]>
Commit | Line | Data |
---|---|---|
78df0fcb | 1 | /* Highly optimized version for x86-64. |
b168057a | 2 | Copyright (C) 1999-2015 Free Software Foundation, Inc. |
78df0fcb AJ |
3 | This file is part of the GNU C Library. |
4 | Based on i686 version contributed by Ulrich Drepper | |
5 | <drepper@cygnus.com>, 1999. | |
7956a3d2 | 6 | Updated with SSE2 support contributed by Intel Corporation. |
78df0fcb AJ |
7 | |
8 | The GNU C Library is free software; you can redistribute it and/or | |
9 | modify it under the terms of the GNU Lesser General Public | |
10 | License as published by the Free Software Foundation; either | |
11 | version 2.1 of the License, or (at your option) any later version. | |
12 | ||
13 | The GNU C Library is distributed in the hope that it will be useful, | |
14 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | Lesser General Public License for more details. | |
17 | ||
18 | You should have received a copy of the GNU Lesser General Public | |
59ba27a6 PE |
19 | License along with the GNU C Library; if not, see |
20 | <http://www.gnu.org/licenses/>. */ | |
78df0fcb AJ |
21 | |
22 | #include <sysdep.h> | |
23 | #include "asm-syntax.h" | |
78df0fcb | 24 | |
7956a3d2 L |
25 | #undef UPDATE_STRNCMP_COUNTER |
26 | ||
27 | #ifndef LABEL | |
28 | #define LABEL(l) L(l) | |
29 | #endif | |
30 | ||
31 | #ifdef USE_AS_STRNCMP | |
f69190e7 UD |
32 | /* The simplified code below is not set up to handle strncmp() so far. |
33 | Should this become necessary it has to be implemented. For now | |
34 | just report the problem. */ | |
4f41c682 | 35 | # if !IS_IN (libc) |
f69190e7 UD |
36 | # error "strncmp not implemented so far" |
37 | # endif | |
38 | ||
7956a3d2 L |
39 | /* Since the counter, %r11, is unsigned, we branch to strcmp_exitz |
40 | if the new counter > the old one or is 0. */ | |
41 | # define UPDATE_STRNCMP_COUNTER \ | |
42 | /* calculate left number to compare */ \ | |
43 | lea -16(%rcx, %r11), %r9; \ | |
44 | cmp %r9, %r11; \ | |
45 | jb LABEL(strcmp_exitz); \ | |
46 | test %r9, %r9; \ | |
47 | je LABEL(strcmp_exitz); \ | |
48 | mov %r9, %r11 | |
49 | ||
42e08a54 UD |
50 | #elif defined USE_AS_STRCASECMP_L |
51 | # include "locale-defines.h" | |
52 | ||
53 | /* No support for strcasecmp outside libc so far since it is not needed. */ | |
4f41c682 | 54 | # if !IS_IN (libc) |
42e08a54 UD |
55 | # error "strcasecmp_l not implemented so far" |
56 | # endif | |
57 | ||
58 | # define UPDATE_STRNCMP_COUNTER | |
e9f82e0d UD |
59 | #elif defined USE_AS_STRNCASECMP_L |
60 | # include "locale-defines.h" | |
61 | ||
62 | /* No support for strncasecmp outside libc so far since it is not needed. */ | |
4f41c682 | 63 | # if !IS_IN (libc) |
e9f82e0d UD |
64 | # error "strncasecmp_l not implemented so far" |
65 | # endif | |
66 | ||
67 | # define UPDATE_STRNCMP_COUNTER \ | |
68 | /* calculate left number to compare */ \ | |
69 | lea -16(%rcx, %r11), %r9; \ | |
70 | cmp %r9, %r11; \ | |
71 | jb LABEL(strcmp_exitz); \ | |
72 | test %r9, %r9; \ | |
73 | je LABEL(strcmp_exitz); \ | |
74 | mov %r9, %r11 | |
7956a3d2 L |
75 | #else |
76 | # define UPDATE_STRNCMP_COUNTER | |
77 | # ifndef STRCMP | |
25244f17 | 78 | # define STRCMP strcmp |
7956a3d2 L |
79 | # endif |
80 | #endif | |
81 | ||
0fda545d | 82 | #ifndef USE_SSSE3 |
7956a3d2 | 83 | .text |
0fda545d | 84 | #else |
f69190e7 | 85 | .section .text.ssse3,"ax",@progbits |
0fda545d UD |
86 | #endif |
87 | ||
42e08a54 | 88 | #ifdef USE_AS_STRCASECMP_L |
73507d3a UD |
89 | # ifndef ENTRY2 |
90 | # define ENTRY2(name) ENTRY (name) | |
91 | # define END2(name) END (name) | |
73507d3a UD |
92 | # endif |
93 | ||
e9f82e0d | 94 | ENTRY2 (__strcasecmp) |
42e08a54 | 95 | movq __libc_tsd_LOCALE@gottpoff(%rip),%rax |
3cb84561 | 96 | mov %fs:(%rax),%RDX_LP |
42e08a54 | 97 | |
73507d3a | 98 | // XXX 5 byte should be before the function |
42e08a54 UD |
99 | /* 5-byte NOP. */ |
100 | .byte 0x0f,0x1f,0x44,0x00,0x00 | |
73507d3a UD |
101 | END2 (__strcasecmp) |
102 | # ifndef NO_NOLOCALE_ALIAS | |
42e08a54 UD |
103 | weak_alias (__strcasecmp, strcasecmp) |
104 | libc_hidden_def (__strcasecmp) | |
73507d3a | 105 | # endif |
42e08a54 | 106 | /* FALLTHROUGH to strcasecmp_l. */ |
e9f82e0d UD |
107 | #elif defined USE_AS_STRNCASECMP_L |
108 | # ifndef ENTRY2 | |
109 | # define ENTRY2(name) ENTRY (name) | |
110 | # define END2(name) END (name) | |
111 | # endif | |
112 | ||
113 | ENTRY2 (__strncasecmp) | |
114 | movq __libc_tsd_LOCALE@gottpoff(%rip),%rax | |
3cb84561 | 115 | mov %fs:(%rax),%RCX_LP |
e9f82e0d UD |
116 | |
117 | // XXX 5 byte should be before the function | |
118 | /* 5-byte NOP. */ | |
119 | .byte 0x0f,0x1f,0x44,0x00,0x00 | |
120 | END2 (__strncasecmp) | |
121 | # ifndef NO_NOLOCALE_ALIAS | |
122 | weak_alias (__strncasecmp, strncasecmp) | |
123 | libc_hidden_def (__strncasecmp) | |
124 | # endif | |
125 | /* FALLTHROUGH to strncasecmp_l. */ | |
42e08a54 UD |
126 | #endif |
127 | ||
29691210 | 128 | ENTRY (STRCMP) |
4f41c682 | 129 | #if !IS_IN (libc) |
7956a3d2 | 130 | /* Simple version since we can't use SSE registers in ld.so. */ |
78df0fcb AJ |
131 | L(oop): movb (%rdi), %al |
132 | cmpb (%rsi), %al | |
133 | jne L(neq) | |
134 | incq %rdi | |
135 | incq %rsi | |
136 | testb %al, %al | |
137 | jnz L(oop) | |
138 | ||
ee618985 | 139 | xorl %eax, %eax |
78df0fcb AJ |
140 | ret |
141 | ||
142 | L(neq): movl $1, %eax | |
143 | movl $-1, %ecx | |
144 | cmovbl %ecx, %eax | |
145 | ret | |
29691210 | 146 | END (STRCMP) |
4f41c682 | 147 | #else /* !IS_IN (libc) */ |
42e08a54 UD |
148 | # ifdef USE_AS_STRCASECMP_L |
149 | /* We have to fall back on the C implementation for locales | |
150 | with encodings not matching ASCII for single bytes. */ | |
151 | # if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0 | |
3cb84561 | 152 | mov LOCALE_T___LOCALES+LC_CTYPE*LP_SIZE(%rdx), %RAX_LP |
42e08a54 | 153 | # else |
3cb84561 | 154 | mov (%rdx), %RAX_LP |
42e08a54 | 155 | # endif |
34372fc6 | 156 | testl $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%rax) |
42e08a54 | 157 | jne __strcasecmp_l_nonascii |
e9f82e0d UD |
158 | # elif defined USE_AS_STRNCASECMP_L |
159 | /* We have to fall back on the C implementation for locales | |
160 | with encodings not matching ASCII for single bytes. */ | |
161 | # if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0 | |
3cb84561 | 162 | mov LOCALE_T___LOCALES+LC_CTYPE*LP_SIZE(%rcx), %RAX_LP |
e9f82e0d | 163 | # else |
3cb84561 | 164 | mov (%rcx), %RAX_LP |
e9f82e0d | 165 | # endif |
34372fc6 | 166 | testl $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%rax) |
e9f82e0d | 167 | jne __strncasecmp_l_nonascii |
42e08a54 UD |
168 | # endif |
169 | ||
7956a3d2 L |
170 | /* |
171 | * This implementation uses SSE to compare up to 16 bytes at a time. | |
172 | */ | |
e9f82e0d | 173 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
174 | test %rdx, %rdx |
175 | je LABEL(strcmp_exitz) | |
176 | cmp $1, %rdx | |
177 | je LABEL(Byte0) | |
178 | mov %rdx, %r11 | |
f69190e7 | 179 | # endif |
7956a3d2 L |
180 | mov %esi, %ecx |
181 | mov %edi, %eax | |
182 | /* Use 64bit AND here to avoid long NOP padding. */ | |
183 | and $0x3f, %rcx /* rsi alignment in cache line */ | |
184 | and $0x3f, %rax /* rdi alignment in cache line */ | |
e9f82e0d | 185 | # if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
42e08a54 UD |
186 | .section .rodata.cst16,"aM",@progbits,16 |
187 | .align 16 | |
188 | .Lbelowupper: | |
189 | .quad 0x4040404040404040 | |
190 | .quad 0x4040404040404040 | |
191 | .Ltopupper: | |
192 | .quad 0x5b5b5b5b5b5b5b5b | |
193 | .quad 0x5b5b5b5b5b5b5b5b | |
194 | .Ltouppermask: | |
195 | .quad 0x2020202020202020 | |
196 | .quad 0x2020202020202020 | |
197 | .previous | |
198 | movdqa .Lbelowupper(%rip), %xmm5 | |
e9f82e0d | 199 | # define UCLOW_reg %xmm5 |
42e08a54 | 200 | movdqa .Ltopupper(%rip), %xmm6 |
e9f82e0d | 201 | # define UCHIGH_reg %xmm6 |
42e08a54 | 202 | movdqa .Ltouppermask(%rip), %xmm7 |
e9f82e0d | 203 | # define LCQWORD_reg %xmm7 |
42e08a54 | 204 | # endif |
7956a3d2 L |
205 | cmp $0x30, %ecx |
206 | ja LABEL(crosscache) /* rsi: 16-byte load will cross cache line */ | |
207 | cmp $0x30, %eax | |
208 | ja LABEL(crosscache) /* rdi: 16-byte load will cross cache line */ | |
209 | movlpd (%rdi), %xmm1 | |
210 | movlpd (%rsi), %xmm2 | |
211 | movhpd 8(%rdi), %xmm1 | |
212 | movhpd 8(%rsi), %xmm2 | |
e9f82e0d | 213 | # if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
42e08a54 UD |
214 | # define TOLOWER(reg1, reg2) \ |
215 | movdqa reg1, %xmm8; \ | |
216 | movdqa UCHIGH_reg, %xmm9; \ | |
217 | movdqa reg2, %xmm10; \ | |
218 | movdqa UCHIGH_reg, %xmm11; \ | |
219 | pcmpgtb UCLOW_reg, %xmm8; \ | |
220 | pcmpgtb reg1, %xmm9; \ | |
221 | pcmpgtb UCLOW_reg, %xmm10; \ | |
222 | pcmpgtb reg2, %xmm11; \ | |
223 | pand %xmm9, %xmm8; \ | |
224 | pand %xmm11, %xmm10; \ | |
225 | pand LCQWORD_reg, %xmm8; \ | |
226 | pand LCQWORD_reg, %xmm10; \ | |
227 | por %xmm8, reg1; \ | |
228 | por %xmm10, reg2 | |
229 | TOLOWER (%xmm1, %xmm2) | |
230 | # else | |
231 | # define TOLOWER(reg1, reg2) | |
232 | # endif | |
7956a3d2 L |
233 | pxor %xmm0, %xmm0 /* clear %xmm0 for null char checks */ |
234 | pcmpeqb %xmm1, %xmm0 /* Any null chars? */ | |
235 | pcmpeqb %xmm2, %xmm1 /* compare first 16 bytes for equality */ | |
236 | psubb %xmm0, %xmm1 /* packed sub of comparison results*/ | |
237 | pmovmskb %xmm1, %edx | |
238 | sub $0xffff, %edx /* if first 16 bytes are same, edx == 0xffff */ | |
239 | jnz LABEL(less16bytes) /* If not, find different value or null char */ | |
e9f82e0d | 240 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
241 | sub $16, %r11 |
242 | jbe LABEL(strcmp_exitz) /* finish comparision */ | |
f69190e7 | 243 | # endif |
7956a3d2 L |
244 | add $16, %rsi /* prepare to search next 16 bytes */ |
245 | add $16, %rdi /* prepare to search next 16 bytes */ | |
246 | ||
247 | /* | |
