aest4-sparcv9.S revision 1.1
1 .text 2 3 .globl aes_t4_encrypt 4 .align 32 5 aes_t4_encrypt: 6 andcc %o0, 7, %g1 ! is input aligned? 7 andn %o0, 7, %o0 8 9 ldx [%o2 + 0], %g4 10 ldx [%o2 + 8], %g5 11 12 ldx [%o0 + 0], %o4 13 bz,pt %icc, 1f 14 ldx [%o0 + 8], %o5 15 ldx [%o0 + 16], %o0 16 sll %g1, 3, %g1 17 sub %g0, %g1, %o3 18 sllx %o4, %g1, %o4 19 sllx %o5, %g1, %g1 20 srlx %o5, %o3, %o5 21 srlx %o0, %o3, %o3 22 or %o5, %o4, %o4 23 or %o3, %g1, %o5 24 1: 25 ld [%o2 + 240], %o3 26 ldd [%o2 + 16], %f12 27 ldd [%o2 + 24], %f14 28 xor %g4, %o4, %o4 29 xor %g5, %o5, %o5 30 .word 0x81b0230c !movxtod %o4,%f0 31 .word 0x85b0230d !movxtod %o5,%f2 32 srl %o3, 1, %o3 33 ldd [%o2 + 32], %f16 34 sub %o3, 1, %o3 35 ldd [%o2 + 40], %f18 36 add %o2, 48, %o2 37 38 .Lenc: 39 .word 0x88cb0400 !aes_eround01 %f12,%f0,%f2,%f4 40 .word 0x84cb8420 !aes_eround23 %f14,%f0,%f2,%f2 41 ldd [%o2 + 0], %f12 42 ldd [%o2 + 8], %f14 43 sub %o3,1,%o3 44 .word 0x80cc0404 !aes_eround01 %f16,%f4,%f2,%f0 45 .word 0x84cc8424 !aes_eround23 %f18,%f4,%f2,%f2 46 ldd [%o2 + 16], %f16 47 ldd [%o2 + 24], %f18 48 brnz,pt %o3, .Lenc 49 add %o2, 32, %o2 50 51 andcc %o1, 7, %o4 ! is output aligned? 52 .word 0x88cb0400 !aes_eround01 %f12,%f0,%f2,%f4 53 .word 0x84cb8420 !aes_eround23 %f14,%f0,%f2,%f2 54 .word 0x80cc0484 !aes_eround01_l %f16,%f4,%f2,%f0 55 .word 0x84cc84a4 !aes_eround23_l %f18,%f4,%f2,%f2 56 57 bnz,pn %icc, 2f 58 nop 59 60 std %f0, [%o1 + 0] 61 retl 62 std %f2, [%o1 + 8] 63 64 2: .word 0x93b24340 !alignaddrl %o1,%g0,%o1 65 mov 0xff, %o5 66 srl %o5, %o4, %o5 67 68 .word 0x89b00900 !faligndata %f0,%f0,%f4 69 .word 0x8db00902 !faligndata %f0,%f2,%f6 70 .word 0x91b08902 !faligndata %f2,%f2,%f8 71 72 stda %f4, [%o1 + %o5]0xc0 ! partial store 73 std %f6, [%o1 + 8] 74 add %o1, 16, %o1 75 orn %g0, %o5, %o5 76 retl 77 stda %f8, [%o1 + %o5]0xc0 ! partial store 78 .type aes_t4_encrypt,#function 79 .size aes_t4_encrypt,.-aes_t4_encrypt 80 81 .globl aes_t4_decrypt 82 .align 32 83 aes_t4_decrypt: 84 andcc %o0, 7, %g1 ! is input aligned? 85 andn %o0, 7, %o0 86 87 ldx [%o2 + 0], %g4 88 ldx [%o2 + 8], %g5 89 90 ldx [%o0 + 0], %o4 91 bz,pt %icc, 1f 92 ldx [%o0 + 8], %o5 93 ldx [%o0 + 16], %o0 94 sll %g1, 3, %g1 95 sub %g0, %g1, %o3 96 sllx %o4, %g1, %o4 97 sllx %o5, %g1, %g1 98 srlx %o5, %o3, %o5 99 srlx %o0, %o3, %o3 100 or %o5, %o4, %o4 101 or %o3, %g1, %o5 102 1: 103 ld [%o2 + 240], %o3 104 ldd [%o2 + 16], %f12 105 ldd [%o2 + 24], %f14 106 xor %g4, %o4, %o4 107 xor %g5, %o5, %o5 108 .word 0x81b0230c !movxtod %o4,%f0 109 .word 0x85b0230d !movxtod %o5,%f2 110 srl %o3, 1, %o3 111 ldd [%o2 + 32], %f16 112 sub %o3, 1, %o3 113 ldd [%o2 + 40], %f18 114 add %o2, 48, %o2 115 116 .Ldec: 117 .word 0x88cb0440 !aes_dround01 %f12,%f0,%f2,%f4 118 .word 0x84cb8460 !aes_dround23 %f14,%f0,%f2,%f2 119 ldd [%o2 + 0], %f12 120 ldd [%o2 + 8], %f14 121 sub %o3,1,%o3 122 .word 0x80cc0444 !aes_dround01 %f16,%f4,%f2,%f0 123 .word 0x84cc8464 !aes_dround23 %f18,%f4,%f2,%f2 124 ldd [%o2 + 16], %f16 125 ldd [%o2 + 24], %f18 126 brnz,pt %o3, .Ldec 127 add %o2, 32, %o2 128 129 andcc %o1, 7, %o4 ! is output aligned? 130 .word 0x88cb0440 !aes_dround01 %f12,%f0,%f2,%f4 131 .word 0x84cb8460 !aes_dround23 %f14,%f0,%f2,%f2 132 .word 0x80cc04c4 !aes_dround01_l %f16,%f4,%f2,%f0 133 .word 0x84cc84e4 !aes_dround23_l %f18,%f4,%f2,%f2 134 135 bnz,pn %icc, 2f 136 nop 137 138 std %f0, [%o1 + 0] 139 retl 140 std %f2, [%o1 + 8] 141 142 2: .word 0x93b24340 !alignaddrl %o1,%g0,%o1 143 mov 0xff, %o5 144 srl %o5, %o4, %o5 145 146 .word 0x89b00900 !faligndata %f0,%f0,%f4 147 .word 0x8db00902 !faligndata %f0,%f2,%f6 148 .word 0x91b08902 !faligndata %f2,%f2,%f8 149 150 stda %f4, [%o1 + %o5]0xc0 ! partial store 151 std %f6, [%o1 + 8] 152 add %o1, 16, %o1 153 orn %g0, %o5, %o5 154 retl 155 stda %f8, [%o1 + %o5]0xc0 ! partial store 156 .type aes_t4_decrypt,#function 157 .size aes_t4_decrypt,.-aes_t4_decrypt 158 .globl aes_t4_set_encrypt_key 159 .align 32 160 aes_t4_set_encrypt_key: 161 .Lset_encrypt_key: 162 and %o0, 7, %o3 163 .word 0x91b20300 !alignaddr %o0,%g0,%o0 164 cmp %o1, 192 165 ldd [%o0 + 0], %f0 166 bl,pt %icc,.L128 167 ldd [%o0 + 8], %f2 168 169 be,pt %icc,.L192 170 ldd [%o0 + 16], %f4 171 brz,pt %o3, .L256aligned 172 ldd [%o0 + 24], %f6 173 174 ldd [%o0 + 32], %f8 175 .word 0x81b00902 !faligndata %f0,%f2,%f0 176 .word 0x85b08904 !faligndata %f2,%f4,%f2 177 .word 0x89b10906 !faligndata %f4,%f6,%f4 178 .word 0x8db18908 !faligndata %f6,%f8,%f6 179 .L256aligned: 180 std %f0, [%o2 + 0] 181 .word 0x80c80106 !aes_kexpand1 %f0,%f6,0,%f0 182 std %f2, [%o2 + 8] 183 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 184 std %f4, [%o2 + 16] 185 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 186 std %f6, [%o2 + 24] 187 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 188 std %f0, [%o2 + 32] 189 .word 0x80c80306 !aes_kexpand1 %f0,%f6,1,%f0 190 std %f2, [%o2 + 40] 191 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 192 std %f4, [%o2 + 48] 193 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 194 std %f6, [%o2 + 56] 195 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 196 std %f0, [%o2 + 64] 197 .word 0x80c80506 !aes_kexpand1 %f0,%f6,2,%f0 198 std %f2, [%o2 + 72] 199 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 200 std %f4, [%o2 + 80] 201 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 202 std %f6, [%o2 + 88] 203 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 204 std %f0, [%o2 + 96] 205 .word 0x80c80706 !aes_kexpand1 %f0,%f6,3,%f0 206 std %f2, [%o2 + 104] 207 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 208 std %f4, [%o2 + 112] 209 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 210 std %f6, [%o2 + 120] 211 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 212 std %f0, [%o2 + 128] 213 .word 0x80c80906 !aes_kexpand1 %f0,%f6,4,%f0 214 std %f2, [%o2 + 136] 215 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 216 std %f4, [%o2 + 144] 217 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 218 std %f6, [%o2 + 152] 219 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 220 std %f0, [%o2 + 160] 221 .word 0x80c80b06 !aes_kexpand1 %f0,%f6,5,%f0 222 std %f2, [%o2 + 168] 223 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 224 std %f4, [%o2 + 176] 225 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 226 std %f6, [%o2 + 184] 227 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 228 std %f0, [%o2 + 192] 229 .word 0x80c80d06 !aes_kexpand1 %f0,%f6,6,%f0 230 std %f2, [%o2 + 200] 231 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 232 std %f4, [%o2 + 208] 233 std %f6, [%o2 + 216] 234 std %f0, [%o2 + 224] 235 std %f2, [%o2 + 232] 236 237 mov 14, %o3 238 st %o3, [%o2 + 240] 239 retl 240 xor %o0, %o0, %o0 241 242 .align 16 243 .L192: 244 brz,pt %o3, .L192aligned 245 nop 246 247 ldd [%o0 + 24], %f6 248 .word 0x81b00902 !faligndata %f0,%f2,%f0 249 .word 0x85b08904 !faligndata %f2,%f4,%f2 250 .word 0x89b10906 !faligndata %f4,%f6,%f4 251 .L192aligned: 252 std %f0, [%o2 + 0] 253 .word 0x80c80104 !aes_kexpand1 %f0,%f4,0,%f0 254 std %f2, [%o2 + 8] 255 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 256 std %f4, [%o2 + 16] 257 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 258 std %f0, [%o2 + 24] 259 .word 0x80c80304 !aes_kexpand1 %f0,%f4,1,%f0 260 std %f2, [%o2 + 32] 261 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 262 std %f4, [%o2 + 40] 263 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 264 std %f0, [%o2 + 48] 265 .word 0x80c80504 !aes_kexpand1 %f0,%f4,2,%f0 266 std %f2, [%o2 + 56] 267 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 268 std %f4, [%o2 + 64] 269 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 270 std %f0, [%o2 + 72] 271 .word 0x80c80704 !aes_kexpand1 %f0,%f4,3,%f0 272 std %f2, [%o2 + 80] 273 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 274 std %f4, [%o2 + 88] 275 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 276 std %f0, [%o2 + 96] 277 .word 0x80c80904 !aes_kexpand1 %f0,%f4,4,%f0 278 std %f2, [%o2 + 104] 279 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 280 std %f4, [%o2 + 112] 281 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 282 std %f0, [%o2 + 120] 283 .word 0x80c80b04 !aes_kexpand1 %f0,%f4,5,%f0 284 std %f2, [%o2 + 128] 285 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 286 std %f4, [%o2 + 136] 287 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 288 std %f0, [%o2 + 144] 289 .word 0x80c80d04 !aes_kexpand1 %f0,%f4,6,%f0 290 std %f2, [%o2 + 152] 291 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 292 std %f4, [%o2 + 160] 293 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 294 std %f0, [%o2 + 168] 295 .word 0x80c80f04 !aes_kexpand1 %f0,%f4,7,%f0 296 std %f2, [%o2 + 176] 297 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 298 std %f4, [%o2 + 184] 299 std %f0, [%o2 + 192] 300 std %f2, [%o2 + 200] 301 302 mov 12, %o3 303 st %o3, [%o2 + 240] 304 retl 305 xor %o0, %o0, %o0 306 307 .align 16 308 .L128: 309 brz,pt %o3, .L128aligned 310 nop 311 312 ldd [%o0 + 16], %f4 313 .word 0x81b00902 !faligndata %f0,%f2,%f0 314 .word 0x85b08904 !faligndata %f2,%f4,%f2 315 .L128aligned: 316 std %f0, [%o2 + 0] 317 .word 0x80c80102 !aes_kexpand1 %f0,%f2,0,%f0 318 std %f2, [%o2 + 8] 319 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 320 std %f0, [%o2 + 16] 321 .word 0x80c80302 !aes_kexpand1 %f0,%f2,1,%f0 322 std %f2, [%o2 + 24] 323 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 324 std %f0, [%o2 + 32] 325 .word 0x80c80502 !aes_kexpand1 %f0,%f2,2,%f0 326 std %f2, [%o2 + 40] 327 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 328 std %f0, [%o2 + 48] 329 .word 0x80c80702 !aes_kexpand1 %f0,%f2,3,%f0 330 std %f2, [%o2 + 56] 331 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 332 std %f0, [%o2 + 64] 333 .word 0x80c80902 !aes_kexpand1 %f0,%f2,4,%f0 334 std %f2, [%o2 + 72] 335 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 336 std %f0, [%o2 + 80] 337 .word 0x80c80b02 !aes_kexpand1 %f0,%f2,5,%f0 338 std %f2, [%o2 + 88] 339 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 340 std %f0, [%o2 + 96] 341 .word 0x80c80d02 !aes_kexpand1 %f0,%f2,6,%f0 342 std %f2, [%o2 + 104] 343 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 344 std %f0, [%o2 + 112] 345 .word 0x80c80f02 !aes_kexpand1 %f0,%f2,7,%f0 346 std %f2, [%o2 + 120] 347 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 348 std %f0, [%o2 + 128] 349 .word 0x80c81102 !aes_kexpand1 %f0,%f2,8,%f0 350 std %f2, [%o2 + 136] 351 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 352 std %f0, [%o2 + 144] 353 .word 0x80c81302 !aes_kexpand1 %f0,%f2,9,%f0 354 std %f2, [%o2 + 152] 355 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 356 std %f0, [%o2 + 160] 357 std %f2, [%o2 + 168] 358 359 mov 10, %o3 360 st %o3, [%o2 + 240] 361 retl 362 xor %o0, %o0, %o0 363 .type aes_t4_set_encrypt_key,#function 364 .size aes_t4_set_encrypt_key,.-aes_t4_set_encrypt_key 365 366 .globl aes_t4_set_decrypt_key 367 .align 32 368 aes_t4_set_decrypt_key: 369 mov %o7, %o5 370 call .Lset_encrypt_key 371 nop 372 373 mov %o5, %o7 374 sll %o3, 4, %o0 ! %o3 is number of rounds 375 add %o3, 2, %o3 376 add %o2, %o0, %o0 ! %o0=%o2+16*rounds 377 srl %o3, 2, %o3 ! %o3=(rounds+2)/4 378 379 .Lkey_flip: 380 ldd [%o2 + 0], %f0 381 ldd [%o2 + 8], %f2 382 ldd [%o2 + 16], %f4 383 ldd [%o2 + 24], %f6 384 ldd [%o0 + 0], %f8 385 ldd [%o0 + 8], %f10 386 ldd [%o0 - 16], %f12 387 ldd [%o0 - 8], %f14 388 sub %o3, 1, %o3 389 std %f0, [%o0 + 0] 390 std %f2, [%o0 + 8] 391 std %f4, [%o0 - 16] 392 std %f6, [%o0 - 8] 393 std %f8, [%o2 + 0] 394 std %f10, [%o2 + 8] 395 std %f12, [%o2 + 16] 396 std %f14, [%o2 + 24] 397 add %o2, 32, %o2 398 brnz %o3, .Lkey_flip 399 sub %o0, 32, %o0 400 401 retl 402 xor %o0, %o0, %o0 403 .type aes_t4_set_decrypt_key,#function 404 .size aes_t4_set_decrypt_key,.-aes_t4_set_decrypt_key 405 .align 32 406 _aes128_encrypt_1x: 407 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 408 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 409 .word 0x80cd0404 !aes_eround01 %f20,%f4,%f2,%f0 410 .word 0x84cd8424 !aes_eround23 %f22,%f4,%f2,%f2 411 .word 0x88ce0400 !aes_eround01 %f24,%f0,%f2,%f4 412 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 413 .word 0x80cf0404 !aes_eround01 %f28,%f4,%f2,%f0 414 .word 0x84cf8424 !aes_eround23 %f30,%f4,%f2,%f2 415 .word 0x88c84400 !aes_eround01 %f32,%f0,%f2,%f4 416 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 417 .word 0x80c94404 !aes_eround01 %f36,%f4,%f2,%f0 418 .word 0x84c9c424 !aes_eround23 %f38,%f4,%f2,%f2 419 .word 0x88ca4400 !aes_eround01 %f40,%f0,%f2,%f4 420 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 421 .word 0x80cb4404 !aes_eround01 %f44,%f4,%f2,%f0 422 .word 0x84cbc424 !aes_eround23 %f46,%f4,%f2,%f2 423 .word 0x88cc4400 !aes_eround01 %f48,%f0,%f2,%f4 424 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 425 .word 0x80cd4484 !aes_eround01_l %f52,%f4,%f2,%f0 426 retl 427 .word 0x84cdc4a4 !aes_eround23_l %f54,%f4,%f2,%f2 428 .type _aes128_encrypt_1x,#function 429 .size _aes128_encrypt_1x,.-_aes128_encrypt_1x 430 431 .align 32 432 _aes128_encrypt_2x: 433 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 434 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 435 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 436 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 437 .word 0x80cd0408 !aes_eround01 %f20,%f8,%f2,%f0 438 .word 0x84cd8428 !aes_eround23 %f22,%f8,%f2,%f2 439 .word 0x88cd0c0a !aes_eround01 %f20,%f10,%f6,%f4 440 .word 0x8ccd8c2a !aes_eround23 %f22,%f10,%f6,%f6 441 .word 0x90ce0400 !aes_eround01 %f24,%f0,%f2,%f8 442 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 443 .word 0x94ce0c04 !aes_eround01 %f24,%f4,%f6,%f10 444 .word 0x8cce8c24 !aes_eround23 %f26,%f4,%f6,%f6 445 .word 0x80cf0408 !aes_eround01 %f28,%f8,%f2,%f0 446 .word 0x84cf8428 !aes_eround23 %f30,%f8,%f2,%f2 447 .word 0x88cf0c0a !aes_eround01 %f28,%f10,%f6,%f4 448 .word 0x8ccf8c2a !aes_eround23 %f30,%f10,%f6,%f6 449 .word 0x90c84400 !aes_eround01 %f32,%f0,%f2,%f8 450 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 451 .word 0x94c84c04 !aes_eround01 %f32,%f4,%f6,%f10 452 .word 0x8cc8cc24 !aes_eround23 %f34,%f4,%f6,%f6 453 .word 0x80c94408 !aes_eround01 %f36,%f8,%f2,%f0 454 .word 0x84c9c428 !aes_eround23 %f38,%f8,%f2,%f2 455 .word 0x88c94c0a !aes_eround01 %f36,%f10,%f6,%f4 456 .word 0x8cc9cc2a !aes_eround23 %f38,%f10,%f6,%f6 457 .word 0x90ca4400 !aes_eround01 %f40,%f0,%f2,%f8 458 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 459 .word 0x94ca4c04 !aes_eround01 %f40,%f4,%f6,%f10 460 .word 0x8ccacc24 !aes_eround23 %f42,%f4,%f6,%f6 461 .word 0x80cb4408 !aes_eround01 %f44,%f8,%f2,%f0 462 .word 0x84cbc428 !aes_eround23 %f46,%f8,%f2,%f2 463 .word 0x88cb4c0a !aes_eround01 %f44,%f10,%f6,%f4 464 .word 0x8ccbcc2a !aes_eround23 %f46,%f10,%f6,%f6 465 .word 0x90cc4400 !aes_eround01 %f48,%f0,%f2,%f8 466 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 467 .word 0x94cc4c04 !aes_eround01 %f48,%f4,%f6,%f10 468 .word 0x8ccccc24 !aes_eround23 %f50,%f4,%f6,%f6 469 .word 0x80cd4488 !aes_eround01_l %f52,%f8,%f2,%f0 470 .word 0x84cdc4a8 !aes_eround23_l %f54,%f8,%f2,%f2 471 .word 0x88cd4c8a !aes_eround01_l %f52,%f10,%f6,%f4 472 retl 473 .word 0x8ccdccaa !aes_eround23_l %f54,%f10,%f6,%f6 474 .type _aes128_encrypt_2x,#function 475 .size _aes128_encrypt_2x,.-_aes128_encrypt_2x 476 477 .align 32 478 _aes128_loadkey: 479 ldx [%i3 + 0], %g4 480 ldx [%i3 + 8], %g5 481 ldd [%i3 + 16], %f16 482 ldd [%i3 + 24], %f18 483 ldd [%i3 + 32], %f20 484 ldd [%i3 + 40], %f22 485 ldd [%i3 + 48], %f24 486 ldd [%i3 + 56], %f26 487 ldd [%i3 + 64], %f28 488 ldd [%i3 + 72], %f30 489 ldd [%i3 + 80], %f32 490 ldd [%i3 + 88], %f34 491 ldd [%i3 + 96], %f36 492 ldd [%i3 + 104], %f38 493 ldd [%i3 + 112], %f40 494 ldd [%i3 + 120], %f42 495 ldd [%i3 + 128], %f44 496 ldd [%i3 + 136], %f46 497 ldd [%i3 + 144], %f48 498 ldd [%i3 + 152], %f50 499 ldd [%i3 + 160], %f52 500 ldd [%i3 + 168], %f54 501 retl 502 nop 503 .type _aes128_loadkey,#function 504 .size _aes128_loadkey,.-_aes128_loadkey 505 _aes128_load_enckey=_aes128_loadkey 506 _aes128_load_deckey=_aes128_loadkey 507 508 .globl aes128_t4_cbc_encrypt 509 .align 32 510 aes128_t4_cbc_encrypt: 511 save %sp, -112, %sp 512 cmp %i2, 0 513 be,pn %icc, .L128_cbc_enc_abort 514 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 515 sub %i0, %i1, %l5 ! %i0!