aest4-sparcv9.S revision 1.1.1.2
1 #include "sparc_arch.h" 2 3 #ifdef __arch64__ 4 .register %g2,#scratch 5 .register %g3,#scratch 6 #endif 7 8 .text 9 10 .globl aes_t4_encrypt 11 .align 32 12 aes_t4_encrypt: 13 andcc %o0, 7, %g1 ! is input aligned? 14 andn %o0, 7, %o0 15 16 ldx [%o2 + 0], %g4 17 ldx [%o2 + 8], %g5 18 19 ldx [%o0 + 0], %o4 20 bz,pt %icc, 1f 21 ldx [%o0 + 8], %o5 22 ldx [%o0 + 16], %o0 23 sll %g1, 3, %g1 24 sub %g0, %g1, %o3 25 sllx %o4, %g1, %o4 26 sllx %o5, %g1, %g1 27 srlx %o5, %o3, %o5 28 srlx %o0, %o3, %o3 29 or %o5, %o4, %o4 30 or %o3, %g1, %o5 31 1: 32 ld [%o2 + 240], %o3 33 ldd [%o2 + 16], %f12 34 ldd [%o2 + 24], %f14 35 xor %g4, %o4, %o4 36 xor %g5, %o5, %o5 37 .word 0x81b0230c !movxtod %o4,%f0 38 .word 0x85b0230d !movxtod %o5,%f2 39 srl %o3, 1, %o3 40 ldd [%o2 + 32], %f16 41 sub %o3, 1, %o3 42 ldd [%o2 + 40], %f18 43 add %o2, 48, %o2 44 45 .Lenc: 46 .word 0x88cb0400 !aes_eround01 %f12,%f0,%f2,%f4 47 .word 0x84cb8420 !aes_eround23 %f14,%f0,%f2,%f2 48 ldd [%o2 + 0], %f12 49 ldd [%o2 + 8], %f14 50 sub %o3,1,%o3 51 .word 0x80cc0404 !aes_eround01 %f16,%f4,%f2,%f0 52 .word 0x84cc8424 !aes_eround23 %f18,%f4,%f2,%f2 53 ldd [%o2 + 16], %f16 54 ldd [%o2 + 24], %f18 55 brnz,pt %o3, .Lenc 56 add %o2, 32, %o2 57 58 andcc %o1, 7, %o4 ! is output aligned? 59 .word 0x88cb0400 !aes_eround01 %f12,%f0,%f2,%f4 60 .word 0x84cb8420 !aes_eround23 %f14,%f0,%f2,%f2 61 .word 0x80cc0484 !aes_eround01_l %f16,%f4,%f2,%f0 62 .word 0x84cc84a4 !aes_eround23_l %f18,%f4,%f2,%f2 63 64 bnz,pn %icc, 2f 65 nop 66 67 std %f0, [%o1 + 0] 68 retl 69 std %f2, [%o1 + 8] 70 71 2: .word 0x93b24340 !alignaddrl %o1,%g0,%o1 72 mov 0xff, %o5 73 srl %o5, %o4, %o5 74 75 .word 0x89b00900 !faligndata %f0,%f0,%f4 76 .word 0x8db00902 !faligndata %f0,%f2,%f6 77 .word 0x91b08902 !faligndata %f2,%f2,%f8 78 79 stda %f4, [%o1 + %o5]0xc0 ! partial store 80 std %f6, [%o1 + 8] 81 add %o1, 16, %o1 82 orn %g0, %o5, %o5 83 retl 84 stda %f8, [%o1 + %o5]0xc0 ! partial store 85 .type aes_t4_encrypt,#function 86 .size aes_t4_encrypt,.-aes_t4_encrypt 87 88 .globl aes_t4_decrypt 89 .align 32 90 aes_t4_decrypt: 91 andcc %o0, 7, %g1 ! is input aligned? 92 andn %o0, 7, %o0 93 94 ldx [%o2 + 0], %g4 95 ldx [%o2 + 8], %g5 96 97 ldx [%o0 + 0], %o4 98 bz,pt %icc, 1f 99 ldx [%o0 + 8], %o5 100 ldx [%o0 + 16], %o0 101 sll %g1, 3, %g1 102 sub %g0, %g1, %o3 103 sllx %o4, %g1, %o4 104 sllx %o5, %g1, %g1 105 srlx %o5, %o3, %o5 106 srlx %o0, %o3, %o3 107 or %o5, %o4, %o4 108 or %o3, %g1, %o5 109 1: 110 ld [%o2 + 240], %o3 111 ldd [%o2 + 16], %f12 112 ldd [%o2 + 24], %f14 113 xor %g4, %o4, %o4 114 xor %g5, %o5, %o5 115 .word 0x81b0230c !movxtod %o4,%f0 116 .word 0x85b0230d !movxtod %o5,%f2 117 srl %o3, 1, %o3 118 ldd [%o2 + 32], %f16 119 sub %o3, 1, %o3 120 ldd [%o2 + 40], %f18 121 add %o2, 48, %o2 122 123 .Ldec: 124 .word 0x88cb0440 !aes_dround01 %f12,%f0,%f2,%f4 125 .word 0x84cb8460 !aes_dround23 %f14,%f0,%f2,%f2 126 ldd [%o2 + 0], %f12 127 ldd [%o2 + 8], %f14 128 sub %o3,1,%o3 129 .word 0x80cc0444 !aes_dround01 %f16,%f4,%f2,%f0 130 .word 0x84cc8464 !aes_dround23 %f18,%f4,%f2,%f2 131 ldd [%o2 + 16], %f16 132 ldd [%o2 + 24], %f18 133 brnz,pt %o3, .Ldec 134 add %o2, 32, %o2 135 136 andcc %o1, 7, %o4 ! is output aligned? 137 .word 0x88cb0440 !aes_dround01 %f12,%f0,%f2,%f4 138 .word 0x84cb8460 !aes_dround23 %f14,%f0,%f2,%f2 139 .word 0x80cc04c4 !aes_dround01_l %f16,%f4,%f2,%f0 140 .word 0x84cc84e4 !aes_dround23_l %f18,%f4,%f2,%f2 141 142 bnz,pn %icc, 2f 143 nop 144 145 std %f0, [%o1 + 0] 146 retl 147 std %f2, [%o1 + 8] 148 149 2: .word 0x93b24340 !alignaddrl %o1,%g0,%o1 150 mov 0xff, %o5 151 srl %o5, %o4, %o5 152 153 .word 0x89b00900 !faligndata %f0,%f0,%f4 154 .word 0x8db00902 !faligndata %f0,%f2,%f6 155 .word 0x91b08902 !faligndata %f2,%f2,%f8 156 157 stda %f4, [%o1 + %o5]0xc0 ! partial store 158 std %f6, [%o1 + 8] 159 add %o1, 16, %o1 160 orn %g0, %o5, %o5 161 retl 162 stda %f8, [%o1 + %o5]0xc0 ! partial store 163 .type aes_t4_decrypt,#function 164 .size aes_t4_decrypt,.-aes_t4_decrypt 165 .globl aes_t4_set_encrypt_key 166 .align 32 167 aes_t4_set_encrypt_key: 168 .Lset_encrypt_key: 169 and %o0, 7, %o3 170 .word 0x91b20300 !alignaddr %o0,%g0,%o0 171 cmp %o1, 192 172 ldd [%o0 + 0], %f0 173 bl,pt %icc,.L128 174 ldd [%o0 + 8], %f2 175 176 be,pt %icc,.L192 177 ldd [%o0 + 16], %f4 178 brz,pt %o3, .L256aligned 179 ldd [%o0 + 24], %f6 180 181 ldd [%o0 + 32], %f8 182 .word 0x81b00902 !faligndata %f0,%f2,%f0 183 .word 0x85b08904 !faligndata %f2,%f4,%f2 184 .word 0x89b10906 !faligndata %f4,%f6,%f4 185 .word 0x8db18908 !faligndata %f6,%f8,%f6 186 .L256aligned: 187 std %f0, [%o2 + 0] 188 .word 0x80c80106 !aes_kexpand1 %f0,%f6,0,%f0 189 std %f2, [%o2 + 8] 190 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 191 std %f4, [%o2 + 16] 192 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 193 std %f6, [%o2 + 24] 194 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 195 std %f0, [%o2 + 32] 196 .word 0x80c80306 !aes_kexpand1 %f0,%f6,1,%f0 197 std %f2, [%o2 + 40] 198 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 199 std %f4, [%o2 + 48] 200 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 201 std %f6, [%o2 + 56] 202 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 203 std %f0, [%o2 + 64] 204 .word 0x80c80506 !aes_kexpand1 %f0,%f6,2,%f0 205 std %f2, [%o2 + 72] 206 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 207 std %f4, [%o2 + 80] 208 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 209 std %f6, [%o2 + 88] 210 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 211 std %f0, [%o2 + 96] 212 .word 0x80c80706 !aes_kexpand1 %f0,%f6,3,%f0 213 std %f2, [%o2 + 104] 214 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 215 std %f4, [%o2 + 112] 216 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 217 std %f6, [%o2 + 120] 218 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 219 std %f0, [%o2 + 128] 220 .word 0x80c80906 !aes_kexpand1 %f0,%f6,4,%f0 221 std %f2, [%o2 + 136] 222 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 223 std %f4, [%o2 + 144] 224 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 225 std %f6, [%o2 + 152] 226 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 227 std %f0, [%o2 + 160] 228 .word 0x80c80b06 !aes_kexpand1 %f0,%f6,5,%f0 229 std %f2, [%o2 + 168] 230 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 231 std %f4, [%o2 + 176] 232 .word 0x89b12602 !aes_kexpand0 %f4,%f2,%f4 233 std %f6, [%o2 + 184] 234 .word 0x8db1a624 !aes_kexpand2 %f6,%f4,%f6 235 std %f0, [%o2 + 192] 236 .word 0x80c80d06 !aes_kexpand1 %f0,%f6,6,%f0 237 std %f2, [%o2 + 200] 238 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 239 std %f4, [%o2 + 208] 240 std %f6, [%o2 + 216] 241 std %f0, [%o2 + 224] 242 std %f2, [%o2 + 232] 243 244 mov 14, %o3 245 st %o3, [%o2 + 240] 246 retl 247 xor %o0, %o0, %o0 248 249 .align 16 250 .L192: 251 brz,pt %o3, .L192aligned 252 nop 253 254 ldd [%o0 + 24], %f6 255 .word 0x81b00902 !faligndata %f0,%f2,%f0 256 .word 0x85b08904 !faligndata %f2,%f4,%f2 257 .word 0x89b10906 !faligndata %f4,%f6,%f4 258 .L192aligned: 259 std %f0, [%o2 + 0] 260 .word 0x80c80104 !aes_kexpand1 %f0,%f4,0,%f0 261 std %f2, [%o2 + 8] 262 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 263 std %f4, [%o2 + 16] 264 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 265 std %f0, [%o2 + 24] 266 .word 0x80c80304 !aes_kexpand1 %f0,%f4,1,%f0 267 std %f2, [%o2 + 32] 268 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 269 std %f4, [%o2 + 40] 270 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 271 std %f0, [%o2 + 48] 272 .word 0x80c80504 !aes_kexpand1 %f0,%f4,2,%f0 273 std %f2, [%o2 + 56] 274 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 275 std %f4, [%o2 + 64] 276 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 277 std %f0, [%o2 + 72] 278 .word 0x80c80704 !aes_kexpand1 %f0,%f4,3,%f0 279 std %f2, [%o2 + 80] 280 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 281 std %f4, [%o2 + 88] 282 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 283 std %f0, [%o2 + 96] 284 .word 0x80c80904 !aes_kexpand1 %f0,%f4,4,%f0 285 std %f2, [%o2 + 104] 286 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 287 std %f4, [%o2 + 112] 288 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 289 std %f0, [%o2 + 120] 290 .word 0x80c80b04 !aes_kexpand1 %f0,%f4,5,%f0 291 std %f2, [%o2 + 128] 292 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 293 std %f4, [%o2 + 136] 294 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 295 std %f0, [%o2 + 144] 296 .word 0x80c80d04 !aes_kexpand1 %f0,%f4,6,%f0 297 std %f2, [%o2 + 152] 298 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 299 std %f4, [%o2 + 160] 300 .word 0x89b12622 !aes_kexpand2 %f4,%f2,%f4 301 std %f0, [%o2 + 168] 302 .word 0x80c80f04 !aes_kexpand1 %f0,%f4,7,%f0 303 std %f2, [%o2 + 176] 304 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 305 std %f4, [%o2 + 184] 306 std %f0, [%o2 + 192] 307 std %f2, [%o2 + 200] 308 309 mov 12, %o3 310 st %o3, [%o2 + 240] 311 retl 312 xor %o0, %o0, %o0 313 314 .align 16 315 .L128: 316 brz,pt %o3, .L128aligned 317 nop 318 319 ldd [%o0 + 16], %f4 320 .word 0x81b00902 !faligndata %f0,%f2,%f0 321 .word 0x85b08904 !faligndata %f2,%f4,%f2 322 .L128aligned: 323 std %f0, [%o2 + 0] 324 .word 0x80c80102 !aes_kexpand1 %f0,%f2,0,%f0 325 std %f2, [%o2 + 8] 326 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 327 std %f0, [%o2 + 16] 328 .word 0x80c80302 !aes_kexpand1 %f0,%f2,1,%f0 329 std %f2, [%o2 + 24] 330 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 331 std %f0, [%o2 + 32] 332 .word 0x80c80502 !aes_kexpand1 %f0,%f2,2,%f0 333 std %f2, [%o2 + 40] 334 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 335 std %f0, [%o2 + 48] 336 .word 0x80c80702 !aes_kexpand1 %f0,%f2,3,%f0 337 std %f2, [%o2 + 56] 338 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 339 std %f0, [%o2 + 64] 340 .word 0x80c80902 !aes_kexpand1 %f0,%f2,4,%f0 341 std %f2, [%o2 + 72] 342 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 343 std %f0, [%o2 + 80] 344 .word 0x80c80b02 !aes_kexpand1 %f0,%f2,5,%f0 345 std %f2, [%o2 + 88] 346 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 347 std %f0, [%o2 + 96] 348 .word 0x80c80d02 !aes_kexpand1 %f0,%f2,6,%f0 349 std %f2, [%o2 + 104] 350 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 351 std %f0, [%o2 + 112] 352 .word 0x80c80f02 !aes_kexpand1 %f0,%f2,7,%f0 353 std %f2, [%o2 + 120] 354 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 355 std %f0, [%o2 + 128] 356 .word 0x80c81102 !aes_kexpand1 %f0,%f2,8,%f0 357 std %f2, [%o2 + 136] 358 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 359 std %f0, [%o2 + 144] 360 .word 0x80c81302 !aes_kexpand1 %f0,%f2,9,%f0 361 std %f2, [%o2 + 152] 362 .word 0x85b0a620 !aes_kexpand2 %f2,%f0,%f2 363 std %f0, [%o2 + 160] 364 std %f2, [%o2 + 168] 365 366 mov 10, %o3 367 st %o3, [%o2 + 240] 368 retl 369 xor %o0, %o0, %o0 370 .type aes_t4_set_encrypt_key,#function 371 .size aes_t4_set_encrypt_key,.