secure/lib/libcrypto/aarch64/keccak1600-armv8.S

   1 /* $FreeBSD$ */
   2 /* Do not modify. This file is auto-generated from keccak1600-armv8.pl. */
   3 .text
   4
   5 .align  8       // strategic alignment and padding that allows to use
   6                 // address value as loop termination condition...
   7 .quad   0,0,0,0,0,0,0,0
   8 .type   iotas,%object
   9 iotas:
  10 .quad   0x0000000000000001
  11 .quad   0x0000000000008082
  12 .quad   0x800000000000808a
  13 .quad   0x8000000080008000
  14 .quad   0x000000000000808b
  15 .quad   0x0000000080000001
  16 .quad   0x8000000080008081
  17 .quad   0x8000000000008009
  18 .quad   0x000000000000008a
  19 .quad   0x0000000000000088
  20 .quad   0x0000000080008009
  21 .quad   0x000000008000000a
  22 .quad   0x000000008000808b
  23 .quad   0x800000000000008b
  24 .quad   0x8000000000008089
  25 .quad   0x8000000000008003
  26 .quad   0x8000000000008002
  27 .quad   0x8000000000000080
  28 .quad   0x000000000000800a
  29 .quad   0x800000008000000a
  30 .quad   0x8000000080008081
  31 .quad   0x8000000000008080
  32 .quad   0x0000000080000001
  33 .quad   0x8000000080008008
  34 .size   iotas,.-iotas
  35 .type   KeccakF1600_int,%function
  36 .align  5
  37 KeccakF1600_int:
  38         adr     x28,iotas
  39         stp     x28,x30,[sp,#16]                // 32 bytes on top are mine
  40         b       .Loop
  41 .align  4
  42 .Loop:
  43         ////////////////////////////////////////// Theta
  44         eor     x26,x0,x5
  45         stp     x4,x9,[sp,#0]   // offload pair...
  46         eor     x27,x1,x6
  47         eor     x28,x2,x7
  48         eor     x30,x3,x8
  49         eor     x4,x4,x9
  50         eor     x26,x26,x10
  51         eor     x27,x27,x11
  52         eor     x28,x28,x12
  53         eor     x30,x30,x13
  54         eor     x4,x4,x14
  55         eor     x26,x26,x15
  56         eor     x27,x27,x16
  57         eor     x28,x28,x17
  58         eor     x30,x30,x25
  59         eor     x4,x4,x19
  60         eor     x26,x26,x20
  61         eor     x28,x28,x22
  62         eor     x27,x27,x21
  63         eor     x30,x30,x23
  64         eor     x4,x4,x24
  65
  66         eor     x9,x26,x28,ror#63
  67
  68         eor     x1,x1,x9
  69         eor     x6,x6,x9
  70         eor     x11,x11,x9
  71         eor     x16,x16,x9
  72         eor     x21,x21,x9
  73
  74         eor     x9,x27,x30,ror#63
  75         eor     x28,x28,x4,ror#63
  76         eor     x30,x30,x26,ror#63
  77         eor     x4,x4,x27,ror#63
  78
  79         eor     x27,   x2,x9            // mov  x27,x2
  80         eor     x7,x7,x9
  81         eor     x12,x12,x9
  82         eor     x17,x17,x9
  83         eor     x22,x22,x9
  84
  85         eor     x0,x0,x4
  86         eor     x5,x5,x4
  87         eor     x10,x10,x4
  88         eor     x15,x15,x4
  89         eor     x20,x20,x4
  90         ldp     x4,x9,[sp,#0]   // re-load offloaded data
  91         eor     x26,   x3,x28           // mov  x26,x3
  92         eor     x8,x8,x28
  93         eor     x13,x13,x28
  94         eor     x25,x25,x28
  95         eor     x23,x23,x28
  96
  97         eor     x28,   x4,x30           // mov  x28,x4
  98         eor     x9,x9,x30
  99         eor     x14,x14,x30
 100         eor     x19,x19,x30
 101         eor     x24,x24,x30
 102
 103         ////////////////////////////////////////// Rho+Pi
 104         mov     x30,x1
 105         ror     x1,x6,#64-44
 106         //mov   x27,x2
 107         ror     x2,x12,#64-43
 108         //mov   x26,x3
 109         ror     x3,x25,#64-21
 110         //mov   x28,x4
 111         ror     x4,x24,#64-14
 112
 113         ror     x6,x9,#64-20
 114         ror     x12,x13,#64-25
 115         ror     x25,x17,#64-15
 116         ror     x24,x21,#64-2
 117
 118         ror     x9,x22,#64-61
 119         ror     x13,x19,#64-8
 120         ror     x17,x11,#64-10
 121         ror     x21,x8,#64-55
 122
 123         ror     x22,x14,#64-39
 124         ror     x19,x23,#64-56
 125         ror     x11,x7,#64-6
 126         ror     x8,x16,#64-45
 127
 128         ror     x14,x20,#64-18
 129         ror     x23,x15,#64-41
 130         ror     x7,x10,#64-3
 131         ror     x16,x5,#64-36
 132
 133         ror     x5,x26,#64-28
 134         ror     x10,x30,#64-1
 135         ror     x15,x28,#64-27
 136         ror     x20,x27,#64-62
 137
 138         ////////////////////////////////////////// Chi+Iota
 139         bic     x26,x2,x1
 140         bic     x27,x3,x2
 141         bic     x28,x0,x4
 142         bic     x30,x1,x0
 143         eor     x0,x0,x26
 144         bic     x26,x4,x3
 145         eor     x1,x1,x27
 146         ldr     x27,[sp,#16]
 147         eor     x3,x3,x28
 148         eor     x4,x4,x30
 149         eor     x2,x2,x26
 150         ldr     x30,[x27],#8            // Iota[i++]
 151
 152         bic     x26,x7,x6
 153         tst     x27,#255                        // are we done?