=%i1 516 ld [%i4 + 0], %f0 517 ld [%i4 + 4], %f1 518 ld [%i4 + 8], %f2 519 ld [%i4 + 12], %f3 520 prefetch [%i0], 20 521 prefetch [%i0 + 63], 20 522 call _aes128_load_enckey 523 and %i0, 7, %l0 524 andn %i0, 7, %i0 525 sll %l0, 3, %l0 526 mov 64, %l1 527 mov 0xff, %l3 528 sub %l1, %l0, %l1 529 and %i1, 7, %l2 530 cmp %i2, 127 531 movrnz %l2, 0, %l5 ! if ( %i1&7 || 532 movleu %icc, 0, %l5 ! %i2<128 || 533 brnz,pn %l5, .L128cbc_enc_blk ! %i0==%i1) 534 srl %l3, %l2, %l3 535 536 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 537 srlx %i2, 4, %i2 538 prefetch [%i1], 22 539 540 .L128_cbc_enc_loop: 541 ldx [%i0 + 0], %o0 542 brz,pt %l0, 4f 543 ldx [%i0 + 8], %o1 544 545 ldx [%i0 + 16], %o2 546 sllx %o0, %l0, %o0 547 srlx %o1, %l1, %g1 548 sllx %o1, %l0, %o1 549 or %g1, %o0, %o0 550 srlx %o2, %l1, %o2 551 or %o2, %o1, %o1 552 4: 553 xor %g4, %o0, %o0 ! ^= rk[0] 554 xor %g5, %o1, %o1 555 .word 0x99b02308 !movxtod %o0,%f12 556 .word 0x9db02309 !movxtod %o1,%f14 557 558 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 559 .word 0x85b38d82 !fxor %f14,%f2,%f2 560 prefetch [%i1 + 63], 22 561 prefetch [%i0 + 16+63], 20 562 call _aes128_encrypt_1x 563 add %i0, 16, %i0 564 565 brnz,pn %l2, 2f 566 sub %i2, 1, %i2 567 568 std %f0, [%i1 + 0] 569 std %f2, [%i1 + 8] 570 brnz,pt %i2, .L128_cbc_enc_loop 571 add %i1, 16, %i1 572 st %f0, [%i4 + 0] 573 st %f1, [%i4 + 4] 574 st %f2, [%i4 + 8] 575 st %f3, [%i4 + 12] 576 .L128_cbc_enc_abort: 577 ret 578 restore 579 580 .align 16 581 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 582 ! and ~3x deterioration 583 ! in inp==out case 584 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 585 .word 0x8db00902 !faligndata %f0,%f2,%f6 586 .word 0x91b08902 !faligndata %f2,%f2,%f8 587 588 stda %f4, [%i1 + %l3]0xc0 ! partial store 589 std %f6, [%i1 + 8] 590 add %i1, 16, %i1 591 orn %g0, %l3, %l3 592 stda %f8, [%i1 + %l3]0xc0 ! partial store 593 594 brnz,pt %i2, .L128_cbc_enc_loop+4 595 orn %g0, %l3, %l3 596 st %f0, [%i4 + 0] 597 st %f1, [%i4 + 4] 598 st %f2, [%i4 + 8] 599 st %f3, [%i4 + 12] 600 ret 601 restore 602 603 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 604 .align 32 605 .L128cbc_enc_blk: 606 add %i1, %i2, %l5 607 and %l5, 63, %l5 ! tail 608 sub %i2, %l5, %i2 609 add %l5, 15, %l5 ! round up to 16n 610 srlx %i2, 4, %i2 611 srl %l5, 4, %l5 612 613 .L128_cbc_enc_blk_loop: 614 ldx [%i0 + 0], %o0 615 brz,pt %l0, 5f 616 ldx [%i0 + 8], %o1 617 618 ldx [%i0 + 16], %o2 619 sllx %o0, %l0, %o0 620 srlx %o1, %l1, %g1 621 sllx %o1, %l0, %o1 622 or %g1, %o0, %o0 623 srlx %o2, %l1, %o2 624 or %o2, %o1, %o1 625 5: 626 xor %g4, %o0, %o0 ! ^= rk[0] 627 xor %g5, %o1, %o1 628 .word 0x99b02308 !movxtod %o0,%f12 629 .word 0x9db02309 !movxtod %o1,%f14 630 631 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 632 .word 0x85b38d82 !fxor %f14,%f2,%f2 633 prefetch [%i0 + 16+63], 20 634 call _aes128_encrypt_1x 635 add %i0, 16, %i0 636 sub %i2, 1, %i2 637 638 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 639 add %i1, 8, %i1 640 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 641 brnz,pt %i2, .L128_cbc_enc_blk_loop 642 add %i1, 8, %i1 643 644 membar #StoreLoad|#StoreStore 645 brnz,pt %l5, .L128_cbc_enc_loop 646 mov %l5, %i2 647 st %f0, [%i4 + 0] 648 st %f1, [%i4 + 4] 649 st %f2, [%i4 + 8] 650 st %f3, [%i4 + 12] 651 ret 652 restore 653 .type aes128_t4_cbc_encrypt,#function 654 .size aes128_t4_cbc_encrypt,.-aes128_t4_cbc_encrypt 655 .globl aes128_t4_ctr32_encrypt 656 .align 32 657 aes128_t4_ctr32_encrypt: 658 save %sp, -112, %sp 659 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 660 661 prefetch [%i0], 20 662 prefetch [%i0 + 63], 20 663 call _aes128_load_enckey 664 sllx %i2, 4, %i2 665 666 ld [%i4 + 0], %l4 ! counter 667 ld [%i4 + 4], %l5 668 ld [%i4 + 8], %l6 669 ld [%i4 + 12], %l7 670 671 sllx %l4, 32, %o5 672 or %l5, %o5, %o5 673 sllx %l6, 32, %g1 674 xor %o5, %g4, %g4 ! ^= rk[0] 675 xor %g1, %g5, %g5 676 .word 0x9db02304 !movxtod %g4,%f14 ! most significant 64 bits 677 678 sub %i0, %i1, %l5 ! %i0!=%i1 679 and %i0, 7, %l0 680 andn %i0, 7, %i0 681 sll %l0, 3, %l0 682 mov 64, %l1 683 mov 0xff, %l3 684 sub %l1, %l0, %l1 685 and %i1, 7, %l2 686 cmp %i2, 255 687 movrnz %l2, 0, %l5 ! if ( %i1&7 || 688 movleu %icc, 0, %l5 ! %i2<256 || 689 brnz,pn %l5, .L128_ctr32_blk ! %i0==%i1) 690 srl %l3, %l2, %l3 691 692 andcc %i2, 16, %g0 ! is number of blocks even? 693 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 694 bz %icc, .L128_ctr32_loop2x 695 srlx %i2, 4, %i2 696 .L128_ctr32_loop: 697 ldx [%i0 + 0], %o0 698 brz,pt %l0, 4f 699 ldx [%i0 + 8], %o1 700 701 ldx [%i0 + 16], %o2 702 sllx %o0, %l0, %o0 703 srlx %o1, %l1, %g1 704 sllx %o1, %l0, %o1 705 or %g1, %o0, %o0 706 srlx %o2, %l1, %o2 707 or %o2, %o1, %o1 708 4: 709 xor %g5, %l7, %g1 ! ^= rk[0] 710 add %l7, 1, %l7 711 .word 0x85b02301 !movxtod %g1,%f2 712 srl %l7, 0, %l7 ! clruw 713 prefetch [%i1 + 63], 22 714 prefetch [%i0 + 16+63], 20 715 .word 0x88cc040e !aes_eround01 %f16,%f14,%f2,%f4 716 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 717 call _aes128_encrypt_1x+8 718 add %i0, 16, %i0 719 720 .word 0x95b02308 !movxtod %o0,%f10 721 .word 0x99b02309 !movxtod %o1,%f12 722 .word 0x81b28d80 !fxor %f10,%f0,%f0 ! ^= inp 723 .word 0x85b30d82 !fxor %f12,%f2,%f2 724 725 brnz,pn %l2, 2f 726 sub %i2, 1, %i2 727 728 std %f0, [%i1 + 0] 729 std %f2, [%i1 + 8] 730 brnz,pt %i2, .L128_ctr32_loop2x 731 add %i1, 16, %i1 732 733 ret 734 restore 735 736 .align 16 737 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 738 ! and ~3x deterioration 739 ! in inp==out case 740 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 741 .word 0x8db00902 !faligndata %f0,%f2,%f6 742 .word 0x91b08902 !faligndata %f2,%f2,%f8 743 stda %f4, [%i1 + %l3]0xc0 ! partial store 744 std %f6, [%i1 + 8] 745 add %i1, 16, %i1 746 orn %g0, %l3, %l3 747 stda %f8, [%i1 + %l3]0xc0 ! partial store 748 749 brnz,pt %i2, .L128_ctr32_loop2x+4 750 orn %g0, %l3, %l3 751 752 ret 753 restore 754 755 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 756 .align 32 757 .L128_ctr32_loop2x: 758 ldx [%i0 + 0], %o0 759 ldx [%i0 + 8], %o1 760 ldx [%i0 + 16], %o2 761 brz,pt %l0, 4f 762 ldx [%i0 + 24], %o3 763 764 ldx [%i0 + 32], %o4 765 sllx %o0, %l0, %o0 766 srlx %o1, %l1, %g1 767 or %g1, %o0, %o0 768 sllx %o1, %l0, %o1 769 srlx %o2, %l1, %g1 770 or %g1, %o1, %o1 771 sllx %o2, %l0, %o2 772 srlx %o3, %l1, %g1 773 or %g1, %o2, %o2 774 sllx %o3, %l0, %o3 775 srlx %o4, %l1, %o4 776 or %o4, %o3, %o3 777 4: 778 xor %g5, %l7, %g1 ! ^= rk[0] 779 add %l7, 1, %l7 780 .word 0x85b02301 !movxtod %g1,%f2 781 srl %l7, 0, %l7 ! clruw 782 xor %g5, %l7, %g1 783 add %l7, 1, %l7 784 .word 0x8db02301 !movxtod %g1,%f6 785 srl %l7, 0, %l7 ! clruw 786 prefetch [%i1 + 63], 22 787 prefetch [%i0 + 32+63], 20 788 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 789 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 790 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 791 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 792 call _aes128_encrypt_2x+16 793 add %i0, 32, %i0 794 795 .word 0x91b02308 !movxtod %o0,%f8 796 .word 0x95b02309 !movxtod %o1,%f10 797 .word 0x99b0230a !movxtod %o2,%f12 798 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 799 .word 0x91b0230b !movxtod %o3,%f8 800 .word 0x85b28d82 !fxor %f10,%f2,%f2 801 .word 0x89b30d84 !fxor %f12,%f4,%f4 802 .word 0x8db20d86 !fxor %f8,%f6,%f6 803 804 brnz,pn %l2, 2f 805 sub %i2, 2, %i2 806 807 std %f0, [%i1 + 0] 808 std %f2, [%i1 + 8] 809 std %f4, [%i1 + 16] 810 std %f6, [%i1 + 24] 811 brnz,pt %i2, .L128_ctr32_loop2x 812 add %i1, 32, %i1 813 814 ret 815 restore 816 817 .align 16 818 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 819 ! and ~3x deterioration 820 ! in inp==out case 821 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 822 .word 0x81b00902 !faligndata %f0,%f2,%f0 823 .word 0x85b08904 !faligndata %f2,%f4,%f2 824 .word 0x89b10906 !faligndata %f4,%f6,%f4 825 .word 0x8db18906 !faligndata %f6,%f6,%f6 826 827 stda %f8, [%i1 + %l3]0xc0 ! partial store 828 std %f0, [%i1 + 8] 829 std %f2, [%i1 + 16] 830 std %f4, [%i1 + 24] 831 add %i1, 32, %i1 832 orn %g0, %l3, %l3 833 stda %f6, [%i1 + %l3]0xc0 ! partial store 834 835 brnz,pt %i2, .L128_ctr32_loop2x+4 836 orn %g0, %l3, %l3 837 838 ret 839 restore 840 841 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 842 .align 32 843 .L128_ctr32_blk: 844 add %i1, %i2, %l5 845 and %l5, 63, %l5 ! tail 846 sub %i2, %l5, %i2 847 add %l5, 15, %l5 ! round up to 16n 848 srlx %i2, 4, %i2 849 srl %l5, 4, %l5 850 sub %i2, 1, %i2 851 add %l5, 1, %l5 852 853 .L128_ctr32_blk_loop2x: 854 ldx [%i0 + 0], %o0 855 ldx [%i0 + 8], %o1 856 ldx [%i0 + 16], %o2 857 brz,pt %l0, 5f 858 ldx [%i0 + 24], %o3 859 860 ldx [%i0 + 32], %o4 861 sllx %o0, %l0, %o0 862 srlx %o1, %l1, %g1 863 or %g1, %o0, %o0 864 sllx %o1, %l0, %o1 865 srlx %o2, %l1, %g1 866 or %g1, %o1, %o1 867 sllx %o2, %l0, %o2 868 srlx %o3, %l1, %g1 869 or %g1, %o2, %o2 870 sllx %o3, %l0, %o3 871 srlx %o4, %l1, %o4 872 or %o4, %o3, %o3 873 5: 874 xor %g5, %l7, %g1 ! ^= rk[0] 875 add %l7, 1, %l7 876 .word 0x85b02301 !movxtod %g1,%f2 877 srl %l7, 0, %l7 ! clruw 878 xor %g5, %l7, %g1 879 add %l7, 1, %l7 880 .word 0x8db02301 !movxtod %g1,%f6 881 srl %l7, 0, %l7 ! clruw 882 prefetch [%i0 + 32+63], 20 883 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 884 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 885 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 886 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 887 call _aes128_encrypt_2x+16 888 add %i0, 32, %i0 889 subcc %i2, 2, %i2 890 891 .word 0x91b02308 !movxtod %o0,%f8 892 .word 0x95b02309 !movxtod %o1,%f10 893 .word 0x99b0230a !movxtod %o2,%f12 894 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 895 .word 0x91b0230b !movxtod %o3,%f8 896 .word 0x85b28d82 !fxor %f10,%f2,%f2 897 .word 0x89b30d84 !fxor %f12,%f4,%f4 898 .word 0x8db20d86 !fxor %f8,%f6,%f6 899 900 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 901 add %i1, 8, %i1 902 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 903 add %i1, 8, %i1 904 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 905 add %i1, 8, %i1 906 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 907 bgu,pt %icc, .L128_ctr32_blk_loop2x 908 add %i1, 8, %i1 909 910 add %l5, %i2, %i2 911 andcc %i2, 1, %g0 ! is number of blocks even? 912 membar #StoreLoad|#StoreStore 913 bnz,pt %icc, .L128_ctr32_loop 914 srl %i2, 0, %i2 915 brnz,pn %i2, .L128_ctr32_loop2x 916 nop 917 918 ret 919 restore 920 .type aes128_t4_ctr32_encrypt,#function 921 .size aes128_t4_ctr32_encrypt,.-aes128_t4_ctr32_encrypt 922 .globl aes128_t4_xts_encrypt 923 .align 32 924 aes128_t4_xts_encrypt: 925 save %sp, -112-16, %sp 926 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 927 928 mov %i5, %o0 929 add %fp, 0-16, %o1 930 call aes_t4_encrypt 931 mov %i4, %o2 932 933 add %fp, 0-16, %l7 934 ldxa [%l7]0x88, %g2 935 add %fp, 0-8, %l7 936 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 937 938 sethi %hi(0x76543210), %l7 939 or %l7, %lo(0x76543210), %l7 940 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 941 942 prefetch [%i0], 20 943 prefetch [%i0 + 63], 20 944 call _aes128_load_enckey 945 and %i2, 15, %i5 946 and %i2, -16, %i2 947 948 sub %i0, %i1, %l5 ! %i0!=%i1 949 and %i0, 7, %l0 950 andn %i0, 7, %i0 951 sll %l0, 3, %l0 952 mov 64, %l1 953 mov 0xff, %l3 954 sub %l1, %l0, %l1 955 and %i1, 7, %l2 956 cmp %i2, 255 957 movrnz %l2, 0, %l5 ! if ( %i1&7 || 958 movleu %icc, 0, %l5 ! %i2<256 || 959 brnz,pn %l5, .L128_xts_enblk ! %i0==%i1) 960 srl %l3, %l2, %l3 961 962 andcc %i2, 16, %g0 ! is number of blocks even? 963 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 964 bz %icc, .L128_xts_enloop2x 965 srlx %i2, 4, %i2 966 .L128_xts_enloop: 967 ldx [%i0 + 0], %o0 968 brz,pt %l0, 4f 969 ldx [%i0 + 8], %o1 970 971 ldx [%i0 + 16], %o2 972 sllx %o0, %l0, %o0 973 srlx %o1, %l1, %g1 974 sllx %o1, %l0, %o1 975 or %g1, %o0, %o0 976 srlx %o2, %l1, %o2 977 or %o2, %o1, %o1 978 4: 979 .word 0x99b02302 !movxtod %g2,%f12 980 .word 0x9db02303 !movxtod %g3,%f14 981 .word 0x99b3098c !bshuffle %f12,%f12,%f12 982 .word 0x9db3898e !bshuffle %f14,%f14,%f14 983 984 xor %g4, %o0, %o0 ! ^= rk[0] 985 xor %g5, %o1, %o1 986 .word 0x81b02308 !movxtod %o0,%f0 987 .word 0x85b02309 !movxtod %o1,%f2 988 989 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 990 .word 0x85b38d82 !fxor %f14,%f2,%f2 991 992 prefetch [%i1 + 63], 22 993 prefetch [%i0 + 16+63], 20 994 call _aes128_encrypt_1x 995 add %i0, 16, %i0 996 997 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 998 .word 0x85b38d82 !fxor %f14,%f2,%f2 999 1000 srax %g3, 63, %l7 ! next tweak value 1001 addcc %g2, %g2, %g2 1002 and %l7, 0x87, %l7 1003 .word 0x87b0c223 !addxc %g3,%g3,%g3 1004 xor %l7, %g2, %g2 1005 1006 brnz,pn %l2, 2f 1007 sub %i2, 1, %i2 1008 1009 std %f0, [%i1 + 0] 1010 std %f2, [%i1 + 8] 1011 brnz,pt %i2, .L128_xts_enloop2x 1012 add %i1, 16, %i1 1013 1014 brnz,pn %i5, .L128_xts_ensteal 1015 nop 1016 1017 ret 1018 restore 1019 1020 .align 16 1021 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1022 ! and ~3x deterioration 1023 ! in inp==out case 1024 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 1025 .word 0x8db00902 !faligndata %f0,%f2,%f6 1026 .word 0x91b08902 !faligndata %f2,%f2,%f8 1027 stda %f4, [%i1 + %l3]0xc0 ! partial store 1028 std %f6, [%i1 + 8] 1029 add %i1, 16, %i1 1030 orn %g0, %l3, %l3 1031 stda %f8, [%i1 + %l3]0xc0 ! partial store 1032 1033 brnz,pt %i2, .L128_xts_enloop2x+4 1034 orn %g0, %l3, %l3 1035 1036 brnz,pn %i5, .L128_xts_ensteal 1037 nop 1038 1039 ret 1040 restore 1041 1042 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1043 .align 32 1044 .L128_xts_enloop2x: 1045 ldx [%i0 + 0], %o0 1046 ldx [%i0 + 8], %o1 1047 ldx [%i0 + 16], %o2 1048 brz,pt %l0, 4f 1049 ldx [%i0 + 24], %o3 1050 1051 ldx [%i0 + 32], %o4 1052 sllx %o0, %l0, %o0 1053 srlx %o1, %l1, %g1 1054 or %g1, %o0, %o0 1055 sllx %o1, %l0, %o1 1056 srlx %o2, %l1, %g1 1057 or %g1, %o1, %o1 1058 sllx %o2, %l0, %o2 1059 srlx %o3, %l1, %g1 1060 or %g1, %o2, %o2 1061 sllx %o3, %l0, %o3 1062 srlx %o4, %l1, %o4 1063 or %o4, %o3, %o3 1064 4: 1065 .word 0x99b02302 !movxtod %g2,%f12 1066 .word 0x9db02303 !movxtod %g3,%f14 1067 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1068 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1069 1070 srax %g3, 63, %l7 ! next tweak value 1071 addcc %g2, %g2, %g2 1072 and %l7, 0x87, %l7 1073 .word 0x87b0c223 !addxc %g3,%g3,%g3 1074 xor %l7, %g2, %g2 1075 1076 .word 0x91b02302 !movxtod %g2,%f8 1077 .word 0x95b02303 !movxtod %g3,%f10 1078 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1079 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1080 1081 xor %g4, %o0, %o0 ! ^= rk[0] 1082 xor %g5, %o1, %o1 1083 xor %g4, %o2, %o2 ! ^= rk[0] 1084 xor %g5, %o3, %o3 1085 .word 0x81b02308 !movxtod %o0,%f0 1086 .word 0x85b02309 !movxtod %o1,%f2 1087 .word 0x89b0230a !movxtod %o2,%f4 1088 .word 0x8db0230b !movxtod %o3,%f6 1089 1090 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1091 .word 0x85b38d82 !fxor %f14,%f2,%f2 1092 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1093 .word 0x8db28d86 !fxor %f10,%f6,%f6 1094 1095 prefetch [%i1 + 63], 22 1096 prefetch [%i0 + 32+63], 20 1097 call _aes128_encrypt_2x 1098 add %i0, 32, %i0 1099 1100 .word 0x91b02302 !movxtod %g2,%f8 1101 .word 0x95b02303 !movxtod %g3,%f10 1102 1103 srax %g3, 63, %l7 ! next tweak value 1104 addcc %g2, %g2, %g2 1105 and %l7, 0x87, %l7 1106 .word 0x87b0c223 !addxc %g3,%g3,%g3 1107 xor %l7, %g2, %g2 1108 1109 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1110 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1111 1112 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1113 .word 0x85b38d82 !fxor %f14,%f2,%f2 1114 .word 0x89b20d84 !fxor %f8,%f4,%f4 1115 .word 0x8db28d86 !fxor %f10,%f6,%f6 1116 1117 brnz,pn %l2, 2f 1118 sub %i2, 2, %i2 1119 1120 std %f0, [%i1 + 0] 1121 std %f2, [%i1 + 8] 1122 std %f4, [%i1 + 16] 1123 std %f6, [%i1 + 24] 1124 brnz,pt %i2, .L128_xts_enloop2x 1125 add %i1, 32, %i1 1126 1127 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1128 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1129 brnz,pn %i5, .L128_xts_ensteal 1130 nop 1131 1132 ret 1133 restore 1134 1135 .align 16 1136 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1137 ! and ~3x deterioration 1138 ! in inp==out case 1139 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 1140 .word 0x95b00902 !faligndata %f0,%f2,%f10 1141 .word 0x99b08904 !faligndata %f2,%f4,%f12 1142 .word 0x9db10906 !faligndata %f4,%f6,%f14 1143 .word 0x81b18906 !faligndata %f6,%f6,%f0 1144 1145 stda %f8, [%i1 + %l3]0xc0 ! partial store 1146 std %f10, [%i1 + 8] 1147 std %f12, [%i1 + 16] 1148 std %f14, [%i1 + 24] 1149 add %i1, 32, %i1 1150 orn %g0, %l3, %l3 1151 stda %f0, [%i1 + %l3]0xc0 ! partial store 1152 1153 brnz,pt %i2, .L128_xts_enloop2x+4 1154 orn %g0, %l3, %l3 1155 1156 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1157 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1158 brnz,pn %i5, .L128_xts_ensteal 1159 nop 1160 1161 ret 1162 restore 1163 1164 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1165 .align 32 1166 .L128_xts_enblk: 1167 add %i1, %i2, %l5 1168 and %l5, 63, %l5 ! tail 1169 sub %i2, %l5, %i2 1170 add %l5, 15, %l5 ! round up to 16n 1171 srlx %i2, 4, %i2 1172 srl %l5, 4, %l5 1173 sub %i2, 1, %i2 1174 add %l5, 1, %l5 1175 1176 .L128_xts_enblk2x: 1177 ldx [%i0 + 0], %o0 1178 ldx [%i0 + 8], %o1 1179 ldx [%i0 + 16], %o2 1180 brz,pt %l0, 5f 1181 ldx [%i0 + 24], %o3 1182 1183 ldx [%i0 + 32], %o4 1184 sllx %o0, %l0, %o0 1185 srlx %o1, %l1, %g1 1186 or %g1, %o0, %o0 1187 sllx %o1, %l0, %o1 1188 srlx %o2, %l1, %g1 1189 or %g1, %o1, %o1 1190 sllx %o2, %l0, %o2 1191 srlx %o3, %l1, %g1 1192 or %g1, %o2, %o2 1193 sllx %o3, %l0, %o3 1194 srlx %o4, %l1, %o4 1195 or %o4, %o3, %o3 1196 5: 1197 .word 0x99b02302 !movxtod %g2,%f12 1198 .word 0x9db02303 !movxtod %g3,%f14 1199 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1200 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1201 1202 srax %g3, 63, %l7 ! next tweak value 1203 addcc %g2, %g2, %g2 1204 and %l7, 0x87, %l7 1205 .word 0x87b0c223 !addxc %g3,%g3,%g3 1206 xor %l7, %g2, %g2 1207 1208 .word 0x91b02302 !movxtod %g2,%f8 1209 .word 0x95b02303 !movxtod %g3,%f10 1210 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1211 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1212 1213 xor %g4, %o0, %o0 ! ^= rk[0] 1214 xor %g5, %o1, %o1 1215 xor %g4, %o2, %o2 ! ^= rk[0] 1216 xor %g5, %o3, %o3 1217 .word 0x81b02308 !movxtod %o0,%f0 1218 .word 0x85b02309 !movxtod %o1,%f2 1219 .word 0x89b0230a !movxtod %o2,%f4 1220 .word 0x8db0230b !movxtod %o3,%f6 1221 1222 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1223 .word 0x85b38d82 !fxor %f14,%f2,%f2 1224 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1225 .word 0x8db28d86 !fxor %f10,%f6,%f6 1226 1227 prefetch [%i0 + 32+63], 20 1228 call _aes128_encrypt_2x 1229 add %i0, 32, %i0 1230 1231 .word 0x91b02302 !movxtod %g2,%f8 1232 .word 0x95b02303 !movxtod %g3,%f10 1233 1234 srax %g3, 63, %l7 ! next tweak value 1235 addcc %g2, %g2, %g2 1236 and %l7, 0x87, %l7 1237 .word 0x87b0c223 !addxc %g3,%g3,%g3 1238 xor %l7, %g2, %g2 1239 1240 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1241 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1242 1243 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1244 .word 0x85b38d82 !fxor %f14,%f2,%f2 1245 .word 0x89b20d84 !fxor %f8,%f4,%f4 1246 .word 0x8db28d86 !fxor %f10,%f6,%f6 1247 1248 subcc %i2, 2, %i2 1249 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1250 add %i1, 8, %i1 1251 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1252 add %i1, 8, %i1 1253 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1254 add %i1, 8, %i1 1255 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1256 bgu,pt %icc, .L128_xts_enblk2x 1257 add %i1, 8, %i1 1258 1259 add %l5, %i2, %i2 1260 andcc %i2, 1, %g0 ! is number of blocks even? 1261 membar #StoreLoad|#StoreStore 1262 bnz,pt %icc, .L128_xts_enloop 1263 srl %i2, 0, %i2 1264 brnz,pn %i2, .L128_xts_enloop2x 1265 nop 1266 1267 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1268 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1269 brnz,pn %i5, .L128_xts_ensteal 1270 nop 1271 1272 ret 1273 restore 1274 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1275 .align 32 1276 .L128_xts_ensteal: 1277 std %f0, [%fp + 0-16] ! copy of output 1278 std %f2, [%fp + 0-8] 1279 1280 srl %l0, 3, %l0 1281 add %fp, 0-16, %l7 1282 add %i0, %l0, %i0 ! original %i0+%i2&-15 1283 add %i1, %l2, %i1 ! original %i1+%i2&-15 1284 mov 0, %l0 1285 nop ! align 1286 1287 .L128_xts_enstealing: 1288 ldub [%i0 + %l0], %o0 1289 ldub [%l7 + %l0], %o1 1290 dec %i5 1291 stb %o0, [%l7 + %l0] 1292 stb %o1, [%i1 + %l0] 1293 brnz %i5, .L128_xts_enstealing 1294 inc %l0 1295 1296 mov %l7, %i0 1297 sub %i1, 16, %i1 1298 mov 0, %l0 1299 sub %i1, %l2, %i1 1300 ba .L128_xts_enloop ! one more time 1301 mov 1, %i2 ! %i5 is 0 1302 ret 1303 restore 1304 .type aes128_t4_xts_encrypt,#function 1305 .size aes128_t4_xts_encrypt,.