248 | * Determine source and destination string offsets from 16-byte alignment. | |
249 | * Use relative offset difference between the two to determine which case | |
250 | * below to use. | |
251 | */ | |
252 | .p2align 4 | |
253 | LABEL(crosscache): | |
254 | and $0xfffffffffffffff0, %rsi /* force %rsi is 16 byte aligned */ | |
255 | and $0xfffffffffffffff0, %rdi /* force %rdi is 16 byte aligned */ | |
256 | mov $0xffff, %edx /* for equivalent offset */ | |
257 | xor %r8d, %r8d | |
258 | and $0xf, %ecx /* offset of rsi */ | |
259 | and $0xf, %eax /* offset of rdi */ | |
260 | cmp %eax, %ecx | |
261 | je LABEL(ashr_0) /* rsi and rdi relative offset same */ | |
262 | ja LABEL(bigger) | |
263 | mov %edx, %r8d /* r8d is offset flag for exit tail */ | |
264 | xchg %ecx, %eax | |
265 | xchg %rsi, %rdi | |
266 | LABEL(bigger): | |
267 | lea 15(%rax), %r9 | |
268 | sub %rcx, %r9 | |
269 | lea LABEL(unaligned_table)(%rip), %r10 | |
270 | movslq (%r10, %r9,4), %r9 | |
271 | lea (%r10, %r9), %r10 | |
272 | jmp *%r10 /* jump to corresponding case */ | |
273 | ||
274 | /* | |
275 | * The following cases will be handled by ashr_0 | |
276 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
277 | * n(0~15) n(0~15) 15(15+ n-n) ashr_0 | |
278 | */ | |
279 | .p2align 4 | |
280 | LABEL(ashr_0): | |
281 | ||
282 | movdqa (%rsi), %xmm1 | |
283 | pxor %xmm0, %xmm0 /* clear %xmm0 for null char check */ | |
284 | pcmpeqb %xmm1, %xmm0 /* Any null chars? */ | |
e9f82e0d | 285 | # if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L |
7956a3d2 | 286 | pcmpeqb (%rdi), %xmm1 /* compare 16 bytes for equality */ |
42e08a54 UD |
287 | # else |
288 | movdqa (%rdi), %xmm2 | |
289 | TOLOWER (%xmm1, %xmm2) | |
290 | pcmpeqb %xmm2, %xmm1 /* compare 16 bytes for equality */ | |
291 | # endif | |
7956a3d2 L |
292 | psubb %xmm0, %xmm1 /* packed sub of comparison results*/ |
293 | pmovmskb %xmm1, %r9d | |
294 | shr %cl, %edx /* adjust 0xffff for offset */ | |
295 | shr %cl, %r9d /* adjust for 16-byte offset */ | |
296 | sub %r9d, %edx | |
297 | /* | |
298 | * edx must be the same with r9d if in left byte (16-rcx) is equal to | |
299 | * the start from (16-rax) and no null char was seen. | |
300 | */ | |
301 | jne LABEL(less32bytes) /* mismatch or null char */ | |
302 | UPDATE_STRNCMP_COUNTER | |
303 | mov $16, %rcx | |
304 | mov $16, %r9 | |
305 | pxor %xmm0, %xmm0 /* clear xmm0, may have changed above */ | |
306 | ||
307 | /* | |
308 | * Now both strings are aligned at 16-byte boundary. Loop over strings | |
309 | * checking 32-bytes per iteration. | |
310 | */ | |
311 | .p2align 4 | |
312 | LABEL(loop_ashr_0): | |
313 | movdqa (%rsi, %rcx), %xmm1 | |
314 | movdqa (%rdi, %rcx), %xmm2 | |
42e08a54 | 315 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
316 | |
317 | pcmpeqb %xmm1, %xmm0 | |
318 | pcmpeqb %xmm2, %xmm1 | |
319 | psubb %xmm0, %xmm1 | |
320 | pmovmskb %xmm1, %edx | |
321 | sub $0xffff, %edx | |
322 | jnz LABEL(exit) /* mismatch or null char seen */ | |
323 | ||
e9f82e0d | 324 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
325 | sub $16, %r11 |
326 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 327 | # endif |
7956a3d2 L |
328 | add $16, %rcx |
329 | movdqa (%rsi, %rcx), %xmm1 | |
330 | movdqa (%rdi, %rcx), %xmm2 | |
42e08a54 | 331 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
332 | |
333 | pcmpeqb %xmm1, %xmm0 | |
334 | pcmpeqb %xmm2, %xmm1 | |
335 | psubb %xmm0, %xmm1 | |
336 | pmovmskb %xmm1, %edx | |
337 | sub $0xffff, %edx | |
338 | jnz LABEL(exit) | |
e9f82e0d | 339 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
340 | sub $16, %r11 |
341 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 342 | # endif |
7956a3d2 L |
343 | add $16, %rcx |
344 | jmp LABEL(loop_ashr_0) | |
345 | ||
346 | /* | |
347 | * The following cases will be handled by ashr_1 | |
42e08a54 | 348 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
7956a3d2 L |
349 | * n(15) n -15 0(15 +(n-15) - n) ashr_1 |
350 | */ | |
351 | .p2align 4 | |
352 | LABEL(ashr_1): | |
353 | pxor %xmm0, %xmm0 | |
354 | movdqa (%rdi), %xmm2 | |
355 | movdqa (%rsi), %xmm1 | |
356 | pcmpeqb %xmm1, %xmm0 /* Any null chars? */ | |
357 | pslldq $15, %xmm2 /* shift first string to align with second */ | |
42e08a54 | 358 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
359 | pcmpeqb %xmm1, %xmm2 /* compare 16 bytes for equality */ |
360 | psubb %xmm0, %xmm2 /* packed sub of comparison results*/ | |
361 | pmovmskb %xmm2, %r9d | |
362 | shr %cl, %edx /* adjust 0xffff for offset */ | |
363 | shr %cl, %r9d /* adjust for 16-byte offset */ | |
364 | sub %r9d, %edx | |
365 | jnz LABEL(less32bytes) /* mismatch or null char seen */ | |
366 | movdqa (%rdi), %xmm3 | |
367 | UPDATE_STRNCMP_COUNTER | |
368 | ||
369 | pxor %xmm0, %xmm0 | |
370 | mov $16, %rcx /* index for loads*/ | |
371 | mov $1, %r9d /* byte position left over from less32bytes case */ | |
372 | /* | |
373 | * Setup %r10 value allows us to detect crossing a page boundary. | |
374 | * When %r10 goes positive we have crossed a page boundary and | |
375 | * need to do a nibble. | |
376 | */ | |
377 | lea 1(%rdi), %r10 | |
378 | and $0xfff, %r10 /* offset into 4K page */ | |
379 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
380 | ||
381 | .p2align 4 | |
382 | LABEL(loop_ashr_1): | |
383 | add $16, %r10 | |
384 | jg LABEL(nibble_ashr_1) /* cross page boundary */ | |
385 | ||
386 | LABEL(gobble_ashr_1): | |
387 | movdqa (%rsi, %rcx), %xmm1 | |
388 | movdqa (%rdi, %rcx), %xmm2 | |
389 | movdqa %xmm2, %xmm4 /* store for next cycle */ | |
390 | ||
f69190e7 | 391 | # ifndef USE_SSSE3 |
7956a3d2 L |
392 | psrldq $1, %xmm3 |
393 | pslldq $15, %xmm2 | |
394 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 395 | # else |
0fda545d | 396 | palignr $1, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 397 | # endif |
42e08a54 | 398 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
399 | |
400 | pcmpeqb %xmm1, %xmm0 | |
401 | pcmpeqb %xmm2, %xmm1 | |
402 | psubb %xmm0, %xmm1 | |
403 | pmovmskb %xmm1, %edx | |
404 | sub $0xffff, %edx | |
405 | jnz LABEL(exit) | |
406 | ||
e9f82e0d | 407 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
408 | sub $16, %r11 |
409 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 410 | # endif |
7956a3d2 L |
411 | add $16, %rcx |
412 | movdqa %xmm4, %xmm3 | |
413 | ||
414 | add $16, %r10 | |
415 | jg LABEL(nibble_ashr_1) /* cross page boundary */ | |
416 | ||
417 | movdqa (%rsi, %rcx), %xmm1 | |
418 | movdqa (%rdi, %rcx), %xmm2 | |
419 | movdqa %xmm2, %xmm4 /* store for next cycle */ | |
420 | ||
f69190e7 | 421 | # ifndef USE_SSSE3 |
7956a3d2 | 422 | psrldq $1, %xmm3 |
0fda545d | 423 | pslldq $15, %xmm2 |
7956a3d2 | 424 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 425 | # else |
0fda545d | 426 | palignr $1, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 427 | # endif |
42e08a54 | 428 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
429 | |
430 | pcmpeqb %xmm1, %xmm0 | |
431 | pcmpeqb %xmm2, %xmm1 | |
432 | psubb %xmm0, %xmm1 | |
433 | pmovmskb %xmm1, %edx | |
434 | sub $0xffff, %edx | |
435 | jnz LABEL(exit) | |
436 | ||
e9f82e0d | 437 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
438 | sub $16, %r11 |
439 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 440 | # endif |
7956a3d2 L |
441 | add $16, %rcx |
442 | movdqa %xmm4, %xmm3 | |
443 | jmp LABEL(loop_ashr_1) | |
444 | ||
445 | /* | |
446 | * Nibble avoids loads across page boundary. This is to avoid a potential | |
447 | * access into unmapped memory. | |
448 | */ | |
449 | .p2align 4 | |
450 | LABEL(nibble_ashr_1): | |
451 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char*/ | |
452 | pmovmskb %xmm0, %edx | |
453 | test $0xfffe, %edx | |
454 | jnz LABEL(ashr_1_exittail) /* find null char*/ | |
455 | ||
e9f82e0d | 456 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 457 | cmp $15, %r11 |
7956a3d2 | 458 | jbe LABEL(ashr_1_exittail) |
f69190e7 | 459 | # endif |
7956a3d2 L |
460 | |
461 | pxor %xmm0, %xmm0 | |
462 | sub $0x1000, %r10 /* substract 4K from %r10 */ | |
463 | jmp LABEL(gobble_ashr_1) | |
464 | ||
465 | /* | |
466 | * Once find null char, determine if there is a string mismatch | |
467 | * before the null char. | |
468 | */ | |
469 | .p2align 4 | |
470 | LABEL(ashr_1_exittail): | |
471 | movdqa (%rsi, %rcx), %xmm1 | |
472 | psrldq $1, %xmm0 | |
473 | psrldq $1, %xmm3 | |
474 | jmp LABEL(aftertail) | |
475 | ||
476 | /* | |
477 | * The following cases will be handled by ashr_2 | |
f69190e7 | 478 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
7956a3d2 L |
479 | * n(14~15) n -14 1(15 +(n-14) - n) ashr_2 |
480 | */ | |
481 | .p2align 4 | |
482 | LABEL(ashr_2): | |
483 | pxor %xmm0, %xmm0 | |
484 | movdqa (%rdi), %xmm2 | |
485 | movdqa (%rsi), %xmm1 | |
486 | pcmpeqb %xmm1, %xmm0 | |
487 | pslldq $14, %xmm2 | |
42e08a54 | 488 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
489 | pcmpeqb %xmm1, %xmm2 |
490 | psubb %xmm0, %xmm2 | |
491 | pmovmskb %xmm2, %r9d | |
492 | shr %cl, %edx | |
493 | shr %cl, %r9d | |
494 | sub %r9d, %edx | |
495 | jnz LABEL(less32bytes) | |
496 | movdqa (%rdi), %xmm3 | |
497 | UPDATE_STRNCMP_COUNTER | |
498 | ||
499 | pxor %xmm0, %xmm0 | |
500 | mov $16, %rcx /* index for loads */ | |
501 | mov $2, %r9d /* byte position left over from less32bytes case */ | |
502 | /* | |
503 | * Setup %r10 value allows us to detect crossing a page boundary. | |
504 | * When %r10 goes positive we have crossed a page boundary and | |
505 | * need to do a nibble. | |
506 | */ | |
507 | lea 2(%rdi), %r10 | |
508 | and $0xfff, %r10 /* offset into 4K page */ | |
509 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
510 | ||
511 | .p2align 4 | |
512 | LABEL(loop_ashr_2): | |
513 | add $16, %r10 | |
514 | jg LABEL(nibble_ashr_2) | |
515 | ||
516 | LABEL(gobble_ashr_2): | |
517 | movdqa (%rsi, %rcx), %xmm1 | |
