keccak1600-armv8.S revision 1.3
1 #include "arm_asm.h" 2 .text 3 4 .align 8 // strategic alignment and padding that allows to use 5 // address value as loop termination condition... 6 .quad 0,0,0,0,0,0,0,0 7 .type iotas,%object 8 iotas: 9 .quad 0x0000000000000001 10 .quad 0x0000000000008082 11 .quad 0x800000000000808a 12 .quad 0x8000000080008000 13 .quad 0x000000000000808b 14 .quad 0x0000000080000001 15 .quad 0x8000000080008081 16 .quad 0x8000000000008009 17 .quad 0x000000000000008a 18 .quad 0x0000000000000088 19 .quad 0x0000000080008009 20 .quad 0x000000008000000a 21 .quad 0x000000008000808b 22 .quad 0x800000000000008b 23 .quad 0x8000000000008089 24 .quad 0x8000000000008003 25 .quad 0x8000000000008002 26 .quad 0x8000000000000080 27 .quad 0x000000000000800a 28 .quad 0x800000008000000a 29 .quad 0x8000000080008081 30 .quad 0x8000000000008080 31 .quad 0x0000000080000001 32 .quad 0x8000000080008008 33 .size iotas,.-iotas 34 .type KeccakF1600_int,%function 35 .align 5 36 KeccakF1600_int: 37 adr x28,iotas 38 .inst 0xd503233f // paciasp 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 .inst 0xd50323bf // autiasp 203 ret 204 .size KeccakF1600_int,.-KeccakF1600_int 205 206 .type KeccakF1600,%function 207 .align 5 208 KeccakF1600: 209 .inst 0xd503233f // paciasp 210 stp x29,x30,[sp,#-128]! 211 add x29,sp,#0 212 stp x19,x20,[sp,#16] 213 stp x21,x22,[sp,#32] 214 stp x23,x24,[sp,#48] 215 stp x25,x26,[sp,#64] 216 stp x27,x28,[sp,#80] 217 sub sp,sp,#48 218 219 str x0,[sp,#32] // offload argument 220 mov x26,x0 221 ldp x0,x1,[x0,#16*0] 222 ldp x2,x3,[x26,#16*1] 223 ldp x4,x5,[x26,#16*2] 224 ldp x6,x7,[x26,#16*3] 225 ldp x8,x9,[x26,#16*4] 226 ldp x10,x11,[x26,#16*5] 227 ldp x12,x13,[x26,#16*6] 228 ldp x14,x15,[x26,#16*7] 229 ldp x16,x17,[x26,#16*8] 230 ldp x25,x19,[x26,#16*9] 231 ldp x20,x21,[x26,#16*10] 232 ldp x22,x23,[x26,#16*11] 233 ldr x24,[x26,#16*12] 234 235 bl KeccakF1600_int 236 237 ldr x26,[sp,#32] 238 stp x0,x1,[x26,#16*0] 239 stp x2,x3,[x26,#16*1] 240 stp x4,x5,[x26,#16*2] 241 stp x6,x7,[x26,#16*3] 242 stp x8,x9,[x26,#16*4] 243 stp x10,x11,[x26,#16*5] 244 stp x12,x13,[x26,#16*6] 245 stp x14,x15,[x26,#16*7] 246 stp x16,x17,[x26,#16*8] 247 stp x25,x19,[x26,#16*9] 248 stp x20,x21,[x26,#16*10] 249 stp x22,x23,[x26,#16*11] 250 str x24,[x26,#16*12] 251 252 ldp x19,x20,[x29,#16] 253 add sp,sp,#48 254 ldp x21,x22,[x29,#32] 255 ldp x23,x24,[x29,#48] 256 ldp x25,x26,[x29,#64] 257 ldp x27,x28,[x29,#80] 258 ldp x29,x30,[sp],#128 259 .inst 0xd50323bf // autiasp 260 ret 261 .size KeccakF1600,.-KeccakF1600 262 263 .globl SHA3_absorb 264 .type SHA3_absorb,%function 265 .align 5 266 SHA3_absorb: 267 .inst 0xd503233f // paciasp 268 stp x29,x30,[sp,#-128]! 269 add x29,sp,#0 270 stp x19,x20,[sp,#16] 271 stp x21,x22,[sp,#32] 272 stp x23,x24,[sp,#48] 273 stp x25,x26,[sp,#64] 274 stp x27,x28,[sp,#80] 275 sub sp,sp,#64 276 277 stp x0,x1,[sp,#32] // offload arguments 278 stp x2,x3,[sp,#48] 279 280 mov x26,x0 // uint64_t A[5][5] 281 mov x27,x1 // const void *inp 282 mov x28,x2 // size_t len 283 mov x30,x3 // size_t bsz 284 ldp x0,x1,[x26,#16*0] 285 ldp x2,x3,[x26,#16*1] 286 ldp x4,x5,[x26,#16*2] 287 ldp x6,x7,[x26,#16*3] 288 ldp x8,x9,[x26,#16*4] 289 ldp x10,x11,[x26,#16*5] 290 