| 1 | /* Highly optimized version for x86-64. |
| 2 | Copyright (C) 1999-2020 Free Software Foundation, Inc. |
| 3 | This file is part of the GNU C Library. |
| 4 | Based on i686 version contributed by Ulrich Drepper |
| 5 | <drepper@cygnus.com>, 1999. |
| 6 | Updated with SSE2 support contributed by Intel Corporation. |
| 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 |
| 19 | License along with the GNU C Library; if not, see |
| 20 | <https://www.gnu.org/licenses/>. */ |
| 21 | |
| 22 | #include <sysdep.h> |
| 23 | #include "asm-syntax.h" |
| 24 | |
| 25 | #undef UPDATE_STRNCMP_COUNTER |
| 26 | |
| 27 | #ifndef LABEL |
| 28 | #define LABEL(l) L(l) |
| 29 | #endif |
| 30 | |
| 31 | #ifdef USE_AS_STRNCMP |
| 32 | /* Since the counter, %r11, is unsigned, we branch to strcmp_exitz |
| 33 | if the new counter > the old one or is 0. */ |
| 34 | # define UPDATE_STRNCMP_COUNTER \ |
| 35 | /* calculate left number to compare */ \ |
| 36 | lea -16(%rcx, %r11), %r9; \ |
| 37 | cmp %r9, %r11; \ |
| 38 | jb LABEL(strcmp_exitz); \ |
| 39 | test %r9, %r9; \ |
| 40 | je LABEL(strcmp_exitz); \ |
| 41 | mov %r9, %r11 |
| 42 | |
| 43 | #elif defined USE_AS_STRCASECMP_L |
| 44 | # include "locale-defines.h" |
| 45 | |
| 46 | # define UPDATE_STRNCMP_COUNTER |
| 47 | #elif defined USE_AS_STRNCASECMP_L |
| 48 | # include "locale-defines.h" |
| 49 | |
| 50 | # define UPDATE_STRNCMP_COUNTER \ |
| 51 | /* calculate left number to compare */ \ |
| 52 | lea -16(%rcx, %r11), %r9; \ |
| 53 | cmp %r9, %r11; \ |
| 54 | jb LABEL(strcmp_exitz); \ |
| 55 | test %r9, %r9; \ |
| 56 | je LABEL(strcmp_exitz); \ |
| 57 | mov %r9, %r11 |
| 58 | #else |
| 59 | # define UPDATE_STRNCMP_COUNTER |
| 60 | # ifndef STRCMP |
| 61 | # define STRCMP strcmp |
| 62 | # endif |
| 63 | #endif |
| 64 | |
| 65 | #ifndef USE_SSSE3 |
| 66 | .text |
| 67 | #else |
| 68 | .section .text.ssse3,"ax" ,@progbits |
| 69 | #endif |
| 70 | |
| 71 | #ifdef USE_AS_STRCASECMP_L |
| 72 | # ifndef ENTRY2 |
| 73 | # define ENTRY2(name) ENTRY (name) |
| 74 | # define END2(name) END (name) |
| 75 | # endif |
| 76 | |
| 77 | ENTRY2 (__strcasecmp) |
| 78 | movq __libc_tsd_LOCALE@gottpoff(%rip),%rax |
| 79 | mov %fs:(%rax),%RDX_LP |
| 80 | |
| 81 | // XXX 5 byte should be before the function |
| 82 | /* 5-byte NOP. */ |
| 83 | .byte 0x0f,0x1f,0x44,0x00,0x00 |
| 84 | END2 (__strcasecmp) |
| 85 | # ifndef NO_NOLOCALE_ALIAS |
| 86 | weak_alias (__strcasecmp, strcasecmp) |
| 87 | libc_hidden_def (__strcasecmp) |
| 88 | # endif |
| 89 | /* FALLTHROUGH to strcasecmp_l. */ |
| 90 | #elif defined USE_AS_STRNCASECMP_L |
| 91 | # ifndef ENTRY2 |
| 92 | # define ENTRY2(name) ENTRY (name) |
| 93 | # define END2(name) END (name) |
| 94 | # endif |
| 95 | |
| 96 | ENTRY2 (__strncasecmp) |
| 97 | movq __libc_tsd_LOCALE@gottpoff(%rip),%rax |
| 98 | mov %fs:(%rax),%RCX_LP |
| 99 | |
| 100 | // XXX 5 byte should be before the function |
| 101 | /* 5-byte NOP. */ |
| 102 | .byte 0x0f,0x1f,0x44,0x00,0x00 |
| 103 | END2 (__strncasecmp) |
| 104 | # ifndef NO_NOLOCALE_ALIAS |
| 105 | weak_alias (__strncasecmp, strncasecmp) |
| 106 | libc_hidden_def (__strncasecmp) |
| 107 | # endif |
| 108 | /* FALLTHROUGH to strncasecmp_l. */ |
| 109 | #endif |
| 110 | |
| 111 | ENTRY (STRCMP) |
| 112 | #ifdef USE_AS_STRCASECMP_L |
| 113 | /* We have to fall back on the C implementation for locales |
| 114 | with encodings not matching ASCII for single bytes. */ |
| 115 | # if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0 |
| 116 | mov LOCALE_T___LOCALES+LC_CTYPE*LP_SIZE(%rdx), %RAX_LP |
| 117 | # else |
| 118 | mov (%rdx), %RAX_LP |
| 119 | # endif |
| 120 | testl $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%rax) |
| 121 | jne __strcasecmp_l_nonascii |
| 122 | #elif defined USE_AS_STRNCASECMP_L |
| 123 | /* We have to fall back on the C implementation for locales |
| 124 | with encodings not matching ASCII for single bytes. */ |
| 125 | # if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0 |
| 126 | mov LOCALE_T___LOCALES+LC_CTYPE*LP_SIZE(%rcx), %RAX_LP |
| 127 | # else |
| 128 | mov (%rcx), %RAX_LP |
| 129 | # endif |
| 130 | testl $1, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%rax) |
| 131 | jne __strncasecmp_l_nonascii |
| 132 | #endif |
| 133 | |
| 134 | /* |
| 135 | * This implementation uses SSE to compare up to 16 bytes at a time. |
| 136 | */ |
| 137 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 138 | test %RDX_LP, %RDX_LP |
| 139 | je LABEL(strcmp_exitz) |
| 140 | cmp $1, %RDX_LP |
| 141 | je LABEL(Byte0) |
| 142 | mov %RDX_LP, %R11_LP |
| 143 | #endif |
| 144 | mov %esi, %ecx |
| 145 | mov %edi, %eax |
| 146 | /* Use 64bit AND here to avoid long NOP padding. */ |
| 147 | and $0x3f, %rcx /* rsi alignment in cache line */ |
| 148 | and $0x3f, %rax /* rdi alignment in cache line */ |
| 149 | #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
| 150 | .section .rodata.cst16,"aM" ,@progbits,16 |
| 151 | .align 16 |
| 152 | .Lbelowupper: |
| 153 | .quad 0x4040404040404040 |
| 154 | .quad 0x4040404040404040 |
| 155 | .Ltopupper: |
| 156 | .quad 0x5b5b5b5b5b5b5b5b |
| 157 | .quad 0x5b5b5b5b5b5b5b5b |
| 158 | .Ltouppermask: |
| 159 | .quad 0x2020202020202020 |
| 160 | .quad 0x2020202020202020 |
| 161 | .previous |
| 162 | movdqa .Lbelowupper(%rip), %xmm5 |
| 163 | # define UCLOW_reg %xmm5 |
| 164 | movdqa .Ltopupper(%rip), %xmm6 |
| 165 | # define UCHIGH_reg %xmm6 |
| 166 | movdqa .Ltouppermask(%rip), %xmm7 |
| 167 | # define LCQWORD_reg %xmm7 |
| 168 | #endif |
| 169 | cmp $0x30, %ecx |
| 170 | ja LABEL(crosscache) /* rsi: 16-byte load will cross cache line */ |
| 171 | cmp $0x30, %eax |
| 172 | ja LABEL(crosscache) /* rdi: 16-byte load will cross cache line */ |
| 173 | movlpd (%rdi), %xmm1 |
| 174 | movlpd (%rsi), %xmm2 |
| 175 | movhpd 8(%rdi), %xmm1 |
| 176 | movhpd 8(%rsi), %xmm2 |
| 177 | #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
| 178 | # define TOLOWER(reg1, reg2) \ |
| 179 | movdqa reg1, %xmm8; \ |
| 180 | movdqa UCHIGH_reg, %xmm9; \ |
| 181 | movdqa reg2, %xmm10; \ |
| 182 | movdqa UCHIGH_reg, %xmm11; \ |
| 183 | pcmpgtb UCLOW_reg, %xmm8; \ |
| 184 | pcmpgtb reg1, %xmm9; \ |
| 185 | pcmpgtb UCLOW_reg, %xmm10; \ |
| 186 | pcmpgtb reg2, %xmm11; \ |
| 187 | pand %xmm9, %xmm8; \ |
| 188 | pand %xmm11, %xmm10; \ |
| 189 | pand LCQWORD_reg, %xmm8; \ |
| 190 | pand LCQWORD_reg, %xmm10; \ |
| 191 | por %xmm8, reg1; \ |
| 192 | por %xmm10, reg2 |
| 193 | TOLOWER (%xmm1, %xmm2) |
| 194 | #else |
| 195 | # define TOLOWER(reg1, reg2) |
| 196 | #endif |
| 197 | pxor %xmm0, %xmm0 /* clear %xmm0 for null char checks */ |
| 198 | pcmpeqb %xmm1, %xmm0 /* Any null chars? */ |
| 199 | pcmpeqb %xmm2, %xmm1 /* compare first 16 bytes for equality */ |
| 200 | psubb %xmm0, %xmm1 /* packed sub of comparison results*/ |
| 201 | pmovmskb %xmm1, %edx |
| 202 | sub $0xffff, %edx /* if first 16 bytes are same, edx == 0xffff */ |
| 203 | jnz LABEL(less16bytes) /* If not, find different value or null char */ |
| 204 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 205 | sub $16, %r11 |
| 206 | jbe LABEL(strcmp_exitz) /* finish comparision */ |
| 207 | #endif |
| 208 | add $16, %rsi /* prepare to search next 16 bytes */ |
| 209 | add $16, %rdi /* prepare to search next 16 bytes */ |
| 210 | |
| 211 | /* |
| 212 | * Determine source and destination string offsets from 16-byte alignment. |
| 213 | * Use relative offset difference between the two to determine which case |
| 214 | * below to use. |
| 215 | */ |
| 216 | .p2align 4 |
| 217 | LABEL(crosscache): |
| 218 | and $0xfffffffffffffff0, %rsi /* force %rsi is 16 byte aligned */ |
| 219 | and $0xfffffffffffffff0, %rdi /* force %rdi is 16 byte aligned */ |
| 220 | mov $0xffff, %edx /* for equivalent offset */ |
| 221 | xor %r8d, %r8d |
| 222 | and $0xf, %ecx /* offset of rsi */ |
| 223 | and $0xf, %eax /* offset of rdi */ |
| 224 | cmp %eax, %ecx |
| 225 | je LABEL(ashr_0) /* rsi and rdi relative offset same */ |
| 226 | ja LABEL(bigger) |
| 227 | mov %edx, %r8d /* r8d is offset flag for exit tail */ |
| 228 | xchg %ecx, %eax |
| 229 | xchg %rsi, %rdi |
| 230 | LABEL(bigger): |
| 231 | lea 15(%rax), %r9 |
| 232 | sub %rcx, %r9 |
| 233 | lea LABEL(unaligned_table)(%rip), %r10 |
| 234 | movslq (%r10, %r9,4), %r9 |
| 235 | lea (%r10, %r9), %r10 |
| 236 | _CET_NOTRACK jmp *%r10 /* jump to corresponding case */ |
| 237 | |
| 238 | /* |
| 239 | * The following cases will be handled by ashr_0 |
| 240 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 241 | * n(0~15) n(0~15) 15(15+ n-n) ashr_0 |
| 242 | */ |
| 243 | .p2align 4 |
| 244 | LABEL(ashr_0): |
| 245 | |
| 246 | movdqa (%rsi), %xmm1 |
| 247 | pxor %xmm0, %xmm0 /* clear %xmm0 for null char check */ |
| 248 | pcmpeqb %xmm1, %xmm0 /* Any null chars? */ |
| 249 | #if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L |
| 250 | pcmpeqb (%rdi), %xmm1 /* compare 16 bytes for equality */ |
| 251 | #else |
