sysdeps/x86_64/multiarch/strcpy-ssse3.S

   1 /* strcpy with SSSE3
   2    Copyright (C) 2011 Free Software Foundation, Inc.
   3    Contributed by Intel Corporation.
   4    This file is part of the GNU C Library.
   5
   6    The GNU C Library is free software; you can redistribute it and/or
   7    modify it under the terms of the GNU Lesser General Public
   8    License as published by the Free Software Foundation; either
   9    version 2.1 of the License, or (at your option) any later version.
  10
  11    The GNU C Library is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14    Lesser General Public License for more details.
  15
  16    You should have received a copy of the GNU Lesser General Public
  17    License along with the GNU C Library; if not, write to the Free
  18    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  19    02111-1307 USA.  */
  20
  21 #ifndef NOT_IN_libc
  22
  23 # ifndef USE_AS_STRCAT
  24 #  include <sysdep.h>
  25
  26 #  ifndef STRCPY
  27 #   define STRCPY  __strcpy_ssse3
  28 #  endif
  29
  30         .section .text.ssse3,"ax",@progbits
  31 ENTRY (STRCPY)
  32         mov     %rsi, %rcx
  33 #  ifdef USE_AS_STRNCPY
  34         mov     %rdx, %r8
  35 #  endif
  36         mov     %rdi, %rdx
  37 #  ifdef USE_AS_STRNCPY
  38         test    %r8, %r8
  39         jz      L(Exit0)
  40         cmp     $8, %r8
  41         jbe     L(StrncpyExit8Bytes)
  42 #  endif
  43         cmpb    $0, (%rcx)
  44         jz      L(Exit1)
  45         cmpb    $0, 1(%rcx)
  46         jz      L(Exit2)
  47         cmpb    $0, 2(%rcx)
  48         jz      L(Exit3)
  49         cmpb    $0, 3(%rcx)
  50         jz      L(Exit4)
  51         cmpb    $0, 4(%rcx)
  52         jz      L(Exit5)
  53         cmpb    $0, 5(%rcx)
  54         jz      L(Exit6)
  55         cmpb    $0, 6(%rcx)
  56         jz      L(Exit7)
  57         cmpb    $0, 7(%rcx)
  58         jz      L(Exit8)
  59 #  ifdef USE_AS_STRNCPY
  60         cmp     $16, %r8
  61         jb      L(StrncpyExit15Bytes)
  62 #  endif
  63         cmpb    $0, 8(%rcx)
  64         jz      L(Exit9)
  65         cmpb    $0, 9(%rcx)
  66         jz      L(Exit10)
  67         cmpb    $0, 10(%rcx)
  68         jz      L(Exit11)
  69         cmpb    $0, 11(%rcx)
  70         jz      L(Exit12)
  71         cmpb    $0, 12(%rcx)
  72         jz      L(Exit13)
  73         cmpb    $0, 13(%rcx)
  74         jz      L(Exit14)
  75         cmpb    $0, 14(%rcx)
  76         jz      L(Exit15)
  77 #  ifdef USE_AS_STRNCPY
  78         cmp     $16, %r8
  79         je      L(Exit16)
  80 #  endif
  81         cmpb    $0, 15(%rcx)
  82         jz      L(Exit16)
  83 # endif
  84
  85 # ifdef USE_AS_STRNCPY
  86         mov     %rcx, %rsi
  87         sub     $16, %r8
  88         and     $0xf, %rsi
  89
  90 /* add 16 bytes rcx_shift to r8 */
  91         add     %rsi, %r8
  92 # endif
  93         lea     16(%rcx), %rsi
  94 /* Now:
  95         rsi     = alignment_16(rcx) + rcx_shift + 16;
  96         rcx_shift = rcx - alignment_16(rcx)
  97 */
  98         and     $-16, %rsi
  99 /* Now:
 100         rsi     = alignment_16(rcx) + 16
 101 */
 102         pxor    %xmm0, %xmm0
 103         mov     (%rcx), %r9
 104         mov     %r9, (%rdx)
 105 /*
 106         look    if there is zero symbol in next 16 bytes of string
 107         from    rsi to rsi + 15 and form mask in xmm0
 108 */
 109         pcmpeqb (%rsi), %xmm0
 110         mov     8(%rcx), %r9
 111         mov     %r9, 8(%rdx)
 112
 113 /* convert byte mask in xmm0 to bit mask */
 114
 115         pmovmskb %xmm0, %rax
 116         sub     %rcx, %rsi
 117
 118 /* rsi = 16 - rcx_shift */
 119
 120 /* rax = 0: there isn't end of string from position rsi to rsi+15 */
 121
 122 # ifdef USE_AS_STRNCPY
 123         sub     $16, %r8
 124         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 125 # endif
 126         test    %rax, %rax
 127         jnz     L(CopyFrom1To16Bytes)
 128
 129         mov     %rdx, %rax
 130         lea     16(%rdx), %rdx
 131 /* Now:
 132         rdx     = rdx + 16 = alignment_16(rdx) + rdx_shift + 16
 133 */
 134         and     $-16, %rdx
 135
 136 /* Now: rdx = alignment_16(rdx) + 16 */
 137
 138         sub     %rdx, %rax
 139
 140 /* Now: rax = rdx_shift - 16 */
 141
 142 # ifdef USE_AS_STRNCPY
 143         add     %rax, %rsi
 144         lea     -1(%rsi), %rsi
 145         and     $1<<31, %esi
 146         test    %rsi, %rsi
 147         jnz     L(ContinueCopy)
 148         lea     16(%r8), %r8
 149
 150 L(ContinueCopy):
 151 # endif
 152         sub     %rax, %rcx
 153 /* Now:
 154         case    rcx_shift >= rdx_shift:
 155         rcx     = alignment_16(rcx) + (rcx_shift  - rdx_shift) + 16
 156         case    rcx_shift < rdx_shift:
 157         rcx     = alignment_16(rcx) + (16 + rcx_shift  - rdx_shift)
 158 */
 159         mov     %rcx, %rax
 160         and     $0xf, %rax
 161 /* Now:
 162         case    rcx_shift >= rdx_shift: rax = rcx_shift  - rdx_shift
 163         case    rcx_shift < rdx_shift: rax = (16 + rcx_shift  - rdx_shift)
 164         rax     can be 0, 1,    ..., 15
 165 */
 166         mov     $0, %rsi
 167
 168 /* case: rcx_shift == rdx_shift */
 169
 170         jz      L(Align16Both)
 171
 172         cmp     $8, %rax
 173         jae     L(ShlHigh8)
 174         cmp     $1, %rax
 175         je      L(Shl1)
 176         cmp     $2, %rax
 177         je      L(Shl2)
 178         cmp     $3, %rax
 179         je      L(Shl3)
 180         cmp     $4, %rax
 181         je      L(Shl4)
 182         cmp     $5, %rax
 183         je      L(Shl5)
 184         cmp     $6, %rax
 185         je      L(Shl6)
 186         jmp     L(Shl7)
 187
 188 L(ShlHigh8):
 189         je      L(Shl8)
 190         cmp     $9, %rax
 191         je      L(Shl9)
 192         cmp     $10, %rax
 193         je      L(Shl10)
 194         cmp     $11, %rax
 195         je      L(Shl11)
 196         cmp     $12, %rax
 197         je      L(Shl12)
 198         cmp     $13, %rax
 199         je      L(Shl13)
 200         cmp     $14, %rax
 201         je      L(Shl14)
 202         jmp     L(Shl15)
 203
 204 L(Align16Both):
 205         movaps  (%rcx), %xmm1
 206         movaps  16(%rcx), %xmm2
 207         movaps  %xmm1, (%rdx)
 208         pcmpeqb %xmm2, %xmm0
 209         pmovmskb %xmm0, %rax
 210         lea     16(%rsi), %rsi
 211 # ifdef USE_AS_STRNCPY
 212         sub     $16, %r8
 213         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 214 # endif
 215         test    %rax, %rax
 216         jnz     L(CopyFrom1To16Bytes)
 217
 218         movaps  16(%rcx, %rsi), %xmm3
 219         movaps  %xmm2, (%rdx, %rsi)
 220         pcmpeqb %xmm3, %xmm0
 221         pmovmskb %xmm0, %rax
 222         lea     16(%rsi), %rsi
 223 # ifdef USE_AS_STRNCPY
 224         sub     $16, %r8
 225         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 226 # endif
 227         test    %rax, %rax
 228         jnz     L(CopyFrom1To16Bytes)
 229
 230         movaps  16(%rcx, %rsi), %xmm4
 231         movaps  %xmm3, (%rdx, %rsi)
 232         pcmpeqb %xmm4, %xmm0
 233         pmovmskb %xmm0, %rax
 234         lea     16(%rsi), %rsi
 235 # ifdef USE_AS_STRNCPY
 236         sub     $16, %r8
 237         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 238 # endif
 239         test    %rax, %rax
 240         jnz     L(CopyFrom1To16Bytes)
 241
 242         movaps  16(%rcx, %rsi), %xmm1
 243         movaps  %xmm4, (%rdx, %rsi)
 244         pcmpeqb %xmm1, %xmm0
 245         pmovmskb %xmm0, %rax
 246         lea     16(%rsi), %rsi
 247 # ifdef USE_AS_STRNCPY
 248         sub     $16, %r8
 249         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 250 # endif
 251         test    %rax, %rax
 252         jnz     L(CopyFrom1To16Bytes)
 253
 254         movaps  16(%rcx, %rsi), %xmm2
 255         movaps  %xmm1, (%rdx, %rsi)
 256         pcmpeqb %xmm2, %xmm0
 257         pmovmskb %xmm0, %rax
 258         lea     16(%rsi), %rsi
 259 # ifdef USE_AS_STRNCPY
 260         sub     $16, %r8
 261         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 262 # endif
 263         test    %rax, %rax
 264         jnz     L(CopyFrom1To16Bytes)
 265
 266         movaps  16(%rcx, %rsi), %xmm3
 267         movaps  %xmm2, (%rdx, %rsi)
 268         pcmpeqb %xmm3, %xmm0
 269         pmovmskb %xmm0, %rax
 270         lea     16(%rsi), %rsi
 271 # ifdef USE_AS_STRNCPY
 272         sub     $16, %r8
 273         jbe     L(CopyFrom1To16BytesCase2OrCase3)
 274 # endif
 275         test    %rax, %rax
 276         jnz     L(CopyFrom1To16Bytes)
 277
 278         movaps  %xmm3, (%rdx, %rsi)
 279         mov     %rcx, %rax
 280         lea     16(%rcx, %rsi), %rcx
 281         and     $-0x40, %rcx
 282         sub     %rcx, %rax
 283         sub     %rax, %rdx
 284 # ifdef USE_AS_STRNCPY
 285         lea     48+64(%r8, %rax), %r8
 286 # endif
 287         mov     $-0x40, %rsi
 288
 289 L(Aligned64Loop):
 290         movaps  (%rcx), %xmm2
 291         movaps  %xmm2, %xmm4
 292         movaps  16(%rcx), %xmm5
 293         movaps  32(%rcx), %xmm3
 294         movaps  %xmm3, %xmm6
 295         movaps  48(%rcx), %xmm7
 296         pminub  %xmm5, %xmm2
 297         pminub  %xmm7, %xmm3
 298         pminub  %xmm2, %xmm3
 299         pcmpeqb %xmm0, %xmm3
 300         pmovmskb %xmm3, %rax
 301         lea     64(%rdx), %rdx
 302         lea     64(%rcx), %rcx
 303 # ifdef USE_AS_STRNCPY
 304         sub     $64, %r8
 305         jbe     L(StrncpyLeaveCase2OrCase3)
 306 # endif
 307         test    %rax, %rax
 308         jnz     L(Aligned64Leave)
 309         movaps  %xmm4, -64(%rdx)
 310         movaps  %xmm5, -48(%rdx)
 311         movaps  %xmm6, -32(%rdx)
 312         movaps  %xmm7, -16(%rdx)
 313         jmp     L(Aligned64Loop)
 314
 315 L(Aligned64Leave):
 316 # ifdef USE_AS_STRNCPY
 317         lea     48(%r8), %r8
 318 # endif
 319         pcmpeqb %xmm4, %xmm0
 320         pmovmskb %xmm0, %rax
 321         test    %rax, %rax
 322         jnz     L(CopyFrom1To16Bytes)
 323
 324         pcmpeqb %xmm5, %xmm0
 325 # ifdef USE_AS_STRNCPY
 326         lea     -16(%r8), %r8
 327 # endif
 328         pmovmskb %xmm0, %rax
 329         movaps  %xmm4, -64(%rdx)
 330         test    %rax, %rax
 331         lea     16(%rsi), %rsi
 332         jnz     L(CopyFrom1To16Bytes)
 333
 334         pcmpeqb %xmm6, %xmm0
 335 # ifdef USE_AS_STRNCPY
 336         lea     -16(%r8), %r8
 337 # endif
 338         pmovmskb %xmm0, %rax
 339         movaps  %xmm5, -48(%rdx)
 340         test    %rax, %rax
 341         lea     16(%rsi), %rsi
 342         jnz     L(CopyFrom1To16Bytes)
 343
 344         movaps  %xmm6, -32(%rdx)
 345         pcmpeqb %xmm7, %xmm0
 346 # ifdef USE_AS_STRNCPY
 347         lea     -16(%r8), %r8
 348 # endif
 349         pmovmskb %xmm0, %rax
 350         lea     16(%rsi), %rsi
 351         jmp     L(CopyFrom1To16Bytes)
 352
 353         .p2align 4
 354 L(Shl1):
 355         movaps  -1(%rcx), %xmm1
 356         movaps  15(%rcx), %xmm2
 357 L(Shl1Start):
 358         pcmpeqb %xmm2, %xmm0
 359         pmovmskb %xmm0, %rax
 360         movaps  %xmm2, %xmm3
 361 # ifdef USE_AS_STRNCPY
 362         sub     $16, %r8
 363         jbe     L(StrncpyExit1Case2OrCase3)
 364 # endif
 365         test    %rax, %rax
 366         jnz     L(Shl1LoopExit)
 367
 368         palignr $1, %xmm1, %xmm2
 369         movaps  %xmm3, %xmm1
 370         movaps  %xmm2, (%rdx)
 371         movaps  31(%rcx), %xmm2
 372
 373         pcmpeqb %xmm2, %xmm0
 374         lea     16(%rdx), %rdx
 375         pmovmskb %xmm0, %rax
 376         lea     16(%rcx), %rcx
 377         movaps  %xmm2, %xmm3
 378 # ifdef USE_AS_STRNCPY
 379         sub     $16, %r8
 380         jbe     L(StrncpyExit1Case2OrCase3)
 381 # endif
 382         test    %rax, %rax
 383         jnz     L(Shl1LoopExit)
 384
 385         palignr $1, %xmm1, %xmm2
 386         movaps  %xmm2, (%rdx)
 387         movaps  31(%rcx), %xmm2
 388         movaps  %xmm3, %xmm1
 389
 390         pcmpeqb %xmm2, %xmm0
 391         lea     16(%rdx), %rdx
 392         pmovmskb %xmm0, %rax
 393         lea     16(%rcx), %rcx
 394         movaps  %xmm2, %xmm3
 395 # ifdef USE_AS_STRNCPY
 396         sub     $16, %r8
 397         jbe     L(StrncpyExit1Case2OrCase3)
 398 # endif
 399         test    %rax, %rax
 400         jnz     L(Shl1LoopExit)
 401
 402         palignr $1, %xmm1, %xmm2
 403         movaps  %xmm3, %xmm1
 404         movaps  %xmm2, (%rdx)
 405         movaps  31(%rcx), %xmm2
 406
 407         pcmpeqb %xmm2, %xmm0
 408         lea     16(%rdx), %rdx
 409         pmovmskb %xmm0, %rax
 410         lea     16(%rcx), %rcx
 411         movaps  %xmm2, %xmm3
 412 # ifdef USE_AS_STRNCPY
 413         sub     $16, %r8
 414         jbe     L(StrncpyExit1Case2OrCase3)
 415 # endif
 416         test    %rax, %rax
 417         jnz     L(Shl1LoopExit)
 418
 419         palignr $1, %xmm1, %xmm2
 420         movaps  %xmm3, %xmm1
 421         movaps  %xmm2, (%rdx)
 422         lea     31(%rcx), %rcx
 423         lea     16(%rdx), %rdx
 424
 425         mov     %rcx, %rax
 426         and     $-0x40, %rcx
 427         sub     %rcx, %rax
 428         lea     -15(%rcx), %rcx
 429         sub     %rax, %rdx
 430 # ifdef USE_AS_STRNCPY
 431         add     %rax, %r8
 432 # endif
 433         movaps  -1(%rcx), %xmm1
 434
 435 L(Shl1LoopStart):
 436         movaps  15(%rcx), %xmm2
 437         movaps  31(%rcx), %xmm3
 438         movaps  %xmm3, %xmm6
 439         movaps  47(%rcx), %xmm4
 440         movaps  %xmm4, %xmm7
 441         movaps  63(%rcx), %xmm5
 442         pminub  %xmm2, %xmm6
 443         pminub  %xmm5, %xmm7
 444         pminub  %xmm6, %xmm7
 445         pcmpeqb %xmm0, %xmm7
 446         pmovmskb %xmm7, %rax
 447         movaps  %xmm5, %xmm7
 448         palignr $1, %xmm4, %xmm5
 449         test    %rax, %rax
 450         palignr $1, %xmm3, %xmm4
 451         jnz     L(Shl1Start)
 452 # ifdef USE_AS_STRNCPY
 453         sub     $64, %r8
 454         jbe     L(StrncpyLeave1)
 455 # endif
 456         palignr $1, %xmm2, %xmm3
 457         lea     64(%rcx), %rcx
 458         palignr $1, %xmm1, %xmm2
 459         movaps  %xmm7, %xmm1