 154         str     x27,[sp,#16]
 155         bic     x27,x8,x7
 156         bic     x28,x5,x9
 157         eor     x0,x0,x30               // A[0][0] ^= Iota
 158         bic     x30,x6,x5
 159         eor     x5,x5,x26
 160         bic     x26,x9,x8
 161         eor     x6,x6,x27
 162         eor     x8,x8,x28
 163         eor     x9,x9,x30
 164         eor     x7,x7,x26
 165
 166         bic     x26,x12,x11
 167         bic     x27,x13,x12
 168         bic     x28,x10,x14
 169         bic     x30,x11,x10
 170         eor     x10,x10,x26
 171         bic     x26,x14,x13
 172         eor     x11,x11,x27
 173         eor     x13,x13,x28
 174         eor     x14,x14,x30
 175         eor     x12,x12,x26
 176
 177         bic     x26,x17,x16
 178         bic     x27,x25,x17
 179         bic     x28,x15,x19
 180         bic     x30,x16,x15
 181         eor     x15,x15,x26
 182         bic     x26,x19,x25
 183         eor     x16,x16,x27
 184         eor     x25,x25,x28
 185         eor     x19,x19,x30
 186         eor     x17,x17,x26
 187
 188         bic     x26,x22,x21
 189         bic     x27,x23,x22
 190         bic     x28,x20,x24
 191         bic     x30,x21,x20
 192         eor     x20,x20,x26
 193         bic     x26,x24,x23
 194         eor     x21,x21,x27
 195         eor     x23,x23,x28
 196         eor     x24,x24,x30
 197         eor     x22,x22,x26
 198
 199         bne     .Loop
 200
 201         ldr     x30,[sp,#24]
 202         ret
 203 .size   KeccakF1600_int,.-KeccakF1600_int
 204
 205 .type   KeccakF1600,%function
 206 .align  5
 207 KeccakF1600:
 208         stp     x29,x30,[sp,#-128]!
 209         add     x29,sp,#0
 210         stp     x19,x20,[sp,#16]
 211         stp     x21,x22,[sp,#32]
 212         stp     x23,x24,[sp,#48]
 213         stp     x25,x26,[sp,#64]
 214         stp     x27,x28,[sp,#80]
 215         sub     sp,sp,#48
 216
 217         str     x0,[sp,#32]                     // offload argument
 218         mov     x26,x0
 219         ldp     x0,x1,[x0,#16*0]
 220         ldp     x2,x3,[x26,#16*1]
 221         ldp     x4,x5,[x26,#16*2]
 222         ldp     x6,x7,[x26,#16*3]
 223         ldp     x8,x9,[x26,#16*4]
 224         ldp     x10,x11,[x26,#16*5]
 225         ldp     x12,x13,[x26,#16*6]
 226         ldp     x14,x15,[x26,#16*7]
 227         ldp     x16,x17,[x26,#16*8]
 228         ldp     x25,x19,[x26,#16*9]
 229         ldp     x20,x21,[x26,#16*10]
 230         ldp     x22,x23,[x26,#16*11]
 231         ldr     x24,[x26,#16*12]
 232
 233         bl      KeccakF1600_int
 234
 235         ldr     x26,[sp,#32]
 236         stp     x0,x1,[x26,#16*0]
 237         stp     x2,x3,[x26,#16*1]
 238         stp     x4,x5,[x26,#16*2]
 239         stp     x6,x7,[x26,#16*3]
 240         stp     x8,x9,[x26,#16*4]
 241         stp     x10,x11,[x26,#16*5]
 242         stp     x12,x13,[x26,#16*6]
 243         stp     x14,x15,[x26,#16*7]
 244         stp     x16,x17,[x26,#16*8]
 245         stp     x25,x19,[x26,#16*9]
 246         stp     x20,x21,[x26,#16*10]
 247         stp     x22,x23,[x26,#16*11]
 248         str     x24,[x26,#16*12]
 249
 250         ldp     x19,x20,[x29,#16]
 251         add     sp,sp,#48
 252         ldp     x21,x22,[x29,#32]
 253         ldp     x23,x24,[x29,#48]
 254         ldp     x25,x26,[x29,#64]
 255         ldp     x27,x28,[x29,#80]
 256         ldp     x29,x30,[sp],#128
 257         ret
 258 .size   KeccakF1600,.-KeccakF1600
 259
 260 .globl  SHA3_absorb
 261 .type   SHA3_absorb,%function
 262 .align  5
 263 SHA3_absorb:
 264         stp     x29,x30,[sp,#-128]!
 265         add     x29,sp,#0
 266         stp     x19,x20,[sp,#16]
 267         stp     x21,x22,[sp,#32]
 268         stp     x23,x24,[sp,#48]
 269         stp     x25,x26,[sp,#64]
 270         stp     x27,x28,[sp,#80]
 271         sub     sp,sp,#64
 272
 273         stp     x0,x1,[sp,#32]                  // offload arguments
 274         stp     x2,x3,[sp,#48]
 275
 276         mov     x26,x0                  // uint64_t A[5][5]
 277         mov     x27,x1                  // const void *inp
 278         mov     x28,x2                  // size_t len
 279         mov     x30,x3                  // size_t bsz
 280         ldp     x0,x1,[x26,#16*0]
 281         ldp     x2,x3,[x26,#16*1]
 282         ldp     x4,x5,[x26,#16*2]
 283         ldp     x6,x7,[x26,#16*3]
 284         ldp     x8,x9,[x26,#16*4]
 285         ldp     x10,x11,[x26,#16*5]
 286         ldp     x12,x13,[x26,#16*6]
 287         ldp     x14,x15,[x26,#16*7]
 288         ldp     x16,x17,[x26,#16*8]
 289         ldp     x25,x19,[x26,#16*9]
 290         ldp     x20,x21,[x26,#16*10]
 291         ldp     x22,x23,[x26,#16*11]
 292         ldr     x24,[x26,#16*12]
 293         b       .Loop_absorb
 294
 295 .align  4
 296 .Loop_absorb:
 297         subs    x26,x28,x30             // len - bsz
 298         blo     .Labsorbed
 299
 300         str     x26,[sp,#48]                    // save len - bsz
 301         ldr     x26,[x27],#8            // *inp++
 302 #ifdef  __AARCH64EB__
 303         rev     x26,x26
 304 #endif
 305         eor     x0,x0,x26
 306         cmp     x30,#8*(0+2)
 307         blo     .Lprocess_block
 308         ldr     x26,[x27],#8            // *inp++
 309 #ifdef  __AARCH64EB__
 310         rev     x26,x26
 311 #endif
 312         eor     x1,x1,x26
 313         beq     .Lprocess_block
 314         ldr     x26,[x27],#8            // *inp++
 315 #ifdef  __AARCH64EB__
 316         rev     x26,x26
 317 #endif
 318         eor     x2,x2,x26
 319         cmp     x30,#8*(2+2)
 320         blo     .Lprocess_block
 321         ldr     x26,[x27],#8            // *inp++
 322 #ifdef  __AARCH64EB__
 323         rev     x26,x26
 324 #endif
 325         eor     x3,x3,x26
