projects / thirdparty / openssl.git / blame
| Commit | Line | Data |
|---|---|---|
| 6aa36e8e RS |
1 | #! /usr/bin/env perl |
| 2 | # Copyright 2012-2016 The OpenSSL Project Authors. All Rights Reserved. | |
| 3 | # | |
| 5593d9c9 | 4 | # Licensed under the Apache License 2.0 (the "License"). You may not use |
| 6aa36e8e RS |
5 | # this file except in compliance with the License. You can obtain a copy |
| 6 | # in the file LICENSE in the source distribution or at | |
| 7 | # https://www.openssl.org/source/license.html | |
| 8 | ||
| 4739ccdb AP |
9 | |
| 10 | # ==================================================================== | |
| e3713c36 RS |
11 | # Written by David S. Miller and Andy Polyakov. |
| 12 | # The module is licensed under 2-clause BSD | |
| 4739ccdb AP |
13 | # license. October 2012. All rights reserved. |
| 14 | # ==================================================================== | |
| 15 | ||
| 16 | ###################################################################### | |
| 17 | # Camellia for SPARC T4. | |
| 18 | # | |
| 19 | # As with AES below results [for aligned data] are virtually identical | |
| 46f4e1be | 20 | # to critical path lengths for 3-cycle instruction latency: |
| 4739ccdb AP |
21 | # |
| 22 | # 128-bit key 192/256- | |
| 23 | # CBC encrypt 4.14/4.21(*) 5.46/5.52 | |
| 24 | # (*) numbers after slash are for | |
| 25 | # misaligned data. | |
| 26 | # | |
| 27 | # As with Intel AES-NI, question is if it's possible to improve | |
| 46f4e1be | 28 | # performance of parallelizable modes by interleaving round |
| 4739ccdb AP |
29 | # instructions. In Camellia every instruction is dependent on |
| 30 | # previous, which means that there is place for 2 additional ones | |
| 31 | # in between two dependent. Can we expect 3x performance improvement? | |
| 32 | # At least one can argue that it should be possible to break 2x | |
| 33 | # barrier... For some reason not even 2x appears to be possible: | |
| 34 | # | |
| 35 | # 128-bit key 192/256- | |
| 36 | # CBC decrypt 2.21/2.74 2.99/3.40 | |
| 37 | # CTR 2.15/2.68(*) 2.93/3.34 | |
| 38 | # (*) numbers after slash are for | |
| 39 | # misaligned data. | |
| 40 | # | |
| 41 | # This is for 2x interleave. But compared to 1x interleave CBC decrypt | |
| 42 | # improved by ... 0% for 128-bit key, and 11% for 192/256-bit one. | |
| 43 | # So that out-of-order execution logic can take non-interleaved code | |
| 44 | # to 1.87x, but can't take 2x interleaved one any further. There | |
| 45 | # surely is some explanation... As result 3x interleave was not even | |
| 46 | # attempted. Instead an effort was made to share specific modes | |
| 47 | # implementations with AES module (therefore sparct4_modes.pl). | |
| 48 | # | |
| 49 | # To anchor to something else, software C implementation processes | |
| 50 | # one byte in 38 cycles with 128-bit key on same processor. | |
| 51 | ||
| 52 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; | |
| 53 | push(@INC,"${dir}","${dir}../../perlasm"); | |
| 54 | require "sparcv9_modes.pl"; | |
| 55 | ||
| 1aa89a7a | 56 | $output = pop and open STDOUT,">$output"; |
| 4739ccdb AP |
57 | |
| 58 | $::evp=1; # if $evp is set to 0, script generates module with | |
| 59 | # Camellia_[en|de]crypt, Camellia_set_key and Camellia_cbc_encrypt | |
| 60 | # entry points. These are fully compatible with openssl/camellia.h. | |
| 61 | ||
| 62 | ###################################################################### | |
| 63 | # single-round subroutines | |
| 64 | # | |
| 65 | { | |
| 66 | my ($inp,$out,$key,$rounds,$tmp,$mask)=map("%o$_",(0..5)); | |
| 67 | ||
| 68 | $code=<<___; | |
| eb77e888 AP |
69 | #include "sparc_arch.h" |
| 70 | ||
| 4739ccdb AP |
71 | .text |
| 72 | ||
| 73 | .globl cmll_t4_encrypt | |
| 74 | .align 32 | |
| 75 | cmll_t4_encrypt: | |
| 76 | andcc $inp, 7, %g1 ! is input aligned? | |
| 77 | andn $inp, 7, $inp | |
| 78 | ||
| 79 | ldx [$key + 0], %g4 | |
| 80 | ldx [$key + 8], %g5 | |
| 81 | ||
| 82 | ldx [$inp + 0], %o4 | |
| 83 | bz,pt %icc, 1f | |
| 84 | ldx [$inp + 8], %o5 | |
| 85 | ldx [$inp + 16], $inp | |
| 86 | sll %g1, 3, %g1 | |
| 87 | sub %g0, %g1, %o3 | |
| 88 | sllx %o4, %g1, %o4 | |
| 89 | sllx %o5, %g1, %g1 | |
