Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 1 | /* |
| 2 | * des3_ede-asm_64.S - x86-64 assembly implementation of 3DES cipher |
| 3 | * |
| 4 | * Copyright © 2014 Jussi Kivilinna <jussi.kivilinna@iki.fi> |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License as published by |
| 8 | * the Free Software Foundation; either version 2 of the License, or |
| 9 | * (at your option) any later version. |
| 10 | * |
| 11 | * This program 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 |
| 14 | * GNU General Public License for more details. |
| 15 | */ |
| 16 | |
| 17 | #include <linux/linkage.h> |
| 18 | |
| 19 | .file "des3_ede-asm_64.S" |
| 20 | .text |
| 21 | |
| 22 | #define s1 .L_s1 |
| 23 | #define s2 ((s1) + (64*8)) |
| 24 | #define s3 ((s2) + (64*8)) |
| 25 | #define s4 ((s3) + (64*8)) |
| 26 | #define s5 ((s4) + (64*8)) |
| 27 | #define s6 ((s5) + (64*8)) |
| 28 | #define s7 ((s6) + (64*8)) |
| 29 | #define s8 ((s7) + (64*8)) |
| 30 | |
| 31 | /* register macros */ |
| 32 | #define CTX %rdi |
| 33 | |
| 34 | #define RL0 %r8 |
| 35 | #define RL1 %r9 |
| 36 | #define RL2 %r10 |
| 37 | |
| 38 | #define RL0d %r8d |
| 39 | #define RL1d %r9d |
| 40 | #define RL2d %r10d |
| 41 | |
| 42 | #define RR0 %r11 |
| 43 | #define RR1 %r12 |
| 44 | #define RR2 %r13 |
| 45 | |
| 46 | #define RR0d %r11d |
| 47 | #define RR1d %r12d |
| 48 | #define RR2d %r13d |
| 49 | |
| 50 | #define RW0 %rax |
| 51 | #define RW1 %rbx |
| 52 | #define RW2 %rcx |
| 53 | |
| 54 | #define RW0d %eax |
| 55 | #define RW1d %ebx |
| 56 | #define RW2d %ecx |
| 57 | |
| 58 | #define RW0bl %al |
| 59 | #define RW1bl %bl |
| 60 | #define RW2bl %cl |
| 61 | |
| 62 | #define RW0bh %ah |
| 63 | #define RW1bh %bh |
| 64 | #define RW2bh %ch |
| 65 | |
| 66 | #define RT0 %r15 |
Josh Poimboeuf | 3ed7b4d | 2017-09-18 14:42:04 -0500 | [diff] [blame] | 67 | #define RT1 %rsi |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 68 | #define RT2 %r14 |
| 69 | #define RT3 %rdx |
| 70 | |
| 71 | #define RT0d %r15d |
Josh Poimboeuf | 3ed7b4d | 2017-09-18 14:42:04 -0500 | [diff] [blame] | 72 | #define RT1d %esi |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 73 | #define RT2d %r14d |
| 74 | #define RT3d %edx |
| 75 | |
| 76 | /*********************************************************************** |
| 77 | * 1-way 3DES |
| 78 | ***********************************************************************/ |
| 79 | #define do_permutation(a, b, offset, mask) \ |
| 80 | movl a, RT0d; \ |
| 81 | shrl $(offset), RT0d; \ |
| 82 | xorl b, RT0d; \ |
| 83 | andl $(mask), RT0d; \ |
| 84 | xorl RT0d, b; \ |
| 85 | shll $(offset), RT0d; \ |
| 86 | xorl RT0d, a; |
| 87 | |
| 88 | #define expand_to_64bits(val, mask) \ |
| 89 | movl val##d, RT0d; \ |
| 90 | rorl $4, RT0d; \ |
| 91 | shlq $32, RT0; \ |
| 92 | orq RT0, val; \ |
| 93 | andq mask, val; |
| 94 | |
| 95 | #define compress_to_64bits(val) \ |
| 96 | movq val, RT0; \ |
| 97 | shrq $32, RT0; \ |
| 98 | roll $4, RT0d; \ |
| 99 | orl RT0d, val##d; |
| 100 | |
| 101 | #define initial_permutation(left, right) \ |
| 102 | do_permutation(left##d, right##d, 4, 0x0f0f0f0f); \ |
| 103 | do_permutation(left##d, right##d, 16, 0x0000ffff); \ |
| 104 | do_permutation(right##d, left##d, 2, 0x33333333); \ |
| 105 | do_permutation(right##d, left##d, 8, 0x00ff00ff); \ |
| 106 | movabs $0x3f3f3f3f3f3f3f3f, RT3; \ |
| 107 | movl left##d, RW0d; \ |
| 108 | roll $1, right##d; \ |
| 109 | xorl right##d, RW0d; \ |
| 110 | andl $0xaaaaaaaa, RW0d; \ |
| 111 | xorl RW0d, left##d; \ |
| 112 | xorl RW0d, right##d; \ |
| 113 | roll $1, left##d; \ |
| 114 | expand_to_64bits(right, RT3); \ |
| 115 | expand_to_64bits(left, RT3); |
| 116 | |
| 117 | #define final_permutation(left, right) \ |
| 118 | compress_to_64bits(right); \ |
| 119 | compress_to_64bits(left); \ |
| 120 | movl right##d, RW0d; \ |
| 121 | rorl $1, left##d; \ |
| 122 | xorl left##d, RW0d; \ |
| 123 | andl $0xaaaaaaaa, RW0d; \ |
| 124 | xorl RW0d, right##d; \ |
| 125 | xorl RW0d, left##d; \ |
| 126 | rorl $1, right##d; \ |
| 127 | do_permutation(right##d, left##d, 8, 0x00ff00ff); \ |
| 128 | do_permutation(right##d, left##d, 2, 0x33333333); \ |
| 129 | do_permutation(left##d, right##d, 16, 0x0000ffff); \ |
| 130 | do_permutation(left##d, right##d, 4, 0x0f0f0f0f); |
| 131 | |