-aes_t4_set_encrypt_key 372 373 .globl aes_t4_set_decrypt_key 374 .align 32 375 aes_t4_set_decrypt_key: 376 mov %o7, %o5 377 call .Lset_encrypt_key 378 nop 379 380 mov %o5, %o7 381 sll %o3, 4, %o0 ! %o3 is number of rounds 382 add %o3, 2, %o3 383 add %o2, %o0, %o0 ! %o0=%o2+16*rounds 384 srl %o3, 2, %o3 ! %o3=(rounds+2)/4 385 386 .Lkey_flip: 387 ldd [%o2 + 0], %f0 388 ldd [%o2 + 8], %f2 389 ldd [%o2 + 16], %f4 390 ldd [%o2 + 24], %f6 391 ldd [%o0 + 0], %f8 392 ldd [%o0 + 8], %f10 393 ldd [%o0 - 16], %f12 394 ldd [%o0 - 8], %f14 395 sub %o3, 1, %o3 396 std %f0, [%o0 + 0] 397 std %f2, [%o0 + 8] 398 std %f4, [%o0 - 16] 399 std %f6, [%o0 - 8] 400 std %f8, [%o2 + 0] 401 std %f10, [%o2 + 8] 402 std %f12, [%o2 + 16] 403 std %f14, [%o2 + 24] 404 add %o2, 32, %o2 405 brnz %o3, .Lkey_flip 406 sub %o0, 32, %o0 407 408 retl 409 xor %o0, %o0, %o0 410 .type aes_t4_set_decrypt_key,#function 411 .size aes_t4_set_decrypt_key,.-aes_t4_set_decrypt_key 412 .align 32 413 _aes128_encrypt_1x: 414 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 415 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 416 .word 0x80cd0404 !aes_eround01 %f20,%f4,%f2,%f0 417 .word 0x84cd8424 !aes_eround23 %f22,%f4,%f2,%f2 418 .word 0x88ce0400 !aes_eround01 %f24,%f0,%f2,%f4 419 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 420 .word 0x80cf0404 !aes_eround01 %f28,%f4,%f2,%f0 421 .word 0x84cf8424 !aes_eround23 %f30,%f4,%f2,%f2 422 .word 0x88c84400 !aes_eround01 %f32,%f0,%f2,%f4 423 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 424 .word 0x80c94404 !aes_eround01 %f36,%f4,%f2,%f0 425 .word 0x84c9c424 !aes_eround23 %f38,%f4,%f2,%f2 426 .word 0x88ca4400 !aes_eround01 %f40,%f0,%f2,%f4 427 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 428 .word 0x80cb4404 !aes_eround01 %f44,%f4,%f2,%f0 429 .word 0x84cbc424 !aes_eround23 %f46,%f4,%f2,%f2 430 .word 0x88cc4400 !aes_eround01 %f48,%f0,%f2,%f4 431 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 432 .word 0x80cd4484 !aes_eround01_l %f52,%f4,%f2,%f0 433 retl 434 .word 0x84cdc4a4 !aes_eround23_l %f54,%f4,%f2,%f2 435 .type _aes128_encrypt_1x,#function 436 .size _aes128_encrypt_1x,.-_aes128_encrypt_1x 437 438 .align 32 439 _aes128_encrypt_2x: 440 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 441 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 442 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 443 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 444 .word 0x80cd0408 !aes_eround01 %f20,%f8,%f2,%f0 445 .word 0x84cd8428 !aes_eround23 %f22,%f8,%f2,%f2 446 .word 0x88cd0c0a !aes_eround01 %f20,%f10,%f6,%f4 447 .word 0x8ccd8c2a !aes_eround23 %f22,%f10,%f6,%f6 448 .word 0x90ce0400 !aes_eround01 %f24,%f0,%f2,%f8 449 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 450 .word 0x94ce0c04 !aes_eround01 %f24,%f4,%f6,%f10 451 .word 0x8cce8c24 !aes_eround23 %f26,%f4,%f6,%f6 452 .word 0x80cf0408 !aes_eround01 %f28,%f8,%f2,%f0 453 .word 0x84cf8428 !aes_eround23 %f30,%f8,%f2,%f2 454 .word 0x88cf0c0a !aes_eround01 %f28,%f10,%f6,%f4 455 .word 0x8ccf8c2a !aes_eround23 %f30,%f10,%f6,%f6 456 .word 0x90c84400 !aes_eround01 %f32,%f0,%f2,%f8 457 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 458 .word 0x94c84c04 !aes_eround01 %f32,%f4,%f6,%f10 459 .word 0x8cc8cc24 !aes_eround23 %f34,%f4,%f6,%f6 460 .word 0x80c94408 !aes_eround01 %f36,%f8,%f2,%f0 461 .word 0x84c9c428 !aes_eround23 %f38,%f8,%f2,%f2 462 .word 0x88c94c0a !aes_eround01 %f36,%f10,%f6,%f4 463 .word 0x8cc9cc2a !aes_eround23 %f38,%f10,%f6,%f6 464 .word 0x90ca4400 !aes_eround01 %f40,%f0,%f2,%f8 465 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 466 .word 0x94ca4c04 !aes_eround01 %f40,%f4,%f6,%f10 467 .word 0x8ccacc24 !aes_eround23 %f42,%f4,%f6,%f6 468 .word 0x80cb4408 !aes_eround01 %f44,%f8,%f2,%f0 469 .word 0x84cbc428 !aes_eround23 %f46,%f8,%f2,%f2 470 .word 0x88cb4c0a !aes_eround01 %f44,%f10,%f6,%f4 471 .word 0x8ccbcc2a !aes_eround23 %f46,%f10,%f6,%f6 472 .word 0x90cc4400 !aes_eround01 %f48,%f0,%f2,%f8 473 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 474 .word 0x94cc4c04 !aes_eround01 %f48,%f4,%f6,%f10 475 .word 0x8ccccc24 !aes_eround23 %f50,%f4,%f6,%f6 476 .word 0x80cd4488 !aes_eround01_l %f52,%f8,%f2,%f0 477 .word 0x84cdc4a8 !aes_eround23_l %f54,%f8,%f2,%f2 478 .word 0x88cd4c8a !aes_eround01_l %f52,%f10,%f6,%f4 479 retl 480 .word 0x8ccdccaa !aes_eround23_l %f54,%f10,%f6,%f6 481 .type _aes128_encrypt_2x,#function 482 .size _aes128_encrypt_2x,.