 460         movaps  %xmm5, 48(%rdx)
 461         movaps  %xmm4, 32(%rdx)
 462         movaps  %xmm3, 16(%rdx)
 463         movaps  %xmm2, (%rdx)
 464         lea     64(%rdx), %rdx
 465         jmp     L(Shl1LoopStart)
 466
 467 L(Shl1LoopExit):
 468         movaps  (%rdx), %xmm6
 469         psrldq  $15, %xmm6
 470         mov     $15, %rsi
 471         palignr $1, %xmm1, %xmm6
 472         movaps  %xmm6, (%rdx)
 473         jmp     L(CopyFrom1To16Bytes)
 474
 475         .p2align 4
 476 L(Shl2):
 477         movaps  -2(%rcx), %xmm1
 478         movaps  14(%rcx), %xmm2
 479 L(Shl2Start):
 480         pcmpeqb %xmm2, %xmm0
 481         pmovmskb %xmm0, %rax
 482         movaps  %xmm2, %xmm3
 483 # ifdef USE_AS_STRNCPY
 484         sub     $16, %r8
 485         jbe     L(StrncpyExit2Case2OrCase3)
 486 # endif
 487         test    %rax, %rax
 488         jnz     L(Shl2LoopExit)
 489
 490         palignr $2, %xmm1, %xmm2
 491         movaps  %xmm3, %xmm1
 492         movaps  %xmm2, (%rdx)
 493         movaps  30(%rcx), %xmm2
 494
 495         pcmpeqb %xmm2, %xmm0
 496         lea     16(%rdx), %rdx
 497         pmovmskb %xmm0, %rax
 498         lea     16(%rcx), %rcx
 499         movaps  %xmm2, %xmm3
 500 # ifdef USE_AS_STRNCPY
 501         sub     $16, %r8
 502         jbe     L(StrncpyExit2Case2OrCase3)
 503 # endif
 504         test    %rax, %rax
 505         jnz     L(Shl2LoopExit)
 506
 507         palignr $2, %xmm1, %xmm2
 508         movaps  %xmm2, (%rdx)
 509         movaps  30(%rcx), %xmm2
 510         movaps  %xmm3, %xmm1
 511
 512         pcmpeqb %xmm2, %xmm0
 513         lea     16(%rdx), %rdx
 514         pmovmskb %xmm0, %rax
 515         lea     16(%rcx), %rcx
 516         movaps  %xmm2, %xmm3
 517 # ifdef USE_AS_STRNCPY
 518         sub     $16, %r8
 519         jbe     L(StrncpyExit2Case2OrCase3)
 520 # endif
 521         test    %rax, %rax
 522         jnz     L(Shl2LoopExit)
 523
 524         palignr $2, %xmm1, %xmm2
 525         movaps  %xmm3, %xmm1
 526         movaps  %xmm2, (%rdx)
 527         movaps  30(%rcx), %xmm2
 528
 529         pcmpeqb %xmm2, %xmm0
 530         lea     16(%rdx), %rdx
 531         pmovmskb %xmm0, %rax
 532         lea     16(%rcx), %rcx
 533         movaps  %xmm2, %xmm3
 534 # ifdef USE_AS_STRNCPY
 535         sub     $16, %r8
 536         jbe     L(StrncpyExit2Case2OrCase3)
 537 # endif
 538         test    %rax, %rax
 539         jnz     L(Shl2LoopExit)
 540
 541         palignr $2, %xmm1, %xmm2
 542         movaps  %xmm3, %xmm1
 543         movaps  %xmm2, (%rdx)
 544         lea     30(%rcx), %rcx
 545         lea     16(%rdx), %rdx
 546
 547         mov     %rcx, %rax
 548         and     $-0x40, %rcx
 549         sub     %rcx, %rax
 550         lea     -14(%rcx), %rcx
 551         sub     %rax, %rdx
 552 # ifdef USE_AS_STRNCPY
 553         add     %rax, %r8
 554 # endif
 555         movaps  -2(%rcx), %xmm1
 556
 557 L(Shl2LoopStart):
 558         movaps  14(%rcx), %xmm2
 559         movaps  30(%rcx), %xmm3
 560         movaps  %xmm3, %xmm6
 561         movaps  46(%rcx), %xmm4
 562         movaps  %xmm4, %xmm7
 563         movaps  62(%rcx), %xmm5
 564         pminub  %xmm2, %xmm6
 565         pminub  %xmm5, %xmm7
 566         pminub  %xmm6, %xmm7
 567         pcmpeqb %xmm0, %xmm7
 568         pmovmskb %xmm7, %rax
 569         movaps  %xmm5, %xmm7
 570         palignr $2, %xmm4, %xmm5
 571         test    %rax, %rax
 572         palignr $2, %xmm3, %xmm4
 573         jnz     L(Shl2Start)
 574 # ifdef USE_AS_STRNCPY
 575         sub     $64, %r8
 576         jbe     L(StrncpyLeave2)
 577 # endif
 578         palignr $2, %xmm2, %xmm3
 579         lea     64(%rcx), %rcx
 580         palignr $2, %xmm1, %xmm2
 581         movaps  %xmm7, %xmm1
 582         movaps  %xmm5, 48(%rdx)
 583         movaps  %xmm4, 32(%rdx)
 584         movaps  %xmm3, 16(%rdx)
 585         movaps  %xmm2, (%rdx)
 586         lea     64(%rdx), %rdx
 587         jmp     L(Shl2LoopStart)
 588
 589 L(Shl2LoopExit):
 590         movaps  (%rdx), %xmm6
 591         psrldq  $14, %xmm6
 592         mov     $14, %rsi
 593         palignr $2, %xmm1, %xmm6
 594         movaps  %xmm6, (%rdx)
 595         jmp     L(CopyFrom1To16Bytes)
 596
 597         .p2align 4
 598 L(Shl3):
 599         movaps  -3(%rcx), %xmm1
 600         movaps  13(%rcx), %xmm2
 601 L(Shl3Start):
 602         pcmpeqb %xmm2, %xmm0
 603         pmovmskb %xmm0, %rax
 604         movaps  %xmm2, %xmm3
 605 # ifdef USE_AS_STRNCPY
 606         sub     $16, %r8
 607         jbe     L(StrncpyExit3Case2OrCase3)
 608 # endif
 609         test    %rax, %rax
 610         jnz     L(Shl3LoopExit)
 611
 612         palignr $3, %xmm1, %xmm2
 613         movaps  %xmm3, %xmm1
 614         movaps  %xmm2, (%rdx)
 615         movaps  29(%rcx), %xmm2
 616
 617         pcmpeqb %xmm2, %xmm0
 618         lea     16(%rdx), %rdx
 619         pmovmskb %xmm0, %rax
 620         lea     16(%rcx), %rcx
 621         movaps  %xmm2, %xmm3
 622 # ifdef USE_AS_STRNCPY
 623         sub     $16, %r8
 624         jbe     L(StrncpyExit3Case2OrCase3)
 625 # endif
 626         test    %rax, %rax
 627         jnz     L(Shl3LoopExit)
 628
 629         palignr $3, %xmm1, %xmm2
 630         movaps  %xmm2, (%rdx)
 631         movaps  29(%rcx), %xmm2
 632         movaps  %xmm3, %xmm1
 633
 634         pcmpeqb %xmm2, %xmm0
 635         lea     16(%rdx), %rdx
 636         pmovmskb %xmm0, %rax
 637         lea     16(%rcx), %rcx
 638         movaps  %xmm2, %xmm3
 639 # ifdef USE_AS_STRNCPY
 640         sub     $16, %r8
 641         jbe     L(StrncpyExit3Case2OrCase3)
 642 # endif
 643         test    %rax, %rax
 644         jnz     L(Shl3LoopExit)
 645
 646         palignr $3, %xmm1, %xmm2
 647         movaps  %xmm3, %xmm1
 648         movaps  %xmm2, (%rdx)
 649         movaps  29(%rcx), %xmm2
 650
 651         pcmpeqb %xmm2, %xmm0
 652         lea     16(%rdx), %rdx
 653         pmovmskb %xmm0, %rax
 654         lea     16(%rcx), %rcx
 655         movaps  %xmm2, %xmm3
 656 # ifdef USE_AS_STRNCPY
 657         sub     $16, %r8
 658         jbe     L(StrncpyExit3Case2OrCase3)
 659 # endif
 660         test    %rax, %rax
 661         jnz     L(Shl3LoopExit)
 662
 663         palignr $3, %xmm1, %xmm2
 664         movaps  %xmm3, %xmm1
 665         movaps  %xmm2, (%rdx)
 666         lea     29(%rcx), %rcx
 667         lea     16(%rdx), %rdx
 668
 669         mov     %rcx, %rax
 670         and     $-0x40, %rcx
 671         sub     %rcx, %rax
 672         lea     -13(%rcx), %rcx
 673         sub     %rax, %rdx
 674 # ifdef USE_AS_STRNCPY
 675         add     %rax, %r8
 676 # endif
 677         movaps  -3(%rcx), %xmm1
 678
 679 L(Shl3LoopStart):
 680         movaps  13(%rcx), %xmm2
 681         movaps  29(%rcx), %xmm3
 682         movaps  %xmm3, %xmm6
 683         movaps  45(%rcx), %xmm4
 684         movaps  %xmm4, %xmm7
 685         movaps  61(%rcx), %xmm5
 686         pminub  %xmm2, %xmm6
 687         pminub  %xmm5, %xmm7
 688         pminub  %xmm6, %xmm7
 689         pcmpeqb %xmm0, %xmm7
 690         pmovmskb %xmm7, %rax
 691         movaps  %xmm5, %xmm7
 692         palignr $3, %xmm4, %xmm5
 693         test    %rax, %rax
 694         palignr $3, %xmm3, %xmm4
 695         jnz     L(Shl3Start)
 696 # ifdef USE_AS_STRNCPY
 697         sub     $64, %r8
 698         jbe     L(StrncpyLeave3)
 699 # endif
 700         palignr $3, %xmm2, %xmm3
 701         lea     64(%rcx), %rcx
 702         palignr $3, %xmm1, %xmm2
 703         movaps  %xmm7, %xmm1
 704         movaps  %xmm5, 48(%rdx)
 705         movaps  %xmm4, 32(%rdx)
 706         movaps  %xmm3, 16(%rdx)
 707         movaps  %xmm2, (%rdx)
 708         lea     64(%rdx), %rdx
 709         jmp     L(Shl3LoopStart)
 710
 711 L(Shl3LoopExit):
 712         movaps  (%rdx), %xmm6
 713         psrldq  $13, %xmm6
 714         mov     $13, %rsi
 715         palignr $3, %xmm1, %xmm6
 716         movaps  %xmm6, (%rdx)
 717         jmp     L(CopyFrom1To16Bytes)
 718
 719         .p2align 4
 720 L(Shl4):
 721         movaps  -4(%rcx), %xmm1
 722         movaps  12(%rcx), %xmm2
 723 L(Shl4Start):
 724         pcmpeqb %xmm2, %xmm0
 725         pmovmskb %xmm0, %rax
 726         movaps  %xmm2, %xmm3
 727 # ifdef USE_AS_STRNCPY
 728         sub     $16, %r8
 729         jbe     L(StrncpyExit4Case2OrCase3)
 730 # endif
 731         test    %rax, %rax
 732         jnz     L(Shl4LoopExit)
 733
 734         palignr $4, %xmm1, %xmm2
 735         movaps  %xmm3, %xmm1
 736         movaps  %xmm2, (%rdx)
 737         movaps  28(%rcx), %xmm2
 738
 739         pcmpeqb %xmm2, %xmm0
 740         lea     16(%rdx), %rdx
 741         pmovmskb %xmm0, %rax
 742         lea     16(%rcx), %rcx
 743         movaps  %xmm2, %xmm3
 744 # ifdef USE_AS_STRNCPY
 745         sub     $16, %r8
 746         jbe     L(StrncpyExit4Case2OrCase3)
 747 # endif
 748         test    %rax, %rax
 749         jnz     L(Shl4LoopExit)
 750
 751         palignr $4, %xmm1, %xmm2
 752         movaps  %xmm2, (%rdx)
 753         movaps  28(%rcx), %xmm2
 754         movaps  %xmm3, %xmm1
 755
 756         pcmpeqb %xmm2, %xmm0
 757         lea     16(%rdx), %rdx
 758         pmovmskb %xmm0, %rax
 759         lea     16(%rcx), %rcx
 760         movaps  %xmm2, %xmm3
 761 # ifdef USE_AS_STRNCPY
 762         sub     $16, %r8
 763         jbe     L(StrncpyExit4Case2OrCase3)
 764 # endif
 765         test    %rax, %rax
 766         jnz     L(Shl4LoopExit)
 767
 768         palignr $4, %xmm1, %xmm2
 769         movaps  %xmm3, %xmm1
 770         movaps  %xmm2, (%rdx)
 771         movaps  28(%rcx), %xmm2
 772
 773         pcmpeqb %xmm2, %xmm0
 774         lea     16(%rdx), %rdx
 775         pmovmskb %xmm0, %rax
 776         lea     16(%rcx), %rcx
 777         movaps  %xmm2, %xmm3
 778 # ifdef USE_AS_STRNCPY
 779         sub     $16, %r8
 780         jbe     L(StrncpyExit4Case2OrCase3)
 781 # endif
 782         test    %rax, %rax
 783         jnz     L(Shl4LoopExit)
 784
 785         palignr $4, %xmm1, %xmm2
 786         movaps  %xmm3, %xmm1
 787         movaps  %xmm2, (%rdx)
 788         lea     28(%rcx), %rcx
 789         lea     16(%rdx), %rdx
 790
 791         mov     %rcx, %rax
 792         and     $-0x40, %rcx
 793         sub     %rcx, %rax
 794         lea     -12(%rcx), %rcx
 795         sub     %rax, %rdx
 796 # ifdef USE_AS_STRNCPY
 797         add     %rax, %r8
 798 # endif
 799         movaps  -4(%rcx), %xmm1
 800
 801 L(Shl4LoopStart):
 802         movaps  12(%rcx), %xmm2
 803         movaps  28(%rcx), %xmm3
 804         movaps  %xmm3, %xmm6
 805         movaps  44(%rcx), %xmm4
 806         movaps  %xmm4, %xmm7
 807         movaps  60(%rcx), %xmm5
 808         pminub  %xmm2, %xmm6
 809         pminub  %xmm5, %xmm7
 810         pminub  %xmm6, %xmm7
 811         pcmpeqb %xmm0, %xmm7
 812         pmovmskb %xmm7, %rax
 813         movaps  %xmm5, %xmm7
 814         palignr $4, %xmm4, %xmm5
 815         test    %rax, %rax
 816         palignr $4, %xmm3, %xmm4
 817         jnz     L(Shl4Start)
 818 # ifdef USE_AS_STRNCPY
 819         sub     $64, %r8
 820         jbe     L(StrncpyLeave4)
 821 # endif
 822         palignr $4, %xmm2, %xmm3
 823         lea     64(%rcx), %rcx
 824         palignr $4, %xmm1, %xmm2
 825         movaps  %xmm7, %xmm1
 826         movaps  %xmm5, 48(%rdx)
 827         movaps  %xmm4, 32(%rdx)
 828         movaps  %xmm3, 16(%rdx)
 829         movaps  %xmm2, (%rdx)
 830         lea     64(%rdx), %rdx
 831         jmp     L(Shl4LoopStart)
 832
 833 L(Shl4LoopExit):
 834         movaps  (%rdx), %xmm6
 835         psrldq  $12, %xmm6
 836         mov     $12, %rsi
 837         palignr $4, %xmm1, %xmm6
 838         movaps  %xmm6, (%rdx)
 839         jmp     L(CopyFrom1To16Bytes)
 840
 841         .p2align 4
 842 L(Shl5):
 843         movaps  -5(%rcx), %xmm1
 844         movaps  11(%rcx), %xmm2
 845 L(Shl5Start):
 846         pcmpeqb %xmm2, %xmm0
 847         pmovmskb %xmm0, %rax
 848         movaps  %xmm2, %xmm3
 849 # ifdef USE_AS_STRNCPY
 850         sub     $16, %r8
 851         jbe     L(StrncpyExit5Case2OrCase3)
 852 # endif
 853         test    %rax, %rax
 854         jnz     L(Shl5LoopExit)
 855
 856         palignr $5, %xmm1, %xmm2
 857         movaps  %xmm3, %xmm1
 858         movaps  %xmm2, (%rdx)
 859         movaps  27(%rcx), %xmm2
 860
 861         pcmpeqb %xmm2, %xmm0
 862         lea     16(%rdx), %rdx
 863         pmovmskb %xmm0, %rax
 864         lea     16(%rcx), %rcx
 865         movaps  %xmm2, %xmm3
 866 # ifdef USE_AS_STRNCPY
 867         sub     $16, %r8
 868         jbe     L(StrncpyExit5Case2OrCase3)
 869 # endif
 870         test    %rax, %rax
 871         jnz     L(Shl5LoopExit)
 872
 873         palignr $5, %xmm1, %xmm2
 874         movaps  %xmm2, (%rdx)
 875         movaps  27(%rcx), %xmm2
 876         movaps  %xmm3, %xmm1
 877
 878         pcmpeqb %xmm2, %xmm0
 879         lea     16(%rdx), %rdx
 880         pmovmskb %xmm0, %rax
 881         lea     16(%rcx), %rcx
 882         movaps  %xmm2, %xmm3
 883 # ifdef USE_AS_STRNCPY
 884         sub     $16, %r8
 885         jbe     L(StrncpyExit5Case2OrCase3)
 886 # endif
 887         test    %rax, %rax
 888         jnz     L(Shl5LoopExit)
 889
 890         palignr $5, %xmm1, %xmm2
 891         movaps  %xmm3, %xmm1
 892         movaps  %xmm2, (%rdx)
 893         movaps  27(%rcx), %xmm2
 894
 895         pcmpeqb %xmm2, %xmm0
 896         lea     16(%rdx), %rdx
 897         pmovmskb %xmm0, %rax
 898         lea     16(%rcx), %rcx
 899         movaps  %xmm2, %xmm3
 900 # ifdef USE_AS_STRNCPY
 901         sub     $16, %r8
 902         jbe     L(StrncpyExit5Case2OrCase3)
 903 # endif
 904         test    %rax, %rax
 905         jnz     L(Shl5LoopExit)
 906
 907         palignr $5, %xmm1, %xmm2
 908         movaps  %xmm3, %xmm1
 909         movaps  %xmm2, (%rdx)
 910         lea     27(%rcx), %rcx
 911         lea     16(%rdx), %rdx
 912
 913         mov     %rcx, %rax
 914         and     $-0x40, %rcx
 915         sub     %rcx, %rax
 916         lea     -11(%rcx), %rcx
 917         sub     %rax, %rdx
 918 # ifdef USE_AS_STRNCPY
 919         add     %rax, %r8
 920 # endif
 921         movaps  -5(%rcx), %xmm1
 922
 923 L(Shl5LoopStart):
 924         movaps  11(%rcx), %xmm2
 925         movaps  27(%rcx), %xmm3