-aes128_t4_xts_encrypt 1306 .globl aes128_t4_xts_decrypt 1307 .align 32 1308 aes128_t4_xts_decrypt: 1309 save %sp, -112-16, %sp 1310 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 1311 1312 mov %i5, %o0 1313 add %fp, 0-16, %o1 1314 call aes_t4_encrypt 1315 mov %i4, %o2 1316 1317 add %fp, 0-16, %l7 1318 ldxa [%l7]0x88, %g2 1319 add %fp, 0-8, %l7 1320 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 1321 1322 sethi %hi(0x76543210), %l7 1323 or %l7, %lo(0x76543210), %l7 1324 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 1325 1326 prefetch [%i0], 20 1327 prefetch [%i0 + 63], 20 1328 call _aes128_load_deckey 1329 and %i2, 15, %i5 1330 and %i2, -16, %i2 1331 mov 0, %l7 1332 movrnz %i5, 16, %l7 1333 sub %i2, %l7, %i2 1334 1335 sub %i0, %i1, %l5 ! %i0!=%i1 1336 and %i0, 7, %l0 1337 andn %i0, 7, %i0 1338 sll %l0, 3, %l0 1339 mov 64, %l1 1340 mov 0xff, %l3 1341 sub %l1, %l0, %l1 1342 and %i1, 7, %l2 1343 cmp %i2, 255 1344 movrnz %l2, 0, %l5 ! if ( %i1&7 || 1345 movleu %icc, 0, %l5 ! %i2<256 || 1346 brnz,pn %l5, .L128_xts_deblk ! %i0==%i1) 1347 srl %l3, %l2, %l3 1348 1349 andcc %i2, 16, %g0 ! is number of blocks even? 1350 brz,pn %i2, .L128_xts_desteal 1351 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 1352 bz %icc, .L128_xts_deloop2x 1353 srlx %i2, 4, %i2 1354 .L128_xts_deloop: 1355 ldx [%i0 + 0], %o0 1356 brz,pt %l0, 4f 1357 ldx [%i0 + 8], %o1 1358 1359 ldx [%i0 + 16], %o2 1360 sllx %o0, %l0, %o0 1361 srlx %o1, %l1, %g1 1362 sllx %o1, %l0, %o1 1363 or %g1, %o0, %o0 1364 srlx %o2, %l1, %o2 1365 or %o2, %o1, %o1 1366 4: 1367 .word 0x99b02302 !movxtod %g2,%f12 1368 .word 0x9db02303 !movxtod %g3,%f14 1369 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1370 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1371 1372 xor %g4, %o0, %o0 ! ^= rk[0] 1373 xor %g5, %o1, %o1 1374 .word 0x81b02308 !movxtod %o0,%f0 1375 .word 0x85b02309 !movxtod %o1,%f2 1376 1377 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1378 .word 0x85b38d82 !fxor %f14,%f2,%f2 1379 1380 prefetch [%i1 + 63], 22 1381 prefetch [%i0 + 16+63], 20 1382 call _aes128_decrypt_1x 1383 add %i0, 16, %i0 1384 1385 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1386 .word 0x85b38d82 !fxor %f14,%f2,%f2 1387 1388 srax %g3, 63, %l7 ! next tweak value 1389 addcc %g2, %g2, %g2 1390 and %l7, 0x87, %l7 1391 .word 0x87b0c223 !addxc %g3,%g3,%g3 1392 xor %l7, %g2, %g2 1393 1394 brnz,pn %l2, 2f 1395 sub %i2, 1, %i2 1396 1397 std %f0, [%i1 + 0] 1398 std %f2, [%i1 + 8] 1399 brnz,pt %i2, .L128_xts_deloop2x 1400 add %i1, 16, %i1 1401 1402 brnz,pn %i5, .L128_xts_desteal 1403 nop 1404 1405 ret 1406 restore 1407 1408 .align 16 1409 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1410 ! and ~3x deterioration 1411 ! in inp==out case 1412 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 1413 .word 0x8db00902 !faligndata %f0,%f2,%f6 1414 .word 0x91b08902 !faligndata %f2,%f2,%f8 1415 stda %f4, [%i1 + %l3]0xc0 ! partial store 1416 std %f6, [%i1 + 8] 1417 add %i1, 16, %i1 1418 orn %g0, %l3, %l3 1419 stda %f8, [%i1 + %l3]0xc0 ! partial store 1420 1421 brnz,pt %i2, .L128_xts_deloop2x+4 1422 orn %g0, %l3, %l3 1423 1424 brnz,pn %i5, .L128_xts_desteal 1425 nop 1426 1427 ret 1428 restore 1429 1430 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1431 .align 32 1432 .L128_xts_deloop2x: 1433 ldx [%i0 + 0], %o0 1434 ldx [%i0 + 8], %o1 1435 ldx [%i0 + 16], %o2 1436 brz,pt %l0, 4f 1437 ldx [%i0 + 24], %o3 1438 1439 ldx [%i0 + 32], %o4 1440 sllx %o0, %l0, %o0 1441 srlx %o1, %l1, %g1 1442 or %g1, %o0, %o0 1443 sllx %o1, %l0, %o1 1444 srlx %o2, %l1, %g1 1445 or %g1, %o1, %o1 1446 sllx %o2, %l0, %o2 1447 srlx %o3, %l1, %g1 1448 or %g1, %o2, %o2 1449 sllx %o3, %l0, %o3 1450 srlx %o4, %l1, %o4 1451 or %o4, %o3, %o3 1452 4: 1453 .word 0x99b02302 !movxtod %g2,%f12 1454 .word 0x9db02303 !movxtod %g3,%f14 1455 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1456 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1457 1458 srax %g3, 63, %l7 ! next tweak value 1459 addcc %g2, %g2, %g2 1460 and %l7, 0x87, %l7 1461 .word 0x87b0c223 !addxc %g3,%g3,%g3 1462 xor %l7, %g2, %g2 1463 1464 .word 0x91b02302 !movxtod %g2,%f8 1465 .word 0x95b02303 !movxtod %g3,%f10 1466 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1467 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1468 1469 xor %g4, %o0, %o0 ! ^= rk[0] 1470 xor %g5, %o1, %o1 1471 xor %g4, %o2, %o2 ! ^= rk[0] 1472 xor %g5, %o3, %o3 1473 .word 0x81b02308 !movxtod %o0,%f0 1474 .word 0x85b02309 !movxtod %o1,%f2 1475 .word 0x89b0230a !movxtod %o2,%f4 1476 .word 0x8db0230b !movxtod %o3,%f6 1477 1478 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1479 .word 0x85b38d82 !fxor %f14,%f2,%f2 1480 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1481 .word 0x8db28d86 !fxor %f10,%f6,%f6 1482 1483 prefetch [%i1 + 63], 22 1484 prefetch [%i0 + 32+63], 20 1485 call _aes128_decrypt_2x 1486 add %i0, 32, %i0 1487 1488 .word 0x91b02302 !movxtod %g2,%f8 1489 .word 0x95b02303 !movxtod %g3,%f10 1490 1491 srax %g3, 63, %l7 ! next tweak value 1492 addcc %g2, %g2, %g2 1493 and %l7, 0x87, %l7 1494 .word 0x87b0c223 !addxc %g3,%g3,%g3 1495 xor %l7, %g2, %g2 1496 1497 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1498 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1499 1500 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1501 .word 0x85b38d82 !fxor %f14,%f2,%f2 1502 .word 0x89b20d84 !fxor %f8,%f4,%f4 1503 .word 0x8db28d86 !fxor %f10,%f6,%f6 1504 1505 brnz,pn %l2, 2f 1506 sub %i2, 2, %i2 1507 1508 std %f0, [%i1 + 0] 1509 std %f2, [%i1 + 8] 1510 std %f4, [%i1 + 16] 1511 std %f6, [%i1 + 24] 1512 brnz,pt %i2, .L128_xts_deloop2x 1513 add %i1, 32, %i1 1514 1515 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1516 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1517 brnz,pn %i5, .L128_xts_desteal 1518 nop 1519 1520 ret 1521 restore 1522 1523 .align 16 1524 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1525 ! and ~3x deterioration 1526 ! in inp==out case 1527 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 1528 .word 0x95b00902 !faligndata %f0,%f2,%f10 1529 .word 0x99b08904 !faligndata %f2,%f4,%f12 1530 .word 0x9db10906 !faligndata %f4,%f6,%f14 1531 .word 0x81b18906 !faligndata %f6,%f6,%f0 1532 1533 stda %f8, [%i1 + %l3]0xc0 ! partial store 1534 std %f10, [%i1 + 8] 1535 std %f12, [%i1 + 16] 1536 std %f14, [%i1 + 24] 1537 add %i1, 32, %i1 1538 orn %g0, %l3, %l3 1539 stda %f0, [%i1 + %l3]0xc0 ! partial store 1540 1541 brnz,pt %i2, .L128_xts_deloop2x+4 1542 orn %g0, %l3, %l3 1543 1544 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1545 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1546 brnz,pn %i5, .L128_xts_desteal 1547 nop 1548 1549 ret 1550 restore 1551 1552 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1553 .align 32 1554 .L128_xts_deblk: 1555 add %i1, %i2, %l5 1556 and %l5, 63, %l5 ! tail 1557 sub %i2, %l5, %i2 1558 add %l5, 15, %l5 ! round up to 16n 1559 srlx %i2, 4, %i2 1560 srl %l5, 4, %l5 1561 sub %i2, 1, %i2 1562 add %l5, 1, %l5 1563 1564 .L128_xts_deblk2x: 1565 ldx [%i0 + 0], %o0 1566 ldx [%i0 + 8], %o1 1567 ldx [%i0 + 16], %o2 1568 brz,pt %l0, 5f 1569 ldx [%i0 + 24], %o3 1570 1571 ldx [%i0 + 32], %o4 1572 sllx %o0, %l0, %o0 1573 srlx %o1, %l1, %g1 1574 or %g1, %o0, %o0 1575 sllx %o1, %l0, %o1 1576 srlx %o2, %l1, %g1 1577 or %g1, %o1, %o1 1578 sllx %o2, %l0, %o2 1579 srlx %o3, %l1, %g1 1580 or %g1, %o2, %o2 1581 sllx %o3, %l0, %o3 1582 srlx %o4, %l1, %o4 1583 or %o4, %o3, %o3 1584 5: 1585 .word 0x99b02302 !movxtod %g2,%f12 1586 .word 0x9db02303 !movxtod %g3,%f14 1587 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1588 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1589 1590 srax %g3, 63, %l7 ! next tweak value 1591 addcc %g2, %g2, %g2 1592 and %l7, 0x87, %l7 1593 .word 0x87b0c223 !addxc %g3,%g3,%g3 1594 xor %l7, %g2, %g2 1595 1596 .word 0x91b02302 !movxtod %g2,%f8 1597 .word 0x95b02303 !movxtod %g3,%f10 1598 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1599 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1600 1601 xor %g4, %o0, %o0 ! ^= rk[0] 1602 xor %g5, %o1, %o1 1603 xor %g4, %o2, %o2 ! ^= rk[0] 1604 xor %g5, %o3, %o3 1605 .word 0x81b02308 !movxtod %o0,%f0 1606 .word 0x85b02309 !movxtod %o1,%f2 1607 .word 0x89b0230a !movxtod %o2,%f4 1608 .word 0x8db0230b !movxtod %o3,%f6 1609 1610 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1611 .word 0x85b38d82 !fxor %f14,%f2,%f2 1612 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1613 .word 0x8db28d86 !fxor %f10,%f6,%f6 1614 1615 prefetch [%i0 + 32+63], 20 1616 call _aes128_decrypt_2x 1617 add %i0, 32, %i0 1618 1619 .word 0x91b02302 !movxtod %g2,%f8 1620 .word 0x95b02303 !movxtod %g3,%f10 1621 1622 srax %g3, 63, %l7 ! next tweak value 1623 addcc %g2, %g2, %g2 1624 and %l7, 0x87, %l7 1625 .word 0x87b0c223 !addxc %g3,%g3,%g3 1626 xor %l7, %g2, %g2 1627 1628 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1629 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1630 1631 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1632 .word 0x85b38d82 !fxor %f14,%f2,%f2 1633 .word 0x89b20d84 !fxor %f8,%f4,%f4 1634 .word 0x8db28d86 !fxor %f10,%f6,%f6 1635 1636 subcc %i2, 2, %i2 1637 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1638 add %i1, 8, %i1 1639 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1640 add %i1, 8, %i1 1641 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1642 add %i1, 8, %i1 1643 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1644 bgu,pt %icc, .L128_xts_deblk2x 1645 add %i1, 8, %i1 1646 1647 add %l5, %i2, %i2 1648 andcc %i2, 1, %g0 ! is number of blocks even? 1649 membar #StoreLoad|#StoreStore 1650 bnz,pt %icc, .L128_xts_deloop 1651 srl %i2, 0, %i2 1652 brnz,pn %i2, .L128_xts_deloop2x 1653 nop 1654 1655 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1656 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1657 brnz,pn %i5, .L128_xts_desteal 1658 nop 1659 1660 ret 1661 restore 1662 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1663 .align 32 1664 .L128_xts_desteal: 1665 ldx [%i0 + 0], %o0 1666 brz,pt %l0, 8f 1667 ldx [%i0 + 8], %o1 1668 1669 ldx [%i0 + 16], %o2 1670 sllx %o0, %l0, %o0 1671 srlx %o1, %l1, %g1 1672 sllx %o1, %l0, %o1 1673 or %g1, %o0, %o0 1674 srlx %o2, %l1, %o2 1675 or %o2, %o1, %o1 1676 8: 1677 srax %g3, 63, %l7 ! next tweak value 1678 addcc %g2, %g2, %o2 1679 and %l7, 0x87, %l7 1680 .word 0x97b0c223 !addxc %g3,%g3,%o3 1681 xor %l7, %o2, %o2 1682 1683 .word 0x99b0230a !movxtod %o2,%f12 1684 .word 0x9db0230b !movxtod %o3,%f14 1685 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1686 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1687 1688 xor %g4, %o0, %o0 ! ^= rk[0] 1689 xor %g5, %o1, %o1 1690 .word 0x81b02308 !movxtod %o0,%f0 1691 .word 0x85b02309 !movxtod %o1,%f2 1692 1693 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1694 .word 0x85b38d82 !fxor %f14,%f2,%f2 1695 1696 call _aes128_decrypt_1x 1697 add %i0, 16, %i0 1698 1699 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1700 .word 0x85b38d82 !fxor %f14,%f2,%f2 1701 1702 std %f0, [%fp + 0-16] 1703 std %f2, [%fp + 0-8] 1704 1705 srl %l0, 3, %l0 1706 add %fp, 0-16, %l7 1707 add %i0, %l0, %i0 ! original %i0+%i2&-15 1708 add %i1, %l2, %i1 ! original %i1+%i2&-15 1709 mov 0, %l0 1710 add %i1, 16, %i1 1711 nop ! align 1712 1713 .L128_xts_destealing: 1714 ldub [%i0 + %l0], %o0 1715 ldub [%l7 + %l0], %o1 1716 dec %i5 1717 stb %o0, [%l7 + %l0] 1718 stb %o1, [%i1 + %l0] 1719 brnz %i5, .L128_xts_destealing 1720 inc %l0 1721 1722 mov %l7, %i0 1723 sub %i1, 16, %i1 1724 mov 0, %l0 1725 sub %i1, %l2, %i1 1726 ba .L128_xts_deloop ! one more time 1727 mov 1, %i2 ! %i5 is 0 1728 ret 1729 restore 1730 .type aes128_t4_xts_decrypt,#function 1731 .size aes128_t4_xts_decrypt,.-aes128_t4_xts_decrypt 1732 .globl aes128_t4_cbc_decrypt 1733 .align 32 1734 aes128_t4_cbc_decrypt: 1735 save %sp, -112, %sp 1736 cmp %i2, 0 1737 be,pn %icc, .L128_cbc_dec_abort 1738 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 1739 sub %i0, %i1, %l5 ! %i0!=%i1 1740 ld [%i4 + 0], %f12 ! load ivec 1741 ld [%i4 + 4], %f13 1742 ld [%i4 + 8], %f14 1743 ld [%i4 + 12], %f15 1744 prefetch [%i0], 20 1745 prefetch [%i0 + 63], 20 1746 call _aes128_load_deckey 1747 and %i0, 7, %l0 1748 andn %i0, 7, %i0 1749 sll %l0, 3, %l0 1750 mov 64, %l1 1751 mov 0xff, %l3 1752 sub %l1, %l0, %l1 1753 and %i1, 7, %l2 1754 cmp %i2, 255 1755 movrnz %l2, 0, %l5 ! if ( %i1&7 || 1756 movleu %icc, 0, %l5 ! %i2<256 || 1757 brnz,pn %l5, .L128cbc_dec_blk ! %i0==%i1) 1758 srl %l3, %l2, %l3 1759 1760 andcc %i2, 16, %g0 ! is number of blocks even? 1761 srlx %i2, 4, %i2 1762 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 1763 bz %icc, .L128_cbc_dec_loop2x 1764 prefetch [%i1], 22 1765 .L128_cbc_dec_loop: 1766 ldx [%i0 + 0], %o0 1767 brz,pt %l0, 4f 1768 ldx [%i0 + 8], %o1 1769 1770 ldx [%i0 + 16], %o2 1771 sllx %o0, %l0, %o0 1772 srlx %o1, %l1, %g1 1773 sllx %o1, %l0, %o1 1774 or %g1, %o0, %o0 1775 srlx %o2, %l1, %o2 1776 or %o2, %o1, %o1 1777 4: 1778 xor %g4, %o0, %o2 ! ^= rk[0] 1779 xor %g5, %o1, %o3 1780 .word 0x81b0230a !movxtod %o2,%f0 1781 .word 0x85b0230b !movxtod %o3,%f2 1782 1783 prefetch [%i1 + 63], 22 1784 prefetch [%i0 + 16+63], 20 1785 call _aes128_decrypt_1x 1786 add %i0, 16, %i0 1787 1788 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 1789 .word 0x85b38d82 !fxor %f14,%f2,%f2 1790 .word 0x99b02308 !movxtod %o0,%f12 1791 .word 0x9db02309 !movxtod %o1,%f14 1792 1793 brnz,pn %l2, 2f 1794 sub %i2, 1, %i2 1795 1796 std %f0, [%i1 + 0] 1797 std %f2, [%i1 + 8] 1798 brnz,pt %i2, .L128_cbc_dec_loop2x 1799 add %i1, 16, %i1 1800 st %f12, [%i4 + 0] 1801 st %f13, [%i4 + 4] 1802 st %f14, [%i4 + 8] 1803 st %f15, [%i4 + 12] 1804 .L128_cbc_dec_abort: 1805 ret 1806 restore 1807 1808 .align 16 1809 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1810 ! and ~3x deterioration 1811 ! in inp==out case 1812 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 1813 .word 0x8db00902 !faligndata %f0,%f2,%f6 1814 .word 0x91b08902 !faligndata %f2,%f2,%f8 1815 1816 stda %f4, [%i1 + %l3]0xc0 ! partial store 1817 std %f6, [%i1 + 8] 1818 add %i1, 16, %i1 1819 orn %g0, %l3, %l3 1820 stda %f8, [%i1 + %l3]0xc0 ! partial store 1821 1822 brnz,pt %i2, .L128_cbc_dec_loop2x+4 1823 orn %g0, %l3, %l3 1824 st %f12, [%i4 + 0] 1825 st %f13, [%i4 + 4] 1826 st %f14, [%i4 + 8] 1827 st %f15, [%i4 + 12] 1828 ret 1829 restore 1830 1831 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1832 .align 32 1833 .L128_cbc_dec_loop2x: 1834 ldx [%i0 + 0], %o0 1835 ldx [%i0 + 8], %o1 1836 ldx [%i0 + 16], %o2 1837 brz,pt %l0, 4f 1838 ldx [%i0 + 24], %o3 1839 1840 ldx [%i0 + 32], %o4 1841 sllx %o0, %l0, %o0 1842 srlx %o1, %l1, %g1 1843 or %g1, %o0, %o0 1844 sllx %o1, %l0, %o1 1845 srlx %o2, %l1, %g1 1846 or %g1, %o1, %o1 1847 sllx %o2, %l0, %o2 1848 srlx %o3, %l1, %g1 1849 or %g1, %o2, %o2 1850 sllx %o3, %l0, %o3 1851 srlx %o4, %l1, %o4 1852 or %o4, %o3, %o3 1853 4: 1854 xor %g4, %o0, %o4 ! ^= rk[0] 