]>
Commit | Line | Data |
---|---|---|
904057bc L |
1 | /* memcmp with SSE4.2 |
2 | Copyright (C) 2010 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, write to the Free | |
18 | Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA | |
19 | 02111-1307 USA. */ | |
20 | ||
21 | #ifndef NOT_IN_libc | |
22 | ||
23 | #include <sysdep.h> | |
24 | #include "asm-syntax.h" | |
25 | ||
26 | #ifndef MEMCMP | |
27 | # define MEMCMP __memcmp_sse4_2 | |
28 | #endif | |
29 | ||
30 | #define CFI_PUSH(REG) \ | |
31 | cfi_adjust_cfa_offset (4); \ | |
32 | cfi_rel_offset (REG, 0) | |
33 | ||
34 | #define CFI_POP(REG) \ | |
35 | cfi_adjust_cfa_offset (-4); \ | |
36 | cfi_restore (REG) | |
37 | ||
38 | #define PUSH(REG) pushl REG; CFI_PUSH (REG) | |
39 | #define POP(REG) popl REG; CFI_POP (REG) | |
40 | ||
41 | #define PARMS 4 | |
42 | #define BLK1 PARMS | |
43 | #define BLK2 BLK1+4 | |
44 | #define LEN BLK2+4 | |
45 | #define RETURN POP (%ebx); ret; CFI_PUSH (%ebx) | |
46 | ||
47 | ||
48 | #ifdef SHARED | |
49 | # define JMPTBL(I, B) I - B | |
50 | ||
51 | /* Load an entry in a jump table into EBX and branch to it. TABLE is a | |
52 | jump table with relative offsets. INDEX is a register contains the | |
53 | index into the jump table. SCALE is the scale of INDEX. */ | |
54 | # define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \ | |
55 | /* We first load PC into EBX. */ \ | |
56 | call __i686.get_pc_thunk.bx; \ | |
57 | /* Get the address of the jump table. */ \ | |
58 | addl $(TABLE - .), %ebx; \ | |
59 | /* Get the entry and convert the relative offset to the \ | |
60 | absolute address. */ \ | |
61 | addl (%ebx,INDEX,SCALE), %ebx; \ | |
62 | /* We loaded the jump table and adjuested EDX/ESI. Go. */ \ | |
63 | jmp *%ebx | |
64 | ||
65 | .section .gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits | |
66 | .globl __i686.get_pc_thunk.bx | |
67 | .hidden __i686.get_pc_thunk.bx | |
68 | ALIGN (4) | |
69 | .type __i686.get_pc_thunk.bx,@function | |
70 | __i686.get_pc_thunk.bx: | |
71 | movl (%esp), %ebx | |
72 | ret | |
73 | #else | |
74 | # define JMPTBL(I, B) I | |
75 | ||
76 | /* Load an entry in a jump table into EBX and branch to it. TABLE is a | |
77 | jump table with relative offsets. INDEX is a register contains the | |
78 | index into the jump table. SCALE is the scale of INDEX. */ | |
79 | # define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \ | |
80 | jmp *TABLE(,INDEX,SCALE) | |
81 | #endif | |
82 | ||
83 | .section .text.sse4.2,"ax",@progbits | |
84 | ENTRY (MEMCMP) | |
85 | movl BLK1(%esp), %eax | |
86 | movl BLK2(%esp), %edx | |
87 | movl LEN(%esp), %ecx | |
88 | cmp $1, %ecx | |
89 | jbe L(less1bytes) | |
90 | pxor %xmm0, %xmm0 | |
91 | cmp $64, %ecx | |
92 | ja L(64bytesormore) | |
93 | cmp $8, %ecx | |
94 | PUSH (%ebx) | |
95 | jb L(less8bytes) | |
96 | add %ecx, %edx | |
97 | add %ecx, %eax | |
98 | BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %ecx, 4) | |
99 | L(less8bytes): | |
100 | ||
101 | mov (%eax), %bl | |
102 | cmpb (%edx), %bl | |
103 | jne L(nonzero) | |
104 | ||
105 | mov 1(%eax), %bl | |
106 | cmpb 1(%edx), %bl | |
107 | jne L(nonzero) | |
6bb74d9f UD |
108 | |
109 | cmp $2, %ecx | |
904057bc L |
110 | jz L(0bytes) |
111 | ||
112 | mov 2(%eax), %bl | |
113 | cmpb 2(%edx), %bl | |
114 | jne L(nonzero) | |
6bb74d9f UD |
115 | |
116 | cmp $3, %ecx | |
904057bc | 117 | jz L(0bytes) |
6bb74d9f | 118 | |
904057bc L |
119 | mov 3(%eax), %bl |
120 | cmpb 3(%edx), %bl | |