ldp x12,x13,[x26,#16*6] 291 ldp x14,x15,[x26,#16*7] 292 ldp x16,x17,[x26,#16*8] 293 ldp x25,x19,[x26,#16*9] 294 ldp x20,x21,[x26,#16*10] 295 ldp x22,x23,[x26,#16*11] 296 ldr x24,[x26,#16*12] 297 b .Loop_absorb 298 299 .align 4 300 .Loop_absorb: 301 subs x26,x28,x30 // len - bsz 302 blo .Labsorbed 303 304 str x26,[sp,#48] // save len - bsz 305 ldr x26,[x27],#8 // *inp++ 306 #ifdef __AARCH64EB__ 307 rev x26,x26 308 #endif 309 eor x0,x0,x26 310 cmp x30,#8*(0+2) 311 blo .Lprocess_block 312 ldr x26,[x27],#8 // *inp++ 313 #ifdef __AARCH64EB__ 314 rev x26,x26 315 #endif 316 eor x1,x1,x26 317 beq .Lprocess_block 318 ldr x26,[x27],#8 // *inp++ 319 #ifdef __AARCH64EB__ 320 rev x26,x26 321 #endif 322 eor x2,x2,x26 323 cmp x30,#8*(2+2) 324 blo .Lprocess_block 325 ldr x26,[x27],#8 // *inp++ 326 #ifdef __AARCH64EB__ 327 rev x26,x26 328 #endif 329 eor x3,x3,x26 330 beq .Lprocess_block 331 ldr x26,[x27],#8 // *inp++ 332 #ifdef __AARCH64EB__ 333 rev x26,x26 334 #endif 335 eor x4,x4,x26 336 cmp x30,#8*(4+2) 337 blo .Lprocess_block 338 ldr x26,[x27],#8 // *inp++ 339 #ifdef __AARCH64EB__ 340 rev x26,x26 341 #endif 342 eor x5,x5,x26 343 beq .Lprocess_block 344 ldr x26,[x27],#8 // *inp++ 345 #ifdef __AARCH64EB__ 346 rev x26,x26 347 #endif 348 eor x6,x6,x26 349 cmp x30,#8*(6+2) 350 blo .Lprocess_block 351 ldr x26,[x27],#8 // *inp++ 352 #ifdef __AARCH64EB__ 353 rev x26,x26 354 #endif 355 eor x7,x7,x26 356 beq .Lprocess_block 357 ldr x26,[x27],#8 // *inp++ 358 #ifdef __AARCH64EB__ 359 rev x26,x26 360 #endif 361 eor x8,x8,x26 362 cmp x30,#8*(8+2) 363 blo .Lprocess_block 364 ldr x26,[x27],#8 // *inp++ 365 #ifdef __AARCH64EB__ 366 rev x26,x26 367 #endif 368 eor x9,x9,x26 369 beq .Lprocess_block 370 ldr x26,[x27],#8 // *inp++ 371 #ifdef __AARCH64EB__ 372 rev x26,x26 373 #endif 374 eor x10,x10,x26 375 cmp x30,#8*(10+2) 376 blo .Lprocess_block 377 ldr x26,[x27],#8 // *inp++ 378 #ifdef __AARCH64EB__ 379 rev x26,x26 380 #endif 381 eor x11,x11,x26 382 beq .Lprocess_block 383 ldr x26,[x27],#8 // *inp++ 384 #ifdef __AARCH64EB__ 385 rev x26,x26 386 #endif 387 eor x12,x12,x26 388 cmp x30,#8*(12+2) 389 blo .Lprocess_block 390 ldr x26,[x27],#8 // *inp++ 391 #ifdef __AARCH64EB__ 392 rev x26,x26 393 #endif 394 eor x13,x13,x26 395 beq .Lprocess_block 396 ldr x26,[x27],#8 // *inp++ 397 #ifdef __AARCH64EB__ 398 rev x26,x26 399 #endif 400 eor x14,x14,x26 401 cmp x30,#8*(14+2) 402 blo .Lprocess_block 403 ldr x26,[x27],#8 // *inp++ 404 #ifdef __AARCH64EB__ 405 rev x26,x26 406 #endif 407 eor x15,x15,x26 408 beq .Lprocess_block 409 ldr x26,[x27],#8 // *inp++ 410 #ifdef __AARCH64EB__ 411 rev x26,x26 412 #endif 413 eor x16,x16,x26 414 cmp x30,#8*(16+2) 415 blo .Lprocess_block 416 ldr x26,[x27],#8 // *inp++ 417 #ifdef __AARCH64EB__ 418 rev x26,x26 419 #endif 420 eor x17,x17,x26 421 beq .Lprocess_block 422 ldr x26,[x27],#8 // *inp++ 423 #ifdef __AARCH64EB__ 424 rev x26,x26 425 #endif 426 eor x25,x25,x26 427 cmp x30,#8*(18+2) 428 blo .Lprocess_block 429 ldr x26,[x27],#8 // *inp++ 430 #ifdef __AARCH64EB__ 431 rev x26,x26 432 #endif 433 eor x19,x19,x26 434 beq .Lprocess_block 435 ldr x26,[x27],#8 // *inp++ 436 #ifdef __AARCH64EB__ 437 rev x26,x26 438 #endif 439 eor x20,x20,x26 440 cmp x30,#8*(20+2) 441 blo .Lprocess_block 442 ldr x26,[x27],#8 // *inp++ 443 #ifdef __AARCH64EB__ 444 rev x26,x26 445 #endif 446 eor x21,x21,x26 447 beq .Lprocess_block 448 ldr x26,[x27],#8 // *inp++ 449 #ifdef __AARCH64EB__ 450 rev x26,x26 451 #endif 452 eor x22,x22,x26 453 cmp x30,#8*(22+2) 454 blo .Lprocess_block 455 ldr x26,[x27],#8 // *inp++ 456 #ifdef __AARCH64EB__ 457 rev x26,x26 458 #endif 459 eor x23,x23,x26 460 beq .Lprocess_block 461 ldr x26,[x27],#8 // *inp++ 462 #ifdef __AARCH64EB__ 463 rev x26,x26 464 #endif 465 eor x24,x24,x26 466 467 .Lprocess_block: 468 str x27,[sp,#40] // save inp 469 470 bl KeccakF1600_int 471 472 ldr x27,[sp,#40] // restore arguments 473 ldp x28,x30,[sp,#48] 474 b .Loop_absorb 475 476 .align 4 477 .Labsorbed: 478 ldr x27,[sp,#32] 479 stp x0,x1,[x27,#16*0] 480 stp x2,x3,[x27,#16*1] 481 stp x4,x5,[x27,#16*2] 482 stp x6,x7,[x27,#16*3] 483 stp x8,x9,[x27,#16*4] 484 stp x10,x11,[x27,#16*5] 485 stp x12,x13,[x27,#16*6] 486 stp x14,x15,[x27,#16*7] 487 stp x16,x17,[x27,#16*8] 488 stp x25,x19,[x27,#16*9] 489 stp x20,x21,[x27,#16*10] 490 stp x22,x23,[x27,#16*11] 491 str x24,[x27,#16*12] 492 493 mov x0,x28 // return value 494 ldp x19,x20,[x29,#16] 495 add sp,sp,#64 496 ldp x21,x22,[x29,#32] 497 ldp x23,x24,[x29,#48] 498 ldp x25,x26,[x29,#64] 499 ldp x27,x28,[x29,#80] 500 ldp x29,x30,[sp],#128 501 .inst 0xd50323bf // autiasp 502 ret 503 .size SHA3_absorb,.-SHA3_absorb 504 .globl SHA3_squeeze 505 .type SHA3_squeeze,%function 506 .align 5 507 SHA3_squeeze: 508 .inst 0xd503233f // paciasp 509 stp x29,x30,[sp,#-48]! 510 add x29,sp,#0 511 stp x19,x20,[sp,#16] 512 stp x21,x22,[sp,#32] 513 514 mov x19,x0 // put aside arguments 515 mov x20,x1 516 mov x21,x2 517 mov x22,x3 518 519 .Loop_squeeze: 520 ldr x4,[x0],#8 521 cmp x21,#8 522 blo .Lsqueeze_tail 523 #ifdef __AARCH64EB__ 524 rev x4,x4 525 #endif 526 str x4,[x20],#8 527 subs x21,x21,#8 528 beq .Lsqueeze_done 529 530 subs x3,x3,#8 531 bhi .Loop_squeeze 532 533 mov x0,x19 534 bl KeccakF1600 535 mov x0,x19 536 mov x3,x22 537 b .Loop_squeeze 538 539 .align 4 540 .Lsqueeze_tail: 541 strb w4,[x20],#1 542 lsr x4,x4,#8 543 subs x21,x21,#1 544 beq .Lsqueeze_done 545 strb w4,[x20],#1 546 lsr x4,x4,#8 547 subs x21,x21,#1 548 beq .Lsqueeze_done 549 strb w4,[x20],#1 550 lsr x4,x4,#8 551 subs x21,x21,#1 552 beq .Lsqueeze_done 553 strb w4,[x20],#1 554 lsr x4,x4,#8 555 subs x21,x21,#1 556 beq .Lsqueeze_done 557 strb w4,[x20],#1 558 lsr x4,x4,#8 559 subs x21,x21,#1 560 beq .Lsqueeze_done 561 strb w4,[x20],#1 562 lsr x4,x4,#8 563 subs x21,x21,#1 564 beq .Lsqueeze_done 565 strb w4,[x20],#1 566 567 .Lsqueeze_done: 568 ldp x19,x20,[sp,#16] 569 ldp x21,x22,[sp,#32] 570 ldp x29,x30,[sp],#48 571 .inst 0xd50323bf // autiasp 572 ret 573 .size SHA3_squeeze,.-SHA3_squeeze 574 .type KeccakF1600_ce,%function 575 .align 5 576 KeccakF1600_ce: 577 mov x9,#12 578 adr x10,iotas 579 b .Loop_ce 580 .align 4 581 .Loop_ce: 582 ////////////////////////////////////////////////// Theta 583 .inst 0xce052819 //eor3 v25.16b,v0.16b,v5.16b,v10.16b 584 .inst 0xce062c3a //eor3 v26.16b,v1.16b,v6.16b,v11.16b 585 .inst 0xce07305b //eor3 v27.16b,v2.16b,v7.16b,v12.16b 586 .inst 0xce08347c //eor3 v28.16b,v3.16b,v8.16b,v13.16b 587 .inst 0xce09389d //eor3 