 926         movaps  %xmm3, %xmm6
 927         movaps  43(%rcx), %xmm4
 928         movaps  %xmm4, %xmm7
 929         movaps  59(%rcx), %xmm5
 930         pminub  %xmm2, %xmm6
 931         pminub  %xmm5, %xmm7
 932         pminub  %xmm6, %xmm7
 933         pcmpeqb %xmm0, %xmm7
 934         pmovmskb %xmm7, %rax
 935         movaps  %xmm5, %xmm7
 936         palignr $5, %xmm4, %xmm5
 937         test    %rax, %rax
 938         palignr $5, %xmm3, %xmm4
 939         jnz     L(Shl5Start)
 940 # ifdef USE_AS_STRNCPY
 941         sub     $64, %r8
 942         jbe     L(StrncpyLeave5)
 943 # endif
 944         palignr $5, %xmm2, %xmm3
 945         lea     64(%rcx), %rcx
 946         palignr $5, %xmm1, %xmm2
 947         movaps  %xmm7, %xmm1
 948         movaps  %xmm5, 48(%rdx)
 949         movaps  %xmm4, 32(%rdx)
 950         movaps  %xmm3, 16(%rdx)
 951         movaps  %xmm2, (%rdx)
 952         lea     64(%rdx), %rdx
 953         jmp     L(Shl5LoopStart)
 954
 955 L(Shl5LoopExit):
 956         movaps  (%rdx), %xmm6
 957         psrldq  $11, %xmm6
 958         mov     $11, %rsi
 959         palignr $5, %xmm1, %xmm6
 960         movaps  %xmm6, (%rdx)
 961         jmp     L(CopyFrom1To16Bytes)
 962
 963         .p2align 4
 964 L(Shl6):
 965         movaps  -6(%rcx), %xmm1
 966         movaps  10(%rcx), %xmm2
 967 L(Shl6Start):
 968         pcmpeqb %xmm2, %xmm0
 969         pmovmskb %xmm0, %rax
 970         movaps  %xmm2, %xmm3
 971 # ifdef USE_AS_STRNCPY
 972         sub     $16, %r8
 973         jbe     L(StrncpyExit6Case2OrCase3)
 974 # endif
 975         test    %rax, %rax
 976         jnz     L(Shl6LoopExit)
 977
 978         palignr $6, %xmm1, %xmm2
 979         movaps  %xmm3, %xmm1
 980         movaps  %xmm2, (%rdx)
 981         movaps  26(%rcx), %xmm2
 982
 983         pcmpeqb %xmm2, %xmm0
 984         lea     16(%rdx), %rdx
 985         pmovmskb %xmm0, %rax
 986         lea     16(%rcx), %rcx
 987         movaps  %xmm2, %xmm3
 988 # ifdef USE_AS_STRNCPY
 989         sub     $16, %r8
 990         jbe     L(StrncpyExit6Case2OrCase3)
 991 # endif
 992         test    %rax, %rax
 993         jnz     L(Shl6LoopExit)
 994
 995         palignr $6, %xmm1, %xmm2
 996         movaps  %xmm2, (%rdx)
 997         movaps  26(%rcx), %xmm2
 998         movaps  %xmm3, %xmm1
 999
1000         pcmpeqb %xmm2, %xmm0
1001         lea     16(%rdx), %rdx
1002         pmovmskb %xmm0, %rax
1003         lea     16(%rcx), %rcx
1004         movaps  %xmm2, %xmm3
1005 # ifdef USE_AS_STRNCPY
1006         sub     $16, %r8
1007         jbe     L(StrncpyExit6Case2OrCase3)
1008 # endif
1009         test    %rax, %rax
1010         jnz     L(Shl6LoopExit)
1011
1012         palignr $6, %xmm1, %xmm2
1013         movaps  %xmm3, %xmm1
1014         movaps  %xmm2, (%rdx)
1015         movaps  26(%rcx), %xmm2
1016
1017         pcmpeqb %xmm2, %xmm0
1018         lea     16(%rdx), %rdx
1019         pmovmskb %xmm0, %rax
1020         lea     16(%rcx), %rcx
1021         movaps  %xmm2, %xmm3
1022 # ifdef USE_AS_STRNCPY
1023         sub     $16, %r8
1024         jbe     L(StrncpyExit6Case2OrCase3)
1025 # endif
1026         test    %rax, %rax
1027         jnz     L(Shl6LoopExit)
1028
1029         palignr $6, %xmm1, %xmm2
1030         movaps  %xmm3, %xmm1
1031         movaps  %xmm2, (%rdx)
1032         lea     26(%rcx), %rcx
1033         lea     16(%rdx), %rdx
1034
1035         mov     %rcx, %rax
1036         and     $-0x40, %rcx
1037         sub     %rcx, %rax
1038         lea     -10(%rcx), %rcx
1039         sub     %rax, %rdx
1040 # ifdef USE_AS_STRNCPY
1041         add     %rax, %r8
1042 # endif
1043         movaps  -6(%rcx), %xmm1
1044
1045 L(Shl6LoopStart):
1046         movaps  10(%rcx), %xmm2
1047         movaps  26(%rcx), %xmm3
1048         movaps  %xmm3, %xmm6
1049         movaps  42(%rcx), %xmm4
1050         movaps  %xmm4, %xmm7
1051         movaps  58(%rcx), %xmm5
1052         pminub  %xmm2, %xmm6
1053         pminub  %xmm5, %xmm7
1054         pminub  %xmm6, %xmm7
1055         pcmpeqb %xmm0, %xmm7
1056         pmovmskb %xmm7, %rax
1057         movaps  %xmm5, %xmm7
1058         palignr $6, %xmm4, %xmm5
1059         test    %rax, %rax
1060         palignr $6, %xmm3, %xmm4
1061         jnz     L(Shl6Start)
1062 # ifdef USE_AS_STRNCPY
1063         sub     $64, %r8
1064         jbe     L(StrncpyLeave6)
1065 # endif
1066         palignr $6, %xmm2, %xmm3
1067         lea     64(%rcx), %rcx
1068         palignr $6, %xmm1, %xmm2
1069         movaps  %xmm7, %xmm1
1070         movaps  %xmm5, 48(%rdx)
1071         movaps  %xmm4, 32(%rdx)
1072         movaps  %xmm3, 16(%rdx)
1073         movaps  %xmm2, (%rdx)
1074         lea     64(%rdx), %rdx
1075         jmp     L(Shl6LoopStart)
1076
1077 L(Shl6LoopExit):
1078         movaps  (%rdx), %xmm6
1079         psrldq  $10, %xmm6
1080         mov     $10, %rsi
1081         palignr $6, %xmm1, %xmm6
1082         movaps  %xmm6, (%rdx)
1083         jmp     L(CopyFrom1To16Bytes)
1084
1085         .p2align 4
1086 L(Shl7):
1087         movaps  -7(%rcx), %xmm1
1088         movaps  9(%rcx), %xmm2
1089 L(Shl7Start):
1090         pcmpeqb %xmm2, %xmm0
1091         pmovmskb %xmm0, %rax
1092         movaps  %xmm2, %xmm3
1093 # ifdef USE_AS_STRNCPY
1094         sub     $16, %r8
1095         jbe     L(StrncpyExit7Case2OrCase3)
1096 # endif
1097         test    %rax, %rax
1098         jnz     L(Shl7LoopExit)
1099
1100         palignr $7, %xmm1, %xmm2
1101         movaps  %xmm3, %xmm1
1102         movaps  %xmm2, (%rdx)
1103         movaps  25(%rcx), %xmm2
1104
1105         pcmpeqb %xmm2, %xmm0
1106         lea     16(%rdx), %rdx
1107         pmovmskb %xmm0, %rax
1108         lea     16(%rcx), %rcx
1109         movaps  %xmm2, %xmm3
1110 # ifdef USE_AS_STRNCPY
1111         sub     $16, %r8
1112         jbe     L(StrncpyExit7Case2OrCase3)
1113 # endif
1114         test    %rax, %rax
1115         jnz     L(Shl7LoopExit)
1116
1117         palignr $7, %xmm1, %xmm2
1118         movaps  %xmm2, (%rdx)
1119         movaps  25(%rcx), %xmm2
1120         movaps  %xmm3, %xmm1
1121
1122         pcmpeqb %xmm2, %xmm0
1123         lea     16(%rdx), %rdx
1124         pmovmskb %xmm0, %rax
1125         lea     16(%rcx), %rcx
1126         movaps  %xmm2, %xmm3
1127 # ifdef USE_AS_STRNCPY
1128         sub     $16, %r8
1129         jbe     L(StrncpyExit7Case2OrCase3)
1130 # endif
1131         test    %rax, %rax
1132         jnz     L(Shl7LoopExit)
1133
1134         palignr $7, %xmm1, %xmm2
1135         movaps  %xmm3, %xmm1
1136         movaps  %xmm2, (%rdx)
1137         movaps  25(%rcx), %xmm2
1138
1139         pcmpeqb %xmm2, %xmm0
1140         lea     16(%rdx), %rdx
1141         pmovmskb %xmm0, %rax
1142         lea     16(%rcx), %rcx
1143         movaps  %xmm2, %xmm3
1144 # ifdef USE_AS_STRNCPY
1145         sub     $16, %r8
1146         jbe     L(StrncpyExit7Case2OrCase3)
1147 # endif
1148         test    %rax, %rax
1149         jnz     L(Shl7LoopExit)
1150
1151         palignr $7, %xmm1, %xmm2
1152         movaps  %xmm3, %xmm1
1153         movaps  %xmm2, (%rdx)
1154         lea     25(%rcx), %rcx
1155         lea     16(%rdx), %rdx
1156
1157         mov     %rcx, %rax
1158         and     $-0x40, %rcx
1159         sub     %rcx, %rax
1160         lea     -9(%rcx), %rcx
1161         sub     %rax, %rdx
1162 # ifdef USE_AS_STRNCPY
1163         add     %rax, %r8
1164 # endif
1165         movaps  -7(%rcx), %xmm1
1166
1167 L(Shl7LoopStart):
1168         movaps  9(%rcx), %xmm2
1169         movaps  25(%rcx), %xmm3
1170         movaps  %xmm3, %xmm6
1171         movaps  41(%rcx), %xmm4
1172         movaps  %xmm4, %xmm7
1173         movaps  57(%rcx), %xmm5
1174         pminub  %xmm2, %xmm6
1175         pminub  %xmm5, %xmm7
1176         pminub  %xmm6, %xmm7
1177         pcmpeqb %xmm0, %xmm7
1178         pmovmskb %xmm7, %rax
1179         movaps  %xmm5, %xmm7
1180         palignr $7, %xmm4, %xmm5
1181         test    %rax, %rax
1182         palignr $7, %xmm3, %xmm4
1183         jnz     L(Shl7Start)
1184 # ifdef USE_AS_STRNCPY
1185         sub     $64, %r8
1186         jbe     L(StrncpyLeave7)
1187 # endif
1188         palignr $7, %xmm2, %xmm3
1189         lea     64(%rcx), %rcx
1190         palignr $7, %xmm1, %xmm2
1191         movaps  %xmm7, %xmm1
1192         movaps  %xmm5, 48(%rdx)
1193         movaps  %xmm4, 32(%rdx)
1194         movaps  %xmm3, 16(%rdx)
1195         movaps  %xmm2, (%rdx)
1196         lea     64(%rdx), %rdx
1197         jmp     L(Shl7LoopStart)
1198
1199 L(Shl7LoopExit):
1200         movaps  (%rdx), %xmm6
1201         psrldq  $9, %xmm6
1202         mov     $9, %rsi
1203         palignr $7, %xmm1, %xmm6
1204         movaps  %xmm6, (%rdx)
1205         jmp     L(CopyFrom1To16Bytes)
1206
1207         .p2align 4
1208 L(Shl8):
1209         movaps  -8(%rcx), %xmm1
1210         movaps  8(%rcx), %xmm2
1211 L(Shl8Start):
1212         pcmpeqb %xmm2, %xmm0
1213         pmovmskb %xmm0, %rax
1214         movaps  %xmm2, %xmm3
1215 # ifdef USE_AS_STRNCPY
1216         sub     $16, %r8
1217         jbe     L(StrncpyExit8Case2OrCase3)
1218 # endif
1219         test    %rax, %rax
1220         jnz     L(Shl8LoopExit)
1221
1222         palignr $8, %xmm1, %xmm2
1223         movaps  %xmm3, %xmm1
1224         movaps  %xmm2, (%rdx)
1225         movaps  24(%rcx), %xmm2
1226
1227         pcmpeqb %xmm2, %xmm0
1228         lea     16(%rdx), %rdx
1229         pmovmskb %xmm0, %rax
1230         lea     16(%rcx), %rcx
1231         movaps  %xmm2, %xmm3
1232 # ifdef USE_AS_STRNCPY
1233         sub     $16, %r8
1234         jbe     L(StrncpyExit8Case2OrCase3)
1235 # endif
1236         test    %rax, %rax
1237         jnz     L(Shl8LoopExit)
1238
1239         palignr $8, %xmm1, %xmm2
1240         movaps  %xmm2, (%rdx)
1241         movaps  24(%rcx), %xmm2
1242         movaps  %xmm3, %xmm1
1243
1244         pcmpeqb %xmm2, %xmm0
1245         lea     16(%rdx), %rdx
1246         pmovmskb %xmm0, %rax
1247         lea     16(%rcx), %rcx
1248         movaps  %xmm2, %xmm3
1249 # ifdef USE_AS_STRNCPY
1250         sub     $16, %r8
1251         jbe     L(StrncpyExit8Case2OrCase3)
1252 # endif
1253         test    %rax, %rax
1254         jnz     L(Shl8LoopExit)
1255
1256         palignr $8, %xmm1, %xmm2
1257         movaps  %xmm3, %xmm1
1258         movaps  %xmm2, (%rdx)
1259         movaps  24(%rcx), %xmm2
1260
1261         pcmpeqb %xmm2, %xmm0
1262         lea     16(%rdx), %rdx
1263         pmovmskb %xmm0, %rax
1264         lea     16(%rcx), %rcx
1265         movaps  %xmm2, %xmm3
1266 # ifdef USE_AS_STRNCPY
1267         sub     $16, %r8
1268         jbe     L(StrncpyExit8Case2OrCase3)
1269 # endif
1270         test    %rax, %rax
1271         jnz     L(Shl8LoopExit)
1272
1273         palignr $8, %xmm1, %xmm2
1274         movaps  %xmm3, %xmm1
1275         movaps  %xmm2, (%rdx)
1276         lea     24(%rcx), %rcx
1277         lea     16(%rdx), %rdx
1278
1279         mov     %rcx, %rax
1280         and     $-0x40, %rcx
1281         sub     %rcx, %rax
1282         lea     -8(%rcx), %rcx
1283         sub     %rax, %rdx
1284 # ifdef USE_AS_STRNCPY
1285         add     %rax, %r8
1286 # endif
1287         movaps  -8(%rcx), %xmm1
1288
1289 L(Shl8LoopStart):
1290         movaps  8(%rcx), %xmm2
1291         movaps  24(%rcx), %xmm3
1292         movaps  %xmm3, %xmm6
1293         movaps  40(%rcx), %xmm4
1294         movaps  %xmm4, %xmm7
1295         movaps  56(%rcx), %xmm5
1296         pminub  %xmm2, %xmm6
1297         pminub  %xmm5, %xmm7
1298         pminub  %xmm6, %xmm7
1299         pcmpeqb %xmm0, %xmm7
1300         pmovmskb %xmm7, %rax
1301         movaps  %xmm5, %xmm7
1302         palignr $8, %xmm4, %xmm5
1303         test    %rax, %rax
1304         palignr $8, %xmm3, %xmm4
1305         jnz     L(Shl8Start)
1306 # ifdef USE_AS_STRNCPY
1307         sub     $64, %r8
1308         jbe     L(StrncpyLeave8)
1309 # endif
1310         palignr $8, %xmm2, %xmm3
1311         lea     64(%rcx), %rcx
1312         palignr $8, %xmm1, %xmm2
1313         movaps  %xmm7, %xmm1
1314         movaps  %xmm5, 48(%rdx)
1315         movaps  %xmm4, 32(%rdx)
1316         movaps  %xmm3, 16(%rdx)
1317         movaps  %xmm2, (%rdx)
1318         lea     64(%rdx), %rdx
1319         jmp     L(Shl8LoopStart)
1320
1321 L(Shl8LoopExit):
1322         movaps  (%rdx), %xmm6
1323         psrldq  $8, %xmm6
1324         mov     $8, %rsi
1325         palignr $8, %xmm1, %xmm6
1326         movaps  %xmm6, (%rdx)
1327         jmp     L(CopyFrom1To16Bytes)
1328
1329         .p2align 4
1330 L(Shl9):
1331         movaps  -9(%rcx), %xmm1
1332         movaps  7(%rcx), %xmm2
1333 L(Shl9Start):
1334         pcmpeqb %xmm2, %xmm0
1335         pmovmskb %xmm0, %rax
1336         movaps  %xmm2, %xmm3
1337 # ifdef USE_AS_STRNCPY
1338         sub     $16, %r8
1339         jbe     L(StrncpyExit9Case2OrCase3)
1340 # endif
1341         test    %rax, %rax
1342         jnz     L(Shl9LoopExit)
1343
1344         palignr $9, %xmm1, %xmm2
1345         movaps  %xmm3, %xmm1
1346         movaps  %xmm2, (%rdx)
1347         movaps  23(%rcx), %xmm2
1348
1349         pcmpeqb %xmm2, %xmm0
1350         lea     16(%rdx), %rdx
1351         pmovmskb %xmm0, %rax
1352         lea     16(%rcx), %rcx
1353         movaps  %xmm2, %xmm3
1354 # ifdef USE_AS_STRNCPY
1355         sub     $16, %r8
1356         jbe     L(StrncpyExit9Case2OrCase3)
1357 # endif
1358         test    %rax, %rax
1359         jnz     L(Shl9LoopExit)
1360
1361         palignr $9, %xmm1, %xmm2
1362         movaps  %xmm2, (%rdx)
1363         movaps  23(%rcx), %xmm2
1364         movaps  %xmm3, %xmm1
1365
1366         pcmpeqb %xmm2, %xmm0
1367         lea     16(%rdx), %rdx
1368         pmovmskb %xmm0, %rax
1369         lea     16(%rcx), %rcx
1370         movaps  %xmm2, %xmm3
1371 # ifdef USE_AS_STRNCPY
1372         sub     $16, %r8
1373         jbe     L(StrncpyExit9Case2OrCase3)
1374 # endif
1375         test    %rax, %rax
1376         jnz     L(Shl9LoopExit)
1377
1378         palignr $9, %xmm1, %xmm2
1379         movaps  %xmm3, %xmm1
1380         movaps  %xmm2, (%rdx)
1381         movaps  23(%rcx), %xmm2
1382
1383         pcmpeqb %xmm2, %xmm0
1384         lea     16(%rdx), %rdx
1385         pmovmskb %xmm0, %rax
1386         lea     16(%rcx), %rcx
1387         movaps  %xmm2, %xmm3