1855 xor %g5, %o1, %o5 1856 .word 0x81b0230c !movxtod %o4,%f0 1857 .word 0x85b0230d !movxtod %o5,%f2 1858 xor %g4, %o2, %o4 1859 xor %g5, %o3, %o5 1860 .word 0x89b0230c !movxtod %o4,%f4 1861 .word 0x8db0230d !movxtod %o5,%f6 1862 1863 prefetch [%i1 + 63], 22 1864 prefetch [%i0 + 32+63], 20 1865 call _aes128_decrypt_2x 1866 add %i0, 32, %i0 1867 1868 .word 0x91b02308 !movxtod %o0,%f8 1869 .word 0x95b02309 !movxtod %o1,%f10 1870 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 1871 .word 0x85b38d82 !fxor %f14,%f2,%f2 1872 .word 0x99b0230a !movxtod %o2,%f12 1873 .word 0x9db0230b !movxtod %o3,%f14 1874 .word 0x89b20d84 !fxor %f8,%f4,%f4 1875 .word 0x8db28d86 !fxor %f10,%f6,%f6 1876 1877 brnz,pn %l2, 2f 1878 sub %i2, 2, %i2 1879 1880 std %f0, [%i1 + 0] 1881 std %f2, [%i1 + 8] 1882 std %f4, [%i1 + 16] 1883 std %f6, [%i1 + 24] 1884 brnz,pt %i2, .L128_cbc_dec_loop2x 1885 add %i1, 32, %i1 1886 st %f12, [%i4 + 0] 1887 st %f13, [%i4 + 4] 1888 st %f14, [%i4 + 8] 1889 st %f15, [%i4 + 12] 1890 ret 1891 restore 1892 1893 .align 16 1894 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1895 ! and ~3x deterioration 1896 ! in inp==out case 1897 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 1898 .word 0x81b00902 !faligndata %f0,%f2,%f0 1899 .word 0x85b08904 !faligndata %f2,%f4,%f2 1900 .word 0x89b10906 !faligndata %f4,%f6,%f4 1901 .word 0x8db18906 !faligndata %f6,%f6,%f6 1902 stda %f8, [%i1 + %l3]0xc0 ! partial store 1903 std %f0, [%i1 + 8] 1904 std %f2, [%i1 + 16] 1905 std %f4, [%i1 + 24] 1906 add %i1, 32, %i1 1907 orn %g0, %l3, %l3 1908 stda %f6, [%i1 + %l3]0xc0 ! partial store 1909 1910 brnz,pt %i2, .L128_cbc_dec_loop2x+4 1911 orn %g0, %l3, %l3 1912 st %f12, [%i4 + 0] 1913 st %f13, [%i4 + 4] 1914 st %f14, [%i4 + 8] 1915 st %f15, [%i4 + 12] 1916 ret 1917 restore 1918 1919 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1920 .align 32 1921 .L128cbc_dec_blk: 1922 add %i1, %i2, %l5 1923 and %l5, 63, %l5 ! tail 1924 sub %i2, %l5, %i2 1925 add %l5, 15, %l5 ! round up to 16n 1926 srlx %i2, 4, %i2 1927 srl %l5, 4, %l5 1928 sub %i2, 1, %i2 1929 add %l5, 1, %l5 1930 1931 .L128_cbc_dec_blk_loop2x: 1932 ldx [%i0 + 0], %o0 1933 ldx [%i0 + 8], %o1 1934 ldx [%i0 + 16], %o2 1935 brz,pt %l0, 5f 1936 ldx [%i0 + 24], %o3 1937 1938 ldx [%i0 + 32], %o4 1939 sllx %o0, %l0, %o0 1940 srlx %o1, %l1, %g1 1941 or %g1, %o0, %o0 1942 sllx %o1, %l0, %o1 1943 srlx %o2, %l1, %g1 1944 or %g1, %o1, %o1 1945 sllx %o2, %l0, %o2 1946 srlx %o3, %l1, %g1 1947 or %g1, %o2, %o2 1948 sllx %o3, %l0, %o3 1949 srlx %o4, %l1, %o4 1950 or %o4, %o3, %o3 1951 5: 1952 xor %g4, %o0, %o4 ! ^= rk[0] 1953 xor %g5, %o1, %o5 1954 .word 0x81b0230c !movxtod %o4,%f0 1955 .word 0x85b0230d !movxtod %o5,%f2 1956 xor %g4, %o2, %o4 1957 xor %g5, %o3, %o5 1958 .word 0x89b0230c !movxtod %o4,%f4 1959 .word 0x8db0230d !movxtod %o5,%f6 1960 1961 prefetch [%i0 + 32+63], 20 1962 call _aes128_decrypt_2x 1963 add %i0, 32, %i0 1964 subcc %i2, 2, %i2 1965 1966 .word 0x91b02308 !movxtod %o0,%f8 1967 .word 0x95b02309 !movxtod %o1,%f10 1968 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 1969 .word 0x85b38d82 !fxor %f14,%f2,%f2 1970 .word 0x99b0230a !movxtod %o2,%f12 1971 .word 0x9db0230b !movxtod %o3,%f14 1972 .word 0x89b20d84 !fxor %f8,%f4,%f4 1973 .word 0x8db28d86 !fxor %f10,%f6,%f6 1974 1975 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1976 add %i1, 8, %i1 1977 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1978 add %i1, 8, %i1 1979 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1980 add %i1, 8, %i1 1981 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1982 bgu,pt %icc, .L128_cbc_dec_blk_loop2x 1983 add %i1, 8, %i1 1984 1985 add %l5, %i2, %i2 1986 andcc %i2, 1, %g0 ! is number of blocks even? 1987 membar #StoreLoad|#StoreStore 1988 bnz,pt %icc, .L128_cbc_dec_loop 1989 srl %i2, 0, %i2 1990 brnz,pn %i2, .L128_cbc_dec_loop2x 1991 nop 1992 st %f12, [%i4 + 0] ! write out ivec 1993 st %f13, [%i4 + 4] 1994 st %f14, [%i4 + 8] 1995 st %f15, [%i4 + 12] 1996 ret 1997 restore 1998 .type aes128_t4_cbc_decrypt,#function 1999 .size aes128_t4_cbc_decrypt,.-aes128_t4_cbc_decrypt 2000 .align 32 2001 _aes128_decrypt_1x: 2002 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 2003 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 2004 .word 0x80cd0444 !aes_dround01 %f20,%f4,%f2,%f0 2005 .word 0x84cd8464 !aes_dround23 %f22,%f4,%f2,%f2 2006 .word 0x88ce0440 !aes_dround01 %f24,%f0,%f2,%f4 2007 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 2008 .word 0x80cf0444 !aes_dround01 %f28,%f4,%f2,%f0 2009 .word 0x84cf8464 !aes_dround23 %f30,%f4,%f2,%f2 2010 .word 0x88c84440 !aes_dround01 %f32,%f0,%f2,%f4 2011 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 2012 .word 0x80c94444 !aes_dround01 %f36,%f4,%f2,%f0 2013 .word 0x84c9c464 !aes_dround23 %f38,%f4,%f2,%f2 2014 .word 0x88ca4440 !aes_dround01 %f40,%f0,%f2,%f4 2015 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 2016 .word 0x80cb4444 !aes_dround01 %f44,%f4,%f2,%f0 2017 .word 0x84cbc464 !aes_dround23 %f46,%f4,%f2,%f2 2018 .word 0x88cc4440 !aes_dround01 %f48,%f0,%f2,%f4 2019 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 2020 .word 0x80cd44c4 !aes_dround01_l %f52,%f4,%f2,%f0 2021 retl 2022 .word 0x84cdc4e4 !aes_dround23_l %f54,%f4,%f2,%f2 2023 .type _aes128_decrypt_1x,#function 2024 .size _aes128_decrypt_1x,.-_aes128_decrypt_1x 2025 2026 .align 32 2027 _aes128_decrypt_2x: 2028 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 2029 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 2030 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 2031 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 2032 .word 0x80cd0448 !aes_dround01 %f20,%f8,%f2,%f0 2033 .word 0x84cd8468 !aes_dround23 %f22,%f8,%f2,%f2 2034 .word 0x88cd0c4a !aes_dround01 %f20,%f10,%f6,%f4 2035 .word 0x8ccd8c6a !aes_dround23 %f22,%f10,%f6,%f6 2036 .word 0x90ce0440 !aes_dround01 %f24,%f0,%f2,%f8 2037 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 2038 .word 0x94ce0c44 !aes_dround01 %f24,%f4,%f6,%f10 2039 .word 0x8cce8c64 !aes_dround23 %f26,%f4,%f6,%f6 2040 .word 0x80cf0448 !aes_dround01 %f28,%f8,%f2,%f0 2041 .word 0x84cf8468 !aes_dround23 %f30,%f8,%f2,%f2 2042 .word 0x88cf0c4a !aes_dround01 %f28,%f10,%f6,%f4 2043 .word 0x8ccf8c6a !aes_dround23 %f30,%f10,%f6,%f6 2044 .word 0x90c84440 !aes_dround01 %f32,%f0,%f2,%f8 2045 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 2046 .word 0x94c84c44 !aes_dround01 %f32,%f4,%f6,%f10 2047 .word 0x8cc8cc64 !aes_dround23 %f34,%f4,%f6,%f6 2048 .word 0x80c94448 !aes_dround01 %f36,%f8,%f2,%f0 2049 .word 0x84c9c468 !aes_dround23 %f38,%f8,%f2,%f2 2050 .word 0x88c94c4a !aes_dround01 %f36,%f10,%f6,%f4 2051 .word 0x8cc9cc6a !aes_dround23 %f38,%f10,%f6,%f6 2052 .word 0x90ca4440 !aes_dround01 %f40,%f0,%f2,%f8 2053 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 2054 .word 0x94ca4c44 !aes_dround01 %f40,%f4,%f6,%f10 2055 .word 0x8ccacc64 !aes_dround23 %f42,%f4,%f6,%f6 2056 .word 0x80cb4448 !aes_dround01 %f44,%f8,%f2,%f0 2057 .word 0x84cbc468 !aes_dround23 %f46,%f8,%f2,%f2 2058 .word 0x88cb4c4a !aes_dround01 %f44,%f10,%f6,%f4 2059 .word 0x8ccbcc6a !aes_dround23 %f46,%f10,%f6,%f6 2060 .word 0x90cc4440 !aes_dround01 %f48,%f0,%f2,%f8 2061 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 2062 .word 0x94cc4c44 !aes_dround01 %f48,%f4,%f6,%f10 2063 .word 0x8ccccc64 !aes_dround23 %f50,%f4,%f6,%f6 2064 .word 0x80cd44c8 !aes_dround01_l %f52,%f8,%f2,%f0 2065 .word 0x84cdc4e8 !aes_dround23_l %f54,%f8,%f2,%f2 2066 .word 0x88cd4cca !aes_dround01_l %f52,%f10,%f6,%f4 2067 retl 2068 .word 0x8ccdccea !aes_dround23_l %f54,%f10,%f6,%f6 2069 .type _aes128_decrypt_2x,#function 2070 .size _aes128_decrypt_2x,.-_aes128_decrypt_2x 2071 .align 32 2072 _aes192_encrypt_1x: 2073 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 2074 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2075 .word 0x80cd0404 !aes_eround01 %f20,%f4,%f2,%f0 2076 .word 0x84cd8424 !aes_eround23 %f22,%f4,%f2,%f2 2077 .word 0x88ce0400 !aes_eround01 %f24,%f0,%f2,%f4 2078 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2079 .word 0x80cf0404 !aes_eround01 %f28,%f4,%f2,%f0 2080 .word 0x84cf8424 !aes_eround23 %f30,%f4,%f2,%f2 2081 .word 0x88c84400 !aes_eround01 %f32,%f0,%f2,%f4 2082 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2083 .word 0x80c94404 !aes_eround01 %f36,%f4,%f2,%f0 2084 .word 0x84c9c424 !aes_eround23 %f38,%f4,%f2,%f2 2085 .word 0x88ca4400 !aes_eround01 %f40,%f0,%f2,%f4 2086 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2087 .word 0x80cb4404 !aes_eround01 %f44,%f4,%f2,%f0 2088 .word 0x84cbc424 !aes_eround23 %f46,%f4,%f2,%f2 2089 .word 0x88cc4400 !aes_eround01 %f48,%f0,%f2,%f4 2090 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2091 .word 0x80cd4404 !aes_eround01 %f52,%f4,%f2,%f0 2092 .word 0x84cdc424 !aes_eround23 %f54,%f4,%f2,%f2 2093 .word 0x88ce4400 !aes_eround01 %f56,%f0,%f2,%f4 2094 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2095 .word 0x80cf4484 !aes_eround01_l %f60,%f4,%f2,%f0 2096 retl 2097 .word 0x84cfc4a4 !aes_eround23_l %f62,%f4,%f2,%f2 2098 .type _aes192_encrypt_1x,#function 2099 .size _aes192_encrypt_1x,.-_aes192_encrypt_1x 2100 2101 .align 32 2102 _aes192_encrypt_2x: 2103 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 2104 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2105 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 2106 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 2107 .word 0x80cd0408 !aes_eround01 %f20,%f8,%f2,%f0 2108 .word 0x84cd8428 !aes_eround23 %f22,%f8,%f2,%f2 2109 .word 0x88cd0c0a !aes_eround01 %f20,%f10,%f6,%f4 2110 .word 0x8ccd8c2a !aes_eround23 %f22,%f10,%f6,%f6 2111 .word 0x90ce0400 !aes_eround01 %f24,%f0,%f2,%f8 2112 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2113 .word 0x94ce0c04 !aes_eround01 %f24,%f4,%f6,%f10 2114 .word 0x8cce8c24 !aes_eround23 %f26,%f4,%f6,%f6 2115 .word 0x80cf0408 !aes_eround01 %f28,%f8,%f2,%f0 2116 .word 0x84cf8428 !aes_eround23 %f30,%f8,%f2,%f2 2117 .word 0x88cf0c0a !aes_eround01 %f28,%f10,%f6,%f4 2118 .word 0x8ccf8c2a !aes_eround23 %f30,%f10,%f6,%f6 2119 .word 0x90c84400 !aes_eround01 %f32,%f0,%f2,%f8 2120 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2121 .word 0x94c84c04 !aes_eround01 %f32,%f4,%f6,%f10 2122 .word 0x8cc8cc24 !aes_eround23 %f34,%f4,%f6,%f6 2123 .word 0x80c94408 !aes_eround01 %f36,%f8,%f2,%f0 2124 .word 0x84c9c428 !aes_eround23 %f38,%f8,%f2,%f2 2125 .word 0x88c94c0a !aes_eround01 %f36,%f10,%f6,%f4 2126 .word 0x8cc9cc2a !aes_eround23 %f38,%f10,%f6,%f6 2127 .word 0x90ca4400 !aes_eround01 %f40,%f0,%f2,%f8 2128 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2129 .word 0x94ca4c04 !aes_eround01 %f40,%f4,%f6,%f10 2130 .word 0x8ccacc24 !aes_eround23 %f42,%f4,%f6,%f6 2131 .word 0x80cb4408 !aes_eround01 %f44,%f8,%f2,%f0 2132 .word 0x84cbc428 !aes_eround23 %f46,%f8,%f2,%f2 2133 .word 0x88cb4c0a !aes_eround01 %f44,%f10,%f6,%f4 2134 .word 0x8ccbcc2a !aes_eround23 %f46,%f10,%f6,%f6 2135 .word 0x90cc4400 !aes_eround01 %f48,%f0,%f2,%f8 2136 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2137 .word 0x94cc4c04 !aes_eround01 %f48,%f4,%f6,%f10 2138 .word 0x8ccccc24 !aes_eround23 %f50,%f4,%f6,%f6 2139 .word 0x80cd4408 !aes_eround01 %f52,%f8,%f2,%f0 2140 .word 0x84cdc428 !aes_eround23 %f54,%f8,%f2,%f2 2141 .word 0x88cd4c0a !aes_eround01 %f52,%f10,%f6,%f4 2142 .word 0x8ccdcc2a !aes_eround23 %f54,%f10,%f6,%f6 2143 .word 0x90ce4400 !aes_eround01 %f56,%f0,%f2,%f8 2144 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2145 .word 0x94ce4c04 !aes_eround01 %f56,%f4,%f6,%f10 2146 .word 0x8ccecc24 !aes_eround23 %f58,%f4,%f6,%f6 2147 .word 0x80cf4488 !aes_eround01_l %f60,%f8,%f2,%f0 2148 .word 0x84cfc4a8 !aes_eround23_l %f62,%f8,%f2,%f2 2149 .word 0x88cf4c8a !aes_eround01_l %f60,%f10,%f6,%f4 2150 retl 2151 .word 0x8ccfccaa !aes_eround23_l %f62,%f10,%f6,%f6 2152 .type _aes192_encrypt_2x,#function 2153 .size _aes192_encrypt_2x,.-_aes192_encrypt_2x 2154 2155 .align 32 2156 _aes256_encrypt_1x: 2157 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 2158 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2159 ldd [%i3 + 208], %f16 2160 ldd [%i3 + 216], %f18 2161 .word 0x80cd0404 !aes_eround01 %f20,%f4,%f2,%f0 2162 .word 0x84cd8424 !aes_eround23 %f22,%f4,%f2,%f2 2163 ldd [%i3 + 224], %f20 2164 ldd [%i3 + 232], %f22 2165 .word 0x88ce0400 !aes_eround01 %f24,%f0,%f2,%f4 2166 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2167 .word 0x80cf0404 !aes_eround01 %f28,%f4,%f2,%f0 2168 .word 0x84cf8424 !aes_eround23 %f30,%f4,%f2,%f2 2169 .word 0x88c84400 !aes_eround01 %f32,%f0,%f2,%f4 2170 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2171 .word 0x80c94404 !aes_eround01 %f36,%f4,%f2,%f0 2172 .word 0x84c9c424 !aes_eround23 %f38,%f4,%f2,%f2 2173 .word 0x88ca4400 !aes_eround01 %f40,%f0,%f2,%f4 2174 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2175 .word 0x80cb4404 !aes_eround01 %f44,%f4,%f2,%f0 2176 .word 0x84cbc424 !aes_eround23 %f46,%f4,%f2,%f2 2177 .word 0x88cc4400 !aes_eround01 %f48,%f0,%f2,%f4 2178 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2179 .word 0x80cd4404 !aes_eround01 %f52,%f4,%f2,%f0 2180 .word 0x84cdc424 !aes_eround23 %f54,%f4,%f2,%f2 2181 .word 0x88ce4400 !aes_eround01 %f56,%f0,%f2,%f4 2182 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2183 .word 0x80cf4404 !aes_eround01 %f60,%f4,%f2,%f0 2184 .word 0x84cfc424 !aes_eround23 %f62,%f4,%f2,%f2 2185 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 2186 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2187 ldd [%i3 + 16], %f16 2188 ldd [%i3 + 24], %f18 2189 .word 0x80cd0484 !aes_eround01_l %f20,%f4,%f2,%f0 2190 .word 0x84cd84a4 !aes_eround23_l %f22,%f4,%f2,%f2 2191 ldd [%i3 + 32], %f20 2192 retl 2193 ldd [%i3 + 40], %f22 2194 .type _aes256_encrypt_1x,#function 2195 .size _aes256_encrypt_1x,.-_aes256_encrypt_1x 2196 2197 .align 32 2198 _aes256_encrypt_2x: 2199 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 2200 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2201 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 2202 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 2203 ldd [%i3 + 208], %f16 2204 ldd [%i3 + 216], %f18 2205 .word 0x80cd0408 !aes_eround01 %f20,%f8,%f2,%f0 2206 .word 0x84cd8428 !aes_eround23 %f22,%f8,%f2,%f2 2207 .word 0x88cd0c0a !aes_eround01 %f20,%f10,%f6,%f4 2208 .word 0x8ccd8c2a !aes_eround23 %f22,%f10,%f6,%f6 2209 ldd [%i3 + 224], %f20 2210 ldd [%i3 + 232], %f22 2211 .word 0x90ce0400 !aes_eround01 %f24,%f0,%f2,%f8 2212 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2213 .word 0x94ce0c04 !aes_eround01 %f24,%f4,%f6,%f10 2214 .word 0x8cce8c24 !aes_eround23 %f26,%f4,%f6,%f6 2215 .word 0x80cf0408 !aes_eround01 %f28,%f8,%f2,%f0 2216 .word 0x84cf8428 !aes_eround23 %f30,%f8,%f2,%f2 2217 .word 0x88cf0c0a !aes_eround01 %f28,%f10,%f6,%f4 2218 .word 0x8ccf8c2a !aes_eround23 %f30,%f10,%f6,%f6 2219 .word 0x90c84400 !aes_eround01 %f32,%f0,%f2,%f8 2220 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2221 .word 0x94c84c04 !aes_eround01 %f32,%f4,%f6,%f10 2222 .word 0x8cc8cc24 !aes_eround23 %f34,%f4,%f6,%f6 2223 .word 0x80c94408 !aes_eround01 %f36,%f8,%f2,%f0 2224 .word 0x84c9c428 !aes_eround23 %f38,%f8,%f2,%f2 2225 .word 0x88c94c0a !aes_eround01 %f36,%f10,%f6,%f4 2226 .word 0x8cc9cc2a !aes_eround23 %f38,%f10,%f6,%f6 2227 .word 0x90ca4400 !aes_eround01 %f40,%f0,%f2,%f8 2228 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2229 .word 0x94ca4c04 !aes_eround01 %f40,%f4,%f6,%f10 2230 .word 0x8ccacc24 !aes_eround23 %f42,%f4,%f6,%f6 2231 .word 0x80cb4408 !aes_eround01 %f44,%f8,%f2,%f0 2232 .word 0x84cbc428 !aes_eround23 %f46,%f8,%f2,%f2 2233 .word 0x88cb4c0a !aes_eround01 %f44,%f10,%f6,%f4 2234 .word 0x8ccbcc2a !aes_eround23 %f46,%f10,%f6,%f6 2235 .word 0x90cc4400 !aes_eround01 %f48,%f0,%f2,%f8 2236 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2237 .word 0x94cc4c04 !aes_eround01 %f48,%f4,%f6,%f10 2238 .word 0x8ccccc24 !aes_eround23 %f50,%f4,%f6,%f6 2239 .word 0x80cd4408 !aes_eround01 %f52,%f8,%f2,%f0 2240 .word 0x84cdc428 !aes_eround23 %f54,%f8,%f2,%f2 2241 .word 0x88cd4c0a !aes_eround01 %f52,%f10,%f6,%f4 2242 .word 0x8ccdcc2a !aes_eround23 %f54,%f10,%f6,%f6 2243 .word 0x90ce4400 !aes_eround01 %f56,%f0,%f2,%f8 2244 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2245 .word 0x94ce4c04 !aes_eround01 %f56,%f4,%f6,%f10 2246 .word 0x8ccecc24 !aes_eround23 %f58,%f4,%f6,%f6 2247 .word 0x80cf4408 !aes_eround01 %f60,%f8,%f2,%f0 2248 .word 0x84cfc428 !aes_eround23 %f62,%f8,%f2,%f2 2249 .word 0x88cf4c0a !aes_eround01 %f60,%f10,%f6,%f4 2250 .word 0x8ccfcc2a !aes_eround23 %f62,%f10,%f6,%f6 2251 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 2252 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2253 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 2254 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 2255 ldd [%i3 + 16], %f16 2256 ldd [%i3 + 24], %f18 2257 .word 0x80cd0488 !aes_eround01_l %f20,%f8,%f2,%f0 2258 .word 0x84cd84a8 !aes_eround23_l %f22,%f8,%f2,%f2 2259 .word 0x88cd0c8a !aes_eround01_l %f20,%f10,%f6,%f4 2260 .word 0x8ccd8caa !aes_eround23_l %f22,%f10,%f6,%f6 2261 ldd [%i3 + 32], %f20 2262 retl 2263 ldd [%i3 + 40], %f22 2264 .type _aes256_encrypt_2x,#function 2265 .size _aes256_encrypt_2x,.