| 252 | movdqa (%rdi), %xmm2 |
| 253 | TOLOWER (%xmm1, %xmm2) |
| 254 | pcmpeqb %xmm2, %xmm1 /* compare 16 bytes for equality */ |
| 255 | #endif |
| 256 | psubb %xmm0, %xmm1 /* packed sub of comparison results*/ |
| 257 | pmovmskb %xmm1, %r9d |
| 258 | shr %cl, %edx /* adjust 0xffff for offset */ |
| 259 | shr %cl, %r9d /* adjust for 16-byte offset */ |
| 260 | sub %r9d, %edx |
| 261 | /* |
| 262 | * edx must be the same with r9d if in left byte (16-rcx) is equal to |
| 263 | * the start from (16-rax) and no null char was seen. |
| 264 | */ |
| 265 | jne LABEL(less32bytes) /* mismatch or null char */ |
| 266 | UPDATE_STRNCMP_COUNTER |
| 267 | mov $16, %rcx |
| 268 | mov $16, %r9 |
| 269 | pxor %xmm0, %xmm0 /* clear xmm0, may have changed above */ |
| 270 | |
| 271 | /* |
| 272 | * Now both strings are aligned at 16-byte boundary. Loop over strings |
| 273 | * checking 32-bytes per iteration. |
| 274 | */ |
| 275 | .p2align 4 |
| 276 | LABEL(loop_ashr_0): |
| 277 | movdqa (%rsi, %rcx), %xmm1 |
| 278 | movdqa (%rdi, %rcx), %xmm2 |
| 279 | TOLOWER (%xmm1, %xmm2) |
| 280 | |
| 281 | pcmpeqb %xmm1, %xmm0 |
| 282 | pcmpeqb %xmm2, %xmm1 |
| 283 | psubb %xmm0, %xmm1 |
| 284 | pmovmskb %xmm1, %edx |
| 285 | sub $0xffff, %edx |
| 286 | jnz LABEL(exit) /* mismatch or null char seen */ |
| 287 | |
| 288 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 289 | sub $16, %r11 |
| 290 | jbe LABEL(strcmp_exitz) |
| 291 | #endif |
| 292 | add $16, %rcx |
| 293 | movdqa (%rsi, %rcx), %xmm1 |
| 294 | movdqa (%rdi, %rcx), %xmm2 |
| 295 | TOLOWER (%xmm1, %xmm2) |
| 296 | |
| 297 | pcmpeqb %xmm1, %xmm0 |
| 298 | pcmpeqb %xmm2, %xmm1 |
| 299 | psubb %xmm0, %xmm1 |
| 300 | pmovmskb %xmm1, %edx |
| 301 | sub $0xffff, %edx |
| 302 | jnz LABEL(exit) |
| 303 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 304 | sub $16, %r11 |
| 305 | jbe LABEL(strcmp_exitz) |
| 306 | #endif |
| 307 | add $16, %rcx |
| 308 | jmp LABEL(loop_ashr_0) |
| 309 | |
| 310 | /* |
| 311 | * The following cases will be handled by ashr_1 |
| 312 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 313 | * n(15) n -15 0(15 +(n-15) - n) ashr_1 |
| 314 | */ |
| 315 | .p2align 4 |
| 316 | LABEL(ashr_1): |
| 317 | pxor %xmm0, %xmm0 |
| 318 | movdqa (%rdi), %xmm2 |
| 319 | movdqa (%rsi), %xmm1 |
| 320 | pcmpeqb %xmm1, %xmm0 /* Any null chars? */ |
| 321 | pslldq $15, %xmm2 /* shift first string to align with second */ |
| 322 | TOLOWER (%xmm1, %xmm2) |
| 323 | pcmpeqb %xmm1, %xmm2 /* compare 16 bytes for equality */ |
| 324 | psubb %xmm0, %xmm2 /* packed sub of comparison results*/ |
| 325 | pmovmskb %xmm2, %r9d |
| 326 | shr %cl, %edx /* adjust 0xffff for offset */ |
| 327 | shr %cl, %r9d /* adjust for 16-byte offset */ |
| 328 | sub %r9d, %edx |
| 329 | jnz LABEL(less32bytes) /* mismatch or null char seen */ |
| 330 | movdqa (%rdi), %xmm3 |
| 331 | UPDATE_STRNCMP_COUNTER |
| 332 | |
| 333 | pxor %xmm0, %xmm0 |
| 334 | mov $16, %rcx /* index for loads*/ |
| 335 | mov $1, %r9d /* byte position left over from less32bytes case */ |
| 336 | /* |
| 337 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 338 | * When %r10 goes positive we have crossed a page boundary and |
| 339 | * need to do a nibble. |
| 340 | */ |
| 341 | lea 1(%rdi), %r10 |
| 342 | and $0xfff, %r10 /* offset into 4K page */ |
| 343 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 344 | |
| 345 | .p2align 4 |
| 346 | LABEL(loop_ashr_1): |
| 347 | add $16, %r10 |
| 348 | jg LABEL(nibble_ashr_1) /* cross page boundary */ |
| 349 | |
| 350 | LABEL(gobble_ashr_1): |
| 351 | movdqa (%rsi, %rcx), %xmm1 |
| 352 | movdqa (%rdi, %rcx), %xmm2 |
| 353 | movdqa %xmm2, %xmm4 /* store for next cycle */ |
| 354 | |
| 355 | #ifndef USE_SSSE3 |
| 356 | psrldq $1, %xmm3 |
| 357 | pslldq $15, %xmm2 |
| 358 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 359 | #else |
| 360 | palignr $1, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 361 | #endif |
| 362 | TOLOWER (%xmm1, %xmm2) |
| 363 | |
| 364 | pcmpeqb %xmm1, %xmm0 |
| 365 | pcmpeqb %xmm2, %xmm1 |
| 366 | psubb %xmm0, %xmm1 |
| 367 | pmovmskb %xmm1, %edx |
| 368 | sub $0xffff, %edx |
| 369 | jnz LABEL(exit) |
| 370 | |
| 371 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 372 | sub $16, %r11 |
| 373 | jbe LABEL(strcmp_exitz) |
| 374 | #endif |
| 375 | add $16, %rcx |
| 376 | movdqa %xmm4, %xmm3 |
| 377 | |
| 378 | add $16, %r10 |
| 379 | jg LABEL(nibble_ashr_1) /* cross page boundary */ |
| 380 | |
| 381 | movdqa (%rsi, %rcx), %xmm1 |
| 382 | movdqa (%rdi, %rcx), %xmm2 |
| 383 | movdqa %xmm2, %xmm4 /* store for next cycle */ |
| 384 | |
| 385 | #ifndef USE_SSSE3 |
| 386 | psrldq $1, %xmm3 |
| 387 | pslldq $15, %xmm2 |
| 388 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 389 | #else |
| 390 | palignr $1, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 391 | #endif |
| 392 | TOLOWER (%xmm1, %xmm2) |
| 393 | |
| 394 | pcmpeqb %xmm1, %xmm0 |
| 395 | pcmpeqb %xmm2, %xmm1 |
| 396 | psubb %xmm0, %xmm1 |
| 397 | pmovmskb %xmm1, %edx |
| 398 | sub $0xffff, %edx |
| 399 | jnz LABEL(exit) |
| 400 | |
| 401 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 402 | sub $16, %r11 |
| 403 | jbe LABEL(strcmp_exitz) |
| 404 | #endif |
| 405 | add $16, %rcx |
| 406 | movdqa %xmm4, %xmm3 |
| 407 | jmp LABEL(loop_ashr_1) |
| 408 | |
| 409 | /* |
| 410 | * Nibble avoids loads across page boundary. This is to avoid a potential |
| 411 | * access into unmapped memory. |
| 412 | */ |
| 413 | .p2align 4 |
| 414 | LABEL(nibble_ashr_1): |
| 415 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char*/ |
| 416 | pmovmskb %xmm0, %edx |
| 417 | test $0xfffe, %edx |
| 418 | jnz LABEL(ashr_1_exittail) /* find null char*/ |
| 419 | |
| 420 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 421 | cmp $15, %r11 |
| 422 | jbe LABEL(ashr_1_exittail) |
| 423 | #endif |
| 424 | |
| 425 | pxor %xmm0, %xmm0 |
| 426 | sub $0x1000, %r10 /* substract 4K from %r10 */ |
| 427 | jmp LABEL(gobble_ashr_1) |
| 428 | |
| 429 | /* |
| 430 | * Once find null char, determine if there is a string mismatch |
| 431 | * before the null char. |
| 432 | */ |
| 433 | .p2align 4 |
| 434 | LABEL(ashr_1_exittail): |
| 435 | movdqa (%rsi, %rcx), %xmm1 |
| 436 | psrldq $1, %xmm0 |
| 437 | psrldq $1, %xmm3 |
| 438 | jmp LABEL(aftertail) |
| 439 | |
| 440 | /* |
| 441 | * The following cases will be handled by ashr_2 |
| 442 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 443 | * n(14~15) n -14 1(15 +(n-14) - n) ashr_2 |
| 444 | */ |
| 445 | .p2align 4 |
| 446 | LABEL(ashr_2): |
| 447 | pxor %xmm0, %xmm0 |
| 448 | movdqa (%rdi), %xmm2 |
| 449 | movdqa (%rsi), %xmm1 |
| 450 | pcmpeqb %xmm1, %xmm0 |
| 451 | pslldq $14, %xmm2 |
| 452 | TOLOWER (%xmm1, %xmm2) |
| 453 | pcmpeqb %xmm1, %xmm2 |
| 454 | psubb %xmm0, %xmm2 |
| 455 | pmovmskb %xmm2, %r9d |
| 456 | shr %cl, %edx |
| 457 | shr %cl, %r9d |
| 458 | sub %r9d, %edx |
| 459 | jnz LABEL(less32bytes) |
| 460 | movdqa (%rdi), %xmm3 |
| 461 | UPDATE_STRNCMP_COUNTER |
| 462 | |
| 463 | pxor %xmm0, %xmm0 |
| 464 | mov $16, %rcx /* index for loads */ |
| 465 | mov $2, %r9d /* byte position left over from less32bytes case */ |
| 466 | /* |
| 467 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 468 | * When %r10 goes positive we have crossed a page boundary and |
| 469 | * need to do a nibble. |
| 470 | */ |
| 471 | lea 2(%rdi), %r10 |
| 472 | and $0xfff, %r10 /* offset into 4K page */ |
| 473 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 474 | |
| 475 | .p2align 4 |
| 476 | LABEL(loop_ashr_2): |
| 477 | add $16, %r10 |
| 478 | jg LABEL(nibble_ashr_2) |
| 479 | |
| 480 | LABEL(gobble_ashr_2): |
| 481 | movdqa (%rsi, %rcx), %xmm1 |
| 482 | movdqa (%rdi, %rcx), %xmm2 |
| 483 | movdqa %xmm2, %xmm4 |
| 484 | |
| 485 | #ifndef USE_SSSE3 |
| 486 | psrldq $2, %xmm3 |
| 487 | pslldq $14, %xmm2 |
| 488 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 489 | #else |
| 490 | palignr $2, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 491 | #endif |
| 492 | TOLOWER (%xmm1, %xmm2) |
| 493 | |
| 494 | pcmpeqb %xmm1, %xmm0 |
| 495 | pcmpeqb %xmm2, %xmm1 |
| 496 | psubb %xmm0, %xmm1 |
| 497 | pmovmskb %xmm1, %edx |
| 498 | sub $0xffff, %edx |
| 499 | jnz LABEL(exit) |
| 500 | |
| 501 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 502 | sub $16, %r11 |
| 503 | jbe LABEL(strcmp_exitz) |
| 504 | #endif |
| 505 | |
| 506 | add $16, %rcx |
| 507 | movdqa %xmm4, %xmm3 |
| 508 | |
| 509 | add $16, %r10 |
| 510 | jg LABEL(nibble_ashr_2) /* cross page boundary */ |
| 511 | |
| 512 | movdqa (%rsi, %rcx), %xmm1 |
| 513 | movdqa (%rdi, %rcx), %xmm2 |
| 514 | movdqa %xmm2, %xmm4 |
| 515 | |
| 516 | #ifndef USE_SSSE3 |
| 517 | psrldq $2, %xmm3 |
| 518 | pslldq $14, %xmm2 |
| 519 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 520 | #else |
| 521 | palignr $2, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 522 | #endif |
| 523 | TOLOWER (%xmm1, %xmm2) |
| 524 | |