518 | movdqa (%rdi, %rcx), %xmm2 | |
519 | movdqa %xmm2, %xmm4 | |
520 | ||
f69190e7 | 521 | # ifndef USE_SSSE3 |
7956a3d2 L |
522 | psrldq $2, %xmm3 |
523 | pslldq $14, %xmm2 | |
0fda545d | 524 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 525 | # else |
0fda545d | 526 | palignr $2, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 527 | # endif |
42e08a54 | 528 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
529 | |
530 | pcmpeqb %xmm1, %xmm0 | |
531 | pcmpeqb %xmm2, %xmm1 | |
532 | psubb %xmm0, %xmm1 | |
533 | pmovmskb %xmm1, %edx | |
534 | sub $0xffff, %edx | |
535 | jnz LABEL(exit) | |
536 | ||
e9f82e0d | 537 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
538 | sub $16, %r11 |
539 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 540 | # endif |
7956a3d2 L |
541 | |
542 | add $16, %rcx | |
543 | movdqa %xmm4, %xmm3 | |
544 | ||
545 | add $16, %r10 | |
546 | jg LABEL(nibble_ashr_2) /* cross page boundary */ | |
547 | ||
548 | movdqa (%rsi, %rcx), %xmm1 | |
549 | movdqa (%rdi, %rcx), %xmm2 | |
550 | movdqa %xmm2, %xmm4 | |
551 | ||
f69190e7 | 552 | # ifndef USE_SSSE3 |
7956a3d2 | 553 | psrldq $2, %xmm3 |
0fda545d UD |
554 | pslldq $14, %xmm2 |
555 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 556 | # else |
0fda545d | 557 | palignr $2, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 558 | # endif |
42e08a54 | 559 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
560 | |
561 | pcmpeqb %xmm1, %xmm0 | |
562 | pcmpeqb %xmm2, %xmm1 | |
563 | psubb %xmm0, %xmm1 | |
564 | pmovmskb %xmm1, %edx | |
565 | sub $0xffff, %edx | |
566 | jnz LABEL(exit) | |
567 | ||
e9f82e0d | 568 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
569 | sub $16, %r11 |
570 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 571 | # endif |
7956a3d2 L |
572 | |
573 | add $16, %rcx | |
574 | movdqa %xmm4, %xmm3 | |
575 | jmp LABEL(loop_ashr_2) | |
576 | ||
577 | .p2align 4 | |
578 | LABEL(nibble_ashr_2): | |
579 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
580 | pmovmskb %xmm0, %edx | |
581 | test $0xfffc, %edx | |
582 | jnz LABEL(ashr_2_exittail) | |
583 | ||
e9f82e0d | 584 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 585 | cmp $14, %r11 |
7956a3d2 | 586 | jbe LABEL(ashr_2_exittail) |
f69190e7 | 587 | # endif |
7956a3d2 L |
588 | |
589 | pxor %xmm0, %xmm0 | |
590 | sub $0x1000, %r10 | |
591 | jmp LABEL(gobble_ashr_2) | |
592 | ||
593 | .p2align 4 | |
594 | LABEL(ashr_2_exittail): | |
595 | movdqa (%rsi, %rcx), %xmm1 | |
596 | psrldq $2, %xmm0 | |
597 | psrldq $2, %xmm3 | |
598 | jmp LABEL(aftertail) | |
599 | ||
600 | /* | |
601 | * The following cases will be handled by ashr_3 | |
602 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
603 | * n(13~15) n -13 2(15 +(n-13) - n) ashr_3 | |
604 | */ | |
605 | .p2align 4 | |
606 | LABEL(ashr_3): | |
607 | pxor %xmm0, %xmm0 | |
608 | movdqa (%rdi), %xmm2 | |
609 | movdqa (%rsi), %xmm1 | |
610 | pcmpeqb %xmm1, %xmm0 | |
611 | pslldq $13, %xmm2 | |
42e08a54 | 612 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
613 | pcmpeqb %xmm1, %xmm2 |
614 | psubb %xmm0, %xmm2 | |
615 | pmovmskb %xmm2, %r9d | |
616 | shr %cl, %edx | |
617 | shr %cl, %r9d | |
618 | sub %r9d, %edx | |
619 | jnz LABEL(less32bytes) | |
620 | movdqa (%rdi), %xmm3 | |
621 | ||
622 | UPDATE_STRNCMP_COUNTER | |
623 | ||
624 | pxor %xmm0, %xmm0 | |
625 | mov $16, %rcx /* index for loads */ | |
626 | mov $3, %r9d /* byte position left over from less32bytes case */ | |
627 | /* | |
628 | * Setup %r10 value allows us to detect crossing a page boundary. | |
629 | * When %r10 goes positive we have crossed a page boundary and | |
630 | * need to do a nibble. | |
631 | */ | |
632 | lea 3(%rdi), %r10 | |
633 | and $0xfff, %r10 /* offset into 4K page */ | |
634 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
635 | ||
636 | .p2align 4 | |
637 | LABEL(loop_ashr_3): | |
638 | add $16, %r10 | |
639 | jg LABEL(nibble_ashr_3) | |
640 | ||
641 | LABEL(gobble_ashr_3): | |
642 | movdqa (%rsi, %rcx), %xmm1 | |
643 | movdqa (%rdi, %rcx), %xmm2 | |
644 | movdqa %xmm2, %xmm4 | |
645 | ||
f69190e7 | 646 | # ifndef USE_SSSE3 |
7956a3d2 L |
647 | psrldq $3, %xmm3 |
648 | pslldq $13, %xmm2 | |
0fda545d | 649 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 650 | # else |
0fda545d | 651 | palignr $3, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 652 | # endif |
42e08a54 | 653 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
654 | |
655 | pcmpeqb %xmm1, %xmm0 | |
656 | pcmpeqb %xmm2, %xmm1 | |
657 | psubb %xmm0, %xmm1 | |
658 | pmovmskb %xmm1, %edx | |
659 | sub $0xffff, %edx | |
660 | jnz LABEL(exit) | |
661 | ||
e9f82e0d | 662 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
663 | sub $16, %r11 |
664 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 665 | # endif |
7956a3d2 L |
666 | |
667 | add $16, %rcx | |
668 | movdqa %xmm4, %xmm3 | |
669 | ||
670 | add $16, %r10 | |
671 | jg LABEL(nibble_ashr_3) /* cross page boundary */ | |
672 | ||
673 | movdqa (%rsi, %rcx), %xmm1 | |
674 | movdqa (%rdi, %rcx), %xmm2 | |
675 | movdqa %xmm2, %xmm4 | |
676 | ||
f69190e7 | 677 | # ifndef USE_SSSE3 |
7956a3d2 | 678 | psrldq $3, %xmm3 |
0fda545d UD |
679 | pslldq $13, %xmm2 |
680 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 681 | # else |
0fda545d | 682 | palignr $3, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 683 | # endif |
42e08a54 | 684 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
685 | |
686 | pcmpeqb %xmm1, %xmm0 | |
687 | pcmpeqb %xmm2, %xmm1 | |
688 | psubb %xmm0, %xmm1 | |
689 | pmovmskb %xmm1, %edx | |
690 | sub $0xffff, %edx | |
691 | jnz LABEL(exit) | |
692 | ||
e9f82e0d | 693 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
694 | sub $16, %r11 |
695 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 696 | # endif |
7956a3d2 L |
697 | |
698 | add $16, %rcx | |
699 | movdqa %xmm4, %xmm3 | |
700 | jmp LABEL(loop_ashr_3) | |
701 | ||
702 | .p2align 4 | |
703 | LABEL(nibble_ashr_3): | |
704 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
705 | pmovmskb %xmm0, %edx | |
706 | test $0xfff8, %edx | |
707 | jnz LABEL(ashr_3_exittail) | |
708 | ||
e9f82e0d | 709 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 710 | cmp $13, %r11 |
7956a3d2 | 711 | jbe LABEL(ashr_3_exittail) |
f69190e7 | 712 | # endif |
7956a3d2 L |
713 | |
714 | pxor %xmm0, %xmm0 | |
715 | sub $0x1000, %r10 | |
716 | jmp LABEL(gobble_ashr_3) | |
717 | ||
718 | .p2align 4 | |
719 | LABEL(ashr_3_exittail): | |
720 | movdqa (%rsi, %rcx), %xmm1 | |
721 | psrldq $3, %xmm0 | |
722 | psrldq $3, %xmm3 | |
723 | jmp LABEL(aftertail) | |
724 | ||
725 | /* | |
726 | * The following cases will be handled by ashr_4 | |
727 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
728 | * n(12~15) n -12 3(15 +(n-12) - n) ashr_4 | |
729 | */ | |
730 | .p2align 4 | |
731 | LABEL(ashr_4): | |
732 | pxor %xmm0, %xmm0 | |
733 | movdqa (%rdi), %xmm2 | |
734 | movdqa (%rsi), %xmm1 | |
735 | pcmpeqb %xmm1, %xmm0 | |
736 | pslldq $12, %xmm2 | |
42e08a54 | 737 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
738 | pcmpeqb %xmm1, %xmm2 |
739 | psubb %xmm0, %xmm2 | |
740 | pmovmskb %xmm2, %r9d | |
741 | shr %cl, %edx | |
742 | shr %cl, %r9d | |
743 | sub %r9d, %edx | |
744 | jnz LABEL(less32bytes) | |
745 | movdqa (%rdi), %xmm3 | |
746 | ||
747 | UPDATE_STRNCMP_COUNTER | |
748 | ||
749 | pxor %xmm0, %xmm0 | |
750 | mov $16, %rcx /* index for loads */ | |
751 | mov $4, %r9d /* byte position left over from less32bytes case */ | |
752 | /* | |
753 | * Setup %r10 value allows us to detect crossing a page boundary. | |
754 | * When %r10 goes positive we have crossed a page boundary and | |
755 | * need to do a nibble. | |
756 | */ | |
757 | lea 4(%rdi), %r10 | |
758 | and $0xfff, %r10 /* offset into 4K page */ | |
759 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
760 | ||
761 | .p2align 4 | |
762 | LABEL(loop_ashr_4): | |
763 | add $16, %r10 | |
764 | jg LABEL(nibble_ashr_4) | |
765 | ||
766 | LABEL(gobble_ashr_4): | |
767 | movdqa (%rsi, %rcx), %xmm1 | |
768 | movdqa (%rdi, %rcx), %xmm2 | |
769 | movdqa %xmm2, %xmm4 | |
770 | ||
f69190e7 | 771 | # ifndef USE_SSSE3 |
7956a3d2 L |
772 | psrldq $4, %xmm3 |
773 | pslldq $12, %xmm2 | |
0fda545d | 774 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 775 | # else |
0fda545d | 776 | palignr $4, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 777 | # endif |
42e08a54 | 778 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
779 | |
780 | pcmpeqb %xmm1, %xmm0 | |
781 | pcmpeqb %xmm2, %xmm1 | |
782 | psubb %xmm0, %xmm1 | |
783 | pmovmskb %xmm1, %edx | |
784 | sub $0xffff, %edx | |
785 | jnz LABEL(exit) | |
786 | ||
e9f82e0d | 787 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
788 | sub $16, %r11 |
789 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 790 | # endif |
7956a3d2 L |
791 | |
792 | add $16, %rcx | |
793 | movdqa %xmm4, %xmm3 | |
794 | ||
795 | add $16, %r10 | |
796 | jg LABEL(nibble_ashr_4) /* cross page boundary */ | |
797 | ||
798 | movdqa (%rsi, %rcx), %xmm1 | |
799 | movdqa (%rdi, %rcx), %xmm2 | |
800 | movdqa %xmm2, %xmm4 | |
801 | ||
f69190e7 | 802 | # ifndef USE_SSSE3 |
7956a3d2 | 803 | psrldq $4, %xmm3 |
0fda545d UD |
804 | pslldq $12, %xmm2 |
805 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 806 | # else |
0fda545d | 807 | palignr $4, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 808 | # endif |
42e08a54 | 809 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
810 | |
811 | pcmpeqb %xmm1, %xmm0 | |
812 | pcmpeqb %xmm2, %xmm1 | |
813 | psubb %xmm0, %xmm1 | |
814 | pmovmskb %xmm1, %edx | |
815 | sub $0xffff, %edx | |
816 | jnz LABEL(exit) | |
817 | ||
e9f82e0d | 818 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
819 | sub $16, %r11 |
820 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 821 | # endif |
7956a3d2 L |
822 | |
823 | add $16, %rcx | |