121 | jne L(nonzero) | |
6bb74d9f UD |
122 | |
123 | cmp $4, %ecx | |
904057bc | 124 | jz L(0bytes) |
6bb74d9f | 125 | |
904057bc L |
126 | mov 4(%eax), %bl |
127 | cmpb 4(%edx), %bl | |
128 | jne L(nonzero) | |
129 | ||
6bb74d9f | 130 | cmp $5, %ecx |
904057bc | 131 | jz L(0bytes) |
6bb74d9f | 132 | |
904057bc L |
133 | mov 5(%eax), %bl |
134 | cmpb 5(%edx), %bl | |
135 | jne L(nonzero) | |
136 | ||
6bb74d9f | 137 | cmp $6, %ecx |
904057bc | 138 | jz L(0bytes) |
6bb74d9f | 139 | |
904057bc L |
140 | mov 6(%eax), %bl |
141 | cmpb 6(%edx), %bl | |
142 | je L(0bytes) | |
143 | L(nonzero): | |
6bb74d9f | 144 | POP (%ebx) |
904057bc L |
145 | mov $1, %eax |
146 | ja L(above) | |
147 | neg %eax | |
148 | L(above): | |
149 | ret | |
150 | CFI_PUSH (%ebx) | |
151 | ||
152 | ALIGN (4) | |
153 | L(0bytes): | |
6bb74d9f | 154 | POP (%ebx) |
904057bc L |
155 | xor %eax, %eax |
156 | ret | |
157 | CFI_PUSH (%ebx) | |
6bb74d9f | 158 | |
904057bc L |
159 | ALIGN (4) |
160 | L(less1bytes): | |
161 | jb L(0bytesend) | |
162 | movzbl (%eax), %eax | |
163 | movzbl (%edx), %edx | |
164 | sub %edx, %eax | |
165 | ret | |
166 | ||
167 | ALIGN (4) | |
168 | L(0bytesend): | |
169 | xor %eax, %eax | |
170 | ret | |
171 | ||
172 | ALIGN (4) | |
173 | L(64bytesormore): | |
174 | PUSH (%ebx) | |
175 | mov %ecx, %ebx | |
176 | mov $64, %ecx | |
177 | sub $64, %ebx | |
178 | L(64bytesormore_loop): | |
179 | movdqu (%eax), %xmm1 | |
180 | movdqu (%edx), %xmm2 | |
181 | pxor %xmm1, %xmm2 | |
182 | ptest %xmm2, %xmm0 | |
183 | jnc L(find_16diff) | |
184 | ||
185 | movdqu 16(%eax), %xmm1 | |
186 | movdqu 16(%edx), %xmm2 | |
187 | pxor %xmm1, %xmm2 | |
188 | ptest %xmm2, %xmm0 | |
189 | jnc L(find_32diff) | |
190 | ||
191 | movdqu 32(%eax), %xmm1 | |
192 | movdqu 32(%edx), %xmm2 | |
193 | pxor %xmm1, %xmm2 | |
194 | ptest %xmm2, %xmm0 | |
195 | jnc L(find_48diff) | |
196 | ||
197 | movdqu 48(%eax), %xmm1 | |
198 | movdqu 48(%edx), %xmm2 | |
199 | pxor %xmm1, %xmm2 | |
200 | ptest %xmm2, %xmm0 | |
201 | jnc L(find_64diff) | |
202 | add %ecx, %eax | |
203 | add %ecx, %edx | |
204 | sub %ecx, %ebx | |
205 | jae L(64bytesormore_loop) | |
206 | add %ebx, %ecx | |
207 | add %ecx, %edx | |
208 | add %ecx, %eax | |
209 | BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %ecx, 4) | |
210 | L(find_16diff): | |
211 | sub $16, %ecx | |
212 | L(find_32diff): | |
213 | sub $16, %ecx | |
214 | L(find_48diff): | |
215 | sub $16, %ecx | |
216 | L(find_64diff): | |
217 | add %ecx, %edx | |
218 | add %ecx, %eax | |
219 | jmp L(16bytes) | |
220 | ALIGN (4) | |
221 | ||
222 | L(16bytes): | |
223 | mov -16(%eax), %ecx | |
224 | mov -16(%edx), %ebx | |
225 | cmp %ebx, %ecx | |
226 | jne L(find_diff) | |
227 | L(12bytes): | |
228 | mov -12(%eax), %ecx | |
229 | mov -12(%edx), %ebx | |
230 | cmp %ebx, %ecx | |
231 | jne L(find_diff) | |
232 | L(8bytes): | |
233 | mov -8(%eax), %ecx | |
234 | mov -8(%edx), %ebx | |
235 | cmp %ebx, %ecx | |
236 | jne L(find_diff) | |
237 | L(4bytes): | |
238 | mov -4(%eax), %ecx | |
239 | mov -4(%edx), %ebx | |
240 | cmp %ebx, %ecx | |
241 | mov $0, %eax | |
242 | jne L(find_diff) | |
243 | RETURN | |
244 | ||
245 | ALIGN (4) | |
246 | L(49bytes): | |
247 | movdqu -49(%eax), %xmm1 | |
248 | movdqu -49(%edx), %xmm2 | |
249 | mov $-49, %ebx | |
250 | pxor %xmm1, %xmm2 | |
251 | ptest %xmm2, %xmm0 | |
252 | jnc L(less16bytes) | |
253 | L(33bytes): | |
254 | movdqu -33(%eax), %xmm1 | |
255 | movdqu -33(%edx), %xmm2 | |
256 | mov $-33, %ebx | |
257 | pxor %xmm1, %xmm2 | |