| 525 | pcmpeqb %xmm1, %xmm0 |
| 526 | pcmpeqb %xmm2, %xmm1 |
| 527 | psubb %xmm0, %xmm1 |
| 528 | pmovmskb %xmm1, %edx |
| 529 | sub $0xffff, %edx |
| 530 | jnz LABEL(exit) |
| 531 | |
| 532 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 533 | sub $16, %r11 |
| 534 | jbe LABEL(strcmp_exitz) |
| 535 | #endif |
| 536 | |
| 537 | add $16, %rcx |
| 538 | movdqa %xmm4, %xmm3 |
| 539 | jmp LABEL(loop_ashr_2) |
| 540 | |
| 541 | .p2align 4 |
| 542 | LABEL(nibble_ashr_2): |
| 543 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 544 | pmovmskb %xmm0, %edx |
| 545 | test $0xfffc, %edx |
| 546 | jnz LABEL(ashr_2_exittail) |
| 547 | |
| 548 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 549 | cmp $14, %r11 |
| 550 | jbe LABEL(ashr_2_exittail) |
| 551 | #endif |
| 552 | |
| 553 | pxor %xmm0, %xmm0 |
| 554 | sub $0x1000, %r10 |
| 555 | jmp LABEL(gobble_ashr_2) |
| 556 | |
| 557 | .p2align 4 |
| 558 | LABEL(ashr_2_exittail): |
| 559 | movdqa (%rsi, %rcx), %xmm1 |
| 560 | psrldq $2, %xmm0 |
| 561 | psrldq $2, %xmm3 |
| 562 | jmp LABEL(aftertail) |
| 563 | |
| 564 | /* |
| 565 | * The following cases will be handled by ashr_3 |
| 566 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 567 | * n(13~15) n -13 2(15 +(n-13) - n) ashr_3 |
| 568 | */ |
| 569 | .p2align 4 |
| 570 | LABEL(ashr_3): |
| 571 | pxor %xmm0, %xmm0 |
| 572 | movdqa (%rdi), %xmm2 |
| 573 | movdqa (%rsi), %xmm1 |
| 574 | pcmpeqb %xmm1, %xmm0 |
| 575 | pslldq $13, %xmm2 |
| 576 | TOLOWER (%xmm1, %xmm2) |
| 577 | pcmpeqb %xmm1, %xmm2 |
| 578 | psubb %xmm0, %xmm2 |
| 579 | pmovmskb %xmm2, %r9d |
| 580 | shr %cl, %edx |
| 581 | shr %cl, %r9d |
| 582 | sub %r9d, %edx |
| 583 | jnz LABEL(less32bytes) |
| 584 | movdqa (%rdi), %xmm3 |
| 585 | |
| 586 | UPDATE_STRNCMP_COUNTER |
| 587 | |
| 588 | pxor %xmm0, %xmm0 |
| 589 | mov $16, %rcx /* index for loads */ |
| 590 | mov $3, %r9d /* byte position left over from less32bytes case */ |
| 591 | /* |
| 592 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 593 | * When %r10 goes positive we have crossed a page boundary and |
| 594 | * need to do a nibble. |
| 595 | */ |
| 596 | lea 3(%rdi), %r10 |
| 597 | and $0xfff, %r10 /* offset into 4K page */ |
| 598 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 599 | |
| 600 | .p2align 4 |
| 601 | LABEL(loop_ashr_3): |
| 602 | add $16, %r10 |
| 603 | jg LABEL(nibble_ashr_3) |
| 604 | |
| 605 | LABEL(gobble_ashr_3): |
| 606 | movdqa (%rsi, %rcx), %xmm1 |
| 607 | movdqa (%rdi, %rcx), %xmm2 |
| 608 | movdqa %xmm2, %xmm4 |
| 609 | |
| 610 | #ifndef USE_SSSE3 |
| 611 | psrldq $3, %xmm3 |
| 612 | pslldq $13, %xmm2 |
| 613 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 614 | #else |
| 615 | palignr $3, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 616 | #endif |
| 617 | TOLOWER (%xmm1, %xmm2) |
| 618 | |
| 619 | pcmpeqb %xmm1, %xmm0 |
| 620 | pcmpeqb %xmm2, %xmm1 |
| 621 | psubb %xmm0, %xmm1 |
| 622 | pmovmskb %xmm1, %edx |
| 623 | sub $0xffff, %edx |
| 624 | jnz LABEL(exit) |
| 625 | |
| 626 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 627 | sub $16, %r11 |
| 628 | jbe LABEL(strcmp_exitz) |
| 629 | #endif |
| 630 | |
| 631 | add $16, %rcx |
| 632 | movdqa %xmm4, %xmm3 |
| 633 | |
| 634 | add $16, %r10 |
| 635 | jg LABEL(nibble_ashr_3) /* cross page boundary */ |
| 636 | |
| 637 | movdqa (%rsi, %rcx), %xmm1 |
| 638 | movdqa (%rdi, %rcx), %xmm2 |
| 639 | movdqa %xmm2, %xmm4 |
| 640 | |
| 641 | #ifndef USE_SSSE3 |
| 642 | psrldq $3, %xmm3 |
| 643 | pslldq $13, %xmm2 |
| 644 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 645 | #else |
| 646 | palignr $3, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 647 | #endif |
| 648 | TOLOWER (%xmm1, %xmm2) |
| 649 | |
| 650 | pcmpeqb %xmm1, %xmm0 |
| 651 | pcmpeqb %xmm2, %xmm1 |
| 652 | psubb %xmm0, %xmm1 |
| 653 | pmovmskb %xmm1, %edx |
| 654 | sub $0xffff, %edx |
| 655 | jnz LABEL(exit) |
| 656 | |
| 657 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 658 | sub $16, %r11 |
| 659 | jbe LABEL(strcmp_exitz) |
| 660 | #endif |
| 661 | |
| 662 | add $16, %rcx |
| 663 | movdqa %xmm4, %xmm3 |
| 664 | jmp LABEL(loop_ashr_3) |
| 665 | |
| 666 | .p2align 4 |
| 667 | LABEL(nibble_ashr_3): |
| 668 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 669 | pmovmskb %xmm0, %edx |
| 670 | test $0xfff8, %edx |
| 671 | jnz LABEL(ashr_3_exittail) |
| 672 | |
| 673 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 674 | cmp $13, %r11 |
| 675 | jbe LABEL(ashr_3_exittail) |
| 676 | #endif |
| 677 | |
| 678 | pxor %xmm0, %xmm0 |
| 679 | sub $0x1000, %r10 |
| 680 | jmp LABEL(gobble_ashr_3) |
| 681 | |
| 682 | .p2align 4 |
| 683 | LABEL(ashr_3_exittail): |
| 684 | movdqa (%rsi, %rcx), %xmm1 |
| 685 | psrldq $3, %xmm0 |
| 686 | psrldq $3, %xmm3 |
| 687 | jmp LABEL(aftertail) |
| 688 | |
| 689 | /* |
| 690 | * The following cases will be handled by ashr_4 |
| 691 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 692 | * n(12~15) n -12 3(15 +(n-12) - n) ashr_4 |
| 693 | */ |
| 694 | .p2align 4 |
| 695 | LABEL(ashr_4): |
| 696 | pxor %xmm0, %xmm0 |
| 697 | movdqa (%rdi), %xmm2 |
| 698 | movdqa (%rsi), %xmm1 |
| 699 | pcmpeqb %xmm1, %xmm0 |
| 700 | pslldq $12, %xmm2 |
| 701 | TOLOWER (%xmm1, %xmm2) |
| 702 | pcmpeqb %xmm1, %xmm2 |
| 703 | psubb %xmm0, %xmm2 |
| 704 | pmovmskb %xmm2, %r9d |
| 705 | shr %cl, %edx |
| 706 | shr %cl, %r9d |
| 707 | sub %r9d, %edx |
| 708 | jnz LABEL(less32bytes) |
| 709 | movdqa (%rdi), %xmm3 |
| 710 | |
| 711 | UPDATE_STRNCMP_COUNTER |
| 712 | |
| 713 | pxor %xmm0, %xmm0 |
| 714 | mov $16, %rcx /* index for loads */ |
| 715 | mov $4, %r9d /* byte position left over from less32bytes case */ |
| 716 | /* |
| 717 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 718 | * When %r10 goes positive we have crossed a page boundary and |
| 719 | * need to do a nibble. |
| 720 | */ |
| 721 | lea 4(%rdi), %r10 |
| 722 | and $0xfff, %r10 /* offset into 4K page */ |
| 723 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 724 | |
| 725 | .p2align 4 |
| 726 | LABEL(loop_ashr_4): |
| 727 | add $16, %r10 |
| 728 | jg LABEL(nibble_ashr_4) |
| 729 | |
| 730 | LABEL(gobble_ashr_4): |
| 731 | movdqa (%rsi, %rcx), %xmm1 |
| 732 | movdqa (%rdi, %rcx), %xmm2 |
| 733 | movdqa %xmm2, %xmm4 |
| 734 | |
| 735 | #ifndef USE_SSSE3 |
| 736 | psrldq $4, %xmm3 |
| 737 | pslldq $12, %xmm2 |
| 738 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 739 | #else |
| 740 | palignr $4, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 741 | #endif |
| 742 | TOLOWER (%xmm1, %xmm2) |
| 743 | |
| 744 | pcmpeqb %xmm1, %xmm0 |
| 745 | pcmpeqb %xmm2, %xmm1 |
| 746 | psubb %xmm0, %xmm1 |
| 747 | pmovmskb %xmm1, %edx |
| 748 | sub $0xffff, %edx |
| 749 | jnz LABEL(exit) |
| 750 | |
| 751 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 752 | sub $16, %r11 |
| 753 | jbe LABEL(strcmp_exitz) |
| 754 | #endif |
| 755 | |
| 756 | add $16, %rcx |
| 757 | movdqa %xmm4, %xmm3 |
| 758 | |
| 759 | add $16, %r10 |
| 760 | jg LABEL(nibble_ashr_4) /* cross page boundary */ |
| 761 | |
| 762 | movdqa (%rsi, %rcx), %xmm1 |
| 763 | movdqa (%rdi, %rcx), %xmm2 |
| 764 | movdqa %xmm2, %xmm4 |
| 765 | |
| 766 | #ifndef USE_SSSE3 |
| 767 | psrldq $4, %xmm3 |
| 768 | pslldq $12, %xmm2 |
| 769 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 770 | #else |
| 771 | palignr $4, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 772 | #endif |
| 773 | TOLOWER (%xmm1, %xmm2) |
| 774 | |
| 775 | pcmpeqb %xmm1, %xmm0 |
| 776 | pcmpeqb %xmm2, %xmm1 |
| 777 | psubb %xmm0, %xmm1 |
| 778 | pmovmskb %xmm1, %edx |
| 779 | sub $0xffff, %edx |
| 780 | jnz LABEL(exit) |
| 781 | |
| 782 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 783 | sub $16, %r11 |
| 784 | jbe LABEL(strcmp_exitz) |
| 785 | #endif |
| 786 | |
| 787 | add $16, %rcx |
| 788 | movdqa %xmm4, %xmm3 |
| 789 | jmp LABEL(loop_ashr_4) |
| 790 | |
| 791 | .p2align 4 |
| 792 | LABEL(nibble_ashr_4): |
| 793 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 794 | pmovmskb %xmm0, %edx |
| 795 | test $0xfff0, %edx |
| 796 | jnz LABEL(ashr_4_exittail) |
| 797 | |
| 798 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 799 | cmp $12, %r11 |
| 800 | jbe LABEL(ashr_4_exittail) |
| 801 | #endif |
| 802 | |
| 803 | pxor %xmm0, %xmm0 |
| 804 | sub $0x1000, %r10 |
| 805 | jmp LABEL(gobble_ashr_4) |
| 806 | |
| 807 | .p2align 4 |
| 808 | LABEL(ashr_4_exittail): |
| 809 | movdqa (%rsi, %rcx), %xmm1 |
| 810 | psrldq $4, %xmm0 |
| 811 | psrldq $4, %xmm3 |
| 812 | jmp LABEL(aftertail) |
| 813 | |
| 814 | /* |
| 815 | * The following cases will be handled by ashr_5 |
| 816 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 817 | * n(11~15) n - 11 4(15 +(n-11) - n) ashr_5 |
| 818 | */ |
| 819 | .p2align 4 |
| 820 | LABEL(ashr_5): |
| 821 | pxor %xmm0, %xmm0 |
| 822 | movdqa (%rdi), %xmm2 |
| 823 | movdqa (%rsi), %xmm1 |
| 824 | pcmpeqb %xmm1, %xmm0 |
| 825 | pslldq $11, %xmm2 |
| 826 | TOLOWER (%xmm1, %xmm2) |