| 90 | srlx %o5, %o3, %o5 | |
| 91 | srlx $inp, %o3, %o3 | |
| 92 | or %o5, %o4, %o4 | |
| 93 | or %o3, %g1, %o5 | |
| 94 | 1: | |
| 95 | ld [$key + 272], $rounds ! grandRounds, 3 or 4 | |
| 96 | ldd [$key + 16], %f12 | |
| 97 | ldd [$key + 24], %f14 | |
| 98 | xor %g4, %o4, %o4 | |
| 99 | xor %g5, %o5, %o5 | |
| 100 | ldd [$key + 32], %f16 | |
| 101 | ldd [$key + 40], %f18 | |
| 102 | movxtod %o4, %f0 | |
| 103 | movxtod %o5, %f2 | |
| 104 | ldd [$key + 48], %f20 | |
| 105 | ldd [$key + 56], %f22 | |
| 106 | sub $rounds, 1, $rounds | |
| 107 | ldd [$key + 64], %f24 | |
| 108 | ldd [$key + 72], %f26 | |
| 109 | add $key, 80, $key | |
| 110 | ||
| 111 | .Lenc: | |
| 112 | camellia_f %f12, %f2, %f0, %f2 | |
| 113 | ldd [$key + 0], %f12 | |
| 114 | sub $rounds,1,$rounds | |
| 115 | camellia_f %f14, %f0, %f2, %f0 | |
| 116 | ldd [$key + 8], %f14 | |
| 117 | camellia_f %f16, %f2, %f0, %f2 | |
| 118 | ldd [$key + 16], %f16 | |
| 119 | camellia_f %f18, %f0, %f2, %f0 | |
| 120 | ldd [$key + 24], %f18 | |
| 121 | camellia_f %f20, %f2, %f0, %f2 | |
| 122 | ldd [$key + 32], %f20 | |
| 123 | camellia_f %f22, %f0, %f2, %f0 | |
| 124 | ldd [$key + 40], %f22 | |
| 125 | camellia_fl %f24, %f0, %f0 | |
| 126 | ldd [$key + 48], %f24 | |
| 127 | camellia_fli %f26, %f2, %f2 | |
| 128 | ldd [$key + 56], %f26 | |
| 129 | brnz,pt $rounds, .Lenc | |
| 130 | add $key, 64, $key | |
| 131 | ||
| 132 | andcc $out, 7, $tmp ! is output aligned? | |
| 133 | camellia_f %f12, %f2, %f0, %f2 | |
| 134 | camellia_f %f14, %f0, %f2, %f0 | |
| 135 | camellia_f %f16, %f2, %f0, %f2 | |
| 136 | camellia_f %f18, %f0, %f2, %f0 | |
| 137 | camellia_f %f20, %f2, %f0, %f4 | |
| 138 | camellia_f %f22, %f0, %f4, %f2 | |
| 139 | fxor %f24, %f4, %f0 | |
| 140 | fxor %f26, %f2, %f2 | |
| 141 | ||
| 142 | bnz,pn %icc, 2f | |
| 143 | nop | |
| 144 | ||
| 145 | std %f0, [$out + 0] | |
| 146 | retl | |
| 147 | std %f2, [$out + 8] | |
| 148 | ||
| 149 | 2: alignaddrl $out, %g0, $out | |
| 150 | mov 0xff, $mask | |
| 151 | srl $mask, $tmp, $mask | |
| 152 | ||
| 153 | faligndata %f0, %f0, %f4 | |
| 154 | faligndata %f0, %f2, %f6 | |
| 155 | faligndata %f2, %f2, %f8 | |
| 156 | ||
| 157 | stda %f4, [$out + $mask]0xc0 ! partial store | |
| 158 | std %f6, [$out + 8] | |
| 159 | add $out, 16, $out | |
| 160 | orn %g0, $mask, $mask | |
| 161 | retl | |
| 162 | stda %f8, [$out + $mask]0xc0 ! partial store | |
| 163 | .type cmll_t4_encrypt,#function | |
| 164 | .size cmll_t4_encrypt,.-cmll_t4_encrypt | |
| 165 | ||
| 166 | .globl cmll_t4_decrypt | |
| 167 | .align 32 | |
| 168 | cmll_t4_decrypt: | |
| 169 | ld [$key + 272], $rounds ! grandRounds, 3 or 4 | |
| 170 | andcc $inp, 7, %g1 ! is input aligned? | |
| 171 | andn $inp, 7, $inp | |
| 172 | ||
| 173 | sll $rounds, 6, $rounds | |
| 174 | add $rounds, $key, $key | |
| 175 | ||
| 176 | ldx [$inp + 0], %o4 | |
| 177 | bz,pt %icc, 1f | |
| 178 | ldx [$inp + 8], %o5 | |
| 179 | ldx [$inp + 16], $inp | |
| 180 | sll %g1, 3, %g1 | |
| 181 | sub %g0, %g1, %g4 | |
| 182 | sllx %o4, %g1, %o4 | |
| 183 | sllx %o5, %g1, %g1 | |
| 184 | srlx %o5, %g4, %o5 | |
| 185 | srlx $inp, %g4, %g4 | |
| 186 | or %o5, %o4, %o4 | |
| 187 | or %g4, %g1, %o5 | |
| 188 | 1: | |
| 189 | ldx [$key + 0], %g4 | |
| 190 | ldx [$key + 8], %g5 | |
| 191 | ldd [$key - 8], %f12 | |
| 192 | ldd [$key - 16], %f14 | |
| 193 | xor %g4, %o4, %o4 | |
| 194 | xor %g5, %o5, %o5 | |
| 195 | ldd [$key - 24], %f16 | |
| 196 | ldd [$key - 32], %f18 | |
| 197 | movxtod %o4, %f0 | |
| 198 | movxtod %o5, %f2 | |
| 199 | ldd [$key - 40], %f20 | |
| 200 | ldd [$key - 48], %f22 | |
| 201 | sub $rounds, 64, $rounds | |
| 202 | ldd [$key - 56], %f24 | |
| 203 | ldd [$key - 64], %f26 | |
| 204 | sub $key, 64, $key | |
| 205 | ||
| 206 | .Ldec: | |
| 207 | camellia_f %f12, %f2, %f0, %f2 | |
| 208 | ldd [$key - 8], %f12 | |
| 209 | sub $rounds, 64, $rounds | |
| 210 | camellia_f %f14, %f0, %f2, %f0 | |
| 211 | ldd [$key - 16], %f14 | |
| 212 | camellia_f %f16, %f2, %f0, %f2 | |
| 213 | ldd [$key - 24], %f16 | |
| 214 | camellia_f %f18, %f0, %f2, %f0 | |
| 215 | ldd [$key - 32], %f18 | |
| 216 | camellia_f %f20, %f2, %f0, %f2 | |
| 217 | ldd [$key - 40], %f20 | |
| 218 | camellia_f %f22, %f0, %f2, %f0 | |
| 219 | ldd [$key - 48], %f22 | |