| 132 | #define round1(n, from, to, load_next_key) \ |
| 133 | xorq from, RW0; \ |
| 134 | \ |
| 135 | movzbl RW0bl, RT0d; \ |
| 136 | movzbl RW0bh, RT1d; \ |
| 137 | shrq $16, RW0; \ |
| 138 | movzbl RW0bl, RT2d; \ |
| 139 | movzbl RW0bh, RT3d; \ |
| 140 | shrq $16, RW0; \ |
| 141 | movq s8(, RT0, 8), RT0; \ |
| 142 | xorq s6(, RT1, 8), to; \ |
| 143 | movzbl RW0bl, RL1d; \ |
| 144 | movzbl RW0bh, RT1d; \ |
| 145 | shrl $16, RW0d; \ |
| 146 | xorq s4(, RT2, 8), RT0; \ |
| 147 | xorq s2(, RT3, 8), to; \ |
| 148 | movzbl RW0bl, RT2d; \ |
| 149 | movzbl RW0bh, RT3d; \ |
| 150 | xorq s7(, RL1, 8), RT0; \ |
| 151 | xorq s5(, RT1, 8), to; \ |
| 152 | xorq s3(, RT2, 8), RT0; \ |
| 153 | load_next_key(n, RW0); \ |
| 154 | xorq RT0, to; \ |
| 155 | xorq s1(, RT3, 8), to; \ |
| 156 | |
| 157 | #define load_next_key(n, RWx) \ |
| 158 | movq (((n) + 1) * 8)(CTX), RWx; |
| 159 | |
| 160 | #define dummy2(a, b) /*_*/ |
| 161 | |
| 162 | #define read_block(io, left, right) \ |
| 163 | movl (io), left##d; \ |
| 164 | movl 4(io), right##d; \ |
| 165 | bswapl left##d; \ |
| 166 | bswapl right##d; |
| 167 | |
| 168 | #define write_block(io, left, right) \ |
| 169 | bswapl left##d; \ |
| 170 | bswapl right##d; \ |
| 171 | movl left##d, (io); \ |
| 172 | movl right##d, 4(io); |
| 173 | |
| 174 | ENTRY(des3_ede_x86_64_crypt_blk) |
| 175 | /* input: |
| 176 | * %rdi: round keys, CTX |
| 177 | * %rsi: dst |
| 178 | * %rdx: src |
| 179 | */ |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 180 | pushq %rbx; |
| 181 | pushq %r12; |
| 182 | pushq %r13; |
| 183 | pushq %r14; |
| 184 | pushq %r15; |
| 185 | |
Josh Poimboeuf | 3ed7b4d | 2017-09-18 14:42:04 -0500 | [diff] [blame] | 186 | pushq %rsi; /* dst */ |
| 187 | |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 188 | read_block(%rdx, RL0, RR0); |
| 189 | initial_permutation(RL0, RR0); |
| 190 | |
| 191 | movq (CTX), RW0; |
| 192 | |
| 193 | round1(0, RR0, RL0, load_next_key); |
| 194 | round1(1, RL0, RR0, load_next_key); |
| 195 | round1(2, RR0, RL0, load_next_key); |
| 196 | round1(3, RL0, RR0, load_next_key); |
| 197 | round1(4, RR0, RL0, load_next_key); |
| 198 | round1(5, RL0, RR0, load_next_key); |
| 199 | round1(6, RR0, RL0, load_next_key); |
| 200 | round1(7, RL0, RR0, load_next_key); |
| 201 | round1(8, RR0, RL0, load_next_key); |
| 202 | round1(9, RL0, RR0, load_next_key); |
| 203 | round1(10, RR0, RL0, load_next_key); |
| 204 | round1(11, RL0, RR0, load_next_key); |
| 205 | round1(12, RR0, RL0, load_next_key); |
| 206 | round1(13, RL0, RR0, load_next_key); |
| 207 | round1(14, RR0, RL0, load_next_key); |
| 208 | round1(15, RL0, RR0, load_next_key); |
| 209 | |
| 210 | round1(16+0, RL0, RR0, load_next_key); |
| 211 | round1(16+1, RR0, RL0, load_next_key); |
| 212 | round1(16+2, RL0, RR0, load_next_key); |
| 213 | round1(16+3, RR0, RL0, load_next_key); |
| 214 | round1(16+4, RL0, RR0, load_next_key); |
| 215 | round1(16+5, RR0, RL0, load_next_key); |
| 216 | round1(16+6, RL0, RR0, load_next_key); |
| 217 | round1(16+7, RR0, RL0, load_next_key); |
| 218 | round1(16+8, RL0, RR0, load_next_key); |
| 219 | round1(16+9, RR0, RL0, load_next_key); |
| 220 | round1(16+10, RL0, RR0, load_next_key); |
| 221 | round1(16+11, RR0, RL0, load_next_key); |
| 222 | round1(16+12, RL0, RR0, load_next_key); |
| 223 | round1(16+13, RR0, RL0, load_next_key); |
| 224 | round1(16+14, RL0, RR0, load_next_key); |
| 225 | round1(16+15, RR0, RL0, load_next_key); |
| 226 | |
| 227 | round1(32+0, RR0, RL0, load_next_key); |
| 228 | round1(32+1, RL0, RR0, load_next_key); |
| 229 | round1(32+2, RR0, RL0, load_next_key); |
| 230 | round1(32+3, RL0, RR0, load_next_key); |
| 231 | round1(32+4, RR0, RL0, load_next_key); |
| 232 | round1(32+5, RL0, RR0, load_next_key); |
| 233 | round1(32+6, RR0, RL0, load_next_key); |
| 234 | round1(32+7, RL0, RR0, load_next_key); |
| 235 | round1(32+8, RR0, RL0, load_next_key); |
| 236 | round1(32+9, RL0, RR0, load_next_key); |
| 237 | round1(32+10, RR0, RL0, load_next_key); |
| 238 | round1(32+11, RL0, RR0, load_next_key); |
| 239 | round1(32+12, RR0, RL0, load_next_key); |
| 240 | round1(32+13, RL0, RR0, load_next_key); |
| 241 | round1(32+14, RR0, RL0, load_next_key); |
| 242 | round1(32+15, RL0, RR0, dummy2); |
| 243 | |
| 244 | final_permutation(RR0, RL0); |