824 | movdqa %xmm4, %xmm3 | |
825 | jmp LABEL(loop_ashr_4) | |
826 | ||
827 | .p2align 4 | |
828 | LABEL(nibble_ashr_4): | |
829 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
830 | pmovmskb %xmm0, %edx | |
831 | test $0xfff0, %edx | |
832 | jnz LABEL(ashr_4_exittail) | |
833 | ||
e9f82e0d | 834 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 835 | cmp $12, %r11 |
7956a3d2 | 836 | jbe LABEL(ashr_4_exittail) |
f69190e7 | 837 | # endif |
7956a3d2 L |
838 | |
839 | pxor %xmm0, %xmm0 | |
840 | sub $0x1000, %r10 | |
841 | jmp LABEL(gobble_ashr_4) | |
842 | ||
843 | .p2align 4 | |
844 | LABEL(ashr_4_exittail): | |
845 | movdqa (%rsi, %rcx), %xmm1 | |
846 | psrldq $4, %xmm0 | |
847 | psrldq $4, %xmm3 | |
848 | jmp LABEL(aftertail) | |
849 | ||
850 | /* | |
851 | * The following cases will be handled by ashr_5 | |
852 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 853 | * n(11~15) n - 11 4(15 +(n-11) - n) ashr_5 |
7956a3d2 L |
854 | */ |
855 | .p2align 4 | |
856 | LABEL(ashr_5): | |
857 | pxor %xmm0, %xmm0 | |
858 | movdqa (%rdi), %xmm2 | |
859 | movdqa (%rsi), %xmm1 | |
860 | pcmpeqb %xmm1, %xmm0 | |
861 | pslldq $11, %xmm2 | |
42e08a54 | 862 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
863 | pcmpeqb %xmm1, %xmm2 |
864 | psubb %xmm0, %xmm2 | |
865 | pmovmskb %xmm2, %r9d | |
866 | shr %cl, %edx | |
867 | shr %cl, %r9d | |
868 | sub %r9d, %edx | |
869 | jnz LABEL(less32bytes) | |
870 | movdqa (%rdi), %xmm3 | |
871 | ||
872 | UPDATE_STRNCMP_COUNTER | |
873 | ||
874 | pxor %xmm0, %xmm0 | |
875 | mov $16, %rcx /* index for loads */ | |
876 | mov $5, %r9d /* byte position left over from less32bytes case */ | |
877 | /* | |
878 | * Setup %r10 value allows us to detect crossing a page boundary. | |
879 | * When %r10 goes positive we have crossed a page boundary and | |
880 | * need to do a nibble. | |
881 | */ | |
882 | lea 5(%rdi), %r10 | |
883 | and $0xfff, %r10 /* offset into 4K page */ | |
884 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
885 | ||
886 | .p2align 4 | |
887 | LABEL(loop_ashr_5): | |
888 | add $16, %r10 | |
889 | jg LABEL(nibble_ashr_5) | |
890 | ||
891 | LABEL(gobble_ashr_5): | |
892 | movdqa (%rsi, %rcx), %xmm1 | |
893 | movdqa (%rdi, %rcx), %xmm2 | |
894 | movdqa %xmm2, %xmm4 | |
895 | ||
f69190e7 | 896 | # ifndef USE_SSSE3 |
7956a3d2 L |
897 | psrldq $5, %xmm3 |
898 | pslldq $11, %xmm2 | |
0fda545d | 899 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 900 | # else |
0fda545d | 901 | palignr $5, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 902 | # endif |
42e08a54 | 903 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
904 | |
905 | pcmpeqb %xmm1, %xmm0 | |
906 | pcmpeqb %xmm2, %xmm1 | |
907 | psubb %xmm0, %xmm1 | |
908 | pmovmskb %xmm1, %edx | |
909 | sub $0xffff, %edx | |
910 | jnz LABEL(exit) | |
911 | ||
8ffcee4a | 912 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
913 | sub $16, %r11 |
914 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 915 | # endif |
7956a3d2 L |
916 | |
917 | add $16, %rcx | |
918 | movdqa %xmm4, %xmm3 | |
919 | ||
920 | add $16, %r10 | |
921 | jg LABEL(nibble_ashr_5) /* cross page boundary */ | |
922 | ||
923 | movdqa (%rsi, %rcx), %xmm1 | |
924 | movdqa (%rdi, %rcx), %xmm2 | |
925 | movdqa %xmm2, %xmm4 | |
926 | ||
f69190e7 | 927 | # ifndef USE_SSSE3 |
7956a3d2 | 928 | psrldq $5, %xmm3 |
0fda545d UD |
929 | pslldq $11, %xmm2 |
930 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 931 | # else |
0fda545d | 932 | palignr $5, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 933 | # endif |
42e08a54 | 934 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
935 | |
936 | pcmpeqb %xmm1, %xmm0 | |
937 | pcmpeqb %xmm2, %xmm1 | |
938 | psubb %xmm0, %xmm1 | |
939 | pmovmskb %xmm1, %edx | |
940 | sub $0xffff, %edx | |
941 | jnz LABEL(exit) | |
942 | ||
e9f82e0d | 943 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
944 | sub $16, %r11 |
945 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 946 | # endif |
7956a3d2 L |
947 | |
948 | add $16, %rcx | |
949 | movdqa %xmm4, %xmm3 | |
950 | jmp LABEL(loop_ashr_5) | |
951 | ||
952 | .p2align 4 | |
953 | LABEL(nibble_ashr_5): | |
954 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
955 | pmovmskb %xmm0, %edx | |
956 | test $0xffe0, %edx | |
957 | jnz LABEL(ashr_5_exittail) | |
958 | ||
e9f82e0d | 959 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 960 | cmp $11, %r11 |
7956a3d2 | 961 | jbe LABEL(ashr_5_exittail) |
f69190e7 | 962 | # endif |
7956a3d2 L |
963 | |
964 | pxor %xmm0, %xmm0 | |
965 | sub $0x1000, %r10 | |
966 | jmp LABEL(gobble_ashr_5) | |
967 | ||
968 | .p2align 4 | |
969 | LABEL(ashr_5_exittail): | |
970 | movdqa (%rsi, %rcx), %xmm1 | |
971 | psrldq $5, %xmm0 | |
972 | psrldq $5, %xmm3 | |
973 | jmp LABEL(aftertail) | |
974 | ||
975 | /* | |
976 | * The following cases will be handled by ashr_6 | |
977 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 978 | * n(10~15) n - 10 5(15 +(n-10) - n) ashr_6 |
7956a3d2 L |
979 | */ |
980 | .p2align 4 | |
981 | LABEL(ashr_6): | |
982 | pxor %xmm0, %xmm0 | |
983 | movdqa (%rdi), %xmm2 | |
984 | movdqa (%rsi), %xmm1 | |
985 | pcmpeqb %xmm1, %xmm0 | |
986 | pslldq $10, %xmm2 | |
42e08a54 | 987 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
988 | pcmpeqb %xmm1, %xmm2 |
989 | psubb %xmm0, %xmm2 | |
990 | pmovmskb %xmm2, %r9d | |
991 | shr %cl, %edx | |
992 | shr %cl, %r9d | |
993 | sub %r9d, %edx | |
994 | jnz LABEL(less32bytes) | |
995 | movdqa (%rdi), %xmm3 | |
996 | ||
997 | UPDATE_STRNCMP_COUNTER | |
998 | ||
999 | pxor %xmm0, %xmm0 | |
1000 | mov $16, %rcx /* index for loads */ | |
1001 | mov $6, %r9d /* byte position left over from less32bytes case */ | |
1002 | /* | |
1003 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1004 | * When %r10 goes positive we have crossed a page boundary and | |
1005 | * need to do a nibble. | |
1006 | */ | |
1007 | lea 6(%rdi), %r10 | |
1008 | and $0xfff, %r10 /* offset into 4K page */ | |
1009 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1010 | ||
1011 | .p2align 4 | |
1012 | LABEL(loop_ashr_6): | |
1013 | add $16, %r10 | |
1014 | jg LABEL(nibble_ashr_6) | |
1015 | ||
1016 | LABEL(gobble_ashr_6): | |
1017 | movdqa (%rsi, %rcx), %xmm1 | |
1018 | movdqa (%rdi, %rcx), %xmm2 | |
1019 | movdqa %xmm2, %xmm4 | |
1020 | ||
f69190e7 | 1021 | # ifndef USE_SSSE3 |
7956a3d2 L |
1022 | psrldq $6, %xmm3 |
1023 | pslldq $10, %xmm2 | |
0fda545d | 1024 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1025 | # else |
0fda545d | 1026 | palignr $6, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1027 | # endif |
42e08a54 | 1028 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1029 | |
1030 | pcmpeqb %xmm1, %xmm0 | |
1031 | pcmpeqb %xmm2, %xmm1 | |
1032 | psubb %xmm0, %xmm1 | |
1033 | pmovmskb %xmm1, %edx | |
1034 | sub $0xffff, %edx | |
1035 | jnz LABEL(exit) | |
1036 | ||
e9f82e0d | 1037 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1038 | sub $16, %r11 |
1039 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1040 | # endif |
7956a3d2 L |
1041 | |
1042 | add $16, %rcx | |
1043 | movdqa %xmm4, %xmm3 | |
1044 | ||
1045 | add $16, %r10 | |
1046 | jg LABEL(nibble_ashr_6) /* cross page boundary */ | |
1047 | ||
1048 | movdqa (%rsi, %rcx), %xmm1 | |
1049 | movdqa (%rdi, %rcx), %xmm2 | |
1050 | movdqa %xmm2, %xmm4 | |
1051 | ||
f69190e7 | 1052 | # ifndef USE_SSSE3 |
7956a3d2 | 1053 | psrldq $6, %xmm3 |
0fda545d UD |
1054 | pslldq $10, %xmm2 |
1055 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1056 | # else |
0fda545d | 1057 | palignr $6, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1058 | # endif |
42e08a54 | 1059 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1060 | |
1061 | pcmpeqb %xmm1, %xmm0 | |
1062 | pcmpeqb %xmm2, %xmm1 | |
1063 | psubb %xmm0, %xmm1 | |
1064 | pmovmskb %xmm1, %edx | |
1065 | sub $0xffff, %edx | |
1066 | jnz LABEL(exit) | |
1067 | ||
e9f82e0d | 1068 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1069 | sub $16, %r11 |
1070 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1071 | # endif |
7956a3d2 L |
1072 | |
1073 | add $16, %rcx | |
1074 | movdqa %xmm4, %xmm3 | |
1075 | jmp LABEL(loop_ashr_6) | |
1076 | ||
1077 | .p2align 4 | |
1078 | LABEL(nibble_ashr_6): | |
1079 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1080 | pmovmskb %xmm0, %edx | |
1081 | test $0xffc0, %edx | |
1082 | jnz LABEL(ashr_6_exittail) | |
1083 | ||
e9f82e0d | 1084 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1085 | cmp $10, %r11 |
7956a3d2 | 1086 | jbe LABEL(ashr_6_exittail) |
f69190e7 | 1087 | # endif |
7956a3d2 L |
1088 | |
1089 | pxor %xmm0, %xmm0 | |
1090 | sub $0x1000, %r10 | |
1091 | jmp LABEL(gobble_ashr_6) | |
1092 | ||
1093 | .p2align 4 | |
1094 | LABEL(ashr_6_exittail): | |
1095 | movdqa (%rsi, %rcx), %xmm1 | |
1096 | psrldq $6, %xmm0 | |
1097 | psrldq $6, %xmm3 | |
1098 | jmp LABEL(aftertail) | |
1099 | ||
1100 | /* | |
1101 | * The following cases will be handled by ashr_7 | |
1102 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1103 | * n(9~15) n - 9 6(15 +(n - 9) - n) ashr_7 |
7956a3d2 L |
1104 | */ |
1105 | .p2align 4 | |
1106 | LABEL(ashr_7): | |
1107 | pxor %xmm0, %xmm0 | |
1108 | movdqa (%rdi), %xmm2 | |
1109 | movdqa (%rsi), %xmm1 | |
1110 | pcmpeqb %xmm1, %xmm0 | |
1111 | pslldq $9, %xmm2 | |
42e08a54 | 1112 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1113 | pcmpeqb %xmm1, %xmm2 |
1114 | psubb %xmm0, %xmm2 | |
1115 | pmovmskb %xmm2, %r9d | |
1116 | shr %cl, %edx | |
1117 | shr %cl, %r9d | |
1118 | sub %r9d, %edx | |
1119 | jnz LABEL(less32bytes) | |
1120 | movdqa (%rdi), %xmm3 | |
1121 | ||
1122 | UPDATE_STRNCMP_COUNTER | |
1123 | ||
1124 | pxor %xmm0, %xmm0 | |
1125 | mov $16, %rcx /* index for loads */ | |
1126 | mov $7, %r9d /* byte position left over from less32bytes case */ | |
1127 | /* | |