| 220 | camellia_fl %f24, %f0, %f0 | |
| 221 | ldd [$key - 56], %f24 | |
| 222 | camellia_fli %f26, %f2, %f2 | |
| 223 | ldd [$key - 64], %f26 | |
| 224 | brnz,pt $rounds, .Ldec | |
| 225 | sub $key, 64, $key | |
| 226 | ||
| 227 | andcc $out, 7, $tmp ! is output aligned? | |
| 228 | camellia_f %f12, %f2, %f0, %f2 | |
| 229 | camellia_f %f14, %f0, %f2, %f0 | |
| 230 | camellia_f %f16, %f2, %f0, %f2 | |
| 231 | camellia_f %f18, %f0, %f2, %f0 | |
| 232 | camellia_f %f20, %f2, %f0, %f4 | |
| 233 | camellia_f %f22, %f0, %f4, %f2 | |
| 234 | fxor %f26, %f4, %f0 | |
| 235 | fxor %f24, %f2, %f2 | |
| 236 | ||
| 237 | bnz,pn %icc, 2f | |
| 238 | nop | |
| 239 | ||
| 240 | std %f0, [$out + 0] | |
| 241 | retl | |
| 242 | std %f2, [$out + 8] | |
| 243 | ||
| 244 | 2: alignaddrl $out, %g0, $out | |
| 245 | mov 0xff, $mask | |
| 246 | srl $mask, $tmp, $mask | |
| 247 | ||
| 248 | faligndata %f0, %f0, %f4 | |
| 249 | faligndata %f0, %f2, %f6 | |
| 250 | faligndata %f2, %f2, %f8 | |
| 251 | ||
| 252 | stda %f4, [$out + $mask]0xc0 ! partial store | |
| 253 | std %f6, [$out + 8] | |
| 254 | add $out, 16, $out | |
| 255 | orn %g0, $mask, $mask | |
| 256 | retl | |
| 257 | stda %f8, [$out + $mask]0xc0 ! partial store | |
| 258 | .type cmll_t4_decrypt,#function | |
| 259 | .size cmll_t4_decrypt,.-cmll_t4_decrypt | |
| 260 | ___ | |
| 261 | } | |
| 262 | ||
| 263 | ###################################################################### | |
| 264 | # key setup subroutines | |
| 265 | # | |
| 266 | { | |
| 267 | sub ROTL128 { | |
| 268 | my $rot = shift; | |
| 269 | ||
| 270 | "srlx %o4, 64-$rot, %g4\n\t". | |
| 271 | "sllx %o4, $rot, %o4\n\t". | |
| 272 | "srlx %o5, 64-$rot, %g5\n\t". | |
| 273 | "sllx %o5, $rot, %o5\n\t". | |
| 274 | "or %o4, %g5, %o4\n\t". | |
| 275 | "or %o5, %g4, %o5"; | |
| 276 | } | |
| 277 | ||
| 278 | my ($inp,$bits,$out,$tmp)=map("%o$_",(0..5)); | |
| 279 | $code.=<<___; | |
| 280 | .globl cmll_t4_set_key | |
| 281 | .align 32 | |
| 282 | cmll_t4_set_key: | |
| 283 | and $inp, 7, $tmp | |
| 284 | alignaddr $inp, %g0, $inp | |
| 285 | cmp $bits, 192 | |
| 286 | ldd [$inp + 0], %f0 | |
| 287 | bl,pt %icc,.L128 | |
| 288 | ldd [$inp + 8], %f2 | |
| 289 | ||
| 290 | be,pt %icc,.L192 | |
| 291 | ldd [$inp + 16], %f4 | |
| 292 | ||
| 293 | brz,pt $tmp, .L256aligned | |
| 294 | ldd [$inp + 24], %f6 | |
| 295 | ||
| 296 | ldd [$inp + 32], %f8 | |
| 297 | faligndata %f0, %f2, %f0 | |
| 298 | faligndata %f2, %f4, %f2 | |
| 299 | faligndata %f4, %f6, %f4 | |
| 300 | b .L256aligned | |
| 301 | faligndata %f6, %f8, %f6 | |
| 302 | ||
| 303 | .align 16 | |
| 304 | .L192: | |
| 305 | brz,a,pt $tmp, .L256aligned | |
| 306 | fnot2 %f4, %f6 | |
| 307 | ||
| 308 | ldd [$inp + 24], %f6 | |
| 309 | nop | |
| 310 | faligndata %f0, %f2, %f0 | |
| 311 | faligndata %f2, %f4, %f2 | |
| 312 | faligndata %f4, %f6, %f4 | |
| 313 | fnot2 %f4, %f6 | |
| 314 | ||
| 315 | .L256aligned: | |
| 316 | std %f0, [$out + 0] ! k[0, 1] | |
| 317 | fsrc2 %f0, %f28 | |
| 318 | std %f2, [$out + 8] ! k[2, 3] | |
| 319 | fsrc2 %f2, %f30 | |
| 320 | fxor %f4, %f0, %f0 | |
| 321 | b .L128key | |
| 322 | fxor %f6, %f2, %f2 | |
| 323 | ||
| 324 | .align 16 | |
| 325 | .L128: | |
| 326 | brz,pt $tmp, .L128aligned | |
| 327 | nop | |
| 328 | ||
| 329 | ldd [$inp + 16], %f4 | |
| 330 | nop | |
| 331 | faligndata %f0, %f2, %f0 | |
| 332 | faligndata %f2, %f4, %f2 | |
| 333 | ||
| 334 | .L128aligned: | |
| 335 | std %f0, [$out + 0] ! k[0, 1] | |
| 336 | fsrc2 %f0, %f28 | |
| 337 | std %f2, [$out + 8] ! k[2, 3] | |
| 338 | fsrc2 %f2, %f30 | |
| 339 | ||
| 340 | .L128key: | |
| 341 | mov %o7, %o5 | |
| 342 | 1: call .+8 | |
| 343 | add %o7, SIGMA-1b, %o4 | |
| 344 | mov %o5, %o7 | |
| 345 | ||
| 346 | ldd [%o4 + 0], %f16 | |
| 347 | ldd [%o4 + 8], %f18 | |
| 348 | ldd [%o4 + 16], %f20 | |
| 349 | ldd [%o4 + 24], %f22 | |
| 350 | ||
| 351 | camellia_f %f16, %f2, %f0, %f2 | |
| 352 | camellia_f %f18, %f0, %f2, %f0 | |
| 353 | fxor %f28, %f0, %f0 | |
| 354 | fxor %f30, %f2, %f2 | |
| 355 | camellia_f %f20, %f2, %f0, %f2 | |
| 356 | camellia_f %f22, %f0, %f2, %f0 | |
| 357 | ||
| 358 | bge,pn %icc, .L256key | |
| 359 | nop | |
| 360 | std %f0, [$out + 0x10] ! k[ 4, 5] | |
| 361 | std %f2, [$out + 0x18] ! k[ 6, 7] | |