-_aes128_encrypt_2x 483 484 .align 32 485 _aes128_loadkey: 486 ldx [%i3 + 0], %g4 487 ldx [%i3 + 8], %g5 488 ldd [%i3 + 16], %f16 489 ldd [%i3 + 24], %f18 490 ldd [%i3 + 32], %f20 491 ldd [%i3 + 40], %f22 492 ldd [%i3 + 48], %f24 493 ldd [%i3 + 56], %f26 494 ldd [%i3 + 64], %f28 495 ldd [%i3 + 72], %f30 496 ldd [%i3 + 80], %f32 497 ldd [%i3 + 88], %f34 498 ldd [%i3 + 96], %f36 499 ldd [%i3 + 104], %f38 500 ldd [%i3 + 112], %f40 501 ldd [%i3 + 120], %f42 502 ldd [%i3 + 128], %f44 503 ldd [%i3 + 136], %f46 504 ldd [%i3 + 144], %f48 505 ldd [%i3 + 152], %f50 506 ldd [%i3 + 160], %f52 507 ldd [%i3 + 168], %f54 508 retl 509 nop 510 .type _aes128_loadkey,#function 511 .size _aes128_loadkey,.-_aes128_loadkey 512 _aes128_load_enckey=_aes128_loadkey 513 _aes128_load_deckey=_aes128_loadkey 514 515 .globl aes128_t4_cbc_encrypt 516 .align 32 517 aes128_t4_cbc_encrypt: 518 save %sp, -STACK_FRAME, %sp 519 cmp %i2, 0 520 be,pn SIZE_T_CC, .L128_cbc_enc_abort 521 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 522 sub %i0, %i1, %l5 ! %i0!=%i1 523 ld [%i4 + 0], %f0 524 ld [%i4 + 4], %f1 525 ld [%i4 + 8], %f2 526 ld [%i4 + 12], %f3 527 prefetch [%i0], 20 528 prefetch [%i0 + 63], 20 529 call _aes128_load_enckey 530 and %i0, 7, %l0 531 andn %i0, 7, %i0 532 sll %l0, 3, %l0 533 mov 64, %l1 534 mov 0xff, %l3 535 sub %l1, %l0, %l1 536 and %i1, 7, %l2 537 cmp %i2, 127 538 movrnz %l2, 0, %l5 ! if ( %i1&7 || 539 movleu SIZE_T_CC, 0, %l5 ! %i2<128 || 540 brnz,pn %l5, .L128cbc_enc_blk ! %i0==%i1) 541 srl %l3, %l2, %l3 542 543 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 544 srlx %i2, 4, %i2 545 prefetch [%i1], 22 546 547 .L128_cbc_enc_loop: 548 ldx [%i0 + 0], %o0 549 brz,pt %l0, 4f 550 ldx [%i0 + 8], %o1 551 552 ldx [%i0 + 16], %o2 553 sllx %o0, %l0, %o0 554 srlx %o1, %l1, %g1 555 sllx %o1, %l0, %o1 556 or %g1, %o0, %o0 557 srlx %o2, %l1, %o2 558 or %o2, %o1, %o1 559 4: 560 xor %g4, %o0, %o0 ! ^= rk[0] 561 xor %g5, %o1, %o1 562 .word 0x99b02308 !movxtod %o0,%f12 563 .word 0x9db02309 !movxtod %o1,%f14 564 565 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 566 .word 0x85b38d82 !fxor %f14,%f2,%f2 567 prefetch [%i1 + 63], 22 568 prefetch [%i0 + 16+63], 20 569 call _aes128_encrypt_1x 570 add %i0, 16, %i0 571 572 brnz,pn %l2, 2f 573 sub %i2, 1, %i2 574 575 std %f0, [%i1 + 0] 576 std %f2, [%i1 + 8] 577 brnz,pt %i2, .L128_cbc_enc_loop 578 add %i1, 16, %i1 579 st %f0, [%i4 + 0] 580 st %f1, [%i4 + 4] 581 st %f2, [%i4 + 8] 582 st %f3, [%i4 + 12] 583 .L128_cbc_enc_abort: 584 ret 585 restore 586 587 .align 16 588 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 589 ! and ~3x deterioration 590 ! in inp==out case 591 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 592 .word 0x8db00902 !faligndata %f0,%f2,%f6 593 .word 0x91b08902 !faligndata %f2,%f2,%f8 594 595 stda %f4, [%i1 + %l3]0xc0 ! partial store 596 std %f6, [%i1 + 8] 597 add %i1, 16, %i1 598 orn %g0, %l3, %l3 599 stda %f8, [%i1 + %l3]0xc0 ! partial store 600 601 brnz,pt %i2, .L128_cbc_enc_loop+4 602 orn %g0, %l3, %l3 603 st %f0, [%i4 + 0] 604 st %f1, [%i4 + 4] 605 st %f2, [%i4 + 8] 606 st %f3, [%i4 + 12] 607 ret 608 restore 609 610 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 611 .align 32 612 .L128cbc_enc_blk: 613 add %i1, %i2, %l5 614 and %l5, 63, %l5 ! tail 615 sub %i2, %l5, %i2 616 add %l5, 15, %l5 ! round up to 16n 617 srlx %i2, 4, %i2 618 srl %l5, 4, %l5 619 620 .L128_cbc_enc_blk_loop: 621 ldx [%i0 + 0], %o0 622 brz,pt %l0, 5f 623 ldx [%i0 + 8], %o1 624 625 ldx [%i0 + 16], %o2 626 sllx %o0, %l0, %o0 627 srlx %o1, %l1, %g1 628 sllx %o1, %l0, %o1 629 or %g1, %o0, %o0 630 srlx %o2, %l1, %o2 631 or %o2, %o1, %o1 632 5: 633 xor %g4, %o0, %o0 ! ^= rk[0] 634 xor %g5, %o1, %o1 635 .word 0x99b02308 !movxtod %o0,%f12 636 .word 0x9db02309 !movxtod %o1,%f14 637 638 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 639 .word 0x85b38d82 !fxor %f14,%f2,%f2 640 prefetch [%i0 + 16+63], 20 641 call _aes128_encrypt_1x 642 add %i0, 16, %i0 643 sub %i2, 1, %i2 644 645 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 646 add %i1, 8, %i1 647 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 648 brnz,pt %i2, .L128_cbc_enc_blk_loop 649 add %i1, 8, %i1 650 651 membar #StoreLoad|#StoreStore 652 brnz,pt %l5, .L128_cbc_enc_loop 653 mov %l5, %i2 654 st %f0, [%i4 + 0] 655 st %f1, [%i4 + 4] 656 st %f2, [%i4 + 8] 657 st %f3, [%i4 + 12] 658 ret 659 restore 660 .type aes128_t4_cbc_encrypt,#function 661 .size aes128_t4_cbc_encrypt,.-aes128_t4_cbc_encrypt 662 .globl aes128_t4_ctr32_encrypt 663 .align 32 664 aes128_t4_ctr32_encrypt: 665 save %sp, -STACK_FRAME, %sp 666 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 667 668 prefetch [%i0], 20 669 prefetch [%i0 + 63], 20 670 call _aes128_load_enckey 671 sllx %i2, 4, %i2 672 673 ld [%i4 + 0], %l4 ! counter 674 ld [%i4 + 4], %l5 675 ld [%i4 + 8], %l6 676 ld [%i4 + 12], %l7 677 678 sllx %l4, 32, %o5 679 or %l5, %o5, %o5 680 sllx %l6, 32, %g1 681 xor %o5, %g4, %g4 ! ^= rk[0] 682 xor %g1, %g5, %g5 683 .word 0x9db02304 !movxtod %g4,%f14 ! most significant 64 bits 684 685 sub %i0, %i1, %l5 ! %i0!