Josh Poimboeuf | 3ed7b4d | 2017-09-18 14:42:04 -0500 | [diff] [blame] | 245 | |
| 246 | popq %rsi /* dst */ |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 247 | write_block(%rsi, RR0, RL0); |
| 248 | |
| 249 | popq %r15; |
| 250 | popq %r14; |
| 251 | popq %r13; |
| 252 | popq %r12; |
| 253 | popq %rbx; |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 254 | |
| 255 | ret; |
| 256 | ENDPROC(des3_ede_x86_64_crypt_blk) |
| 257 | |
| 258 | /*********************************************************************** |
| 259 | * 3-way 3DES |
| 260 | ***********************************************************************/ |
| 261 | #define expand_to_64bits(val, mask) \ |
| 262 | movl val##d, RT0d; \ |
| 263 | rorl $4, RT0d; \ |
| 264 | shlq $32, RT0; \ |
| 265 | orq RT0, val; \ |
| 266 | andq mask, val; |
| 267 | |
| 268 | #define compress_to_64bits(val) \ |
| 269 | movq val, RT0; \ |
| 270 | shrq $32, RT0; \ |
| 271 | roll $4, RT0d; \ |
| 272 | orl RT0d, val##d; |
| 273 | |
| 274 | #define initial_permutation3(left, right) \ |
| 275 | do_permutation(left##0d, right##0d, 4, 0x0f0f0f0f); \ |
| 276 | do_permutation(left##0d, right##0d, 16, 0x0000ffff); \ |
| 277 | do_permutation(left##1d, right##1d, 4, 0x0f0f0f0f); \ |
| 278 | do_permutation(left##1d, right##1d, 16, 0x0000ffff); \ |
| 279 | do_permutation(left##2d, right##2d, 4, 0x0f0f0f0f); \ |
| 280 | do_permutation(left##2d, right##2d, 16, 0x0000ffff); \ |
| 281 | \ |
| 282 | do_permutation(right##0d, left##0d, 2, 0x33333333); \ |
| 283 | do_permutation(right##0d, left##0d, 8, 0x00ff00ff); \ |
| 284 | do_permutation(right##1d, left##1d, 2, 0x33333333); \ |
| 285 | do_permutation(right##1d, left##1d, 8, 0x00ff00ff); \ |
| 286 | do_permutation(right##2d, left##2d, 2, 0x33333333); \ |
| 287 | do_permutation(right##2d, left##2d, 8, 0x00ff00ff); \ |
| 288 | \ |
| 289 | movabs $0x3f3f3f3f3f3f3f3f, RT3; \ |
| 290 | \ |
| 291 | movl left##0d, RW0d; \ |
| 292 | roll $1, right##0d; \ |
| 293 | xorl right##0d, RW0d; \ |
| 294 | andl $0xaaaaaaaa, RW0d; \ |
| 295 | xorl RW0d, left##0d; \ |
| 296 | xorl RW0d, right##0d; \ |
| 297 | roll $1, left##0d; \ |
| 298 | expand_to_64bits(right##0, RT3); \ |
| 299 | expand_to_64bits(left##0, RT3); \ |
| 300 | movl left##1d, RW1d; \ |
| 301 | roll $1, right##1d; \ |
| 302 | xorl right##1d, RW1d; \ |
| 303 | andl $0xaaaaaaaa, RW1d; \ |
| 304 | xorl RW1d, left##1d; \ |
| 305 | xorl RW1d, right##1d; \ |
| 306 | roll $1, left##1d; \ |
| 307 | expand_to_64bits(right##1, RT3); \ |
| 308 | expand_to_64bits(left##1, RT3); \ |
| 309 | movl left##2d, RW2d; \ |
| 310 | roll $1, right##2d; \ |
| 311 | xorl right##2d, RW2d; \ |
| 312 | andl $0xaaaaaaaa, RW2d; \ |
| 313 | xorl RW2d, left##2d; \ |
| 314 | xorl RW2d, right##2d; \ |
| 315 | roll $1, left##2d; \ |
| 316 | expand_to_64bits(right##2, RT3); \ |
| 317 | expand_to_64bits(left##2, RT3); |
| 318 | |
| 319 | #define final_permutation3(left, right) \ |
| 320 | compress_to_64bits(right##0); \ |
| 321 | compress_to_64bits(left##0); \ |
| 322 | movl right##0d, RW0d; \ |
| 323 | rorl $1, left##0d; \ |
| 324 | xorl left##0d, RW0d; \ |
| 325 | andl $0xaaaaaaaa, RW0d; \ |
| 326 | xorl RW0d, right##0d; \ |
| 327 | xorl RW0d, left##0d; \ |
| 328 | rorl $1, right##0d; \ |
| 329 | compress_to_64bits(right##1); \ |
| 330 | compress_to_64bits(left##1); \ |
| 331 | movl right##1d, RW1d; \ |
| 332 | rorl $1, left##1d; \ |
| 333 | xorl left##1d, RW1d; \ |
| 334 | andl $0xaaaaaaaa, RW1d; \ |
| 335 | xorl RW1d, right##1d; \ |
| 336 | xorl RW1d, left##1d; \ |
| 337 | rorl $1, right##1d; \ |
| 338 | compress_to_64bits(right##2); \ |
| 339 | compress_to_64bits(left##2); \ |
| 340 | movl right##2d, RW2d; \ |
| 341 | rorl $1, left##2d; \ |
| 342 | xorl left##2d, RW2d; \ |
| 343 | andl $0xaaaaaaaa, RW2d; \ |
| 344 | xorl RW2d, right##2d; \ |
| 345 | xorl RW2d, left##2d; \ |
| 346 | rorl $1, right##2d; \ |
| 347 | \ |
| 348 | do_permutation(right##0d, left##0d, 8, 0x00ff00ff); \ |
| 349 | do_permutation(right##0d, left##0d, 2, 0x33333333); \ |
| 350 | do_permutation(right##1d, left##1d, 8, 0x00ff00ff); \ |
| 351 | do_permutation(right##1d, left##1d, 2, 0x33333333); \ |
| 352 | do_permutation(right##2d, left##2d, 8, 0x00ff00ff); \ |
| 353 | do_permutation(right##2d, left##2d, 2, 0x33333333); \ |
| 354 | \ |
| 355 | do_permutation(left##0d, right##0d, 16, 0x0000ffff); \ |