| 362 | ||
| 363 | movdtox %f0, %o4 | |
| 364 | movdtox %f2, %o5 | |
| 365 | `&ROTL128(15)` | |
| 366 | stx %o4, [$out + 0x30] ! k[12, 13] | |
| 367 | stx %o5, [$out + 0x38] ! k[14, 15] | |
| 368 | `&ROTL128(15)` | |
| 369 | stx %o4, [$out + 0x40] ! k[16, 17] | |
| 370 | stx %o5, [$out + 0x48] ! k[18, 19] | |
| 371 | `&ROTL128(15)` | |
| 372 | stx %o4, [$out + 0x60] ! k[24, 25] | |
| 373 | `&ROTL128(15)` | |
| 374 | stx %o4, [$out + 0x70] ! k[28, 29] | |
| 375 | stx %o5, [$out + 0x78] ! k[30, 31] | |
| 376 | `&ROTL128(34)` | |
| 377 | stx %o4, [$out + 0xa0] ! k[40, 41] | |
| 378 | stx %o5, [$out + 0xa8] ! k[42, 43] | |
| 379 | `&ROTL128(17)` | |
| 380 | stx %o4, [$out + 0xc0] ! k[48, 49] | |
| 381 | stx %o5, [$out + 0xc8] ! k[50, 51] | |
| 382 | ||
| 383 | movdtox %f28, %o4 ! k[ 0, 1] | |
| 384 | movdtox %f30, %o5 ! k[ 2, 3] | |
| 385 | `&ROTL128(15)` | |
| 386 | stx %o4, [$out + 0x20] ! k[ 8, 9] | |
| 387 | stx %o5, [$out + 0x28] ! k[10, 11] | |
| 388 | `&ROTL128(30)` | |
| 389 | stx %o4, [$out + 0x50] ! k[20, 21] | |
| 390 | stx %o5, [$out + 0x58] ! k[22, 23] | |
| 391 | `&ROTL128(15)` | |
| 392 | stx %o5, [$out + 0x68] ! k[26, 27] | |
| 393 | `&ROTL128(17)` | |
| 394 | stx %o4, [$out + 0x80] ! k[32, 33] | |
| 395 | stx %o5, [$out + 0x88] ! k[34, 35] | |
| 396 | `&ROTL128(17)` | |
| 397 | stx %o4, [$out + 0x90] ! k[36, 37] | |
| 398 | stx %o5, [$out + 0x98] ! k[38, 39] | |
| 399 | `&ROTL128(17)` | |
| 400 | stx %o4, [$out + 0xb0] ! k[44, 45] | |
| 401 | stx %o5, [$out + 0xb8] ! k[46, 47] | |
| 402 | ||
| 403 | mov 3, $tmp | |
| 404 | st $tmp, [$out + 0x110] | |
| 405 | retl | |
| 406 | xor %o0, %o0, %o0 | |
| 407 | ||
| 408 | .align 16 | |
| 409 | .L256key: | |
| 410 | ldd [%o4 + 32], %f24 | |
| 411 | ldd [%o4 + 40], %f26 | |
| 412 | ||
| 413 | std %f0, [$out + 0x30] ! k[12, 13] | |
| 414 | std %f2, [$out + 0x38] ! k[14, 15] | |
| 415 | ||
| 416 | fxor %f4, %f0, %f0 | |
| 417 | fxor %f6, %f2, %f2 | |
| 418 | camellia_f %f24, %f2, %f0, %f2 | |
| 419 | camellia_f %f26, %f0, %f2, %f0 | |
| 420 | ||
| 421 | std %f0, [$out + 0x10] ! k[ 4, 5] | |
| 422 | std %f2, [$out + 0x18] ! k[ 6, 7] | |
| 423 | ||
| 424 | movdtox %f0, %o4 | |
| 425 | movdtox %f2, %o5 | |
| 426 | `&ROTL128(30)` | |
| 427 | stx %o4, [$out + 0x50] ! k[20, 21] | |
| 428 | stx %o5, [$out + 0x58] ! k[22, 23] | |
| 429 | `&ROTL128(30)` | |
| 430 | stx %o4, [$out + 0xa0] ! k[40, 41] | |
| 431 | stx %o5, [$out + 0xa8] ! k[42, 43] | |
| 432 | `&ROTL128(51)` | |
| 433 | stx %o4, [$out + 0x100] ! k[64, 65] | |
| 434 | stx %o5, [$out + 0x108] ! k[66, 67] | |
| 435 | ||
| 436 | movdtox %f4, %o4 ! k[ 8, 9] | |
| 437 | movdtox %f6, %o5 ! k[10, 11] | |
| 438 | `&ROTL128(15)` | |
| 439 | stx %o4, [$out + 0x20] ! k[ 8, 9] | |
| 440 | stx %o5, [$out + 0x28] ! k[10, 11] | |
| 441 | `&ROTL128(15)` | |
| 442 | stx %o4, [$out + 0x40] ! k[16, 17] | |
| 443 | stx %o5, [$out + 0x48] ! k[18, 19] | |
| 444 | `&ROTL128(30)` | |
| 445 | stx %o4, [$out + 0x90] ! k[36, 37] | |
| 446 | stx %o5, [$out + 0x98] ! k[38, 39] | |
| 447 | `&ROTL128(34)` | |
| 448 | stx %o4, [$out + 0xd0] ! k[52, 53] | |
| 449 | stx %o5, [$out + 0xd8] ! k[54, 55] | |
| 450 | ldx [$out + 0x30], %o4 ! k[12, 13] | |
| 451 | ldx [$out + 0x38], %o5 ! k[14, 15] | |
| 452 | `&ROTL128(15)` | |
| 453 | stx %o4, [$out + 0x30] ! k[12, 13] | |
| 454 | stx %o5, [$out + 0x38] ! k[14, 15] | |
| 455 | `&ROTL128(30)` | |
| 456 | stx %o4, [$out + 0x70] ! k[28, 29] | |
| 457 | stx %o5, [$out + 0x78] ! k[30, 31] | |
| 458 | srlx %o4, 32, %g4 | |
| 459 | srlx %o5, 32, %g5 | |
| 460 | st %o4, [$out + 0xc0] ! k[48] | |
| 461 | st %g5, [$out + 0xc4] ! k[49] | |
| 462 | st %o5, [$out + 0xc8] ! k[50] | |
| 463 | st %g4, [$out + 0xcc] ! k[51] | |
| 464 | `&ROTL128(49)` | |
| 465 | stx %o4, [$out + 0xe0] ! k[56, 57] | |
| 466 | stx %o5, [$out + 0xe8] ! k[58, 59] | |
| 467 | ||
| 468 | movdtox %f28, %o4 ! k[ 0, 1] | |
| 469 | movdtox %f30, %o5 ! k[ 2, 3] | |
| 470 | `&ROTL128(45)` | |
| 471 | stx %o4, [$out + 0x60] ! k[24, 25] | |
| 472 | stx %o5, [$out + 0x68] ! k[26, 27] | |
| 473 | `&ROTL128(15)` | |
| 474 | stx %o4, [$out + 0x80] ! k[32, 33] | |
| 475 | stx %o5, [$out + 0x88] ! k[34, 35] | |
| 476 | `&ROTL128(17)` | |
| 477 | stx %o4, [$out + 0xb0] ! k[44, 45] | |