=%i1 686 and %i0, 7, %l0 687 andn %i0, 7, %i0 688 sll %l0, 3, %l0 689 mov 64, %l1 690 mov 0xff, %l3 691 sub %l1, %l0, %l1 692 and %i1, 7, %l2 693 cmp %i2, 255 694 movrnz %l2, 0, %l5 ! if ( %i1&7 || 695 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 696 brnz,pn %l5, .L128_ctr32_blk ! %i0==%i1) 697 srl %l3, %l2, %l3 698 699 andcc %i2, 16, %g0 ! is number of blocks even? 700 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 701 bz %icc, .L128_ctr32_loop2x 702 srlx %i2, 4, %i2 703 .L128_ctr32_loop: 704 ldx [%i0 + 0], %o0 705 brz,pt %l0, 4f 706 ldx [%i0 + 8], %o1 707 708 ldx [%i0 + 16], %o2 709 sllx %o0, %l0, %o0 710 srlx %o1, %l1, %g1 711 sllx %o1, %l0, %o1 712 or %g1, %o0, %o0 713 srlx %o2, %l1, %o2 714 or %o2, %o1, %o1 715 4: 716 xor %g5, %l7, %g1 ! ^= rk[0] 717 add %l7, 1, %l7 718 .word 0x85b02301 !movxtod %g1,%f2 719 srl %l7, 0, %l7 ! clruw 720 prefetch [%i1 + 63], 22 721 prefetch [%i0 + 16+63], 20 722 .word 0x88cc040e !aes_eround01 %f16,%f14,%f2,%f4 723 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 724 call _aes128_encrypt_1x+8 725 add %i0, 16, %i0 726 727 .word 0x95b02308 !movxtod %o0,%f10 728 .word 0x99b02309 !movxtod %o1,%f12 729 .word 0x81b28d80 !fxor %f10,%f0,%f0 ! ^= inp 730 .word 0x85b30d82 !fxor %f12,%f2,%f2 731 732 brnz,pn %l2, 2f 733 sub %i2, 1, %i2 734 735 std %f0, [%i1 + 0] 736 std %f2, [%i1 + 8] 737 brnz,pt %i2, .L128_ctr32_loop2x 738 add %i1, 16, %i1 739 740 ret 741 restore 742 743 .align 16 744 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 745 ! and ~3x deterioration 746 ! in inp==out case 747 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 748 .word 0x8db00902 !faligndata %f0,%f2,%f6 749 .word 0x91b08902 !faligndata %f2,%f2,%f8 750 stda %f4, [%i1 + %l3]0xc0 ! partial store 751 std %f6, [%i1 + 8] 752 add %i1, 16, %i1 753 orn %g0, %l3, %l3 754 stda %f8, [%i1 + %l3]0xc0 ! partial store 755 756 brnz,pt %i2, .L128_ctr32_loop2x+4 757 orn %g0, %l3, %l3 758 759 ret 760 restore 761 762 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 763 .align 32 764 .L128_ctr32_loop2x: 765 ldx [%i0 + 0], %o0 766 ldx [%i0 + 8], %o1 767 ldx [%i0 + 16], %o2 768 brz,pt %l0, 4f 769 ldx [%i0 + 24], %o3 770 771 ldx [%i0 + 32], %o4 772 sllx %o0, %l0, %o0 773 srlx %o1, %l1, %g1 774 or %g1, %o0, %o0 775 sllx %o1, %l0, %o1 776 srlx %o2, %l1, %g1 777 or %g1, %o1, %o1 778 sllx %o2, %l0, %o2 779 srlx %o3, %l1, %g1 780 or %g1, %o2, %o2 781 sllx %o3, %l0, %o3 782 srlx %o4, %l1, %o4 783 or %o4, %o3, %o3 784 4: 785 xor %g5, %l7, %g1 ! ^= rk[0] 786 add %l7, 1, %l7 787 .word 0x85b02301 !movxtod %g1,%f2 788 srl %l7, 0, %l7 ! clruw 789 xor %g5, %l7, %g1 790 add %l7, 1, %l7 791 .word 0x8db02301 !movxtod %g1,%f6 792 srl %l7, 0, %l7 ! clruw 793 prefetch [%i1 + 63], 22 794 prefetch [%i0 + 32+63], 20 795 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 796 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 797 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 798 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 799 call _aes128_encrypt_2x+16 800 add %i0, 32, %i0 801 802 .word 0x91b02308 !movxtod %o0,%f8 803 .word 0x95b02309 !movxtod %o1,%f10 804 .word 0x99b0230a !movxtod %o2,%f12 805 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 806 .word 0x91b0230b !movxtod %o3,%f8 807 .word 0x85b28d82 !fxor %f10,%f2,%f2 808 .word 0x89b30d84 !fxor %f12,%f4,%f4 809 .word 0x8db20d86 !fxor %f8,%f6,%f6 810 811 brnz,pn %l2, 2f 812 sub %i2, 2, %i2 813 814 std %f0, [%i1 + 0] 815 std %f2, [%i1 + 8] 816 std %f4, [%i1 + 16] 817 std %f6, [%i1 + 24] 818 brnz,pt %i2, .L128_ctr32_loop2x 819 add %i1, 32, %i1 820 821 ret 822 restore 823 824 .align 16 825 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 826 ! and ~3x deterioration 827 ! in inp==out case 828 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 829 .word 0x81b00902 !faligndata %f0,%f2,%f0 830 .word 0x85b08904 !faligndata %f2,%f4,%f2 831 .word 0x89b10906 !faligndata %f4,%f6,%f4 832 .word 0x8db18906 !faligndata %f6,%f6,%f6 833 834 stda %f8, [%i1 + %l3]0xc0 ! partial store 835 std %f0, [%i1 + 8] 836 std %f2, [%i1 + 16] 837 std %f4, [%i1 + 24] 838 add %i1, 32, %i1 839 orn %g0, %l3, %l3 840 stda %f6, [%i1 + %l3]0xc0 ! partial store 841 842 brnz,pt %i2, .L128_ctr32_loop2x+4 843 orn %g0, %l3, %l3 844 845 ret 846 restore 847 848 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 849 .align 32 850 .L128_ctr32_blk: 851 add %i1, %i2, %l5 852 and %l5, 63, %l5 ! tail 853 sub %i2, %l5, %i2 854 add %l5, 15, %l5 ! round up to 16n 855 srlx %i2, 4, %i2 856 srl %l5, 4, %l5 857 sub %i2, 1, %i2 858 add %l5, 1, %l5 859 860 .L128_ctr32_blk_loop2x: 861 ldx [%i0 + 0], %o0 862 ldx [%i0 + 8], %o1 863 ldx [%i0 + 16], %o2 864 brz,pt %l0, 5f 865 ldx [%i0 + 24], %o3 866 867 ldx [%i0 + 32], %o4 868 sllx %o0, %l0, %o0 869 srlx %o1, %l1, %g1 870 or %g1, %o0, %o0 871 sllx %o1, %l0, %o1 872 srlx %o2, %l1, %g1 873 or %g1, %o1, %o1 874 sllx %o2, %l0, %o2 875 srlx %o3, %l1, %g1 876 or %g1, %o2, %o2 877 sllx %o3, %l0, %o3 878 srlx %o4, %l1, %o4 879 or %o4, %o3, %o3 880 5: 881 xor %g5, %l7, %g1 ! ^= rk[0] 882 add %l7, 1, %l7 883 .word 0x85b02301 !movxtod %g1,%f2 884 srl %l7, 0, %l7 ! clruw 885 xor %g5, %l7, %g1 886 add %l7, 1, %l7 887 .word 0x8db02301 !movxtod %g1,%f6 888 srl %l7, 0, %l7 ! clruw 889 