1388 # ifdef USE_AS_STRNCPY
1389         sub     $16, %r8
1390         jbe     L(StrncpyExit9Case2OrCase3)
1391 # endif
1392         test    %rax, %rax
1393         jnz     L(Shl9LoopExit)
1394
1395         palignr $9, %xmm1, %xmm2
1396         movaps  %xmm3, %xmm1
1397         movaps  %xmm2, (%rdx)
1398         lea     23(%rcx), %rcx
1399         lea     16(%rdx), %rdx
1400
1401         mov     %rcx, %rax
1402         and     $-0x40, %rcx
1403         sub     %rcx, %rax
1404         lea     -7(%rcx), %rcx
1405         sub     %rax, %rdx
1406 # ifdef USE_AS_STRNCPY
1407         add     %rax, %r8
1408 # endif
1409         movaps  -9(%rcx), %xmm1
1410
1411 L(Shl9LoopStart):
1412         movaps  7(%rcx), %xmm2
1413         movaps  23(%rcx), %xmm3
1414         movaps  %xmm3, %xmm6
1415         movaps  39(%rcx), %xmm4
1416         movaps  %xmm4, %xmm7
1417         movaps  55(%rcx), %xmm5
1418         pminub  %xmm2, %xmm6
1419         pminub  %xmm5, %xmm7
1420         pminub  %xmm6, %xmm7
1421         pcmpeqb %xmm0, %xmm7
1422         pmovmskb %xmm7, %rax
1423         movaps  %xmm5, %xmm7
1424         palignr $9, %xmm4, %xmm5
1425         test    %rax, %rax
1426         palignr $9, %xmm3, %xmm4
1427         jnz     L(Shl9Start)
1428 # ifdef USE_AS_STRNCPY
1429         sub     $64, %r8
1430         jbe     L(StrncpyLeave9)
1431 # endif
1432         palignr $9, %xmm2, %xmm3
1433         lea     64(%rcx), %rcx
1434         palignr $9, %xmm1, %xmm2
1435         movaps  %xmm7, %xmm1
1436         movaps  %xmm5, 48(%rdx)
1437         movaps  %xmm4, 32(%rdx)
1438         movaps  %xmm3, 16(%rdx)
1439         movaps  %xmm2, (%rdx)
1440         lea     64(%rdx), %rdx
1441         jmp     L(Shl9LoopStart)
1442
1443 L(Shl9LoopExit):
1444         movaps  (%rdx), %xmm6
1445         psrldq  $7, %xmm6
1446         mov     $7, %rsi
1447         palignr $9, %xmm1, %xmm6
1448         movaps  %xmm6, (%rdx)
1449         jmp     L(CopyFrom1To16Bytes)
1450
1451         .p2align 4
1452 L(Shl10):
1453         movaps  -10(%rcx), %xmm1
1454         movaps  6(%rcx), %xmm2
1455 L(Shl10Start):
1456         pcmpeqb %xmm2, %xmm0
1457         pmovmskb %xmm0, %rax
1458         movaps  %xmm2, %xmm3
1459 # ifdef USE_AS_STRNCPY
1460         sub     $16, %r8
1461         jbe     L(StrncpyExit10Case2OrCase3)
1462 # endif
1463         test    %rax, %rax
1464         jnz     L(Shl10LoopExit)
1465
1466         palignr $10, %xmm1, %xmm2
1467         movaps  %xmm3, %xmm1
1468         movaps  %xmm2, (%rdx)
1469         movaps  22(%rcx), %xmm2
1470
1471         pcmpeqb %xmm2, %xmm0
1472         lea     16(%rdx), %rdx
1473         pmovmskb %xmm0, %rax
1474         lea     16(%rcx), %rcx
1475         movaps  %xmm2, %xmm3
1476 # ifdef USE_AS_STRNCPY
1477         sub     $16, %r8
1478         jbe     L(StrncpyExit10Case2OrCase3)
1479 # endif
1480         test    %rax, %rax
1481         jnz     L(Shl10LoopExit)
1482
1483         palignr $10, %xmm1, %xmm2
1484         movaps  %xmm2, (%rdx)
1485         movaps  22(%rcx), %xmm2
1486         movaps  %xmm3, %xmm1
1487
1488         pcmpeqb %xmm2, %xmm0
1489         lea     16(%rdx), %rdx
1490         pmovmskb %xmm0, %rax
1491         lea     16(%rcx), %rcx
1492         movaps  %xmm2, %xmm3
1493 # ifdef USE_AS_STRNCPY
1494         sub     $16, %r8
1495         jbe     L(StrncpyExit10Case2OrCase3)
1496 # endif
1497         test    %rax, %rax
1498         jnz     L(Shl10LoopExit)
1499
1500         palignr $10, %xmm1, %xmm2
1501         movaps  %xmm3, %xmm1
1502         movaps  %xmm2, (%rdx)
1503         movaps  22(%rcx), %xmm2
1504
1505         pcmpeqb %xmm2, %xmm0
1506         lea     16(%rdx), %rdx
1507         pmovmskb %xmm0, %rax
1508         lea     16(%rcx), %rcx
1509         movaps  %xmm2, %xmm3
1510 # ifdef USE_AS_STRNCPY
1511         sub     $16, %r8
1512         jbe     L(StrncpyExit10Case2OrCase3)
1513 # endif
1514         test    %rax, %rax
1515         jnz     L(Shl10LoopExit)
1516
1517         palignr $10, %xmm1, %xmm2
1518         movaps  %xmm3, %xmm1
1519         movaps  %xmm2, (%rdx)
1520         lea     22(%rcx), %rcx
1521         lea     16(%rdx), %rdx
1522
1523         mov     %rcx, %rax
1524         and     $-0x40, %rcx
1525         sub     %rcx, %rax
1526         lea     -6(%rcx), %rcx
1527         sub     %rax, %rdx
1528 # ifdef USE_AS_STRNCPY
1529         add     %rax, %r8
1530 # endif
1531         movaps  -10(%rcx), %xmm1
1532
1533 L(Shl10LoopStart):
1534         movaps  6(%rcx), %xmm2
1535         movaps  22(%rcx), %xmm3
1536         movaps  %xmm3, %xmm6
1537         movaps  38(%rcx), %xmm4
1538         movaps  %xmm4, %xmm7
1539         movaps  54(%rcx), %xmm5
1540         pminub  %xmm2, %xmm6
1541         pminub  %xmm5, %xmm7
1542         pminub  %xmm6, %xmm7
1543         pcmpeqb %xmm0, %xmm7
1544         pmovmskb %xmm7, %rax
1545         movaps  %xmm5, %xmm7
1546         palignr $10, %xmm4, %xmm5
1547         test    %rax, %rax
1548         palignr $10, %xmm3, %xmm4
1549         jnz     L(Shl10Start)
1550 # ifdef USE_AS_STRNCPY
1551         sub     $64, %r8
1552         jbe     L(StrncpyLeave10)
1553 # endif
1554         palignr $10, %xmm2, %xmm3
1555         lea     64(%rcx), %rcx
1556         palignr $10, %xmm1, %xmm2
1557         movaps  %xmm7, %xmm1
1558         movaps  %xmm5, 48(%rdx)
1559         movaps  %xmm4, 32(%rdx)
1560         movaps  %xmm3, 16(%rdx)
1561         movaps  %xmm2, (%rdx)
1562         lea     64(%rdx), %rdx
1563         jmp     L(Shl10LoopStart)
1564
1565 L(Shl10LoopExit):
1566         movaps  (%rdx), %xmm6
1567         psrldq  $6, %xmm6
1568         mov     $6, %rsi
1569         palignr $10, %xmm1, %xmm6
1570         movaps  %xmm6, (%rdx)
1571         jmp     L(CopyFrom1To16Bytes)
1572
1573         .p2align 4
1574 L(Shl11):
1575         movaps  -11(%rcx), %xmm1
1576         movaps  5(%rcx), %xmm2
1577 L(Shl11Start):
1578         pcmpeqb %xmm2, %xmm0
1579         pmovmskb %xmm0, %rax
1580         movaps  %xmm2, %xmm3
1581 # ifdef USE_AS_STRNCPY
1582         sub     $16, %r8
1583         jbe     L(StrncpyExit11Case2OrCase3)
1584 # endif
1585         test    %rax, %rax
1586         jnz     L(Shl11LoopExit)
1587
1588         palignr $11, %xmm1, %xmm2
1589         movaps  %xmm3, %xmm1
1590         movaps  %xmm2, (%rdx)
1591         movaps  21(%rcx), %xmm2
1592
1593         pcmpeqb %xmm2, %xmm0
1594         lea     16(%rdx), %rdx
1595         pmovmskb %xmm0, %rax
1596         lea     16(%rcx), %rcx
1597         movaps  %xmm2, %xmm3
1598 # ifdef USE_AS_STRNCPY
1599         sub     $16, %r8
1600         jbe     L(StrncpyExit11Case2OrCase3)
1601 # endif
1602         test    %rax, %rax
1603         jnz     L(Shl11LoopExit)
1604
1605         palignr $11, %xmm1, %xmm2
1606         movaps  %xmm2, (%rdx)
1607         movaps  21(%rcx), %xmm2
1608         movaps  %xmm3, %xmm1
1609
1610         pcmpeqb %xmm2, %xmm0
1611         lea     16(%rdx), %rdx
1612         pmovmskb %xmm0, %rax
1613         lea     16(%rcx), %rcx
1614         movaps  %xmm2, %xmm3
1615 # ifdef USE_AS_STRNCPY
1616         sub     $16, %r8
1617         jbe     L(StrncpyExit11Case2OrCase3)
1618 # endif
1619         test    %rax, %rax
1620         jnz     L(Shl11LoopExit)
1621
1622         palignr $11, %xmm1, %xmm2
1623         movaps  %xmm3, %xmm1
1624         movaps  %xmm2, (%rdx)
1625         movaps  21(%rcx), %xmm2
1626
1627         pcmpeqb %xmm2, %xmm0
1628         lea     16(%rdx), %rdx
1629         pmovmskb %xmm0, %rax
1630         lea     16(%rcx), %rcx
1631         movaps  %xmm2, %xmm3
1632 # ifdef USE_AS_STRNCPY
1633         sub     $16, %r8
1634         jbe     L(StrncpyExit11Case2OrCase3)
1635 # endif
1636         test    %rax, %rax
1637         jnz     L(Shl11LoopExit)
1638
1639         palignr $11, %xmm1, %xmm2
1640         movaps  %xmm3, %xmm1
1641         movaps  %xmm2, (%rdx)
1642         lea     21(%rcx), %rcx
1643         lea     16(%rdx), %rdx
1644
1645         mov     %rcx, %rax
1646         and     $-0x40, %rcx
1647         sub     %rcx, %rax
1648         lea     -5(%rcx), %rcx
1649         sub     %rax, %rdx
1650 # ifdef USE_AS_STRNCPY
1651         add     %rax, %r8
1652 # endif
1653         movaps  -11(%rcx), %xmm1
1654
1655 L(Shl11LoopStart):
1656         movaps  5(%rcx), %xmm2
1657         movaps  21(%rcx), %xmm3
1658         movaps  %xmm3, %xmm6
1659         movaps  37(%rcx), %xmm4
1660         movaps  %xmm4, %xmm7
1661         movaps  53(%rcx), %xmm5
1662         pminub  %xmm2, %xmm6
1663         pminub  %xmm5, %xmm7
1664         pminub  %xmm6, %xmm7
1665         pcmpeqb %xmm0, %xmm7
1666         pmovmskb %xmm7, %rax
1667         movaps  %xmm5, %xmm7
1668         palignr $11, %xmm4, %xmm5
1669         test    %rax, %rax
1670         palignr $11, %xmm3, %xmm4
1671         jnz     L(Shl11Start)
1672 # ifdef USE_AS_STRNCPY
1673         sub     $64, %r8
1674         jbe     L(StrncpyLeave11)
1675 # endif
1676         palignr $11, %xmm2, %xmm3
1677         lea     64(%rcx), %rcx
1678         palignr $11, %xmm1, %xmm2
1679         movaps  %xmm7, %xmm1
1680         movaps  %xmm5, 48(%rdx)
1681         movaps  %xmm4, 32(%rdx)
1682         movaps  %xmm3, 16(%rdx)
1683         movaps  %xmm2, (%rdx)
1684         lea     64(%rdx), %rdx
1685         jmp     L(Shl11LoopStart)
1686
1687 L(Shl11LoopExit):
1688         movaps  (%rdx), %xmm6
1689         psrldq  $5, %xmm6
1690         mov     $5, %rsi
1691         palignr $11, %xmm1, %xmm6
1692         movaps  %xmm6, (%rdx)
1693         jmp     L(CopyFrom1To16Bytes)
1694
1695         .p2align 4
1696 L(Shl12):
1697         movaps  -12(%rcx), %xmm1
1698         movaps  4(%rcx), %xmm2
1699 L(Shl12Start):
1700         pcmpeqb %xmm2, %xmm0
1701         pmovmskb %xmm0, %rax
1702         movaps  %xmm2, %xmm3
1703 # ifdef USE_AS_STRNCPY
1704         sub     $16, %r8
1705         jbe     L(StrncpyExit12Case2OrCase3)
1706 # endif
1707         test    %rax, %rax
1708         jnz     L(Shl12LoopExit)
1709
1710         palignr $12, %xmm1, %xmm2
1711         movaps  %xmm3, %xmm1
1712         movaps  %xmm2, (%rdx)
1713         movaps  20(%rcx), %xmm2
1714
1715         pcmpeqb %xmm2, %xmm0
1716         lea     16(%rdx), %rdx
1717         pmovmskb %xmm0, %rax
1718         lea     16(%rcx), %rcx
1719         movaps  %xmm2, %xmm3
1720 # ifdef USE_AS_STRNCPY
1721         sub     $16, %r8
1722         jbe     L(StrncpyExit12Case2OrCase3)
1723 # endif
1724         test    %rax, %rax
1725         jnz     L(Shl12LoopExit)
1726
1727         palignr $12, %xmm1, %xmm2
1728         movaps  %xmm2, (%rdx)
1729         movaps  20(%rcx), %xmm2
1730         movaps  %xmm3, %xmm1
1731
1732         pcmpeqb %xmm2, %xmm0
1733         lea     16(%rdx), %rdx
1734         pmovmskb %xmm0, %rax
1735         lea     16(%rcx), %rcx
1736         movaps  %xmm2, %xmm3
1737 # ifdef USE_AS_STRNCPY
1738         sub     $16, %r8
1739         jbe     L(StrncpyExit12Case2OrCase3)
1740 # endif
1741         test    %rax, %rax
1742         jnz     L(Shl12LoopExit)
1743
1744         palignr $12, %xmm1, %xmm2
1745         movaps  %xmm3, %xmm1
1746         movaps  %xmm2, (%rdx)
1747         movaps  20(%rcx), %xmm2
1748
1749         pcmpeqb %xmm2, %xmm0
1750         lea     16(%rdx), %rdx
1751         pmovmskb %xmm0, %rax
1752         lea     16(%rcx), %rcx
1753         movaps  %xmm2, %xmm3
1754 # ifdef USE_AS_STRNCPY
1755         sub     $16, %r8
1756         jbe     L(StrncpyExit12Case2OrCase3)
1757 # endif
1758         test    %rax, %rax
1759         jnz     L(Shl12LoopExit)
1760
1761         palignr $12, %xmm1, %xmm2
1762         movaps  %xmm3, %xmm1
1763         movaps  %xmm2, (%rdx)
1764         lea     20(%rcx), %rcx
1765         lea     16(%rdx), %rdx
1766
1767         mov     %rcx, %rax
1768         and     $-0x40, %rcx
1769         sub     %rcx, %rax
1770         lea     -4(%rcx), %rcx
1771         sub     %rax, %rdx
1772 # ifdef USE_AS_STRNCPY
1773         add     %rax, %r8
1774 # endif
1775         movaps  -12(%rcx), %xmm1
1776
1777 L(Shl12LoopStart):
1778         movaps  4(%rcx), %xmm2
1779         movaps  20(%rcx), %xmm3
1780         movaps  %xmm3, %xmm6
1781         movaps  36(%rcx), %xmm4
1782         movaps  %xmm4, %xmm7
1783         movaps  52(%rcx), %xmm5
1784         pminub  %xmm2, %xmm6
1785         pminub  %xmm5, %xmm7
1786         pminub  %xmm6, %xmm7
1787         pcmpeqb %xmm0, %xmm7
1788         pmovmskb %xmm7, %rax
1789         movaps  %xmm5, %xmm7
1790         palignr $12, %xmm4, %xmm5
1791         test    %rax, %rax
1792         palignr $12, %xmm3, %xmm4
1793         jnz     L(Shl12Start)
1794 # ifdef USE_AS_STRNCPY
1795         sub     $64, %r8
1796         jbe     L(StrncpyLeave12)
1797 # endif
1798         palignr $12, %xmm2, %xmm3
1799         lea     64(%rcx), %rcx
1800         palignr $12, %xmm1, %xmm2
1801         movaps  %xmm7, %xmm1
1802         movaps  %xmm5, 48(%rdx)
1803         movaps  %xmm4, 32(%rdx)
1804         movaps  %xmm3, 16(%rdx)
1805         movaps  %xmm2, (%rdx)
1806         lea     64(%rdx), %rdx
1807         jmp     L(Shl12LoopStart)
1808
1809 L(Shl12LoopExit):
1810         movaps  (%rdx), %xmm6
1811         psrldq  $4, %xmm6
1812         mov     $4, %rsi
1813         palignr $12, %xmm1, %xmm6
1814         movaps  %xmm6, (%rdx)
1815         jmp     L(CopyFrom1To16Bytes)
1816
1817         .p2align 4
1818 L(Shl13):
1819         movaps  -13(%rcx), %xmm1
1820         movaps  3(%rcx), %xmm2
1821 L(Shl13Start):
1822         pcmpeqb %xmm2, %xmm0
1823         pmovmskb %xmm0, %rax
1824         movaps  %xmm2, %xmm3
1825 # ifdef USE_AS_STRNCPY
1826         sub     $16, %r8
1827         jbe     L(StrncpyExit13Case2OrCase3)
1828 # endif
1829         test    %rax, %rax
1830         jnz     L(Shl13LoopExit)
1831
1832         palignr $13, %xmm1, %xmm2
1833         movaps  %xmm3, %xmm1
1834         movaps  %xmm2, (%rdx)
1835         movaps  19(%rcx), %xmm2
1836
1837         pcmpeqb %xmm2, %xmm0
1838         lea     16(%rdx), %rdx
1839         pmovmskb %xmm0, %rax
1840         lea     16(%rcx), %rcx
1841         movaps  %xmm2, %xmm3
1842 # ifdef USE_AS_STRNCPY
1843         sub     $16, %r8
1844         jbe     L(StrncpyExit13Case2OrCase3)
1845 # endif
1846         test    %rax, %rax
1847         jnz     L(Shl13LoopExit)
1848