| 478 | stx %o5, [$out + 0xb8] ! k[46, 47] | |
| 479 | `&ROTL128(34)` | |
| 480 | stx %o4, [$out + 0xf0] ! k[60, 61] | |
| 481 | stx %o5, [$out + 0xf8] ! k[62, 63] | |
| 482 | ||
| 483 | mov 4, $tmp | |
| 484 | st $tmp, [$out + 0x110] | |
| 485 | retl | |
| 486 | xor %o0, %o0, %o0 | |
| 487 | .type cmll_t4_set_key,#function | |
| 488 | .size cmll_t4_set_key,.-cmll_t4_set_key | |
| 489 | .align 32 | |
| 490 | SIGMA: | |
| 491 | .long 0xa09e667f, 0x3bcc908b, 0xb67ae858, 0x4caa73b2 | |
| 492 | .long 0xc6ef372f, 0xe94f82be, 0x54ff53a5, 0xf1d36f1c | |
| 493 | .long 0x10e527fa, 0xde682d1d, 0xb05688c2, 0xb3e6c1fd | |
| 494 | .type SIGMA,#object | |
| 495 | .size SIGMA,.-SIGMA | |
| 496 | .asciz "Camellia for SPARC T4, David S. Miller, Andy Polyakov" | |
| 497 | ___ | |
| 498 | } | |
| 499 | ||
| 500 | {{{ | |
| 501 | my ($inp,$out,$len,$key,$ivec,$enc)=map("%i$_",(0..5)); | |
| 502 | my ($ileft,$iright,$ooff,$omask,$ivoff)=map("%l$_",(1..7)); | |
| 503 | ||
| 504 | $code.=<<___; | |
| 505 | .align 32 | |
| 506 | _cmll128_load_enckey: | |
| 507 | ldx [$key + 0], %g4 | |
| 508 | ldx [$key + 8], %g5 | |
| 509 | ___ | |
| 510 | for ($i=2; $i<26;$i++) { # load key schedule | |
| 511 | $code.=<<___; | |
| 512 | ldd [$key + `8*$i`], %f`12+2*$i` | |
| 513 | ___ | |
| 514 | } | |
| 515 | $code.=<<___; | |
| 516 | retl | |
| 517 | nop | |
| 518 | .type _cmll128_load_enckey,#function | |
| 519 | .size _cmll128_load_enckey,.-_cmll128_load_enckey | |
| 520 | _cmll256_load_enckey=_cmll128_load_enckey | |
| 521 | ||
| 522 | .align 32 | |
| 523 | _cmll256_load_deckey: | |
| 524 | ldd [$key + 64], %f62 | |
| 525 | ldd [$key + 72], %f60 | |
| 526 | b .Load_deckey | |
| 527 | add $key, 64, $key | |
| 528 | _cmll128_load_deckey: | |
| 529 | ldd [$key + 0], %f60 | |
| 530 | ldd [$key + 8], %f62 | |
| 531 | .Load_deckey: | |
| 532 | ___ | |
| 533 | for ($i=2; $i<24;$i++) { # load key schedule | |
| 534 | $code.=<<___; | |
| 535 | ldd [$key + `8*$i`], %f`62-2*$i` | |
| 536 | ___ | |
| 537 | } | |
| 538 | $code.=<<___; | |
| 539 | ldx [$key + 192], %g4 | |
| 540 | retl | |
| 541 | ldx [$key + 200], %g5 | |
| 542 | .type _cmll256_load_deckey,#function | |
| 543 | .size _cmll256_load_deckey,.-_cmll256_load_deckey | |
| 544 | ||
| 545 | .align 32 | |
| 546 | _cmll128_encrypt_1x: | |
| 547 | ___ | |
| 548 | for ($i=0; $i<3; $i++) { | |
| 549 | $code.=<<___; | |
| 550 | camellia_f %f`16+16*$i+0`, %f2, %f0, %f2 | |
| 551 | camellia_f %f`16+16*$i+2`, %f0, %f2, %f0 | |
| 552 | camellia_f %f`16+16*$i+4`, %f2, %f0, %f2 | |
| 553 | camellia_f %f`16+16*$i+6`, %f0, %f2, %f0 | |
| 554 | ___ | |
| 555 | $code.=<<___ if ($i<2); | |
| 556 | camellia_f %f`16+16*$i+8`, %f2, %f0, %f2 | |
| 557 | camellia_f %f`16+16*$i+10`, %f0, %f2, %f0 | |
| 558 | camellia_fl %f`16+16*$i+12`, %f0, %f0 | |
| 559 | camellia_fli %f`16+16*$i+14`, %f2, %f2 | |
| 560 | ___ | |
| 561 | } | |
| 562 | $code.=<<___; | |
| 563 | camellia_f %f56, %f2, %f0, %f4 | |
| 564 | camellia_f %f58, %f0, %f4, %f2 | |
| 565 | fxor %f60, %f4, %f0 | |
| 566 | retl | |
| 567 | fxor %f62, %f2, %f2 | |
| 568 | .type _cmll128_encrypt_1x,#function | |
| 569 | .size _cmll128_encrypt_1x,.-_cmll128_encrypt_1x | |
| 570 | _cmll128_decrypt_1x=_cmll128_encrypt_1x | |
| 571 | ||
| 572 | .align 32 | |
| 573 | _cmll128_encrypt_2x: | |
| 574 | ___ | |
| 575 | for ($i=0; $i<3; $i++) { | |
| 576 | $code.=<<___; | |
| 577 | camellia_f %f`16+16*$i+0`, %f2, %f0, %f2 | |
| 578 | camellia_f %f`16+16*$i+0`, %f6, %f4, %f6 | |
| 579 | camellia_f %f`16+16*$i+2`, %f0, %f2, %f0 | |
| 580 | camellia_f %f`16+16*$i+2`, %f4, %f6, %f4 | |
| 581 | camellia_f %f`16+16*$i+4`, %f2, %f0, %f2 | |
| 582 | camellia_f %f`16+16*$i+4`, %f6, %f4, %f6 | |
| 583 | camellia_f %f`16+16*$i+6`, %f0, %f2, %f0 | |
| 584 | camellia_f %f`16+16*$i+6`, %f4, %f6, %f4 | |
| 585 | ___ | |
| 586 | $code.=<<___ if ($i<2); | |
| 587 | camellia_f %f`16+16*$i+8`, %f2, %f0, %f2 | |
| 588 | camellia_f %f`16+16*$i+8`, %f6, %f4, %f6 | |
| 589 | camellia_f %f`16+16*$i+10`, %f0, %f2, %f0 | |
| 590 | camellia_f %f`16+16*$i+10`, %f4, %f6, %f4 | |
| 591 | camellia_fl %f`16+16*$i+12`, %f0, %f0 | |
| 592 | camellia_fl %f`16+16*$i+12`, %f4, %f4 | |
| 593 | camellia_fli %f`16+16*$i+14`, %f2, %f2 | |
| 594 | camellia_fli %f`16+16*$i+14`, %f6, %f6 | |