| 827 | pcmpeqb %xmm1, %xmm2 |
| 828 | psubb %xmm0, %xmm2 |
| 829 | pmovmskb %xmm2, %r9d |
| 830 | shr %cl, %edx |
| 831 | shr %cl, %r9d |
| 832 | sub %r9d, %edx |
| 833 | jnz LABEL(less32bytes) |
| 834 | movdqa (%rdi), %xmm3 |
| 835 | |
| 836 | UPDATE_STRNCMP_COUNTER |
| 837 | |
| 838 | pxor %xmm0, %xmm0 |
| 839 | mov $16, %rcx /* index for loads */ |
| 840 | mov $5, %r9d /* byte position left over from less32bytes case */ |
| 841 | /* |
| 842 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 843 | * When %r10 goes positive we have crossed a page boundary and |
| 844 | * need to do a nibble. |
| 845 | */ |
| 846 | lea 5(%rdi), %r10 |
| 847 | and $0xfff, %r10 /* offset into 4K page */ |
| 848 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 849 | |
| 850 | .p2align 4 |
| 851 | LABEL(loop_ashr_5): |
| 852 | add $16, %r10 |
| 853 | jg LABEL(nibble_ashr_5) |
| 854 | |
| 855 | LABEL(gobble_ashr_5): |
| 856 | movdqa (%rsi, %rcx), %xmm1 |
| 857 | movdqa (%rdi, %rcx), %xmm2 |
| 858 | movdqa %xmm2, %xmm4 |
| 859 | |
| 860 | #ifndef USE_SSSE3 |
| 861 | psrldq $5, %xmm3 |
| 862 | pslldq $11, %xmm2 |
| 863 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 864 | #else |
| 865 | palignr $5, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 866 | #endif |
| 867 | TOLOWER (%xmm1, %xmm2) |
| 868 | |
| 869 | pcmpeqb %xmm1, %xmm0 |
| 870 | pcmpeqb %xmm2, %xmm1 |
| 871 | psubb %xmm0, %xmm1 |
| 872 | pmovmskb %xmm1, %edx |
| 873 | sub $0xffff, %edx |
| 874 | jnz LABEL(exit) |
| 875 | |
| 876 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 877 | sub $16, %r11 |
| 878 | jbe LABEL(strcmp_exitz) |
| 879 | #endif |
| 880 | |
| 881 | add $16, %rcx |
| 882 | movdqa %xmm4, %xmm3 |
| 883 | |
| 884 | add $16, %r10 |
| 885 | jg LABEL(nibble_ashr_5) /* cross page boundary */ |
| 886 | |
| 887 | movdqa (%rsi, %rcx), %xmm1 |
| 888 | movdqa (%rdi, %rcx), %xmm2 |
| 889 | movdqa %xmm2, %xmm4 |
| 890 | |
| 891 | #ifndef USE_SSSE3 |
| 892 | psrldq $5, %xmm3 |
| 893 | pslldq $11, %xmm2 |
| 894 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 895 | #else |
| 896 | palignr $5, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 897 | #endif |
| 898 | TOLOWER (%xmm1, %xmm2) |
| 899 | |
| 900 | pcmpeqb %xmm1, %xmm0 |
| 901 | pcmpeqb %xmm2, %xmm1 |
| 902 | psubb %xmm0, %xmm1 |
| 903 | pmovmskb %xmm1, %edx |
| 904 | sub $0xffff, %edx |
| 905 | jnz LABEL(exit) |
| 906 | |
| 907 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 908 | sub $16, %r11 |
| 909 | jbe LABEL(strcmp_exitz) |
| 910 | #endif |
| 911 | |
| 912 | add $16, %rcx |
| 913 | movdqa %xmm4, %xmm3 |
| 914 | jmp LABEL(loop_ashr_5) |
| 915 | |
| 916 | .p2align 4 |
| 917 | LABEL(nibble_ashr_5): |
| 918 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 919 | pmovmskb %xmm0, %edx |
| 920 | test $0xffe0, %edx |
| 921 | jnz LABEL(ashr_5_exittail) |
| 922 | |
| 923 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 924 | cmp $11, %r11 |
| 925 | jbe LABEL(ashr_5_exittail) |
| 926 | #endif |
| 927 | |
| 928 | pxor %xmm0, %xmm0 |
| 929 | sub $0x1000, %r10 |
| 930 | jmp LABEL(gobble_ashr_5) |
| 931 | |
| 932 | .p2align 4 |
| 933 | LABEL(ashr_5_exittail): |
| 934 | movdqa (%rsi, %rcx), %xmm1 |
| 935 | psrldq $5, %xmm0 |
| 936 | psrldq $5, %xmm3 |
| 937 | jmp LABEL(aftertail) |
| 938 | |
| 939 | /* |
| 940 | * The following cases will be handled by ashr_6 |
| 941 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 942 | * n(10~15) n - 10 5(15 +(n-10) - n) ashr_6 |
| 943 | */ |
| 944 | .p2align 4 |
| 945 | LABEL(ashr_6): |
| 946 | pxor %xmm0, %xmm0 |
| 947 | movdqa (%rdi), %xmm2 |
| 948 | movdqa (%rsi), %xmm1 |
| 949 | pcmpeqb %xmm1, %xmm0 |
| 950 | pslldq $10, %xmm2 |
| 951 | TOLOWER (%xmm1, %xmm2) |
| 952 | pcmpeqb %xmm1, %xmm2 |
| 953 | psubb %xmm0, %xmm2 |
| 954 | pmovmskb %xmm2, %r9d |
| 955 | shr %cl, %edx |
| 956 | shr %cl, %r9d |
| 957 | sub %r9d, %edx |
| 958 | jnz LABEL(less32bytes) |
| 959 | movdqa (%rdi), %xmm3 |
| 960 | |
| 961 | UPDATE_STRNCMP_COUNTER |
| 962 | |
| 963 | pxor %xmm0, %xmm0 |
| 964 | mov $16, %rcx /* index for loads */ |
| 965 | mov $6, %r9d /* byte position left over from less32bytes case */ |
| 966 | /* |
| 967 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 968 | * When %r10 goes positive we have crossed a page boundary and |
| 969 | * need to do a nibble. |
| 970 | */ |
| 971 | lea 6(%rdi), %r10 |
| 972 | and $0xfff, %r10 /* offset into 4K page */ |
| 973 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 974 | |
| 975 | .p2align 4 |
| 976 | LABEL(loop_ashr_6): |
| 977 | add $16, %r10 |
| 978 | jg LABEL(nibble_ashr_6) |
| 979 | |
| 980 | LABEL(gobble_ashr_6): |
| 981 | movdqa (%rsi, %rcx), %xmm1 |
| 982 | movdqa (%rdi, %rcx), %xmm2 |
| 983 | movdqa %xmm2, %xmm4 |
| 984 | |
| 985 | #ifndef USE_SSSE3 |
| 986 | psrldq $6, %xmm3 |
| 987 | pslldq $10, %xmm2 |
| 988 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 989 | #else |
| 990 | palignr $6, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 991 | #endif |
| 992 | TOLOWER (%xmm1, %xmm2) |
| 993 | |
| 994 | pcmpeqb %xmm1, %xmm0 |
| 995 | pcmpeqb %xmm2, %xmm1 |
| 996 | psubb %xmm0, %xmm1 |
| 997 | pmovmskb %xmm1, %edx |
| 998 | sub $0xffff, %edx |
| 999 | jnz LABEL(exit) |
| 1000 | |
| 1001 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1002 | sub $16, %r11 |
| 1003 | jbe LABEL(strcmp_exitz) |
| 1004 | #endif |
| 1005 | |
| 1006 | add $16, %rcx |
| 1007 | movdqa %xmm4, %xmm3 |
| 1008 | |
| 1009 | add $16, %r10 |
| 1010 | jg LABEL(nibble_ashr_6) /* cross page boundary */ |
| 1011 | |
| 1012 | movdqa (%rsi, %rcx), %xmm1 |
| 1013 | movdqa (%rdi, %rcx), %xmm2 |
| 1014 | movdqa %xmm2, %xmm4 |
| 1015 | |
| 1016 | #ifndef USE_SSSE3 |
| 1017 | psrldq $6, %xmm3 |
| 1018 | pslldq $10, %xmm2 |
| 1019 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1020 | #else |
| 1021 | palignr $6, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1022 | #endif |
| 1023 | TOLOWER (%xmm1, %xmm2) |
| 1024 | |
| 1025 | pcmpeqb %xmm1, %xmm0 |
| 1026 | pcmpeqb %xmm2, %xmm1 |
| 1027 | psubb %xmm0, %xmm1 |
| 1028 | pmovmskb %xmm1, %edx |
| 1029 | sub $0xffff, %edx |
| 1030 | jnz LABEL(exit) |
| 1031 | |
| 1032 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1033 | sub $16, %r11 |
| 1034 | jbe LABEL(strcmp_exitz) |
| 1035 | #endif |
| 1036 | |
| 1037 | add $16, %rcx |
| 1038 | movdqa %xmm4, %xmm3 |
| 1039 | jmp LABEL(loop_ashr_6) |
| 1040 | |
| 1041 | .p2align 4 |
| 1042 | LABEL(nibble_ashr_6): |
| 1043 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1044 | pmovmskb %xmm0, %edx |
| 1045 | test $0xffc0, %edx |
| 1046 | jnz LABEL(ashr_6_exittail) |
| 1047 | |
| 1048 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1049 | cmp $10, %r11 |
| 1050 | jbe LABEL(ashr_6_exittail) |
| 1051 | #endif |
| 1052 | |
| 1053 | pxor %xmm0, %xmm0 |
| 1054 | sub $0x1000, %r10 |
| 1055 | jmp LABEL(gobble_ashr_6) |
| 1056 | |
| 1057 | .p2align 4 |
| 1058 | LABEL(ashr_6_exittail): |
| 1059 | movdqa (%rsi, %rcx), %xmm1 |
| 1060 | psrldq $6, %xmm0 |
| 1061 | psrldq $6, %xmm3 |
| 1062 | jmp LABEL(aftertail) |
| 1063 | |
| 1064 | /* |
| 1065 | * The following cases will be handled by ashr_7 |
| 1066 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1067 | * n(9~15) n - 9 6(15 +(n - 9) - n) ashr_7 |
| 1068 | */ |
| 1069 | .p2align 4 |
| 1070 | LABEL(ashr_7): |
| 1071 | pxor %xmm0, %xmm0 |
| 1072 | movdqa (%rdi), %xmm2 |
| 1073 | movdqa (%rsi), %xmm1 |
| 1074 | pcmpeqb %xmm1, %xmm0 |
| 1075 | pslldq $9, %xmm2 |
| 1076 | TOLOWER (%xmm1, %xmm2) |
| 1077 | pcmpeqb %xmm1, %xmm2 |
| 1078 | psubb %xmm0, %xmm2 |
| 1079 | pmovmskb %xmm2, %r9d |
| 1080 | shr %cl, %edx |
| 1081 | shr %cl, %r9d |
| 1082 | sub %r9d, %edx |
| 1083 | jnz LABEL(less32bytes) |
| 1084 | movdqa (%rdi), %xmm3 |
| 1085 | |
| 1086 | UPDATE_STRNCMP_COUNTER |
| 1087 | |
| 1088 | pxor %xmm0, %xmm0 |
| 1089 | mov $16, %rcx /* index for loads */ |
| 1090 | mov $7, %r9d /* byte position left over from less32bytes case */ |
| 1091 | /* |
| 1092 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1093 | * When %r10 goes positive we have crossed a page boundary and |
| 1094 | * need to do a nibble. |
| 1095 | */ |
| 1096 | lea 7(%rdi), %r10 |
| 1097 | and $0xfff, %r10 /* offset into 4K page */ |
| 1098 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1099 | |
| 1100 | .p2align 4 |
| 1101 | LABEL(loop_ashr_7): |
| 1102 | add $16, %r10 |
| 1103 | jg LABEL(nibble_ashr_7) |
| 1104 | |
| 1105 | LABEL(gobble_ashr_7): |
| 1106 | movdqa (%rsi, %rcx), %xmm1 |
| 1107 | movdqa (%rdi, %rcx), %xmm2 |
| 1108 | movdqa %xmm2, %xmm4 |
| 1109 | |
| 1110 | #ifndef USE_SSSE3 |
| 1111 | psrldq $7, %xmm3 |
| 1112 | pslldq $9, %xmm2 |
| 1113 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1114 | #else |
| 1115 | palignr $7, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1116 | #endif |
| 1117 | TOLOWER (%xmm1, %xmm2) |
| 1118 | |
| 1119 | pcmpeqb %xmm1, %xmm0 |
| 1120 | pcmpeqb %xmm2, %xmm1 |