| 356 | do_permutation(left##0d, right##0d, 4, 0x0f0f0f0f); \ |
| 357 | do_permutation(left##1d, right##1d, 16, 0x0000ffff); \ |
| 358 | do_permutation(left##1d, right##1d, 4, 0x0f0f0f0f); \ |
| 359 | do_permutation(left##2d, right##2d, 16, 0x0000ffff); \ |
| 360 | do_permutation(left##2d, right##2d, 4, 0x0f0f0f0f); |
| 361 | |
| 362 | #define round3(n, from, to, load_next_key, do_movq) \ |
| 363 | xorq from##0, RW0; \ |
| 364 | movzbl RW0bl, RT3d; \ |
| 365 | movzbl RW0bh, RT1d; \ |
| 366 | shrq $16, RW0; \ |
| 367 | xorq s8(, RT3, 8), to##0; \ |
| 368 | xorq s6(, RT1, 8), to##0; \ |
| 369 | movzbl RW0bl, RT3d; \ |
| 370 | movzbl RW0bh, RT1d; \ |
| 371 | shrq $16, RW0; \ |
| 372 | xorq s4(, RT3, 8), to##0; \ |
| 373 | xorq s2(, RT1, 8), to##0; \ |
| 374 | movzbl RW0bl, RT3d; \ |
| 375 | movzbl RW0bh, RT1d; \ |
| 376 | shrl $16, RW0d; \ |
| 377 | xorq s7(, RT3, 8), to##0; \ |
| 378 | xorq s5(, RT1, 8), to##0; \ |
| 379 | movzbl RW0bl, RT3d; \ |
| 380 | movzbl RW0bh, RT1d; \ |
| 381 | load_next_key(n, RW0); \ |
| 382 | xorq s3(, RT3, 8), to##0; \ |
| 383 | xorq s1(, RT1, 8), to##0; \ |
| 384 | xorq from##1, RW1; \ |
| 385 | movzbl RW1bl, RT3d; \ |
| 386 | movzbl RW1bh, RT1d; \ |
| 387 | shrq $16, RW1; \ |
| 388 | xorq s8(, RT3, 8), to##1; \ |
| 389 | xorq s6(, RT1, 8), to##1; \ |
| 390 | movzbl RW1bl, RT3d; \ |
| 391 | movzbl RW1bh, RT1d; \ |
| 392 | shrq $16, RW1; \ |
| 393 | xorq s4(, RT3, 8), to##1; \ |
| 394 | xorq s2(, RT1, 8), to##1; \ |
| 395 | movzbl RW1bl, RT3d; \ |
| 396 | movzbl RW1bh, RT1d; \ |
| 397 | shrl $16, RW1d; \ |
| 398 | xorq s7(, RT3, 8), to##1; \ |
| 399 | xorq s5(, RT1, 8), to##1; \ |
| 400 | movzbl RW1bl, RT3d; \ |
| 401 | movzbl RW1bh, RT1d; \ |
| 402 | do_movq(RW0, RW1); \ |
| 403 | xorq s3(, RT3, 8), to##1; \ |
| 404 | xorq s1(, RT1, 8), to##1; \ |
| 405 | xorq from##2, RW2; \ |
| 406 | movzbl RW2bl, RT3d; \ |
| 407 | movzbl RW2bh, RT1d; \ |
| 408 | shrq $16, RW2; \ |
| 409 | xorq s8(, RT3, 8), to##2; \ |
| 410 | xorq s6(, RT1, 8), to##2; \ |
| 411 | movzbl RW2bl, RT3d; \ |
| 412 | movzbl RW2bh, RT1d; \ |
| 413 | shrq $16, RW2; \ |
| 414 | xorq s4(, RT3, 8), to##2; \ |
| 415 | xorq s2(, RT1, 8), to##2; \ |
| 416 | movzbl RW2bl, RT3d; \ |
| 417 | movzbl RW2bh, RT1d; \ |
| 418 | shrl $16, RW2d; \ |
| 419 | xorq s7(, RT3, 8), to##2; \ |
| 420 | xorq s5(, RT1, 8), to##2; \ |
| 421 | movzbl RW2bl, RT3d; \ |
| 422 | movzbl RW2bh, RT1d; \ |
| 423 | do_movq(RW0, RW2); \ |
| 424 | xorq s3(, RT3, 8), to##2; \ |
| 425 | xorq s1(, RT1, 8), to##2; |
| 426 | |
| 427 | #define __movq(src, dst) \ |
| 428 | movq src, dst; |
| 429 | |
| 430 | ENTRY(des3_ede_x86_64_crypt_blk_3way) |
| 431 | /* input: |
| 432 | * %rdi: ctx, round keys |
| 433 | * %rsi: dst (3 blocks) |
| 434 | * %rdx: src (3 blocks) |
| 435 | */ |
| 436 | |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 437 | pushq %rbx; |
| 438 | pushq %r12; |
| 439 | pushq %r13; |
| 440 | pushq %r14; |
| 441 | pushq %r15; |
| 442 | |
Josh Poimboeuf | 3ed7b4d | 2017-09-18 14:42:04 -0500 | [diff] [blame] | 443 | pushq %rsi /* dst */ |
| 444 | |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 445 | /* load input */ |
| 446 | movl 0 * 4(%rdx), RL0d; |
| 447 | movl 1 * 4(%rdx), RR0d; |
| 448 | movl 2 * 4(%rdx), RL1d; |
| 449 | movl 3 * 4(%rdx), RR1d; |
| 450 | movl 4 * 4(%rdx), RL2d; |
| 451 | movl 5 * 4(%rdx), RR2d; |
| 452 | |
| 453 | bswapl RL0d; |
| 454 | bswapl RR0d; |
| 455 | bswapl RL1d; |
| 456 | bswapl RR1d; |
| 457 | bswapl RL2d; |
| 458 | bswapl RR2d; |
| 459 | |
| 460 | initial_permutation3(RL, RR); |
| 461 | |
| 462 | movq 0(CTX), RW0; |
| 463 | movq RW0, RW1; |
| 464 | movq RW0, RW2; |
| 465 | |
| 466 | round3(0, RR, RL, load_next_key, __movq); |
| 467 | round3(1, RL, RR, load_next_key, __movq); |
| 468 | round3(2, RR, RL, load_next_key, __movq); |
| 469 | round3(3, RL, RR, load_next_key, __movq); |
| 470 | round3(4, RR, RL, load_next_key, __movq); |
| 471 | round3(5, RL, RR, load_next_key, __movq); |
| 472 | round3(6, RR, RL, load_next_key, __movq); |
| 473 | round3(7, RL, RR, load_next_key, __movq); |
| 474 | round3(8, RR, RL, load_next_key, __movq); |
| 475 | round3(9, RL, RR, load_next_key, __movq); |
| 476 | round3(10, RR, RL, load_next_key, __movq); |
| 477 | round3(11, RL, RR, load_next_key, __movq); |