 326         beq     .Lprocess_block
 327         ldr     x26,[x27],#8            // *inp++
 328 #ifdef  __AARCH64EB__
 329         rev     x26,x26
 330 #endif
 331         eor     x4,x4,x26
 332         cmp     x30,#8*(4+2)
 333         blo     .Lprocess_block
 334         ldr     x26,[x27],#8            // *inp++
 335 #ifdef  __AARCH64EB__
 336         rev     x26,x26
 337 #endif
 338         eor     x5,x5,x26
 339         beq     .Lprocess_block
 340         ldr     x26,[x27],#8            // *inp++
 341 #ifdef  __AARCH64EB__
 342         rev     x26,x26
 343 #endif
 344         eor     x6,x6,x26
 345         cmp     x30,#8*(6+2)
 346         blo     .Lprocess_block
 347         ldr     x26,[x27],#8            // *inp++
 348 #ifdef  __AARCH64EB__
 349         rev     x26,x26
 350 #endif
 351         eor     x7,x7,x26
 352         beq     .Lprocess_block
 353         ldr     x26,[x27],#8            // *inp++
 354 #ifdef  __AARCH64EB__
 355         rev     x26,x26
 356 #endif
 357         eor     x8,x8,x26
 358         cmp     x30,#8*(8+2)
 359         blo     .Lprocess_block
 360         ldr     x26,[x27],#8            // *inp++
 361 #ifdef  __AARCH64EB__
 362         rev     x26,x26
 363 #endif
 364         eor     x9,x9,x26
 365         beq     .Lprocess_block
 366         ldr     x26,[x27],#8            // *inp++
 367 #ifdef  __AARCH64EB__
 368         rev     x26,x26
 369 #endif
 370         eor     x10,x10,x26
 371         cmp     x30,#8*(10+2)
 372         blo     .Lprocess_block
 373         ldr     x26,[x27],#8            // *inp++
 374 #ifdef  __AARCH64EB__
 375         rev     x26,x26
 376 #endif
 377         eor     x11,x11,x26
 378         beq     .Lprocess_block
 379         ldr     x26,[x27],#8            // *inp++
 380 #ifdef  __AARCH64EB__
 381         rev     x26,x26
 382 #endif
 383         eor     x12,x12,x26
 384         cmp     x30,#8*(12+2)
 385         blo     .Lprocess_block
 386         ldr     x26,[x27],#8            // *inp++
 387 #ifdef  __AARCH64EB__
 388         rev     x26,x26
 389 #endif
 390         eor     x13,x13,x26
 391         beq     .Lprocess_block
 392         ldr     x26,[x27],#8            // *inp++
 393 #ifdef  __AARCH64EB__
 394         rev     x26,x26
 395 #endif
 396         eor     x14,x14,x26
 397         cmp     x30,#8*(14+2)
 398         blo     .Lprocess_block
 399         ldr     x26,[x27],#8            // *inp++
 400 #ifdef  __AARCH64EB__
 401         rev     x26,x26
 402 #endif
 403         eor     x15,x15,x26
 404         beq     .Lprocess_block
 405         ldr     x26,[x27],#8            // *inp++
 406 #ifdef  __AARCH64EB__
 407         rev     x26,x26
 408 #endif
 409         eor     x16,x16,x26
 410         cmp     x30,#8*(16+2)
 411         blo     .Lprocess_block
 412         ldr     x26,[x27],#8            // *inp++
 413 #ifdef  __AARCH64EB__
 414         rev     x26,x26
 415 #endif
 416         eor     x17,x17,x26
 417         beq     .Lprocess_block
 418         ldr     x26,[x27],#8            // *inp++
 419 #ifdef  __AARCH64EB__
 420         rev     x26,x26
 421 #endif
 422         eor     x25,x25,x26
 423         cmp     x30,#8*(18+2)
 424         blo     .Lprocess_block
 425         ldr     x26,[x27],#8            // *inp++
 426 #ifdef  __AARCH64EB__
 427         rev     x26,x26
 428 #endif
 429         eor     x19,x19,x26
 430         beq     .Lprocess_block
 431         ldr     x26,[x27],#8            // *inp++
 432 #ifdef  __AARCH64EB__
 433         rev     x26,x26
 434 #endif
 435         eor     x20,x20,x26
 436         cmp     x30,#8*(20+2)
 437         blo     .Lprocess_block
 438         ldr     x26,[x27],#8            // *inp++
 439 #ifdef  __AARCH64EB__
 440         rev     x26,x26
 441 #endif
 442         eor     x21,x21,x26
 443         beq     .Lprocess_block
 444         ldr     x26,[x27],#8            // *inp++
 445 #ifdef  __AARCH64EB__
 446         rev     x26,x26
 447 #endif
 448         eor     x22,x22,x26
 449         cmp     x30,#8*(22+2)
 450         blo     .Lprocess_block
 451         ldr     x26,[x27],#8            // *inp++
 452 #ifdef  __AARCH64EB__
 453         rev     x26,x26
 454 #endif
 455         eor     x23,x23,x26
 456         beq     .Lprocess_block
 457         ldr     x26,[x27],#8            // *inp++
 458 #ifdef  __AARCH64EB__
 459         rev     x26,x26
 460 #endif
 461         eor     x24,x24,x26
 462
 463 .Lprocess_block:
 464         str     x27,[sp,#40]                    // save inp
 465
 466         bl      KeccakF1600_int
 467
 468         ldr     x27,[sp,#40]                    // restore arguments
 469         ldp     x28,x30,[sp,#48]
 470         b       .Loop_absorb
 471
 472 .align  4
 473 .Labsorbed:
 474         ldr     x27,[sp,#32]
 475         stp     x0,x1,[x27,#16*0]
 476         stp     x2,x3,[x27,#16*1]
 477         stp     x4,x5,[x27,#16*2]
 478         stp     x6,x7,[x27,#16*3]
 479         stp     x8,x9,[x27,#16*4]
 480         stp     x10,x11,[x27,#16*5]
 481         stp     x12,x13,[x27,#16*6]
 482         stp     x14,x15,[x27,#16*7]
 483         stp     x16,x17,[x27,#16*8]
 484         stp     x25,x19,[x27,#16*9]
 485         stp     x20,x21,[x27,#16*10]
 486         stp     x22,x23,[x27,#16*11]
 487         str     x24,[x27,#16*12]
 488
 489         mov     x0,x28                  // return value
 490         ldp     x19,x20,[x29,#16]
 491         add     sp,sp,#64
 492         ldp     x21,x22,[x29,#32]
 493         ldp     x23,x24,[x29,#48]
 494         ldp     x25,x26,[x29,#64]
 495         ldp     x27,x28,[x29,#80]
 496         ldp     x29,x30,[sp],#128
 497         ret
 498 .size   SHA3_absorb,.-SHA3_absorb
 499 .globl  SHA3_squeeze
 500 .type   SHA3_squeeze,%function
 501 .align  5
 502 SHA3_squeeze:
 503         stp     x29,x30,[sp,#-48]!