prefetch [%i0 + 32+63], 20 890 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 891 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 892 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 893 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 894 call _aes128_encrypt_2x+16 895 add %i0, 32, %i0 896 subcc %i2, 2, %i2 897 898 .word 0x91b02308 !movxtod %o0,%f8 899 .word 0x95b02309 !movxtod %o1,%f10 900 .word 0x99b0230a !movxtod %o2,%f12 901 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 902 .word 0x91b0230b !movxtod %o3,%f8 903 .word 0x85b28d82 !fxor %f10,%f2,%f2 904 .word 0x89b30d84 !fxor %f12,%f4,%f4 905 .word 0x8db20d86 !fxor %f8,%f6,%f6 906 907 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 908 add %i1, 8, %i1 909 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 910 add %i1, 8, %i1 911 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 912 add %i1, 8, %i1 913 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 914 bgu,pt SIZE_T_CC, .L128_ctr32_blk_loop2x 915 add %i1, 8, %i1 916 917 add %l5, %i2, %i2 918 andcc %i2, 1, %g0 ! is number of blocks even? 919 membar #StoreLoad|#StoreStore 920 bnz,pt %icc, .L128_ctr32_loop 921 srl %i2, 0, %i2 922 brnz,pn %i2, .L128_ctr32_loop2x 923 nop 924 925 ret 926 restore 927 .type aes128_t4_ctr32_encrypt,#function 928 .size aes128_t4_ctr32_encrypt,.-aes128_t4_ctr32_encrypt 929 .globl aes128_t4_xts_encrypt 930 .align 32 931 aes128_t4_xts_encrypt: 932 save %sp, -STACK_FRAME-16, %sp 933 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 934 935 mov %i5, %o0 936 add %fp, STACK_BIAS-16, %o1 937 call aes_t4_encrypt 938 mov %i4, %o2 939 940 add %fp, STACK_BIAS-16, %l7 941 ldxa [%l7]0x88, %g2 942 add %fp, STACK_BIAS-8, %l7 943 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 944 945 sethi %hi(0x76543210), %l7 946 or %l7, %lo(0x76543210), %l7 947 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 948 949 prefetch [%i0], 20 950 prefetch [%i0 + 63], 20 951 call _aes128_load_enckey 952 and %i2, 15, %i5 953 and %i2, -16, %i2 954 955 sub %i0, %i1, %l5 ! %i0!=%i1 956 and %i0, 7, %l0 957 andn %i0, 7, %i0 958 sll %l0, 3, %l0 959 mov 64, %l1 960 mov 0xff, %l3 961 sub %l1, %l0, %l1 962 and %i1, 7, %l2 963 cmp %i2, 255 964 movrnz %l2, 0, %l5 ! if ( %i1&7 || 965 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 966 brnz,pn %l5, .L128_xts_enblk ! %i0==%i1) 967 srl %l3, %l2, %l3 968 969 andcc %i2, 16, %g0 ! is number of blocks even? 970 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 971 bz %icc, .L128_xts_enloop2x 972 srlx %i2, 4, %i2 973 .L128_xts_enloop: 974 ldx [%i0 + 0], %o0 975 brz,pt %l0, 4f 976 ldx [%i0 + 8], %o1 977 978 ldx [%i0 + 16], %o2 979 sllx %o0, %l0, %o0 980 srlx %o1, %l1, %g1 981 sllx %o1, %l0, %o1 982 or %g1, %o0, %o0 983 srlx %o2, %l1, %o2 984 or %o2, %o1, %o1 985 4: 986 .word 0x99b02302 !movxtod %g2,%f12 987 .word 0x9db02303 !movxtod %g3,%f14 988 .word 0x99b3098c !bshuffle %f12,%f12,%f12 989 .word 0x9db3898e !bshuffle %f14,%f14,%f14 990 991 xor %g4, %o0, %o0 ! ^= rk[0] 992 xor %g5, %o1, %o1 993 .word 0x81b02308 !movxtod %o0,%f0 994 .word 0x85b02309 !movxtod %o1,%f2 995 996 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 997 .word 0x85b38d82 !fxor %f14,%f2,%f2 998 999 prefetch [%i1 + 63], 22 1000 prefetch [%i0 + 16+63], 20 1001 call _aes128_encrypt_1x 1002 add %i0, 16, %i0 1003 1004 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1005 .word 0x85b38d82 !fxor %f14,%f2,%f2 1006 1007 srax %g3, 63, %l7 ! next tweak value 1008 addcc %g2, %g2, %g2 1009 and %l7, 0x87, %l7 1010 .word 0x87b0c223 !addxc %g3,%g3,%g3 1011 xor %l7, %g2, %g2 1012 1013 brnz,pn %l2, 2f 1014 sub %i2, 1, %i2 1015 1016 std %f0, [%i1 + 0] 1017 std %f2, [%i1 + 8] 1018 brnz,pt %i2, .L128_xts_enloop2x 1019 add %i1, 16, %i1 1020 1021 brnz,pn %i5, .L128_xts_ensteal 1022 nop 1023 1024 ret 1025 restore 1026 1027 .align 16 1028 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1029 ! and ~3x deterioration 1030 ! in inp==out case 1031 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 1032 .word 0x8db00902 !faligndata %f0,%f2,%f6 1033 .word 0x91b08902 !faligndata %f2,%f2,%f8 1034 stda %f4, [%i1 + %l3]0xc0 ! partial store 1035 std %f6, [%i1 + 8] 1036 add %i1, 16, %i1 1037 orn %g0, %l3, %l3 1038 stda %f8, [%i1 + %l3]0xc0 ! partial store 1039 1040 brnz,pt %i2, .L128_xts_enloop2x+4 1041 orn %g0, %l3, %l3 1042 1043 brnz,pn %i5, .L128_xts_ensteal 1044 nop 1045 1046 ret 1047 restore 1048 1049 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1050 .align 32 1051 .L128_xts_enloop2x: 1052 ldx [%i0 + 0], %o0 1053 ldx [%i0 + 8], %o1 1054 ldx [%i0 + 16], %o2 1055 brz,pt %l0, 4f 1056 ldx [%i0 + 24], %o3 1057 1058 ldx [%i0 + 32], %o4 1059 sllx %o0, %l0, %o0 1060 srlx %o1, %l1, %g1 1061 or %g1, %o0, %o0 1062 sllx %o1, %l0, %o1 1063 srlx %o2, %l1, %g1 1064 or %g1, %o1, %o1 1065 sllx %o2, %l0, %o2 1066 srlx %o3, %l1, %g1 1067 or %g1, %o2, %o2 1068 sllx %o3, %l0, %o3 1069 srlx %o4, %l1, %o4 1070 or %o4, %o3, %o3 1071 4: 1072 .word 0x99b02302 !movxtod %g2,%f12 1073 .word 0x9db02303 !movxtod %g3,%f14 1074 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1075 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1076 