| 478 | round3(12, RR, RL, load_next_key, __movq); |
| 479 | round3(13, RL, RR, load_next_key, __movq); |
| 480 | round3(14, RR, RL, load_next_key, __movq); |
| 481 | round3(15, RL, RR, load_next_key, __movq); |
| 482 | |
| 483 | round3(16+0, RL, RR, load_next_key, __movq); |
| 484 | round3(16+1, RR, RL, load_next_key, __movq); |
| 485 | round3(16+2, RL, RR, load_next_key, __movq); |
| 486 | round3(16+3, RR, RL, load_next_key, __movq); |
| 487 | round3(16+4, RL, RR, load_next_key, __movq); |
| 488 | round3(16+5, RR, RL, load_next_key, __movq); |
| 489 | round3(16+6, RL, RR, load_next_key, __movq); |
| 490 | round3(16+7, RR, RL, load_next_key, __movq); |
| 491 | round3(16+8, RL, RR, load_next_key, __movq); |
| 492 | round3(16+9, RR, RL, load_next_key, __movq); |
| 493 | round3(16+10, RL, RR, load_next_key, __movq); |
| 494 | round3(16+11, RR, RL, load_next_key, __movq); |
| 495 | round3(16+12, RL, RR, load_next_key, __movq); |
| 496 | round3(16+13, RR, RL, load_next_key, __movq); |
| 497 | round3(16+14, RL, RR, load_next_key, __movq); |
| 498 | round3(16+15, RR, RL, load_next_key, __movq); |
| 499 | |
| 500 | round3(32+0, RR, RL, load_next_key, __movq); |
| 501 | round3(32+1, RL, RR, load_next_key, __movq); |
| 502 | round3(32+2, RR, RL, load_next_key, __movq); |
| 503 | round3(32+3, RL, RR, load_next_key, __movq); |
| 504 | round3(32+4, RR, RL, load_next_key, __movq); |
| 505 | round3(32+5, RL, RR, load_next_key, __movq); |
| 506 | round3(32+6, RR, RL, load_next_key, __movq); |
| 507 | round3(32+7, RL, RR, load_next_key, __movq); |
| 508 | round3(32+8, RR, RL, load_next_key, __movq); |
| 509 | round3(32+9, RL, RR, load_next_key, __movq); |
| 510 | round3(32+10, RR, RL, load_next_key, __movq); |
| 511 | round3(32+11, RL, RR, load_next_key, __movq); |
| 512 | round3(32+12, RR, RL, load_next_key, __movq); |
| 513 | round3(32+13, RL, RR, load_next_key, __movq); |
| 514 | round3(32+14, RR, RL, load_next_key, __movq); |
| 515 | round3(32+15, RL, RR, dummy2, dummy2); |
| 516 | |
| 517 | final_permutation3(RR, RL); |
| 518 | |
| 519 | bswapl RR0d; |
| 520 | bswapl RL0d; |
| 521 | bswapl RR1d; |
| 522 | bswapl RL1d; |
| 523 | bswapl RR2d; |
| 524 | bswapl RL2d; |
| 525 | |
Josh Poimboeuf | 3ed7b4d | 2017-09-18 14:42:04 -0500 | [diff] [blame] | 526 | popq %rsi /* dst */ |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 527 | movl RR0d, 0 * 4(%rsi); |
| 528 | movl RL0d, 1 * 4(%rsi); |
| 529 | movl RR1d, 2 * 4(%rsi); |
| 530 | movl RL1d, 3 * 4(%rsi); |
| 531 | movl RR2d, 4 * 4(%rsi); |
| 532 | movl RL2d, 5 * 4(%rsi); |
| 533 | |
| 534 | popq %r15; |
| 535 | popq %r14; |
| 536 | popq %r13; |
| 537 | popq %r12; |
| 538 | popq %rbx; |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 539 | |
| 540 | ret; |
| 541 | ENDPROC(des3_ede_x86_64_crypt_blk_3way) |
| 542 | |
Denys Vlasenko | e183914 | 2017-01-19 22:33:04 +0100 | [diff] [blame] | 543 | .section .rodata, "a", @progbits |
Jussi Kivilinna | 6574e6c | 2014-06-09 20:59:54 +0300 | [diff] [blame] | 544 | .align 16 |
| 545 | .L_s1: |
| 546 | .quad 0x0010100001010400, 0x0000000000000000 |
| 547 | .quad 0x0000100000010000, 0x0010100001010404 |
| 548 | .quad 0x0010100001010004, 0x0000100000010404 |
| 549 | .quad 0x0000000000000004, 0x0000100000010000 |
| 550 | .quad 0x0000000000000400, 0x0010100001010400 |
| 551 | .quad 0x0010100001010404, 0x0000000000000400 |
| 552 | .quad 0x0010000001000404, 0x0010100001010004 |
| 553 | .quad 0x0010000001000000, 0x0000000000000004 |
| 554 | .quad 0x0000000000000404, 0x0010000001000400 |
| 555 | .quad 0x0010000001000400, 0x0000100000010400 |
| 556 | .quad 0x0000100000010400, 0x0010100001010000 |
| 557 | .quad 0x0010100001010000, 0x0010000001000404 |
| 558 | .quad 0x0000100000010004, 0x0010000001000004 |
| 559 | .quad 0x0010000001000004, 0x0000100000010004 |
| 560 | .quad 0x0000000000000000, 0x0000000000000404 |
| 561 | .quad 0x0000100000010404, 0x0010000001000000 |
| 562 | .quad 0x0000100000010000, 0x0010100001010404 |
| 563 | .quad 0x0000000000000004, 0x0010100001010000 |
| 564 | .quad 0x0010100001010400, 0x0010000001000000 |
| 565 | .quad 0x0010000001000000, 0x0000000000000400 |
| 566 | .quad 0x0010100001010004, 0x0000100000010000 |
| 567 | .quad 0x0000100000010400, 0x0010000001000004 |
| 568 | .quad 0x0000000000000400, 0x0000000000000004 |