 504         add     x29,sp,#0
 505         stp     x19,x20,[sp,#16]
 506         stp     x21,x22,[sp,#32]
 507
 508         mov     x19,x0                  // put aside arguments
 509         mov     x20,x1
 510         mov     x21,x2
 511         mov     x22,x3
 512
 513 .Loop_squeeze:
 514         ldr     x4,[x0],#8
 515         cmp     x21,#8
 516         blo     .Lsqueeze_tail
 517 #ifdef  __AARCH64EB__
 518         rev     x4,x4
 519 #endif
 520         str     x4,[x20],#8
 521         subs    x21,x21,#8
 522         beq     .Lsqueeze_done
 523
 524         subs    x3,x3,#8
 525         bhi     .Loop_squeeze
 526
 527         mov     x0,x19
 528         bl      KeccakF1600
 529         mov     x0,x19
 530         mov     x3,x22
 531         b       .Loop_squeeze
 532
 533 .align  4
 534 .Lsqueeze_tail:
 535         strb    w4,[x20],#1
 536         lsr     x4,x4,#8
 537         subs    x21,x21,#1
 538         beq     .Lsqueeze_done
 539         strb    w4,[x20],#1
 540         lsr     x4,x4,#8
 541         subs    x21,x21,#1
 542         beq     .Lsqueeze_done
 543         strb    w4,[x20],#1
 544         lsr     x4,x4,#8
 545         subs    x21,x21,#1
 546         beq     .Lsqueeze_done
 547         strb    w4,[x20],#1
 548         lsr     x4,x4,#8
 549         subs    x21,x21,#1
 550         beq     .Lsqueeze_done
 551         strb    w4,[x20],#1
 552         lsr     x4,x4,#8
 553         subs    x21,x21,#1
 554         beq     .Lsqueeze_done
 555         strb    w4,[x20],#1
 556         lsr     x4,x4,#8
 557         subs    x21,x21,#1
 558         beq     .Lsqueeze_done
 559         strb    w4,[x20],#1
 560
 561 .Lsqueeze_done:
 562         ldp     x19,x20,[sp,#16]
 563         ldp     x21,x22,[sp,#32]
 564         ldp     x29,x30,[sp],#48
 565         ret
 566 .size   SHA3_squeeze,.-SHA3_squeeze
 567 .type   KeccakF1600_ce,%function
 568 .align  5
 569 KeccakF1600_ce:
 570         mov     x9,#12
 571         adr     x10,iotas
 572         b       .Loop_ce
 573 .align  4
 574 .Loop_ce:
 575         ////////////////////////////////////////////////// Theta
 576 .inst   0xce052819      //eor3 v25.16b,v0.16b,v5.16b,v10.16b
 577 .inst   0xce062c3a      //eor3 v26.16b,v1.16b,v6.16b,v11.16b
 578 .inst   0xce07305b      //eor3 v27.16b,v2.16b,v7.16b,v12.16b
 579 .inst   0xce08347c      //eor3 v28.16b,v3.16b,v8.16b,v13.16b
 580 .inst   0xce09389d      //eor3 v29.16b,v4.16b,v9.16b,v14.16b
 581 .inst   0xce0f5339      //eor3 v25.16b,v25.16b,   v15.16b,v20.16b
 582 .inst   0xce10575a      //eor3 v26.16b,v26.16b,   v16.16b,v21.16b
 583 .inst   0xce115b7b      //eor3 v27.16b,v27.16b,   v17.16b,v22.16b
 584 .inst   0xce125f9c      //eor3 v28.16b,v28.16b,   v18.16b,v23.16b
 585 .inst   0xce1363bd      //eor3 v29.16b,v29.16b,   v19.16b,v24.16b
 586
 587 .inst   0xce7b8f3e      //rax1 v30.16b,v25.16b,v27.16b                  // D[1]
 588 .inst   0xce7c8f5f      //rax1 v31.16b,v26.16b,v28.16b                  // D[2]
 589 .inst   0xce7d8f7b      //rax1 v27.16b,v27.16b,v29.16b                  // D[3]
 590 .inst   0xce798f9c      //rax1 v28.16b,v28.16b,v25.16b                  // D[4]
 591 .inst   0xce7a8fbd      //rax1 v29.16b,v29.16b,v26.16b                  // D[0]
 592
 593         ////////////////////////////////////////////////// Theta+Rho+Pi
 594 .inst   0xce9e50d9      //xar v25.16b,   v6.16b,v30.16b,#64-44  // C[0]=A[0][1]
 595 .inst   0xce9cb126      //xar v6.16b,v9.16b,v28.16b,#64-20
 596 .inst   0xce9f0ec9      //xar v9.16b,v22.16b,v31.16b,#64-61
 597 .inst   0xce9c65d6      //xar v22.16b,v14.16b,v28.16b,#64-39
 598 .inst   0xce9dba8e      //xar v14.16b,v20.16b,v29.16b,#64-18
 599
 600 .inst   0xce9f0854      //xar v20.16b,v2.16b,v31.16b,#64-62
 601
 602 .inst   0xce9f5582      //xar v2.16b,v12.16b,v31.16b,#64-43
 603 .inst   0xce9b9dac      //xar v12.16b,v13.16b,v27.16b,#64-25
 604 .inst   0xce9ce26d      //xar v13.16b,v19.16b,v28.16b,#64-8
 605 .inst   0xce9b22f3      //xar v19.16b,v23.16b,v27.16b,#64-56
 606 .inst   0xce9d5df7      //xar v23.16b,v15.16b,v29.16b,#64-41
 607
 608 .inst   0xce9c948f      //xar v15.16b,v4.16b,v28.16b,#64-27
 609
 610         eor     v0.16b,v0.16b,v29.16b
 611         ldr     x11,[x10],#8
 612
 613 .inst   0xce9bae5a      //xar v26.16b,   v18.16b,v27.16b,#64-21 // C[1]=A[0][3]
 614 .inst   0xce9fc632      //xar v18.16b,v17.16b,v31.16b,#64-15
 615 .inst   0xce9ed971      //xar v17.16b,v11.16b,v30.16b,#64-10
 616 .inst   0xce9fe8eb      //xar v11.16b,v7.16b,v31.16b,#64-6
 617 .inst   0xce9df547      //xar v7.16b,v10.16b,v29.16b,#64-3
 618
 619 .inst   0xce9efc2a      //xar v10.16b,v1.16b,v30.16b,#64-1      // *
 620
 621 .inst   0xce9ccb04      //xar v4.16b,v24.16b,v28.16b,#64-14