v29.16b,v4.16b,v9.16b,v14.16b 588 .inst 0xce0f5339 //eor3 v25.16b,v25.16b, v15.16b,v20.16b 589 .inst 0xce10575a //eor3 v26.16b,v26.16b, v16.16b,v21.16b 590 .inst 0xce115b7b //eor3 v27.16b,v27.16b, v17.16b,v22.16b 591 .inst 0xce125f9c //eor3 v28.16b,v28.16b, v18.16b,v23.16b 592 .inst 0xce1363bd //eor3 v29.16b,v29.16b, v19.16b,v24.16b 593 594 .inst 0xce7b8f3e //rax1 v30.16b,v25.16b,v27.16b // D[1] 595 .inst 0xce7c8f5f //rax1 v31.16b,v26.16b,v28.16b // D[2] 596 .inst 0xce7d8f7b //rax1 v27.16b,v27.16b,v29.16b // D[3] 597 .inst 0xce798f9c //rax1 v28.16b,v28.16b,v25.16b // D[4] 598 .inst 0xce7a8fbd //rax1 v29.16b,v29.16b,v26.16b // D[0] 599 600 ////////////////////////////////////////////////// Theta+Rho+Pi 601 .inst 0xce9e50d9 //xar v25.16b, v6.16b,v30.16b,#64-44 // C[0]=A[0][1] 602 .inst 0xce9cb126 //xar v6.16b,v9.16b,v28.16b,#64-20 603 .inst 0xce9f0ec9 //xar v9.16b,v22.16b,v31.16b,#64-61 604 .inst 0xce9c65d6 //xar v22.16b,v14.16b,v28.16b,#64-39 605 .inst 0xce9dba8e //xar v14.16b,v20.16b,v29.16b,#64-18 606 607 .inst 0xce9f0854 //xar v20.16b,v2.16b,v31.16b,#64-62 608 609 .inst 0xce9f5582 //xar v2.16b,v12.16b,v31.16b,#64-43 610 .inst 0xce9b9dac //xar v12.16b,v13.16b,v27.16b,#64-25 611 .inst 0xce9ce26d //xar v13.16b,v19.16b,v28.16b,#64-8 612 .inst 0xce9b22f3 //xar v19.16b,v23.16b,v27.16b,#64-56 613 .inst 0xce9d5df7 //xar v23.16b,v15.16b,v29.16b,#64-41 614 615 .inst 0xce9c948f //xar v15.16b,v4.16b,v28.16b,#64-27 616 617 eor v0.16b,v0.16b,v29.16b 618 ldr x11,[x10],#8 619 620 .inst 0xce9bae5a //xar v26.16b, v18.16b,v27.16b,#64-21 // C[1]=A[0][3] 621 .inst 0xce9fc632 //xar v18.16b,v17.16b,v31.16b,#64-15 622 .inst 0xce9ed971 //xar v17.16b,v11.16b,v30.16b,#64-10 623 .inst 0xce9fe8eb //xar v11.16b,v7.16b,v31.16b,#64-6 624 .inst 0xce9df547 //xar v7.16b,v10.16b,v29.16b,#64-3 625 626 .inst 0xce9efc2a //xar v10.16b,v1.16b,v30.16b,#64-1 // * 627 628 .inst 0xce9ccb04 //xar v4.16b,v24.16b,v28.16b,#64-14 629 .inst 0xce9efab8 //xar v24.16b,v21.16b,v30.16b,#64-2 630 .inst 0xce9b2515 //xar v21.16b,v8.16b,v27.16b,#64-55 631 .inst 0xce9e4e08 //xar v8.16b,v16.16b,v30.16b,#64-45 632 .inst 0xce9d70b0 //xar v16.16b,v5.16b,v29.16b,#64-36 633 634 .inst 0xce9b907b //xar v27.16b, v3.16b,v27.16b,#64-28 // C[2]=A[1][0] 635 636 ////////////////////////////////////////////////// Chi+Iota 637 dup v31.2d,x11 // borrow C[6] 638 .inst 0xce22641c //bcax v28.16b, v0.16b,v2.16b,v25.16b // * 639 .inst 0xce3a0b21 //bcax v1.16b,v25.16b, v26.16b, v2.16b // * 640 .inst 0xce246842 //bcax v2.16b,v2.16b,v4.16b,v26.16b 641 .inst 0xce201343 //bcax v3.16b,v26.16b, v0.16b,v4.16b 642 .inst 0xce390084 //bcax v4.16b,v4.16b,v25.16b, v0.16b 643 644 .inst 0xce271b65 //bcax v5.16b,v27.16b, v7.16b,v6.16b // * 645 .inst 0xce281cd9 //bcax v25.16b, v6.16b,v8.16b,v7.16b // * 646 .inst 0xce2920e7 //bcax v7.16b,v7.16b,v9.16b,v8.16b 647 .inst 0xce3b2508 //bcax v8.16b,v8.16b,v27.16b, v9.16b 648 .inst 0xce266d29 //bcax v9.16b,v9.16b,v6.16b,v27.16b 649 650 eor v0.16b,v28.16b,v31.16b // Iota 651 652 .inst 0xce2c2d5a //bcax v26.16b, v10.16b,v12.16b,v11.16b // * 653 .inst 0xce2d317b //bcax v27.16b, v11.16b,v13.16b,v12.16b // * 654 .inst 0xce2e358c //bcax v12.16b,v12.16b,v14.16b,v13.16b 655 .inst 0xce2a39ad //bcax v13.16b,v13.16b,v10.16b,v14.16b 