1849         palignr $13, %xmm1, %xmm2
1850         movaps  %xmm2, (%rdx)
1851         movaps  19(%rcx), %xmm2
1852         movaps  %xmm3, %xmm1
1853
1854         pcmpeqb %xmm2, %xmm0
1855         lea     16(%rdx), %rdx
1856         pmovmskb %xmm0, %rax
1857         lea     16(%rcx), %rcx
1858         movaps  %xmm2, %xmm3
1859 # ifdef USE_AS_STRNCPY
1860         sub     $16, %r8
1861         jbe     L(StrncpyExit13Case2OrCase3)
1862 # endif
1863         test    %rax, %rax
1864         jnz     L(Shl13LoopExit)
1865
1866         palignr $13, %xmm1, %xmm2
1867         movaps  %xmm3, %xmm1
1868         movaps  %xmm2, (%rdx)
1869         movaps  19(%rcx), %xmm2
1870
1871         pcmpeqb %xmm2, %xmm0
1872         lea     16(%rdx), %rdx
1873         pmovmskb %xmm0, %rax
1874         lea     16(%rcx), %rcx
1875         movaps  %xmm2, %xmm3
1876 # ifdef USE_AS_STRNCPY
1877         sub     $16, %r8
1878         jbe     L(StrncpyExit13Case2OrCase3)
1879 # endif
1880         test    %rax, %rax
1881         jnz     L(Shl13LoopExit)
1882
1883         palignr $13, %xmm1, %xmm2
1884         movaps  %xmm3, %xmm1
1885         movaps  %xmm2, (%rdx)
1886         lea     19(%rcx), %rcx
1887         lea     16(%rdx), %rdx
1888
1889         mov     %rcx, %rax
1890         and     $-0x40, %rcx
1891         sub     %rcx, %rax
1892         lea     -3(%rcx), %rcx
1893         sub     %rax, %rdx
1894 # ifdef USE_AS_STRNCPY
1895         add     %rax, %r8
1896 # endif
1897         movaps  -13(%rcx), %xmm1
1898
1899 L(Shl13LoopStart):
1900         movaps  3(%rcx), %xmm2
1901         movaps  19(%rcx), %xmm3
1902         movaps  %xmm3, %xmm6
1903         movaps  35(%rcx), %xmm4
1904         movaps  %xmm4, %xmm7
1905         movaps  51(%rcx), %xmm5
1906         pminub  %xmm2, %xmm6
1907         pminub  %xmm5, %xmm7
1908         pminub  %xmm6, %xmm7
1909         pcmpeqb %xmm0, %xmm7
1910         pmovmskb %xmm7, %rax
1911         movaps  %xmm5, %xmm7
1912         palignr $13, %xmm4, %xmm5
1913         test    %rax, %rax
1914         palignr $13, %xmm3, %xmm4
1915         jnz     L(Shl13Start)
1916 # ifdef USE_AS_STRNCPY
1917         sub     $64, %r8
1918         jbe     L(StrncpyLeave13)
1919 # endif
1920         palignr $13, %xmm2, %xmm3
1921         lea     64(%rcx), %rcx
1922         palignr $13, %xmm1, %xmm2
1923         movaps  %xmm7, %xmm1
1924         movaps  %xmm5, 48(%rdx)
1925         movaps  %xmm4, 32(%rdx)
1926         movaps  %xmm3, 16(%rdx)
1927         movaps  %xmm2, (%rdx)
1928         lea     64(%rdx), %rdx
1929         jmp     L(Shl13LoopStart)
1930
1931 L(Shl13LoopExit):
1932         movaps  (%rdx), %xmm6
1933         psrldq  $3, %xmm6
1934         mov     $3, %rsi
1935         palignr $13, %xmm1, %xmm6
1936         movaps  %xmm6, (%rdx)
1937         jmp     L(CopyFrom1To16Bytes)
1938
1939         .p2align 4
1940 L(Shl14):
1941         movaps  -14(%rcx), %xmm1
1942         movaps  2(%rcx), %xmm2
1943 L(Shl14Start):
1944         pcmpeqb %xmm2, %xmm0
1945         pmovmskb %xmm0, %rax
1946         movaps  %xmm2, %xmm3
1947 # ifdef USE_AS_STRNCPY
1948         sub     $16, %r8
1949         jbe     L(StrncpyExit14Case2OrCase3)
1950 # endif
1951         test    %rax, %rax
1952         jnz     L(Shl14LoopExit)
1953
1954         palignr $14, %xmm1, %xmm2
1955         movaps  %xmm3, %xmm1
1956         movaps  %xmm2, (%rdx)
1957         movaps  18(%rcx), %xmm2
1958
1959         pcmpeqb %xmm2, %xmm0
1960         lea     16(%rdx), %rdx
1961         pmovmskb %xmm0, %rax
1962         lea     16(%rcx), %rcx
1963         movaps  %xmm2, %xmm3
1964 # ifdef USE_AS_STRNCPY
1965         sub     $16, %r8
1966         jbe     L(StrncpyExit14Case2OrCase3)
1967 # endif
1968         test    %rax, %rax
1969         jnz     L(Shl14LoopExit)
1970
1971         palignr $14, %xmm1, %xmm2
1972         movaps  %xmm2, (%rdx)
1973         movaps  18(%rcx), %xmm2
1974         movaps  %xmm3, %xmm1
1975
1976         pcmpeqb %xmm2, %xmm0
1977         lea     16(%rdx), %rdx
1978         pmovmskb %xmm0, %rax
1979         lea     16(%rcx), %rcx
1980         movaps  %xmm2, %xmm3
1981 # ifdef USE_AS_STRNCPY
1982         sub     $16, %r8
1983         jbe     L(StrncpyExit14Case2OrCase3)
1984 # endif
1985         test    %rax, %rax
1986         jnz     L(Shl14LoopExit)
1987
1988         palignr $14, %xmm1, %xmm2
1989         movaps  %xmm3, %xmm1
1990         movaps  %xmm2, (%rdx)
1991         movaps  18(%rcx), %xmm2
1992
1993         pcmpeqb %xmm2, %xmm0
1994         lea     16(%rdx), %rdx
1995         pmovmskb %xmm0, %rax
1996         lea     16(%rcx), %rcx
1997         movaps  %xmm2, %xmm3
1998 # ifdef USE_AS_STRNCPY
1999         sub     $16, %r8
2000         jbe     L(StrncpyExit14Case2OrCase3)
2001 # endif
2002         test    %rax, %rax
2003         jnz     L(Shl14LoopExit)
2004
2005         palignr $14, %xmm1, %xmm2
2006         movaps  %xmm3, %xmm1
2007         movaps  %xmm2, (%rdx)
2008         lea     18(%rcx), %rcx
2009         lea     16(%rdx), %rdx
2010
2011         mov     %rcx, %rax
2012         and     $-0x40, %rcx
2013         sub     %rcx, %rax
2014         lea     -2(%rcx), %rcx
2015         sub     %rax, %rdx
2016 # ifdef USE_AS_STRNCPY
2017         add     %rax, %r8
2018 # endif
2019         movaps  -14(%rcx), %xmm1
2020
2021 L(Shl14LoopStart):
2022         movaps  2(%rcx), %xmm2
2023         movaps  18(%rcx), %xmm3
2024         movaps  %xmm3, %xmm6
2025         movaps  34(%rcx), %xmm4
2026         movaps  %xmm4, %xmm7
2027         movaps  50(%rcx), %xmm5
2028         pminub  %xmm2, %xmm6
2029         pminub  %xmm5, %xmm7
2030         pminub  %xmm6, %xmm7
2031         pcmpeqb %xmm0, %xmm7
2032         pmovmskb %xmm7, %rax
2033         movaps  %xmm5, %xmm7
2034         palignr $14, %xmm4, %xmm5
2035         test    %rax, %rax
2036         palignr $14, %xmm3, %xmm4
2037         jnz     L(Shl14Start)
2038 # ifdef USE_AS_STRNCPY
2039         sub     $64, %r8
2040         jbe     L(StrncpyLeave14)
2041 # endif
2042         palignr $14, %xmm2, %xmm3
2043         lea     64(%rcx), %rcx
2044         palignr $14, %xmm1, %xmm2
2045         movaps  %xmm7, %xmm1
2046         movaps  %xmm5, 48(%rdx)
2047         movaps  %xmm4, 32(%rdx)
2048         movaps  %xmm3, 16(%rdx)
2049         movaps  %xmm2, (%rdx)
2050         lea     64(%rdx), %rdx
2051         jmp     L(Shl14LoopStart)
2052
2053 L(Shl14LoopExit):
2054         movaps  (%rdx), %xmm6
2055         psrldq  $2, %xmm6
2056         mov     $2, %rsi
2057         palignr $14, %xmm1, %xmm6
2058         movaps  %xmm6, (%rdx)
2059         jmp     L(CopyFrom1To16Bytes)
2060
2061         .p2align 4
2062 L(Shl15):
2063         movaps  -15(%rcx), %xmm1
2064         movaps  1(%rcx), %xmm2
2065 L(Shl15Start):
2066         pcmpeqb %xmm2, %xmm0
2067         pmovmskb %xmm0, %rax
2068         movaps  %xmm2, %xmm3
2069 # ifdef USE_AS_STRNCPY
2070         sub     $16, %r8
2071         jbe     L(StrncpyExit15Case2OrCase3)
2072 # endif
2073         test    %rax, %rax
2074         jnz     L(Shl15LoopExit)
2075
2076         palignr $15, %xmm1, %xmm2
2077         movaps  %xmm3, %xmm1
2078         movaps  %xmm2, (%rdx)
2079         movaps  17(%rcx), %xmm2
2080
2081         pcmpeqb %xmm2, %xmm0
2082         lea     16(%rdx), %rdx
2083         pmovmskb %xmm0, %rax
2084         lea     16(%rcx), %rcx
2085         movaps  %xmm2, %xmm3
2086 # ifdef USE_AS_STRNCPY
2087         sub     $16, %r8
2088         jbe     L(StrncpyExit15Case2OrCase3)
2089 # endif
2090         test    %rax, %rax
2091         jnz     L(Shl15LoopExit)
2092
2093         palignr $15, %xmm1, %xmm2
2094         movaps  %xmm2, (%rdx)
2095         movaps  17(%rcx), %xmm2
2096         movaps  %xmm3, %xmm1
2097
2098         pcmpeqb %xmm2, %xmm0
2099         lea     16(%rdx), %rdx
2100         pmovmskb %xmm0, %rax
2101         lea     16(%rcx), %rcx
2102         movaps  %xmm2, %xmm3
2103 # ifdef USE_AS_STRNCPY
2104         sub     $16, %r8
2105         jbe     L(StrncpyExit15Case2OrCase3)
2106 # endif
2107         test    %rax, %rax
2108         jnz     L(Shl15LoopExit)
2109
2110         palignr $15, %xmm1, %xmm2
2111         movaps  %xmm3, %xmm1
2112         movaps  %xmm2, (%rdx)
2113         movaps  17(%rcx), %xmm2
2114
2115         pcmpeqb %xmm2, %xmm0
2116         lea     16(%rdx), %rdx
2117         pmovmskb %xmm0, %rax
2118         lea     16(%rcx), %rcx
2119         movaps  %xmm2, %xmm3
2120 # ifdef USE_AS_STRNCPY
2121         sub     $16, %r8
2122         jbe     L(StrncpyExit15Case2OrCase3)
2123 # endif
2124         test    %rax, %rax
2125         jnz     L(Shl15LoopExit)
2126
2127         palignr $15, %xmm1, %xmm2
2128         movaps  %xmm3, %xmm1
2129         movaps  %xmm2, (%rdx)
2130         lea     17(%rcx), %rcx
2131         lea     16(%rdx), %rdx
2132
2133         mov     %rcx, %rax
2134         and     $-0x40, %rcx
2135         sub     %rcx, %rax
2136         lea     -1(%rcx), %rcx
2137         sub     %rax, %rdx
2138 # ifdef USE_AS_STRNCPY
2139         add     %rax, %r8
2140 # endif
2141         movaps  -15(%rcx), %xmm1
2142
2143 L(Shl15LoopStart):
2144         movaps  1(%rcx), %xmm2
2145         movaps  17(%rcx), %xmm3
2146         movaps  %xmm3, %xmm6
2147         movaps  33(%rcx), %xmm4
2148         movaps  %xmm4, %xmm7
2149         movaps  49(%rcx), %xmm5
2150         pminub  %xmm2, %xmm6
2151         pminub  %xmm5, %xmm7
2152         pminub  %xmm6, %xmm7
2153         pcmpeqb %xmm0, %xmm7
2154         pmovmskb %xmm7, %rax
2155         movaps  %xmm5, %xmm7
2156         palignr $15, %xmm4, %xmm5
2157         test    %rax, %rax
2158         palignr $15, %xmm3, %xmm4
2159         jnz     L(Shl15Start)
2160 # ifdef USE_AS_STRNCPY
2161         sub     $64, %r8
2162         jbe     L(StrncpyLeave15)
2163 # endif
2164         palignr $15, %xmm2, %xmm3
2165         lea     64(%rcx), %rcx
2166         palignr $15, %xmm1, %xmm2
2167         movaps  %xmm7, %xmm1
2168         movaps  %xmm5, 48(%rdx)
2169         movaps  %xmm4, 32(%rdx)
2170         movaps  %xmm3, 16(%rdx)
2171         movaps  %xmm2, (%rdx)
2172         lea     64(%rdx), %rdx
2173         jmp     L(Shl15LoopStart)
2174
2175 L(Shl15LoopExit):
2176         movaps  (%rdx), %xmm6
2177         psrldq  $1, %xmm6
2178         mov     $1, %rsi
2179         palignr $15, %xmm1, %xmm6
2180         movaps  %xmm6, (%rdx)
2181 # ifdef USE_AS_STRCAT
2182         jmp     L(CopyFrom1To16Bytes)
2183 # endif
2184
2185 # ifndef USE_AS_STRCAT
2186         .p2align 4
2187 L(CopyFrom1To16Bytes):
2188 #  ifdef USE_AS_STRNCPY
2189         add     $16, %r8
2190 #  endif
2191         add     %rsi, %rdx
2192         add     %rsi, %rcx
2193
2194         test    %al, %al
2195         jz      L(ExitHigh)
2196         test    $0x01, %al
2197         jnz     L(Exit1)
2198         test    $0x02, %al
2199         jnz     L(Exit2)
2200         test    $0x04, %al
2201         jnz     L(Exit3)
2202         test    $0x08, %al
2203         jnz     L(Exit4)
2204         test    $0x10, %al
2205         jnz     L(Exit5)
2206         test    $0x20, %al
2207         jnz     L(Exit6)
2208         test    $0x40, %al
2209         jnz     L(Exit7)
2210
2211         .p2align 4
2212 L(Exit8):
2213         mov     (%rcx), %rax
2214         mov     %rax, (%rdx)
2215 #  ifdef USE_AS_STPCPY
2216         lea     7(%rdx), %rax
2217 #  else
2218         mov     %rdi, %rax
2219 #  endif
2220 #  ifdef USE_AS_STRNCPY
2221         sub     $8, %r8
2222         lea     8(%rdx), %rcx
2223         jnz     L(StrncpyFillTailWithZero1)
2224 #   ifdef USE_AS_STPCPY
2225         cmpb    $1, (%rax)
2226         sbb     $-1, %rax
2227 #   endif
2228 #  endif
2229         ret
2230
2231         .p2align 4
2232 L(ExitHigh):
2233         test    $0x01, %ah
2234         jnz     L(Exit9)
2235         test    $0x02, %ah
2236         jnz     L(Exit10)
2237         test    $0x04, %ah
2238         jnz     L(Exit11)
2239         test    $0x08, %ah
2240         jnz     L(Exit12)
2241         test    $0x10, %ah
2242         jnz     L(Exit13)
2243         test    $0x20, %ah
2244         jnz     L(Exit14)
2245         test    $0x40, %ah
2246         jnz     L(Exit15)
2247
2248         .p2align 4
2249 L(Exit16):
2250         mov     (%rcx), %rax
2251         mov     %rax, (%rdx)
2252         mov     8(%rcx), %rax
2253         mov     %rax, 8(%rdx)
2254 #  ifdef USE_AS_STPCPY
2255         lea     15(%rdx), %rax
2256 #  else
2257         mov     %rdi, %rax
2258 #  endif
2259 #  ifdef USE_AS_STRNCPY
2260         sub     $16, %r8
2261         lea     16(%rdx), %rcx
2262         jnz     L(StrncpyFillTailWithZero1)
2263 #   ifdef USE_AS_STPCPY
2264         cmpb    $1, (%rax)
2265         sbb     $-1, %rax
2266 #   endif
2267 #  endif
2268         ret
2269
2270 #  ifdef USE_AS_STRNCPY
2271
2272         .p2align 4
2273 L(CopyFrom1To16BytesCase2):
2274         add     $16, %r8
2275         add     %rsi, %rcx
2276         lea     (%rsi, %rdx), %rsi
2277         lea     -9(%r8), %rdx
2278         and     $1<<7, %dh
2279         or      %al, %dh
2280         test    %dh, %dh
2281         lea     (%rsi), %rdx
2282         jz      L(ExitHighCase2)
2283
2284         cmp     $1, %r8
2285         je      L(Exit1)
2286         test    $0x01, %al
2287         jnz     L(Exit1)
2288         cmp     $2, %r8
2289         je      L(Exit2)
2290         test    $0x02, %al
2291         jnz     L(Exit2)
2292         cmp     $3, %r8
2293         je      L(Exit3)
2294         test    $0x04, %al
2295         jnz     L(Exit3)
2296         cmp     $4, %r8
2297         je      L(Exit4)
2298         test    $0x08, %al
2299         jnz     L(Exit4)
2300         cmp     $5, %r8
2301         je      L(Exit5)
2302         test    $0x10, %al
2303         jnz     L(Exit5)
2304         cmp     $6, %r8
2305         je      L(Exit6)
2306         test    $0x20, %al
2307         jnz     L(Exit6)
2308         cmp     $7, %r8
2309         je      L(Exit7)
2310         test    $0x40, %al
2311         jnz     L(Exit7)
2312         jmp     L(Exit8)
2313
2314         .p2align 4
2315 L(ExitHighCase2):
2316         cmp     $9, %r8
2317         je      L(Exit9)
2318         test    $0x01, %ah
2319         jnz     L(Exit9)
2320         cmp     $10, %r8
2321         je      L(Exit10)
2322         test    $0x02, %ah