258 | ptest %xmm2, %xmm0 | |
259 | jnc L(less16bytes) | |
260 | L(17bytes): | |
261 | mov -17(%eax), %ecx | |
262 | mov -17(%edx), %ebx | |
263 | cmp %ebx, %ecx | |
264 | jne L(find_diff) | |
265 | L(13bytes): | |
266 | mov -13(%eax), %ecx | |
267 | mov -13(%edx), %ebx | |
268 | cmp %ebx, %ecx | |
269 | jne L(find_diff) | |
270 | L(9bytes): | |
271 | mov -9(%eax), %ecx | |
272 | mov -9(%edx), %ebx | |
273 | cmp %ebx, %ecx | |
274 | jne L(find_diff) | |
275 | L(5bytes): | |
276 | mov -5(%eax), %ecx | |
277 | mov -5(%edx), %ebx | |
278 | cmp %ebx, %ecx | |
279 | jne L(find_diff) | |
280 | movzbl -1(%eax), %ecx | |
281 | cmp -1(%edx), %cl | |
282 | mov $0, %eax | |
283 | jne L(end) | |
284 | RETURN | |
285 | ||
286 | ALIGN (4) | |
287 | L(50bytes): | |
288 | mov $-50, %ebx | |
289 | movdqu -50(%eax), %xmm1 | |
290 | movdqu -50(%edx), %xmm2 | |
291 | pxor %xmm1, %xmm2 | |
292 | ptest %xmm2, %xmm0 | |
293 | jnc L(less16bytes) | |
294 | L(34bytes): | |
295 | mov $-34, %ebx | |
296 | movdqu -34(%eax), %xmm1 | |
297 | movdqu -34(%edx), %xmm2 | |
298 | pxor %xmm1, %xmm2 | |
299 | ptest %xmm2, %xmm0 | |
300 | jnc L(less16bytes) | |
301 | L(18bytes): | |
302 | mov -18(%eax), %ecx | |
303 | mov -18(%edx), %ebx | |
304 | cmp %ebx, %ecx | |
305 | jne L(find_diff) | |
306 | L(14bytes): | |
307 | mov -14(%eax), %ecx | |
308 | mov -14(%edx), %ebx | |
309 | cmp %ebx, %ecx | |
310 | jne L(find_diff) | |
311 | L(10bytes): | |
312 | mov -10(%eax), %ecx | |
313 | mov -10(%edx), %ebx | |
314 | cmp %ebx, %ecx | |
315 | jne L(find_diff) | |
316 | L(6bytes): | |
317 | mov -6(%eax), %ecx | |
318 | mov -6(%edx), %ebx | |
319 | cmp %ebx, %ecx | |
320 | jne L(find_diff) | |
321 | L(2bytes): | |
322 | movzwl -2(%eax), %ecx | |
323 | movzwl -2(%edx), %ebx | |
324 | cmp %bl, %cl | |
325 | jne L(end) | |
326 | cmp %bh, %ch | |
327 | mov $0, %eax | |
328 | jne L(end) | |
329 | RETURN | |
330 | ||
331 | ALIGN (4) | |
332 | L(51bytes): | |
333 | mov $-51, %ebx | |
334 | movdqu -51(%eax), %xmm1 | |
335 | movdqu -51(%edx), %xmm2 | |
336 | pxor %xmm1, %xmm2 | |
337 | ptest %xmm2, %xmm0 | |
338 | jnc L(less16bytes) | |
339 | L(35bytes): | |
340 | mov $-35, %ebx | |
341 | movdqu -35(%eax), %xmm1 | |
342 | movdqu -35(%edx), %xmm2 | |
343 | pxor %xmm1, %xmm2 | |
344 | ptest %xmm2, %xmm0 | |
345 | jnc L(less16bytes) | |
346 | L(19bytes): | |
347 | movl -19(%eax), %ecx | |
348 | movl -19(%edx), %ebx | |
349 | cmp %ebx, %ecx | |
350 | jne L(find_diff) | |
351 | L(15bytes): | |
352 | movl -15(%eax), %ecx | |
353 | movl -15(%edx), %ebx | |
354 | cmp %ebx, %ecx | |
355 | jne L(find_diff) | |
356 | L(11bytes): | |
357 | movl -11(%eax), %ecx | |
358 | movl -11(%edx), %ebx | |
359 | cmp %ebx, %ecx | |
360 | jne L(find_diff) | |
361 | L(7bytes): | |
362 | movl -7(%eax), %ecx | |
363 | movl -7(%edx), %ebx | |
364 | cmp %ebx, %ecx | |
365 | jne L(find_diff) | |
366 | L(3bytes): | |
367 | movzwl -3(%eax), %ecx | |
368 | movzwl -3(%edx), %ebx | |
369 | cmpb %bl, %cl | |
370 | jne L(end) | |
371 | cmp %bx, %cx | |
372 | jne L(end) | |
373 | L(1bytes): | |
374 | movzbl -1(%eax), %eax | |
375 | cmpb -1(%edx), %al | |
376 | mov $0, %eax | |
377 | jne L(end) | |
378 | RETURN | |
379 | ||
380 | ||
381 | L(52bytes): | |
382 | movdqu -52(%eax), %xmm1 | |
383 | movdqu -52(%edx), %xmm2 | |
384 | mov $-52, %ebx | |
385 | pxor %xmm1, %xmm2 | |
386 | ptest %xmm2, %xmm0 | |
387 | jnc L(less16bytes) | |
388 | L(36bytes): | |
389 | movdqu -36(%eax), %xmm1 | |
390 | movdqu -36(%edx), %xmm2 | |
