string/memmem.s

   1         .text
   2         .p2align 4,,15
   3         .type   two_way_long_needle, @function
   4 two_way_long_needle:
   5         pushq   %r15
   6         pushq   %r14
   7         movq    %rdi, %r15
   8         pushq   %r13
   9         pushq   %r12
  10         movq    %rsi, %r14
  11         pushq   %rbp
  12         pushq   %rbx
  13         movq    %rdx, %rbp
  14         movq    %rcx, %rbx
  15         movl    $1, %edi
  16         movl    $1, %edx
  17         subq    $2088, %rsp
  18         xorl    %ecx, %ecx
  19         movq    $-1, %rsi
  20 .L2:
  21         leaq    (%rdx,%rcx), %rax
  22         cmpq    %rax, %rbx
  23         jbe     .L71
  24 .L7:
  25         leaq    0(%rbp,%rsi), %r8
  26         movzbl  (%r8,%rdx), %r11d
  27         cmpb    %r11b, 0(%rbp,%rax)
  28         jnb     .L3
  29         movq    %rax, %rcx
  30         movl    $1, %edx
  31         movq    %rax, %rdi
  32         leaq    (%rdx,%rcx), %rax
  33         subq    %rsi, %rdi
  34         cmpq    %rax, %rbx
  35         ja      .L7
  36 .L71:
  37         movl    $1, %r8d
  38         movl    $1, %edx
  39         xorl    %ecx, %ecx
  40         movq    $-1, %r13
  41 .L8:
  42         leaq    (%rdx,%rcx), %rax
  43         cmpq    %rax, %rbx
  44         jbe     .L72
  45 .L13:
  46         leaq    0(%rbp,%r13), %r9
  47         movzbl  (%r9,%rdx), %r10d
  48         cmpb    %r10b, 0(%rbp,%rax)
  49         jbe     .L9
  50         movq    %rax, %rcx
  51         movl    $1, %edx
  52         movq    %rax, %r8
  53         leaq    (%rdx,%rcx), %rax
  54         subq    %r13, %r8
  55         cmpq    %rax, %rbx
  56         ja      .L13
  57 .L72:
  58         addq    $1, %r13
  59         addq    $1, %rsi
  60         cmpq    %rsi, %r13
  61         jnb     .L14
  62         movq    %rdi, %r8
  63         movq    %rsi, %r13
  64 .L14:
  65         leaq    32(%rsp), %rax
  66         leaq    2080(%rsp), %rdx
  67         .p2align 4,,10
  68         .p2align 3
  69 .L15:
  70         movq    %rbx, (%rax)
  71         addq    $8, %rax
  72         cmpq    %rax, %rdx
  73         jne     .L15
  74         testq   %rbx, %rbx
  75         leaq    -1(%rbx), %r12
  76         je      .L16
  77         movq    %rbp, %rdx
  78         movq    %r12, %rax
  79         .p2align 4,,10
  80         .p2align 3
  81 .L17:
  82         movzbl  (%rdx), %ecx
  83         addq    $1, %rdx
  84         movq    %rax, 32(%rsp,%rcx,8)
  85         subq    $1, %rax
  86         cmpq    $-1, %rax
  87         jne     .L17
  88 .L16:
  89         leaq    0(%rbp,%r8), %rsi
  90         movq    %r13, %rdx
  91         movq    %rbp, %rdi
  92         movq    %r8, 8(%rsp)
  93         call    memcmp@PLT
  94         testl   %eax, %eax
  95         jne     .L18
  96         leaq    -1(%r13), %r11
  97         movq    8(%rsp), %r8
  98         movq    %rbx, %r10
  99         xorl    %esi, %esi
 100         xorl    %edx, %edx
 101         subq    %rbx, %r14
 102         leaq    0(%rbp,%r11), %rax
 103         subq    %r8, %r10
 104         movq    %rax, 24(%rsp)
 105         movl    $1, %eax
 106         subq    %r13, %rax
 107         movq    %rax, 8(%rsp)
 108         jmp     .L29
 109         .p2align 4,,10
 110         .p2align 3
 111 .L73:
 112         cmpq    %r8, %rcx
 113         jnb     .L67
 114         testq   %rsi, %rsi
 115         cmovne  %r10, %rcx
 116 .L67:
 117         addq    %rcx, %rdx
 118         xorl    %esi, %esi
 119         cmpq    %rdx, %r14