-_aes256_encrypt_2x 2266 2267 .align 32 2268 _aes192_loadkey: 2269 ldx [%i3 + 0], %g4 2270 ldx [%i3 + 8], %g5 2271 ldd [%i3 + 16], %f16 2272 ldd [%i3 + 24], %f18 2273 ldd [%i3 + 32], %f20 2274 ldd [%i3 + 40], %f22 2275 ldd [%i3 + 48], %f24 2276 ldd [%i3 + 56], %f26 2277 ldd [%i3 + 64], %f28 2278 ldd [%i3 + 72], %f30 2279 ldd [%i3 + 80], %f32 2280 ldd [%i3 + 88], %f34 2281 ldd [%i3 + 96], %f36 2282 ldd [%i3 + 104], %f38 2283 ldd [%i3 + 112], %f40 2284 ldd [%i3 + 120], %f42 2285 ldd [%i3 + 128], %f44 2286 ldd [%i3 + 136], %f46 2287 ldd [%i3 + 144], %f48 2288 ldd [%i3 + 152], %f50 2289 ldd [%i3 + 160], %f52 2290 ldd [%i3 + 168], %f54 2291 ldd [%i3 + 176], %f56 2292 ldd [%i3 + 184], %f58 2293 ldd [%i3 + 192], %f60 2294 ldd [%i3 + 200], %f62 2295 retl 2296 nop 2297 .type _aes192_loadkey,#function 2298 .size _aes192_loadkey,.-_aes192_loadkey 2299 _aes256_loadkey=_aes192_loadkey 2300 _aes192_load_enckey=_aes192_loadkey 2301 _aes192_load_deckey=_aes192_loadkey 2302 _aes256_load_enckey=_aes192_loadkey 2303 _aes256_load_deckey=_aes192_loadkey 2304 .globl aes256_t4_cbc_encrypt 2305 .align 32 2306 aes256_t4_cbc_encrypt: 2307 save %sp, -112, %sp 2308 cmp %i2, 0 2309 be,pn %icc, .L256_cbc_enc_abort 2310 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2311 sub %i0, %i1, %l5 ! %i0!=%i1 2312 ld [%i4 + 0], %f0 2313 ld [%i4 + 4], %f1 2314 ld [%i4 + 8], %f2 2315 ld [%i4 + 12], %f3 2316 prefetch [%i0], 20 2317 prefetch [%i0 + 63], 20 2318 call _aes256_load_enckey 2319 and %i0, 7, %l0 2320 andn %i0, 7, %i0 2321 sll %l0, 3, %l0 2322 mov 64, %l1 2323 mov 0xff, %l3 2324 sub %l1, %l0, %l1 2325 and %i1, 7, %l2 2326 cmp %i2, 127 2327 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2328 movleu %icc, 0, %l5 ! %i2<128 || 2329 brnz,pn %l5, .L256cbc_enc_blk ! %i0==%i1) 2330 srl %l3, %l2, %l3 2331 2332 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2333 srlx %i2, 4, %i2 2334 prefetch [%i1], 22 2335 2336 .L256_cbc_enc_loop: 2337 ldx [%i0 + 0], %o0 2338 brz,pt %l0, 4f 2339 ldx [%i0 + 8], %o1 2340 2341 ldx [%i0 + 16], %o2 2342 sllx %o0, %l0, %o0 2343 srlx %o1, %l1, %g1 2344 sllx %o1, %l0, %o1 2345 or %g1, %o0, %o0 2346 srlx %o2, %l1, %o2 2347 or %o2, %o1, %o1 2348 4: 2349 xor %g4, %o0, %o0 ! ^= rk[0] 2350 xor %g5, %o1, %o1 2351 .word 0x99b02308 !movxtod %o0,%f12 2352 .word 0x9db02309 !movxtod %o1,%f14 2353 2354 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2355 .word 0x85b38d82 !fxor %f14,%f2,%f2 2356 prefetch [%i1 + 63], 22 2357 prefetch [%i0 + 16+63], 20 2358 call _aes256_encrypt_1x 2359 add %i0, 16, %i0 2360 2361 brnz,pn %l2, 2f 2362 sub %i2, 1, %i2 2363 2364 std %f0, [%i1 + 0] 2365 std %f2, [%i1 + 8] 2366 brnz,pt %i2, .L256_cbc_enc_loop 2367 add %i1, 16, %i1 2368 st %f0, [%i4 + 0] 2369 st %f1, [%i4 + 4] 2370 st %f2, [%i4 + 8] 2371 st %f3, [%i4 + 12] 2372 .L256_cbc_enc_abort: 2373 ret 2374 restore 2375 2376 .align 16 2377 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2378 ! and ~3x deterioration 2379 ! in inp==out case 2380 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2381 .word 0x8db00902 !faligndata %f0,%f2,%f6 2382 .word 0x91b08902 !faligndata %f2,%f2,%f8 2383 2384 stda %f4, [%i1 + %l3]0xc0 ! partial store 2385 std %f6, [%i1 + 8] 2386 add %i1, 16, %i1 2387 orn %g0, %l3, %l3 2388 stda %f8, [%i1 + %l3]0xc0 ! partial store 2389 2390 brnz,pt %i2, .L256_cbc_enc_loop+4 2391 orn %g0, %l3, %l3 2392 st %f0, [%i4 + 0] 2393 st %f1, [%i4 + 4] 2394 st %f2, [%i4 + 8] 2395 st %f3, [%i4 + 12] 2396 ret 2397 restore 2398 2399 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2400 .align 32 2401 .L256cbc_enc_blk: 2402 add %i1, %i2, %l5 2403 and %l5, 63, %l5 ! tail 2404 sub %i2, %l5, %i2 2405 add %l5, 15, %l5 ! round up to 16n 2406 srlx %i2, 4, %i2 2407 srl %l5, 4, %l5 2408 2409 .L256_cbc_enc_blk_loop: 2410 ldx [%i0 + 0], %o0 2411 brz,pt %l0, 5f 2412 ldx [%i0 + 8], %o1 2413 2414 ldx [%i0 + 16], %o2 2415 sllx %o0, %l0, %o0 2416 srlx %o1, %l1, %g1 2417 sllx %o1, %l0, %o1 2418 or %g1, %o0, %o0 2419 srlx %o2, %l1, %o2 2420 or %o2, %o1, %o1 2421 5: 2422 xor %g4, %o0, %o0 ! ^= rk[0] 2423 xor %g5, %o1, %o1 2424 .word 0x99b02308 !movxtod %o0,%f12 2425 .word 0x9db02309 !movxtod %o1,%f14 2426 2427 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2428 .word 0x85b38d82 !fxor %f14,%f2,%f2 2429 prefetch [%i0 + 16+63], 20 2430 call _aes256_encrypt_1x 2431 add %i0, 16, %i0 2432 sub %i2, 1, %i2 2433 2434 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2435 add %i1, 8, %i1 2436 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2437 brnz,pt %i2, .L256_cbc_enc_blk_loop 2438 add %i1, 8, %i1 2439 2440 membar #StoreLoad|#StoreStore 2441 brnz,pt %l5, .L256_cbc_enc_loop 2442 mov %l5, %i2 2443 st %f0, [%i4 + 0] 2444 st %f1, [%i4 + 4] 2445 st %f2, [%i4 + 8] 2446 st %f3, [%i4 + 12] 2447 ret 2448 restore 2449 .type aes256_t4_cbc_encrypt,#function 2450 .size aes256_t4_cbc_encrypt,.-aes256_t4_cbc_encrypt 2451 .globl aes192_t4_cbc_encrypt 2452 .align 32 2453 aes192_t4_cbc_encrypt: 2454 save %sp, -112, %sp 2455 cmp %i2, 0 2456 be,pn %icc, .L192_cbc_enc_abort 2457 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2458 sub %i0, %i1, %l5 ! %i0!=%i1 2459 ld [%i4 + 0], %f0 2460 ld [%i4 + 4], %f1 2461 ld [%i4 + 8], %f2 2462 ld [%i4 + 12], %f3 2463 prefetch [%i0], 20 2464 prefetch [%i0 + 63], 20 2465 call _aes192_load_enckey 2466 and %i0, 7, %l0 2467 andn %i0, 7, %i0 2468 sll %l0, 3, %l0 2469 mov 64, %l1 2470 mov 0xff, %l3 2471 sub %l1, %l0, %l1 2472 and %i1, 7, %l2 2473 cmp %i2, 127 2474 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2475 movleu %icc, 0, %l5 ! %i2<128 || 2476 brnz,pn %l5, .L192cbc_enc_blk ! %i0==%i1) 2477 srl %l3, %l2, %l3 2478 2479 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2480 srlx %i2, 4, %i2 2481 prefetch [%i1], 22 2482 2483 .L192_cbc_enc_loop: 2484 ldx [%i0 + 0], %o0 2485 brz,pt %l0, 4f 2486 ldx [%i0 + 8], %o1 2487 2488 ldx [%i0 + 16], %o2 2489 sllx %o0, %l0, %o0 2490 srlx %o1, %l1, %g1 2491 sllx %o1, %l0, %o1 2492 or %g1, %o0, %o0 2493 srlx %o2, %l1, %o2 2494 or %o2, %o1, %o1 2495 4: 2496 xor %g4, %o0, %o0 ! ^= rk[0] 2497 xor %g5, %o1, %o1 2498 .word 0x99b02308 !movxtod %o0,%f12 2499 .word 0x9db02309 !movxtod %o1,%f14 2500 2501 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2502 .word 0x85b38d82 !fxor %f14,%f2,%f2 2503 prefetch [%i1 + 63], 22 2504 prefetch [%i0 + 16+63], 20 2505 call _aes192_encrypt_1x 2506 add %i0, 16, %i0 2507 2508 brnz,pn %l2, 2f 2509 sub %i2, 1, %i2 2510 2511 std %f0, [%i1 + 0] 2512 std %f2, [%i1 + 8] 2513 brnz,pt %i2, .L192_cbc_enc_loop 2514 add %i1, 16, %i1 2515 st %f0, [%i4 + 0] 2516 st %f1, [%i4 + 4] 2517 st %f2, [%i4 + 8] 2518 st %f3, [%i4 + 12] 2519 .L192_cbc_enc_abort: 2520 ret 2521 restore 2522 2523 .align 16 2524 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2525 ! and ~3x deterioration 2526 ! in inp==out case 2527 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2528 .word 0x8db00902 !faligndata %f0,%f2,%f6 2529 .word 0x91b08902 !faligndata %f2,%f2,%f8 2530 2531 stda %f4, [%i1 + %l3]0xc0 ! partial store 2532 std %f6, [%i1 + 8] 2533 add %i1, 16, %i1 2534 orn %g0, %l3, %l3 2535 stda %f8, [%i1 + %l3]0xc0 ! partial store 2536 2537 brnz,pt %i2, .L192_cbc_enc_loop+4 2538 orn %g0, %l3, %l3 2539 st %f0, [%i4 + 0] 2540 st %f1, [%i4 + 4] 2541 st %f2, [%i4 + 8] 2542 st %f3, [%i4 + 12] 2543 ret 2544 restore 2545 2546 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2547 .align 32 2548 .L192cbc_enc_blk: 2549 add %i1, %i2, %l5 2550 and %l5, 63, %l5 ! tail 2551 sub %i2, %l5, %i2 2552 add %l5, 15, %l5 ! round up to 16n 2553 srlx %i2, 4, %i2 2554 srl %l5, 4, %l5 2555 2556 .L192_cbc_enc_blk_loop: 2557 ldx [%i0 + 0], %o0 2558 brz,pt %l0, 5f 2559 ldx [%i0 + 8], %o1 2560 2561 ldx [%i0 + 16], %o2 2562 sllx %o0, %l0, %o0 2563 srlx %o1, %l1, %g1 2564 sllx %o1, %l0, %o1 2565 or %g1, %o0, %o0 2566 srlx %o2, %l1, %o2 2567 or %o2, %o1, %o1 2568 5: 2569 xor %g4, %o0, %o0 ! ^= rk[0] 2570 xor %g5, %o1, %o1 2571 .word 0x99b02308 !movxtod %o0,%f12 2572 .word 0x9db02309 !movxtod %o1,%f14 2573 2574 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2575 .word 0x85b38d82 !fxor %f14,%f2,%f2 2576 prefetch [%i0 + 16+63], 20 2577 call _aes192_encrypt_1x 2578 add %i0, 16, %i0 2579 sub %i2, 1, %i2 2580 2581 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2582 add %i1, 8, %i1 2583 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2584 brnz,pt %i2, .L192_cbc_enc_blk_loop 2585 add %i1, 8, %i1 2586 2587 membar #StoreLoad|#StoreStore 2588 brnz,pt %l5, .L192_cbc_enc_loop 2589 mov %l5, %i2 2590 st %f0, [%i4 + 0] 2591 st %f1, [%i4 + 4] 2592 st %f2, [%i4 + 8] 2593 st %f3, [%i4 + 12] 2594 ret 2595 restore 2596 .type aes192_t4_cbc_encrypt,#function 2597 .size aes192_t4_cbc_encrypt,.-aes192_t4_cbc_encrypt 2598 .globl aes256_t4_ctr32_encrypt 2599 .align 32 2600 aes256_t4_ctr32_encrypt: 2601 save %sp, -112, %sp 2602 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2603 2604 prefetch [%i0], 20 2605 prefetch [%i0 + 63], 20 2606 call _aes256_load_enckey 2607 sllx %i2, 4, %i2 2608 2609 ld [%i4 + 0], %l4 ! counter 2610 ld [%i4 + 4], %l5 2611 ld [%i4 + 8], %l6 2612 ld [%i4 + 12], %l7 2613 2614 sllx %l4, 32, %o5 2615 or %l5, %o5, %o5 2616 sllx %l6, 32, %g1 2617 xor %o5, %g4, %g4 ! ^= rk[0] 2618 xor %g1, %g5, %g5 2619 .word 0x9db02304 !movxtod %g4,%f14 ! most significant 64 bits 2620 2621 sub %i0, %i1, %l5 ! %i0!=%i1 2622 and %i0, 7, %l0 2623 andn %i0, 7, %i0 2624 sll %l0, 3, %l0 2625 mov 64, %l1 2626 mov 0xff, %l3 2627 sub %l1, %l0, %l1 2628 and %i1, 7, %l2 2629 cmp %i2, 255 2630 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2631 movleu %icc, 0, %l5 ! %i2<256 || 2632 brnz,pn %l5, .L256_ctr32_blk ! %i0==%i1) 2633 srl %l3, %l2, %l3 2634 2635 andcc %i2, 16, %g0 ! is number of blocks even? 2636 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2637 bz %icc, .L256_ctr32_loop2x 2638 srlx %i2, 4, %i2 2639 .L256_ctr32_loop: 2640 ldx [%i0 + 0], %o0 2641 brz,pt %l0, 4f 2642 ldx [%i0 + 8], %o1 2643 2644 ldx [%i0 + 16], %o2 2645 sllx %o0, %l0, %o0 2646 srlx %o1, %l1, %g1 2647 sllx %o1, %l0, %o1 2648 or %g1, %o0, %o0 2649 srlx %o2, %l1, %o2 2650 or %o2, %o1, %o1 2651 4: 2652 xor %g5, %l7, %g1 ! ^= rk[0] 2653 add %l7, 1, %l7 2654 .word 0x85b02301 !movxtod %g1,%f2 2655 srl %l7, 0, %l7 ! clruw 2656 prefetch [%i1 + 63], 22 2657 prefetch [%i0 + 16+63], 20 2658 .word 0x88cc040e !aes_eround01 %f16,%f14,%f2,%f4 2659 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 2660 call _aes256_encrypt_1x+8 2661 add %i0, 16, %i0 2662 2663 .word 0x95b02308 !movxtod %o0,%f10 2664 .word 0x99b02309 !movxtod %o1,%f12 2665 .word 0x81b28d80 !fxor %f10,%f0,%f0 ! ^= inp 2666 .word 0x85b30d82 !fxor %f12,%f2,%f2 2667 2668 brnz,pn %l2, 2f 2669 sub %i2, 1, %i2 2670 2671 std %f0, [%i1 + 0] 2672 std %f2, [%i1 + 8] 2673 brnz,pt %i2, .L256_ctr32_loop2x 2674 add %i1, 16, %i1 2675 2676 ret 2677 restore 2678 2679 .align 16 2680 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2681 ! and ~3x deterioration 2682 ! in inp==out case 2683 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2684 .word 0x8db00902 !faligndata %f0,%f2,%f6 2685 .word 0x91b08902 !faligndata %f2,%f2,%f8 2686 stda %f4, [%i1 + %l3]0xc0 ! partial store 2687 std %f6, [%i1 + 8] 2688 add %i1, 16, %i1 2689 orn %g0, %l3, %l3 2690 stda %f8, [%i1 + %l3]0xc0 ! partial store 2691 2692 brnz,pt %i2, .L256_ctr32_loop2x+4 2693 orn %g0, %l3, %l3 2694 2695 ret 2696 restore 2697 2698 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2699 .align 32 2700 .L256_ctr32_loop2x: 2701 ldx [%i0 + 0], %o0 2702 ldx [%i0 + 8], %o1 2703 ldx [%i0 + 16], %o2 2704 brz,pt %l0, 4f 2705 ldx [%i0 + 24], %o3 2706 2707 ldx [%i0 + 32], %o4 2708 sllx %o0, %l0, %o0 2709 srlx %o1, %l1, %g1 2710 or %g1, %o0, %o0 2711 sllx %o1, %l0, %o1 2712 srlx %o2, %l1, %g1 2713 or %g1, %o1, %o1 2714 sllx %o2, %l0, %o2 2715 srlx %o3, %l1, %g1 2716 or %g1, %o2, %o2 2717 sllx %o3, %l0, %o3 2718 srlx %o4, %l1, %o4 2719 or %o4, %o3, %o3 2720 4: 2721 xor %g5, %l7, %g1 ! ^= rk[0] 2722 add %l7, 1, %l7 2723 .word 0x85b02301 !movxtod %g1,%f2 2724 srl %l7, 0, %l7 ! clruw 2725 xor %g5, %l7, %g1 2726 add %l7, 1, %l7 2727 .word 0x8db02301 !movxtod %g1,%f6 2728 srl %l7, 0, %l7 ! clruw 2729 prefetch [%i1 + 63], 22 2730 prefetch [%i0 + 32+63], 20 2731 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 2732 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 2733 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 2734 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 2735 call _aes256_encrypt_2x+16 2736 add %i0, 32, %i0 2737 2738 .word 0x91b02308 !movxtod %o0,%f8 2739 .word 0x95b02309 !movxtod %o1,%f10 2740 .word 0x99b0230a !movxtod %o2,%f12 2741 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 2742 .word 0x91b0230b !movxtod %o3,%f8 2743 .word 0x85b28d82 !fxor %f10,%f2,%f2 2744 .word 0x89b30d84 !fxor %f12,%f4,%f4 2745 .word 0x8db20d86 !fxor %f8,%f6,%f6 2746 2747 brnz,pn %l2, 2f 2748 sub %i2, 2, %i2 2749 2750 std %f0, [%i1 + 0] 2751 std %f2, [%i1 + 8] 2752 std %f4, [%i1 + 16] 2753 std %f6, [%i1 + 24] 2754 brnz,pt %i2, .L256_ctr32_loop2x 2755 add %i1, 32, %i1 2756 2757 ret 2758 restore 2759 2760 .align 16 2761 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2762 ! and ~3x deterioration 2763 ! in inp==out case 2764 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 2765 .word 0x81b00902 !faligndata %f0,%f2,%f0 2766 .word 0x85b08904 !faligndata %f2,%f4,%f2 2767 .word 0x89b10906 !faligndata %f4,%f6,%f4 2768 .word 0x8db18906 !faligndata %f6,%f6,%f6 2769 2770 stda %f8, [%i1 + %l3]0xc0 ! partial store 2771 std %f0, [%i1 + 8] 2772 std %f2, [%i1 + 16] 2773 std %f4, [%i1 + 24] 2774 add %i1, 32, %i1 2775 orn %g0, %l3, %l3 2776 stda %f6, [%i1 + %l3]0xc0 ! partial store 2777 2778 brnz,pt %i2, .L256_ctr32_loop2x+4 2779 orn %g0, %l3, %l3 2780 2781 ret 2782 restore 2783 2784 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2785 .align 32 2786 .L256_ctr32_blk: 2787 add %i1, %i2, %l5 2788 and %l5, 63, %l5 ! tail 2789 sub %i2, %l5, %i2 2790 add %l5, 15, %l5 ! round up to 16n 2791 srlx %i2, 4, %i2 2792 srl %l5, 4, %l5 2793 sub %i2, 1, %i2 2794 add %l5, 1, %l5 2795 2796 .L256_ctr32_blk_loop2x: 2797 ldx [%i0 + 0], %o0 2798 ldx [%i0 + 8], %o1 2799 ldx [%i0 + 16], %o2 2800 brz,pt %l0, 5f 2801 ldx [%i0 + 24], %o3 2802 2803 ldx [%i0 + 32], %o4 2804 sllx %o0, %l0, %o0 2805 srlx %o1, %l1, %g1 2806 or %g1, %o0, %o0 2807 sllx %o1, %l0, %o1 2808 srlx %o2, %l1, %g1 2809 or %g1, %o1, %o1 2810 sllx %o2, %l0, %o2 2811 srlx %o3, %l1, %g1 2812 or %g1, %o2, %o2 2813 sllx %o3, %l0, %o3 2814 srlx %o4, %l1, %o4 2815 or %o4, %o3, %o3 2816 5: 2817 xor %g5, %l7, %g1 ! ^= rk[0] 2818 add %l7, 1, %l7 2819 .word 0x85b02301 !movxtod %g1,%f2 2820 srl %l7, 0, %l7 ! clruw 2821 xor %g5, %l7, %g1 2822 add %l7, 1, %l7 2823 .word 0x8db02301 !movxtod %g1,%f6 2824 srl %l7, 0, %l7 ! clruw 2825 prefetch [%i0 + 32+63], 20 2826 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 2827 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 2828 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 2829 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 2830 call _aes256_encrypt_2x+16 2831 add %i0, 32, %i0 2832 subcc %i2, 2, %i2 2833 2834 .word 0x91b02308 !movxtod %o0,%f8 2835 .word 0x95b02309 !movxtod %o1,%f10 2836 .word 0x99b0230a !movxtod %o2,%f12 2837 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 2838 .word 0x91b0230b !movxtod %o3,%f8 2839 .word 0x85b28d82 !fxor %f10,%f2,%f2 2840 .word 0x89b30d84 !fxor %f12,%f4,%f4 2841 .word 0x8db20d86 !fxor %f8,%f6,%f6 2842 2843 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2844 add %i1, 8, %i1 2845 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2846 add %i1, 8, %i1 2847 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2848 add %i1, 8, %i1 2849 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2850 bgu,pt %icc, .L256_ctr32_blk_loop2x 2851 add %i1, 8, %i1 2852 2853 add %l5, %i2, %i2 2854 andcc %i2, 1, %g0 ! is number of blocks even? 2855 membar #StoreLoad|#StoreStore 2856 bnz,pt %icc, .L256_ctr32_loop 2857 srl %i2, 0, %i2 2858 brnz,pn %i2, .L256_ctr32_loop2x 2859 nop 2860 2861 ret 2862 restore 2863 .type aes256_t4_ctr32_encrypt,#function 2864 .size aes256_t4_ctr32_encrypt,.-aes256_t4_ctr32_encrypt 2865 .globl aes256_t4_xts_encrypt 2866 .align 32 2867 aes256_t4_xts_encrypt: 2868 save %sp, -112-16, %sp 2869 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2870 2871 mov %i5, %o0 2872 add %fp, 0-16, %o1 2873 call aes_t4_encrypt 2874 mov %i4, %o2 2875 2876 add %fp, 0-16, %l7 2877 ldxa [%l7]0x88, %g2 2878 add %fp, 0-8, %l7 2879 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 2880 2881 sethi %hi(0x76543210), %l7 2882 or %l7, %lo(0x76543210), %l7 2883 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 2884 2885 prefetch [%i0], 20 2886 prefetch [%i0 + 63], 20 2887 call _aes256_load_enckey 2888 and %i2, 15, %i5 2889 and %i2, -16, %i2 2890 2891 sub %i0, %i1, %l5 ! %i0!