391 | mov $-36, %ebx | |
392 | pxor %xmm1, %xmm2 | |
393 | ptest %xmm2, %xmm0 | |
394 | jnc L(less16bytes) | |
395 | L(20bytes): | |
396 | movdqu -20(%eax), %xmm1 | |
397 | movdqu -20(%edx), %xmm2 | |
398 | mov $-20, %ebx | |
399 | pxor %xmm1, %xmm2 | |
400 | ptest %xmm2, %xmm0 | |
401 | jnc L(less16bytes) | |
402 | mov -4(%eax), %ecx | |
403 | mov -4(%edx), %ebx | |
404 | cmp %ebx, %ecx | |
405 | mov $0, %eax | |
406 | jne L(find_diff) | |
407 | RETURN | |
408 | ||
409 | L(53bytes): | |
410 | movdqu -53(%eax), %xmm1 | |
411 | movdqu -53(%edx), %xmm2 | |
412 | mov $-53, %ebx | |
413 | pxor %xmm1, %xmm2 | |
414 | ptest %xmm2, %xmm0 | |
415 | jnc L(less16bytes) | |
416 | L(37bytes): | |
417 | mov $-37, %ebx | |
418 | movdqu -37(%eax), %xmm1 | |
419 | movdqu -37(%edx), %xmm2 | |
420 | pxor %xmm1, %xmm2 | |
421 | ptest %xmm2, %xmm0 | |
422 | jnc L(less16bytes) | |
423 | L(21bytes): | |
424 | mov $-21, %ebx | |
425 | movdqu -21(%eax), %xmm1 | |
426 | movdqu -21(%edx), %xmm2 | |
427 | pxor %xmm1, %xmm2 | |
428 | ptest %xmm2, %xmm0 | |
429 | jnc L(less16bytes) | |
430 | mov -5(%eax), %ecx | |
431 | mov -5(%edx), %ebx | |
432 | cmp %ebx, %ecx | |
433 | jne L(find_diff) | |
434 | movzbl -1(%eax), %ecx | |
435 | cmp -1(%edx), %cl | |
436 | mov $0, %eax | |
437 | jne L(end) | |
438 | RETURN | |
439 | ||
440 | L(54bytes): | |
441 | movdqu -54(%eax), %xmm1 | |
442 | movdqu -54(%edx), %xmm2 | |
443 | mov $-54, %ebx | |
444 | pxor %xmm1, %xmm2 | |
445 | ptest %xmm2, %xmm0 | |
446 | jnc L(less16bytes) | |
447 | L(38bytes): | |
448 | mov $-38, %ebx | |
449 | movdqu -38(%eax), %xmm1 | |
450 | movdqu -38(%edx), %xmm2 | |
451 | pxor %xmm1, %xmm2 | |
452 | ptest %xmm2, %xmm0 | |
453 | jnc L(less16bytes) | |
454 | L(22bytes): | |
455 | mov $-22, %ebx | |
456 | movdqu -22(%eax), %xmm1 | |
457 | movdqu -22(%edx), %xmm2 | |
458 | pxor %xmm1, %xmm2 | |
459 | ptest %xmm2, %xmm0 | |
460 | jnc L(less16bytes) | |
461 | ||
462 | mov -6(%eax), %ecx | |
463 | mov -6(%edx), %ebx | |
464 | cmp %ebx, %ecx | |
465 | jne L(find_diff) | |
466 | movzwl -2(%eax), %ecx | |
467 | movzwl -2(%edx), %ebx | |
468 | cmp %bl, %cl | |
469 | jne L(end) | |
470 | cmp %bh, %ch | |
471 | mov $0, %eax | |
472 | jne L(end) | |
473 | RETURN | |
474 | ||
475 | L(55bytes): | |
476 | movdqu -55(%eax), %xmm1 | |
477 | movdqu -55(%edx), %xmm2 | |
478 | mov $-55, %ebx | |
479 | pxor %xmm1, %xmm2 | |
480 | ptest %xmm2, %xmm0 | |
481 | jnc L(less16bytes) | |
482 | L(39bytes): | |
483 | mov $-39, %ebx | |
484 | movdqu -39(%eax), %xmm1 | |
485 | movdqu -39(%edx), %xmm2 | |
486 | pxor %xmm1, %xmm2 | |
487 | ptest %xmm2, %xmm0 | |
488 | jnc L(less16bytes) | |
489 | L(23bytes): | |
490 | mov $-23, %ebx | |
491 | movdqu -23(%eax), %xmm1 | |
492 | movdqu -23(%edx), %xmm2 | |
493 | pxor %xmm1, %xmm2 | |
494 | ptest %xmm2, %xmm0 | |
495 | jnc L(less16bytes) | |
496 | movl -7(%eax), %ecx | |
497 | movl -7(%edx), %ebx | |
498 | cmp %ebx, %ecx | |
499 | jne L(find_diff) | |
500 | movzwl -3(%eax), %ecx | |
501 | movzwl -3(%edx), %ebx | |
502 | cmpb %bl, %cl | |
503 | jne L(end) | |
504 | cmp %bx, %cx | |
505 | jne L(end) | |
506 | movzbl -1(%eax), %eax | |
507 | cmpb -1(%edx), %al | |
508 | mov $0, %eax | |
509 | jne L(end) | |
510 | RETURN | |
511 | ||
512 | L(56bytes): | |
513 | movdqu -56(%eax), %xmm1 | |
514 | movdqu -56(%edx), %xmm2 | |
515 | mov $-56, %ebx | |
516 | pxor %xmm1, %xmm2 | |
517 | ptest %xmm2, %xmm0 | |
518 | jnc L(less16bytes) | |
519 | L(40bytes): | |
520 | mov $-40, %ebx | |