1128 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1129 | * When %r10 goes positive we have crossed a page boundary and | |
1130 | * need to do a nibble. | |
1131 | */ | |
1132 | lea 7(%rdi), %r10 | |
1133 | and $0xfff, %r10 /* offset into 4K page */ | |
1134 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1135 | ||
1136 | .p2align 4 | |
1137 | LABEL(loop_ashr_7): | |
1138 | add $16, %r10 | |
1139 | jg LABEL(nibble_ashr_7) | |
1140 | ||
1141 | LABEL(gobble_ashr_7): | |
1142 | movdqa (%rsi, %rcx), %xmm1 | |
1143 | movdqa (%rdi, %rcx), %xmm2 | |
1144 | movdqa %xmm2, %xmm4 | |
1145 | ||
f69190e7 | 1146 | # ifndef USE_SSSE3 |
7956a3d2 L |
1147 | psrldq $7, %xmm3 |
1148 | pslldq $9, %xmm2 | |
0fda545d | 1149 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1150 | # else |
0fda545d | 1151 | palignr $7, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1152 | # endif |
42e08a54 | 1153 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1154 | |
1155 | pcmpeqb %xmm1, %xmm0 | |
1156 | pcmpeqb %xmm2, %xmm1 | |
1157 | psubb %xmm0, %xmm1 | |
1158 | pmovmskb %xmm1, %edx | |
1159 | sub $0xffff, %edx | |
1160 | jnz LABEL(exit) | |
1161 | ||
e9f82e0d | 1162 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1163 | sub $16, %r11 |
1164 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1165 | # endif |
7956a3d2 L |
1166 | |
1167 | add $16, %rcx | |
1168 | movdqa %xmm4, %xmm3 | |
1169 | ||
1170 | add $16, %r10 | |
1171 | jg LABEL(nibble_ashr_7) /* cross page boundary */ | |
1172 | ||
1173 | movdqa (%rsi, %rcx), %xmm1 | |
1174 | movdqa (%rdi, %rcx), %xmm2 | |
1175 | movdqa %xmm2, %xmm4 | |
1176 | ||
f69190e7 | 1177 | # ifndef USE_SSSE3 |
7956a3d2 | 1178 | psrldq $7, %xmm3 |
0fda545d UD |
1179 | pslldq $9, %xmm2 |
1180 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1181 | # else |
0fda545d | 1182 | palignr $7, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1183 | # endif |
42e08a54 | 1184 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1185 | |
1186 | pcmpeqb %xmm1, %xmm0 | |
1187 | pcmpeqb %xmm2, %xmm1 | |
1188 | psubb %xmm0, %xmm1 | |
1189 | pmovmskb %xmm1, %edx | |
1190 | sub $0xffff, %edx | |
1191 | jnz LABEL(exit) | |
1192 | ||
e9f82e0d | 1193 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1194 | sub $16, %r11 |
1195 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1196 | # endif |
7956a3d2 L |
1197 | |
1198 | add $16, %rcx | |
1199 | movdqa %xmm4, %xmm3 | |
1200 | jmp LABEL(loop_ashr_7) | |
1201 | ||
1202 | .p2align 4 | |
1203 | LABEL(nibble_ashr_7): | |
1204 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1205 | pmovmskb %xmm0, %edx | |
1206 | test $0xff80, %edx | |
1207 | jnz LABEL(ashr_7_exittail) | |
1208 | ||
e9f82e0d | 1209 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1210 | cmp $9, %r11 |
7956a3d2 | 1211 | jbe LABEL(ashr_7_exittail) |
f69190e7 | 1212 | # endif |
7956a3d2 L |
1213 | |
1214 | pxor %xmm0, %xmm0 | |
1215 | sub $0x1000, %r10 | |
1216 | jmp LABEL(gobble_ashr_7) | |
1217 | ||
1218 | .p2align 4 | |
1219 | LABEL(ashr_7_exittail): | |
1220 | movdqa (%rsi, %rcx), %xmm1 | |
1221 | psrldq $7, %xmm0 | |
1222 | psrldq $7, %xmm3 | |
1223 | jmp LABEL(aftertail) | |
1224 | ||
1225 | /* | |
1226 | * The following cases will be handled by ashr_8 | |
1227 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1228 | * n(8~15) n - 8 7(15 +(n - 8) - n) ashr_8 |
7956a3d2 L |
1229 | */ |
1230 | .p2align 4 | |
1231 | LABEL(ashr_8): | |
1232 | pxor %xmm0, %xmm0 | |
1233 | movdqa (%rdi), %xmm2 | |
1234 | movdqa (%rsi), %xmm1 | |
1235 | pcmpeqb %xmm1, %xmm0 | |
1236 | pslldq $8, %xmm2 | |
42e08a54 | 1237 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1238 | pcmpeqb %xmm1, %xmm2 |
1239 | psubb %xmm0, %xmm2 | |
1240 | pmovmskb %xmm2, %r9d | |
1241 | shr %cl, %edx | |
1242 | shr %cl, %r9d | |
1243 | sub %r9d, %edx | |
1244 | jnz LABEL(less32bytes) | |
1245 | movdqa (%rdi), %xmm3 | |
1246 | ||
1247 | UPDATE_STRNCMP_COUNTER | |
1248 | ||
1249 | pxor %xmm0, %xmm0 | |
1250 | mov $16, %rcx /* index for loads */ | |
1251 | mov $8, %r9d /* byte position left over from less32bytes case */ | |
1252 | /* | |
1253 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1254 | * When %r10 goes positive we have crossed a page boundary and | |
1255 | * need to do a nibble. | |
1256 | */ | |
1257 | lea 8(%rdi), %r10 | |
1258 | and $0xfff, %r10 /* offset into 4K page */ | |
1259 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1260 | ||
1261 | .p2align 4 | |
1262 | LABEL(loop_ashr_8): | |
1263 | add $16, %r10 | |
1264 | jg LABEL(nibble_ashr_8) | |
1265 | ||
1266 | LABEL(gobble_ashr_8): | |
1267 | movdqa (%rsi, %rcx), %xmm1 | |
1268 | movdqa (%rdi, %rcx), %xmm2 | |
1269 | movdqa %xmm2, %xmm4 | |
1270 | ||
f69190e7 | 1271 | # ifndef USE_SSSE3 |
7956a3d2 L |
1272 | psrldq $8, %xmm3 |
1273 | pslldq $8, %xmm2 | |
0fda545d | 1274 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1275 | # else |
0fda545d | 1276 | palignr $8, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1277 | # endif |
42e08a54 | 1278 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1279 | |
1280 | pcmpeqb %xmm1, %xmm0 | |
1281 | pcmpeqb %xmm2, %xmm1 | |
1282 | psubb %xmm0, %xmm1 | |
1283 | pmovmskb %xmm1, %edx | |
1284 | sub $0xffff, %edx | |
1285 | jnz LABEL(exit) | |
1286 | ||
e9f82e0d | 1287 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1288 | sub $16, %r11 |
1289 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1290 | # endif |
7956a3d2 L |
1291 | |
1292 | add $16, %rcx | |
1293 | movdqa %xmm4, %xmm3 | |
1294 | ||
1295 | add $16, %r10 | |
1296 | jg LABEL(nibble_ashr_8) /* cross page boundary */ | |
1297 | ||
1298 | movdqa (%rsi, %rcx), %xmm1 | |
1299 | movdqa (%rdi, %rcx), %xmm2 | |
1300 | movdqa %xmm2, %xmm4 | |
1301 | ||
f69190e7 | 1302 | # ifndef USE_SSSE3 |
7956a3d2 | 1303 | psrldq $8, %xmm3 |
0fda545d UD |
1304 | pslldq $8, %xmm2 |
1305 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1306 | # else |
0fda545d | 1307 | palignr $8, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1308 | # endif |
42e08a54 | 1309 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1310 | |
1311 | pcmpeqb %xmm1, %xmm0 | |
1312 | pcmpeqb %xmm2, %xmm1 | |
1313 | psubb %xmm0, %xmm1 | |
1314 | pmovmskb %xmm1, %edx | |
1315 | sub $0xffff, %edx | |
1316 | jnz LABEL(exit) | |
1317 | ||
e9f82e0d | 1318 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1319 | sub $16, %r11 |
1320 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1321 | # endif |
7956a3d2 L |
1322 | |
1323 | add $16, %rcx | |
1324 | movdqa %xmm4, %xmm3 | |
1325 | jmp LABEL(loop_ashr_8) | |
1326 | ||
1327 | .p2align 4 | |
1328 | LABEL(nibble_ashr_8): | |
1329 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1330 | pmovmskb %xmm0, %edx | |
1331 | test $0xff00, %edx | |
1332 | jnz LABEL(ashr_8_exittail) | |
1333 | ||
e9f82e0d | 1334 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1335 | cmp $8, %r11 |
7956a3d2 | 1336 | jbe LABEL(ashr_8_exittail) |
f69190e7 | 1337 | # endif |
7956a3d2 L |
1338 | |
1339 | pxor %xmm0, %xmm0 | |
1340 | sub $0x1000, %r10 | |
1341 | jmp LABEL(gobble_ashr_8) | |
1342 | ||
1343 | .p2align 4 | |
1344 | LABEL(ashr_8_exittail): | |
1345 | movdqa (%rsi, %rcx), %xmm1 | |
1346 | psrldq $8, %xmm0 | |
1347 | psrldq $8, %xmm3 | |
1348 | jmp LABEL(aftertail) | |
1349 | ||
1350 | /* | |
1351 | * The following cases will be handled by ashr_9 | |
1352 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1353 | * n(7~15) n - 7 8(15 +(n - 7) - n) ashr_9 |
7956a3d2 L |
1354 | */ |
1355 | .p2align 4 | |
1356 | LABEL(ashr_9): | |
1357 | pxor %xmm0, %xmm0 | |
1358 | movdqa (%rdi), %xmm2 | |
1359 | movdqa (%rsi), %xmm1 | |
1360 | pcmpeqb %xmm1, %xmm0 | |
1361 | pslldq $7, %xmm2 | |
42e08a54 | 1362 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1363 | pcmpeqb %xmm1, %xmm2 |
1364 | psubb %xmm0, %xmm2 | |
1365 | pmovmskb %xmm2, %r9d | |
1366 | shr %cl, %edx | |
1367 | shr %cl, %r9d | |
1368 | sub %r9d, %edx | |
1369 | jnz LABEL(less32bytes) | |
1370 | movdqa (%rdi), %xmm3 | |
1371 | ||
1372 | UPDATE_STRNCMP_COUNTER | |
1373 | ||
1374 | pxor %xmm0, %xmm0 | |
1375 | mov $16, %rcx /* index for loads */ | |
1376 | mov $9, %r9d /* byte position left over from less32bytes case */ | |
1377 | /* | |
1378 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1379 | * When %r10 goes positive we have crossed a page boundary and | |
1380 | * need to do a nibble. | |
1381 | */ | |
1382 | lea 9(%rdi), %r10 | |
1383 | and $0xfff, %r10 /* offset into 4K page */ | |
1384 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1385 | ||
1386 | .p2align 4 | |
1387 | LABEL(loop_ashr_9): | |
1388 | add $16, %r10 | |
1389 | jg LABEL(nibble_ashr_9) | |
1390 | ||
1391 | LABEL(gobble_ashr_9): | |
1392 | movdqa (%rsi, %rcx), %xmm1 | |
1393 | movdqa (%rdi, %rcx), %xmm2 | |
1394 | movdqa %xmm2, %xmm4 | |
1395 | ||
f69190e7 | 1396 | # ifndef USE_SSSE3 |
7956a3d2 L |
1397 | psrldq $9, %xmm3 |
1398 | pslldq $7, %xmm2 | |
0fda545d | 1399 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1400 | # else |
0fda545d | 1401 | palignr $9, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1402 | # endif |
42e08a54 | 1403 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1404 | |
1405 | pcmpeqb %xmm1, %xmm0 | |
1406 | pcmpeqb %xmm2, %xmm1 | |
1407 | psubb %xmm0, %xmm1 | |
1408 | pmovmskb %xmm1, %edx | |
1409 | sub $0xffff, %edx | |
1410 | jnz LABEL(exit) | |
1411 | ||
e9f82e0d | 1412 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1413 | sub $16, %r11 |
1414 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1415 | # endif |
7956a3d2 L |
1416 | |
1417 | add $16, %rcx | |
1418 | movdqa %xmm4, %xmm3 | |
1419 | ||
1420 | add $16, %r10 | |
1421 | jg LABEL(nibble_ashr_9) /* cross page boundary */ | |
1422 | ||
1423 | movdqa (%rsi, %rcx), %xmm1 | |
1424 | movdqa (%rdi, %rcx), %xmm2 | |
1425 | movdqa %xmm2, %xmm4 | |
1426 | ||
f69190e7 | 1427 | # ifndef USE_SSSE3 |