=%i1 2892 and %i0, 7, %l0 2893 andn %i0, 7, %i0 2894 sll %l0, 3, %l0 2895 mov 64, %l1 2896 mov 0xff, %l3 2897 sub %l1, %l0, %l1 2898 and %i1, 7, %l2 2899 cmp %i2, 255 2900 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2901 movleu %icc, 0, %l5 ! %i2<256 || 2902 brnz,pn %l5, .L256_xts_enblk ! %i0==%i1) 2903 srl %l3, %l2, %l3 2904 2905 andcc %i2, 16, %g0 ! is number of blocks even? 2906 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2907 bz %icc, .L256_xts_enloop2x 2908 srlx %i2, 4, %i2 2909 .L256_xts_enloop: 2910 ldx [%i0 + 0], %o0 2911 brz,pt %l0, 4f 2912 ldx [%i0 + 8], %o1 2913 2914 ldx [%i0 + 16], %o2 2915 sllx %o0, %l0, %o0 2916 srlx %o1, %l1, %g1 2917 sllx %o1, %l0, %o1 2918 or %g1, %o0, %o0 2919 srlx %o2, %l1, %o2 2920 or %o2, %o1, %o1 2921 4: 2922 .word 0x99b02302 !movxtod %g2,%f12 2923 .word 0x9db02303 !movxtod %g3,%f14 2924 .word 0x99b3098c !bshuffle %f12,%f12,%f12 2925 .word 0x9db3898e !bshuffle %f14,%f14,%f14 2926 2927 xor %g4, %o0, %o0 ! ^= rk[0] 2928 xor %g5, %o1, %o1 2929 .word 0x81b02308 !movxtod %o0,%f0 2930 .word 0x85b02309 !movxtod %o1,%f2 2931 2932 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 2933 .word 0x85b38d82 !fxor %f14,%f2,%f2 2934 2935 prefetch [%i1 + 63], 22 2936 prefetch [%i0 + 16+63], 20 2937 call _aes256_encrypt_1x 2938 add %i0, 16, %i0 2939 2940 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 2941 .word 0x85b38d82 !fxor %f14,%f2,%f2 2942 2943 srax %g3, 63, %l7 ! next tweak value 2944 addcc %g2, %g2, %g2 2945 and %l7, 0x87, %l7 2946 .word 0x87b0c223 !addxc %g3,%g3,%g3 2947 xor %l7, %g2, %g2 2948 2949 brnz,pn %l2, 2f 2950 sub %i2, 1, %i2 2951 2952 std %f0, [%i1 + 0] 2953 std %f2, [%i1 + 8] 2954 brnz,pt %i2, .L256_xts_enloop2x 2955 add %i1, 16, %i1 2956 2957 brnz,pn %i5, .L256_xts_ensteal 2958 nop 2959 2960 ret 2961 restore 2962 2963 .align 16 2964 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2965 ! and ~3x deterioration 2966 ! in inp==out case 2967 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2968 .word 0x8db00902 !faligndata %f0,%f2,%f6 2969 .word 0x91b08902 !faligndata %f2,%f2,%f8 2970 stda %f4, [%i1 + %l3]0xc0 ! partial store 2971 std %f6, [%i1 + 8] 2972 add %i1, 16, %i1 2973 orn %g0, %l3, %l3 2974 stda %f8, [%i1 + %l3]0xc0 ! partial store 2975 2976 brnz,pt %i2, .L256_xts_enloop2x+4 2977 orn %g0, %l3, %l3 2978 2979 brnz,pn %i5, .L256_xts_ensteal 2980 nop 2981 2982 ret 2983 restore 2984 2985 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2986 .align 32 2987 .L256_xts_enloop2x: 2988 ldx [%i0 + 0], %o0 2989 ldx [%i0 + 8], %o1 2990 ldx [%i0 + 16], %o2 2991 brz,pt %l0, 4f 2992 ldx [%i0 + 24], %o3 2993 2994 ldx [%i0 + 32], %o4 2995 sllx %o0, %l0, %o0 2996 srlx %o1, %l1, %g1 2997 or %g1, %o0, %o0 2998 sllx %o1, %l0, %o1 2999 srlx %o2, %l1, %g1 3000 or %g1, %o1, %o1 3001 sllx %o2, %l0, %o2 3002 srlx %o3, %l1, %g1 3003 or %g1, %o2, %o2 3004 sllx %o3, %l0, %o3 3005 srlx %o4, %l1, %o4 3006 or %o4, %o3, %o3 3007 4: 3008 .word 0x99b02302 !movxtod %g2,%f12 3009 .word 0x9db02303 !movxtod %g3,%f14 3010 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3011 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3012 3013 srax %g3, 63, %l7 ! next tweak value 3014 addcc %g2, %g2, %g2 3015 and %l7, 0x87, %l7 3016 .word 0x87b0c223 !addxc %g3,%g3,%g3 3017 xor %l7, %g2, %g2 3018 3019 .word 0x91b02302 !movxtod %g2,%f8 3020 .word 0x95b02303 !movxtod %g3,%f10 3021 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3022 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3023 3024 xor %g4, %o0, %o0 ! ^= rk[0] 3025 xor %g5, %o1, %o1 3026 xor %g4, %o2, %o2 ! ^= rk[0] 3027 xor %g5, %o3, %o3 3028 .word 0x81b02308 !movxtod %o0,%f0 3029 .word 0x85b02309 !movxtod %o1,%f2 3030 .word 0x89b0230a !movxtod %o2,%f4 3031 .word 0x8db0230b !movxtod %o3,%f6 3032 3033 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3034 .word 0x85b38d82 !fxor %f14,%f2,%f2 3035 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3036 .word 0x8db28d86 !fxor %f10,%f6,%f6 3037 3038 prefetch [%i1 + 63], 22 3039 prefetch [%i0 + 32+63], 20 3040 call _aes256_encrypt_2x 3041 add %i0, 32, %i0 3042 3043 .word 0x91b02302 !movxtod %g2,%f8 3044 .word 0x95b02303 !movxtod %g3,%f10 3045 3046 srax %g3, 63, %l7 ! next tweak value 3047 addcc %g2, %g2, %g2 3048 and %l7, 0x87, %l7 3049 .word 0x87b0c223 !addxc %g3,%g3,%g3 3050 xor %l7, %g2, %g2 3051 3052 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3053 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3054 3055 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3056 .word 0x85b38d82 !fxor %f14,%f2,%f2 3057 .word 0x89b20d84 !fxor %f8,%f4,%f4 3058 .word 0x8db28d86 !fxor %f10,%f6,%f6 3059 3060 brnz,pn %l2, 2f 3061 sub %i2, 2, %i2 3062 3063 std %f0, [%i1 + 0] 3064 std %f2, [%i1 + 8] 3065 std %f4, [%i1 + 16] 3066 std %f6, [%i1 + 24] 3067 brnz,pt %i2, .L256_xts_enloop2x 3068 add %i1, 32, %i1 3069 3070 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3071 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3072 brnz,pn %i5, .L256_xts_ensteal 3073 nop 3074 3075 ret 3076 restore 3077 3078 .align 16 3079 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3080 ! and ~3x deterioration 3081 ! in inp==out case 3082 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 3083 .word 0x95b00902 !faligndata %f0,%f2,%f10 3084 .word 0x99b08904 !faligndata %f2,%f4,%f12 3085 .word 0x9db10906 !faligndata %f4,%f6,%f14 3086 .word 0x81b18906 !faligndata %f6,%f6,%f0 3087 3088 stda %f8, [%i1 + %l3]0xc0 ! partial store 3089 std %f10, [%i1 + 8] 3090 std %f12, [%i1 + 16] 3091 std %f14, [%i1 + 24] 3092 add %i1, 32, %i1 3093 orn %g0, %l3, %l3 3094 stda %f0, [%i1 + %l3]0xc0 ! partial store 3095 3096 brnz,pt %i2, .L256_xts_enloop2x+4 3097 orn %g0, %l3, %l3 3098 3099 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3100 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3101 brnz,pn %i5, .L256_xts_ensteal 3102 nop 3103 3104 ret 3105 restore 3106 3107 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3108 .align 32 3109 .L256_xts_enblk: 3110 add %i1, %i2, %l5 3111 and %l5, 63, %l5 ! tail 3112 sub %i2, %l5, %i2 3113 add %l5, 15, %l5 ! round up to 16n 3114 srlx %i2, 4, %i2 3115 srl %l5, 4, %l5 3116 sub %i2, 1, %i2 3117 add %l5, 1, %l5 3118 3119 .L256_xts_enblk2x: 3120 ldx [%i0 + 0], %o0 3121 ldx [%i0 + 8], %o1 3122 ldx [%i0 + 16], %o2 3123 brz,pt %l0, 5f 3124 ldx [%i0 + 24], %o3 3125 3126 ldx [%i0 + 32], %o4 3127 sllx %o0, %l0, %o0 3128 srlx %o1, %l1, %g1 3129 or %g1, %o0, %o0 3130 sllx %o1, %l0, %o1 3131 srlx %o2, %l1, %g1 3132 or %g1, %o1, %o1 3133 sllx %o2, %l0, %o2 3134 srlx %o3, %l1, %g1 3135 or %g1, %o2, %o2 3136 sllx %o3, %l0, %o3 3137 srlx %o4, %l1, %o4 3138 or %o4, %o3, %o3 3139 5: 3140 .word 0x99b02302 !movxtod %g2,%f12 3141 .word 0x9db02303 !movxtod %g3,%f14 3142 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3143 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3144 3145 srax %g3, 63, %l7 ! next tweak value 3146 addcc %g2, %g2, %g2 3147 and %l7, 0x87, %l7 3148 .word 0x87b0c223 !addxc %g3,%g3,%g3 3149 xor %l7, %g2, %g2 3150 3151 .word 0x91b02302 !movxtod %g2,%f8 3152 .word 0x95b02303 !movxtod %g3,%f10 3153 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3154 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3155 3156 xor %g4, %o0, %o0 ! ^= rk[0] 3157 xor %g5, %o1, %o1 3158 xor %g4, %o2, %o2 ! ^= rk[0] 3159 xor %g5, %o3, %o3 3160 .word 0x81b02308 !movxtod %o0,%f0 3161 .word 0x85b02309 !movxtod %o1,%f2 3162 .word 0x89b0230a !movxtod %o2,%f4 3163 .word 0x8db0230b !movxtod %o3,%f6 3164 3165 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3166 .word 0x85b38d82 !fxor %f14,%f2,%f2 3167 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3168 .word 0x8db28d86 !fxor %f10,%f6,%f6 3169 3170 prefetch [%i0 + 32+63], 20 3171 call _aes256_encrypt_2x 3172 add %i0, 32, %i0 3173 3174 .word 0x91b02302 !movxtod %g2,%f8 3175 .word 0x95b02303 !movxtod %g3,%f10 3176 3177 srax %g3, 63, %l7 ! next tweak value 3178 addcc %g2, %g2, %g2 3179 and %l7, 0x87, %l7 3180 .word 0x87b0c223 !addxc %g3,%g3,%g3 3181 xor %l7, %g2, %g2 3182 3183 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3184 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3185 3186 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3187 .word 0x85b38d82 !fxor %f14,%f2,%f2 3188 .word 0x89b20d84 !fxor %f8,%f4,%f4 3189 .word 0x8db28d86 !fxor %f10,%f6,%f6 3190 3191 subcc %i2, 2, %i2 3192 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3193 add %i1, 8, %i1 3194 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3195 add %i1, 8, %i1 3196 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3197 add %i1, 8, %i1 3198 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3199 bgu,pt %icc, .L256_xts_enblk2x 3200 add %i1, 8, %i1 3201 3202 add %l5, %i2, %i2 3203 andcc %i2, 1, %g0 ! is number of blocks even? 3204 membar #StoreLoad|#StoreStore 3205 bnz,pt %icc, .L256_xts_enloop 3206 srl %i2, 0, %i2 3207 brnz,pn %i2, .L256_xts_enloop2x 3208 nop 3209 3210 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3211 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3212 brnz,pn %i5, .L256_xts_ensteal 3213 nop 3214 3215 ret 3216 restore 3217 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3218 .align 32 3219 .L256_xts_ensteal: 3220 std %f0, [%fp + 0-16] ! copy of output 3221 std %f2, [%fp + 0-8] 3222 3223 srl %l0, 3, %l0 3224 add %fp, 0-16, %l7 3225 add %i0, %l0, %i0 ! original %i0+%i2&-15 3226 add %i1, %l2, %i1 ! original %i1+%i2&-15 3227 mov 0, %l0 3228 nop ! align 3229 3230 .L256_xts_enstealing: 3231 ldub [%i0 + %l0], %o0 3232 ldub [%l7 + %l0], %o1 3233 dec %i5 3234 stb %o0, [%l7 + %l0] 3235 stb %o1, [%i1 + %l0] 3236 brnz %i5, .L256_xts_enstealing 3237 inc %l0 3238 3239 mov %l7, %i0 3240 sub %i1, 16, %i1 3241 mov 0, %l0 3242 sub %i1, %l2, %i1 3243 ba .L256_xts_enloop ! one more time 3244 mov 1, %i2 ! %i5 is 0 3245 ret 3246 restore 3247 .type aes256_t4_xts_encrypt,#function 3248 .size aes256_t4_xts_encrypt,.-aes256_t4_xts_encrypt 3249 .globl aes256_t4_xts_decrypt 3250 .align 32 3251 aes256_t4_xts_decrypt: 3252 save %sp, -112-16, %sp 3253 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 3254 3255 mov %i5, %o0 3256 add %fp, 0-16, %o1 3257 call aes_t4_encrypt 3258 mov %i4, %o2 3259 3260 add %fp, 0-16, %l7 3261 ldxa [%l7]0x88, %g2 3262 add %fp, 0-8, %l7 3263 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 3264 3265 sethi %hi(0x76543210), %l7 3266 or %l7, %lo(0x76543210), %l7 3267 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 3268 3269 prefetch [%i0], 20 3270 prefetch [%i0 + 63], 20 3271 call _aes256_load_deckey 3272 and %i2, 15, %i5 3273 and %i2, -16, %i2 3274 mov 0, %l7 3275 movrnz %i5, 16, %l7 3276 sub %i2, %l7, %i2 3277 3278 sub %i0, %i1, %l5 ! %i0!=%i1 3279 and %i0, 7, %l0 3280 andn %i0, 7, %i0 3281 sll %l0, 3, %l0 3282 mov 64, %l1 3283 mov 0xff, %l3 3284 sub %l1, %l0, %l1 3285 and %i1, 7, %l2 3286 cmp %i2, 255 3287 movrnz %l2, 0, %l5 ! if ( %i1&7 || 3288 movleu %icc, 0, %l5 ! %i2<256 || 3289 brnz,pn %l5, .L256_xts_deblk ! %i0==%i1) 3290 srl %l3, %l2, %l3 3291 3292 andcc %i2, 16, %g0 ! is number of blocks even? 3293 brz,pn %i2, .L256_xts_desteal 3294 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 3295 bz %icc, .L256_xts_deloop2x 3296 srlx %i2, 4, %i2 3297 .L256_xts_deloop: 3298 ldx [%i0 + 0], %o0 3299 brz,pt %l0, 4f 3300 ldx [%i0 + 8], %o1 3301 3302 ldx [%i0 + 16], %o2 3303 sllx %o0, %l0, %o0 3304 srlx %o1, %l1, %g1 3305 sllx %o1, %l0, %o1 3306 or %g1, %o0, %o0 3307 srlx %o2, %l1, %o2 3308 or %o2, %o1, %o1 3309 4: 3310 .word 0x99b02302 !movxtod %g2,%f12 3311 .word 0x9db02303 !movxtod %g3,%f14 3312 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3313 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3314 3315 xor %g4, %o0, %o0 ! ^= rk[0] 3316 xor %g5, %o1, %o1 3317 .word 0x81b02308 !movxtod %o0,%f0 3318 .word 0x85b02309 !movxtod %o1,%f2 3319 3320 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3321 .word 0x85b38d82 !fxor %f14,%f2,%f2 3322 3323 prefetch [%i1 + 63], 22 3324 prefetch [%i0 + 16+63], 20 3325 call _aes256_decrypt_1x 3326 add %i0, 16, %i0 3327 3328 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3329 .word 0x85b38d82 !fxor %f14,%f2,%f2 3330 3331 srax %g3, 63, %l7 ! next tweak value 3332 addcc %g2, %g2, %g2 3333 and %l7, 0x87, %l7 3334 .word 0x87b0c223 !addxc %g3,%g3,%g3 3335 xor %l7, %g2, %g2 3336 3337 brnz,pn %l2, 2f 3338 sub %i2, 1, %i2 3339 3340 std %f0, [%i1 + 0] 3341 std %f2, [%i1 + 8] 3342 brnz,pt %i2, .L256_xts_deloop2x 3343 add %i1, 16, %i1 3344 3345 brnz,pn %i5, .L256_xts_desteal 3346 nop 3347 3348 ret 3349 restore 3350 3351 .align 16 3352 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3353 ! and ~3x deterioration 3354 ! in inp==out case 3355 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 3356 .word 0x8db00902 !faligndata %f0,%f2,%f6 3357 .word 0x91b08902 !faligndata %f2,%f2,%f8 3358 stda %f4, [%i1 + %l3]0xc0 ! partial store 3359 std %f6, [%i1 + 8] 3360 add %i1, 16, %i1 3361 orn %g0, %l3, %l3 3362 stda %f8, [%i1 + %l3]0xc0 ! partial store 3363 3364 brnz,pt %i2, .L256_xts_deloop2x+4 3365 orn %g0, %l3, %l3 3366 3367 brnz,pn %i5, .L256_xts_desteal 3368 nop 3369 3370 ret 3371 restore 3372 3373 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3374 .align 32 3375 .L256_xts_deloop2x: 3376 ldx [%i0 + 0], %o0 3377 ldx [%i0 + 8], %o1 3378 ldx [%i0 + 16], %o2 3379 brz,pt %l0, 4f 3380 ldx [%i0 + 24], %o3 3381 3382 ldx [%i0 + 32], %o4 3383 sllx %o0, %l0, %o0 3384 srlx %o1, %l1, %g1 3385 or %g1, %o0, %o0 3386 sllx %o1, %l0, %o1 3387 srlx %o2, %l1, %g1 3388 or %g1, %o1, %o1 3389 sllx %o2, %l0, %o2 3390 srlx %o3, %l1, %g1 3391 or %g1, %o2, %o2 3392 sllx %o3, %l0, %o3 3393 srlx %o4, %l1, %o4 3394 or %o4, %o3, %o3 3395 4: 3396 .word 0x99b02302 !movxtod %g2,%f12 3397 .word 0x9db02303 !movxtod %g3,%f14 3398 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3399 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3400 3401 srax %g3, 63, %l7 ! next tweak value 3402 addcc %g2, %g2, %g2 3403 and %l7, 0x87, %l7 3404 .word 0x87b0c223 !addxc %g3,%g3,%g3 3405 xor %l7, %g2, %g2 3406 3407 .word 0x91b02302 !movxtod %g2,%f8 3408 .word 0x95b02303 !movxtod %g3,%f10 3409 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3410 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3411 3412 xor %g4, %o0, %o0 ! ^= rk[0] 3413 xor %g5, %o1, %o1 3414 xor %g4, %o2, %o2 ! ^= rk[0] 3415 xor %g5, %o3, %o3 3416 .word 0x81b02308 !movxtod %o0,%f0 3417 .word 0x85b02309 !movxtod %o1,%f2 3418 .word 0x89b0230a !movxtod %o2,%f4 3419 .word 0x8db0230b !movxtod %o3,%f6 3420 3421 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3422 .word 0x85b38d82 !fxor %f14,%f2,%f2 3423 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3424 .word 0x8db28d86 !fxor %f10,%f6,%f6 3425 3426 prefetch [%i1 + 63], 22 3427 prefetch [%i0 + 32+63], 20 3428 call _aes256_decrypt_2x 3429 add %i0, 32, %i0 3430 3431 .word 0x91b02302 !movxtod %g2,%f8 3432 .word 0x95b02303 !movxtod %g3,%f10 3433 3434 srax %g3, 63, %l7 ! next tweak value 3435 addcc %g2, %g2, %g2 3436 and %l7, 0x87, %l7 3437 .word 0x87b0c223 !addxc %g3,%g3,%g3 3438 xor %l7, %g2, %g2 3439 3440 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3441 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3442 3443 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3444 .word 0x85b38d82 !fxor %f14,%f2,%f2 3445 .word 0x89b20d84 !fxor %f8,%f4,%f4 3446 .word 0x8db28d86 !fxor %f10,%f6,%f6 3447 3448 brnz,pn %l2, 2f 3449 sub %i2, 2, %i2 3450 3451 std %f0, [%i1 + 0] 3452 std %f2, [%i1 + 8] 3453 std %f4, [%i1 + 16] 3454 std %f6, [%i1 + 24] 3455 brnz,pt %i2, .L256_xts_deloop2x 3456 add %i1, 32, %i1 3457 3458 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3459 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3460 brnz,pn %i5, .L256_xts_desteal 3461 nop 3462 3463 ret 3464 restore 3465 3466 .align 16 3467 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3468 ! and ~3x deterioration 3469 ! in inp==out case 3470 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 3471 .word 0x95b00902 !faligndata %f0,%f2,%f10 3472 .word 0x99b08904 !faligndata %f2,%f4,%f12 3473 .word 0x9db10906 !faligndata %f4,%f6,%f14 3474 .word 0x81b18906 !faligndata %f6,%f6,%f0 3475 3476 stda %f8, [%i1 + %l3]0xc0 ! partial store 3477 std %f10, [%i1 + 8] 3478 std %f12, [%i1 + 16] 3479 std %f14, [%i1 + 24] 3480 add %i1, 32, %i1 3481 orn %g0, %l3, %l3 3482 stda %f0, [%i1 + %l3]0xc0 ! partial store 3483 3484 brnz,pt %i2, .L256_xts_deloop2x+4 3485 orn %g0, %l3, %l3 3486 3487 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3488 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3489 brnz,pn %i5, .L256_xts_desteal 3490 nop 3491 3492 ret 3493 restore 3494 3495 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3496 .align 32 3497 .L256_xts_deblk: 3498 add %i1, %i2, %l5 3499 and %l5, 63, %l5 ! tail 3500 sub %i2, %l5, %i2 3501 add %l5, 15, %l5 ! round up to 16n 3502 srlx %i2, 4, %i2 3503 srl %l5, 4, %l5 3504 sub %i2, 1, %i2 3505 add %l5, 1, %l5 3506 3507 .L256_xts_deblk2x: 3508 ldx [%i0 + 0], %o0 3509 ldx [%i0 + 8], %o1 3510 ldx [%i0 + 16], %o2 3511 brz,pt %l0, 5f 3512 ldx [%i0 + 24], %o3 3513 3514 ldx [%i0 + 32], %o4 3515 sllx %o0, %l0, %o0 3516 srlx %o1, %l1, %g1 3517 or %g1, %o0, %o0 3518 sllx %o1, %l0, %o1 3519 srlx %o2, %l1, %g1 3520 or %g1, %o1, %o1 3521 sllx %o2, %l0, %o2 3522 srlx %o3, %l1, %g1 3523 or %g1, %o2, %o2 3524 sllx %o3, %l0, %o3 3525 srlx %o4, %l1, %o4 3526 or %o4, %o3, %o3 3527 5: 3528 .word 0x99b02302 !movxtod %g2,%f12 3529 .word 0x9db02303 !movxtod %g3,%f14 3530 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3531 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3532 3533 srax %g3, 63, %l7 ! next tweak value 3534 addcc %g2, %g2, %g2 3535 and %l7, 0x87, %l7 3536 .word 0x87b0c223 !addxc %g3,%g3,%g3 3537 xor %l7, %g2, %g2 3538 3539 .word 0x91b02302 !movxtod %g2,%f8 3540 .word 0x95b02303 !movxtod %g3,%f10 3541 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3542 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3543 3544 xor %g4, %o0, %o0 ! ^= rk[0] 3545 xor %g5, %o1, %o1 3546 xor %g4, %o2, %o2 ! ^= rk[0] 3547 xor %g5, %o3, %o3 3548 .word 0x81b02308 !movxtod %o0,%f0 3549 .word 0x85b02309 !movxtod %o1,%f2 3550 .word 0x89b0230a !movxtod %o2,%f4 3551 .word 0x8db0230b !movxtod %o3,%f6 3552 3553 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3554 .word 0x85b38d82 !fxor %f14,%f2,%f2 3555 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3556 .word 0x8db28d86 !fxor %f10,%f6,%f6 3557 3558 prefetch [%i0 + 32+63], 20 3559 call _aes256_decrypt_2x 3560 add %i0, 32, %i0 3561 3562 .word 0x91b02302 !movxtod %g2,%f8 3563 .word 0x95b02303 !movxtod %g3,%f10 3564 3565 srax %g3, 63, %l7 ! next tweak value 3566 addcc %g2, %g2, %g2 3567 and %l7, 0x87, %l7 3568 .word 0x87b0c223 !addxc %g3,%g3,%g3 3569 xor %l7, %g2, %g2 3570 3571 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3572 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3573 3574 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3575 .word 0x85b38d82 !fxor %f14,%f2,%f2 3576 .word 0x89b20d84 !fxor %f8,%f4,%f4 3577 .word 0x8db28d86 !fxor %f10,%f6,%f6 3578 3579 subcc %i2, 2, %i2 3580 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3581 add %i1, 8, %i1 3582 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3583 add %i1, 8, %i1 3584 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3585 add %i1, 8, %i1 3586 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3587 bgu,pt %icc, .L256_xts_deblk2x 3588 add %i1, 8, %i1 3589 3590 add %l5, %i2, %i2 3591 andcc %i2, 1, %g0 ! is number of blocks even? 3592 membar #StoreLoad|#StoreStore 3593 bnz,pt %icc, .L256_xts_deloop 3594 srl %i2, 0, %i2 3595 brnz,pn %i2, .L256_xts_deloop2x 3596 nop 3597 3598 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3599 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3600 brnz,pn %i5, .L256_xts_desteal 3601 nop 3602 3603 ret 3604 restore 3605 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3606 .align 32 3607 .L256_xts_desteal: 3608 ldx [%i0 + 0], %o0 3609 brz,pt %l0, 8f 3610 ldx [%i0 + 8], %o1 3611 3612 ldx [%i0 + 16], %o2 3613 sllx %o0, %l0, %o0 3614 srlx %o1, %l1, %g1 3615 sllx %o1, %l0, %o1 3616 or %g1, %o0, %o0 3617 srlx %o2, %l1, %o2 3618 or %o2, %o1, %o1 3619 8: 3620 srax %g3, 63, %l7 ! next tweak value 3621 addcc %g2, %g2, %o2 3622 and %l7, 0x87, %l7 3623 .word 0x97b0c223 !addxc %g3,%g3,%o3 3624 xor %l7, %o2, %o2 3625 3626 .word 0x99b0230a !movxtod %o2,%f12 3627 .word 0x9db0230b !movxtod %o3,%f14 3628 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3629 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3630 3631 xor %g4, %o0, %o0 ! ^= rk[0] 3632 xor %g5, %o1, %o1 3633 .word 0x81b02308 !movxtod %o0,%f0 3634 .word 0x85b02309 !movxtod %o1,%f2 3635 3636 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3637 .word 0x85b38d82 !fxor %f14,%f2,%f2 3638 3639 call _aes256_decrypt_1x 3640 add %i0, 16, %i0 3641 3642 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3643 .word 0x85b38d82 !fxor %f14,%f2,%f2 3644 3645 std %f0, [%fp + 0-16] 3646 std %f2, [%fp + 0-8] 3647 3648 srl %l0, 3, %l0 3649 add %fp, 0-16, %l7 3650 add %i0, %l0, %i0 ! original %i0+%i2&-15 3651 add %i1, %l2, %i1 ! original %i1+%i2&-15 3652 mov 0, %l0 3653 add %i1, 16, %i1 3654 nop ! align 3655 3656 .L256_xts_destealing: 3657 ldub [%i0 + %l0], %o0 3658 ldub [%l7 + %l0], %o1 3659 dec %i5 3660 stb %o0, [%l7 + %l0] 3661 stb %o1, [%i1 + %l0] 3662 brnz %i5, .L256_xts_destealing 3663 inc %l0 3664 3665 mov %l7, %i0 3666 sub %i1, 16, %i1 3667 mov 0, %l0 3668 sub %i1, %l2, %i1 3669 ba .L256_xts_deloop ! one more time 3670 mov 1, %i2 ! %i5 is 0 3671 ret 3672 restore 3673 .type aes256_t4_xts_decrypt,#function 3674 .size aes256_t4_xts_decrypt,.-aes256_t4_xts_decrypt 3675 .globl aes192_t4_ctr32_encrypt 3676 .align 32 3677 aes192_t4_ctr32_encrypt: 3678 save %sp, -112, %sp 3679 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 3680 3681 prefetch [%i0], 20 3682 prefetch [%i0 + 63], 20 3683 call _aes192_load_enckey 3684 sllx %i2, 4, %i2 3685 3686 ld [%i4 + 0], %l4 ! counter 3687 ld [%i4 + 4], %l5 3688 ld [%i4 + 8], %l6 3689 ld [%i4 + 12], %l7 3690 3691 sllx %l4, 32, %o5 3692 or %l5, %o5, %o5 3693 sllx %l6, 32, %g1 3694 xor %o5, %g4, %g4 ! ^= rk[0] 3695 xor %g1, %g5, %g5 3696 .word 0x9db02304 !movxtod %g4,%f14 ! most significant 64 bits 3697 3698 sub %i0, %i1, %l5 ! %i0!=%i1 3699 and %i0, 7, %l0 3700 andn %i0, 7, %i0 3701 sll %l0, 3, %l0 3702 mov 64, %l1 3703 mov 0xff, %l3 3704 sub %l1, %l0, %l1 3705 and %i1, 7, %l2 3706 cmp %i2, 255 3707 movrnz %l2, 0, %l5 ! if ( %i1&7 || 3708 movleu %icc, 0, %l5 ! %i2<256 || 3709 brnz,pn %l5, .L192_ctr32_blk ! %i0==%i1) 3710 srl %l3, %l2, %l3 3711 3712 andcc %i2, 16, %g0 ! is number of blocks even? 3713 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 3714 bz %icc, .L192_ctr32_loop2x 3715 srlx %i2, 4, %i2 3716 .L192_ctr32_loop: 3717 ldx [%i0 + 0], %o0 3718 brz,pt %l0, 4f 3719 ldx [%i0 + 8], %o1 3720 3721 ldx [%i0 + 16], %o2 3722 sllx %o0, %l0, %o0 3723 srlx %o1, %l1, %g1 3724 sllx %o1, %l0, %o1 3725 or %g1, %o0, %o0 3726 srlx %o2, %l1, %o2 3727 or %o2, %o1, %o1 3728 4: 3729 xor %g5, %l7, %g1 ! ^= rk[0] 3730 add %l7, 1, %l7 3731 .word 0x85b02301 !movxtod %g1,%f2 3732 srl %l7, 0, %l7 ! clruw 3733 prefetch [%i1 + 63], 22 3734 prefetch [%i0 + 16+63], 20 3735 .word 0x88cc040e !aes_eround01 %f16,%f14,%f2,%f4 3736 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 3737 call _aes192_encrypt_1x+8 3738 add %i0, 16, %i0 3739 3740 .word 0x95b02308 !movxtod %o0,%f10 3741 .word 0x99b02309 !movxtod %o1,%f12 3742 .word 0x81b28d80 !fxor %f10,%f0,%f0 ! ^= inp 3743 .word 0x85b30d82 !fxor %f12,%f2,%f2 3744 3745 brnz,pn %l2, 2f 3746 sub %i2, 1, %i2 3747 3748 std %f0, [%i1 + 0] 3749 std %f2, [%i1 + 8] 3750 brnz,pt %i2, .L192_ctr32_loop2x 3751 add %i1, 16, %i1 3752 3753 ret 3754 restore 3755 3756 .align 16 3757 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3758 ! and ~3x deterioration 3759 ! in inp==out case 3760 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 3761 .word 0x8db00902 !faligndata %f0,%f2,%f6 3762 .word 0x91b08902 !faligndata %f2,%f2,%f8 3763 stda %f4, [%i1 + %l3]0xc0 ! partial store 3764 std %f6, [%i1 + 8] 3765 add %i1, 16, %i1 3766 orn %g0, %l3, %l3 3767 stda %f8, [%i1 + %l3]0xc0 ! partial store 3768 3769 brnz,pt %i2, .L192_ctr32_loop2x+4 3770 orn %g0, %l3, %l3 3771 3772 ret 3773 restore 3774 3775 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3776 .align 32 3777 .L192_ctr32_loop2x: 3778 ldx [%i0 + 0], %o0 3779 ldx [%i0 + 8], %o1 3780 ldx [%i0 + 16], %o2 3781 brz,pt %l0, 4f 3782 ldx [%i0 + 24], %o3 3783 3784 ldx [%i0 + 32], %o4 3785 sllx %o0, %l0, %o0 3786 srlx %o1, %l1, %g1 3787 or %g1, %o0, %o0 3788 sllx %o1, %l0, %o1 3789 srlx %o2, %l1, %g1 3790 or %g1, %o1, %o1 3791 sllx %o2, %l0, %o2 3792 srlx %o3, %l1, %g1 3793 or %g1, %o2, %o2 3794 sllx %o3, %l0, %o3 3795 srlx %o4, %l1, %o4 3796 or %o4, %o3, %o3 3797 4: 3798 xor %g5, %l7, %g1 ! ^= rk[0] 3799 add %l7, 1, %l7 3800 .word 0x85b02301 !movxtod %g1,%f2 3801 srl %l7, 0, %l7 ! clruw 3802 xor %g5, %l7, %g1 3803 add %l7, 1, %l7 3804 .word 0x8db02301 !movxtod %g1,%f6 3805 srl %l7, 0, %l7 ! clruw 3806 prefetch [%i1 + 63], 22 3807 prefetch [%i0 + 32+63], 20 3808 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 3809 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 3810 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 3811 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 3812 call _aes192_encrypt_2x+16 3813 add %i0, 32, %i0 3814 3815 .word 0x91b02308 !movxtod %o0,%f8 3816 .word 0x95b02309 !movxtod %o1,%f10 3817 .word 0x99b0230a !movxtod %o2,%f12 3818 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 3819 .word 0x91b0230b !movxtod %o3,%f8 3820 .word 0x85b28d82 !fxor %f10,%f2,%f2 3821 .word 0x89b30d84 !fxor %f12,%f4,%f4 3822 .word 0x8db20d86 !fxor %f8,%f6,%f6 3823 3824 brnz,pn %l2, 2f 3825 sub %i2, 2, %i2 3826 3827 std %f0, [%i1 + 0] 3828 std %f2, [%i1 + 8] 3829 std %f4, [%i1 + 16] 3830 std %f6, [%i1 + 24] 3831 brnz,pt %i2, .L192_ctr32_loop2x 3832 add %i1, 32, %i1 3833 3834 ret 3835 restore 3836 3837 .align 16 3838 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3839 ! and ~3x deterioration 3840 ! in inp==out case 3841 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 3842 .word 0x81b00902 !faligndata %f0,%f2,%f0 3843 .word 0x85b08904 !faligndata %f2,%f4,%f2 3844 .word 0x89b10906 !faligndata %f4,%f6,%f4 3845 .word 0x8db18906 !faligndata %f6,%f6,%f6 3846 3847 stda %f8, [%i1 + %l3]0xc0 ! partial store 3848 std %f0, [%i1 + 8] 3849 std %f2, [%i1 + 16] 3850 std %f4, [%i1 + 24] 3851 add %i1, 32, %i1 3852 orn %g0, %l3, %l3 3853 stda %f6, [%i1 + %l3]0xc0 ! partial store 3854 3855 brnz,pt %i2, .L192_ctr32_loop2x+4 3856 orn %g0, %l3, %l3 3857 3858 ret 3859 restore 3860 3861 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3862 .align 32 3863 .L192_ctr32_blk: 3864 add %i1, %i2, %l5 3865 and %l5, 63, %l5 ! tail 3866 sub %i2, %l5, %i2 3867 add %l5, 15, %l5 ! round up to 16n 3868 srlx %i2, 4, %i2 3869 srl %l5, 4, %l5 3870 sub %i2, 1, %i2 3871 add %l5, 1, %l5 3872 3873 .L192_ctr32_blk_loop2x: 3874 ldx [%i0 + 0], %o0 3875 ldx [%i0 + 8], %o1 3876 ldx [%i0 + 16], %o2 3877 brz,pt %l0, 5f 3878 ldx [%i0 + 24], %o3 3879 3880 ldx [%i0 + 32], %o4 3881 sllx %o0, %l0, %o0 3882 srlx %o1, %l1, %g1 3883 or %g1, %o0, %o0 3884 sllx %o1, %l0, %o1 3885 srlx %o2, %l1, %g1 3886 or %g1, %o1, %o1 3887 sllx %o2, %l0, %o2 3888 srlx %o3, %l1, %g1 3889 or %g1, %o2, %o2 3890 sllx %o3, %l0, %o3 3891 srlx %o4, %l1, %o4 3892 or %o4, %o3, %o3 3893 5: 3894 xor %g5, %l7, %g1 ! ^= rk[0] 3895 add %l7, 1, %l7 3896 .word 0x85b02301 !movxtod %g1,%f2 3897 srl %l7, 0, %l7 ! clruw 3898 xor %g5, %l7, %g1 3899 add %l7, 1, %l7 3900 .word 0x8db02301 !movxtod %g1,%f6 3901 srl %l7, 0, %l7 ! clruw 3902 prefetch [%i0 + 32+63], 20 3903 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 3904 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 3905 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 3906 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 3907 call _aes192_encrypt_2x+16 3908 add %i0, 32, %i0 3909 subcc %i2, 2, %i2 3910 3911 .word 0x91b02308 !movxtod %o0,%f8 3912 .word 0x95b02309 !movxtod %o1,%f10 3913 .word 0x99b0230a !movxtod %o2,%f12 3914 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 3915 .word 0x91b0230b !movxtod %o3,%f8 3916 .word 0x85b28d82 !fxor %f10,%f2,%f2 3917 .word 0x89b30d84 !fxor %f12,%f4,%f4 3918 .word 0x8db20d86 !fxor %f8,%f6,%f6 3919 3920 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3921 add %i1, 8, %i1 3922 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3923 add %i1, 8, %i1 3924 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3925 add %i1, 8, %i1 3926 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3927 bgu,pt %icc, .L192_ctr32_blk_loop2x 3928 add %i1, 8, %i1 3929 3930 add %l5, %i2, %i2 3931 andcc %i2, 1, %g0 ! is number of blocks even? 3932 membar #StoreLoad|#StoreStore 3933 bnz,pt %icc, .L192_ctr32_loop 3934 srl %i2, 0, %i2 3935 brnz,pn %i2, .L192_ctr32_loop2x 3936 nop 3937 3938 ret 3939 restore 3940 .type aes192_t4_ctr32_encrypt,#function 3941 .size aes192_t4_ctr32_encrypt,.-aes192_t4_ctr32_encrypt 3942 .globl aes192_t4_cbc_decrypt 3943 .align 32 3944 aes192_t4_cbc_decrypt: 3945 save %sp, -112, %sp 3946 cmp %i2, 0 3947 be,pn %icc, .L192_cbc_dec_abort 3948 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 3949 sub %i0, %i1, %l5 ! %i0!=%i1 3950 ld [%i4 + 0], %f12 ! load ivec 3951 ld [%i4 + 4], %f13 3952 ld [%i4 + 8], %f14 3953 ld [%i4 + 12], %f15 3954 prefetch [%i0], 20 3955 prefetch [%i0 + 63], 20 3956 call _aes192_load_deckey 3957 and %i0, 7, %l0 3958 andn %i0, 7, %i0 3959 sll %l0, 3, %l0 3960 mov 64, %l1 3961 mov 0xff, %l3 3962 sub %l1, %l0, %l1 3963 and %i1, 7, %l2 3964 cmp %i2, 255 3965 movrnz %l2, 0, %l5 ! if ( %i1&7 || 3966 movleu %icc, 0, %l5 ! %i2<256 || 3967 brnz,pn %l5, .L192cbc_dec_blk ! %i0==%i1) 3968 srl %l3, %l2, %l3 3969 3970 andcc %i2, 16, %g0 ! is number of blocks even? 3971 srlx %i2, 4, %i2 3972 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 3973 bz %icc, .L192_cbc_dec_loop2x 3974 prefetch [%i1], 22 3975 .L192_cbc_dec_loop: 3976 ldx [%i0 + 0], %o0 3977 brz,pt %l0, 4f 3978 ldx [%i0 + 8], %o1 3979 3980 ldx [%i0 + 16], %o2 3981 sllx %o0, %l0, %o0 3982 srlx %o1, %l1, %g1 3983 sllx %o1, %l0, %o1 3984 or %g1, %o0, %o0 3985 srlx %o2, %l1, %o2 3986 or %o2, %o1, %o1 3987 4: 3988 xor %g4, %o0, %o2 ! ^= rk[0] 3989 xor %g5, %o1, %o3 3990 .word 0x81b0230a !movxtod %o2,%f0 3991 .word 0x85b0230b !movxtod %o3,%f2 3992 3993 prefetch [%i1 + 63], 22 3994 prefetch [%i0 + 16+63], 20 3995 call _aes192_decrypt_1x 3996 add %i0, 16, %i0 3997 3998 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 3999 .word 0x85b38d82 !fxor %f14,%f2,%f2 4000 .word 0x99b02308 !movxtod %o0,%f12 4001 .word 0x9db02309 !movxtod %o1,%f14 4002 4003 brnz,pn %l2, 2f 4004 sub %i2, 1, %i2 4005 4006 std %f0, [%i1 + 0] 4007 std %f2, [%i1 + 8] 4008 brnz,pt %i2, .L192_cbc_dec_loop2x 4009 add %i1, 16, %i1 4010 st %f12, [%i4 + 0] 4011 st %f13, [%i4 + 4] 4012 st %f14, [%i4 + 8] 4013 st %f15, [%i4 + 12] 4014 .L192_cbc_dec_abort: 4015 ret 4016 restore 4017 4018 .align 16 4019 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4020 ! and ~3x deterioration 4021 ! in inp==out case 4022 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 4023 .word 0x8db00902 !faligndata %f0,%f2,%f6 4024 .word 0x91b08902 !faligndata %f2,%f2,%f8 4025 4026 stda %f4, [%i1 + %l3]0xc0 ! partial store 4027 std %f6, [%i1 + 8] 4028 add %i1, 16, %i1 4029 orn %g0, %l3, %l3 4030 stda %f8, [%i1 + %l3]0xc0 ! partial store 4031 4032 brnz,pt %i2, .L192_cbc_dec_loop2x+4 4033 orn %g0, %l3, %l3 4034 st %f12, [%i4 + 0] 4035 st %f13, [%i4 + 4] 4036 st %f14, [%i4 + 8] 4037 st %f15, [%i4 + 12] 4038 ret 4039 restore 4040 4041 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4042 .align 32 4043 .L192_cbc_dec_loop2x: 4044 ldx [%i0 + 0], %o0 4045 ldx [%i0 + 8], %o1 4046 ldx [%i0 + 16], %o2 4047 brz,pt %l0, 4f 4048 ldx [%i0 + 24], %o3 4049 4050 ldx [%i0 + 32], %o4 4051 sllx %o0, %l0, %o0 4052 srlx %o1, %l1, %g1 4053 or %g1, %o0, %o0 4054 sllx %o1, %l0, %o1 4055 srlx %o2, %l1, %g1 4056 or %g1, %o1, %o1 4057 sllx %o2, %l0, %o2 4058 srlx %o3, %l1, %g1 4059 or %g1, %o2, %o2 4060 sllx %o3, %l0, %o3 4061 srlx %o4, %l1, %o4 4062 or %o4, %o3, %o3 4063 4: 4064 xor %g4, %o0, %o4 ! ^= rk[0] 4065 xor %g5, %o1, %o5 4066 .word 0x81b0230c !movxtod %o4,%f0 4067 .word 0x85b0230d !movxtod %o5,%f2 4068 xor %g4, %o2, %o4 4069 xor %g5, %o3, %o5 4070 .word 0x89b0230c !movxtod %o4,%f4 4071 .word 0x8db0230d !movxtod %o5,%f6 4072 4073 prefetch [%i1 + 63], 22 4074 prefetch [%i0 + 32+63], 20 4075 call _aes192_decrypt_2x 4076 add %i0, 32, %i0 4077 4078 .word 0x91b02308 !movxtod %o0,%f8 4079 .word 0x95b02309 !movxtod %o1,%f10 4080 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4081 .word 0x85b38d82 !fxor %f14,%f2,%f2 4082 .word 0x99b0230a !movxtod %o2,%f12 4083 .word 0x9db0230b !movxtod %o3,%f14 4084 .word 0x89b20d84 !fxor %f8,%f4,%f4 4085 .word 0x8db28d86 !fxor %f10,%f6,%f6 4086 4087 brnz,pn %l2, 2f 4088 sub %i2, 2, %i2 4089 4090 std %f0, [%i1 + 0] 4091 std %f2, [%i1 + 8] 4092 std %f4, [%i1 + 16] 4093 std %f6, [%i1 + 24] 4094 brnz,pt %i2, .L192_cbc_dec_loop2x 4095 add %i1, 32, %i1 4096 st %f12, [%i4 + 0] 4097 st %f13, [%i4 + 4] 4098 st %f14, [%i4 + 8] 4099 st %f15, [%i4 + 12] 4100 ret 4101 restore 4102 4103 .align 16 4104 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4105 ! and ~3x deterioration 4106 ! in inp==out case 4107 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 4108 .word 0x81b00902 !faligndata %f0,%f2,%f0 4109 .word 0x85b08904 !faligndata %f2,%f4,%f2 4110 .word 0x89b10906 !faligndata %f4,%f6,%f4 4111 .word 0x8db18906 !faligndata %f6,%f6,%f6 4112 stda %f8, [%i1 + %l3]0xc0 ! partial store 4113 std %f0, [%i1 + 8] 4114 std %f2, [%i1 + 16] 4115 std %f4, [%i1 + 24] 4116 add %i1, 32, %i1 4117 orn %g0, %l3, %l3 4118 stda %f6, [%i1 + %l3]0xc0 ! partial store 4119 4120 brnz,pt %i2, .L192_cbc_dec_loop2x+4 4121 orn %g0, %l3, %l3 4122 st %f12, [%i4 + 0] 4123 st %f13, [%i4 + 4] 4124 st %f14, [%i4 + 8] 4125 st %f15, [%i4 + 12] 4126 ret 4127 restore 4128 4129 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4130 .align 32 4131 .L192cbc_dec_blk: 4132 add %i1, %i2, %l5 4133 and %l5, 63, %l5 ! tail 4134 sub %i2, %l5, %i2 4135 add %l5, 15, %l5 ! round up to 16n 4136 srlx %i2, 4, %i2 4137 srl %l5, 4, %l5 4138 sub %i2, 1, %i2 4139 add %l5, 1, %l5 4140 4141 .L192_cbc_dec_blk_loop2x: 4142 ldx [%i0 + 0], %o0 4143 ldx [%i0 + 8], %o1 4144 ldx [%i0 + 16], %o2 4145 brz,pt %l0, 5f 4146 ldx [%i0 + 24], %o3 4147 4148 ldx [%i0 + 32], %o4 4149 sllx %o0, %l0, %o0 4150 srlx %o1, %l1, %g1 4151 or %g1, %o0, %o0 4152 sllx %o1, %l0, %o1 4153 srlx %o2, %l1, %g1 4154 or %g1, %o1, %o1 4155 sllx %o2, %l0, %o2 4156 srlx %o3, %l1, %g1 4157 or %g1, %o2, %o2 4158 sllx %o3, %l0, %o3 4159 srlx %o4, %l1, %o4 4160 or %o4, %o3, %o3 4161 5: 4162 xor %g4, %o0, %o4 ! ^= rk[0] 4163 xor %g5, %o1, %o5 4164 .word 0x81b0230c !movxtod %o4,%f0 4165 .word 0x85b0230d !movxtod %o5,%f2 4166 xor %g4, %o2, %o4 4167 xor %g5, %o3, %o5 4168 .word 0x89b0230c !movxtod %o4,%f4 4169 .word 0x8db0230d !movxtod %o5,%f6 4170 4171 prefetch [%i0 + 32+63], 20 4172 call _aes192_decrypt_2x 4173 add %i0, 32, %i0 4174 subcc %i2, 2, %i2 4175 4176 .word 0x91b02308 !movxtod %o0,%f8 4177 .word 0x95b02309 !movxtod %o1,%f10 4178 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4179 .word 0x85b38d82 !fxor %f14,%f2,%f2 4180 .word 0x99b0230a !movxtod %o2,%f12 4181 .word 0x9db0230b !movxtod %o3,%f14 4182 .word 0x89b20d84 !fxor %f8,%f4,%f4 4183 .word 0x8db28d86 !fxor %f10,%f6,%f6 4184 4185 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4186 add %i1, 8, %i1 4187 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4188 add %i1, 8, %i1 4189 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4190 add %i1, 8, %i1 4191 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4192 bgu,pt %icc, .L192_cbc_dec_blk_loop2x 4193 add %i1, 8, %i1 4194 4195 add %l5, %i2, %i2 4196 andcc %i2, 1, %g0 ! is number of blocks even? 4197 membar #StoreLoad|#StoreStore 4198 bnz,pt %icc, .L192_cbc_dec_loop 4199 srl %i2, 0, %i2 4200 brnz,pn %i2, .L192_cbc_dec_loop2x 4201 nop 4202 st %f12, [%i4 + 0] ! write out ivec 4203 st %f13, [%i4 + 4] 4204 st %f14, [%i4 + 8] 4205 st %f15, [%i4 + 12] 4206 ret 4207 restore 4208 .type aes192_t4_cbc_decrypt,#function 4209 .size aes192_t4_cbc_decrypt,.