521 | movdqu -40(%eax), %xmm1 | |
522 | movdqu -40(%edx), %xmm2 | |
523 | pxor %xmm1, %xmm2 | |
524 | ptest %xmm2, %xmm0 | |
525 | jnc L(less16bytes) | |
526 | L(24bytes): | |
527 | mov $-24, %ebx | |
528 | movdqu -24(%eax), %xmm1 | |
529 | movdqu -24(%edx), %xmm2 | |
530 | pxor %xmm1, %xmm2 | |
531 | ptest %xmm2, %xmm0 | |
532 | jnc L(less16bytes) | |
533 | ||
534 | mov -8(%eax), %ecx | |
535 | mov -8(%edx), %ebx | |
536 | cmp %ebx, %ecx | |
537 | jne L(find_diff) | |
538 | ||
539 | mov -4(%eax), %ecx | |
540 | mov -4(%edx), %ebx | |
541 | cmp %ebx, %ecx | |
542 | mov $0, %eax | |
543 | jne L(find_diff) | |
544 | RETURN | |
545 | ||
546 | L(57bytes): | |
547 | movdqu -57(%eax), %xmm1 | |
548 | movdqu -57(%edx), %xmm2 | |
549 | mov $-57, %ebx | |
550 | pxor %xmm1, %xmm2 | |
551 | ptest %xmm2, %xmm0 | |
552 | jnc L(less16bytes) | |
553 | L(41bytes): | |
554 | mov $-41, %ebx | |
555 | movdqu -41(%eax), %xmm1 | |
556 | movdqu -41(%edx), %xmm2 | |
557 | pxor %xmm1, %xmm2 | |
558 | ptest %xmm2, %xmm0 | |
559 | jnc L(less16bytes) | |
560 | L(25bytes): | |
561 | mov $-25, %ebx | |
562 | movdqu -25(%eax), %xmm1 | |
563 | movdqu -25(%edx), %xmm2 | |
564 | pxor %xmm1, %xmm2 | |
565 | ptest %xmm2, %xmm0 | |
566 | jnc L(less16bytes) | |
567 | mov -9(%eax), %ecx | |
568 | mov -9(%edx), %ebx | |
569 | cmp %ebx, %ecx | |
570 | jne L(find_diff) | |
571 | mov -5(%eax), %ecx | |
572 | mov -5(%edx), %ebx | |
573 | cmp %ebx, %ecx | |
574 | jne L(find_diff) | |
575 | movzbl -1(%eax), %ecx | |
576 | cmp -1(%edx), %cl | |
577 | mov $0, %eax | |
578 | jne L(end) | |
579 | RETURN | |
580 | ||
581 | L(58bytes): | |
582 | movdqu -58(%eax), %xmm1 | |
583 | movdqu -58(%edx), %xmm2 | |
584 | mov $-58, %ebx | |
585 | pxor %xmm1, %xmm2 | |
586 | ptest %xmm2, %xmm0 | |
587 | jnc L(less16bytes) | |
588 | L(42bytes): | |
589 | mov $-42, %ebx | |
590 | movdqu -42(%eax), %xmm1 | |
591 | movdqu -42(%edx), %xmm2 | |
592 | pxor %xmm1, %xmm2 | |
593 | ptest %xmm2, %xmm0 | |
594 | jnc L(less16bytes) | |
595 | L(26bytes): | |
596 | mov $-26, %ebx | |
597 | movdqu -26(%eax), %xmm1 | |
598 | movdqu -26(%edx), %xmm2 | |
599 | pxor %xmm1, %xmm2 | |
600 | ptest %xmm2, %xmm0 | |
601 | jnc L(less16bytes) | |
602 | ||
603 | mov -10(%eax), %ecx | |
604 | mov -10(%edx), %ebx | |
605 | cmp %ebx, %ecx | |
606 | jne L(find_diff) | |
607 | ||
608 | mov -6(%eax), %ecx | |
609 | mov -6(%edx), %ebx | |
610 | cmp %ebx, %ecx | |
611 | jne L(find_diff) | |
6bb74d9f | 612 | |
904057bc L |
613 | movzwl -2(%eax), %ecx |
614 | movzwl -2(%edx), %ebx | |
615 | cmp %bl, %cl | |
616 | jne L(end) | |
617 | cmp %bh, %ch | |
618 | mov $0, %eax | |
619 | jne L(end) | |
620 | RETURN | |
621 | ||
622 | L(59bytes): | |
623 | movdqu -59(%eax), %xmm1 | |
624 | movdqu -59(%edx), %xmm2 | |
625 | mov $-59, %ebx | |
626 | pxor %xmm1, %xmm2 | |
627 | ptest %xmm2, %xmm0 | |
628 | jnc L(less16bytes) | |
629 | L(43bytes): | |
630 | mov $-43, %ebx | |
631 | movdqu -43(%eax), %xmm1 | |
632 | movdqu -43(%edx), %xmm2 | |
633 | pxor %xmm1, %xmm2 | |
634 | ptest %xmm2, %xmm0 | |
635 | jnc L(less16bytes) | |
636 | L(27bytes): | |
637 | mov $-27, %ebx | |
638 | movdqu -27(%eax), %xmm1 | |
639 | movdqu -27(%edx), %xmm2 | |
640 | pxor %xmm1, %xmm2 | |
641 | ptest %xmm2, %xmm0 | |
642 | jnc L(less16bytes) | |
643 | movl -11(%eax), %ecx | |
644 | movl -11(%edx), %ebx | |
645 | cmp %ebx, %ecx | |
646 | jne L(find_diff) | |
647 | movl -7(%eax), %ecx | |
648 | movl -7(%edx), %ebx | |
649 | cmp %ebx, %ecx | |