| 595 | ___ | |
| 596 | } | |
| 597 | $code.=<<___; | |
| 598 | camellia_f %f56, %f2, %f0, %f8 | |
| 599 | camellia_f %f56, %f6, %f4, %f10 | |
| 600 | camellia_f %f58, %f0, %f8, %f2 | |
| 601 | camellia_f %f58, %f4, %f10, %f6 | |
| 602 | fxor %f60, %f8, %f0 | |
| 603 | fxor %f60, %f10, %f4 | |
| 604 | fxor %f62, %f2, %f2 | |
| 605 | retl | |
| 606 | fxor %f62, %f6, %f6 | |
| 607 | .type _cmll128_encrypt_2x,#function | |
| 608 | .size _cmll128_encrypt_2x,.-_cmll128_encrypt_2x | |
| 609 | _cmll128_decrypt_2x=_cmll128_encrypt_2x | |
| 610 | ||
| 611 | .align 32 | |
| 612 | _cmll256_encrypt_1x: | |
| 613 | camellia_f %f16, %f2, %f0, %f2 | |
| 614 | camellia_f %f18, %f0, %f2, %f0 | |
| 615 | ldd [$key + 208], %f16 | |
| 616 | ldd [$key + 216], %f18 | |
| 617 | camellia_f %f20, %f2, %f0, %f2 | |
| 618 | camellia_f %f22, %f0, %f2, %f0 | |
| 619 | ldd [$key + 224], %f20 | |
| 620 | ldd [$key + 232], %f22 | |
| 621 | camellia_f %f24, %f2, %f0, %f2 | |
| 622 | camellia_f %f26, %f0, %f2, %f0 | |
| 623 | ldd [$key + 240], %f24 | |
| 624 | ldd [$key + 248], %f26 | |
| 625 | camellia_fl %f28, %f0, %f0 | |
| 626 | camellia_fli %f30, %f2, %f2 | |
| 627 | ldd [$key + 256], %f28 | |
| 628 | ldd [$key + 264], %f30 | |
| 629 | ___ | |
| 630 | for ($i=1; $i<3; $i++) { | |
| 631 | $code.=<<___; | |
| 632 | camellia_f %f`16+16*$i+0`, %f2, %f0, %f2 | |
| 633 | camellia_f %f`16+16*$i+2`, %f0, %f2, %f0 | |
| 634 | camellia_f %f`16+16*$i+4`, %f2, %f0, %f2 | |
| 635 | camellia_f %f`16+16*$i+6`, %f0, %f2, %f0 | |
| 636 | camellia_f %f`16+16*$i+8`, %f2, %f0, %f2 | |
| 637 | camellia_f %f`16+16*$i+10`, %f0, %f2, %f0 | |
| 638 | camellia_fl %f`16+16*$i+12`, %f0, %f0 | |
| 639 | camellia_fli %f`16+16*$i+14`, %f2, %f2 | |
| 640 | ___ | |
| 641 | } | |
| 642 | $code.=<<___; | |
| 643 | camellia_f %f16, %f2, %f0, %f2 | |
| 644 | camellia_f %f18, %f0, %f2, %f0 | |
| 645 | ldd [$key + 16], %f16 | |
| 646 | ldd [$key + 24], %f18 | |
| 647 | camellia_f %f20, %f2, %f0, %f2 | |
| 648 | camellia_f %f22, %f0, %f2, %f0 | |
| 649 | ldd [$key + 32], %f20 | |
| 650 | ldd [$key + 40], %f22 | |
| 651 | camellia_f %f24, %f2, %f0, %f4 | |
| 652 | camellia_f %f26, %f0, %f4, %f2 | |
| 653 | ldd [$key + 48], %f24 | |
| 654 | ldd [$key + 56], %f26 | |
| 655 | fxor %f28, %f4, %f0 | |
| 656 | fxor %f30, %f2, %f2 | |
| 657 | ldd [$key + 64], %f28 | |
| 658 | retl | |
| 659 | ldd [$key + 72], %f30 | |
| 660 | .type _cmll256_encrypt_1x,#function | |
| 661 | .size _cmll256_encrypt_1x,.-_cmll256_encrypt_1x | |
| 662 | ||
| 663 | .align 32 | |
| 664 | _cmll256_encrypt_2x: | |
| 665 | camellia_f %f16, %f2, %f0, %f2 | |
| 666 | camellia_f %f16, %f6, %f4, %f6 | |
| 667 | camellia_f %f18, %f0, %f2, %f0 | |
| 668 | camellia_f %f18, %f4, %f6, %f4 | |
| 669 | ldd [$key + 208], %f16 | |
| 670 | ldd [$key + 216], %f18 | |
| 671 | camellia_f %f20, %f2, %f0, %f2 | |
| 672 | camellia_f %f20, %f6, %f4, %f6 | |
| 673 | camellia_f %f22, %f0, %f2, %f0 | |
| 674 | camellia_f %f22, %f4, %f6, %f4 | |
| 675 | ldd [$key + 224], %f20 | |
| 676 | ldd [$key + 232], %f22 | |
| 677 | camellia_f %f24, %f2, %f0, %f2 | |
| 678 | camellia_f %f24, %f6, %f4, %f6 | |
| 679 | camellia_f %f26, %f0, %f2, %f0 | |
| 680 | camellia_f %f26, %f4, %f6, %f4 | |
| 681 | ldd [$key + 240], %f24 | |
| 682 | ldd [$key + 248], %f26 | |
| 683 | camellia_fl %f28, %f0, %f0 | |
| 684 | camellia_fl %f28, %f4, %f4 | |
| 685 | camellia_fli %f30, %f2, %f2 | |
| 686 | camellia_fli %f30, %f6, %f6 | |
| 687 | ldd [$key + 256], %f28 | |
| 688 | ldd [$key + 264], %f30 | |
| 689 | ___ | |
| 690 | for ($i=1; $i<3; $i++) { | |
| 691 | $code.=<<___; | |
| 692 | camellia_f %f`16+16*$i+0`, %f2, %f0, %f2 | |
| 693 | camellia_f %f`16+16*$i+0`, %f6, %f4, %f6 | |
| 694 | camellia_f %f`16+16*$i+2`, %f0, %f2, %f0 | |
| 695 | camellia_f %f`16+16*$i+2`, %f4, %f6, %f4 | |
| 696 | camellia_f %f`16+16*$i+4`, %f2, %f0, %f2 | |
| 697 | camellia_f %f`16+16*$i+4`, %f6, %f4, %f6 | |
| 698 | camellia_f %f`16+16*$i+6`, %f0, %f2, %f0 | |
| 699 | camellia_f %f`16+16*$i+6`, %f4, %f6, %f4 | |
| 700 | camellia_f %f`16+16*$i+8`, %f2, %f0, %f2 | |
| 701 | camellia_f %f`16+16*$i+8`, %f6, %f4, %f6 | |
| 702 | camellia_f %f`16+16*$i+10`, %f0, %f2, %f0 | |
| 703 | camellia_f %f`16+16*$i+10`, %f4, %f6, %f4 | |