| 569 | .quad 0x0010000001000404, 0x0000100000010404 |
| 570 | .quad 0x0010100001010404, 0x0000100000010004 |
| 571 | .quad 0x0010100001010000, 0x0010000001000404 |
| 572 | .quad 0x0010000001000004, 0x0000000000000404 |
| 573 | .quad 0x0000100000010404, 0x0010100001010400 |
| 574 | .quad 0x0000000000000404, 0x0010000001000400 |
| 575 | .quad 0x0010000001000400, 0x0000000000000000 |
| 576 | .quad 0x0000100000010004, 0x0000100000010400 |
| 577 | .quad 0x0000000000000000, 0x0010100001010004 |
| 578 | .L_s2: |
| 579 | .quad 0x0801080200100020, 0x0800080000000000 |
| 580 | .quad 0x0000080000000000, 0x0001080200100020 |
| 581 | .quad 0x0001000000100000, 0x0000000200000020 |
| 582 | .quad 0x0801000200100020, 0x0800080200000020 |
| 583 | .quad 0x0800000200000020, 0x0801080200100020 |
| 584 | .quad 0x0801080000100000, 0x0800000000000000 |
| 585 | .quad 0x0800080000000000, 0x0001000000100000 |
| 586 | .quad 0x0000000200000020, 0x0801000200100020 |
| 587 | .quad 0x0001080000100000, 0x0001000200100020 |
| 588 | .quad 0x0800080200000020, 0x0000000000000000 |
| 589 | .quad 0x0800000000000000, 0x0000080000000000 |
| 590 | .quad 0x0001080200100020, 0x0801000000100000 |
| 591 | .quad 0x0001000200100020, 0x0800000200000020 |
| 592 | .quad 0x0000000000000000, 0x0001080000100000 |
| 593 | .quad 0x0000080200000020, 0x0801080000100000 |
| 594 | .quad 0x0801000000100000, 0x0000080200000020 |
| 595 | .quad 0x0000000000000000, 0x0001080200100020 |
| 596 | .quad 0x0801000200100020, 0x0001000000100000 |
| 597 | .quad 0x0800080200000020, 0x0801000000100000 |
| 598 | .quad 0x0801080000100000, 0x0000080000000000 |
| 599 | .quad 0x0801000000100000, 0x0800080000000000 |
| 600 | .quad 0x0000000200000020, 0x0801080200100020 |
| 601 | .quad 0x0001080200100020, 0x0000000200000020 |
| 602 | .quad 0x0000080000000000, 0x0800000000000000 |
| 603 | .quad 0x0000080200000020, 0x0801080000100000 |
| 604 | .quad 0x0001000000100000, 0x0800000200000020 |
| 605 | .quad 0x0001000200100020, 0x0800080200000020 |
| 606 | .quad 0x0800000200000020, 0x0001000200100020 |
| 607 | .quad 0x0001080000100000, 0x0000000000000000 |
| 608 | .quad 0x0800080000000000, 0x0000080200000020 |
| 609 | .quad 0x0800000000000000, 0x0801000200100020 |
| 610 | .quad 0x0801080200100020, 0x0001080000100000 |
| 611 | .L_s3: |
| 612 | .quad 0x0000002000000208, 0x0000202008020200 |
| 613 | .quad 0x0000000000000000, 0x0000200008020008 |
| 614 | .quad 0x0000002008000200, 0x0000000000000000 |
| 615 | .quad 0x0000202000020208, 0x0000002008000200 |
| 616 | .quad 0x0000200000020008, 0x0000000008000008 |
| 617 | .quad 0x0000000008000008, 0x0000200000020000 |
| 618 | .quad 0x0000202008020208, 0x0000200000020008 |
| 619 | .quad 0x0000200008020000, 0x0000002000000208 |
| 620 | .quad 0x0000000008000000, 0x0000000000000008 |
| 621 | .quad 0x0000202008020200, 0x0000002000000200 |
| 622 | .quad 0x0000202000020200, 0x0000200008020000 |
| 623 | .quad 0x0000200008020008, 0x0000202000020208 |
| 624 | .quad 0x0000002008000208, 0x0000202000020200 |
| 625 | .quad 0x0000200000020000, 0x0000002008000208 |
| 626 | .quad 0x0000000000000008, 0x0000202008020208 |
| 627 | .quad 0x0000002000000200, 0x0000000008000000 |
| 628 | .quad 0x0000202008020200, 0x0000000008000000 |
| 629 | .quad 0x0000200000020008, 0x0000002000000208 |
| 630 | .quad 0x0000200000020000, 0x0000202008020200 |
| 631 | .quad 0x0000002008000200, 0x0000000000000000 |
| 632 | .quad 0x0000002000000200, 0x0000200000020008 |
| 633 | .quad 0x0000202008020208, 0x0000002008000200 |
| 634 | .quad 0x0000000008000008, 0x0000002000000200 |
| 635 | .quad 0x0000000000000000, 0x0000200008020008 |
| 636 | .quad 0x0000002008000208, 0x0000200000020000 |
| 637 | .quad 0x0000000008000000, 0x0000202008020208 |
| 638 | .quad 0x0000000000000008, 0x0000202000020208 |
| 639 | .quad 0x0000202000020200, 0x0000000008000008 |
| 640 | .quad 0x0000200008020000, 0x0000002008000208 |
| 641 | .quad 0x0000002000000208, 0x0000200008020000 |
| 642 | .quad 0x0000202000020208, 0x0000000000000008 |
| 643 | .quad 0x0000200008020008, 0x0000202000020200 |
| 644 | .L_s4: |
| 645 | .quad 0x1008020000002001, 0x1000020800002001 |
| 646 | .quad 0x1000020800002001, 0x0000000800000000 |
| 647 | .quad 0x0008020800002000, 0x1008000800000001 |
| 648 | .quad 0x1008000000000001, 0x1000020000002001 |
| 649 | .quad 0x0000000000000000, 0x0008020000002000 |