656 .inst 0xce2b29ce //bcax v14.16b,v14.16b,v11.16b,v10.16b 657 658 .inst 0xce3141fc //bcax v28.16b, v15.16b,v17.16b,v16.16b // * 659 .inst 0xce32461d //bcax v29.16b, v16.16b,v18.16b,v17.16b // * 660 .inst 0xce334a31 //bcax v17.16b,v17.16b,v19.16b,v18.16b 661 .inst 0xce2f4e52 //bcax v18.16b,v18.16b,v15.16b,v19.16b 662 .inst 0xce303e73 //bcax v19.16b,v19.16b,v16.16b,v15.16b 663 664 .inst 0xce36569e //bcax v30.16b, v20.16b,v22.16b,v21.16b // * 665 .inst 0xce375abf //bcax v31.16b, v21.16b,v23.16b,v22.16b // * 666 .inst 0xce385ed6 //bcax v22.16b,v22.16b,v24.16b,v23.16b 667 .inst 0xce3462f7 //bcax v23.16b,v23.16b,v20.16b,v24.16b 668 .inst 0xce355318 //bcax v24.16b,v24.16b,v21.16b,v20.16b 669 ////////////////////////////////////////////////// Theta 670 .inst 0xce056806 //eor3 v6.16b,v0.16b,v5.16b,v26.16b 671 .inst 0xce196c2a //eor3 v10.16b,v1.16b,v25.16b,v27.16b 672 .inst 0xce07304b //eor3 v11.16b,v2.16b,v7.16b,v12.16b 673 .inst 0xce08346f //eor3 v15.16b,v3.16b,v8.16b,v13.16b 674 .inst 0xce093890 //eor3 v16.16b,v4.16b,v9.16b,v14.16b 675 .inst 0xce1c78c6 //eor3 v6.16b,v6.16b, v28.16b,v30.16b 676 .inst 0xce1d7d4a //eor3 v10.16b,v10.16b, v29.16b,v31.16b 677 .inst 0xce11596b //eor3 v11.16b,v11.16b, v17.16b,v22.16b 678 .inst 0xce125def //eor3 v15.16b,v15.16b, v18.16b,v23.16b 679 .inst 0xce136210 //eor3 v16.16b,v16.16b, v19.16b,v24.16b 680 681 .inst 0xce6b8cd4 //rax1 v20.16b,v6.16b,v11.16b // D[1] 682 .inst 0xce6f8d55 //rax1 v21.16b,v10.16b,v15.16b // D[2] 683 .inst 0xce708d6b //rax1 v11.16b,v11.16b,v16.16b // D[3] 684 .inst 0xce668def //rax1 v15.16b,v15.16b,v6.16b // D[4] 685 .inst 0xce6a8e10 //rax1 v16.16b,v16.16b,v10.16b // D[0] 686 687 ////////////////////////////////////////////////// Theta+Rho+Pi 688 .inst 0xce945326 //xar v6.16b, v25.16b,v20.16b,#64-44 // C[0]=A[0][1] 689 .inst 0xce8fb139 //xar v25.16b,v9.16b,v15.16b,#64-20 690 .inst 0xce950ec9 //xar v9.16b,v22.16b,v21.16b,#64-61 691 .inst 0xce8f65d6 //xar v22.16b,v14.16b,v15.16b,#64-39 692 .inst 0xce90bbce //xar v14.16b,v30.16b,v16.16b,#64-18 693 694 .inst 0xce95085e //xar v30.16b,v2.16b,v21.16b,#64-62 695 696 .inst 0xce955582 //xar v2.16b,v12.16b,v21.16b,#64-43 697 .inst 0xce8b9dac //xar v12.16b,v13.16b,v11.16b,#64-25 698 .inst 0xce8fe26d //xar v13.16b,v19.16b,v15.16b,#64-8 699 .inst 0xce8b22f3 //xar v19.16b,v23.16b,v11.16b,#64-56 700 .inst 0xce905f97 //xar v23.16b,v28.16b,v16.16b,#64-41 701 702 .inst 0xce8f949c //xar v28.16b,v4.16b,v15.16b,#64-27 703 704 eor v0.16b,v0.16b,v16.16b 705 ldr x11,[x10],#8 706 707 .inst 0xce8bae4a //xar v10.16b, v18.16b,v11.16b,#64-21 // C[1]=A[0][3] 708 .inst 0xce95c632 //xar v18.16b,v17.16b,v21.16b,#64-15 709 .inst 0xce94db71 //xar v17.16b,v27.16b,v20.16b,#64-10 710 .inst 0xce95e8fb //xar v27.16b,v7.16b,v21.16b,#64-6 711 .inst 0xce90f747 //xar v7.16b,v26.16b,v16.16b,#64-3 712 713 .inst 0xce94fc3a //xar v26.16b,v1.16b,v20.16b,#64-1 // * 714 715 .inst 0xce8fcb04 //xar v4.16b,v24.16b,v15.16b,#64-14 716 .inst 0xce94fbf8 //xar v24.16b,v31.16b,v20.16b,#64-2 717 .inst 0xce8b251f //xar v31.16b,v8.16b,v11.16b,#64-55 718 .inst 0xce944fa8 //xar v8.16b,v29.16b,v20.16b,#64-45 719 .inst 0xce9070bd //xar v29.16b,v5.16b,v16.16b,#64-36 720 721 .inst 0xce8b906b //xar v11.16b, v3.16b,v11.16b,#64-28 // C[2]=A[1][0] 722 723 ////////////////////////////////////////////////// Chi+Iota 724 dup v21.2d,x11 // borrow C[6] 725 .inst 0xce22180f //bcax v15.16b, v0.16b,v2.16b,v6.16b // * 726 .inst 0xce2a08c1 //bcax v1.16b,v6.16b, v10.16b, v2.16b // * 727 .inst 0xce242842 //bcax v2.16b,v2.16b,v4.16b,v10.16b 728 .inst 0xce201143 //bcax v3.16b,v10.16b, v0.16b,v4.16b 729 .inst 0xce260084 //bcax v4.16b,v4.16b,v6.16b, v0.16b 730 731 .inst 0xce276565 //bcax v5.16b,v11.16b, v7.16b,v25.16b // * 732 .inst 0xce281f26 //bcax v6.16b, v25.16b,v8.16b,v7.16b // * 733 .inst 0xce2920e7 //bcax v7.16b,v7.16b,v9.16b,v8.16b 734 .inst 0xce2b2508 //bcax v8.16b,v8.16b,v11.16b, v9.16b 735 .inst 0xce392d29 //bcax v9.16b,v9.16b,v25.16b,v11.16b 736 737 eor v0.16b,v15.16b,v21.16b // Iota 738 739 .inst 0xce2c6f4a //bcax v10.16b, v26.16b,v12.16b,v27.16b // * 740 .inst 0xce2d336b //bcax v11.16b, v27.16b,v13.16b,v12.16b // * 741 .inst 0xce2e358c //bcax v12.16b,v12.16b,v14.16b,v13.16b 742 .inst 0xce3a39ad //bcax v13.16b,v13.16b,v26.16b,v14.16b 743 .inst 0xce3b69ce //bcax v14.16b,v14.16b,v27.16b,v26.16b 744 745 .inst 0xce31778f //bcax v15.16b, v28.16b,v17.16b,v29.16b // * 746 .inst 0xce3247b0 //bcax v16.16b, v29.16b,v18.16b,v17.16b // * 747 .inst 0xce334a31 //bcax v17.16b,v17.16b,v19.16b,v18.16b 748 .inst 0xce3c4e52 //bcax v18.16b,v18.16b,v28.16b,v19.16b 749 .inst 0xce3d7273 //bcax v19.16b,v19.16b,v29.16b,v28.16b 750 751 .inst 0xce367fd4 //bcax v20.16b, v30.16b,v22.16b,v31.16b // * 752 .inst 0xce375bf5 //bcax v21.16b, v31.16b,v23.16b,v22.16b // * 753 .inst 0xce385ed6 //bcax v22.16b,v22.16b,v24.16b,v23.16b 754 .inst 0xce3e62f7 //bcax v23.16b,v23.16b,v30.16b,v24.16b 755 .inst 0xce3f7b18 //bcax v24.16b,v24.16b,v31.16b,v30.16b 756 subs x9,x9,#1 757 bne .Loop_ce 758 759 ret 760 .size KeccakF1600_ce,.-KeccakF1600_ce 761 762 .type KeccakF1600_cext,%function 763 .align 5 764 KeccakF1600_cext: 765 .inst 0xd503233f // paciasp 766 stp x29,x30,[sp,#-80]! 767 add x29,sp,#0 768 stp d8,d9,[sp,#16] // per ABI requirement 769 stp d10,d11,[sp,#32] 770 stp d12,d13,[sp,#48] 771 stp d14,d15,[sp,#64] 772 ldp d0,d1,[x0,#8*0] 773 ldp d2,d3,[x0,#8*2] 774 ldp d4,d5,[x0,#8*4] 775 ldp d6,d7,[x0,#8*6] 776 ldp d8,d9,[x0,#8*8] 777 ldp d10,d11,[x0,#8*10] 778 ldp d12,d13,[x0,#8*12] 779 ldp d14,d15,[x0,#8*14] 780 ldp d16,d17,[x0,#8*16] 781 ldp d18,d19,[x0,#8*18] 782 ldp d20,d21,[x0,#8*20] 783 ldp d22,d23,[x0,#8*22] 784 ldr d24,[x0,#8*24] 785 bl KeccakF1600_ce 786 ldr x30,[sp,#8] 787 stp d0,d1,[x0,#8*0] 788 stp d2,d3,[x0,#8*2] 789 stp d4,d5,[x0,#8*4] 790 stp d6,d7,[x0,#8*6] 791 stp d8,d9,[x0,#8*8] 792 stp d10,d11,[x0,#8*10] 793 stp d12,d13,[x0,#8*12] 794 stp d14,d15,[x0,#8*14] 795 stp d16,d17,[x0,#8*16] 796 stp d18,d19,[x0,#8*18] 797 stp d20,d21,[x0,#8*20] 798 stp d22,d23,[x0,#8*22] 799 str d24,[x0,#8*24] 800 801 ldp d8,d9,[sp,#16] 802 ldp d10,d11,[sp,#32] 803 ldp d12,d13,[sp,#48] 804 ldp d14,d15,[sp,#64] 805 ldr x29,[sp],#80 806 .inst 0xd50323bf // autiasp 807 ret 808 .size KeccakF1600_cext,.-KeccakF1600_cext 809 .globl SHA3_absorb_cext 810 .type SHA3_absorb_cext,%function 811 .align 5 812 SHA3_absorb_cext: 813 .inst 0xd503233f // paciasp 814 stp x29,x30,[sp,#-80]! 