 622 .inst   0xce9efab8      //xar v24.16b,v21.16b,v30.16b,#64-2
 623 .inst   0xce9b2515      //xar v21.16b,v8.16b,v27.16b,#64-55
 624 .inst   0xce9e4e08      //xar v8.16b,v16.16b,v30.16b,#64-45
 625 .inst   0xce9d70b0      //xar v16.16b,v5.16b,v29.16b,#64-36
 626
 627 .inst   0xce9b907b      //xar v27.16b,   v3.16b,v27.16b,#64-28  // C[2]=A[1][0]
 628
 629         ////////////////////////////////////////////////// Chi+Iota
 630         dup     v31.2d,x11                              // borrow C[6]
 631 .inst   0xce22641c      //bcax v28.16b,   v0.16b,v2.16b,v25.16b // *
 632 .inst   0xce3a0b21      //bcax v1.16b,v25.16b,   v26.16b,   v2.16b      // *
 633 .inst   0xce246842      //bcax v2.16b,v2.16b,v4.16b,v26.16b
 634 .inst   0xce201343      //bcax v3.16b,v26.16b,   v0.16b,v4.16b
 635 .inst   0xce390084      //bcax v4.16b,v4.16b,v25.16b,   v0.16b
 636
 637 .inst   0xce271b65      //bcax v5.16b,v27.16b,   v7.16b,v6.16b  // *
 638 .inst   0xce281cd9      //bcax v25.16b,   v6.16b,v8.16b,v7.16b  // *
 639 .inst   0xce2920e7      //bcax v7.16b,v7.16b,v9.16b,v8.16b
 640 .inst   0xce3b2508      //bcax v8.16b,v8.16b,v27.16b,   v9.16b
 641 .inst   0xce266d29      //bcax v9.16b,v9.16b,v6.16b,v27.16b
 642
 643         eor     v0.16b,v28.16b,v31.16b                  // Iota
 644
 645 .inst   0xce2c2d5a      //bcax v26.16b,   v10.16b,v12.16b,v11.16b       // *
 646 .inst   0xce2d317b      //bcax v27.16b,   v11.16b,v13.16b,v12.16b       // *
 647 .inst   0xce2e358c      //bcax v12.16b,v12.16b,v14.16b,v13.16b
 648 .inst   0xce2a39ad      //bcax v13.16b,v13.16b,v10.16b,v14.16b
 649 .inst   0xce2b29ce      //bcax v14.16b,v14.16b,v11.16b,v10.16b
 650
 651 .inst   0xce3141fc      //bcax v28.16b,   v15.16b,v17.16b,v16.16b       // *
 652 .inst   0xce32461d      //bcax v29.16b,   v16.16b,v18.16b,v17.16b       // *
 653 .inst   0xce334a31      //bcax v17.16b,v17.16b,v19.16b,v18.16b
 654 .inst   0xce2f4e52      //bcax v18.16b,v18.16b,v15.16b,v19.16b
 655 .inst   0xce303e73      //bcax v19.16b,v19.16b,v16.16b,v15.16b
 656
 657 .inst   0xce36569e      //bcax v30.16b,   v20.16b,v22.16b,v21.16b       // *
 658 .inst   0xce375abf      //bcax v31.16b,   v21.16b,v23.16b,v22.16b       // *
 659 .inst   0xce385ed6      //bcax v22.16b,v22.16b,v24.16b,v23.16b
 660 .inst   0xce3462f7      //bcax v23.16b,v23.16b,v20.16b,v24.16b
 661 .inst   0xce355318      //bcax v24.16b,v24.16b,v21.16b,v20.16b
 662         ////////////////////////////////////////////////// Theta
 663 .inst   0xce056806      //eor3 v6.16b,v0.16b,v5.16b,v26.16b
 664 .inst   0xce196c2a      //eor3 v10.16b,v1.16b,v25.16b,v27.16b
 665 .inst   0xce07304b      //eor3 v11.16b,v2.16b,v7.16b,v12.16b
 666 .inst   0xce08346f      //eor3 v15.16b,v3.16b,v8.16b,v13.16b
 667 .inst   0xce093890      //eor3 v16.16b,v4.16b,v9.16b,v14.16b
 668 .inst   0xce1c78c6      //eor3 v6.16b,v6.16b,   v28.16b,v30.16b
 669 .inst   0xce1d7d4a      //eor3 v10.16b,v10.16b,   v29.16b,v31.16b
 670 .inst   0xce11596b      //eor3 v11.16b,v11.16b,   v17.16b,v22.16b
 671 .inst   0xce125def      //eor3 v15.16b,v15.16b,   v18.16b,v23.16b
 672 .inst   0xce136210      //eor3 v16.16b,v16.16b,   v19.16b,v24.16b
 673
 674 .inst   0xce6b8cd4      //rax1 v20.16b,v6.16b,v11.16b                   // D[1]
 675 .inst   0xce6f8d55      //rax1 v21.16b,v10.16b,v15.16b                  // D[2]
 676 .inst   0xce708d6b      //rax1 v11.16b,v11.16b,v16.16b                  // D[3]
 677 .inst   0xce668def      //rax1 v15.16b,v15.16b,v6.16b                   // D[4]
 678 .inst   0xce6a8e10      //rax1 v16.16b,v16.16b,v10.16b                  // D[0]
 679
 680         ////////////////////////////////////////////////// Theta+Rho+Pi
 681 .inst   0xce945326      //xar v6.16b,   v25.16b,v20.16b,#64-44  // C[0]=A[0][1]
 682 .inst   0xce8fb139      //xar v25.16b,v9.16b,v15.16b,#64-20
 683 .inst   0xce950ec9      //xar v9.16b,v22.16b,v21.16b,#64-61
 684 .inst   0xce8f65d6      //xar v22.16b,v14.16b,v15.16b,#64-39
 685 .inst   0xce90bbce      //xar v14.16b,v30.16b,v16.16b,#64-18
 686
 687 .inst   0xce95085e      //xar v30.16b,v2.16b,v21.16b,#64-62
 688
 689 .inst   0xce955582      //xar v2.16b,v12.16b,v21.16b,#64-43
 690 .inst   0xce8b9dac      //xar v12.16b,v13.16b,v11.16b,#64-25
 691 .inst   0xce8fe26d      //xar v13.16b,v19.16b,v15.16b,#64-8
 692 .inst   0xce8b22f3      //xar v19.16b,v23.16b,v11.16b,#64-56
 693 .inst   0xce905f97      //xar v23.16b,v28.16b,v16.16b,#64-41
 694
 695 .inst   0xce8f949c      //xar v28.16b,v4.16b,v15.16b,#64-27
 696
 697         eor     v0.16b,v0.16b,v16.16b
 698         ldr     x11,[x10],#8
 699
 700 .inst   0xce8bae4a      //xar v10.16b,   v18.16b,v11.16b,#64-21 // C[1]=A[0][3]
 701 .inst   0xce95c632      //xar v18.16b,v17.16b,v21.16b,#64-15