| 650 | .quad 0x0008020000002000, 0x1008020800002001 |
| 651 | .quad 0x1000000800000001, 0x0000000000000000 |
| 652 | .quad 0x0008000800000000, 0x1008000000000001 |
| 653 | .quad 0x1000000000000001, 0x0000020000002000 |
| 654 | .quad 0x0008000000000000, 0x1008020000002001 |
| 655 | .quad 0x0000000800000000, 0x0008000000000000 |
| 656 | .quad 0x1000020000002001, 0x0000020800002000 |
| 657 | .quad 0x1008000800000001, 0x1000000000000001 |
| 658 | .quad 0x0000020800002000, 0x0008000800000000 |
| 659 | .quad 0x0000020000002000, 0x0008020800002000 |
| 660 | .quad 0x1008020800002001, 0x1000000800000001 |
| 661 | .quad 0x0008000800000000, 0x1008000000000001 |
| 662 | .quad 0x0008020000002000, 0x1008020800002001 |
| 663 | .quad 0x1000000800000001, 0x0000000000000000 |
| 664 | .quad 0x0000000000000000, 0x0008020000002000 |
| 665 | .quad 0x0000020800002000, 0x0008000800000000 |
| 666 | .quad 0x1008000800000001, 0x1000000000000001 |
| 667 | .quad 0x1008020000002001, 0x1000020800002001 |
| 668 | .quad 0x1000020800002001, 0x0000000800000000 |
| 669 | .quad 0x1008020800002001, 0x1000000800000001 |
| 670 | .quad 0x1000000000000001, 0x0000020000002000 |
| 671 | .quad 0x1008000000000001, 0x1000020000002001 |
| 672 | .quad 0x0008020800002000, 0x1008000800000001 |
| 673 | .quad 0x1000020000002001, 0x0000020800002000 |
| 674 | .quad 0x0008000000000000, 0x1008020000002001 |
| 675 | .quad 0x0000000800000000, 0x0008000000000000 |
| 676 | .quad 0x0000020000002000, 0x0008020800002000 |
| 677 | .L_s5: |
| 678 | .quad 0x0000001000000100, 0x0020001002080100 |
| 679 | .quad 0x0020000002080000, 0x0420001002000100 |
| 680 | .quad 0x0000000000080000, 0x0000001000000100 |
| 681 | .quad 0x0400000000000000, 0x0020000002080000 |
| 682 | .quad 0x0400001000080100, 0x0000000000080000 |
| 683 | .quad 0x0020001002000100, 0x0400001000080100 |
| 684 | .quad 0x0420001002000100, 0x0420000002080000 |
| 685 | .quad 0x0000001000080100, 0x0400000000000000 |
| 686 | .quad 0x0020000002000000, 0x0400000000080000 |
| 687 | .quad 0x0400000000080000, 0x0000000000000000 |
| 688 | .quad 0x0400001000000100, 0x0420001002080100 |
| 689 | .quad 0x0420001002080100, 0x0020001002000100 |
| 690 | .quad 0x0420000002080000, 0x0400001000000100 |
| 691 | .quad 0x0000000000000000, 0x0420000002000000 |
| 692 | .quad 0x0020001002080100, 0x0020000002000000 |
| 693 | .quad 0x0420000002000000, 0x0000001000080100 |
| 694 | .quad 0x0000000000080000, 0x0420001002000100 |
| 695 | .quad 0x0000001000000100, 0x0020000002000000 |
| 696 | .quad 0x0400000000000000, 0x0020000002080000 |
| 697 | .quad 0x0420001002000100, 0x0400001000080100 |
| 698 | .quad 0x0020001002000100, 0x0400000000000000 |
| 699 | .quad 0x0420000002080000, 0x0020001002080100 |
| 700 | .quad 0x0400001000080100, 0x0000001000000100 |
| 701 | .quad 0x0020000002000000, 0x0420000002080000 |
| 702 | .quad 0x0420001002080100, 0x0000001000080100 |
| 703 | .quad 0x0420000002000000, 0x0420001002080100 |
| 704 | .quad 0x0020000002080000, 0x0000000000000000 |
| 705 | .quad 0x0400000000080000, 0x0420000002000000 |
| 706 | .quad 0x0000001000080100, 0x0020001002000100 |
| 707 | .quad 0x0400001000000100, 0x0000000000080000 |
| 708 | .quad 0x0000000000000000, 0x0400000000080000 |
| 709 | .quad 0x0020001002080100, 0x0400001000000100 |
| 710 | .L_s6: |
| 711 | .quad 0x0200000120000010, 0x0204000020000000 |
| 712 | .quad 0x0000040000000000, 0x0204040120000010 |
| 713 | .quad 0x0204000020000000, 0x0000000100000010 |
| 714 | .quad 0x0204040120000010, 0x0004000000000000 |
| 715 | .quad 0x0200040020000000, 0x0004040100000010 |
| 716 | .quad 0x0004000000000000, 0x0200000120000010 |
| 717 | .quad 0x0004000100000010, 0x0200040020000000 |
| 718 | .quad 0x0200000020000000, 0x0000040100000010 |
| 719 | .quad 0x0000000000000000, 0x0004000100000010 |
| 720 | .quad 0x0200040120000010, 0x0000040000000000 |
| 721 | .quad 0x0004040000000000, 0x0200040120000010 |
| 722 | .quad 0x0000000100000010, 0x0204000120000010 |
| 723 | .quad 0x0204000120000010, 0x0000000000000000 |
| 724 | .quad 0x0004040100000010, 0x0204040020000000 |
| 725 | .quad 0x0000040100000010, 0x0004040000000000 |
| 726 | .quad 0x0204040020000000, 0x0200000020000000 |
| 727 | .quad 0x0200040020000000, 0x0000000100000010 |
| 728 | .quad 0x0204000120000010, 0x0004040000000000 |