 120         jb      .L69
 121 .L29:
 122         leaq    (%r15,%rdx), %rax
 123         movzbl  (%rax,%r12), %ecx
 124         movq    32(%rsp,%rcx,8), %rcx
 125         testq   %rcx, %rcx
 126         jne     .L73
 127         cmpq    %rsi, %r13
 128         movq    %rsi, %rcx
 129         cmovnb  %r13, %rcx
 130         leaq    (%rcx,%rdx), %rdi
 131         leaq    0(%rbp,%rcx), %r9
 132         addq    %r15, %rdi
 133         cmpq    %rcx, %r12
 134         jbe     .L22
 135         movzbl  (%r9), %ebx
 136         cmpb    %bl, (%rdi)
 137         je      .L24
 138         jmp     .L23
 139         .p2align 4,,10
 140         .p2align 3
 141 .L74:
 142         movzbl  (%rax,%rcx), %edi
 143         cmpb    %dil, 0(%rbp,%rcx)
 144         jne     .L23
 145 .L24:
 146         addq    $1, %rcx
 147         cmpq    %rcx, %r12
 148         ja      .L74
 149 .L22:
 150         leaq    (%r11,%rdx), %r9
 151         movq    %r11, %rdi
 152         addq    %r15, %r9
 153         cmpq    %rsi, %r13
 154         jbe     .L40
 155         movq    24(%rsp), %rbx
 156         movzbl  (%rbx), %ebx
 157         cmpb    %bl, (%r9)
 158         jne     .L40
 159         leaq    -1(%rsi), %rbx
 160         subq    %r13, %r9
 161         movq    %rax, 16(%rsp)
 162         jmp     .L26
 163         .p2align 4,,10
 164         .p2align 3
 165 .L75:
 166         movzbl  1(%rcx,%r9), %eax
 167         cmpb    %al, 0(%rbp,%rcx)
 168         jne     .L65
 169         movq    %rcx, %rdi
 170 .L26:
 171         leaq    -1(%rdi), %rcx
 172         cmpq    %rcx, %rbx
 173         jne     .L75
 174 .L65:
 175         movq    16(%rsp), %rax
 176 .L25:
 177         addq    $1, %rsi
 178         cmpq    %rdi, %rsi
 179         ja      .L1
 180         addq    %r8, %rdx
 181         movq    %r10, %rsi
 182         cmpq    %rdx, %r14
 183         jnb     .L29
 184 .L69:
 185         xorl    %eax, %eax
 186 .L1:
 187         addq    $2088, %rsp
 188         popq    %rbx
 189         popq    %rbp
 190         popq    %r12
 191         popq    %r13
 192         popq    %r14
 193         popq    %r15
 194         ret
 195         .p2align 4,,10
 196         .p2align 3
 197 .L3:
 198         je      .L76
 199         movq    %rcx, %rsi
 200         movl    $1, %edi
 201         addq    $1, %rcx
 202         movl    $1, %edx
 203         jmp     .L2
 204         .p2align 4,,10
 205         .p2align 3
 206 .L9:
 207         je      .L77
 208         movq    %rcx, %r13
 209         movl    $1, %r8d
 210         addq    $1, %rcx
 211         movl    $1, %edx
 212         jmp     .L8
 213         .p2align 4,,10
 214         .p2align 3
 215 .L23:
 216         addq    8(%rsp), %rdx
 217         jmp     .L67
 218 .L77:
 219         cmpq    %r8, %rdx
 220         je      .L12
 221         addq    $1, %rdx
 222         jmp     .L8
 223 .L76:
 224         cmpq    %rdi, %rdx
 225         je      .L6
 226         addq    $1, %rdx
 227         jmp     .L2
 228 .L18:
 229         movq    %rbx, %rax
 230         movl    $1, %edi
 231         leaq    0(%rbp,%r13), %r8
 232         subq    %r13, %rax
 233         cmpq    %r13, %rax
 234         cmovb   %r13, %rax
 235         xorl    %ecx, %ecx
 236         subq    %r13, %rdi
 237         leaq    1(%rax), %rsi
 238         subq    %rbx, %r14
 239         .p2align 4,,10
 240         .p2align 3
 241 .L38:
 242         leaq    (%r15,%rcx), %rax
 243         movzbl  (%rax,%r12), %edx
 244         movq    32(%rsp,%rdx,8), %rdx
 245         testq   %rdx, %rdx