| 1121 | psubb %xmm0, %xmm1 |
| 1122 | pmovmskb %xmm1, %edx |
| 1123 | sub $0xffff, %edx |
| 1124 | jnz LABEL(exit) |
| 1125 | |
| 1126 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1127 | sub $16, %r11 |
| 1128 | jbe LABEL(strcmp_exitz) |
| 1129 | #endif |
| 1130 | |
| 1131 | add $16, %rcx |
| 1132 | movdqa %xmm4, %xmm3 |
| 1133 | |
| 1134 | add $16, %r10 |
| 1135 | jg LABEL(nibble_ashr_7) /* cross page boundary */ |
| 1136 | |
| 1137 | movdqa (%rsi, %rcx), %xmm1 |
| 1138 | movdqa (%rdi, %rcx), %xmm2 |
| 1139 | movdqa %xmm2, %xmm4 |
| 1140 | |
| 1141 | #ifndef USE_SSSE3 |
| 1142 | psrldq $7, %xmm3 |
| 1143 | pslldq $9, %xmm2 |
| 1144 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1145 | #else |
| 1146 | palignr $7, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1147 | #endif |
| 1148 | TOLOWER (%xmm1, %xmm2) |
| 1149 | |
| 1150 | pcmpeqb %xmm1, %xmm0 |
| 1151 | pcmpeqb %xmm2, %xmm1 |
| 1152 | psubb %xmm0, %xmm1 |
| 1153 | pmovmskb %xmm1, %edx |
| 1154 | sub $0xffff, %edx |
| 1155 | jnz LABEL(exit) |
| 1156 | |
| 1157 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1158 | sub $16, %r11 |
| 1159 | jbe LABEL(strcmp_exitz) |
| 1160 | #endif |
| 1161 | |
| 1162 | add $16, %rcx |
| 1163 | movdqa %xmm4, %xmm3 |
| 1164 | jmp LABEL(loop_ashr_7) |
| 1165 | |
| 1166 | .p2align 4 |
| 1167 | LABEL(nibble_ashr_7): |
| 1168 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1169 | pmovmskb %xmm0, %edx |
| 1170 | test $0xff80, %edx |
| 1171 | jnz LABEL(ashr_7_exittail) |
| 1172 | |
| 1173 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1174 | cmp $9, %r11 |
| 1175 | jbe LABEL(ashr_7_exittail) |
| 1176 | #endif |
| 1177 | |
| 1178 | pxor %xmm0, %xmm0 |
| 1179 | sub $0x1000, %r10 |
| 1180 | jmp LABEL(gobble_ashr_7) |
| 1181 | |
| 1182 | .p2align 4 |
| 1183 | LABEL(ashr_7_exittail): |
| 1184 | movdqa (%rsi, %rcx), %xmm1 |
| 1185 | psrldq $7, %xmm0 |
| 1186 | psrldq $7, %xmm3 |
| 1187 | jmp LABEL(aftertail) |
| 1188 | |
| 1189 | /* |
| 1190 | * The following cases will be handled by ashr_8 |
| 1191 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1192 | * n(8~15) n - 8 7(15 +(n - 8) - n) ashr_8 |
| 1193 | */ |
| 1194 | .p2align 4 |
| 1195 | LABEL(ashr_8): |
| 1196 | pxor %xmm0, %xmm0 |
| 1197 | movdqa (%rdi), %xmm2 |
| 1198 | movdqa (%rsi), %xmm1 |
| 1199 | pcmpeqb %xmm1, %xmm0 |
| 1200 | pslldq $8, %xmm2 |
| 1201 | TOLOWER (%xmm1, %xmm2) |
| 1202 | pcmpeqb %xmm1, %xmm2 |
| 1203 | psubb %xmm0, %xmm2 |
| 1204 | pmovmskb %xmm2, %r9d |
| 1205 | shr %cl, %edx |
| 1206 | shr %cl, %r9d |
| 1207 | sub %r9d, %edx |
| 1208 | jnz LABEL(less32bytes) |
| 1209 | movdqa (%rdi), %xmm3 |
| 1210 | |
| 1211 | UPDATE_STRNCMP_COUNTER |
| 1212 | |
| 1213 | pxor %xmm0, %xmm0 |
| 1214 | mov $16, %rcx /* index for loads */ |
| 1215 | mov $8, %r9d /* byte position left over from less32bytes case */ |
| 1216 | /* |
| 1217 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1218 | * When %r10 goes positive we have crossed a page boundary and |
| 1219 | * need to do a nibble. |
| 1220 | */ |
| 1221 | lea 8(%rdi), %r10 |
| 1222 | and $0xfff, %r10 /* offset into 4K page */ |
| 1223 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1224 | |
| 1225 | .p2align 4 |
| 1226 | LABEL(loop_ashr_8): |
| 1227 | add $16, %r10 |
| 1228 | jg LABEL(nibble_ashr_8) |
| 1229 | |
| 1230 | LABEL(gobble_ashr_8): |
| 1231 | movdqa (%rsi, %rcx), %xmm1 |
| 1232 | movdqa (%rdi, %rcx), %xmm2 |
| 1233 | movdqa %xmm2, %xmm4 |
| 1234 | |
| 1235 | #ifndef USE_SSSE3 |
| 1236 | psrldq $8, %xmm3 |
| 1237 | pslldq $8, %xmm2 |
| 1238 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1239 | #else |
| 1240 | palignr $8, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1241 | #endif |
| 1242 | TOLOWER (%xmm1, %xmm2) |
| 1243 | |
| 1244 | pcmpeqb %xmm1, %xmm0 |
| 1245 | pcmpeqb %xmm2, %xmm1 |
| 1246 | psubb %xmm0, %xmm1 |
| 1247 | pmovmskb %xmm1, %edx |
| 1248 | sub $0xffff, %edx |
| 1249 | jnz LABEL(exit) |
| 1250 | |
| 1251 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1252 | sub $16, %r11 |
| 1253 | jbe LABEL(strcmp_exitz) |
| 1254 | #endif |
| 1255 | |
| 1256 | add $16, %rcx |
| 1257 | movdqa %xmm4, %xmm3 |
| 1258 | |
| 1259 | add $16, %r10 |
| 1260 | jg LABEL(nibble_ashr_8) /* cross page boundary */ |
| 1261 | |
| 1262 | movdqa (%rsi, %rcx), %xmm1 |
| 1263 | movdqa (%rdi, %rcx), %xmm2 |
| 1264 | movdqa %xmm2, %xmm4 |
| 1265 | |
| 1266 | #ifndef USE_SSSE3 |
| 1267 | psrldq $8, %xmm3 |
| 1268 | pslldq $8, %xmm2 |
| 1269 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1270 | #else |
| 1271 | palignr $8, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1272 | #endif |
| 1273 | TOLOWER (%xmm1, %xmm2) |
| 1274 | |
| 1275 | pcmpeqb %xmm1, %xmm0 |
| 1276 | pcmpeqb %xmm2, %xmm1 |
| 1277 | psubb %xmm0, %xmm1 |
| 1278 | pmovmskb %xmm1, %edx |
| 1279 | sub $0xffff, %edx |
| 1280 | jnz LABEL(exit) |
| 1281 | |
| 1282 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1283 | sub $16, %r11 |
| 1284 | jbe LABEL(strcmp_exitz) |
| 1285 | #endif |
| 1286 | |
| 1287 | add $16, %rcx |
| 1288 | movdqa %xmm4, %xmm3 |
| 1289 | jmp LABEL(loop_ashr_8) |
| 1290 | |
| 1291 | .p2align 4 |
| 1292 | LABEL(nibble_ashr_8): |
| 1293 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1294 | pmovmskb %xmm0, %edx |
| 1295 | test $0xff00, %edx |
| 1296 | jnz LABEL(ashr_8_exittail) |
| 1297 | |
| 1298 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1299 | cmp $8, %r11 |
| 1300 | jbe LABEL(ashr_8_exittail) |
| 1301 | #endif |
| 1302 | |
| 1303 | pxor %xmm0, %xmm0 |
| 1304 | sub $0x1000, %r10 |
| 1305 | jmp LABEL(gobble_ashr_8) |
| 1306 | |
| 1307 | .p2align 4 |
| 1308 | LABEL(ashr_8_exittail): |
| 1309 | movdqa (%rsi, %rcx), %xmm1 |
| 1310 | psrldq $8, %xmm0 |
| 1311 | psrldq $8, %xmm3 |
| 1312 | jmp LABEL(aftertail) |
| 1313 | |
| 1314 | /* |
| 1315 | * The following cases will be handled by ashr_9 |
| 1316 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1317 | * n(7~15) n - 7 8(15 +(n - 7) - n) ashr_9 |
| 1318 | */ |
| 1319 | .p2align 4 |
| 1320 | LABEL(ashr_9): |
| 1321 | pxor %xmm0, %xmm0 |
| 1322 | movdqa (%rdi), %xmm2 |
| 1323 | movdqa (%rsi), %xmm1 |
| 1324 | pcmpeqb %xmm1, %xmm0 |
| 1325 | pslldq $7, %xmm2 |
| 1326 | TOLOWER (%xmm1, %xmm2) |
| 1327 | pcmpeqb %xmm1, %xmm2 |
| 1328 | psubb %xmm0, %xmm2 |
| 1329 | pmovmskb %xmm2, %r9d |
| 1330 | shr %cl, %edx |
| 1331 | shr %cl, %r9d |
| 1332 | sub %r9d, %edx |
| 1333 | jnz LABEL(less32bytes) |
| 1334 | movdqa (%rdi), %xmm3 |
| 1335 | |
| 1336 | UPDATE_STRNCMP_COUNTER |
| 1337 | |
| 1338 | pxor %xmm0, %xmm0 |
| 1339 | mov $16, %rcx /* index for loads */ |
| 1340 | mov $9, %r9d /* byte position left over from less32bytes case */ |
| 1341 | /* |
| 1342 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1343 | * When %r10 goes positive we have crossed a page boundary and |
| 1344 | * need to do a nibble. |
| 1345 | */ |
| 1346 | lea 9(%rdi), %r10 |
| 1347 | and $0xfff, %r10 /* offset into 4K page */ |
| 1348 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1349 | |
| 1350 | .p2align 4 |
| 1351 | LABEL(loop_ashr_9): |
| 1352 | add $16, %r10 |
| 1353 | jg LABEL(nibble_ashr_9) |
| 1354 | |
| 1355 | LABEL(gobble_ashr_9): |
| 1356 | movdqa (%rsi, %rcx), %xmm1 |
| 1357 | movdqa (%rdi, %rcx), %xmm2 |
| 1358 | movdqa %xmm2, %xmm4 |
| 1359 | |
| 1360 | #ifndef USE_SSSE3 |
| 1361 | psrldq $9, %xmm3 |
| 1362 | pslldq $7, %xmm2 |
| 1363 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1364 | #else |
| 1365 | palignr $9, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1366 | #endif |
| 1367 | TOLOWER (%xmm1, %xmm2) |
| 1368 | |
| 1369 | pcmpeqb %xmm1, %xmm0 |
| 1370 | pcmpeqb %xmm2, %xmm1 |
| 1371 | psubb %xmm0, %xmm1 |
| 1372 | pmovmskb %xmm1, %edx |
| 1373 | sub $0xffff, %edx |
| 1374 | jnz LABEL(exit) |
| 1375 | |
| 1376 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1377 | sub $16, %r11 |
| 1378 | jbe LABEL(strcmp_exitz) |
| 1379 | #endif |
| 1380 | |
| 1381 | add $16, %rcx |
| 1382 | movdqa %xmm4, %xmm3 |
| 1383 | |
| 1384 | add $16, %r10 |
| 1385 | jg LABEL(nibble_ashr_9) /* cross page boundary */ |
| 1386 | |
| 1387 | movdqa (%rsi, %rcx), %xmm1 |
| 1388 | movdqa (%rdi, %rcx), %xmm2 |
| 1389 | movdqa %xmm2, %xmm4 |
| 1390 | |
| 1391 | #ifndef USE_SSSE3 |
| 1392 | psrldq $9, %xmm3 |
| 1393 | pslldq $7, %xmm2 |
| 1394 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1395 | #else |
| 1396 | palignr $9, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1397 | #endif |
| 1398 | TOLOWER (%xmm1, %xmm2) |
| 1399 | |
| 1400 | pcmpeqb %xmm1, %xmm0 |
| 1401 | pcmpeqb %xmm2, %xmm1 |
| 1402 | psubb %xmm0, %xmm1 |
| 1403 | pmovmskb %xmm1, %edx |
| 1404 | sub $0xffff, %edx |
| 1405 | jnz LABEL(exit) |
| 1406 | |
| 1407 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1408 | sub $16, %r11 |
| 1409 | jbe LABEL(strcmp_exitz) |
| 1410 | #endif |
| 1411 | |
| 1412 | add $16, %rcx |
| 1413 | movdqa %xmm4, %xmm3 /* store for next cycle */ |