650 | jne L(find_diff) | |
651 | movzwl -3(%eax), %ecx | |
652 | movzwl -3(%edx), %ebx | |
653 | cmpb %bl, %cl | |
654 | jne L(end) | |
655 | cmp %bx, %cx | |
656 | jne L(end) | |
657 | movzbl -1(%eax), %eax | |
658 | cmpb -1(%edx), %al | |
659 | mov $0, %eax | |
660 | jne L(end) | |
661 | RETURN | |
662 | ||
663 | L(60bytes): | |
664 | movdqu -60(%eax), %xmm1 | |
665 | movdqu -60(%edx), %xmm2 | |
666 | mov $-60, %ebx | |
667 | pxor %xmm1, %xmm2 | |
668 | ptest %xmm2, %xmm0 | |
669 | jnc L(less16bytes) | |
670 | L(44bytes): | |
671 | mov $-44, %ebx | |
672 | movdqu -44(%eax), %xmm1 | |
673 | movdqu -44(%edx), %xmm2 | |
674 | pxor %xmm1, %xmm2 | |
675 | ptest %xmm2, %xmm0 | |
676 | jnc L(less16bytes) | |
677 | L(28bytes): | |
678 | mov $-28, %ebx | |
679 | movdqu -28(%eax), %xmm1 | |
680 | movdqu -28(%edx), %xmm2 | |
681 | pxor %xmm1, %xmm2 | |
682 | ptest %xmm2, %xmm0 | |
683 | jnc L(less16bytes) | |
684 | mov -12(%eax), %ecx | |
685 | mov -12(%edx), %ebx | |
686 | cmp %ebx, %ecx | |
687 | jne L(find_diff) | |
688 | mov -8(%eax), %ecx | |
689 | mov -8(%edx), %ebx | |
690 | cmp %ebx, %ecx | |
691 | jne L(find_diff) | |
692 | mov -4(%eax), %ecx | |
693 | mov -4(%edx), %ebx | |
694 | cmp %ebx, %ecx | |
695 | mov $0, %eax | |
696 | jne L(find_diff) | |
697 | RETURN | |
698 | ||
699 | L(61bytes): | |
700 | movdqu -61(%eax), %xmm1 | |
701 | movdqu -61(%edx), %xmm2 | |
702 | mov $-61, %ebx | |
703 | pxor %xmm1, %xmm2 | |
704 | ptest %xmm2, %xmm0 | |
705 | jnc L(less16bytes) | |
706 | L(45bytes): | |
707 | mov $-45, %ebx | |
708 | movdqu -45(%eax), %xmm1 | |
709 | movdqu -45(%edx), %xmm2 | |
710 | pxor %xmm1, %xmm2 | |
711 | ptest %xmm2, %xmm0 | |
712 | jnc L(less16bytes) | |
713 | L(29bytes): | |
714 | mov $-29, %ebx | |
715 | movdqu -29(%eax), %xmm1 | |
716 | movdqu -29(%edx), %xmm2 | |
717 | pxor %xmm1, %xmm2 | |
718 | ptest %xmm2, %xmm0 | |
719 | jnc L(less16bytes) | |
720 | ||
721 | mov -13(%eax), %ecx | |
722 | mov -13(%edx), %ebx | |
723 | cmp %ebx, %ecx | |
724 | jne L(find_diff) | |
725 | ||
726 | mov -9(%eax), %ecx | |
727 | mov -9(%edx), %ebx | |
728 | cmp %ebx, %ecx | |
729 | jne L(find_diff) | |
730 | ||
731 | mov -5(%eax), %ecx | |
732 | mov -5(%edx), %ebx | |
733 | cmp %ebx, %ecx | |
734 | jne L(find_diff) | |
735 | movzbl -1(%eax), %ecx | |
736 | cmp -1(%edx), %cl | |
737 | mov $0, %eax | |
738 | jne L(end) | |
739 | RETURN | |
740 | ||
741 | L(62bytes): | |
742 | movdqu -62(%eax), %xmm1 | |
743 | movdqu -62(%edx), %xmm2 | |
744 | mov $-62, %ebx | |
745 | pxor %xmm1, %xmm2 | |
746 | ptest %xmm2, %xmm0 | |
747 | jnc L(less16bytes) | |
748 | L(46bytes): | |
749 | mov $-46, %ebx | |
750 | movdqu -46(%eax), %xmm1 | |
751 | movdqu -46(%edx), %xmm2 | |
752 | pxor %xmm1, %xmm2 | |
753 | ptest %xmm2, %xmm0 | |
754 | jnc L(less16bytes) | |
755 | L(30bytes): | |
756 | mov $-30, %ebx | |
757 | movdqu -30(%eax), %xmm1 | |
758 | movdqu -30(%edx), %xmm2 | |
759 | pxor %xmm1, %xmm2 | |
760 | ptest %xmm2, %xmm0 | |
761 | jnc L(less16bytes) | |
762 | mov -14(%eax), %ecx | |
763 | mov -14(%edx), %ebx | |
764 | cmp %ebx, %ecx | |
765 | jne L(find_diff) | |
766 | mov -10(%eax), %ecx | |
767 | mov -10(%edx), %ebx | |
768 | cmp %ebx, %ecx | |
769 | jne L(find_diff) | |
770 | mov -6(%eax), %ecx | |
771 | mov -6(%edx), %ebx | |
772 | cmp %ebx, %ecx | |
773 | jne L(find_diff) | |
774 | movzwl -2(%eax), %ecx | |
775 | movzwl -2(%edx), %ebx | |
776 | cmp %bl, %cl | |
777 | jne L(end) | |
778 | cmp %bh, %ch | |
779 | mov $0, %eax | |