 702 .inst   0xce94db71      //xar v17.16b,v27.16b,v20.16b,#64-10
 703 .inst   0xce95e8fb      //xar v27.16b,v7.16b,v21.16b,#64-6
 704 .inst   0xce90f747      //xar v7.16b,v26.16b,v16.16b,#64-3
 705
 706 .inst   0xce94fc3a      //xar v26.16b,v1.16b,v20.16b,#64-1      // *
 707
 708 .inst   0xce8fcb04      //xar v4.16b,v24.16b,v15.16b,#64-14
 709 .inst   0xce94fbf8      //xar v24.16b,v31.16b,v20.16b,#64-2
 710 .inst   0xce8b251f      //xar v31.16b,v8.16b,v11.16b,#64-55
 711 .inst   0xce944fa8      //xar v8.16b,v29.16b,v20.16b,#64-45
 712 .inst   0xce9070bd      //xar v29.16b,v5.16b,v16.16b,#64-36
 713
 714 .inst   0xce8b906b      //xar v11.16b,   v3.16b,v11.16b,#64-28  // C[2]=A[1][0]
 715
 716         ////////////////////////////////////////////////// Chi+Iota
 717         dup     v21.2d,x11                              // borrow C[6]
 718 .inst   0xce22180f      //bcax v15.16b,   v0.16b,v2.16b,v6.16b  // *
 719 .inst   0xce2a08c1      //bcax v1.16b,v6.16b,   v10.16b,   v2.16b       // *
 720 .inst   0xce242842      //bcax v2.16b,v2.16b,v4.16b,v10.16b
 721 .inst   0xce201143      //bcax v3.16b,v10.16b,   v0.16b,v4.16b
 722 .inst   0xce260084      //bcax v4.16b,v4.16b,v6.16b,   v0.16b
 723
 724 .inst   0xce276565      //bcax v5.16b,v11.16b,   v7.16b,v25.16b // *
 725 .inst   0xce281f26      //bcax v6.16b,   v25.16b,v8.16b,v7.16b  // *
 726 .inst   0xce2920e7      //bcax v7.16b,v7.16b,v9.16b,v8.16b
 727 .inst   0xce2b2508      //bcax v8.16b,v8.16b,v11.16b,   v9.16b
 728 .inst   0xce392d29      //bcax v9.16b,v9.16b,v25.16b,v11.16b
 729
 730         eor     v0.16b,v15.16b,v21.16b                  // Iota
 731
 732 .inst   0xce2c6f4a      //bcax v10.16b,   v26.16b,v12.16b,v27.16b       // *
 733 .inst   0xce2d336b      //bcax v11.16b,   v27.16b,v13.16b,v12.16b       // *
 734 .inst   0xce2e358c      //bcax v12.16b,v12.16b,v14.16b,v13.16b
 735 .inst   0xce3a39ad      //bcax v13.16b,v13.16b,v26.16b,v14.16b
 736 .inst   0xce3b69ce      //bcax v14.16b,v14.16b,v27.16b,v26.16b
 737
 738 .inst   0xce31778f      //bcax v15.16b,   v28.16b,v17.16b,v29.16b       // *
 739 .inst   0xce3247b0      //bcax v16.16b,   v29.16b,v18.16b,v17.16b       // *
 740 .inst   0xce334a31      //bcax v17.16b,v17.16b,v19.16b,v18.16b
 741 .inst   0xce3c4e52      //bcax v18.16b,v18.16b,v28.16b,v19.16b
 742 .inst   0xce3d7273      //bcax v19.16b,v19.16b,v29.16b,v28.16b
 743
 744 .inst   0xce367fd4      //bcax v20.16b,   v30.16b,v22.16b,v31.16b       // *
 745 .inst   0xce375bf5      //bcax v21.16b,   v31.16b,v23.16b,v22.16b       // *
 746 .inst   0xce385ed6      //bcax v22.16b,v22.16b,v24.16b,v23.16b
 747 .inst   0xce3e62f7      //bcax v23.16b,v23.16b,v30.16b,v24.16b
 748 .inst   0xce3f7b18      //bcax v24.16b,v24.16b,v31.16b,v30.16b
 749         subs    x9,x9,#1
 750         bne     .Loop_ce
 751
 752         ret
 753 .size   KeccakF1600_ce,.-KeccakF1600_ce
 754
 755 .type   KeccakF1600_cext,%function
 756 .align  5
 757 KeccakF1600_cext:
 758         stp     x29,x30,[sp,#-80]!
 759         add     x29,sp,#0
 760         stp     d8,d9,[sp,#16]          // per ABI requirement
 761         stp     d10,d11,[sp,#32]
 762         stp     d12,d13,[sp,#48]
 763         stp     d14,d15,[sp,#64]
 764         ldp     d0,d1,[x0,#8*0]
 765         ldp     d2,d3,[x0,#8*2]
 766         ldp     d4,d5,[x0,#8*4]
 767         ldp     d6,d7,[x0,#8*6]
 768         ldp     d8,d9,[x0,#8*8]
 769         ldp     d10,d11,[x0,#8*10]
 770         ldp     d12,d13,[x0,#8*12]
 771         ldp     d14,d15,[x0,#8*14]
 772         ldp     d16,d17,[x0,#8*16]
 773         ldp     d18,d19,[x0,#8*18]
 774         ldp     d20,d21,[x0,#8*20]
 775         ldp     d22,d23,[x0,#8*22]
 776         ldr     d24,[x0,#8*24]
 777         bl      KeccakF1600_ce
 778         ldr     x30,[sp,#8]
 779         stp     d0,d1,[x0,#8*0]
 780         stp     d2,d3,[x0,#8*2]
 781         stp     d4,d5,[x0,#8*4]
 782         stp     d6,d7,[x0,#8*6]
 783         stp     d8,d9,[x0,#8*8]
 784         stp     d10,d11,[x0,#8*10]
 785         stp     d12,d13,[x0,#8*12]
 786         stp     d14,d15,[x0,#8*14]
 787         stp     d16,d17,[x0,#8*16]
 788         stp     d18,d19,[x0,#8*18]
 789         stp     d20,d21,[x0,#8*20]
 790         stp     d22,d23,[x0,#8*22]
 791         str     d24,[x0,#8*24]
 792
 793         ldp     d8,d9,[sp,#16]
 794         ldp     d10,d11,[sp,#32]
 795         ldp     d12,d13,[sp,#48]
 796         ldp     d14,d15,[sp,#64]
 797         ldr     x29,[sp],#80
 798         ret
 799 .size   KeccakF1600_cext,.-KeccakF1600_cext
 800 .globl  SHA3_absorb_cext
 801 .type   SHA3_absorb_cext,%function
 802 .align  5
 803 SHA3_absorb_cext:
 804         stp     x29,x30,[sp,#-80]!