2323         jnz     L(Exit10)
2324         cmp     $11, %r8
2325         je      L(Exit11)
2326         test    $0x04, %ah
2327         jnz     L(Exit11)
2328         cmp     $12, %r8
2329         je      L(Exit12)
2330         test    $0x8, %ah
2331         jnz     L(Exit12)
2332         cmp     $13, %r8
2333         je      L(Exit13)
2334         test    $0x10, %ah
2335         jnz     L(Exit13)
2336         cmp     $14, %r8
2337         je      L(Exit14)
2338         test    $0x20, %ah
2339         jnz     L(Exit14)
2340         cmp     $15, %r8
2341         je      L(Exit15)
2342         test    $0x40, %ah
2343         jnz     L(Exit15)
2344         jmp     L(Exit16)
2345
2346 L(CopyFrom1To16BytesCase2OrCase3):
2347         test    %rax, %rax
2348         jnz     L(CopyFrom1To16BytesCase2)
2349
2350         .p2align 4
2351 L(CopyFrom1To16BytesCase3):
2352         add     $16, %r8
2353         add     %rsi, %rdx
2354         add     %rsi, %rcx
2355
2356         cmp     $16, %r8
2357         je      L(Exit16)
2358         cmp     $8, %r8
2359         je      L(Exit8)
2360         jg      L(More8Case3)
2361         cmp     $4, %r8
2362         je      L(Exit4)
2363         jg      L(More4Case3)
2364         cmp     $2, %r8
2365         jl      L(Exit1)
2366         je      L(Exit2)
2367         jg      L(Exit3)
2368 L(More8Case3): /* but less than 16 */
2369         cmp     $12, %r8
2370         je      L(Exit12)
2371         jl      L(Less12Case3)
2372         cmp     $14, %r8
2373         jl      L(Exit13)
2374         je      L(Exit14)
2375         jg      L(Exit15)
2376 L(More4Case3): /* but less than 8 */
2377         cmp     $6, %r8
2378         jl      L(Exit5)
2379         je      L(Exit6)
2380         jg      L(Exit7)
2381 L(Less12Case3): /* but more than 8 */
2382         cmp     $10, %r8
2383         jl      L(Exit9)
2384         je      L(Exit10)
2385         jg      L(Exit11)
2386 #  endif
2387
2388         .p2align 4
2389 L(Exit1):
2390         movb    (%rcx), %al
2391         movb    %al, (%rdx)
2392 #  ifdef USE_AS_STPCPY
2393         lea     (%rdx), %rax
2394 #  else
2395         mov     %rdi, %rax
2396 #  endif
2397 #  ifdef USE_AS_STRNCPY
2398         sub     $1, %r8
2399         lea     1(%rdx), %rcx
2400         jnz     L(StrncpyFillTailWithZero1)
2401 #   ifdef USE_AS_STPCPY
2402         cmpb    $1, (%rax)
2403         sbb     $-1, %rax
2404 #   endif
2405 #  endif
2406         ret
2407
2408         .p2align 4
2409 L(Exit2):
2410         movw    (%rcx), %ax
2411         movw    %ax, (%rdx)
2412 #  ifdef USE_AS_STPCPY
2413         lea     1(%rdx), %rax
2414 #  else
2415         mov     %rdi, %rax
2416 #  endif
2417 #  ifdef USE_AS_STRNCPY
2418         sub     $2, %r8
2419         lea     2(%rdx), %rcx
2420         jnz     L(StrncpyFillTailWithZero1)
2421 #   ifdef USE_AS_STPCPY
2422         cmpb    $1, (%rax)
2423         sbb     $-1, %rax
2424 #   endif
2425 #  endif
2426         ret
2427
2428         .p2align 4
2429 L(Exit3):
2430         movw    (%rcx), %ax
2431         movw    %ax, (%rdx)
2432         movb    2(%rcx), %al
2433         movb    %al, 2(%rdx)
2434 #  ifdef USE_AS_STPCPY
2435         lea     2(%rdx), %rax
2436 #  else
2437         mov     %rdi, %rax
2438 #  endif
2439 #  ifdef USE_AS_STRNCPY
2440         sub     $3, %r8
2441         lea     3(%rdx), %rcx
2442         jnz     L(StrncpyFillTailWithZero1)
2443 #   ifdef USE_AS_STPCPY
2444         cmpb    $1, (%rax)
2445         sbb     $-1, %rax
2446 #   endif
2447 #  endif
2448         ret
2449
2450         .p2align 4
2451 L(Exit4):
2452         movl    (%rcx), %eax
2453         movl    %eax, (%rdx)
2454 #  ifdef USE_AS_STPCPY
2455         lea     3(%rdx), %rax
2456 #  else
2457         mov     %rdi, %rax
2458 #  endif
2459 #  ifdef USE_AS_STRNCPY
2460         sub     $4, %r8
2461         lea     4(%rdx), %rcx
2462         jnz     L(StrncpyFillTailWithZero1)
2463 #   ifdef USE_AS_STPCPY
2464         cmpb    $1, (%rax)
2465         sbb     $-1, %rax
2466 #  endif
2467 #  endif
2468         ret
2469
2470         .p2align 4
2471 L(Exit5):
2472         movl    (%rcx), %eax
2473         movl    %eax, (%rdx)
2474         movb    4(%rcx), %al
2475         movb    %al, 4(%rdx)
2476 #  ifdef USE_AS_STPCPY
2477         lea     4(%rdx), %rax
2478 #  else
2479         mov     %rdi, %rax
2480 #  endif
2481 #  ifdef USE_AS_STRNCPY
2482         sub     $5, %r8
2483         lea     5(%rdx), %rcx
2484         jnz     L(StrncpyFillTailWithZero1)
2485 #   ifdef USE_AS_STPCPY
2486         cmpb    $1, (%rax)
2487         sbb     $-1, %rax
2488 #   endif
2489 #  endif
2490         ret
2491
2492         .p2align 4
2493 L(Exit6):
2494         movl    (%rcx), %eax
2495         movl    %eax, (%rdx)
2496         movw    4(%rcx), %ax
2497         movw    %ax, 4(%rdx)
2498 #  ifdef USE_AS_STPCPY
2499         lea     5(%rdx), %rax
2500 #  else
2501         mov     %rdi, %rax
2502 #  endif
2503 #  ifdef USE_AS_STRNCPY
2504         sub     $6, %r8
2505         lea     6(%rdx), %rcx
2506         jnz     L(StrncpyFillTailWithZero1)
2507 #   ifdef USE_AS_STPCPY
2508         cmpb    $1, (%rax)
2509         sbb     $-1, %rax
2510 #   endif
2511 #  endif
2512         ret
2513
2514         .p2align 4
2515 L(Exit7):
2516         movl    (%rcx), %eax
2517         movl    %eax, (%rdx)
2518         movl    3(%rcx), %eax
2519         movl    %eax, 3(%rdx)
2520 #  ifdef USE_AS_STPCPY
2521         lea     6(%rdx), %rax
2522 #  else
2523         mov     %rdi, %rax
2524 #  endif
2525 #  ifdef USE_AS_STRNCPY
2526         sub     $7, %r8
2527         lea     7(%rdx), %rcx
2528         jnz     L(StrncpyFillTailWithZero1)
2529 #   ifdef USE_AS_STPCPY
2530         cmpb    $1, (%rax)
2531         sbb     $-1, %rax
2532 #   endif
2533 #  endif
2534         ret
2535
2536         .p2align 4
2537 L(Exit9):
2538         mov     (%rcx), %rax
2539         mov     %rax, (%rdx)
2540         mov     5(%rcx), %eax
2541         mov     %eax, 5(%rdx)
2542 #  ifdef USE_AS_STPCPY
2543         lea     8(%rdx), %rax
2544 #  else
2545         mov     %rdi, %rax
2546 #  endif
2547 #  ifdef USE_AS_STRNCPY
2548         sub     $9, %r8
2549         lea     9(%rdx), %rcx
2550         jnz     L(StrncpyFillTailWithZero1)
2551 #   ifdef USE_AS_STPCPY
2552         cmpb    $1, (%rax)
2553         sbb     $-1, %rax
2554 #   endif
2555 #  endif
2556         ret
2557
2558         .p2align 4
2559 L(Exit10):
2560         mov     (%rcx), %rax
2561         mov     %rax, (%rdx)
2562         mov     6(%rcx), %eax
2563         mov     %eax, 6(%rdx)
2564 #  ifdef USE_AS_STPCPY
2565         lea     9(%rdx), %rax
2566 #  else
2567         mov     %rdi, %rax
2568 #  endif
2569 #  ifdef USE_AS_STRNCPY
2570         sub     $10, %r8
2571         lea     10(%rdx), %rcx
2572         jnz     L(StrncpyFillTailWithZero1)
2573 #   ifdef USE_AS_STPCPY
2574         cmpb    $1, (%rax)
2575         sbb     $-1, %rax
2576 #   endif
2577 #  endif
2578         ret
2579
2580         .p2align 4
2581 L(Exit11):
2582         mov     (%rcx), %rax
2583         mov     %rax, (%rdx)
2584         mov     7(%rcx), %eax
2585         mov     %eax, 7(%rdx)
2586 #  ifdef USE_AS_STPCPY
2587         lea     10(%rdx), %rax
2588 #  else
2589         mov     %rdi, %rax
2590 #  endif
2591 #  ifdef USE_AS_STRNCPY
2592         sub     $11, %r8
2593         lea     11(%rdx), %rcx
2594         jnz     L(StrncpyFillTailWithZero1)
2595 #   ifdef USE_AS_STPCPY
2596         cmpb    $1, (%rax)
2597         sbb     $-1, %rax
2598 #   endif
2599 #  endif
2600         ret
2601
2602         .p2align 4
2603 L(Exit12):
2604         mov     (%rcx), %rax
2605         mov     %rax, (%rdx)
2606         mov     8(%rcx), %eax
2607         mov     %eax, 8(%rdx)
2608 #  ifdef USE_AS_STPCPY
2609         lea     11(%rdx), %rax
2610 #  else
2611         mov     %rdi, %rax
2612 #  endif
2613 #  ifdef USE_AS_STRNCPY
2614         sub     $12, %r8
2615         lea     12(%rdx), %rcx
2616         jnz     L(StrncpyFillTailWithZero1)
2617 #   ifdef USE_AS_STPCPY
2618         cmpb    $1, (%rax)
2619         sbb     $-1, %rax
2620 #   endif
2621 #  endif
2622         ret
2623
2624         .p2align 4
2625 L(Exit13):
2626         mov     (%rcx), %rax
2627         mov     %rax, (%rdx)
2628         mov     5(%rcx), %rax
2629         mov     %rax, 5(%rdx)
2630 #  ifdef USE_AS_STPCPY
2631         lea     12(%rdx), %rax
2632 #  else
2633         mov     %rdi, %rax
2634 #  endif
2635 #  ifdef USE_AS_STRNCPY
2636         sub     $13, %r8
2637         lea     13(%rdx), %rcx
2638         jnz     L(StrncpyFillTailWithZero1)
2639 #   ifdef USE_AS_STPCPY
2640         cmpb    $1, (%rax)
2641         sbb     $-1, %rax
2642 #   endif
2643 #  endif
2644         ret
2645
2646         .p2align 4
2647 L(Exit14):
2648         mov     (%rcx), %rax
2649         mov     %rax, (%rdx)
2650         mov     6(%rcx), %rax
2651         mov     %rax, 6(%rdx)
2652 #  ifdef USE_AS_STPCPY
2653         lea     13(%rdx), %rax
2654 #  else
2655         mov     %rdi, %rax
2656 #  endif
2657 #  ifdef USE_AS_STRNCPY
2658         sub     $14, %r8
2659         lea     14(%rdx), %rcx
2660         jnz     L(StrncpyFillTailWithZero1)
2661 #   ifdef USE_AS_STPCPY
2662         cmpb    $1, (%rax)
2663         sbb     $-1, %rax
2664 #   endif
2665 #  endif
2666         ret
2667
2668         .p2align 4
2669 L(Exit15):
2670         mov     (%rcx), %rax
2671         mov     %rax, (%rdx)
2672         mov     7(%rcx), %rax
2673         mov     %rax, 7(%rdx)
2674 #  ifdef USE_AS_STPCPY
2675         lea     14(%rdx), %rax
2676 #  else
2677         mov     %rdi, %rax
2678 #  endif
2679 #  ifdef USE_AS_STRNCPY
2680         sub     $15, %r8
2681         lea     15(%rdx), %rcx
2682         jnz     L(StrncpyFillTailWithZero1)
2683 #   ifdef USE_AS_STPCPY
2684         cmpb    $1, (%rax)
2685         sbb     $-1, %rax
2686 #   endif
2687 #  endif
2688         ret
2689
2690 #  ifdef USE_AS_STRNCPY
2691         .p2align 4
2692 L(Fill0):
2693         ret
2694
2695         .p2align 4
2696 L(Fill1):
2697         movb    %dl, (%rcx)
2698         ret
2699
2700         .p2align 4
2701 L(Fill2):
2702         movw    %dx, (%rcx)
2703         ret
2704
2705         .p2align 4
2706 L(Fill3):
2707         movw    %dx, (%rcx)
2708         movb    %dl, 2(%rcx)
2709         ret
2710
2711         .p2align 4
2712 L(Fill4):
2713         movl    %edx, (%rcx)
2714         ret
2715
2716         .p2align 4
2717 L(Fill5):
2718         movl    %edx, (%rcx)
2719         movb    %dl, 4(%rcx)
2720         ret
2721
2722         .p2align 4
2723 L(Fill6):
2724         movl    %edx, (%rcx)
2725         movw    %dx, 4(%rcx)
2726         ret
2727
2728         .p2align 4
2729 L(Fill7):
2730         movl    %edx, (%rcx)
2731         movl    %edx, 3(%rcx)
2732         ret
2733
2734         .p2align 4
2735 L(Fill8):
2736         mov     %rdx, (%rcx)
2737         ret
2738
2739         .p2align 4
2740 L(Fill9):
2741         mov     %rdx, (%rcx)
2742         movb    %dl, 8(%rcx)
2743         ret
2744
2745         .p2align 4
2746 L(Fill10):
2747         mov     %rdx, (%rcx)
2748         movw    %dx, 8(%rcx)
2749         ret
2750
2751         .p2align 4
2752 L(Fill11):
2753         mov     %rdx, (%rcx)
2754         movl    %edx, 7(%rcx)
2755         ret
2756
2757         .p2align 4
2758 L(Fill12):
2759         mov     %rdx, (%rcx)
2760         movl    %edx, 8(%rcx)
2761         ret
2762
2763         .p2align 4
2764 L(Fill13):
2765         mov     %rdx, (%rcx)
2766         mov     %rdx, 5(%rcx)
2767         ret
2768
2769         .p2align 4
2770 L(Fill14):
2771         mov     %rdx, (%rcx)
2772         mov     %rdx, 6(%rcx)
2773         ret
2774
2775         .p2align 4
2776 L(Fill15):
2777         mov     %rdx, (%rcx)
2778         mov     %rdx, 7(%rcx)
2779         ret
2780
2781         .p2align 4
2782 L(Fill16):
2783         mov     %rdx, (%rcx)
2784         mov     %rdx, 8(%rcx)
2785         ret
2786
2787         .p2align 4
2788 L(StrncpyFillExit1):
2789         lea     16(%r8), %r8
2790 L(FillFrom1To16Bytes):
2791         test    %r8, %r8
2792         jz      L(Fill0)
2793         cmp     $16, %r8
2794         je      L(Fill16)
2795         cmp     $8, %r8
2796         je      L(Fill8)
2797         jg      L(FillMore8)
2798         cmp     $4, %r8
2799         je      L(Fill4)
2800         jg      L(FillMore4)
2801         cmp     $2, %r8
2802         jl      L(Fill1)
2803         je      L(Fill2)
2804         jg      L(Fill3)
2805 L(FillMore8): /* but less than 16 */
2806         cmp     $12, %r8
2807         je      L(Fill12)
2808         jl      L(FillLess12)
2809         cmp     $14, %r8
2810         jl      L(Fill13)
2811         je      L(Fill14)
2812         jg      L(Fill15)
2813 L(FillMore4): /* but less than 8 */
2814         cmp     $6, %r8
2815         jl      L(Fill5)
2816         je      L(Fill6)
2817         jg      L(Fill7)
2818 L(FillLess12): /* but more than 8 */
2819         cmp     $10, %r8
2820         jl      L(Fill9)
2821         je      L(Fill10)
2822         jmp     L(Fill11)
2823
2824         .p2align 4
2825 L(StrncpyFillTailWithZero1):
2826         xor     %rdx, %rdx
2827         sub     $16, %r8
2828         jbe     L(StrncpyFillExit1)
2829
2830         pxor    %xmm0, %xmm0
2831         mov     %rdx, (%rcx)
2832         mov     %rdx, 8(%rcx)
2833
2834         lea     16(%rcx), %rcx
2835
2836         mov     %rcx, %rdx
2837         and     $0xf, %rdx
2838         sub     %rdx, %rcx
2839         add     %rdx, %r8
2840         xor     %rdx, %rdx
2841         sub     $64, %r8
2842         jb      L(StrncpyFillLess64)
2843
2844 L(StrncpyFillLoopMovdqa):
2845         movdqa  %xmm0, (%rcx)
2846         movdqa  %xmm0, 16(%rcx)
2847         movdqa  %xmm0, 32(%rcx)
2848         movdqa  %xmm0, 48(%rcx)
2849         lea     64(%rcx), %rcx
2850         sub     $64, %r8
2851         jae     L(StrncpyFillLoopMovdqa)
2852
2853 L(StrncpyFillLess64):
2854         add     $32, %r8
2855         jl      L(StrncpyFillLess32)
2856         movdqa  %xmm0, (%rcx)
2857         movdqa  %xmm0, 16(%rcx)
2858         lea     32(%rcx), %rcx
2859         sub     $16, %r8
2860         jl      L(StrncpyFillExit1)
2861         movdqa  %xmm0, (%rcx)