780 | jne L(end) | |
781 | RETURN | |
782 | ||
783 | L(63bytes): | |
784 | movdqu -63(%eax), %xmm1 | |
785 | movdqu -63(%edx), %xmm2 | |
786 | mov $-63, %ebx | |
787 | pxor %xmm1, %xmm2 | |
788 | ptest %xmm2, %xmm0 | |
789 | jnc L(less16bytes) | |
790 | L(47bytes): | |
791 | mov $-47, %ebx | |
792 | movdqu -47(%eax), %xmm1 | |
793 | movdqu -47(%edx), %xmm2 | |
794 | pxor %xmm1, %xmm2 | |
795 | ptest %xmm2, %xmm0 | |
796 | jnc L(less16bytes) | |
797 | L(31bytes): | |
798 | mov $-31, %ebx | |
799 | movdqu -31(%eax), %xmm1 | |
800 | movdqu -31(%edx), %xmm2 | |
801 | pxor %xmm1, %xmm2 | |
802 | ptest %xmm2, %xmm0 | |
803 | jnc L(less16bytes) | |
804 | ||
805 | movl -15(%eax), %ecx | |
806 | movl -15(%edx), %ebx | |
807 | cmp %ebx, %ecx | |
808 | jne L(find_diff) | |
809 | movl -11(%eax), %ecx | |
810 | movl -11(%edx), %ebx | |
811 | cmp %ebx, %ecx | |
812 | jne L(find_diff) | |
813 | movl -7(%eax), %ecx | |
814 | movl -7(%edx), %ebx | |
815 | cmp %ebx, %ecx | |
816 | jne L(find_diff) | |
817 | movzwl -3(%eax), %ecx | |
818 | movzwl -3(%edx), %ebx | |
819 | cmpb %bl, %cl | |
820 | jne L(end) | |
821 | cmp %bx, %cx | |
822 | jne L(end) | |
823 | movzbl -1(%eax), %eax | |
824 | cmpb -1(%edx), %al | |
825 | mov $0, %eax | |
826 | jne L(end) | |
827 | RETURN | |
828 | ||
829 | L(64bytes): | |
830 | movdqu -64(%eax), %xmm1 | |
831 | movdqu -64(%edx), %xmm2 | |
832 | mov $-64, %ebx | |
833 | pxor %xmm1, %xmm2 | |
834 | ptest %xmm2, %xmm0 | |
835 | jnc L(less16bytes) | |
836 | L(48bytes): | |
837 | movdqu -48(%eax), %xmm1 | |
838 | movdqu -48(%edx), %xmm2 | |
839 | mov $-48, %ebx | |
840 | pxor %xmm1, %xmm2 | |
841 | ptest %xmm2, %xmm0 | |
842 | jnc L(less16bytes) | |
843 | L(32bytes): | |
844 | movdqu -32(%eax), %xmm1 | |
845 | movdqu -32(%edx), %xmm2 | |
846 | mov $-32, %ebx | |
847 | pxor %xmm1, %xmm2 | |
848 | ptest %xmm2, %xmm0 | |
849 | jnc L(less16bytes) | |
850 | ||
851 | mov -16(%eax), %ecx | |
852 | mov -16(%edx), %ebx | |
853 | cmp %ebx, %ecx | |
854 | jne L(find_diff) | |
855 | ||
856 | mov -12(%eax), %ecx | |
857 | mov -12(%edx), %ebx | |
858 | cmp %ebx, %ecx | |
859 | jne L(find_diff) | |
860 | ||
861 | mov -8(%eax), %ecx | |
862 | mov -8(%edx), %ebx | |
863 | cmp %ebx, %ecx | |
864 | jne L(find_diff) | |
865 | ||
866 | mov -4(%eax), %ecx | |
867 | mov -4(%edx), %ebx | |
868 | cmp %ebx, %ecx | |
869 | mov $0, %eax | |
870 | jne L(find_diff) | |
871 | RETURN | |
872 | ||
873 | L(less16bytes): | |
874 | add %ebx, %eax | |
875 | add %ebx, %edx | |
6bb74d9f | 876 | |
904057bc L |
877 | mov (%eax), %ecx |
878 | mov (%edx), %ebx | |
879 | cmp %ebx, %ecx | |
880 | jne L(find_diff) | |
881 | ||
882 | mov 4(%eax), %ecx | |
883 | mov 4(%edx), %ebx | |
884 | cmp %ebx, %ecx | |
885 | jne L(find_diff) | |
886 | ||
887 | mov 8(%eax), %ecx | |
888 | mov 8(%edx), %ebx | |
889 | cmp %ebx, %ecx | |
890 | jne L(find_diff) | |
891 | ||
892 | mov 12(%eax), %ecx | |
893 | mov 12(%edx), %ebx | |
894 | cmp %ebx, %ecx | |
895 | mov $0, %eax | |
896 | jne L(find_diff) | |
897 | RETURN | |
898 | ||
899 | ALIGN (4) | |
900 | L(find_diff): | |
901 | cmpb %bl, %cl | |
902 | jne L(end) | |
903 | cmp %bx, %cx | |
904 | jne L(end) | |
905 | shr $16,%ecx | |
906 | shr $16,%ebx | |
907 | cmp %bl, %cl | |
908 | jne L(end) | |
909 | cmp %bx, %cx | |
910 | L(end): | |
6bb74d9f | 911 | POP (%ebx) |
904057bc L |
912 | mov $1, %eax |
913 | ja L(bigger) | |
914 | neg %eax | |
915 | L(bigger): | |
916 | ret | |
917 | ||
918 | ALIGN (2) | |
919 | L(table_64bytes): | |