 805         add     x29,sp,#0
 806         stp     d8,d9,[sp,#16]          // per ABI requirement
 807         stp     d10,d11,[sp,#32]
 808         stp     d12,d13,[sp,#48]
 809         stp     d14,d15,[sp,#64]
 810         ldp     d0,d1,[x0,#8*0]
 811         ldp     d2,d3,[x0,#8*2]
 812         ldp     d4,d5,[x0,#8*4]
 813         ldp     d6,d7,[x0,#8*6]
 814         ldp     d8,d9,[x0,#8*8]
 815         ldp     d10,d11,[x0,#8*10]
 816         ldp     d12,d13,[x0,#8*12]
 817         ldp     d14,d15,[x0,#8*14]
 818         ldp     d16,d17,[x0,#8*16]
 819         ldp     d18,d19,[x0,#8*18]
 820         ldp     d20,d21,[x0,#8*20]
 821         ldp     d22,d23,[x0,#8*22]
 822         ldr     d24,[x0,#8*24]
 823         b       .Loop_absorb_ce
 824
 825 .align  4
 826 .Loop_absorb_ce:
 827         subs    x2,x2,x3                // len - bsz
 828         blo     .Labsorbed_ce
 829         ldr     d31,[x1],#8             // *inp++
 830 #ifdef  __AARCH64EB__
 831         rev64   v31.16b,v31.16b
 832 #endif
 833         eor     v0.16b,v0.16b,v31.16b
 834         cmp     x3,#8*(0+2)
 835         blo     .Lprocess_block_ce
 836         ldr     d31,[x1],#8             // *inp++
 837 #ifdef  __AARCH64EB__
 838         rev     v31.16b,v31.16b
 839 #endif
 840         eor     v1.16b,v1.16b,v31.16b
 841         beq     .Lprocess_block_ce
 842         ldr     d31,[x1],#8             // *inp++
 843 #ifdef  __AARCH64EB__
 844         rev64   v31.16b,v31.16b
 845 #endif
 846         eor     v2.16b,v2.16b,v31.16b
 847         cmp     x3,#8*(2+2)
 848         blo     .Lprocess_block_ce
 849         ldr     d31,[x1],#8             // *inp++
 850 #ifdef  __AARCH64EB__
 851         rev     v31.16b,v31.16b
 852 #endif
 853         eor     v3.16b,v3.16b,v31.16b
 854         beq     .Lprocess_block_ce
 855         ldr     d31,[x1],#8             // *inp++
 856 #ifdef  __AARCH64EB__
 857         rev64   v31.16b,v31.16b
 858 #endif
 859         eor     v4.16b,v4.16b,v31.16b
 860         cmp     x3,#8*(4+2)
 861         blo     .Lprocess_block_ce
 862         ldr     d31,[x1],#8             // *inp++
 863 #ifdef  __AARCH64EB__
 864         rev     v31.16b,v31.16b
 865 #endif
 866         eor     v5.16b,v5.16b,v31.16b
 867         beq     .Lprocess_block_ce
 868         ldr     d31,[x1],#8             // *inp++
 869 #ifdef  __AARCH64EB__
 870         rev64   v31.16b,v31.16b
 871 #endif
 872         eor     v6.16b,v6.16b,v31.16b
 873         cmp     x3,#8*(6+2)
 874         blo     .Lprocess_block_ce
 875         ldr     d31,[x1],#8             // *inp++
 876 #ifdef  __AARCH64EB__
 877         rev     v31.16b,v31.16b
 878 #endif
 879         eor     v7.16b,v7.16b,v31.16b
 880         beq     .Lprocess_block_ce
 881         ldr     d31,[x1],#8             // *inp++
 882 #ifdef  __AARCH64EB__
 883         rev64   v31.16b,v31.16b
 884 #endif
 885         eor     v8.16b,v8.16b,v31.16b
 886         cmp     x3,#8*(8+2)
 887         blo     .Lprocess_block_ce
 888         ldr     d31,[x1],#8             // *inp++
 889 #ifdef  __AARCH64EB__
 890         rev     v31.16b,v31.16b
 891 #endif
 892         eor     v9.16b,v9.16b,v31.16b
 893         beq     .Lprocess_block_ce
 894         ldr     d31,[x1],#8             // *inp++
 895 #ifdef  __AARCH64EB__
 896         rev64   v31.16b,v31.16b
 897 #endif
 898         eor     v10.16b,v10.16b,v31.16b
 899         cmp     x3,#8*(10+2)
 900         blo     .Lprocess_block_ce
 901         ldr     d31,[x1],#8             // *inp++
 902 #ifdef  __AARCH64EB__
 903         rev     v31.16b,v31.16b
 904 #endif
 905         eor     v11.16b,v11.16b,v31.16b
 906         beq     .Lprocess_block_ce
 907         ldr     d31,[x1],#8             // *inp++
 908 #ifdef  __AARCH64EB__
 909         rev64   v31.16b,v31.16b
 910 #endif
 911         eor     v12.16b,v12.16b,v31.16b
 912         cmp     x3,#8*(12+2)
 913         blo     .Lprocess_block_ce
 914         ldr     d31,[x1],#8             // *inp++
 915 #ifdef  __AARCH64EB__
 916         rev     v31.16b,v31.16b
 917 #endif
 918         eor     v13.16b,v13.16b,v31.16b
 919         beq     .Lprocess_block_ce
 920         ldr     d31,[x1],#8             // *inp++
 921 #ifdef  __AARCH64EB__
 922         rev64   v31.16b,v31.16b
 923 #endif
 924         eor     v14.16b,v14.16b,v31.16b
 925         cmp     x3,#8*(14+2)
 926         blo     .Lprocess_block_ce
 927         ldr     d31,[x1],#8             // *inp++
 928 #ifdef  __AARCH64EB__
 929         rev     v31.16b,v31.16b
 930 #endif
 931         eor     v15.16b,v15.16b,v31.16b
 932         beq     .Lprocess_block_ce
 933         ldr     d31,[x1],#8             // *inp++
 934 #ifdef  __AARCH64EB__
 935         rev64   v31.16b,v31.16b
 936 #endif
 937         eor     v16.16b,v16.16b,v31.16b