| 1414 | jmp LABEL(loop_ashr_9) |
| 1415 | |
| 1416 | .p2align 4 |
| 1417 | LABEL(nibble_ashr_9): |
| 1418 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1419 | pmovmskb %xmm0, %edx |
| 1420 | test $0xfe00, %edx |
| 1421 | jnz LABEL(ashr_9_exittail) |
| 1422 | |
| 1423 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1424 | cmp $7, %r11 |
| 1425 | jbe LABEL(ashr_9_exittail) |
| 1426 | #endif |
| 1427 | |
| 1428 | pxor %xmm0, %xmm0 |
| 1429 | sub $0x1000, %r10 |
| 1430 | jmp LABEL(gobble_ashr_9) |
| 1431 | |
| 1432 | .p2align 4 |
| 1433 | LABEL(ashr_9_exittail): |
| 1434 | movdqa (%rsi, %rcx), %xmm1 |
| 1435 | psrldq $9, %xmm0 |
| 1436 | psrldq $9, %xmm3 |
| 1437 | jmp LABEL(aftertail) |
| 1438 | |
| 1439 | /* |
| 1440 | * The following cases will be handled by ashr_10 |
| 1441 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1442 | * n(6~15) n - 6 9(15 +(n - 6) - n) ashr_10 |
| 1443 | */ |
| 1444 | .p2align 4 |
| 1445 | LABEL(ashr_10): |
| 1446 | pxor %xmm0, %xmm0 |
| 1447 | movdqa (%rdi), %xmm2 |
| 1448 | movdqa (%rsi), %xmm1 |
| 1449 | pcmpeqb %xmm1, %xmm0 |
| 1450 | pslldq $6, %xmm2 |
| 1451 | TOLOWER (%xmm1, %xmm2) |
| 1452 | pcmpeqb %xmm1, %xmm2 |
| 1453 | psubb %xmm0, %xmm2 |
| 1454 | pmovmskb %xmm2, %r9d |
| 1455 | shr %cl, %edx |
| 1456 | shr %cl, %r9d |
| 1457 | sub %r9d, %edx |
| 1458 | jnz LABEL(less32bytes) |
| 1459 | movdqa (%rdi), %xmm3 |
| 1460 | |
| 1461 | UPDATE_STRNCMP_COUNTER |
| 1462 | |
| 1463 | pxor %xmm0, %xmm0 |
| 1464 | mov $16, %rcx /* index for loads */ |
| 1465 | mov $10, %r9d /* byte position left over from less32bytes case */ |
| 1466 | /* |
| 1467 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1468 | * When %r10 goes positive we have crossed a page boundary and |
| 1469 | * need to do a nibble. |
| 1470 | */ |
| 1471 | lea 10(%rdi), %r10 |
| 1472 | and $0xfff, %r10 /* offset into 4K page */ |
| 1473 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1474 | |
| 1475 | .p2align 4 |
| 1476 | LABEL(loop_ashr_10): |
| 1477 | add $16, %r10 |
| 1478 | jg LABEL(nibble_ashr_10) |
| 1479 | |
| 1480 | LABEL(gobble_ashr_10): |
| 1481 | movdqa (%rsi, %rcx), %xmm1 |
| 1482 | movdqa (%rdi, %rcx), %xmm2 |
| 1483 | movdqa %xmm2, %xmm4 |
| 1484 | |
| 1485 | #ifndef USE_SSSE3 |
| 1486 | psrldq $10, %xmm3 |
| 1487 | pslldq $6, %xmm2 |
| 1488 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1489 | #else |
| 1490 | palignr $10, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1491 | #endif |
| 1492 | TOLOWER (%xmm1, %xmm2) |
| 1493 | |
| 1494 | pcmpeqb %xmm1, %xmm0 |
| 1495 | pcmpeqb %xmm2, %xmm1 |
| 1496 | psubb %xmm0, %xmm1 |
| 1497 | pmovmskb %xmm1, %edx |
| 1498 | sub $0xffff, %edx |
| 1499 | jnz LABEL(exit) |
| 1500 | |
| 1501 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1502 | sub $16, %r11 |
| 1503 | jbe LABEL(strcmp_exitz) |
| 1504 | #endif |
| 1505 | |
| 1506 | add $16, %rcx |
| 1507 | movdqa %xmm4, %xmm3 |
| 1508 | |
| 1509 | add $16, %r10 |
| 1510 | jg LABEL(nibble_ashr_10) /* cross page boundary */ |
| 1511 | |
| 1512 | movdqa (%rsi, %rcx), %xmm1 |
| 1513 | movdqa (%rdi, %rcx), %xmm2 |
| 1514 | movdqa %xmm2, %xmm4 |
| 1515 | |
| 1516 | #ifndef USE_SSSE3 |
| 1517 | psrldq $10, %xmm3 |
| 1518 | pslldq $6, %xmm2 |
| 1519 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1520 | #else |
| 1521 | palignr $10, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1522 | #endif |
| 1523 | TOLOWER (%xmm1, %xmm2) |
| 1524 | |
| 1525 | pcmpeqb %xmm1, %xmm0 |
| 1526 | pcmpeqb %xmm2, %xmm1 |
| 1527 | psubb %xmm0, %xmm1 |
| 1528 | pmovmskb %xmm1, %edx |
| 1529 | sub $0xffff, %edx |
| 1530 | jnz LABEL(exit) |
| 1531 | |
| 1532 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1533 | sub $16, %r11 |
| 1534 | jbe LABEL(strcmp_exitz) |
| 1535 | #endif |
| 1536 | |
| 1537 | add $16, %rcx |
| 1538 | movdqa %xmm4, %xmm3 |
| 1539 | jmp LABEL(loop_ashr_10) |
| 1540 | |
| 1541 | .p2align 4 |
| 1542 | LABEL(nibble_ashr_10): |
| 1543 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1544 | pmovmskb %xmm0, %edx |
| 1545 | test $0xfc00, %edx |
| 1546 | jnz LABEL(ashr_10_exittail) |
| 1547 | |
| 1548 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1549 | cmp $6, %r11 |
| 1550 | jbe LABEL(ashr_10_exittail) |
| 1551 | #endif |
| 1552 | |
| 1553 | pxor %xmm0, %xmm0 |
| 1554 | sub $0x1000, %r10 |
| 1555 | jmp LABEL(gobble_ashr_10) |
| 1556 | |
| 1557 | .p2align 4 |
| 1558 | LABEL(ashr_10_exittail): |
| 1559 | movdqa (%rsi, %rcx), %xmm1 |
| 1560 | psrldq $10, %xmm0 |
| 1561 | psrldq $10, %xmm3 |
| 1562 | jmp LABEL(aftertail) |
| 1563 | |
| 1564 | /* |
| 1565 | * The following cases will be handled by ashr_11 |
| 1566 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1567 | * n(5~15) n - 5 10(15 +(n - 5) - n) ashr_11 |
| 1568 | */ |
| 1569 | .p2align 4 |
| 1570 | LABEL(ashr_11): |
| 1571 | pxor %xmm0, %xmm0 |
| 1572 | movdqa (%rdi), %xmm2 |
| 1573 | movdqa (%rsi), %xmm1 |
| 1574 | pcmpeqb %xmm1, %xmm0 |
| 1575 | pslldq $5, %xmm2 |
| 1576 | TOLOWER (%xmm1, %xmm2) |
| 1577 | pcmpeqb %xmm1, %xmm2 |
| 1578 | psubb %xmm0, %xmm2 |
| 1579 | pmovmskb %xmm2, %r9d |
| 1580 | shr %cl, %edx |
| 1581 | shr %cl, %r9d |
| 1582 | sub %r9d, %edx |
| 1583 | jnz LABEL(less32bytes) |
| 1584 | movdqa (%rdi), %xmm3 |
| 1585 | |
| 1586 | UPDATE_STRNCMP_COUNTER |
| 1587 | |
| 1588 | pxor %xmm0, %xmm0 |
| 1589 | mov $16, %rcx /* index for loads */ |
| 1590 | mov $11, %r9d /* byte position left over from less32bytes case */ |
| 1591 | /* |
| 1592 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1593 | * When %r10 goes positive we have crossed a page boundary and |
| 1594 | * need to do a nibble. |
| 1595 | */ |
| 1596 | lea 11(%rdi), %r10 |
| 1597 | and $0xfff, %r10 /* offset into 4K page */ |
| 1598 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1599 | |
| 1600 | .p2align 4 |
| 1601 | LABEL(loop_ashr_11): |
| 1602 | add $16, %r10 |
| 1603 | jg LABEL(nibble_ashr_11) |
| 1604 | |
| 1605 | LABEL(gobble_ashr_11): |
| 1606 | movdqa (%rsi, %rcx), %xmm1 |
| 1607 | movdqa (%rdi, %rcx), %xmm2 |
| 1608 | movdqa %xmm2, %xmm4 |
| 1609 | |
| 1610 | #ifndef USE_SSSE3 |
| 1611 | psrldq $11, %xmm3 |
| 1612 | pslldq $5, %xmm2 |
| 1613 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1614 | #else |
| 1615 | palignr $11, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1616 | #endif |
| 1617 | TOLOWER (%xmm1, %xmm2) |
| 1618 | |
| 1619 | pcmpeqb %xmm1, %xmm0 |
| 1620 | pcmpeqb %xmm2, %xmm1 |
| 1621 | psubb %xmm0, %xmm1 |
| 1622 | pmovmskb %xmm1, %edx |
| 1623 | sub $0xffff, %edx |
| 1624 | jnz LABEL(exit) |
| 1625 | |
| 1626 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1627 | sub $16, %r11 |
| 1628 | jbe LABEL(strcmp_exitz) |
| 1629 | #endif |
| 1630 | |
| 1631 | add $16, %rcx |
| 1632 | movdqa %xmm4, %xmm3 |
| 1633 | |
| 1634 | add $16, %r10 |
| 1635 | jg LABEL(nibble_ashr_11) /* cross page boundary */ |
| 1636 | |
| 1637 | movdqa (%rsi, %rcx), %xmm1 |
| 1638 | movdqa (%rdi, %rcx), %xmm2 |
| 1639 | movdqa %xmm2, %xmm4 |
| 1640 | |
| 1641 | #ifndef USE_SSSE3 |
| 1642 | psrldq $11, %xmm3 |
| 1643 | pslldq $5, %xmm2 |
| 1644 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1645 | #else |
| 1646 | palignr $11, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1647 | #endif |
| 1648 | TOLOWER (%xmm1, %xmm2) |
| 1649 | |
| 1650 | pcmpeqb %xmm1, %xmm0 |
| 1651 | pcmpeqb %xmm2, %xmm1 |
| 1652 | psubb %xmm0, %xmm1 |
| 1653 | pmovmskb %xmm1, %edx |
| 1654 | sub $0xffff, %edx |
| 1655 | jnz LABEL(exit) |
| 1656 | |
| 1657 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1658 | sub $16, %r11 |
| 1659 | jbe LABEL(strcmp_exitz) |
| 1660 | #endif |
| 1661 | |
| 1662 | add $16, %rcx |
| 1663 | movdqa %xmm4, %xmm3 |
| 1664 | jmp LABEL(loop_ashr_11) |
| 1665 | |
| 1666 | .p2align 4 |
| 1667 | LABEL(nibble_ashr_11): |
| 1668 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1669 | pmovmskb %xmm0, %edx |
| 1670 | test $0xf800, %edx |
| 1671 | jnz LABEL(ashr_11_exittail) |
| 1672 | |
| 1673 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1674 | cmp $5, %r11 |
| 1675 | jbe LABEL(ashr_11_exittail) |
| 1676 | #endif |
| 1677 | |
| 1678 | pxor %xmm0, %xmm0 |
| 1679 | sub $0x1000, %r10 |
| 1680 | jmp LABEL(gobble_ashr_11) |
| 1681 | |
| 1682 | .p2align 4 |
| 1683 | LABEL(ashr_11_exittail): |
| 1684 | movdqa (%rsi, %rcx), %xmm1 |
| 1685 | psrldq $11, %xmm0 |
| 1686 | psrldq $11, %xmm3 |
| 1687 | jmp LABEL(aftertail) |
| 1688 | |
| 1689 | /* |
| 1690 | * The following cases will be handled by ashr_12 |
| 1691 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1692 | * n(4~15) n - 4 11(15 +(n - 4) - n) ashr_12 |
| 1693 | */ |
| 1694 | .p2align 4 |
| 1695 | LABEL(ashr_12): |
| 1696 | pxor %xmm0, %xmm0 |
| 1697 | movdqa (%rdi), %xmm2 |
| 1698 | movdqa (%rsi), %xmm1 |
| 1699 | pcmpeqb %xmm1, %xmm0 |
| 1700 | pslldq $4, %xmm2 |
| 1701 | TOLOWER (%xmm1, %xmm2) |
| 1702 | pcmpeqb %xmm1, %xmm2 |
| 1703 | psubb %xmm0, %xmm2 |