-aes192_t4_cbc_decrypt 4210 .globl aes256_t4_cbc_decrypt 4211 .align 32 4212 aes256_t4_cbc_decrypt: 4213 save %sp, -112, %sp 4214 cmp %i2, 0 4215 be,pn %icc, .L256_cbc_dec_abort 4216 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 4217 sub %i0, %i1, %l5 ! %i0!=%i1 4218 ld [%i4 + 0], %f12 ! load ivec 4219 ld [%i4 + 4], %f13 4220 ld [%i4 + 8], %f14 4221 ld [%i4 + 12], %f15 4222 prefetch [%i0], 20 4223 prefetch [%i0 + 63], 20 4224 call _aes256_load_deckey 4225 and %i0, 7, %l0 4226 andn %i0, 7, %i0 4227 sll %l0, 3, %l0 4228 mov 64, %l1 4229 mov 0xff, %l3 4230 sub %l1, %l0, %l1 4231 and %i1, 7, %l2 4232 cmp %i2, 255 4233 movrnz %l2, 0, %l5 ! if ( %i1&7 || 4234 movleu %icc, 0, %l5 ! %i2<256 || 4235 brnz,pn %l5, .L256cbc_dec_blk ! %i0==%i1) 4236 srl %l3, %l2, %l3 4237 4238 andcc %i2, 16, %g0 ! is number of blocks even? 4239 srlx %i2, 4, %i2 4240 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 4241 bz %icc, .L256_cbc_dec_loop2x 4242 prefetch [%i1], 22 4243 .L256_cbc_dec_loop: 4244 ldx [%i0 + 0], %o0 4245 brz,pt %l0, 4f 4246 ldx [%i0 + 8], %o1 4247 4248 ldx [%i0 + 16], %o2 4249 sllx %o0, %l0, %o0 4250 srlx %o1, %l1, %g1 4251 sllx %o1, %l0, %o1 4252 or %g1, %o0, %o0 4253 srlx %o2, %l1, %o2 4254 or %o2, %o1, %o1 4255 4: 4256 xor %g4, %o0, %o2 ! ^= rk[0] 4257 xor %g5, %o1, %o3 4258 .word 0x81b0230a !movxtod %o2,%f0 4259 .word 0x85b0230b !movxtod %o3,%f2 4260 4261 prefetch [%i1 + 63], 22 4262 prefetch [%i0 + 16+63], 20 4263 call _aes256_decrypt_1x 4264 add %i0, 16, %i0 4265 4266 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4267 .word 0x85b38d82 !fxor %f14,%f2,%f2 4268 .word 0x99b02308 !movxtod %o0,%f12 4269 .word 0x9db02309 !movxtod %o1,%f14 4270 4271 brnz,pn %l2, 2f 4272 sub %i2, 1, %i2 4273 4274 std %f0, [%i1 + 0] 4275 std %f2, [%i1 + 8] 4276 brnz,pt %i2, .L256_cbc_dec_loop2x 4277 add %i1, 16, %i1 4278 st %f12, [%i4 + 0] 4279 st %f13, [%i4 + 4] 4280 st %f14, [%i4 + 8] 4281 st %f15, [%i4 + 12] 4282 .L256_cbc_dec_abort: 4283 ret 4284 restore 4285 4286 .align 16 4287 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4288 ! and ~3x deterioration 4289 ! in inp==out case 4290 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 4291 .word 0x8db00902 !faligndata %f0,%f2,%f6 4292 .word 0x91b08902 !faligndata %f2,%f2,%f8 4293 4294 stda %f4, [%i1 + %l3]0xc0 ! partial store 4295 std %f6, [%i1 + 8] 4296 add %i1, 16, %i1 4297 orn %g0, %l3, %l3 4298 stda %f8, [%i1 + %l3]0xc0 ! partial store 4299 4300 brnz,pt %i2, .L256_cbc_dec_loop2x+4 4301 orn %g0, %l3, %l3 4302 st %f12, [%i4 + 0] 4303 st %f13, [%i4 + 4] 4304 st %f14, [%i4 + 8] 4305 st %f15, [%i4 + 12] 4306 ret 4307 restore 4308 4309 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4310 .align 32 4311 .L256_cbc_dec_loop2x: 4312 ldx [%i0 + 0], %o0 4313 ldx [%i0 + 8], %o1 4314 ldx [%i0 + 16], %o2 4315 brz,pt %l0, 4f 4316 ldx [%i0 + 24], %o3 4317 4318 ldx [%i0 + 32], %o4 4319 sllx %o0, %l0, %o0 4320 srlx %o1, %l1, %g1 4321 or %g1, %o0, %o0 4322 sllx %o1, %l0, %o1 4323 srlx %o2, %l1, %g1 4324 or %g1, %o1, %o1 4325 sllx %o2, %l0, %o2 4326 srlx %o3, %l1, %g1 4327 or %g1, %o2, %o2 4328 sllx %o3, %l0, %o3 4329 srlx %o4, %l1, %o4 4330 or %o4, %o3, %o3 4331 4: 4332 xor %g4, %o0, %o4 ! ^= rk[0] 4333 xor %g5, %o1, %o5 4334 .word 0x81b0230c !movxtod %o4,%f0 4335 .word 0x85b0230d !movxtod %o5,%f2 4336 xor %g4, %o2, %o4 4337 xor %g5, %o3, %o5 4338 .word 0x89b0230c !movxtod %o4,%f4 4339 .word 0x8db0230d !movxtod %o5,%f6 4340 4341 prefetch [%i1 + 63], 22 4342 prefetch [%i0 + 32+63], 20 4343 call _aes256_decrypt_2x 4344 add %i0, 32, %i0 4345 4346 .word 0x91b02308 !movxtod %o0,%f8 4347 .word 0x95b02309 !movxtod %o1,%f10 4348 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4349 .word 0x85b38d82 !fxor %f14,%f2,%f2 4350 .word 0x99b0230a !movxtod %o2,%f12 4351 .word 0x9db0230b !movxtod %o3,%f14 4352 .word 0x89b20d84 !fxor %f8,%f4,%f4 4353 .word 0x8db28d86 !fxor %f10,%f6,%f6 4354 4355 brnz,pn %l2, 2f 4356 sub %i2, 2, %i2 4357 4358 std %f0, [%i1 + 0] 4359 std %f2, [%i1 + 8] 4360 std %f4, [%i1 + 16] 4361 std %f6, [%i1 + 24] 4362 brnz,pt %i2, .L256_cbc_dec_loop2x 4363 add %i1, 32, %i1 4364 st %f12, [%i4 + 0] 4365 st %f13, [%i4 + 4] 4366 st %f14, [%i4 + 8] 4367 st %f15, [%i4 + 12] 4368 ret 4369 restore 4370 4371 .align 16 4372 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4373 ! and ~3x deterioration 4374 ! in inp==out case 4375 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 4376 .word 0x81b00902 !faligndata %f0,%f2,%f0 4377 .word 0x85b08904 !faligndata %f2,%f4,%f2 4378 .word 0x89b10906 !faligndata %f4,%f6,%f4 4379 .word 0x8db18906 !faligndata %f6,%f6,%f6 4380 stda %f8, [%i1 + %l3]0xc0 ! partial store 4381 std %f0, [%i1 + 8] 4382 std %f2, [%i1 + 16] 4383 std %f4, [%i1 + 24] 4384 add %i1, 32, %i1 4385 orn %g0, %l3, %l3 4386 stda %f6, [%i1 + %l3]0xc0 ! partial store 4387 4388 brnz,pt %i2, .L256_cbc_dec_loop2x+4 4389 orn %g0, %l3, %l3 4390 st %f12, [%i4 + 0] 4391 st %f13, [%i4 + 4] 4392 st %f14, [%i4 + 8] 4393 st %f15, [%i4 + 12] 4394 ret 4395 restore 4396 4397 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4398 .align 32 4399 .L256cbc_dec_blk: 4400 add %i1, %i2, %l5 4401 and %l5, 63, %l5 ! tail 4402 sub %i2, %l5, %i2 4403 add %l5, 15, %l5 ! round up to 16n 4404 srlx %i2, 4, %i2 4405 srl %l5, 4, %l5 4406 sub %i2, 1, %i2 4407 add %l5, 1, %l5 4408 4409 .L256_cbc_dec_blk_loop2x: 4410 ldx [%i0 + 0], %o0 4411 ldx [%i0 + 8], %o1 4412 ldx [%i0 + 16], %o2 4413 brz,pt %l0, 5f 4414 ldx [%i0 + 24], %o3 4415 4416 ldx [%i0 + 32], %o4 4417 sllx %o0, %l0, %o0 4418 srlx %o1, %l1, %g1 4419 or %g1, %o0, %o0 4420 sllx %o1, %l0, %o1 4421 srlx %o2, %l1, %g1 4422 or %g1, %o1, %o1 4423 sllx %o2, %l0, %o2 4424 srlx %o3, %l1, %g1 4425 or %g1, %o2, %o2 4426 sllx %o3, %l0, %o3 4427 srlx %o4, %l1, %o4 4428 or %o4, %o3, %o3 4429 5: 4430 xor %g4, %o0, %o4 ! ^= rk[0] 4431 xor %g5, %o1, %o5 4432 .word 0x81b0230c !movxtod %o4,%f0 4433 .word 0x85b0230d !movxtod %o5,%f2 4434 xor %g4, %o2, %o4 4435 xor %g5, %o3, %o5 4436 .word 0x89b0230c !movxtod %o4,%f4 4437 .word 0x8db0230d !movxtod %o5,%f6 4438 4439 prefetch [%i0 + 32+63], 20 4440 call _aes256_decrypt_2x 4441 add %i0, 32, %i0 4442 subcc %i2, 2, %i2 4443 4444 .word 0x91b02308 !movxtod %o0,%f8 4445 .word 0x95b02309 !movxtod %o1,%f10 4446 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4447 .word 0x85b38d82 !fxor %f14,%f2,%f2 4448 .word 0x99b0230a !movxtod %o2,%f12 4449 .word 0x9db0230b !movxtod %o3,%f14 4450 .word 0x89b20d84 !fxor %f8,%f4,%f4 4451 .word 0x8db28d86 !fxor %f10,%f6,%f6 4452 4453 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4454 add %i1, 8, %i1 4455 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4456 add %i1, 8, %i1 4457 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4458 add %i1, 8, %i1 4459 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4460 bgu,pt %icc, .L256_cbc_dec_blk_loop2x 4461 add %i1, 8, %i1 4462 4463 add %l5, %i2, %i2 4464 andcc %i2, 1, %g0 ! is number of blocks even? 4465 membar #StoreLoad|#StoreStore 4466 bnz,pt %icc, .L256_cbc_dec_loop 4467 srl %i2, 0, %i2 4468 brnz,pn %i2, .L256_cbc_dec_loop2x 4469 nop 4470 st %f12, [%i4 + 0] ! write out ivec 4471 st %f13, [%i4 + 4] 4472 st %f14, [%i4 + 8] 4473 st %f15, [%i4 + 12] 4474 ret 4475 restore 4476 .type aes256_t4_cbc_decrypt,#function 4477 .size aes256_t4_cbc_decrypt,.-aes256_t4_cbc_decrypt 4478 .align 32 4479 _aes256_decrypt_1x: 4480 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 4481 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4482 ldd [%i3 + 208], %f16 4483 ldd [%i3 + 216], %f18 4484 .word 0x80cd0444 !aes_dround01 %f20,%f4,%f2,%f0 4485 .word 0x84cd8464 !aes_dround23 %f22,%f4,%f2,%f2 4486 ldd [%i3 + 224], %f20 4487 ldd [%i3 + 232], %f22 4488 .word 0x88ce0440 !aes_dround01 %f24,%f0,%f2,%f4 4489 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4490 .word 0x80cf0444 !aes_dround01 %f28,%f4,%f2,%f0 4491 .word 0x84cf8464 !aes_dround23 %f30,%f4,%f2,%f2 4492 .word 0x88c84440 !aes_dround01 %f32,%f0,%f2,%f4 4493 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4494 .word 0x80c94444 !aes_dround01 %f36,%f4,%f2,%f0 4495 .word 0x84c9c464 !aes_dround23 %f38,%f4,%f2,%f2 4496 .word 0x88ca4440 !aes_dround01 %f40,%f0,%f2,%f4 4497 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4498 .word 0x80cb4444 !aes_dround01 %f44,%f4,%f2,%f0 4499 .word 0x84cbc464 !aes_dround23 %f46,%f4,%f2,%f2 4500 .word 0x88cc4440 !aes_dround01 %f48,%f0,%f2,%f4 4501 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4502 .word 0x80cd4444 !aes_dround01 %f52,%f4,%f2,%f0 4503 .word 0x84cdc464 !aes_dround23 %f54,%f4,%f2,%f2 4504 .word 0x88ce4440 !aes_dround01 %f56,%f0,%f2,%f4 4505 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4506 .word 0x80cf4444 !aes_dround01 %f60,%f4,%f2,%f0 4507 .word 0x84cfc464 !aes_dround23 %f62,%f4,%f2,%f2 4508 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 4509 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4510 ldd [%i3 + 16], %f16 4511 ldd [%i3 + 24], %f18 4512 .word 0x80cd04c4 !aes_dround01_l %f20,%f4,%f2,%f0 4513 .word 0x84cd84e4 !aes_dround23_l %f22,%f4,%f2,%f2 4514 ldd [%i3 + 32], %f20 4515 retl 4516 ldd [%i3 + 40], %f22 4517 .type _aes256_decrypt_1x,#function 4518 .size _aes256_decrypt_1x,.-_aes256_decrypt_1x 4519 4520 .align 32 4521 _aes256_decrypt_2x: 4522 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 4523 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4524 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 4525 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 4526 ldd [%i3 + 208], %f16 4527 ldd [%i3 + 216], %f18 4528 .word 0x80cd0448 !aes_dround01 %f20,%f8,%f2,%f0 4529 .word 0x84cd8468 !aes_dround23 %f22,%f8,%f2,%f2 4530 .word 0x88cd0c4a !aes_dround01 %f20,%f10,%f6,%f4 4531 .word 0x8ccd8c6a !aes_dround23 %f22,%f10,%f6,%f6 4532 ldd [%i3 + 224], %f20 4533 ldd [%i3 + 232], %f22 4534 .word 0x90ce0440 !aes_dround01 %f24,%f0,%f2,%f8 4535 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4536 .word 0x94ce0c44 !aes_dround01 %f24,%f4,%f6,%f10 4537 .word 0x8cce8c64 !aes_dround23 %f26,%f4,%f6,%f6 4538 .word 0x80cf0448 !aes_dround01 %f28,%f8,%f2,%f0 4539 .word 0x84cf8468 !aes_dround23 %f30,%f8,%f2,%f2 4540 .word 0x88cf0c4a !aes_dround01 %f28,%f10,%f6,%f4 4541 .word 0x8ccf8c6a !aes_dround23 %f30,%f10,%f6,%f6 4542 .word 0x90c84440 !aes_dround01 %f32,%f0,%f2,%f8 4543 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4544 .word 0x94c84c44 !aes_dround01 %f32,%f4,%f6,%f10 4545 .word 0x8cc8cc64 !aes_dround23 %f34,%f4,%f6,%f6 4546 .word 0x80c94448 !aes_dround01 %f36,%f8,%f2,%f0 4547 .word 0x84c9c468 !aes_dround23 %f38,%f8,%f2,%f2 4548 .word 0x88c94c4a !aes_dround01 %f36,%f10,%f6,%f4 4549 .word 0x8cc9cc6a !aes_dround23 %f38,%f10,%f6,%f6 4550 .word 0x90ca4440 !aes_dround01 %f40,%f0,%f2,%f8 4551 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4552 .word 0x94ca4c44 !aes_dround01 %f40,%f4,%f6,%f10 4553 .word 0x8ccacc64 !aes_dround23 %f42,%f4,%f6,%f6 4554 .word 0x80cb4448 !aes_dround01 %f44,%f8,%f2,%f0 4555 .word 0x84cbc468 !aes_dround23 %f46,%f8,%f2,%f2 4556 .word 0x88cb4c4a !aes_dround01 %f44,%f10,%f6,%f4 4557 .word 0x8ccbcc6a !aes_dround23 %f46,%f10,%f6,%f6 4558 .word 0x90cc4440 !aes_dround01 %f48,%f0,%f2,%f8 4559 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4560 .word 0x94cc4c44 !aes_dround01 %f48,%f4,%f6,%f10 4561 .word 0x8ccccc64 !aes_dround23 %f50,%f4,%f6,%f6 4562 .word 0x80cd4448 !aes_dround01 %f52,%f8,%f2,%f0 4563 .word 0x84cdc468 !aes_dround23 %f54,%f8,%f2,%f2 4564 .word 0x88cd4c4a !aes_dround01 %f52,%f10,%f6,%f4 4565 .word 0x8ccdcc6a !aes_dround23 %f54,%f10,%f6,%f6 4566 .word 0x90ce4440 !aes_dround01 %f56,%f0,%f2,%f8 4567 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4568 .word 0x94ce4c44 !aes_dround01 %f56,%f4,%f6,%f10 4569 .word 0x8ccecc64 !aes_dround23 %f58,%f4,%f6,%f6 4570 .word 0x80cf4448 !aes_dround01 %f60,%f8,%f2,%f0 4571 .word 0x84cfc468 !aes_dround23 %f62,%f8,%f2,%f2 4572 .word 0x88cf4c4a !aes_dround01 %f60,%f10,%f6,%f4 4573 .word 0x8ccfcc6a !aes_dround23 %f62,%f10,%f6,%f6 4574 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 4575 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4576 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 4577 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 4578 ldd [%i3 + 16], %f16 4579 ldd [%i3 + 24], %f18 4580 .word 0x80cd04c8 !aes_dround01_l %f20,%f8,%f2,%f0 4581 .word 0x84cd84e8 !aes_dround23_l %f22,%f8,%f2,%f2 4582 .word 0x88cd0cca !aes_dround01_l %f20,%f10,%f6,%f4 4583 .word 0x8ccd8cea !aes_dround23_l %f22,%f10,%f6,%f6 4584 ldd [%i3 + 32], %f20 4585 retl 4586 ldd [%i3 + 40], %f22 4587 .type _aes256_decrypt_2x,#function 4588 .size _aes256_decrypt_2x,.-_aes256_decrypt_2x 4589 4590 .align 32 4591 _aes192_decrypt_1x: 4592 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 4593 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4594 .word 0x80cd0444 !aes_dround01 %f20,%f4,%f2,%f0 4595 .word 0x84cd8464 !aes_dround23 %f22,%f4,%f2,%f2 4596 .word 0x88ce0440 !aes_dround01 %f24,%f0,%f2,%f4 4597 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4598 .word 0x80cf0444 !aes_dround01 %f28,%f4,%f2,%f0 4599 .word 0x84cf8464 !aes_dround23 %f30,%f4,%f2,%f2 4600 .word 0x88c84440 !aes_dround01 %f32,%f0,%f2,%f4 4601 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4602 .word 0x80c94444 !aes_dround01 %f36,%f4,%f2,%f0 4603 .word 0x84c9c464 !aes_dround23 %f38,%f4,%f2,%f2 4604 .word 0x88ca4440 !aes_dround01 %f40,%f0,%f2,%f4 4605 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4606 .word 0x80cb4444 !aes_dround01 %f44,%f4,%f2,%f0 4607 .word 0x84cbc464 !aes_dround23 %f46,%f4,%f2,%f2 4608 .word 0x88cc4440 !aes_dround01 %f48,%f0,%f2,%f4 4609 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4610 .word 0x80cd4444 !aes_dround01 %f52,%f4,%f2,%f0 4611 .word 0x84cdc464 !aes_dround23 %f54,%f4,%f2,%f2 4612 .word 0x88ce4440 !aes_dround01 %f56,%f0,%f2,%f4 4613 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4614 .word 0x80cf44c4 !aes_dround01_l %f60,%f4,%f2,%f0 4615 retl 4616 .word 0x84cfc4e4 !aes_dround23_l %f62,%f4,%f2,%f2 4617 .type _aes192_decrypt_1x,#function 4618 .size _aes192_decrypt_1x,.-_aes192_decrypt_1x 4619 4620 .align 32 4621 _aes192_decrypt_2x: 4622 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 4623 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4624 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 4625 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 4626 .word 0x80cd0448 !aes_dround01 %f20,%f8,%f2,%f0 4627 .word 0x84cd8468 !aes_dround23 %f22,%f8,%f2,%f2 4628 .word 0x88cd0c4a !aes_dround01 %f20,%f10,%f6,%f4 4629 .word 0x8ccd8c6a !aes_dround23 %f22,%f10,%f6,%f6 4630 .word 0x90ce0440 !aes_dround01 %f24,%f0,%f2,%f8 4631 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4632 .word 0x94ce0c44 !aes_dround01 %f24,%f4,%f6,%f10 4633 .word 0x8cce8c64 !aes_dround23 %f26,%f4,%f6,%f6 4634 .word 0x80cf0448 !aes_dround01 %f28,%f8,%f2,%f0 4635 .word 0x84cf8468 !aes_dround23 %f30,%f8,%f2,%f2 4636 .word 0x88cf0c4a !aes_dround01 %f28,%f10,%f6,%f4 4637 .word 0x8ccf8c6a !aes_dround23 %f30,%f10,%f6,%f6 4638 .word 0x90c84440 !aes_dround01 %f32,%f0,%f2,%f8 4639 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4640 .word 0x94c84c44 !aes_dround01 %f32,%f4,%f6,%f10 4641 .word 0x8cc8cc64 !aes_dround23 %f34,%f4,%f6,%f6 4642 .word 0x80c94448 !aes_dround01 %f36,%f8,%f2,%f0 4643 .word 0x84c9c468 !aes_dround23 %f38,%f8,%f2,%f2 4644 .word 0x88c94c4a !aes_dround01 %f36,%f10,%f6,%f4 4645 .word 0x8cc9cc6a !aes_dround23 %f38,%f10,%f6,%f6 4646 .word 0x90ca4440 !aes_dround01 %f40,%f0,%f2,%f8 4647 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4648 .word 0x94ca4c44 !aes_dround01 %f40,%f4,%f6,%f10 4649 .word 0x8ccacc64 !aes_dround23 %f42,%f4,%f6,%f6 4650 .word 0x80cb4448 !aes_dround01 %f44,%f8,%f2,%f0 4651 .word 0x84cbc468 !aes_dround23 %f46,%f8,%f2,%f2 4652 .word 0x88cb4c4a !aes_dround01 %f44,%f10,%f6,%f4 4653 .word 0x8ccbcc6a !aes_dround23 %f46,%f10,%f6,%f6 4654 .word 0x90cc4440 !aes_dround01 %f48,%f0,%f2,%f8 4655 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4656 .word 0x94cc4c44 !aes_dround01 %f48,%f4,%f6,%f10 4657 .word 0x8ccccc64 !aes_dround23 %f50,%f4,%f6,%f6 4658 .word 0x80cd4448 !aes_dround01 %f52,%f8,%f2,%f0 4659 .word 0x84cdc468 !aes_dround23 %f54,%f8,%f2,%f2 4660 .word 0x88cd4c4a !aes_dround01 %f52,%f10,%f6,%f4 4661 .word 0x8ccdcc6a !aes_dround23 %f54,%f10,%f6,%f6 4662 .word 0x90ce4440 !aes_dround01 %f56,%f0,%f2,%f8 4663 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4664 .word 0x94ce4c44 !aes_dround01 %f56,%f4,%f6,%f10 4665 .word 0x8ccecc64 !aes_dround23 %f58,%f4,%f6,%f6 4666 .word 0x80cf44c8 !aes_dround01_l %f60,%f8,%f2,%f0 4667 .word 0x84cfc4e8 !aes_dround23_l %f62,%f8,%f2,%f2 4668 .word 0x88cf4cca !aes_dround01_l %f60,%f10,%f6,%f4 4669 retl 4670 .word 0x8ccfccea !aes_dround23_l %f62,%f10,%f6,%f6 4671 .type _aes192_decrypt_2x,#function 4672 .size _aes192_decrypt_2x,.-_aes192_decrypt_2x 4673 .asciz "AES for SPARC T4, David S. Miller, Andy Polyakov" 4674 .align 4 4675
Indexes created Fri Mar 06 19:51:54 UTC 2026