815 add x29,sp,#0 816 stp d8,d9,[sp,#16] // per ABI requirement 817 stp d10,d11,[sp,#32] 818 stp d12,d13,[sp,#48] 819 stp d14,d15,[sp,#64] 820 ldp d0,d1,[x0,#8*0] 821 ldp d2,d3,[x0,#8*2] 822 ldp d4,d5,[x0,#8*4] 823 ldp d6,d7,[x0,#8*6] 824 ldp d8,d9,[x0,#8*8] 825 ldp d10,d11,[x0,#8*10] 826 ldp d12,d13,[x0,#8*12] 827 ldp d14,d15,[x0,#8*14] 828 ldp d16,d17,[x0,#8*16] 829 ldp d18,d19,[x0,#8*18] 830 ldp d20,d21,[x0,#8*20] 831 ldp d22,d23,[x0,#8*22] 832 ldr d24,[x0,#8*24] 833 b .Loop_absorb_ce 834 835 .align 4 836 .Loop_absorb_ce: 837 subs x2,x2,x3 // len - bsz 838 blo .Labsorbed_ce 839 ldr d31,[x1],#8 // *inp++ 840 #ifdef __AARCH64EB__ 841 rev64 v31.16b,v31.16b 842 #endif 843 eor v0.16b,v0.16b,v31.16b 844 cmp x3,#8*(0+2) 845 blo .Lprocess_block_ce 846 ldr d31,[x1],#8 // *inp++ 847 #ifdef __AARCH64EB__ 848 rev64 v31.16b,v31.16b 849 #endif 850 eor v1.16b,v1.16b,v31.16b 851 beq .Lprocess_block_ce 852 ldr d31,[x1],#8 // *inp++ 853 #ifdef __AARCH64EB__ 854 rev64 v31.16b,v31.16b 855 #endif 856 eor v2.16b,v2.16b,v31.16b 857 cmp x3,#8*(2+2) 858 blo .Lprocess_block_ce 859 ldr d31,[x1],#8 // *inp++ 860 #ifdef __AARCH64EB__ 861 rev64 v31.16b,v31.16b 862 #endif 863 eor v3.16b,v3.16b,v31.16b 864 beq .Lprocess_block_ce 865 ldr d31,[x1],#8 // *inp++ 866 #ifdef __AARCH64EB__ 867 rev64 v31.16b,v31.16b 868 #endif 869 eor v4.16b,v4.16b,v31.16b 870 cmp x3,#8*(4+2) 871 blo .Lprocess_block_ce 872 ldr d31,[x1],#8 // *inp++ 873 #ifdef __AARCH64EB__ 874 rev64 v31.16b,v31.16b 875 #endif 876 eor v5.16b,v5.16b,v31.16b 877 beq .Lprocess_block_ce 878 ldr d31,[x1],#8 // *inp++ 879 #ifdef __AARCH64EB__ 880 rev64 v31.16b,v31.16b 881 #endif 882 eor v6.16b,v6.16b,v31.16b 883 cmp x3,#8*(6+2) 884 blo .Lprocess_block_ce 885 ldr d31,[x1],#8 // *inp++ 886 #ifdef __AARCH64EB__ 887 rev64 v31.16b,v31.16b 888 #endif 889 eor v7.16b,v7.16b,v31.16b 890 beq .Lprocess_block_ce 891 ldr d31,[x1],#8 // *inp++ 892 #ifdef __AARCH64EB__ 893 rev64 v31.16b,v31.16b 894 #endif 895 eor v8.16b,v8.16b,v31.16b 896 cmp x3,#8*(8+2) 897 blo .Lprocess_block_ce 898 ldr d31,[x1],#8 // *inp++ 899 #ifdef __AARCH64EB__ 900 rev64 v31.16b,v31.16b 901 #endif 902 eor v9.16b,v9.16b,v31.16b 903 beq .Lprocess_block_ce 904 ldr d31,[x1],#8 // *inp++ 905 #ifdef __AARCH64EB__ 906 rev64 v31.16b,v31.16b 907 #endif 908 eor v10.16b,v10.16b,v31.16b 909 cmp x3,#8*(10+2) 910 blo .Lprocess_block_ce 911 ldr d31,[x1],#8 // *inp++ 912 #ifdef __AARCH64EB__ 913 rev64 v31.16b,v31.16b 914 #endif 915 eor v11.16b,v11.16b,v31.16b 916 beq .Lprocess_block_ce 917 ldr d31,[x1],#8 // *inp++ 918 #ifdef __AARCH64EB__ 919 rev64 v31.16b,v31.16b 920 #endif 921 eor v12.16b,v12.16b,v31.16b 922 cmp x3,#8*(12+2) 923 blo .Lprocess_block_ce 924 ldr d31,[x1],#8 // *inp++ 925 #ifdef __AARCH64EB__ 926 rev64 v31.16b,v31.16b 927 #endif 928 eor v13.16b,v13.16b,v31.16b 929 beq .Lprocess_block_ce 930 ldr d31,[x1],#8 // *inp++ 931 #ifdef __AARCH64EB__ 932 rev64 v31.16b,v31.16b 933 #endif 934 eor v14.16b,v14.16b,v31.16b 935 cmp x3,#8*(14+2) 936 blo .Lprocess_block_ce 937 ldr d31,[x1],#8 // *inp++ 938 #ifdef __AARCH64EB__ 939 rev64 v31.16b,v31.16b 940 #endif 941 eor v15.16b,v15.16b,v31.16b 942 beq .Lprocess_block_ce 943 ldr d31,[x1],#8 // *inp++ 944 #ifdef __AARCH64EB__ 945 rev64 v31.16b,v31.16b 946 #endif 947 eor v16.16b,v16.16b,v31.16b 