2862         lea     16(%rcx), %rcx
2863         jmp     L(FillFrom1To16Bytes)
2864
2865 L(StrncpyFillLess32):
2866         add     $16, %r8
2867         jl      L(StrncpyFillExit1)
2868         movdqa  %xmm0, (%rcx)
2869         lea     16(%rcx), %rcx
2870         jmp     L(FillFrom1To16Bytes)
2871
2872         .p2align 4
2873 L(Exit0):
2874         mov     %rdx, %rax
2875         ret
2876
2877         .p2align 4
2878 L(StrncpyExit15Bytes):
2879         cmp     $9, %r8
2880         je      L(Exit9)
2881         cmpb    $0, 8(%rcx)
2882         jz      L(Exit9)
2883         cmp     $10, %r8
2884         je      L(Exit10)
2885         cmpb    $0, 9(%rcx)
2886         jz      L(Exit10)
2887         cmp     $11, %r8
2888         je      L(Exit11)
2889         cmpb    $0, 10(%rcx)
2890         jz      L(Exit11)
2891         cmp     $12, %r8
2892         je      L(Exit12)
2893         cmpb    $0, 11(%rcx)
2894         jz      L(Exit12)
2895         cmp     $13, %r8
2896         je      L(Exit13)
2897         cmpb    $0, 12(%rcx)
2898         jz      L(Exit13)
2899         cmp     $14, %r8
2900         je      L(Exit14)
2901         cmpb    $0, 13(%rcx)
2902         jz      L(Exit14)
2903         mov     (%rcx), %rax
2904         mov     %rax, (%rdx)
2905         mov     7(%rcx), %rax
2906         mov     %rax, 7(%rdx)
2907 #   ifdef USE_AS_STPCPY
2908         lea     14(%rdx), %rax
2909         cmpb    $1, (%rax)
2910         sbb     $-1, %rax
2911 #   else
2912         mov     %rdi, %rax
2913 #   endif
2914         ret
2915
2916         .p2align 4
2917 L(StrncpyExit8Bytes):
2918         cmp     $1, %r8
2919         je      L(Exit1)
2920         cmpb    $0, (%rcx)
2921         jz      L(Exit1)
2922         cmp     $2, %r8
2923         je      L(Exit2)
2924         cmpb    $0, 1(%rcx)
2925         jz      L(Exit2)
2926         cmp     $3, %r8
2927         je      L(Exit3)
2928         cmpb    $0, 2(%rcx)
2929         jz      L(Exit3)
2930         cmp     $4, %r8
2931         je      L(Exit4)
2932         cmpb    $0, 3(%rcx)
2933         jz      L(Exit4)
2934         cmp     $5, %r8
2935         je      L(Exit5)
2936         cmpb    $0, 4(%rcx)
2937         jz      L(Exit5)
2938         cmp     $6, %r8
2939         je      L(Exit6)
2940         cmpb    $0, 5(%rcx)
2941         jz      L(Exit6)
2942         cmp     $7, %r8
2943         je      L(Exit7)
2944         cmpb    $0, 6(%rcx)
2945         jz      L(Exit7)
2946         mov     (%rcx), %rax
2947         mov     %rax, (%rdx)
2948 #   ifdef USE_AS_STPCPY
2949         lea     7(%rdx), %rax
2950         cmpb    $1, (%rax)
2951         sbb     $-1, %rax
2952 #   else
2953         mov     %rdi, %rax
2954 #   endif
2955         ret
2956
2957 #  endif
2958
2959 # endif
2960
2961 # ifdef USE_AS_STRNCPY
2962
2963 L(StrncpyLeaveCase2OrCase3):
2964         test    %rax, %rax
2965         jnz     L(Aligned64LeaveCase2)
2966
2967 L(Aligned64LeaveCase3):
2968         lea     64(%r8), %r8
2969         sub     $16, %r8
2970         jbe     L(CopyFrom1To16BytesCase3)
2971         movaps  %xmm4, -64(%rdx)
2972         lea     16(%rsi), %rsi
2973         sub     $16, %r8
2974         jbe     L(CopyFrom1To16BytesCase3)
2975         movaps  %xmm5, -48(%rdx)
2976         lea     16(%rsi), %rsi
2977         sub     $16, %r8
2978         jbe     L(CopyFrom1To16BytesCase3)
2979         movaps  %xmm6, -32(%rdx)
2980         lea     16(%rsi), %rsi
2981         lea     -16(%r8), %r8
2982         jmp     L(CopyFrom1To16BytesCase3)
2983
2984 L(Aligned64LeaveCase2):
2985         pcmpeqb %xmm4, %xmm0
2986         pmovmskb %xmm0, %rax
2987         add     $48, %r8
2988         jle     L(CopyFrom1To16BytesCase2OrCase3)
2989         test    %rax, %rax
2990         jnz     L(CopyFrom1To16Bytes)
2991
2992         pcmpeqb %xmm5, %xmm0
2993         pmovmskb %xmm0, %rax
2994         movaps  %xmm4, -64(%rdx)
2995         lea     16(%rsi), %rsi
2996         sub     $16, %r8
2997         jbe     L(CopyFrom1To16BytesCase2OrCase3)
2998         test    %rax, %rax
2999         jnz     L(CopyFrom1To16Bytes)
3000
3001         pcmpeqb %xmm6, %xmm0
3002         pmovmskb %xmm0, %rax
3003         movaps  %xmm5, -48(%rdx)
3004         lea     16(%rsi), %rsi
3005         sub     $16, %r8
3006         jbe     L(CopyFrom1To16BytesCase2OrCase3)
3007         test    %rax, %rax
3008         jnz     L(CopyFrom1To16Bytes)
3009
3010         pcmpeqb %xmm7, %xmm0
3011         pmovmskb %xmm0, %rax
3012         movaps  %xmm6, -32(%rdx)
3013         lea     16(%rsi), %rsi
3014         lea     -16(%r8), %r8
3015         jmp     L(CopyFrom1To16BytesCase2)
3016 /*--------------------------------------------------*/
3017 L(StrncpyExit1Case2OrCase3):
3018         movaps  (%rdx), %xmm6
3019         psrldq  $15, %xmm6
3020         mov     $15, %rsi
3021         palignr $1, %xmm1, %xmm6
3022         movaps  %xmm6, (%rdx)
3023         test    %rax, %rax
3024         jnz     L(CopyFrom1To16BytesCase2)
3025         jmp     L(CopyFrom1To16BytesCase3)
3026
3027 L(StrncpyExit2Case2OrCase3):
3028         movaps  (%rdx), %xmm6
3029         psrldq  $14, %xmm6
3030         mov     $14, %rsi
3031         palignr $2, %xmm1, %xmm6
3032         movaps  %xmm6, (%rdx)
3033         test    %rax, %rax
3034         jnz     L(CopyFrom1To16BytesCase2)
3035         jmp     L(CopyFrom1To16BytesCase3)
3036
3037 L(StrncpyExit3Case2OrCase3):
3038         movaps  (%rdx), %xmm6
3039         psrldq  $13, %xmm6
3040         mov     $13, %rsi
3041         palignr $3, %xmm1, %xmm6
3042         movaps  %xmm6, (%rdx)
3043         test    %rax, %rax
3044         jnz     L(CopyFrom1To16BytesCase2)
3045         jmp     L(CopyFrom1To16BytesCase3)
3046
3047 L(StrncpyExit4Case2OrCase3):
3048         movaps  (%rdx), %xmm6
3049         psrldq  $12, %xmm6
3050         mov     $12, %rsi
3051         palignr $4, %xmm1, %xmm6
3052         movaps  %xmm6, (%rdx)
3053         test    %rax, %rax
3054         jnz     L(CopyFrom1To16BytesCase2)
3055         jmp     L(CopyFrom1To16BytesCase3)
3056
3057 L(StrncpyExit5Case2OrCase3):
3058         movaps  (%rdx), %xmm6
3059         psrldq  $11, %xmm6
3060         mov     $11, %rsi
3061         palignr $5, %xmm1, %xmm6
3062         movaps  %xmm6, (%rdx)
3063         test    %rax, %rax
3064         jnz     L(CopyFrom1To16BytesCase2)
3065         jmp     L(CopyFrom1To16BytesCase3)
3066
3067 L(StrncpyExit6Case2OrCase3):
3068         movaps  (%rdx), %xmm6
3069         psrldq  $10, %xmm6
3070         mov     $10, %rsi
3071         palignr $6, %xmm1, %xmm6
3072         movaps  %xmm6, (%rdx)
3073         test    %rax, %rax
3074         jnz     L(CopyFrom1To16BytesCase2)
3075         jmp     L(CopyFrom1To16BytesCase3)
3076
3077 L(StrncpyExit7Case2OrCase3):
3078         movaps  (%rdx), %xmm6
3079         psrldq  $9, %xmm6
3080         mov     $9, %rsi
3081         palignr $7, %xmm1, %xmm6
3082         movaps  %xmm6, (%rdx)
3083         test    %rax, %rax
3084         jnz     L(CopyFrom1To16BytesCase2)
3085         jmp     L(CopyFrom1To16BytesCase3)
3086
3087 L(StrncpyExit8Case2OrCase3):
3088         movaps  (%rdx), %xmm6
3089         psrldq  $8, %xmm6
3090         mov     $8, %rsi
3091         palignr $8, %xmm1, %xmm6
3092         movaps  %xmm6, (%rdx)
3093         test    %rax, %rax
3094         jnz     L(CopyFrom1To16BytesCase2)
3095         jmp     L(CopyFrom1To16BytesCase3)
3096
3097 L(StrncpyExit9Case2OrCase3):
3098         movaps  (%rdx), %xmm6
3099         psrldq  $7, %xmm6
3100         mov     $7, %rsi
3101         palignr $9, %xmm1, %xmm6
3102         movaps  %xmm6, (%rdx)
3103         test    %rax, %rax
3104         jnz     L(CopyFrom1To16BytesCase2)
3105         jmp     L(CopyFrom1To16BytesCase3)
3106
3107 L(StrncpyExit10Case2OrCase3):
3108         movaps  (%rdx), %xmm6
3109         psrldq  $6, %xmm6
3110         mov     $6, %rsi
3111         palignr $10, %xmm1, %xmm6
3112         movaps  %xmm6, (%rdx)
3113         test    %rax, %rax
3114         jnz     L(CopyFrom1To16BytesCase2)
3115         jmp     L(CopyFrom1To16BytesCase3)
3116
3117 L(StrncpyExit11Case2OrCase3):
3118         movaps  (%rdx), %xmm6
3119         psrldq  $5, %xmm6
3120         mov     $5, %rsi
3121         palignr $11, %xmm1, %xmm6
3122         movaps  %xmm6, (%rdx)
3123         test    %rax, %rax
3124         jnz     L(CopyFrom1To16BytesCase2)
3125         jmp     L(CopyFrom1To16BytesCase3)
3126
3127 L(StrncpyExit12Case2OrCase3):
3128         movaps  (%rdx), %xmm6
3129         psrldq  $4, %xmm6
3130         mov     $4, %rsi
3131         palignr $12, %xmm1, %xmm6
3132         movaps  %xmm6, (%rdx)
3133         test    %rax, %rax
3134         jnz     L(CopyFrom1To16BytesCase2)
3135         jmp     L(CopyFrom1To16BytesCase3)
3136
3137 L(StrncpyExit13Case2OrCase3):
3138         movaps  (%rdx), %xmm6
3139         psrldq  $3, %xmm6
3140         mov     $3, %rsi
3141         palignr $13, %xmm1, %xmm6
3142         movaps  %xmm6, (%rdx)
3143         test    %rax, %rax
3144         jnz     L(CopyFrom1To16BytesCase2)
3145         jmp     L(CopyFrom1To16BytesCase3)
3146
3147 L(StrncpyExit14Case2OrCase3):
3148         movaps  (%rdx), %xmm6
3149         psrldq  $2, %xmm6
3150         mov     $2, %rsi
3151         palignr $14, %xmm1, %xmm6
3152         movaps  %xmm6, (%rdx)
3153         test    %rax, %rax
3154         jnz     L(CopyFrom1To16BytesCase2)
3155         jmp     L(CopyFrom1To16BytesCase3)
3156
3157 L(StrncpyExit15Case2OrCase3):
3158         movaps  (%rdx), %xmm6
3159         psrldq  $1, %xmm6
3160         mov     $1, %rsi
3161         palignr $15, %xmm1, %xmm6
3162         movaps  %xmm6, (%rdx)
3163         test    %rax, %rax
3164         jnz     L(CopyFrom1To16BytesCase2)
3165         jmp     L(CopyFrom1To16BytesCase3)
3166
3167 L(StrncpyLeave1):
3168         movaps  %xmm2, %xmm3
3169         add     $48, %r8
3170         jle     L(StrncpyExit1)
3171         palignr $1, %xmm1, %xmm2
3172         movaps  %xmm3, %xmm1
3173         movaps  %xmm2, (%rdx)
3174         movaps  31(%rcx), %xmm2
3175         lea     16(%rsi), %rsi
3176         movaps  %xmm2, %xmm3
3177         sub     $16, %r8
3178         jbe     L(StrncpyExit1)
3179         palignr $1, %xmm1, %xmm2
3180         movaps  %xmm2, 16(%rdx)
3181         movaps  31+16(%rcx), %xmm2
3182         movaps  %xmm3, %xmm1
3183         lea     16(%rsi), %rsi
3184         sub     $16, %r8
3185         jbe     L(StrncpyExit1)
3186         movaps  %xmm2, %xmm1
3187         movaps  %xmm4, 32(%rdx)
3188         lea     16(%rsi), %rsi
3189         sub     $16, %r8
3190         jbe     L(StrncpyExit1)
3191         movaps  %xmm7, %xmm1
3192         movaps  %xmm5, 48(%rdx)
3193         lea     16(%rsi), %rsi
3194         lea     -16(%r8), %r8
3195
3196 L(StrncpyExit1):
3197         movaps  (%rdx, %rsi), %xmm6
3198         psrldq  $15, %xmm6
3199         palignr $1, %xmm1, %xmm6
3200         movaps  %xmm6, (%rdx, %rsi)
3201         lea     15(%rsi), %rsi
3202         jmp     L(CopyFrom1To16BytesCase3)
3203
3204 L(StrncpyLeave2):
3205         movaps  %xmm2, %xmm3
3206         add     $48, %r8
3207         jle     L(StrncpyExit2)
3208         palignr $2, %xmm1, %xmm2
3209         movaps  %xmm3, %xmm1
3210         movaps  %xmm2, (%rdx)
3211         movaps  30(%rcx), %xmm2
3212         lea     16(%rsi), %rsi
3213         movaps  %xmm2, %xmm3
3214         sub     $16, %r8
3215         jbe     L(StrncpyExit2)
3216         palignr $2, %xmm1, %xmm2
3217         movaps  %xmm2, 16(%rdx)
3218         movaps  30+16(%rcx), %xmm2
3219         movaps  %xmm3, %xmm1
3220         lea     16(%rsi), %rsi
3221         sub     $16, %r8
3222         jbe     L(StrncpyExit2)
3223         movaps  %xmm2, %xmm1
3224         movaps  %xmm4, 32(%rdx)
3225         lea     16(%rsi), %rsi
3226         sub     $16, %r8
3227         jbe     L(StrncpyExit2)
3228         movaps  %xmm7, %xmm1
3229         movaps  %xmm5, 48(%rdx)
3230         lea     16(%rsi), %rsi
3231         lea     -16(%r8), %r8
3232
3233 L(StrncpyExit2):
3234         movaps  (%rdx, %rsi), %xmm6
3235         psrldq  $14, %xmm6
3236         palignr $2, %xmm1, %xmm6
3237         movaps  %xmm6, (%rdx, %rsi)
3238         lea     14(%rsi), %rsi
3239         jmp     L(CopyFrom1To16BytesCase3)
3240
3241 L(StrncpyLeave3):
3242         movaps  %xmm2, %xmm3
3243         add     $48, %r8
3244         jle     L(StrncpyExit3)
3245         palignr $3, %xmm1, %xmm2
3246         movaps  %xmm3, %xmm1
3247         movaps  %xmm2, (%rdx)
3248         movaps  29(%rcx), %xmm2
3249         lea     16(%rsi), %rsi
3250         movaps  %xmm2, %xmm3
3251         sub     $16, %r8
3252         jbe     L(StrncpyExit3)
3253         palignr $3, %xmm1, %xmm2
3254         movaps  %xmm2, 16(%rdx)
3255         movaps  29+16(%rcx), %xmm2
3256         movaps  %xmm3, %xmm1
3257         lea     16(%rsi), %rsi
3258         sub     $16, %r8
3259         jbe     L(StrncpyExit3)
3260         movaps  %xmm2, %xmm1
3261         movaps  %xmm4, 32(%rdx)
3262         lea     16(%rsi), %rsi
3263         sub     $16, %r8
3264         jbe     L(StrncpyExit3)
3265         movaps  %xmm7, %xmm1
3266         movaps  %xmm5, 48(%rdx)
3267         lea     16(%rsi), %rsi
3268         lea     -16(%r8), %r8
3269
3270 L(StrncpyExit3):
3271         movaps  (%rdx, %rsi), %xmm6
3272         psrldq  $13, %xmm6
3273         palignr $3, %xmm1, %xmm6
3274         movaps  %xmm6, (%rdx, %rsi)
3275         lea     13(%rsi), %rsi
3276         jmp     L(CopyFrom1To16BytesCase3)
3277
3278 L(StrncpyLeave4):
3279         movaps  %xmm2, %xmm3
3280         add     $48, %r8
3281         jle     L(StrncpyExit4)
3282         palignr $4, %xmm1, %xmm2
3283         movaps  %xmm3, %xmm1
3284         movaps  %xmm2, (%rdx)
3285         movaps  28(%rcx), %xmm2
3286         lea     16(%rsi), %rsi
3287         movaps  %xmm2, %xmm3
3288         sub     $16, %r8
3289         jbe     L(StrncpyExit4)
3290         palignr $4, %xmm1, %xmm2
3291         movaps  %xmm2, 16(%rdx)
3292         movaps  28+16(%rcx), %xmm2
3293         movaps  %xmm3, %xmm1
3294         lea     16(%rsi), %rsi
3295         sub     $16, %r8
3296         jbe     L(StrncpyExit4)
3297         movaps  %xmm2, %xmm1
3298         movaps  %xmm4, 32(%rdx)
3299         lea     16(%rsi), %rsi