| 704 | camellia_fl %f`16+16*$i+12`, %f0, %f0 | |
| 705 | camellia_fl %f`16+16*$i+12`, %f4, %f4 | |
| 706 | camellia_fli %f`16+16*$i+14`, %f2, %f2 | |
| 707 | camellia_fli %f`16+16*$i+14`, %f6, %f6 | |
| 708 | ___ | |
| 709 | } | |
| 710 | $code.=<<___; | |
| 711 | camellia_f %f16, %f2, %f0, %f2 | |
| 712 | camellia_f %f16, %f6, %f4, %f6 | |
| 713 | camellia_f %f18, %f0, %f2, %f0 | |
| 714 | camellia_f %f18, %f4, %f6, %f4 | |
| 715 | ldd [$key + 16], %f16 | |
| 716 | ldd [$key + 24], %f18 | |
| 717 | camellia_f %f20, %f2, %f0, %f2 | |
| 718 | camellia_f %f20, %f6, %f4, %f6 | |
| 719 | camellia_f %f22, %f0, %f2, %f0 | |
| 720 | camellia_f %f22, %f4, %f6, %f4 | |
| 721 | ldd [$key + 32], %f20 | |
| 722 | ldd [$key + 40], %f22 | |
| 723 | camellia_f %f24, %f2, %f0, %f8 | |
| 724 | camellia_f %f24, %f6, %f4, %f10 | |
| 725 | camellia_f %f26, %f0, %f8, %f2 | |
| 726 | camellia_f %f26, %f4, %f10, %f6 | |
| 727 | ldd [$key + 48], %f24 | |
| 728 | ldd [$key + 56], %f26 | |
| 729 | fxor %f28, %f8, %f0 | |
| 730 | fxor %f28, %f10, %f4 | |
| 731 | fxor %f30, %f2, %f2 | |
| 732 | fxor %f30, %f6, %f6 | |
| 733 | ldd [$key + 64], %f28 | |
| 734 | retl | |
| 735 | ldd [$key + 72], %f30 | |
| 736 | .type _cmll256_encrypt_2x,#function | |
| 737 | .size _cmll256_encrypt_2x,.-_cmll256_encrypt_2x | |
| 738 | ||
| 739 | .align 32 | |
| 740 | _cmll256_decrypt_1x: | |
| 741 | camellia_f %f16, %f2, %f0, %f2 | |
| 742 | camellia_f %f18, %f0, %f2, %f0 | |
| 743 | ldd [$key - 8], %f16 | |
| 744 | ldd [$key - 16], %f18 | |
| 745 | camellia_f %f20, %f2, %f0, %f2 | |
| 746 | camellia_f %f22, %f0, %f2, %f0 | |
| 747 | ldd [$key - 24], %f20 | |
| 748 | ldd [$key - 32], %f22 | |
| 749 | camellia_f %f24, %f2, %f0, %f2 | |
| 750 | camellia_f %f26, %f0, %f2, %f0 | |
| 751 | ldd [$key - 40], %f24 | |
| 752 | ldd [$key - 48], %f26 | |
| 753 | camellia_fl %f28, %f0, %f0 | |
| 754 | camellia_fli %f30, %f2, %f2 | |
| 755 | ldd [$key - 56], %f28 | |
| 756 | ldd [$key - 64], %f30 | |
| 757 | ___ | |
| 758 | for ($i=1; $i<3; $i++) { | |
| 759 | $code.=<<___; | |
| 760 | camellia_f %f`16+16*$i+0`, %f2, %f0, %f2 | |
| 761 | camellia_f %f`16+16*$i+2`, %f0, %f2, %f0 | |
| 762 | camellia_f %f`16+16*$i+4`, %f2, %f0, %f2 | |
| 763 | camellia_f %f`16+16*$i+6`, %f0, %f2, %f0 | |
| 764 | camellia_f %f`16+16*$i+8`, %f2, %f0, %f2 | |
| 765 | camellia_f %f`16+16*$i+10`, %f0, %f2, %f0 | |
| 766 | camellia_fl %f`16+16*$i+12`, %f0, %f0 | |
| 767 | camellia_fli %f`16+16*$i+14`, %f2, %f2 | |
| 768 | ___ | |
| 769 | } | |
| 770 | $code.=<<___; | |
| 771 | camellia_f %f16, %f2, %f0, %f2 | |
| 772 | camellia_f %f18, %f0, %f2, %f0 | |
| 773 | ldd [$key + 184], %f16 | |
| 774 | ldd [$key + 176], %f18 | |
| 775 | camellia_f %f20, %f2, %f0, %f2 | |
| 776 | camellia_f %f22, %f0, %f2, %f0 | |
| 777 | ldd [$key + 168], %f20 | |
| 778 | ldd [$key + 160], %f22 | |
| 779 | camellia_f %f24, %f2, %f0, %f4 | |
| 780 | camellia_f %f26, %f0, %f4, %f2 | |
| 781 | ldd [$key + 152], %f24 | |
| 782 | ldd [$key + 144], %f26 | |
| 783 | fxor %f30, %f4, %f0 | |
| 784 | fxor %f28, %f2, %f2 | |
| 785 | ldd [$key + 136], %f28 | |
| 786 | retl | |
| 787 | ldd [$key + 128], %f30 | |
| 788 | .type _cmll256_decrypt_1x,#function | |
| 789 | .size _cmll256_decrypt_1x,.-_cmll256_decrypt_1x | |
| 790 | ||
| 791 | .align 32 | |
| 792 | _cmll256_decrypt_2x: | |
| 793 | camellia_f %f16, %f2, %f0, %f2 | |
| 794 | camellia_f %f16, %f6, %f4, %f6 | |
| 795 | camellia_f %f18, %f0, %f2, %f0 | |
| 796 | camellia_f %f18, %f4, %f6, %f4 | |
| 797 | ldd [$key - 8], %f16 | |
| 798 | ldd [$key - 16], %f18 | |
| 799 | camellia_f %f20, %f2, %f0, %f2 | |
| 800 | camellia_f %f20, %f6, %f4, %f6 | |
| 801 | camellia_f %f22, %f0, %f2, %f0 | |
| 802 | camellia_f %f22, %f4, %f6, %f4 | |
| 803 | ldd [$key - 24], %f20 | |
| 804 | ldd [$key - 32], %f22 | |
| 805 | camellia_f %f24, %f2, %f0, %f2 | |
| 806 | camellia_f %f24, %f6, %f4, %f6 | |
| 807 | camellia_f %f26, %f0, %f2, %f0 | |
| 808 | camellia_f %f26, %f4, %f6, %f4 | |
| 809 | ldd [$key - 40], %f24 | |
| 810 | ldd [$key - 48], %f26 | |
| 811 | camellia_fl %f28, %f0, %f0 | |
| 812 | camellia_fl %f28, %f4, %f4 | |
| 813 | camellia_fli %f30, %f2, %f2 | |
| 814 | camellia_fli %f30, %f6, %f6 | |
| 815 | ldd [$key - 56], %f28 | |