920 | .int JMPTBL (L(0bytes), L(table_64bytes)) | |
921 | .int JMPTBL (L(1bytes), L(table_64bytes)) | |
922 | .int JMPTBL (L(2bytes), L(table_64bytes)) | |
923 | .int JMPTBL (L(3bytes), L(table_64bytes)) | |
924 | .int JMPTBL (L(4bytes), L(table_64bytes)) | |
925 | .int JMPTBL (L(5bytes), L(table_64bytes)) | |
926 | .int JMPTBL (L(6bytes), L(table_64bytes)) | |
927 | .int JMPTBL (L(7bytes), L(table_64bytes)) | |
928 | .int JMPTBL (L(8bytes), L(table_64bytes)) | |
929 | .int JMPTBL (L(9bytes), L(table_64bytes)) | |
930 | .int JMPTBL (L(10bytes), L(table_64bytes)) | |
931 | .int JMPTBL (L(11bytes), L(table_64bytes)) | |
932 | .int JMPTBL (L(12bytes), L(table_64bytes)) | |
933 | .int JMPTBL (L(13bytes), L(table_64bytes)) | |
934 | .int JMPTBL (L(14bytes), L(table_64bytes)) | |
935 | .int JMPTBL (L(15bytes), L(table_64bytes)) | |
936 | .int JMPTBL (L(16bytes), L(table_64bytes)) | |
937 | .int JMPTBL (L(17bytes), L(table_64bytes)) | |
938 | .int JMPTBL (L(18bytes), L(table_64bytes)) | |
939 | .int JMPTBL (L(19bytes), L(table_64bytes)) | |
940 | .int JMPTBL (L(20bytes), L(table_64bytes)) | |
941 | .int JMPTBL (L(21bytes), L(table_64bytes)) | |
942 | .int JMPTBL (L(22bytes), L(table_64bytes)) | |
943 | .int JMPTBL (L(23bytes), L(table_64bytes)) | |
944 | .int JMPTBL (L(24bytes), L(table_64bytes)) | |
945 | .int JMPTBL (L(25bytes), L(table_64bytes)) | |
946 | .int JMPTBL (L(26bytes), L(table_64bytes)) | |
947 | .int JMPTBL (L(27bytes), L(table_64bytes)) | |
948 | .int JMPTBL (L(28bytes), L(table_64bytes)) | |
949 | .int JMPTBL (L(29bytes), L(table_64bytes)) | |
950 | .int JMPTBL (L(30bytes), L(table_64bytes)) | |
951 | .int JMPTBL (L(31bytes), L(table_64bytes)) | |
952 | .int JMPTBL (L(32bytes), L(table_64bytes)) | |
953 | .int JMPTBL (L(33bytes), L(table_64bytes)) | |
954 | .int JMPTBL (L(34bytes), L(table_64bytes)) | |
955 | .int JMPTBL (L(35bytes), L(table_64bytes)) | |
956 | .int JMPTBL (L(36bytes), L(table_64bytes)) | |
957 | .int JMPTBL (L(37bytes), L(table_64bytes)) | |
958 | .int JMPTBL (L(38bytes), L(table_64bytes)) | |
959 | .int JMPTBL (L(39bytes), L(table_64bytes)) | |
960 | .int JMPTBL (L(40bytes), L(table_64bytes)) | |
961 | .int JMPTBL (L(41bytes), L(table_64bytes)) | |
962 | .int JMPTBL (L(42bytes), L(table_64bytes)) | |
963 | .int JMPTBL (L(43bytes), L(table_64bytes)) | |
964 | .int JMPTBL (L(44bytes), L(table_64bytes)) | |
965 | .int JMPTBL (L(45bytes), L(table_64bytes)) | |
966 | .int JMPTBL (L(46bytes), L(table_64bytes)) | |
967 | .int JMPTBL (L(47bytes), L(table_64bytes)) | |
968 | .int JMPTBL (L(48bytes), L(table_64bytes)) | |
969 | .int JMPTBL (L(49bytes), L(table_64bytes)) | |
970 | .int JMPTBL (L(50bytes), L(table_64bytes)) | |
971 | .int JMPTBL (L(51bytes), L(table_64bytes)) | |
972 | .int JMPTBL (L(52bytes), L(table_64bytes)) | |
973 | .int JMPTBL (L(53bytes), L(table_64bytes)) | |
974 | .int JMPTBL (L(54bytes), L(table_64bytes)) | |
975 | .int JMPTBL (L(55bytes), L(table_64bytes)) | |
976 | .int JMPTBL (L(56bytes), L(table_64bytes)) | |
977 | .int JMPTBL (L(57bytes), L(table_64bytes)) | |
978 | .int JMPTBL (L(58bytes), L(table_64bytes)) | |
979 | .int JMPTBL (L(59bytes), L(table_64bytes)) | |
980 | .int JMPTBL (L(60bytes), L(table_64bytes)) | |
981 | .int JMPTBL (L(61bytes), L(table_64bytes)) | |
982 | .int JMPTBL (L(62bytes), L(table_64bytes)) | |
983 | .int JMPTBL (L(63bytes), L(table_64bytes)) | |
984 | .int JMPTBL (L(64bytes), L(table_64bytes)) | |
985 | ||
986 | END (MEMCMP) | |
987 | ||
988 | #endif |