| 1704 | pmovmskb %xmm2, %r9d |
| 1705 | shr %cl, %edx |
| 1706 | shr %cl, %r9d |
| 1707 | sub %r9d, %edx |
| 1708 | jnz LABEL(less32bytes) |
| 1709 | movdqa (%rdi), %xmm3 |
| 1710 | |
| 1711 | UPDATE_STRNCMP_COUNTER |
| 1712 | |
| 1713 | pxor %xmm0, %xmm0 |
| 1714 | mov $16, %rcx /* index for loads */ |
| 1715 | mov $12, %r9d /* byte position left over from less32bytes case */ |
| 1716 | /* |
| 1717 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1718 | * When %r10 goes positive we have crossed a page boundary and |
| 1719 | * need to do a nibble. |
| 1720 | */ |
| 1721 | lea 12(%rdi), %r10 |
| 1722 | and $0xfff, %r10 /* offset into 4K page */ |
| 1723 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1724 | |
| 1725 | .p2align 4 |
| 1726 | LABEL(loop_ashr_12): |
| 1727 | add $16, %r10 |
| 1728 | jg LABEL(nibble_ashr_12) |
| 1729 | |
| 1730 | LABEL(gobble_ashr_12): |
| 1731 | movdqa (%rsi, %rcx), %xmm1 |
| 1732 | movdqa (%rdi, %rcx), %xmm2 |
| 1733 | movdqa %xmm2, %xmm4 |
| 1734 | |
| 1735 | #ifndef USE_SSSE3 |
| 1736 | psrldq $12, %xmm3 |
| 1737 | pslldq $4, %xmm2 |
| 1738 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1739 | #else |
| 1740 | palignr $12, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1741 | #endif |
| 1742 | TOLOWER (%xmm1, %xmm2) |
| 1743 | |
| 1744 | pcmpeqb %xmm1, %xmm0 |
| 1745 | pcmpeqb %xmm2, %xmm1 |
| 1746 | psubb %xmm0, %xmm1 |
| 1747 | pmovmskb %xmm1, %edx |
| 1748 | sub $0xffff, %edx |
| 1749 | jnz LABEL(exit) |
| 1750 | |
| 1751 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1752 | sub $16, %r11 |
| 1753 | jbe LABEL(strcmp_exitz) |
| 1754 | #endif |
| 1755 | |
| 1756 | add $16, %rcx |
| 1757 | movdqa %xmm4, %xmm3 |
| 1758 | |
| 1759 | add $16, %r10 |
| 1760 | jg LABEL(nibble_ashr_12) /* cross page boundary */ |
| 1761 | |
| 1762 | movdqa (%rsi, %rcx), %xmm1 |
| 1763 | movdqa (%rdi, %rcx), %xmm2 |
| 1764 | movdqa %xmm2, %xmm4 |
| 1765 | |
| 1766 | #ifndef USE_SSSE3 |
| 1767 | psrldq $12, %xmm3 |
| 1768 | pslldq $4, %xmm2 |
| 1769 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1770 | #else |
| 1771 | palignr $12, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1772 | #endif |
| 1773 | TOLOWER (%xmm1, %xmm2) |
| 1774 | |
| 1775 | pcmpeqb %xmm1, %xmm0 |
| 1776 | pcmpeqb %xmm2, %xmm1 |
| 1777 | psubb %xmm0, %xmm1 |
| 1778 | pmovmskb %xmm1, %edx |
| 1779 | sub $0xffff, %edx |
| 1780 | jnz LABEL(exit) |
| 1781 | |
| 1782 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1783 | sub $16, %r11 |
| 1784 | jbe LABEL(strcmp_exitz) |
| 1785 | #endif |
| 1786 | |
| 1787 | add $16, %rcx |
| 1788 | movdqa %xmm4, %xmm3 |
| 1789 | jmp LABEL(loop_ashr_12) |
| 1790 | |
| 1791 | .p2align 4 |
| 1792 | LABEL(nibble_ashr_12): |
| 1793 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1794 | pmovmskb %xmm0, %edx |
| 1795 | test $0xf000, %edx |
| 1796 | jnz LABEL(ashr_12_exittail) |
| 1797 | |
| 1798 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1799 | cmp $4, %r11 |
| 1800 | jbe LABEL(ashr_12_exittail) |
| 1801 | #endif |
| 1802 | |
| 1803 | pxor %xmm0, %xmm0 |
| 1804 | sub $0x1000, %r10 |
| 1805 | jmp LABEL(gobble_ashr_12) |
| 1806 | |
| 1807 | .p2align 4 |
| 1808 | LABEL(ashr_12_exittail): |
| 1809 | movdqa (%rsi, %rcx), %xmm1 |
| 1810 | psrldq $12, %xmm0 |
| 1811 | psrldq $12, %xmm3 |
| 1812 | jmp LABEL(aftertail) |
| 1813 | |
| 1814 | /* |
| 1815 | * The following cases will be handled by ashr_13 |
| 1816 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1817 | * n(3~15) n - 3 12(15 +(n - 3) - n) ashr_13 |
| 1818 | */ |
| 1819 | .p2align 4 |
| 1820 | LABEL(ashr_13): |
| 1821 | pxor %xmm0, %xmm0 |
| 1822 | movdqa (%rdi), %xmm2 |
| 1823 | movdqa (%rsi), %xmm1 |
| 1824 | pcmpeqb %xmm1, %xmm0 |
| 1825 | pslldq $3, %xmm2 |
| 1826 | TOLOWER (%xmm1, %xmm2) |
| 1827 | pcmpeqb %xmm1, %xmm2 |
| 1828 | psubb %xmm0, %xmm2 |
| 1829 | pmovmskb %xmm2, %r9d |
| 1830 | shr %cl, %edx |
| 1831 | shr %cl, %r9d |
| 1832 | sub %r9d, %edx |
| 1833 | jnz LABEL(less32bytes) |
| 1834 | movdqa (%rdi), %xmm3 |
| 1835 | |
| 1836 | UPDATE_STRNCMP_COUNTER |
| 1837 | |
| 1838 | pxor %xmm0, %xmm0 |
| 1839 | mov $16, %rcx /* index for loads */ |
| 1840 | mov $13, %r9d /* byte position left over from less32bytes case */ |
| 1841 | /* |
| 1842 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1843 | * When %r10 goes positive we have crossed a page boundary and |
| 1844 | * need to do a nibble. |
| 1845 | */ |
| 1846 | lea 13(%rdi), %r10 |
| 1847 | and $0xfff, %r10 /* offset into 4K page */ |
| 1848 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1849 | |
| 1850 | .p2align 4 |
| 1851 | LABEL(loop_ashr_13): |
| 1852 | add $16, %r10 |
| 1853 | jg LABEL(nibble_ashr_13) |
| 1854 | |
| 1855 | LABEL(gobble_ashr_13): |
| 1856 | movdqa (%rsi, %rcx), %xmm1 |
| 1857 | movdqa (%rdi, %rcx), %xmm2 |
| 1858 | movdqa %xmm2, %xmm4 |
| 1859 | |
| 1860 | #ifndef USE_SSSE3 |
| 1861 | psrldq $13, %xmm3 |
| 1862 | pslldq $3, %xmm2 |
| 1863 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1864 | #else |
| 1865 | palignr $13, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1866 | #endif |
| 1867 | TOLOWER (%xmm1, %xmm2) |
| 1868 | |
| 1869 | pcmpeqb %xmm1, %xmm0 |
| 1870 | pcmpeqb %xmm2, %xmm1 |
| 1871 | psubb %xmm0, %xmm1 |
| 1872 | pmovmskb %xmm1, %edx |
| 1873 | sub $0xffff, %edx |
| 1874 | jnz LABEL(exit) |
| 1875 | |
| 1876 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1877 | sub $16, %r11 |
| 1878 | jbe LABEL(strcmp_exitz) |
| 1879 | #endif |
| 1880 | |
| 1881 | add $16, %rcx |
| 1882 | movdqa %xmm4, %xmm3 |
| 1883 | |
| 1884 | add $16, %r10 |
| 1885 | jg LABEL(nibble_ashr_13) /* cross page boundary */ |
| 1886 | |
| 1887 | movdqa (%rsi, %rcx), %xmm1 |
| 1888 | movdqa (%rdi, %rcx), %xmm2 |
| 1889 | movdqa %xmm2, %xmm4 |
| 1890 | |
| 1891 | #ifndef USE_SSSE3 |
| 1892 | psrldq $13, %xmm3 |
| 1893 | pslldq $3, %xmm2 |
| 1894 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1895 | #else |
| 1896 | palignr $13, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1897 | #endif |
| 1898 | TOLOWER (%xmm1, %xmm2) |
| 1899 | |
| 1900 | pcmpeqb %xmm1, %xmm0 |
| 1901 | pcmpeqb %xmm2, %xmm1 |
| 1902 | psubb %xmm0, %xmm1 |
| 1903 | pmovmskb %xmm1, %edx |
| 1904 | sub $0xffff, %edx |
| 1905 | jnz LABEL(exit) |
| 1906 | |
| 1907 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1908 | sub $16, %r11 |
| 1909 | jbe LABEL(strcmp_exitz) |
| 1910 | #endif |
| 1911 | |
| 1912 | add $16, %rcx |
| 1913 | movdqa %xmm4, %xmm3 |
| 1914 | jmp LABEL(loop_ashr_13) |
| 1915 | |
| 1916 | .p2align 4 |
| 1917 | LABEL(nibble_ashr_13): |
| 1918 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 1919 | pmovmskb %xmm0, %edx |
| 1920 | test $0xe000, %edx |
| 1921 | jnz LABEL(ashr_13_exittail) |
| 1922 | |
| 1923 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 1924 | cmp $3, %r11 |
| 1925 | jbe LABEL(ashr_13_exittail) |
| 1926 | #endif |
| 1927 | |
| 1928 | pxor %xmm0, %xmm0 |
| 1929 | sub $0x1000, %r10 |
| 1930 | jmp LABEL(gobble_ashr_13) |
| 1931 | |
| 1932 | .p2align 4 |
| 1933 | LABEL(ashr_13_exittail): |
| 1934 | movdqa (%rsi, %rcx), %xmm1 |
| 1935 | psrldq $13, %xmm0 |
| 1936 | psrldq $13, %xmm3 |
| 1937 | jmp LABEL(aftertail) |
| 1938 | |
| 1939 | /* |
| 1940 | * The following cases will be handled by ashr_14 |
| 1941 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 1942 | * n(2~15) n - 2 13(15 +(n - 2) - n) ashr_14 |
| 1943 | */ |
| 1944 | .p2align 4 |
| 1945 | LABEL(ashr_14): |
| 1946 | pxor %xmm0, %xmm0 |
| 1947 | movdqa (%rdi), %xmm2 |
| 1948 | movdqa (%rsi), %xmm1 |
| 1949 | pcmpeqb %xmm1, %xmm0 |
| 1950 | pslldq $2, %xmm2 |
| 1951 | TOLOWER (%xmm1, %xmm2) |
| 1952 | pcmpeqb %xmm1, %xmm2 |
| 1953 | psubb %xmm0, %xmm2 |
| 1954 | pmovmskb %xmm2, %r9d |
| 1955 | shr %cl, %edx |
| 1956 | shr %cl, %r9d |
| 1957 | sub %r9d, %edx |
| 1958 | jnz LABEL(less32bytes) |
| 1959 | movdqa (%rdi), %xmm3 |
| 1960 | |
| 1961 | UPDATE_STRNCMP_COUNTER |
| 1962 | |
| 1963 | pxor %xmm0, %xmm0 |
| 1964 | mov $16, %rcx /* index for loads */ |
| 1965 | mov $14, %r9d /* byte position left over from less32bytes case */ |
| 1966 | /* |
| 1967 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 1968 | * When %r10 goes positive we have crossed a page boundary and |
| 1969 | * need to do a nibble. |
| 1970 | */ |
| 1971 | lea 14(%rdi), %r10 |
| 1972 | and $0xfff, %r10 /* offset into 4K page */ |
| 1973 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 1974 | |
| 1975 | .p2align 4 |
| 1976 | LABEL(loop_ashr_14): |
| 1977 | add $16, %r10 |
| 1978 | jg LABEL(nibble_ashr_14) |
| 1979 | |
| 1980 | LABEL(gobble_ashr_14): |
| 1981 | movdqa (%rsi, %rcx), %xmm1 |
| 1982 | movdqa (%rdi, %rcx), %xmm2 |
| 1983 | movdqa %xmm2, %xmm4 |
| 1984 | |
| 1985 | #ifndef USE_SSSE3 |
| 1986 | psrldq $14, %xmm3 |
| 1987 | pslldq $2, %xmm2 |
| 1988 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1989 | #else |
| 1990 | palignr $14, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 1991 | #endif |