7956a3d2 | 1428 | psrldq $9, %xmm3 |
0fda545d UD |
1429 | pslldq $7, %xmm2 |
1430 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1431 | # else |
0fda545d | 1432 | palignr $9, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1433 | # endif |
42e08a54 | 1434 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1435 | |
1436 | pcmpeqb %xmm1, %xmm0 | |
1437 | pcmpeqb %xmm2, %xmm1 | |
1438 | psubb %xmm0, %xmm1 | |
1439 | pmovmskb %xmm1, %edx | |
1440 | sub $0xffff, %edx | |
1441 | jnz LABEL(exit) | |
1442 | ||
e9f82e0d | 1443 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1444 | sub $16, %r11 |
1445 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1446 | # endif |
7956a3d2 L |
1447 | |
1448 | add $16, %rcx | |
1449 | movdqa %xmm4, %xmm3 /* store for next cycle */ | |
1450 | jmp LABEL(loop_ashr_9) | |
1451 | ||
1452 | .p2align 4 | |
1453 | LABEL(nibble_ashr_9): | |
1454 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1455 | pmovmskb %xmm0, %edx | |
1456 | test $0xfe00, %edx | |
1457 | jnz LABEL(ashr_9_exittail) | |
1458 | ||
e9f82e0d | 1459 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1460 | cmp $7, %r11 |
7956a3d2 | 1461 | jbe LABEL(ashr_9_exittail) |
f69190e7 | 1462 | # endif |
7956a3d2 L |
1463 | |
1464 | pxor %xmm0, %xmm0 | |
1465 | sub $0x1000, %r10 | |
1466 | jmp LABEL(gobble_ashr_9) | |
1467 | ||
1468 | .p2align 4 | |
1469 | LABEL(ashr_9_exittail): | |
1470 | movdqa (%rsi, %rcx), %xmm1 | |
1471 | psrldq $9, %xmm0 | |
1472 | psrldq $9, %xmm3 | |
1473 | jmp LABEL(aftertail) | |
1474 | ||
1475 | /* | |
1476 | * The following cases will be handled by ashr_10 | |
1477 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1478 | * n(6~15) n - 6 9(15 +(n - 6) - n) ashr_10 |
7956a3d2 L |
1479 | */ |
1480 | .p2align 4 | |
1481 | LABEL(ashr_10): | |
1482 | pxor %xmm0, %xmm0 | |
1483 | movdqa (%rdi), %xmm2 | |
1484 | movdqa (%rsi), %xmm1 | |
1485 | pcmpeqb %xmm1, %xmm0 | |
1486 | pslldq $6, %xmm2 | |
42e08a54 | 1487 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1488 | pcmpeqb %xmm1, %xmm2 |
1489 | psubb %xmm0, %xmm2 | |
1490 | pmovmskb %xmm2, %r9d | |
1491 | shr %cl, %edx | |
1492 | shr %cl, %r9d | |
1493 | sub %r9d, %edx | |
1494 | jnz LABEL(less32bytes) | |
1495 | movdqa (%rdi), %xmm3 | |
1496 | ||
1497 | UPDATE_STRNCMP_COUNTER | |
1498 | ||
1499 | pxor %xmm0, %xmm0 | |
1500 | mov $16, %rcx /* index for loads */ | |
1501 | mov $10, %r9d /* byte position left over from less32bytes case */ | |
1502 | /* | |
1503 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1504 | * When %r10 goes positive we have crossed a page boundary and | |
1505 | * need to do a nibble. | |
1506 | */ | |
1507 | lea 10(%rdi), %r10 | |
1508 | and $0xfff, %r10 /* offset into 4K page */ | |
1509 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1510 | ||
1511 | .p2align 4 | |
1512 | LABEL(loop_ashr_10): | |
1513 | add $16, %r10 | |
1514 | jg LABEL(nibble_ashr_10) | |
1515 | ||
1516 | LABEL(gobble_ashr_10): | |
1517 | movdqa (%rsi, %rcx), %xmm1 | |
1518 | movdqa (%rdi, %rcx), %xmm2 | |
1519 | movdqa %xmm2, %xmm4 | |
1520 | ||
f69190e7 | 1521 | # ifndef USE_SSSE3 |
7956a3d2 L |
1522 | psrldq $10, %xmm3 |
1523 | pslldq $6, %xmm2 | |
0fda545d | 1524 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1525 | # else |
0fda545d | 1526 | palignr $10, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1527 | # endif |
42e08a54 | 1528 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1529 | |
1530 | pcmpeqb %xmm1, %xmm0 | |
1531 | pcmpeqb %xmm2, %xmm1 | |
1532 | psubb %xmm0, %xmm1 | |
1533 | pmovmskb %xmm1, %edx | |
1534 | sub $0xffff, %edx | |
1535 | jnz LABEL(exit) | |
1536 | ||
e9f82e0d | 1537 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1538 | sub $16, %r11 |
1539 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1540 | # endif |
7956a3d2 L |
1541 | |
1542 | add $16, %rcx | |
1543 | movdqa %xmm4, %xmm3 | |
1544 | ||
1545 | add $16, %r10 | |
1546 | jg LABEL(nibble_ashr_10) /* cross page boundary */ | |
1547 | ||
1548 | movdqa (%rsi, %rcx), %xmm1 | |
1549 | movdqa (%rdi, %rcx), %xmm2 | |
1550 | movdqa %xmm2, %xmm4 | |
1551 | ||
f69190e7 | 1552 | # ifndef USE_SSSE3 |
7956a3d2 | 1553 | psrldq $10, %xmm3 |
0fda545d UD |
1554 | pslldq $6, %xmm2 |
1555 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1556 | # else |
0fda545d | 1557 | palignr $10, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1558 | # endif |
42e08a54 | 1559 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1560 | |
1561 | pcmpeqb %xmm1, %xmm0 | |
1562 | pcmpeqb %xmm2, %xmm1 | |
1563 | psubb %xmm0, %xmm1 | |
1564 | pmovmskb %xmm1, %edx | |
1565 | sub $0xffff, %edx | |
1566 | jnz LABEL(exit) | |
1567 | ||
e9f82e0d | 1568 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1569 | sub $16, %r11 |
1570 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1571 | # endif |
7956a3d2 L |
1572 | |
1573 | add $16, %rcx | |
1574 | movdqa %xmm4, %xmm3 | |
1575 | jmp LABEL(loop_ashr_10) | |
1576 | ||
1577 | .p2align 4 | |
1578 | LABEL(nibble_ashr_10): | |
1579 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1580 | pmovmskb %xmm0, %edx | |
1581 | test $0xfc00, %edx | |
1582 | jnz LABEL(ashr_10_exittail) | |
1583 | ||
e9f82e0d | 1584 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1585 | cmp $6, %r11 |
7956a3d2 | 1586 | jbe LABEL(ashr_10_exittail) |
f69190e7 | 1587 | # endif |
7956a3d2 L |
1588 | |
1589 | pxor %xmm0, %xmm0 | |
1590 | sub $0x1000, %r10 | |
1591 | jmp LABEL(gobble_ashr_10) | |
1592 | ||
1593 | .p2align 4 | |
1594 | LABEL(ashr_10_exittail): | |
1595 | movdqa (%rsi, %rcx), %xmm1 | |
1596 | psrldq $10, %xmm0 | |
1597 | psrldq $10, %xmm3 | |
1598 | jmp LABEL(aftertail) | |
1599 | ||
1600 | /* | |
1601 | * The following cases will be handled by ashr_11 | |
1602 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1603 | * n(5~15) n - 5 10(15 +(n - 5) - n) ashr_11 |
7956a3d2 L |
1604 | */ |
1605 | .p2align 4 | |
1606 | LABEL(ashr_11): | |
1607 | pxor %xmm0, %xmm0 | |
1608 | movdqa (%rdi), %xmm2 | |
1609 | movdqa (%rsi), %xmm1 | |
1610 | pcmpeqb %xmm1, %xmm0 | |
1611 | pslldq $5, %xmm2 | |
42e08a54 | 1612 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1613 | pcmpeqb %xmm1, %xmm2 |
1614 | psubb %xmm0, %xmm2 | |
1615 | pmovmskb %xmm2, %r9d | |
1616 | shr %cl, %edx | |
1617 | shr %cl, %r9d | |
1618 | sub %r9d, %edx | |
1619 | jnz LABEL(less32bytes) | |
1620 | movdqa (%rdi), %xmm3 | |
1621 | ||
1622 | UPDATE_STRNCMP_COUNTER | |
1623 | ||
1624 | pxor %xmm0, %xmm0 | |
1625 | mov $16, %rcx /* index for loads */ | |
1626 | mov $11, %r9d /* byte position left over from less32bytes case */ | |
1627 | /* | |
1628 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1629 | * When %r10 goes positive we have crossed a page boundary and | |
1630 | * need to do a nibble. | |
1631 | */ | |
1632 | lea 11(%rdi), %r10 | |
1633 | and $0xfff, %r10 /* offset into 4K page */ | |
1634 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1635 | ||
1636 | .p2align 4 | |
1637 | LABEL(loop_ashr_11): | |
1638 | add $16, %r10 | |
1639 | jg LABEL(nibble_ashr_11) | |
1640 | ||
1641 | LABEL(gobble_ashr_11): | |
1642 | movdqa (%rsi, %rcx), %xmm1 | |
1643 | movdqa (%rdi, %rcx), %xmm2 | |
1644 | movdqa %xmm2, %xmm4 | |
1645 | ||
f69190e7 | 1646 | # ifndef USE_SSSE3 |
7956a3d2 L |
1647 | psrldq $11, %xmm3 |
1648 | pslldq $5, %xmm2 | |
0fda545d | 1649 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1650 | # else |
0fda545d | 1651 | palignr $11, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1652 | # endif |
42e08a54 | 1653 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1654 | |
1655 | pcmpeqb %xmm1, %xmm0 | |
1656 | pcmpeqb %xmm2, %xmm1 | |
1657 | psubb %xmm0, %xmm1 | |
1658 | pmovmskb %xmm1, %edx | |
1659 | sub $0xffff, %edx | |
1660 | jnz LABEL(exit) | |
1661 | ||
e9f82e0d | 1662 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1663 | sub $16, %r11 |
1664 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1665 | # endif |
7956a3d2 L |
1666 | |
1667 | add $16, %rcx | |
1668 | movdqa %xmm4, %xmm3 | |
1669 | ||
1670 | add $16, %r10 | |
1671 | jg LABEL(nibble_ashr_11) /* cross page boundary */ | |
1672 | ||
1673 | movdqa (%rsi, %rcx), %xmm1 | |
1674 | movdqa (%rdi, %rcx), %xmm2 | |
1675 | movdqa %xmm2, %xmm4 | |
1676 | ||
f69190e7 | 1677 | # ifndef USE_SSSE3 |
7956a3d2 | 1678 | psrldq $11, %xmm3 |
0fda545d UD |
1679 | pslldq $5, %xmm2 |
1680 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1681 | # else |
0fda545d | 1682 | palignr $11, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1683 | # endif |
42e08a54 | 1684 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1685 | |
1686 | pcmpeqb %xmm1, %xmm0 | |
1687 | pcmpeqb %xmm2, %xmm1 | |
1688 | psubb %xmm0, %xmm1 | |
1689 | pmovmskb %xmm1, %edx | |
1690 | sub $0xffff, %edx | |
1691 | jnz LABEL(exit) | |
1692 | ||
e9f82e0d | 1693 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1694 | sub $16, %r11 |
1695 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1696 | # endif |
7956a3d2 L |
1697 | |
1698 | add $16, %rcx | |
1699 | movdqa %xmm4, %xmm3 | |
1700 | jmp LABEL(loop_ashr_11) | |
1701 | ||
1702 | .p2align 4 | |
1703 | LABEL(nibble_ashr_11): | |
1704 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1705 | pmovmskb %xmm0, %edx | |
1706 | test $0xf800, %edx | |
1707 | jnz LABEL(ashr_11_exittail) | |
1708 | ||
e9f82e0d | 1709 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1710 | cmp $5, %r11 |
7956a3d2 | 1711 | jbe LABEL(ashr_11_exittail) |
f69190e7 | 1712 | # endif |
7956a3d2 L |
1713 | |
1714 | pxor %xmm0, %xmm0 | |
1715 | sub $0x1000, %r10 | |
1716 | jmp LABEL(gobble_ashr_11) | |
1717 | ||
1718 | .p2align 4 | |
1719 | LABEL(ashr_11_exittail): | |
1720 | movdqa (%rsi, %rcx), %xmm1 | |
1721 | psrldq $11, %xmm0 | |
1722 | psrldq $11, %xmm3 | |
1723 | jmp LABEL(aftertail) | |
1724 | ||
1725 | /* | |
1726 | * The following cases will be handled by ashr_12 | |