| 816 | ldd [$key - 64], %f30 | |
| 817 | ___ | |
| 818 | for ($i=1; $i<3; $i++) { | |
| 819 | $code.=<<___; | |
| 820 | camellia_f %f`16+16*$i+0`, %f2, %f0, %f2 | |
| 821 | camellia_f %f`16+16*$i+0`, %f6, %f4, %f6 | |
| 822 | camellia_f %f`16+16*$i+2`, %f0, %f2, %f0 | |
| 823 | camellia_f %f`16+16*$i+2`, %f4, %f6, %f4 | |
| 824 | camellia_f %f`16+16*$i+4`, %f2, %f0, %f2 | |
| 825 | camellia_f %f`16+16*$i+4`, %f6, %f4, %f6 | |
| 826 | camellia_f %f`16+16*$i+6`, %f0, %f2, %f0 | |
| 827 | camellia_f %f`16+16*$i+6`, %f4, %f6, %f4 | |
| 828 | camellia_f %f`16+16*$i+8`, %f2, %f0, %f2 | |
| 829 | camellia_f %f`16+16*$i+8`, %f6, %f4, %f6 | |
| 830 | camellia_f %f`16+16*$i+10`, %f0, %f2, %f0 | |
| 831 | camellia_f %f`16+16*$i+10`, %f4, %f6, %f4 | |
| 832 | camellia_fl %f`16+16*$i+12`, %f0, %f0 | |
| 833 | camellia_fl %f`16+16*$i+12`, %f4, %f4 | |
| 834 | camellia_fli %f`16+16*$i+14`, %f2, %f2 | |
| 835 | camellia_fli %f`16+16*$i+14`, %f6, %f6 | |
| 836 | ___ | |
| 837 | } | |
| 838 | $code.=<<___; | |
| 839 | camellia_f %f16, %f2, %f0, %f2 | |
| 840 | camellia_f %f16, %f6, %f4, %f6 | |
| 841 | camellia_f %f18, %f0, %f2, %f0 | |
| 842 | camellia_f %f18, %f4, %f6, %f4 | |
| 843 | ldd [$key + 184], %f16 | |
| 844 | ldd [$key + 176], %f18 | |
| 845 | camellia_f %f20, %f2, %f0, %f2 | |
| 846 | camellia_f %f20, %f6, %f4, %f6 | |
| 847 | camellia_f %f22, %f0, %f2, %f0 | |
| 848 | camellia_f %f22, %f4, %f6, %f4 | |
| 849 | ldd [$key + 168], %f20 | |
| 850 | ldd [$key + 160], %f22 | |
| 851 | camellia_f %f24, %f2, %f0, %f8 | |
| 852 | camellia_f %f24, %f6, %f4, %f10 | |
| 853 | camellia_f %f26, %f0, %f8, %f2 | |
| 854 | camellia_f %f26, %f4, %f10, %f6 | |
| 855 | ldd [$key + 152], %f24 | |
| 856 | ldd [$key + 144], %f26 | |
| 857 | fxor %f30, %f8, %f0 | |
| 858 | fxor %f30, %f10, %f4 | |
| 859 | fxor %f28, %f2, %f2 | |
| 860 | fxor %f28, %f6, %f6 | |
| 861 | ldd [$key + 136], %f28 | |
| 862 | retl | |
| 863 | ldd [$key + 128], %f30 | |
| 864 | .type _cmll256_decrypt_2x,#function | |
| 865 | .size _cmll256_decrypt_2x,.-_cmll256_decrypt_2x | |
| 866 | ___ | |
| 867 | ||
| 868 | &alg_cbc_encrypt_implement("cmll",128); | |
| 869 | &alg_cbc_encrypt_implement("cmll",256); | |
| 870 | ||
| 871 | &alg_cbc_decrypt_implement("cmll",128); | |
| 872 | &alg_cbc_decrypt_implement("cmll",256); | |
| 873 | ||
| 874 | if ($::evp) { | |
| 875 | &alg_ctr32_implement("cmll",128); | |
| 876 | &alg_ctr32_implement("cmll",256); | |
| 877 | } | |
| 878 | }}} | |
| 879 | ||
| 880 | if (!$::evp) { | |
| 881 | $code.=<<___; | |
| 882 | .global Camellia_encrypt | |
| 883 | Camellia_encrypt=cmll_t4_encrypt | |
| 884 | .global Camellia_decrypt | |
| 885 | Camellia_decrypt=cmll_t4_decrypt | |
| 886 | .global Camellia_set_key | |
| 887 | .align 32 | |
| 888 | Camellia_set_key: | |
| 889 | andcc %o2, 7, %g0 ! double-check alignment | |
| 890 | bnz,a,pn %icc, 1f | |
| 891 | mov -1, %o0 | |
| 892 | brz,a,pn %o0, 1f | |
| 893 | mov -1, %o0 | |
| 894 | brz,a,pn %o2, 1f | |
| 895 | mov -1, %o0 | |
| 896 | andncc %o1, 0x1c0, %g0 | |
| 897 | bnz,a,pn %icc, 1f | |
| 898 | mov -2, %o0 | |
| 899 | cmp %o1, 128 | |
| 900 | bl,a,pn %icc, 1f | |
| 901 | mov -2, %o0 | |
| 902 | b cmll_t4_set_key | |
| 903 | nop | |
| 904 | 1: retl | |
| 905 | nop | |
| 906 | .type Camellia_set_key,#function | |
| 907 | .size Camellia_set_key,.-Camellia_set_key | |
| 908 | ___ | |
| 909 | ||
| 910 | my ($inp,$out,$len,$key,$ivec,$enc)=map("%o$_",(0..5)); | |
| 911 | ||
| 912 | $code.=<<___; | |
| 913 | .globl Camellia_cbc_encrypt | |
| 914 | .align 32 | |
| 915 | Camellia_cbc_encrypt: | |
| 916 | ld [$key + 272], %g1 | |
| 917 | nop | |
| 918 | brz $enc, .Lcbc_decrypt | |
| 919 | cmp %g1, 3 | |
| 920 | ||
| 921 | be,pt %icc, cmll128_t4_cbc_encrypt | |
| 922 | nop | |
| 923 | ba cmll256_t4_cbc_encrypt | |
| 924 | nop | |
| 925 | ||
| 926 | .Lcbc_decrypt: | |
| 927 | be,pt %icc, cmll128_t4_cbc_decrypt | |
| 928 | nop | |
| 929 | ba cmll256_t4_cbc_decrypt | |
| 930 | nop | |
| 931 | .type Camellia_cbc_encrypt,#function | |
| 932 | .size Camellia_cbc_encrypt,.-Camellia_cbc_encrypt | |
| 933 | ___ | |
| 934 | } | |
| 935 | ||
| 936 | &emit_assembler(); | |
| 937 | ||
| a21314db | 938 | close STDOUT or die "error closing STDOUT: $!"; |