| 729 | .quad 0x0204040120000010, 0x0004000000000000 |
| 730 | .quad 0x0000040100000010, 0x0200000120000010 |
| 731 | .quad 0x0004000000000000, 0x0200040020000000 |
| 732 | .quad 0x0200000020000000, 0x0000040100000010 |
| 733 | .quad 0x0200000120000010, 0x0204040120000010 |
| 734 | .quad 0x0004040000000000, 0x0204000020000000 |
| 735 | .quad 0x0004040100000010, 0x0204040020000000 |
| 736 | .quad 0x0000000000000000, 0x0204000120000010 |
| 737 | .quad 0x0000000100000010, 0x0000040000000000 |
| 738 | .quad 0x0204000020000000, 0x0004040100000010 |
| 739 | .quad 0x0000040000000000, 0x0004000100000010 |
| 740 | .quad 0x0200040120000010, 0x0000000000000000 |
| 741 | .quad 0x0204040020000000, 0x0200000020000000 |
| 742 | .quad 0x0004000100000010, 0x0200040120000010 |
| 743 | .L_s7: |
| 744 | .quad 0x0002000000200000, 0x2002000004200002 |
| 745 | .quad 0x2000000004000802, 0x0000000000000000 |
| 746 | .quad 0x0000000000000800, 0x2000000004000802 |
| 747 | .quad 0x2002000000200802, 0x0002000004200800 |
| 748 | .quad 0x2002000004200802, 0x0002000000200000 |
| 749 | .quad 0x0000000000000000, 0x2000000004000002 |
| 750 | .quad 0x2000000000000002, 0x0000000004000000 |
| 751 | .quad 0x2002000004200002, 0x2000000000000802 |
| 752 | .quad 0x0000000004000800, 0x2002000000200802 |
| 753 | .quad 0x2002000000200002, 0x0000000004000800 |
| 754 | .quad 0x2000000004000002, 0x0002000004200000 |
| 755 | .quad 0x0002000004200800, 0x2002000000200002 |
| 756 | .quad 0x0002000004200000, 0x0000000000000800 |
| 757 | .quad 0x2000000000000802, 0x2002000004200802 |
| 758 | .quad 0x0002000000200800, 0x2000000000000002 |
| 759 | .quad 0x0000000004000000, 0x0002000000200800 |
| 760 | .quad 0x0000000004000000, 0x0002000000200800 |
| 761 | .quad 0x0002000000200000, 0x2000000004000802 |
| 762 | .quad 0x2000000004000802, 0x2002000004200002 |
| 763 | .quad 0x2002000004200002, 0x2000000000000002 |
| 764 | .quad 0x2002000000200002, 0x0000000004000000 |
| 765 | .quad 0x0000000004000800, 0x0002000000200000 |
| 766 | .quad 0x0002000004200800, 0x2000000000000802 |
| 767 | .quad 0x2002000000200802, 0x0002000004200800 |
| 768 | .quad 0x2000000000000802, 0x2000000004000002 |
| 769 | .quad 0x2002000004200802, 0x0002000004200000 |
| 770 | .quad 0x0002000000200800, 0x0000000000000000 |
| 771 | .quad 0x2000000000000002, 0x2002000004200802 |
| 772 | .quad 0x0000000000000000, 0x2002000000200802 |
| 773 | .quad 0x0002000004200000, 0x0000000000000800 |
| 774 | .quad 0x2000000004000002, 0x0000000004000800 |
| 775 | .quad 0x0000000000000800, 0x2002000000200002 |
| 776 | .L_s8: |
| 777 | .quad 0x0100010410001000, 0x0000010000001000 |
| 778 | .quad 0x0000000000040000, 0x0100010410041000 |
| 779 | .quad 0x0100000010000000, 0x0100010410001000 |
| 780 | .quad 0x0000000400000000, 0x0100000010000000 |
| 781 | .quad 0x0000000400040000, 0x0100000010040000 |
| 782 | .quad 0x0100010410041000, 0x0000010000041000 |
| 783 | .quad 0x0100010010041000, 0x0000010400041000 |
| 784 | .quad 0x0000010000001000, 0x0000000400000000 |
| 785 | .quad 0x0100000010040000, 0x0100000410000000 |
| 786 | .quad 0x0100010010001000, 0x0000010400001000 |
| 787 | .quad 0x0000010000041000, 0x0000000400040000 |
| 788 | .quad 0x0100000410040000, 0x0100010010041000 |
| 789 | .quad 0x0000010400001000, 0x0000000000000000 |
| 790 | .quad 0x0000000000000000, 0x0100000410040000 |
| 791 | .quad 0x0100000410000000, 0x0100010010001000 |
| 792 | .quad 0x0000010400041000, 0x0000000000040000 |
| 793 | .quad 0x0000010400041000, 0x0000000000040000 |
| 794 | .quad 0x0100010010041000, 0x0000010000001000 |
| 795 | .quad 0x0000000400000000, 0x0100000410040000 |
| 796 | .quad 0x0000010000001000, 0x0000010400041000 |
| 797 | .quad 0x0100010010001000, 0x0000000400000000 |
| 798 | .quad 0x0100000410000000, 0x0100000010040000 |
| 799 | .quad 0x0100000410040000, 0x0100000010000000 |
| 800 | .quad 0x0000000000040000, 0x0100010410001000 |
| 801 | .quad 0x0000000000000000, 0x0100010410041000 |
| 802 | .quad 0x0000000400040000, 0x0100000410000000 |
| 803 | .quad 0x0100000010040000, 0x0100010010001000 |
| 804 | .quad 0x0100010410001000, 0x0000000000000000 |
| 805 | .quad 0x0100010410041000, 0x0000010000041000 |
| 806 | .quad 0x0000010000041000, 0x0000010400001000 |
| 807 | .quad 0x0000010400001000, 0x0000000400040000 |
| 808 | .quad 0x0100000010000000, 0x0100010010041000 |