1727 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1728 | * n(4~15) n - 4 11(15 +(n - 4) - n) ashr_12 |
7956a3d2 L |
1729 | */ |
1730 | .p2align 4 | |
1731 | LABEL(ashr_12): | |
1732 | pxor %xmm0, %xmm0 | |
1733 | movdqa (%rdi), %xmm2 | |
1734 | movdqa (%rsi), %xmm1 | |
1735 | pcmpeqb %xmm1, %xmm0 | |
1736 | pslldq $4, %xmm2 | |
42e08a54 | 1737 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1738 | pcmpeqb %xmm1, %xmm2 |
1739 | psubb %xmm0, %xmm2 | |
1740 | pmovmskb %xmm2, %r9d | |
1741 | shr %cl, %edx | |
1742 | shr %cl, %r9d | |
1743 | sub %r9d, %edx | |
1744 | jnz LABEL(less32bytes) | |
1745 | movdqa (%rdi), %xmm3 | |
1746 | ||
1747 | UPDATE_STRNCMP_COUNTER | |
1748 | ||
1749 | pxor %xmm0, %xmm0 | |
1750 | mov $16, %rcx /* index for loads */ | |
1751 | mov $12, %r9d /* byte position left over from less32bytes case */ | |
1752 | /* | |
1753 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1754 | * When %r10 goes positive we have crossed a page boundary and | |
1755 | * need to do a nibble. | |
1756 | */ | |
1757 | lea 12(%rdi), %r10 | |
1758 | and $0xfff, %r10 /* offset into 4K page */ | |
1759 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1760 | ||
1761 | .p2align 4 | |
1762 | LABEL(loop_ashr_12): | |
1763 | add $16, %r10 | |
1764 | jg LABEL(nibble_ashr_12) | |
1765 | ||
1766 | LABEL(gobble_ashr_12): | |
1767 | movdqa (%rsi, %rcx), %xmm1 | |
1768 | movdqa (%rdi, %rcx), %xmm2 | |
1769 | movdqa %xmm2, %xmm4 | |
1770 | ||
f69190e7 | 1771 | # ifndef USE_SSSE3 |
7956a3d2 L |
1772 | psrldq $12, %xmm3 |
1773 | pslldq $4, %xmm2 | |
0fda545d | 1774 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1775 | # else |
0fda545d | 1776 | palignr $12, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1777 | # endif |
42e08a54 | 1778 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1779 | |
1780 | pcmpeqb %xmm1, %xmm0 | |
1781 | pcmpeqb %xmm2, %xmm1 | |
1782 | psubb %xmm0, %xmm1 | |
1783 | pmovmskb %xmm1, %edx | |
1784 | sub $0xffff, %edx | |
1785 | jnz LABEL(exit) | |
1786 | ||
e9f82e0d | 1787 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1788 | sub $16, %r11 |
1789 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1790 | # endif |
7956a3d2 L |
1791 | |
1792 | add $16, %rcx | |
1793 | movdqa %xmm4, %xmm3 | |
1794 | ||
1795 | add $16, %r10 | |
1796 | jg LABEL(nibble_ashr_12) /* cross page boundary */ | |
1797 | ||
1798 | movdqa (%rsi, %rcx), %xmm1 | |
1799 | movdqa (%rdi, %rcx), %xmm2 | |
1800 | movdqa %xmm2, %xmm4 | |
1801 | ||
f69190e7 | 1802 | # ifndef USE_SSSE3 |
7956a3d2 | 1803 | psrldq $12, %xmm3 |
0fda545d UD |
1804 | pslldq $4, %xmm2 |
1805 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1806 | # else |
0fda545d | 1807 | palignr $12, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1808 | # endif |
42e08a54 | 1809 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1810 | |
1811 | pcmpeqb %xmm1, %xmm0 | |
1812 | pcmpeqb %xmm2, %xmm1 | |
1813 | psubb %xmm0, %xmm1 | |
1814 | pmovmskb %xmm1, %edx | |
1815 | sub $0xffff, %edx | |
1816 | jnz LABEL(exit) | |
1817 | ||
e9f82e0d | 1818 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1819 | sub $16, %r11 |
1820 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1821 | # endif |
7956a3d2 L |
1822 | |
1823 | add $16, %rcx | |
1824 | movdqa %xmm4, %xmm3 | |
1825 | jmp LABEL(loop_ashr_12) | |
1826 | ||
1827 | .p2align 4 | |
1828 | LABEL(nibble_ashr_12): | |
1829 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1830 | pmovmskb %xmm0, %edx | |
1831 | test $0xf000, %edx | |
1832 | jnz LABEL(ashr_12_exittail) | |
1833 | ||
e9f82e0d | 1834 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1835 | cmp $4, %r11 |
7956a3d2 | 1836 | jbe LABEL(ashr_12_exittail) |
f69190e7 | 1837 | # endif |
7956a3d2 L |
1838 | |
1839 | pxor %xmm0, %xmm0 | |
1840 | sub $0x1000, %r10 | |
1841 | jmp LABEL(gobble_ashr_12) | |
1842 | ||
1843 | .p2align 4 | |
1844 | LABEL(ashr_12_exittail): | |
1845 | movdqa (%rsi, %rcx), %xmm1 | |
1846 | psrldq $12, %xmm0 | |
1847 | psrldq $12, %xmm3 | |
1848 | jmp LABEL(aftertail) | |
1849 | ||
1850 | /* | |
1851 | * The following cases will be handled by ashr_13 | |
1852 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1853 | * n(3~15) n - 3 12(15 +(n - 3) - n) ashr_13 |
7956a3d2 L |
1854 | */ |
1855 | .p2align 4 | |
1856 | LABEL(ashr_13): | |
1857 | pxor %xmm0, %xmm0 | |
1858 | movdqa (%rdi), %xmm2 | |
1859 | movdqa (%rsi), %xmm1 | |
1860 | pcmpeqb %xmm1, %xmm0 | |
1861 | pslldq $3, %xmm2 | |
42e08a54 | 1862 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1863 | pcmpeqb %xmm1, %xmm2 |
1864 | psubb %xmm0, %xmm2 | |
1865 | pmovmskb %xmm2, %r9d | |
1866 | shr %cl, %edx | |
1867 | shr %cl, %r9d | |
1868 | sub %r9d, %edx | |
1869 | jnz LABEL(less32bytes) | |
1870 | movdqa (%rdi), %xmm3 | |
1871 | ||
1872 | UPDATE_STRNCMP_COUNTER | |
1873 | ||
1874 | pxor %xmm0, %xmm0 | |
1875 | mov $16, %rcx /* index for loads */ | |
1876 | mov $13, %r9d /* byte position left over from less32bytes case */ | |
1877 | /* | |
1878 | * Setup %r10 value allows us to detect crossing a page boundary. | |
1879 | * When %r10 goes positive we have crossed a page boundary and | |
1880 | * need to do a nibble. | |
1881 | */ | |
1882 | lea 13(%rdi), %r10 | |
1883 | and $0xfff, %r10 /* offset into 4K page */ | |
1884 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
1885 | ||
1886 | .p2align 4 | |
1887 | LABEL(loop_ashr_13): | |
1888 | add $16, %r10 | |
1889 | jg LABEL(nibble_ashr_13) | |
1890 | ||
1891 | LABEL(gobble_ashr_13): | |
1892 | movdqa (%rsi, %rcx), %xmm1 | |
1893 | movdqa (%rdi, %rcx), %xmm2 | |
1894 | movdqa %xmm2, %xmm4 | |
1895 | ||
f69190e7 | 1896 | # ifndef USE_SSSE3 |
7956a3d2 L |
1897 | psrldq $13, %xmm3 |
1898 | pslldq $3, %xmm2 | |
0fda545d | 1899 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1900 | # else |
0fda545d | 1901 | palignr $13, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1902 | # endif |
42e08a54 | 1903 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1904 | |
1905 | pcmpeqb %xmm1, %xmm0 | |
1906 | pcmpeqb %xmm2, %xmm1 | |
1907 | psubb %xmm0, %xmm1 | |
1908 | pmovmskb %xmm1, %edx | |
1909 | sub $0xffff, %edx | |
1910 | jnz LABEL(exit) | |
1911 | ||
e9f82e0d | 1912 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1913 | sub $16, %r11 |
1914 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1915 | # endif |
7956a3d2 L |
1916 | |
1917 | add $16, %rcx | |
1918 | movdqa %xmm4, %xmm3 | |
1919 | ||
1920 | add $16, %r10 | |
1921 | jg LABEL(nibble_ashr_13) /* cross page boundary */ | |
1922 | ||
1923 | movdqa (%rsi, %rcx), %xmm1 | |
1924 | movdqa (%rdi, %rcx), %xmm2 | |
1925 | movdqa %xmm2, %xmm4 | |
1926 | ||
f69190e7 | 1927 | # ifndef USE_SSSE3 |
7956a3d2 | 1928 | psrldq $13, %xmm3 |
0fda545d UD |
1929 | pslldq $3, %xmm2 |
1930 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 1931 | # else |
0fda545d | 1932 | palignr $13, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 1933 | # endif |
42e08a54 | 1934 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1935 | |
1936 | pcmpeqb %xmm1, %xmm0 | |
1937 | pcmpeqb %xmm2, %xmm1 | |
1938 | psubb %xmm0, %xmm1 | |
1939 | pmovmskb %xmm1, %edx | |
1940 | sub $0xffff, %edx | |
1941 | jnz LABEL(exit) | |
1942 | ||
e9f82e0d | 1943 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
1944 | sub $16, %r11 |
1945 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 1946 | # endif |
7956a3d2 L |
1947 | |
1948 | add $16, %rcx | |
1949 | movdqa %xmm4, %xmm3 | |
1950 | jmp LABEL(loop_ashr_13) | |
1951 | ||
1952 | .p2align 4 | |
1953 | LABEL(nibble_ashr_13): | |
1954 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
1955 | pmovmskb %xmm0, %edx | |
1956 | test $0xe000, %edx | |
1957 | jnz LABEL(ashr_13_exittail) | |
1958 | ||
e9f82e0d | 1959 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 1960 | cmp $3, %r11 |
7956a3d2 | 1961 | jbe LABEL(ashr_13_exittail) |
f69190e7 | 1962 | # endif |
7956a3d2 L |
1963 | |
1964 | pxor %xmm0, %xmm0 | |
1965 | sub $0x1000, %r10 | |
1966 | jmp LABEL(gobble_ashr_13) | |
1967 | ||
1968 | .p2align 4 | |
1969 | LABEL(ashr_13_exittail): | |
1970 | movdqa (%rsi, %rcx), %xmm1 | |
1971 | psrldq $13, %xmm0 | |
1972 | psrldq $13, %xmm3 | |
1973 | jmp LABEL(aftertail) | |
1974 | ||
1975 | /* | |
1976 | * The following cases will be handled by ashr_14 | |
1977 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 1978 | * n(2~15) n - 2 13(15 +(n - 2) - n) ashr_14 |
7956a3d2 L |
1979 | */ |
1980 | .p2align 4 | |
1981 | LABEL(ashr_14): | |
1982 | pxor %xmm0, %xmm0 | |
1983 | movdqa (%rdi), %xmm2 | |
1984 | movdqa (%rsi), %xmm1 | |
1985 | pcmpeqb %xmm1, %xmm0 | |
1986 | pslldq $2, %xmm2 | |
42e08a54 | 1987 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
1988 | pcmpeqb %xmm1, %xmm2 |
1989 | psubb %xmm0, %xmm2 | |
1990 | pmovmskb %xmm2, %r9d | |
1991 | shr %cl, %edx | |
1992 | shr %cl, %r9d | |
1993 | sub %r9d, %edx | |
1994 | jnz LABEL(less32bytes) | |
1995 | movdqa (%rdi), %xmm3 | |
1996 | ||
1997 | UPDATE_STRNCMP_COUNTER | |
1998 | ||
1999 | pxor %xmm0, %xmm0 | |
2000 | mov $16, %rcx /* index for loads */ | |
2001 | mov $14, %r9d /* byte position left over from less32bytes case */ | |
2002 | /* | |
2003 | * Setup %r10 value allows us to detect crossing a page boundary. | |
2004 | * When %r10 goes positive we have crossed a page boundary and | |
2005 | * need to do a nibble. | |
2006 | */ | |
2007 | lea 14(%rdi), %r10 | |
2008 | and $0xfff, %r10 /* offset into 4K page */ | |
2009 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
2010 | ||
2011 | .p2align 4 | |
2012 | LABEL(loop_ashr_14): | |
2013 | add $16, %r10 | |
2014 | jg LABEL(nibble_ashr_14) | |
2015 | ||
2016 | LABEL(gobble_ashr_14): | |
2017 | movdqa (%rsi, %rcx), %xmm1 | |
2018 | movdqa (%rdi, %rcx), %xmm2 | |
2019 | movdqa %xmm2, %xmm4 | |
2020 | ||
f69190e7 | 2021 | # ifndef USE_SSSE3 |
7956a3d2 L |