948 cmp x3,#8*(16+2) 949 blo .Lprocess_block_ce 950 ldr d31,[x1],#8 // *inp++ 951 #ifdef __AARCH64EB__ 952 rev64 v31.16b,v31.16b 953 #endif 954 eor v17.16b,v17.16b,v31.16b 955 beq .Lprocess_block_ce 956 ldr d31,[x1],#8 // *inp++ 957 #ifdef __AARCH64EB__ 958 rev64 v31.16b,v31.16b 959 #endif 960 eor v18.16b,v18.16b,v31.16b 961 cmp x3,#8*(18+2) 962 blo .Lprocess_block_ce 963 ldr d31,[x1],#8 // *inp++ 964 #ifdef __AARCH64EB__ 965 rev64 v31.16b,v31.16b 966 #endif 967 eor v19.16b,v19.16b,v31.16b 968 beq .Lprocess_block_ce 969 ldr d31,[x1],#8 // *inp++ 970 #ifdef __AARCH64EB__ 971 rev64 v31.16b,v31.16b 972 #endif 973 eor v20.16b,v20.16b,v31.16b 974 cmp x3,#8*(20+2) 975 blo .Lprocess_block_ce 976 ldr d31,[x1],#8 // *inp++ 977 #ifdef __AARCH64EB__ 978 rev64 v31.16b,v31.16b 979 #endif 980 eor v21.16b,v21.16b,v31.16b 981 beq .Lprocess_block_ce 982 ldr d31,[x1],#8 // *inp++ 983 #ifdef __AARCH64EB__ 984 rev64 v31.16b,v31.16b 985 #endif 986 eor v22.16b,v22.16b,v31.16b 987 cmp x3,#8*(22+2) 988 blo .Lprocess_block_ce 989 ldr d31,[x1],#8 // *inp++ 990 #ifdef __AARCH64EB__ 991 rev64 v31.16b,v31.16b 992 #endif 993 eor v23.16b,v23.16b,v31.16b 994 beq .Lprocess_block_ce 995 ldr d31,[x1],#8 // *inp++ 996 #ifdef __AARCH64EB__ 997 rev64 v31.16b,v31.16b 998 #endif 999 eor v24.16b,v24.16b,v31.16b 1000 1001 .Lprocess_block_ce: 1002 1003 bl KeccakF1600_ce 1004 1005 b .Loop_absorb_ce 1006 1007 .align 4 1008 .Labsorbed_ce: 1009 stp d0,d1,[x0,#8*0] 1010 stp d2,d3,[x0,#8*2] 1011 stp d4,d5,[x0,#8*4] 1012 stp d6,d7,[x0,#8*6] 1013 stp d8,d9,[x0,#8*8] 1014 stp d10,d11,[x0,#8*10] 1015 stp d12,d13,[x0,#8*12] 1016 stp d14,d15,[x0,#8*14] 1017 stp d16,d17,[x0,#8*16] 1018 stp d18,d19,[x0,#8*18] 1019 stp d20,d21,[x0,#8*20] 1020 stp d22,d23,[x0,#8*22] 1021 str d24,[x0,#8*24] 1022 add x0,x2,x3 // return value 1023 1024 ldp d8,d9,[sp,#16] 1025 ldp d10,d11,[sp,#32] 1026 ldp d12,d13,[sp,#48] 1027 ldp d14,d15,[sp,#64] 1028 ldp x29,x30,[sp],#80 1029 .inst 0xd50323bf // autiasp 1030 ret 1031 .size SHA3_absorb_cext,.-SHA3_absorb_cext 1032 .globl SHA3_squeeze_cext 1033 .type SHA3_squeeze_cext,%function 1034 .align 5 1035 SHA3_squeeze_cext: 1036 .inst 0xd503233f // paciasp 1037 stp x29,x30,[sp,#-16]! 1038 add x29,sp,#0 1039 mov x9,x0 1040 mov x10,x3 1041 1042 .Loop_squeeze_ce: 1043 ldr x4,[x9],#8 1044 cmp x2,#8 1045 blo .Lsqueeze_tail_ce 1046 #ifdef __AARCH64EB__ 1047 rev x4,x4 1048 #endif 1049 str x4,[x1],#8 1050 beq .Lsqueeze_done_ce 1051 1052 sub x2,x2,#8 1053 subs x10,x10,#8 1054 bhi .Loop_squeeze_ce 1055 1056 bl KeccakF1600_cext 1057 ldr x30,[sp,#8] 1058 mov x9,x0 1059 mov x10,x3 1060 b .Loop_squeeze_ce 1061 1062 .align 4 1063 .Lsqueeze_tail_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 lsr x4,x4,#8 1078 subs x2,x2,#1 1079 beq .Lsqueeze_done_ce 1080 strb w4,[x1],#1 1081 lsr x4,x4,#8 1082 subs x2,x2,#1 1083 beq .Lsqueeze_done_ce 1084 strb w4,[x1],#1 1085 lsr x4,x4,#8 1086 subs x2,x2,#1 1087 beq .Lsqueeze_done_ce 1088 strb w4,[x1],#1 1089 1090 .Lsqueeze_done_ce: 1091 ldr x29,[sp],#16 1092 .inst 0xd50323bf // autiasp 1093 ret 1094 .size SHA3_squeeze_cext,.-SHA3_squeeze_cext 1095 .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 1096 .align 2 1097
Indexes created Mon Apr 27 00:23:16 UTC 2026