| 1992 | TOLOWER (%xmm1, %xmm2) |
| 1993 | |
| 1994 | pcmpeqb %xmm1, %xmm0 |
| 1995 | pcmpeqb %xmm2, %xmm1 |
| 1996 | psubb %xmm0, %xmm1 |
| 1997 | pmovmskb %xmm1, %edx |
| 1998 | sub $0xffff, %edx |
| 1999 | jnz LABEL(exit) |
| 2000 | |
| 2001 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 2002 | sub $16, %r11 |
| 2003 | jbe LABEL(strcmp_exitz) |
| 2004 | #endif |
| 2005 | |
| 2006 | add $16, %rcx |
| 2007 | movdqa %xmm4, %xmm3 |
| 2008 | |
| 2009 | add $16, %r10 |
| 2010 | jg LABEL(nibble_ashr_14) /* cross page boundary */ |
| 2011 | |
| 2012 | movdqa (%rsi, %rcx), %xmm1 |
| 2013 | movdqa (%rdi, %rcx), %xmm2 |
| 2014 | movdqa %xmm2, %xmm4 |
| 2015 | |
| 2016 | #ifndef USE_SSSE3 |
| 2017 | psrldq $14, %xmm3 |
| 2018 | pslldq $2, %xmm2 |
| 2019 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 2020 | #else |
| 2021 | palignr $14, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 2022 | #endif |
| 2023 | TOLOWER (%xmm1, %xmm2) |
| 2024 | |
| 2025 | pcmpeqb %xmm1, %xmm0 |
| 2026 | pcmpeqb %xmm2, %xmm1 |
| 2027 | psubb %xmm0, %xmm1 |
| 2028 | pmovmskb %xmm1, %edx |
| 2029 | sub $0xffff, %edx |
| 2030 | jnz LABEL(exit) |
| 2031 | |
| 2032 | #if defined USE_AS_STRNCMP | defined USE_AS_STRNCASECMP_L |
| 2033 | sub $16, %r11 |
| 2034 | jbe LABEL(strcmp_exitz) |
| 2035 | #endif |
| 2036 | |
| 2037 | add $16, %rcx |
| 2038 | movdqa %xmm4, %xmm3 |
| 2039 | jmp LABEL(loop_ashr_14) |
| 2040 | |
| 2041 | .p2align 4 |
| 2042 | LABEL(nibble_ashr_14): |
| 2043 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 2044 | pmovmskb %xmm0, %edx |
| 2045 | test $0xc000, %edx |
| 2046 | jnz LABEL(ashr_14_exittail) |
| 2047 | |
| 2048 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 2049 | cmp $2, %r11 |
| 2050 | jbe LABEL(ashr_14_exittail) |
| 2051 | #endif |
| 2052 | |
| 2053 | pxor %xmm0, %xmm0 |
| 2054 | sub $0x1000, %r10 |
| 2055 | jmp LABEL(gobble_ashr_14) |
| 2056 | |
| 2057 | .p2align 4 |
| 2058 | LABEL(ashr_14_exittail): |
| 2059 | movdqa (%rsi, %rcx), %xmm1 |
| 2060 | psrldq $14, %xmm0 |
| 2061 | psrldq $14, %xmm3 |
| 2062 | jmp LABEL(aftertail) |
| 2063 | |
| 2064 | /* |
| 2065 | * The following cases will be handled by ashr_15 |
| 2066 | * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case |
| 2067 | * n(1~15) n - 1 14(15 +(n - 1) - n) ashr_15 |
| 2068 | */ |
| 2069 | .p2align 4 |
| 2070 | LABEL(ashr_15): |
| 2071 | pxor %xmm0, %xmm0 |
| 2072 | movdqa (%rdi), %xmm2 |
| 2073 | movdqa (%rsi), %xmm1 |
| 2074 | pcmpeqb %xmm1, %xmm0 |
| 2075 | pslldq $1, %xmm2 |
| 2076 | TOLOWER (%xmm1, %xmm2) |
| 2077 | pcmpeqb %xmm1, %xmm2 |
| 2078 | psubb %xmm0, %xmm2 |
| 2079 | pmovmskb %xmm2, %r9d |
| 2080 | shr %cl, %edx |
| 2081 | shr %cl, %r9d |
| 2082 | sub %r9d, %edx |
| 2083 | jnz LABEL(less32bytes) |
| 2084 | |
| 2085 | movdqa (%rdi), %xmm3 |
| 2086 | |
| 2087 | UPDATE_STRNCMP_COUNTER |
| 2088 | |
| 2089 | pxor %xmm0, %xmm0 |
| 2090 | mov $16, %rcx /* index for loads */ |
| 2091 | mov $15, %r9d /* byte position left over from less32bytes case */ |
| 2092 | /* |
| 2093 | * Setup %r10 value allows us to detect crossing a page boundary. |
| 2094 | * When %r10 goes positive we have crossed a page boundary and |
| 2095 | * need to do a nibble. |
| 2096 | */ |
| 2097 | lea 15(%rdi), %r10 |
| 2098 | and $0xfff, %r10 /* offset into 4K page */ |
| 2099 | |
| 2100 | sub $0x1000, %r10 /* subtract 4K pagesize */ |
| 2101 | |
| 2102 | .p2align 4 |
| 2103 | LABEL(loop_ashr_15): |
| 2104 | add $16, %r10 |
| 2105 | jg LABEL(nibble_ashr_15) |
| 2106 | |
| 2107 | LABEL(gobble_ashr_15): |
| 2108 | movdqa (%rsi, %rcx), %xmm1 |
| 2109 | movdqa (%rdi, %rcx), %xmm2 |
| 2110 | movdqa %xmm2, %xmm4 |
| 2111 | |
| 2112 | #ifndef USE_SSSE3 |
| 2113 | psrldq $15, %xmm3 |
| 2114 | pslldq $1, %xmm2 |
| 2115 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 2116 | #else |
| 2117 | palignr $15, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 2118 | #endif |
| 2119 | TOLOWER (%xmm1, %xmm2) |
| 2120 | |
| 2121 | pcmpeqb %xmm1, %xmm0 |
| 2122 | pcmpeqb %xmm2, %xmm1 |
| 2123 | psubb %xmm0, %xmm1 |
| 2124 | pmovmskb %xmm1, %edx |
| 2125 | sub $0xffff, %edx |
| 2126 | jnz LABEL(exit) |
| 2127 | |
| 2128 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 2129 | sub $16, %r11 |
| 2130 | jbe LABEL(strcmp_exitz) |
| 2131 | #endif |
| 2132 | |
| 2133 | add $16, %rcx |
| 2134 | movdqa %xmm4, %xmm3 |
| 2135 | |
| 2136 | add $16, %r10 |
| 2137 | jg LABEL(nibble_ashr_15) /* cross page boundary */ |
| 2138 | |
| 2139 | movdqa (%rsi, %rcx), %xmm1 |
| 2140 | movdqa (%rdi, %rcx), %xmm2 |
| 2141 | movdqa %xmm2, %xmm4 |
| 2142 | |
| 2143 | #ifndef USE_SSSE3 |
| 2144 | psrldq $15, %xmm3 |
| 2145 | pslldq $1, %xmm2 |
| 2146 | por %xmm3, %xmm2 /* merge into one 16byte value */ |
| 2147 | #else |
| 2148 | palignr $15, %xmm3, %xmm2 /* merge into one 16byte value */ |
| 2149 | #endif |
| 2150 | TOLOWER (%xmm1, %xmm2) |
| 2151 | |
| 2152 | pcmpeqb %xmm1, %xmm0 |
| 2153 | pcmpeqb %xmm2, %xmm1 |
| 2154 | psubb %xmm0, %xmm1 |
| 2155 | pmovmskb %xmm1, %edx |
| 2156 | sub $0xffff, %edx |
| 2157 | jnz LABEL(exit) |
| 2158 | |
| 2159 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 2160 | sub $16, %r11 |
| 2161 | jbe LABEL(strcmp_exitz) |
| 2162 | #endif |
| 2163 | |
| 2164 | add $16, %rcx |
| 2165 | movdqa %xmm4, %xmm3 |
| 2166 | jmp LABEL(loop_ashr_15) |
| 2167 | |
| 2168 | .p2align 4 |
| 2169 | LABEL(nibble_ashr_15): |
| 2170 | pcmpeqb %xmm3, %xmm0 /* check nibble for null char */ |
| 2171 | pmovmskb %xmm0, %edx |
| 2172 | test $0x8000, %edx |
| 2173 | jnz LABEL(ashr_15_exittail) |
| 2174 | |
| 2175 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 2176 | cmpq $1, %r11 |
| 2177 | jbe LABEL(ashr_15_exittail) |
| 2178 | #endif |
| 2179 | |
| 2180 | pxor %xmm0, %xmm0 |
| 2181 | sub $0x1000, %r10 |
| 2182 | jmp LABEL(gobble_ashr_15) |
| 2183 | |
| 2184 | .p2align 4 |
| 2185 | LABEL(ashr_15_exittail): |
| 2186 | movdqa (%rsi, %rcx), %xmm1 |
| 2187 | psrldq $15, %xmm3 |
| 2188 | psrldq $15, %xmm0 |
| 2189 | |
| 2190 | .p2align 4 |
| 2191 | LABEL(aftertail): |
| 2192 | TOLOWER (%xmm1, %xmm3) |
| 2193 | pcmpeqb %xmm3, %xmm1 |
| 2194 | psubb %xmm0, %xmm1 |
| 2195 | pmovmskb %xmm1, %edx |
| 2196 | not %edx |
| 2197 | |
| 2198 | .p2align 4 |
| 2199 | LABEL(exit): |
| 2200 | lea -16(%r9, %rcx), %rax /* locate the exact offset for rdi */ |
| 2201 | LABEL(less32bytes): |
| 2202 | lea (%rdi, %rax), %rdi /* locate the exact address for first operand(rdi) */ |
| 2203 | lea (%rsi, %rcx), %rsi /* locate the exact address for second operand(rsi) */ |
| 2204 | test %r8d, %r8d |
| 2205 | jz LABEL(ret) |
| 2206 | xchg %rsi, %rdi /* recover original order according to flag(%r8d) */ |
| 2207 | |
| 2208 | .p2align 4 |
| 2209 | LABEL(ret): |
| 2210 | LABEL(less16bytes): |
| 2211 | bsf %rdx, %rdx /* find and store bit index in %rdx */ |
| 2212 | |
| 2213 | #if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L |
| 2214 | sub %rdx, %r11 |
| 2215 | jbe LABEL(strcmp_exitz) |
| 2216 | #endif |
| 2217 | movzbl (%rsi, %rdx), %ecx |
| 2218 | movzbl (%rdi, %rdx), %eax |
| 2219 | |
| 2220 | #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
| 2221 | leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rdx |
| 2222 | movl (%rdx,%rcx,4), %ecx |
| 2223 | movl (%rdx,%rax,4), %eax |
| 2224 | #endif |
| 2225 | |
| 2226 | sub %ecx, %eax |
| 2227 | ret |
| 2228 | |
| 2229 | LABEL(strcmp_exitz): |
| 2230 | xor %eax, %eax |
| 2231 | ret |
| 2232 | |
| 2233 | .p2align 4 |
| 2234 | LABEL(Byte0): |
| 2235 | movzx (%rsi), %ecx |
| 2236 | movzx (%rdi), %eax |
| 2237 | |
| 2238 | #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L |
| 2239 | leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rdx |
| 2240 | movl (%rdx,%rcx,4), %ecx |
| 2241 | movl (%rdx,%rax,4), %eax |
| 2242 | #endif |
| 2243 | |
| 2244 | sub %ecx, %eax |
| 2245 | ret |
| 2246 | END (STRCMP) |
| 2247 | |
| 2248 | .section .rodata,"a" ,@progbits |
| 2249 | .p2align 3 |
| 2250 | LABEL(unaligned_table): |
| 2251 | .int LABEL(ashr_1) - LABEL(unaligned_table) |
| 2252 | .int LABEL(ashr_2) - LABEL(unaligned_table) |
| 2253 | .int LABEL(ashr_3) - LABEL(unaligned_table) |
| 2254 | .int LABEL(ashr_4) - LABEL(unaligned_table) |
| 2255 | .int LABEL(ashr_5) - LABEL(unaligned_table) |
| 2256 | .int LABEL(ashr_6) - LABEL(unaligned_table) |
| 2257 | .int LABEL(ashr_7) - LABEL(unaligned_table) |
| 2258 | .int LABEL(ashr_8) - LABEL(unaligned_table) |
| 2259 | .int LABEL(ashr_9) - LABEL(unaligned_table) |
| 2260 | .int LABEL(ashr_10) - LABEL(unaligned_table) |
| 2261 | .int LABEL(ashr_11) - LABEL(unaligned_table) |
| 2262 | .int LABEL(ashr_12) - LABEL(unaligned_table) |
| 2263 | .int LABEL(ashr_13) - LABEL(unaligned_table) |
| 2264 | .int LABEL(ashr_14) - LABEL(unaligned_table) |
| 2265 | .int LABEL(ashr_15) - LABEL(unaligned_table) |
| 2266 | .int LABEL(ashr_0) - LABEL(unaligned_table) |
| 2267 | libc_hidden_builtin_def (STRCMP) |
| 2268 | |