 246         jne     .L68
 247         leaq    (%rcx,%r13), %rdx
 248         addq    %r15, %rdx
 249         cmpq    %r13, %r12
 250         jbe     .L32
 251         movzbl  (%rdx), %ebx
 252         movq    %r13, %rdx
 253         cmpb    %bl, (%r8)
 254         je      .L34
 255         jmp     .L33
 256         .p2align 4,,10
 257         .p2align 3
 258 .L78:
 259         movzbl  (%rax,%rdx), %ebx
 260         cmpb    %bl, 0(%rbp,%rdx)
 261         jne     .L33
 262 .L34:
 263         addq    $1, %rdx
 264         cmpq    %rdx, %r12
 265         ja      .L78
 266 .L32:
 267         leaq    -1(%r13), %rdx
 268         leaq    (%rcx,%rdx), %r9
 269         leaq    0(%rbp,%rdx), %r10
 270         addq    %r15, %r9
 271         cmpq    $-1, %rdx
 272         je      .L1
 273         movzbl  (%r10), %ebx
 274         cmpb    %bl, (%r9)
 275         jne     .L36
 276         subq    %r13, %r9
 277         jmp     .L37
 278         .p2align 4,,10
 279         .p2align 3
 280 .L79:
 281         movzbl  1(%rdx,%r9), %ebx
 282         cmpb    %bl, 0(%rbp,%rdx)
 283         jne     .L36
 284 .L37:
 285         subq    $1, %rdx
 286         cmpq    $-1, %rdx
 287         jne     .L79
 288         jmp     .L1
 289         .p2align 4,,10
 290         .p2align 3
 291 .L33:
 292         addq    %rdi, %rcx
 293 .L68:
 294         addq    %rdx, %rcx
 295 .L31:
 296         cmpq    %r14, %rcx
 297         jbe     .L38
 298         xorl    %eax, %eax
 299         jmp     .L1
 300 .L36:
 301         addq    %rsi, %rcx
 302         jmp     .L31
 303 .L12:
 304         movq    %rax, %rcx
 305         movl    $1, %edx
 306         jmp     .L8
 307 .L6:
 308         movq    %rax, %rcx
 309         movl    $1, %edx
 310         jmp     .L2
 311 .L40:
 312         movq    %r13, %rdi
 313         jmp     .L25
 314         .size   two_way_long_needle, .-two_way_long_needle
 315         .p2align 4,,15
 316         .globl  __memmem
 317         .hidden __memmem
 318         .type   __memmem, @function
 319 __memmem:
 320         pushq   %r15
 321         pushq   %r14
 322         pushq   %r13
 323         pushq   %r12
 324         pushq   %rbp
 325         pushq   %rbx
 326         movq    %rdi, %rbx
 327         subq    $280, %rsp
 328         testq   %rcx, %rcx
 329         je      .L100
 330         cmpq    $1, %rcx
 331         movq    %rdx, %r15
 332         movq    %rcx, %rbp
 333         je      .L123
 334         cmpq    %rsi, %rcx
 335         ja      .L101
 336         movq    %rsi, %r14
 337         subq    %rcx, %r14
 338         addq    %rdi, %r14
 339         cmpq    $2, %rcx
 340         je      .L124
 341         cmpq    $256, %rcx
 342         ja      .L125
 343         leaq    -1(%rcx), %r13
 344         leaq    16(%rsp), %rdx
 345         xorl    %eax, %eax
 346         movl    $32, %ecx
 347         movq    %rdx, %rdi
 348         cmpq    $1, %r13
 349         rep stosq
 350         jbe     .L90
 351         movl    $1, %eax
 352         .p2align 4,,10
 353         .p2align 3
 354 .L91:
 355         movzbl  -1(%r15,%rax), %ecx
 356         movzbl  (%r15,%rax), %edx
 357         salq    $3, %rcx
 358         subq    %rcx, %rdx
 359         movzbl  %dl, %edx
 360         movb    %al, 16(%rsp,%rdx)
 361         addq    $1, %rax
 362         cmpq    %rax, %r13
 363         jne     .L91
 364 .L90:
 365         movzbl  -2(%r15,%rbp), %edx
 366         movzbl  -1(%r15,%rbp), %eax
 367         movq    %r13, %rdi
 368         xorl    %r12d, %r12d
 369         subq    $9, %rbp