 938         cmp     x3,#8*(16+2)
 939         blo     .Lprocess_block_ce
 940         ldr     d31,[x1],#8             // *inp++
 941 #ifdef  __AARCH64EB__
 942         rev     v31.16b,v31.16b
 943 #endif
 944         eor     v17.16b,v17.16b,v31.16b
 945         beq     .Lprocess_block_ce
 946         ldr     d31,[x1],#8             // *inp++
 947 #ifdef  __AARCH64EB__
 948         rev64   v31.16b,v31.16b
 949 #endif
 950         eor     v18.16b,v18.16b,v31.16b
 951         cmp     x3,#8*(18+2)
 952         blo     .Lprocess_block_ce
 953         ldr     d31,[x1],#8             // *inp++
 954 #ifdef  __AARCH64EB__
 955         rev     v31.16b,v31.16b
 956 #endif
 957         eor     v19.16b,v19.16b,v31.16b
 958         beq     .Lprocess_block_ce
 959         ldr     d31,[x1],#8             // *inp++
 960 #ifdef  __AARCH64EB__
 961         rev64   v31.16b,v31.16b
 962 #endif
 963         eor     v20.16b,v20.16b,v31.16b
 964         cmp     x3,#8*(20+2)
 965         blo     .Lprocess_block_ce
 966         ldr     d31,[x1],#8             // *inp++
 967 #ifdef  __AARCH64EB__
 968         rev     v31.16b,v31.16b
 969 #endif
 970         eor     v21.16b,v21.16b,v31.16b
 971         beq     .Lprocess_block_ce
 972         ldr     d31,[x1],#8             // *inp++
 973 #ifdef  __AARCH64EB__
 974         rev64   v31.16b,v31.16b
 975 #endif
 976         eor     v22.16b,v22.16b,v31.16b
 977         cmp     x3,#8*(22+2)
 978         blo     .Lprocess_block_ce
 979         ldr     d31,[x1],#8             // *inp++
 980 #ifdef  __AARCH64EB__
 981         rev     v31.16b,v31.16b
 982 #endif
 983         eor     v23.16b,v23.16b,v31.16b
 984         beq     .Lprocess_block_ce
 985         ldr     d31,[x1],#8             // *inp++
 986 #ifdef  __AARCH64EB__
 987         rev     v31.16b,v31.16b
 988 #endif
 989         eor     v24.16b,v24.16b,v31.16b
 990
 991 .Lprocess_block_ce:
 992
 993         bl      KeccakF1600_ce
 994
 995         b       .Loop_absorb_ce
 996
 997 .align  4
 998 .Labsorbed_ce:
 999         stp     d0,d1,[x0,#8*0]
1000         stp     d2,d3,[x0,#8*2]
1001         stp     d4,d5,[x0,#8*4]
1002         stp     d6,d7,[x0,#8*6]
1003         stp     d8,d9,[x0,#8*8]
1004         stp     d10,d11,[x0,#8*10]
1005         stp     d12,d13,[x0,#8*12]
1006         stp     d14,d15,[x0,#8*14]
1007         stp     d16,d17,[x0,#8*16]
1008         stp     d18,d19,[x0,#8*18]
1009         stp     d20,d21,[x0,#8*20]
1010         stp     d22,d23,[x0,#8*22]
1011         str     d24,[x0,#8*24]
1012         add     x0,x2,x3                // return value
1013
1014         ldp     d8,d9,[sp,#16]
1015         ldp     d10,d11,[sp,#32]
1016         ldp     d12,d13,[sp,#48]
1017         ldp     d14,d15,[sp,#64]
1018         ldp     x29,x30,[sp],#80
1019         ret
1020 .size   SHA3_absorb_cext,.-SHA3_absorb_cext
1021 .globl  SHA3_squeeze_cext
1022 .type   SHA3_squeeze_cext,%function
1023 .align  5
1024 SHA3_squeeze_cext:
1025         stp     x29,x30,[sp,#-16]!
1026         add     x29,sp,#0
1027         mov     x9,x0
1028         mov     x10,x3
1029
1030 .Loop_squeeze_ce:
1031         ldr     x4,[x9],#8
1032         cmp     x2,#8
1033         blo     .Lsqueeze_tail_ce
1034 #ifdef  __AARCH64EB__
1035         rev     x4,x4
1036 #endif
1037         str     x4,[x1],#8
1038         beq     .Lsqueeze_done_ce
1039
1040         sub     x2,x2,#8
1041         subs    x10,x10,#8
1042         bhi     .Loop_squeeze_ce
1043
1044         bl      KeccakF1600_cext
1045         ldr     x30,[sp,#8]
1046         mov     x9,x0
1047         mov     x10,x3
1048         b       .Loop_squeeze_ce
1049
1050 .align  4
1051 .Lsqueeze_tail_ce:
1052         strb    w4,[x1],#1
1053         lsr     x4,x4,#8
1054         subs    x2,x2,#1
1055         beq     .Lsqueeze_done_ce
1056         strb    w4,[x1],#1
1057         lsr     x4,x4,#8
1058         subs    x2,x2,#1
1059         beq     .Lsqueeze_done_ce
1060         strb    w4,[x1],#1
1061         lsr     x4,x4,#8
1062         subs    x2,x2,#1
1063         beq     .Lsqueeze_done_ce
1064         strb    w4,[x1],#1
1065         lsr     x4,x4,#8
1066         subs    x2,x2,#1
1067         beq     .Lsqueeze_done_ce
1068         strb    w4,[x1],#1
1069         lsr     x4,x4,#8
1070         subs    x2,x2,#1
1071         beq     .Lsqueeze_done_ce
1072         strb    w4,[x1],#1
1073         lsr     x4,x4,#8
1074         subs    x2,x2,#1
1075         beq     .Lsqueeze_done_ce
1076         strb    w4,[x1],#1
1077
1078 .Lsqueeze_done_ce:
1079         ldr     x29,[sp],#16
1080         ret
1081 .size   SHA3_squeeze_cext,.-SHA3_squeeze_cext
1082 .byte   75,101,99,99,97,107,45,49,54,48,48,32,97,98,115,111,114,98,32,97,110,100,32,115,113,117,101,101,122,101,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
1083 .align  2