2022 | psrldq $14, %xmm3 |
2023 | pslldq $2, %xmm2 | |
0fda545d | 2024 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 2025 | # else |
0fda545d | 2026 | palignr $14, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 2027 | # endif |
42e08a54 | 2028 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
2029 | |
2030 | pcmpeqb %xmm1, %xmm0 | |
2031 | pcmpeqb %xmm2, %xmm1 | |
2032 | psubb %xmm0, %xmm1 | |
2033 | pmovmskb %xmm1, %edx | |
2034 | sub $0xffff, %edx | |
2035 | jnz LABEL(exit) | |
2036 | ||
e9f82e0d | 2037 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
2038 | sub $16, %r11 |
2039 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 2040 | # endif |
7956a3d2 L |
2041 | |
2042 | add $16, %rcx | |
2043 | movdqa %xmm4, %xmm3 | |
2044 | ||
2045 | add $16, %r10 | |
2046 | jg LABEL(nibble_ashr_14) /* cross page boundary */ | |
2047 | ||
2048 | movdqa (%rsi, %rcx), %xmm1 | |
2049 | movdqa (%rdi, %rcx), %xmm2 | |
2050 | movdqa %xmm2, %xmm4 | |
2051 | ||
f69190e7 | 2052 | # ifndef USE_SSSE3 |
7956a3d2 | 2053 | psrldq $14, %xmm3 |
0fda545d UD |
2054 | pslldq $2, %xmm2 |
2055 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 2056 | # else |
0fda545d | 2057 | palignr $14, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 2058 | # endif |
42e08a54 | 2059 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
2060 | |
2061 | pcmpeqb %xmm1, %xmm0 | |
2062 | pcmpeqb %xmm2, %xmm1 | |
2063 | psubb %xmm0, %xmm1 | |
2064 | pmovmskb %xmm1, %edx | |
2065 | sub $0xffff, %edx | |
2066 | jnz LABEL(exit) | |
2067 | ||
e9f82e0d | 2068 | # if defined USE_AS_STRNCMP | defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
2069 | sub $16, %r11 |
2070 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 2071 | # endif |
7956a3d2 L |
2072 | |
2073 | add $16, %rcx | |
2074 | movdqa %xmm4, %xmm3 | |
2075 | jmp LABEL(loop_ashr_14) | |
2076 | ||
2077 | .p2align 4 | |
2078 | LABEL(nibble_ashr_14): | |
2079 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
2080 | pmovmskb %xmm0, %edx | |
2081 | test $0xc000, %edx | |
2082 | jnz LABEL(ashr_14_exittail) | |
2083 | ||
e9f82e0d | 2084 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 | 2085 | cmp $2, %r11 |
7956a3d2 | 2086 | jbe LABEL(ashr_14_exittail) |
f69190e7 | 2087 | # endif |
7956a3d2 L |
2088 | |
2089 | pxor %xmm0, %xmm0 | |
2090 | sub $0x1000, %r10 | |
2091 | jmp LABEL(gobble_ashr_14) | |
2092 | ||
2093 | .p2align 4 | |
2094 | LABEL(ashr_14_exittail): | |
2095 | movdqa (%rsi, %rcx), %xmm1 | |
2096 | psrldq $14, %xmm0 | |
2097 | psrldq $14, %xmm3 | |
2098 | jmp LABEL(aftertail) | |
2099 | ||
2100 | /* | |
2101 | * The following cases will be handled by ashr_15 | |
2102 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case | |
f69190e7 | 2103 | * n(1~15) n - 1 14(15 +(n - 1) - n) ashr_15 |
7956a3d2 L |
2104 | */ |
2105 | .p2align 4 | |
2106 | LABEL(ashr_15): | |
2107 | pxor %xmm0, %xmm0 | |
2108 | movdqa (%rdi), %xmm2 | |
2109 | movdqa (%rsi), %xmm1 | |
2110 | pcmpeqb %xmm1, %xmm0 | |
2111 | pslldq $1, %xmm2 | |
42e08a54 | 2112 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
2113 | pcmpeqb %xmm1, %xmm2 |
2114 | psubb %xmm0, %xmm2 | |
2115 | pmovmskb %xmm2, %r9d | |
2116 | shr %cl, %edx | |
2117 | shr %cl, %r9d | |
2118 | sub %r9d, %edx | |
2119 | jnz LABEL(less32bytes) | |
2120 | ||
2121 | movdqa (%rdi), %xmm3 | |
2122 | ||
2123 | UPDATE_STRNCMP_COUNTER | |
2124 | ||
2125 | pxor %xmm0, %xmm0 | |
2126 | mov $16, %rcx /* index for loads */ | |
2127 | mov $15, %r9d /* byte position left over from less32bytes case */ | |
2128 | /* | |
2129 | * Setup %r10 value allows us to detect crossing a page boundary. | |
2130 | * When %r10 goes positive we have crossed a page boundary and | |
2131 | * need to do a nibble. | |
2132 | */ | |
2133 | lea 15(%rdi), %r10 | |
2134 | and $0xfff, %r10 /* offset into 4K page */ | |
2135 | ||
2136 | sub $0x1000, %r10 /* subtract 4K pagesize */ | |
2137 | ||
2138 | .p2align 4 | |
2139 | LABEL(loop_ashr_15): | |
2140 | add $16, %r10 | |
2141 | jg LABEL(nibble_ashr_15) | |
2142 | ||
2143 | LABEL(gobble_ashr_15): | |
2144 | movdqa (%rsi, %rcx), %xmm1 | |
2145 | movdqa (%rdi, %rcx), %xmm2 | |
2146 | movdqa %xmm2, %xmm4 | |
2147 | ||
f69190e7 | 2148 | # ifndef USE_SSSE3 |
7956a3d2 L |
2149 | psrldq $15, %xmm3 |
2150 | pslldq $1, %xmm2 | |
0fda545d | 2151 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 2152 | # else |
0fda545d | 2153 | palignr $15, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 2154 | # endif |
42e08a54 | 2155 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
2156 | |
2157 | pcmpeqb %xmm1, %xmm0 | |
2158 | pcmpeqb %xmm2, %xmm1 | |
2159 | psubb %xmm0, %xmm1 | |
2160 | pmovmskb %xmm1, %edx | |
2161 | sub $0xffff, %edx | |
2162 | jnz LABEL(exit) | |
2163 | ||
e9f82e0d | 2164 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
2165 | sub $16, %r11 |
2166 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 2167 | # endif |
7956a3d2 L |
2168 | |
2169 | add $16, %rcx | |
2170 | movdqa %xmm4, %xmm3 | |
2171 | ||
2172 | add $16, %r10 | |
2173 | jg LABEL(nibble_ashr_15) /* cross page boundary */ | |
2174 | ||
2175 | movdqa (%rsi, %rcx), %xmm1 | |
2176 | movdqa (%rdi, %rcx), %xmm2 | |
2177 | movdqa %xmm2, %xmm4 | |
2178 | ||
f69190e7 | 2179 | # ifndef USE_SSSE3 |
7956a3d2 | 2180 | psrldq $15, %xmm3 |
0fda545d UD |
2181 | pslldq $1, %xmm2 |
2182 | por %xmm3, %xmm2 /* merge into one 16byte value */ | |
f69190e7 | 2183 | # else |
0fda545d | 2184 | palignr $15, %xmm3, %xmm2 /* merge into one 16byte value */ |
f69190e7 | 2185 | # endif |
42e08a54 | 2186 | TOLOWER (%xmm1, %xmm2) |
7956a3d2 L |
2187 | |
2188 | pcmpeqb %xmm1, %xmm0 | |
2189 | pcmpeqb %xmm2, %xmm1 | |
2190 | psubb %xmm0, %xmm1 | |
2191 | pmovmskb %xmm1, %edx | |
2192 | sub $0xffff, %edx | |
2193 | jnz LABEL(exit) | |
2194 | ||
e9f82e0d | 2195 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
2196 | sub $16, %r11 |
2197 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 2198 | # endif |
7956a3d2 L |
2199 | |
2200 | add $16, %rcx | |
2201 | movdqa %xmm4, %xmm3 | |
2202 | jmp LABEL(loop_ashr_15) | |
2203 | ||
2204 | .p2align 4 | |
2205 | LABEL(nibble_ashr_15): | |
2206 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ | |
2207 | pmovmskb %xmm0, %edx | |
2208 | test $0x8000, %edx | |
2209 | jnz LABEL(ashr_15_exittail) | |
2210 | ||
e9f82e0d | 2211 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
45db99c7 UD |
2212 | cmpq $1, %r11 |
2213 | jbe LABEL(ashr_15_exittail) | |
f69190e7 | 2214 | # endif |
7956a3d2 L |
2215 | |
2216 | pxor %xmm0, %xmm0 | |
2217 | sub $0x1000, %r10 | |
2218 | jmp LABEL(gobble_ashr_15) | |
2219 | ||
2220 | .p2align 4 | |
2221 | LABEL(ashr_15_exittail): | |
2222 | movdqa (%rsi, %rcx), %xmm1 | |
2223 | psrldq $15, %xmm3 | |
2224 | psrldq $15, %xmm0 | |
2225 | ||
2226 | .p2align 4 | |
2227 | LABEL(aftertail): | |
42e08a54 | 2228 | TOLOWER (%xmm1, %xmm3) |
7956a3d2 L |
2229 | pcmpeqb %xmm3, %xmm1 |
2230 | psubb %xmm0, %xmm1 | |
2231 | pmovmskb %xmm1, %edx | |
2232 | not %edx | |
2233 | ||
2234 | .p2align 4 | |
2235 | LABEL(exit): | |
2236 | lea -16(%r9, %rcx), %rax /* locate the exact offset for rdi */ | |
2237 | LABEL(less32bytes): | |
2238 | lea (%rdi, %rax), %rdi /* locate the exact address for first operand(rdi) */ | |
2239 | lea (%rsi, %rcx), %rsi /* locate the exact address for second operand(rsi) */ | |
2240 | test %r8d, %r8d | |
2241 | jz LABEL(ret) | |
2242 | xchg %rsi, %rdi /* recover original order according to flag(%r8d) */ | |
2243 | ||
2244 | .p2align 4 | |
2245 | LABEL(ret): | |
2246 | LABEL(less16bytes): | |
2247 | bsf %rdx, %rdx /* find and store bit index in %rdx */ | |
2248 | ||
e9f82e0d | 2249 | # if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
7956a3d2 L |
2250 | sub %rdx, %r11 |
2251 | jbe LABEL(strcmp_exitz) | |
f69190e7 | 2252 | # endif |
7956a3d2 L |
2253 | movzbl (%rsi, %rdx), %ecx |
2254 | movzbl (%rdi, %rdx), %eax | |
2255 | ||
e9f82e0d | 2256 | # if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
42e08a54 UD |
2257 | leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rdx |
2258 | movl (%rdx,%rcx,4), %ecx | |
2259 | movl (%rdx,%rax,4), %eax | |
2260 | # endif | |
2261 | ||
7956a3d2 L |
2262 | sub %ecx, %eax |
2263 | ret | |
2264 | ||
2265 | LABEL(strcmp_exitz): | |
2266 | xor %eax, %eax | |
2267 | ret | |
2268 | ||
2269 | .p2align 4 | |
2270 | LABEL(Byte0): | |
2271 | movzx (%rsi), %ecx | |
2272 | movzx (%rdi), %eax | |
2273 | ||
e9f82e0d UD |
2274 | # if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
2275 | leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rdx | |
2276 | movl (%rdx,%rcx,4), %ecx | |
2277 | movl (%rdx,%rax,4), %eax | |
2278 | # endif | |
2279 | ||
7956a3d2 L |
2280 | sub %ecx, %eax |
2281 | ret | |
29691210 | 2282 | END (STRCMP) |
7956a3d2 L |
2283 | |
2284 | .section .rodata,"a",@progbits | |
2285 | .p2align 3 | |
2286 | LABEL(unaligned_table): | |
2287 | .int LABEL(ashr_1) - LABEL(unaligned_table) | |
2288 | .int LABEL(ashr_2) - LABEL(unaligned_table) | |
2289 | .int LABEL(ashr_3) - LABEL(unaligned_table) | |
2290 | .int LABEL(ashr_4) - LABEL(unaligned_table) | |
2291 | .int LABEL(ashr_5) - LABEL(unaligned_table) | |
2292 | .int LABEL(ashr_6) - LABEL(unaligned_table) | |
2293 | .int LABEL(ashr_7) - LABEL(unaligned_table) | |
2294 | .int LABEL(ashr_8) - LABEL(unaligned_table) | |
2295 | .int LABEL(ashr_9) - LABEL(unaligned_table) | |
2296 | .int LABEL(ashr_10) - LABEL(unaligned_table) | |
2297 | .int LABEL(ashr_11) - LABEL(unaligned_table) | |
2298 | .int LABEL(ashr_12) - LABEL(unaligned_table) | |
2299 | .int LABEL(ashr_13) - LABEL(unaligned_table) | |
2300 | .int LABEL(ashr_14) - LABEL(unaligned_table) | |
2301 | .int LABEL(ashr_15) - LABEL(unaligned_table) | |
2302 | .int LABEL(ashr_0) - LABEL(unaligned_table) | |
4f41c682 | 2303 | #endif /* !IS_IN (libc) */ |
7956a3d2 | 2304 | libc_hidden_builtin_def (STRCMP) |