3300         sub     $16, %r8
3301         jbe     L(StrncpyExit4)
3302         movaps  %xmm7, %xmm1
3303         movaps  %xmm5, 48(%rdx)
3304         lea     16(%rsi), %rsi
3305         lea     -16(%r8), %r8
3306
3307 L(StrncpyExit4):
3308         movaps  (%rdx, %rsi), %xmm6
3309         psrldq  $12, %xmm6
3310         palignr $4, %xmm1, %xmm6
3311         movaps  %xmm6, (%rdx, %rsi)
3312         lea     12(%rsi), %rsi
3313         jmp     L(CopyFrom1To16BytesCase3)
3314
3315 L(StrncpyLeave5):
3316         movaps  %xmm2, %xmm3
3317         add     $48, %r8
3318         jle     L(StrncpyExit5)
3319         palignr $5, %xmm1, %xmm2
3320         movaps  %xmm3, %xmm1
3321         movaps  %xmm2, (%rdx)
3322         movaps  27(%rcx), %xmm2
3323         lea     16(%rsi), %rsi
3324         movaps  %xmm2, %xmm3
3325         sub     $16, %r8
3326         jbe     L(StrncpyExit5)
3327         palignr $5, %xmm1, %xmm2
3328         movaps  %xmm2, 16(%rdx)
3329         movaps  27+16(%rcx), %xmm2
3330         movaps  %xmm3, %xmm1
3331         lea     16(%rsi), %rsi
3332         sub     $16, %r8
3333         jbe     L(StrncpyExit5)
3334         movaps  %xmm2, %xmm1
3335         movaps  %xmm4, 32(%rdx)
3336         lea     16(%rsi), %rsi
3337         sub     $16, %r8
3338         jbe     L(StrncpyExit5)
3339         movaps  %xmm7, %xmm1
3340         movaps  %xmm5, 48(%rdx)
3341         lea     16(%rsi), %rsi
3342         lea     -16(%r8), %r8
3343
3344 L(StrncpyExit5):
3345         movaps  (%rdx, %rsi), %xmm6
3346         psrldq  $11, %xmm6
3347         palignr $5, %xmm1, %xmm6
3348         movaps  %xmm6, (%rdx, %rsi)
3349         lea     11(%rsi), %rsi
3350         jmp     L(CopyFrom1To16BytesCase3)
3351
3352 L(StrncpyLeave6):
3353         movaps  %xmm2, %xmm3
3354         add     $48, %r8
3355         jle     L(StrncpyExit6)
3356         palignr $6, %xmm1, %xmm2
3357         movaps  %xmm3, %xmm1
3358         movaps  %xmm2, (%rdx)
3359         movaps  26(%rcx), %xmm2
3360         lea     16(%rsi), %rsi
3361         movaps  %xmm2, %xmm3
3362         sub     $16, %r8
3363         jbe     L(StrncpyExit6)
3364         palignr $6, %xmm1, %xmm2
3365         movaps  %xmm2, 16(%rdx)
3366         movaps  26+16(%rcx), %xmm2
3367         movaps  %xmm3, %xmm1
3368         lea     16(%rsi), %rsi
3369         sub     $16, %r8
3370         jbe     L(StrncpyExit6)
3371         movaps  %xmm2, %xmm1
3372         movaps  %xmm4, 32(%rdx)
3373         lea     16(%rsi), %rsi
3374         sub     $16, %r8
3375         jbe     L(StrncpyExit6)
3376         movaps  %xmm7, %xmm1
3377         movaps  %xmm5, 48(%rdx)
3378         lea     16(%rsi), %rsi
3379         lea     -16(%r8), %r8
3380
3381 L(StrncpyExit6):
3382         movaps  (%rdx, %rsi), %xmm6
3383         psrldq  $10, %xmm6
3384         palignr $6, %xmm1, %xmm6
3385         movaps  %xmm6, (%rdx, %rsi)
3386         lea     10(%rsi), %rsi
3387         jmp     L(CopyFrom1To16BytesCase3)
3388
3389 L(StrncpyLeave7):
3390         movaps  %xmm2, %xmm3
3391         add     $48, %r8
3392         jle     L(StrncpyExit7)
3393         palignr $7, %xmm1, %xmm2
3394         movaps  %xmm3, %xmm1
3395         movaps  %xmm2, (%rdx)
3396         movaps  25(%rcx), %xmm2
3397         lea     16(%rsi), %rsi
3398         movaps  %xmm2, %xmm3
3399         sub     $16, %r8
3400         jbe     L(StrncpyExit7)
3401         palignr $7, %xmm1, %xmm2
3402         movaps  %xmm2, 16(%rdx)
3403         movaps  25+16(%rcx), %xmm2
3404         movaps  %xmm3, %xmm1
3405         lea     16(%rsi), %rsi
3406         sub     $16, %r8
3407         jbe     L(StrncpyExit7)
3408         movaps  %xmm2, %xmm1
3409         movaps  %xmm4, 32(%rdx)
3410         lea     16(%rsi), %rsi
3411         sub     $16, %r8
3412         jbe     L(StrncpyExit7)
3413         movaps  %xmm7, %xmm1
3414         movaps  %xmm5, 48(%rdx)
3415         lea     16(%rsi), %rsi
3416         lea     -16(%r8), %r8
3417
3418 L(StrncpyExit7):
3419         movaps  (%rdx, %rsi), %xmm6
3420         psrldq  $9, %xmm6
3421         palignr $7, %xmm1, %xmm6
3422         movaps  %xmm6, (%rdx, %rsi)
3423         lea     9(%rsi), %rsi
3424         jmp     L(CopyFrom1To16BytesCase3)
3425
3426 L(StrncpyLeave8):
3427         movaps  %xmm2, %xmm3
3428         add     $48, %r8
3429         jle     L(StrncpyExit8)
3430         palignr $8, %xmm1, %xmm2
3431         movaps  %xmm3, %xmm1
3432         movaps  %xmm2, (%rdx)
3433         movaps  24(%rcx), %xmm2
3434         lea     16(%rsi), %rsi
3435         movaps  %xmm2, %xmm3
3436         sub     $16, %r8
3437         jbe     L(StrncpyExit8)
3438         palignr $8, %xmm1, %xmm2
3439         movaps  %xmm2, 16(%rdx)
3440         movaps  24+16(%rcx), %xmm2
3441         movaps  %xmm3, %xmm1
3442         lea     16(%rsi), %rsi
3443         sub     $16, %r8
3444         jbe     L(StrncpyExit8)
3445         movaps  %xmm2, %xmm1
3446         movaps  %xmm4, 32(%rdx)
3447         lea     16(%rsi), %rsi
3448         sub     $16, %r8
3449         jbe     L(StrncpyExit8)
3450         movaps  %xmm7, %xmm1
3451         movaps  %xmm5, 48(%rdx)
3452         lea     16(%rsi), %rsi
3453         lea     -16(%r8), %r8
3454
3455 L(StrncpyExit8):
3456         movaps  (%rdx, %rsi), %xmm6
3457         psrldq  $8, %xmm6
3458         palignr $8, %xmm1, %xmm6
3459         movaps  %xmm6, (%rdx, %rsi)
3460         lea     8(%rsi), %rsi
3461         jmp     L(CopyFrom1To16BytesCase3)
3462
3463 L(StrncpyLeave9):
3464         movaps  %xmm2, %xmm3
3465         add     $48, %r8
3466         jle     L(StrncpyExit9)
3467         palignr $9, %xmm1, %xmm2
3468         movaps  %xmm3, %xmm1
3469         movaps  %xmm2, (%rdx)
3470         movaps  23(%rcx), %xmm2
3471         lea     16(%rsi), %rsi
3472         movaps  %xmm2, %xmm3
3473         sub     $16, %r8
3474         jbe     L(StrncpyExit9)
3475         palignr $9, %xmm1, %xmm2
3476         movaps  %xmm2, 16(%rdx)
3477         movaps  23+16(%rcx), %xmm2
3478         movaps  %xmm3, %xmm1
3479         lea     16(%rsi), %rsi
3480         sub     $16, %r8
3481         jbe     L(StrncpyExit9)
3482         movaps  %xmm2, %xmm1
3483         movaps  %xmm4, 32(%rdx)
3484         lea     16(%rsi), %rsi
3485         sub     $16, %r8
3486         jbe     L(StrncpyExit9)
3487         movaps  %xmm7, %xmm1
3488         movaps  %xmm5, 48(%rdx)
3489         lea     16(%rsi), %rsi
3490         lea     -16(%r8), %r8
3491
3492 L(StrncpyExit9):
3493         movaps  (%rdx, %rsi), %xmm6
3494         psrldq  $7, %xmm6
3495         palignr $9, %xmm1, %xmm6
3496         movaps  %xmm6, (%rdx, %rsi)
3497         lea     7(%rsi), %rsi
3498         jmp     L(CopyFrom1To16BytesCase3)
3499
3500 L(StrncpyLeave10):
3501         movaps  %xmm2, %xmm3
3502         add     $48, %r8
3503         jle     L(StrncpyExit10)
3504         palignr $10, %xmm1, %xmm2
3505         movaps  %xmm3, %xmm1
3506         movaps  %xmm2, (%rdx)
3507         movaps  22(%rcx), %xmm2
3508         lea     16(%rsi), %rsi
3509         movaps  %xmm2, %xmm3
3510         sub     $16, %r8
3511         jbe     L(StrncpyExit10)
3512         palignr $10, %xmm1, %xmm2
3513         movaps  %xmm2, 16(%rdx)
3514         movaps  22+16(%rcx), %xmm2
3515         movaps  %xmm3, %xmm1
3516         lea     16(%rsi), %rsi
3517         sub     $16, %r8
3518         jbe     L(StrncpyExit10)
3519         movaps  %xmm2, %xmm1
3520         movaps  %xmm4, 32(%rdx)
3521         lea     16(%rsi), %rsi
3522         sub     $16, %r8
3523         jbe     L(StrncpyExit10)
3524         movaps  %xmm7, %xmm1
3525         movaps  %xmm5, 48(%rdx)
3526         lea     16(%rsi), %rsi
3527         lea     -16(%r8), %r8
3528
3529 L(StrncpyExit10):
3530         movaps  (%rdx, %rsi), %xmm6
3531         psrldq  $6, %xmm6
3532         palignr $10, %xmm1, %xmm6
3533         movaps  %xmm6, (%rdx, %rsi)
3534         lea     6(%rsi), %rsi
3535         jmp     L(CopyFrom1To16BytesCase3)
3536
3537 L(StrncpyLeave11):
3538         movaps  %xmm2, %xmm3
3539         add     $48, %r8
3540         jle     L(StrncpyExit11)
3541         palignr $11, %xmm1, %xmm2
3542         movaps  %xmm3, %xmm1
3543         movaps  %xmm2, (%rdx)
3544         movaps  21(%rcx), %xmm2
3545         lea     16(%rsi), %rsi
3546         movaps  %xmm2, %xmm3
3547         sub     $16, %r8
3548         jbe     L(StrncpyExit11)
3549         palignr $11, %xmm1, %xmm2
3550         movaps  %xmm2, 16(%rdx)
3551         movaps  21+16(%rcx), %xmm2
3552         movaps  %xmm3, %xmm1
3553         lea     16(%rsi), %rsi
3554         sub     $16, %r8
3555         jbe     L(StrncpyExit11)
3556         movaps  %xmm2, %xmm1
3557         movaps  %xmm4, 32(%rdx)
3558         lea     16(%rsi), %rsi
3559         sub     $16, %r8
3560         jbe     L(StrncpyExit11)
3561         movaps  %xmm7, %xmm1
3562         movaps  %xmm5, 48(%rdx)
3563         lea     16(%rsi), %rsi
3564         lea     -16(%r8), %r8
3565
3566 L(StrncpyExit11):
3567         movaps  (%rdx, %rsi), %xmm6
3568         psrldq  $5, %xmm6
3569         palignr $11, %xmm1, %xmm6
3570         movaps  %xmm6, (%rdx, %rsi)
3571         lea     5(%rsi), %rsi
3572         jmp     L(CopyFrom1To16BytesCase3)
3573
3574 L(StrncpyLeave12):
3575         movaps  %xmm2, %xmm3
3576         add     $48, %r8
3577         jle     L(StrncpyExit12)
3578         palignr $12, %xmm1, %xmm2
3579         movaps  %xmm3, %xmm1
3580         movaps  %xmm2, (%rdx)
3581         movaps  20(%rcx), %xmm2
3582         lea     16(%rsi), %rsi
3583         movaps  %xmm2, %xmm3
3584         sub     $16, %r8
3585         jbe     L(StrncpyExit12)
3586         palignr $12, %xmm1, %xmm2
3587         movaps  %xmm2, 16(%rdx)
3588         movaps  20+16(%rcx), %xmm2
3589         movaps  %xmm3, %xmm1
3590         lea     16(%rsi), %rsi
3591         sub     $16, %r8
3592         jbe     L(StrncpyExit12)
3593         movaps  %xmm2, %xmm1
3594         movaps  %xmm4, 32(%rdx)
3595         lea     16(%rsi), %rsi
3596         sub     $16, %r8
3597         jbe     L(StrncpyExit12)
3598         movaps  %xmm7, %xmm1
3599         movaps  %xmm5, 48(%rdx)
3600         lea     16(%rsi), %rsi
3601         lea     -16(%r8), %r8
3602
3603 L(StrncpyExit12):
3604         movaps  (%rdx, %rsi), %xmm6
3605         psrldq  $4, %xmm6
3606         palignr $12, %xmm1, %xmm6
3607         movaps  %xmm6, (%rdx, %rsi)
3608         lea     4(%rsi), %rsi
3609         jmp     L(CopyFrom1To16BytesCase3)
3610
3611 L(StrncpyLeave13):
3612         movaps  %xmm2, %xmm3
3613         add     $48, %r8
3614         jle     L(StrncpyExit13)
3615         palignr $13, %xmm1, %xmm2
3616         movaps  %xmm3, %xmm1
3617         movaps  %xmm2, (%rdx)
3618         movaps  19(%rcx), %xmm2
3619         lea     16(%rsi), %rsi
3620         movaps  %xmm2, %xmm3
3621         sub     $16, %r8
3622         jbe     L(StrncpyExit13)
3623         palignr $13, %xmm1, %xmm2
3624         movaps  %xmm2, 16(%rdx)
3625         movaps  19+16(%rcx), %xmm2
3626         movaps  %xmm3, %xmm1
3627         lea     16(%rsi), %rsi
3628         sub     $16, %r8
3629         jbe     L(StrncpyExit13)
3630         movaps  %xmm2, %xmm1
3631         movaps  %xmm4, 32(%rdx)
3632         lea     16(%rsi), %rsi
3633         sub     $16, %r8
3634         jbe     L(StrncpyExit13)
3635         movaps  %xmm7, %xmm1
3636         movaps  %xmm5, 48(%rdx)
3637         lea     16(%rsi), %rsi
3638         lea     -16(%r8), %r8
3639
3640 L(StrncpyExit13):
3641         movaps  (%rdx, %rsi), %xmm6
3642         psrldq  $3, %xmm6
3643         palignr $13, %xmm1, %xmm6
3644         movaps  %xmm6, (%rdx, %rsi)
3645         lea     3(%rsi), %rsi
3646         jmp     L(CopyFrom1To16BytesCase3)
3647
3648 L(StrncpyLeave14):
3649         movaps  %xmm2, %xmm3
3650         add     $48, %r8
3651         jle     L(StrncpyExit14)
3652         palignr $14, %xmm1, %xmm2
3653         movaps  %xmm3, %xmm1
3654         movaps  %xmm2, (%rdx)
3655         movaps  18(%rcx), %xmm2
3656         lea     16(%rsi), %rsi
3657         movaps  %xmm2, %xmm3
3658         sub     $16, %r8
3659         jbe     L(StrncpyExit14)
3660         palignr $14, %xmm1, %xmm2
3661         movaps  %xmm2, 16(%rdx)
3662         movaps  18+16(%rcx), %xmm2
3663         movaps  %xmm3, %xmm1
3664         lea     16(%rsi), %rsi
3665         sub     $16, %r8
3666         jbe     L(StrncpyExit14)
3667         movaps  %xmm2, %xmm1
3668         movaps  %xmm4, 32(%rdx)
3669         lea     16(%rsi), %rsi
3670         sub     $16, %r8
3671         jbe     L(StrncpyExit14)
3672         movaps  %xmm7, %xmm1
3673         movaps  %xmm5, 48(%rdx)
3674         lea     16(%rsi), %rsi
3675         lea     -16(%r8), %r8
3676
3677 L(StrncpyExit14):
3678         movaps  (%rdx, %rsi), %xmm6
3679         psrldq  $2, %xmm6
3680         palignr $14, %xmm1, %xmm6
3681         movaps  %xmm6, (%rdx, %rsi)
3682         lea     2(%rsi), %rsi
3683         jmp     L(CopyFrom1To16BytesCase3)
3684
3685 L(StrncpyLeave15):
3686         movaps  %xmm2, %xmm3
3687         add     $48, %r8
3688         jle     L(StrncpyExit15)
3689         palignr $15, %xmm1, %xmm2
3690         movaps  %xmm3, %xmm1
3691         movaps  %xmm2, (%rdx)
3692         movaps  17(%rcx), %xmm2
3693         lea     16(%rsi), %rsi
3694         movaps  %xmm2, %xmm3
3695         sub     $16, %r8
3696         jbe     L(StrncpyExit15)
3697         palignr $15, %xmm1, %xmm2
3698         movaps  %xmm2, 16(%rdx)
3699         movaps  17+16(%rcx), %xmm2
3700         movaps  %xmm3, %xmm1
3701         lea     16(%rsi), %rsi
3702         sub     $16, %r8
3703         jbe     L(StrncpyExit15)
3704         movaps  %xmm2, %xmm1
3705         movaps  %xmm4, 32(%rdx)
3706         lea     16(%rsi), %rsi
3707         sub     $16, %r8
3708         jbe     L(StrncpyExit15)
3709         movaps  %xmm7, %xmm1
3710         movaps  %xmm5, 48(%rdx)
3711         lea     16(%rsi), %rsi
3712         lea     -16(%r8), %r8
3713
3714 L(StrncpyExit15):
3715         movaps  (%rdx, %rsi), %xmm6
3716         psrldq  $1, %xmm6
3717         palignr $15, %xmm1, %xmm6
3718         movaps  %xmm6, (%rdx, %rsi)
3719         lea     1(%rsi), %rsi
3720         jmp     L(CopyFrom1To16BytesCase3)
3721 # endif
3722 # ifndef USE_AS_STRCAT
3723 END (STRCPY)
3724 # endif
3725 #endif