 370         salq    $3, %rdx
 371         subq    %rdx, %rax
 372         movzbl  %al, %eax
 373         movzbl  16(%rsp,%rax), %edx
 374         movb    %r13b, 16(%rsp,%rax)
 375         subq    %rdx, %rdi
 376         movq    %rdi, 8(%rsp)
 377 .L95:
 378         cmpq    %rbx, %r14
 379         jnb     .L93
 380         jmp     .L101
 381         .p2align 4,,10
 382         .p2align 3
 383 .L126:
 384         cmpq    %rbx, %r14
 385         jb      .L103
 386 .L93:
 387         addq    %r13, %rbx
 388         movzbl  -1(%rbx), %edx
 389         movzbl  (%rbx), %eax
 390         salq    $3, %rdx
 391         subq    %rdx, %rax
 392         movzbl  %al, %eax
 393         movzbl  16(%rsp,%rax), %eax
 394         testq   %rax, %rax
 395         je      .L126
 396 .L103:
 397         subq    %rax, %rbx
 398         cmpq    %rax, %r13
 399         ja      .L95
 400         cmpq    $14, %r13
 401         jbe     .L96
 402         movq    (%r15,%r12), %rax
 403         cmpq    %rax, (%rbx,%r12)
 404         jne     .L97
 405 .L96:
 406         movq    %r13, %rdx
 407         movq    %r15, %rsi
 408         movq    %rbx, %rdi
 409         call    memcmp@PLT
 410         testl   %eax, %eax
 411         je      .L100
 412         leaq    -8(%r12), %rax
 413         cmpq    $8, %r12
 414         cmovb   %rbp, %rax
 415         movq    %rax, %r12
 416 .L97:
 417         addq    8(%rsp), %rbx
 418         cmpq    %rbx, %r14
 419         jnb     .L93
 420 .L101:
 421         xorl    %eax, %eax
 422         jmp     .L80
 423         .p2align 4,,10
 424         .p2align 3
 425 .L100:
 426         movq    %rbx, %rax
 427 .L80:
 428         addq    $280, %rsp
 429         popq    %rbx
 430         popq    %rbp
 431         popq    %r12
 432         popq    %r13
 433         popq    %r14
 434         popq    %r15
 435         ret
 436         .p2align 4,,10
 437         .p2align 3
 438 .L123:
 439         movzbl  (%rdx), %eax
 440         addq    $280, %rsp
 441         movq    %rsi, %rdx
 442         popq    %rbx
 443         popq    %rbp
 444         popq    %r12
 445         popq    %r13
 446         popq    %r14
 447         popq    %r15
 448         movl    %eax, %esi
 449         jmp     memchr
 450         .p2align 4,,10
 451         .p2align 3
 452 .L124:
 453         movzbl  (%rdx), %edx
 454         movzbl  1(%r15), %eax
 455         addq    $1, %rbx
 456         movzbl  1(%rdi), %ecx
 457         sall    $16, %edx
 458         orl     %eax, %edx
 459         movzbl  (%rdi), %eax
 460         sall    $16, %eax
 461         orl     %ecx, %eax
 462         cmpq    %rbx, %r14
 463         jnb     .L122
 464         jmp     .L84
 465         .p2align 4,,10
 466         .p2align 3
 467 .L127:
 468         addq    $1, %rbx
 469         movzbl  (%rbx), %ecx
 470         sall    $16, %eax
 471         orl     %ecx, %eax
 472         cmpq    %rbx, %r14
 473         jb      .L84
 474 .L122:
 475         cmpl    %eax, %edx
 476         jne     .L127
 477 .L84:
 478         cmpl    %eax, %edx
 479         jne     .L101
 480         leaq    -1(%rbx), %rax
 481         jmp     .L80
 482 .L125:
 483         addq    $280, %rsp
 484         popq    %rbx
 485         popq    %rbp
 486         popq    %r12
 487         popq    %r13
 488         popq    %r14
 489         popq    %r15
 490         jmp     two_way_long_needle
 491         .size   __memmem, .-__memmem
 492         .weak   memmem
 493         .hidden memmem
 494         .set    memmem,__memmem
 495         .hidden memchr