # endif | |
2107 | RETURN1 | |
2108 | ||
2109 | .p2align 4 | |
2110 | L(ExitTail7): | |
2111 | movl (%ecx), %eax | |
2112 | movl %eax, (%edx) | |
2113 | movl 3(%ecx), %eax | |
2114 | movl %eax, 3(%edx) | |
2115 | # ifdef USE_AS_STPCPY | |
2116 | lea 6(%edx), %eax | |
2117 | # else | |
2118 | movl %edx, %eax | |
2119 | # endif | |
2120 | RETURN1 | |
2121 | ||
2122 | .p2align 4 | |
2123 | L(ExitTail8): | |
2124 | movl (%ecx), %eax | |
2125 | movl %eax, (%edx) | |
2126 | movl 4(%ecx), %eax | |
2127 | movl %eax, 4(%edx) | |
2128 | # ifdef USE_AS_STPCPY | |
2129 | lea 7(%edx), %eax | |
2130 | # else | |
2131 | movl %edx, %eax | |
2132 | # endif | |
2133 | RETURN1 | |
2134 | ||
2135 | .p2align 4 | |
2136 | L(ExitTail9): | |
2137 | movl (%ecx), %eax | |
2138 | movl %eax, (%edx) | |
2139 | movl 4(%ecx), %eax | |
2140 | movl %eax, 4(%edx) | |
2141 | movb 8(%ecx), %al | |
2142 | movb %al, 8(%edx) | |
2143 | # ifdef USE_AS_STPCPY | |
2144 | lea 8(%edx), %eax | |
2145 | # else | |
2146 | movl %edx, %eax | |
2147 | # endif | |
2148 | RETURN1 | |
2149 | ||
2150 | .p2align 4 | |
2151 | L(ExitTail10): | |
2152 | movl (%ecx), %eax | |
2153 | movl %eax, (%edx) | |
2154 | movl 4(%ecx), %eax | |
2155 | movl %eax, 4(%edx) | |
2156 | movw 8(%ecx), %ax | |
2157 | movw %ax, 8(%edx) | |
2158 | # ifdef USE_AS_STPCPY | |
2159 | lea 9(%edx), %eax | |
2160 | # else | |
2161 | movl %edx, %eax | |
2162 | # endif | |
2163 | RETURN1 | |
2164 | ||
2165 | .p2align 4 | |
2166 | L(ExitTail11): | |
2167 | movl (%ecx), %eax | |
2168 | movl %eax, (%edx) | |
2169 | movl 4(%ecx), %eax | |
2170 | movl %eax, 4(%edx) | |
2171 | movl 7(%ecx), %eax | |
2172 | movl %eax, 7(%edx) | |
2173 | # ifdef USE_AS_STPCPY | |
2174 | lea 10(%edx), %eax | |
2175 | # else | |
2176 | movl %edx, %eax | |
2177 | # endif | |
2178 | RETURN1 | |
2179 | ||
2180 | .p2align 4 | |
2181 | L(ExitTail12): | |
2182 | movl (%ecx), %eax | |
2183 | movl %eax, (%edx) | |
2184 | movl 4(%ecx), %eax | |
2185 | movl %eax, 4(%edx) | |
2186 | movl 8(%ecx), %eax | |
2187 | movl %eax, 8(%edx) | |
2188 | # ifdef USE_AS_STPCPY | |
2189 | lea 11(%edx), %eax | |
2190 | # else | |
2191 | movl %edx, %eax | |
2192 | # endif | |
2193 | RETURN1 | |
2194 | ||
2195 | .p2align 4 | |
2196 | L(ExitTail13): | |
2197 | movlpd (%ecx), %xmm0 | |
2198 | movlpd %xmm0, (%edx) | |
2199 | movlpd 5(%ecx), %xmm0 | |
2200 | movlpd %xmm0, 5(%edx) | |
2201 | # ifdef USE_AS_STPCPY | |
2202 | lea 12(%edx), %eax | |
2203 | # else | |
2204 | movl %edx, %eax | |
2205 | # endif | |
2206 | RETURN1 | |
2207 | ||
2208 | .p2align 4 | |
2209 | L(ExitTail14): | |
2210 | movlpd (%ecx), %xmm0 | |
2211 | movlpd %xmm0, (%edx) | |
2212 | movlpd 6(%ecx), %xmm0 | |
2213 | movlpd %xmm0, 6(%edx) | |
2214 | # ifdef USE_AS_STPCPY | |
2215 | lea 13(%edx), %eax | |
2216 | # else | |
2217 | movl %edx, %eax | |
2218 | # endif | |
2219 | RETURN1 | |
2220 | ||
2221 | .p2align 4 | |
2222 | L(ExitTail15): | |
2223 | movlpd (%ecx), %xmm0 | |
2224 | movlpd %xmm0, (%edx) | |
2225 | movlpd 7(%ecx), %xmm0 | |
2226 | movlpd %xmm0, 7(%edx) | |
2227 | # ifdef USE_AS_STPCPY | |
2228 | lea 14(%edx), %eax | |
2229 | # else | |
2230 | movl %edx, %eax | |
2231 | # endif | |
2232 | RETURN1 | |
2233 | ||
2234 | .p2align 4 | |
2235 | L(ExitTail16): | |
2236 | movlpd (%ecx), %xmm0 | |
2237 | movlpd %xmm0, (%edx) | |
2238 | movlpd 8(%ecx), %xmm0 | |
2239 | movlpd %xmm0, 8(%edx) | |
2240 | # ifdef USE_AS_STPCPY | |
2241 | lea 15(%edx), %eax | |
2242 | # else | |
2243 | movl %edx, %eax | |
2244 | # endif | |
2245 | RETURN1 | |
2246 | ||
2247 | END (STRCPY) | |
2248 | # endif | |
2249 | ||
2250 | #endif |