1077 srax %g3, 63, %l7 ! next tweak value 1078 addcc %g2, %g2, %g2 1079 and %l7, 0x87, %l7 1080 .word 0x87b0c223 !addxc %g3,%g3,%g3 1081 xor %l7, %g2, %g2 1082 1083 .word 0x91b02302 !movxtod %g2,%f8 1084 .word 0x95b02303 !movxtod %g3,%f10 1085 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1086 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1087 1088 xor %g4, %o0, %o0 ! ^= rk[0] 1089 xor %g5, %o1, %o1 1090 xor %g4, %o2, %o2 ! ^= rk[0] 1091 xor %g5, %o3, %o3 1092 .word 0x81b02308 !movxtod %o0,%f0 1093 .word 0x85b02309 !movxtod %o1,%f2 1094 .word 0x89b0230a !movxtod %o2,%f4 1095 .word 0x8db0230b !movxtod %o3,%f6 1096 1097 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1098 .word 0x85b38d82 !fxor %f14,%f2,%f2 1099 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1100 .word 0x8db28d86 !fxor %f10,%f6,%f6 1101 1102 prefetch [%i1 + 63], 22 1103 prefetch [%i0 + 32+63], 20 1104 call _aes128_encrypt_2x 1105 add %i0, 32, %i0 1106 1107 .word 0x91b02302 !movxtod %g2,%f8 1108 .word 0x95b02303 !movxtod %g3,%f10 1109 1110 srax %g3, 63, %l7 ! next tweak value 1111 addcc %g2, %g2, %g2 1112 and %l7, 0x87, %l7 1113 .word 0x87b0c223 !addxc %g3,%g3,%g3 1114 xor %l7, %g2, %g2 1115 1116 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1117 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1118 1119 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1120 .word 0x85b38d82 !fxor %f14,%f2,%f2 1121 .word 0x89b20d84 !fxor %f8,%f4,%f4 1122 .word 0x8db28d86 !fxor %f10,%f6,%f6 1123 1124 brnz,pn %l2, 2f 1125 sub %i2, 2, %i2 1126 1127 std %f0, [%i1 + 0] 1128 std %f2, [%i1 + 8] 1129 std %f4, [%i1 + 16] 1130 std %f6, [%i1 + 24] 1131 brnz,pt %i2, .L128_xts_enloop2x 1132 add %i1, 32, %i1 1133 1134 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1135 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1136 brnz,pn %i5, .L128_xts_ensteal 1137 nop 1138 1139 ret 1140 restore 1141 1142 .align 16 1143 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1144 ! and ~3x deterioration 1145 ! in inp==out case 1146 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 1147 .word 0x95b00902 !faligndata %f0,%f2,%f10 1148 .word 0x99b08904 !faligndata %f2,%f4,%f12 1149 .word 0x9db10906 !faligndata %f4,%f6,%f14 1150 .word 0x81b18906 !faligndata %f6,%f6,%f0 1151 1152 stda %f8, [%i1 + %l3]0xc0 ! partial store 1153 std %f10, [%i1 + 8] 1154 std %f12, [%i1 + 16] 1155 std %f14, [%i1 + 24] 1156 add %i1, 32, %i1 1157 orn %g0, %l3, %l3 1158 stda %f0, [%i1 + %l3]0xc0 ! partial store 1159 1160 brnz,pt %i2, .L128_xts_enloop2x+4 1161 orn %g0, %l3, %l3 1162 1163 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1164 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1165 brnz,pn %i5, .L128_xts_ensteal 1166 nop 1167 1168 ret 1169 restore 1170 1171 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1172 .align 32 1173 .L128_xts_enblk: 1174 add %i1, %i2, %l5 1175 and %l5, 63, %l5 ! tail 1176 sub %i2, %l5, %i2 1177 add %l5, 15, %l5 ! round up to 16n 1178 srlx %i2, 4, %i2 1179 srl %l5, 4, %l5 1180 sub %i2, 1, %i2 1181 add %l5, 1, %l5 1182 1183 .L128_xts_enblk2x: 1184 ldx [%i0 + 0], %o0 1185 ldx [%i0 + 8], %o1 1186 ldx [%i0 + 16], %o2 1187 brz,pt %l0, 5f 1188 ldx [%i0 + 24], %o3 1189 1190 ldx [%i0 + 32], %o4 1191 sllx %o0, %l0, %o0 1192 srlx %o1, %l1, %g1 1193 or %g1, %o0, %o0 1194 sllx %o1, %l0, %o1 1195 srlx %o2, %l1, %g1 1196 or %g1, %o1, %o1 1197 sllx %o2, %l0, %o2 1198 srlx %o3, %l1, %g1 1199 or %g1, %o2, %o2 1200 sllx %o3, %l0, %o3 1201 srlx %o4, %l1, %o4 1202 or %o4, %o3, %o3 1203 5: 1204 .word 0x99b02302 !movxtod %g2,%f12 1205 .word 0x9db02303 !movxtod %g3,%f14 1206 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1207 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1208 1209 srax %g3, 63, %l7 ! next tweak value 1210 addcc %g2, %g2, %g2 1211 and %l7, 0x87, %l7 1212 .word 0x87b0c223 !addxc %g3,%g3,%g3 1213 xor %l7, %g2, %g2 1214 1215 .word 0x91b02302 !movxtod %g2,%f8 1216 .word 0x95b02303 !movxtod %g3,%f10 1217 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1218 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1219 1220 xor %g4, %o0, %o0 ! ^= rk[0] 1221 xor %g5, %o1, %o1 1222 xor %g4, %o2, %o2 ! ^= rk[0] 1223 xor %g5, %o3, %o3 1224 .word 0x81b02308 !movxtod %o0,%f0 1225 .word 0x85b02309 !movxtod %o1,%f2 1226 .word 0x89b0230a !movxtod %o2,%f4 1227 .word 0x8db0230b !movxtod %o3,%f6 1228 1229 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1230 .word 0x85b38d82 !fxor %f14,%f2,%f2 1231 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1232 .word 0x8db28d86 !fxor %f10,%f6,%f6 1233 1234 prefetch [%i0 + 32+63], 20 1235 call _aes128_encrypt_2x 1236 add %i0, 32, %i0 1237 1238 .word 0x91b02302 !movxtod %g2,%f8 1239 .word 0x95b02303 !movxtod %g3,%f10 1240 1241 srax %g3, 63, %l7 ! next tweak value 1242 addcc %g2, %g2, %g2 1243 and %l7, 0x87, %l7 1244 .word 0x87b0c223 !addxc %g3,%g3,%g3 1245 xor %l7, %g2, %g2 1246 1247 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1248 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1249 1250 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1251 .word 0x85b38d82 !fxor %f14,%f2,%f2 1252 .word 0x89b20d84 !fxor %f8,%f4,%f4 1253 .word 0x8db28d86 !fxor %f10,%f6,%f6 1254 1255 subcc %i2, 2, %i2 1256 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1257 add %i1, 8, %i1 1258 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1259 add %i1, 8, %i1 1260 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1261 add %i1, 8, %i1 1262 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1263 bgu,pt SIZE_T_CC, .L128_xts_enblk2x 1264 add %i1, 8, %i1 1265 1266 add %l5, %i2, %i2 1267 andcc %i2, 1, %g0 ! is number of blocks even? 1268 membar #StoreLoad|#StoreStore 1269 bnz,pt %icc, .L128_xts_enloop 1270 srl %i2, 0, %i2 1271 brnz,pn %i2, .L128_xts_enloop2x 1272 nop 1273 1274 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1275 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1276 brnz,pn %i5, .L128_xts_ensteal 1277 nop 1278 1279 ret 1280 restore 1281 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1282 .align 32 1283 .L128_xts_ensteal: 1284 std %f0, [%fp + STACK_BIAS-16] ! copy of output 1285 std %f2, [%fp + STACK_BIAS-8] 1286 1287 srl %l0, 3, %l0 1288 add %fp, STACK_BIAS-16, %l7 1289 add %i0, %l0, %i0 ! original %i0+%i2&-15 1290 add %i1, %l2, %i1 ! original %i1+%i2&-15 1291 mov 0, %l0 1292 nop ! align 1293 1294 .L128_xts_enstealing: 1295 ldub [%i0 + %l0], %o0 1296 ldub [%l7 + %l0], %o1 1297 dec %i5 1298 stb %o0, [%l7 + %l0] 1299 stb %o1, [%i1 + %l0] 1300 brnz %i5, .L128_xts_enstealing 1301 inc %l0 1302 1303 mov %l7, %i0 1304 sub %i1, 16, %i1 1305 mov 0, %l0 1306 sub %i1, %l2, %i1 1307 ba .L128_xts_enloop ! one more time 1308 mov 1, %i2 ! %i5 is 0 1309 ret 1310 restore 1311 .type aes128_t4_xts_encrypt,#function 1312 .size aes128_t4_xts_encrypt,.-aes128_t4_xts_encrypt 1313 .globl aes128_t4_xts_decrypt 1314 .align 32 1315 aes128_t4_xts_decrypt: 1316 save %sp, -STACK_FRAME-16, %sp 1317 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 1318 1319 mov %i5, %o0 1320 add %fp, STACK_BIAS-16, %o1 1321 call aes_t4_encrypt 1322 mov %i4, %o2 1323 1324 add %fp, STACK_BIAS-16, %l7 1325 ldxa [%l7]0x88, %g2 1326 add %fp, STACK_BIAS-8, %l7 1327 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 1328 1329 sethi %hi(0x76543210), %l7 1330 or %l7, %lo(0x76543210), %l7 1331 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 1332 1333 prefetch [%i0], 20 1334 prefetch [%i0 + 63], 20 1335 call _aes128_load_deckey 1336 and %i2, 15, %i5 1337 and %i2, -16, %i2 1338 mov 0, %l7 1339 movrnz %i5, 16, %l7 1340 sub %i2, %l7, %i2 1341 1342 sub %i0, %i1, %l5 ! %i0!=%i1 1343 and %i0, 7, %l0 1344 andn %i0, 7, %i0 1345 sll %l0, 3, %l0 1346 mov 64, %l1 1347 mov 0xff, %l3 1348 sub %l1, %l0, %l1 1349 and %i1, 7, %l2 1350 cmp %i2, 255 1351 movrnz %l2, 0, %l5 ! if ( %i1&7 || 1352 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 1353 brnz,pn %l5, .L128_xts_deblk ! %i0==%i1) 1354 srl %l3, %l2, %l3 1355 1356 andcc %i2, 16, %g0 ! is number of blocks even? 1357 brz,pn %i2, .L128_xts_desteal 1358 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 1359 bz %icc, .L128_xts_deloop2x 1360 srlx %i2, 4, %i2 1361 .L128_xts_deloop: 1362 ldx [%i0 + 0], %o0 1363 brz,pt %l0, 4f 1364 ldx [%i0 + 8], %o1 1365 1366 ldx [%i0 + 16], %o2 1367 sllx %o0, %l0, %o0 1368 srlx %o1, %l1, %g1 1369 sllx %o1, %l0, %o1 1370 or %g1, %o0, %o0 1371 srlx %o2, %l1, %o2 1372 or %o2, %o1, %o1 1373 4: 1374 .word 0x99b02302 !movxtod %g2,%f12 1375 .word 0x9db02303 !movxtod %g3,%f14 1376 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1377 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1378 1379 xor %g4, %o0, %o0 ! ^= rk[0] 1380 xor %g5, %o1, %o1 1381 .word 0x81b02308 !movxtod %o0,%f0 1382 .word 0x85b02309 !movxtod %o1,%f2 1383 1384 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1385 .word 0x85b38d82 !fxor %f14,%f2,%f2 1386 1387 prefetch [%i1 + 63], 22 1388 prefetch [%i0 + 16+63], 20 1389 call _aes128_decrypt_1x 1390 add %i0, 16, %i0 1391 1392 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1393 .word 0x85b38d82 !fxor %f14,%f2,%f2 1394 1395 srax %g3, 63, %l7 ! next tweak value 1396 addcc %g2, %g2, %g2 1397 and %l7, 0x87, %l7 1398 .word 0x87b0c223 !addxc %g3,%g3,%g3 1399 xor %l7, %g2, %g2 1400 1401 brnz,pn %l2, 2f 1402 sub %i2, 1, %i2 1403 1404 std %f0, [%i1 + 0] 1405 std %f2, [%i1 + 8] 1406 brnz,pt %i2, .L128_xts_deloop2x 1407 add %i1, 16, %i1 1408 1409 brnz,pn %i5, .L128_xts_desteal 1410 nop 1411 1412 ret 1413 restore 1414 1415 .align 16 1416 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1417 ! and ~3x deterioration 1418 ! in inp==out case 1419 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 1420 .word 0x8db00902 !faligndata %f0,%f2,%f6 1421 .word 0x91b08902 !faligndata %f2,%f2,%f8 1422 stda %f4, [%i1 + %l3]0xc0 ! partial store 1423 std %f6, [%i1 + 8] 1424 add %i1, 16, %i1 1425 orn %g0, %l3, %l3 1426 stda %f8, [%i1 + %l3]0xc0 ! partial store 1427 1428 brnz,pt %i2, .L128_xts_deloop2x+4 1429 orn %g0, %l3, %l3 1430 1431 brnz,pn %i5, .L128_xts_desteal 1432 nop 1433 1434 ret 1435 restore 1436 1437 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1438 .align 32 1439 .L128_xts_deloop2x: 1440 ldx [%i0 + 0], %o0 1441 ldx [%i0 + 8], %o1 1442 ldx [%i0 + 16], %o2 1443 brz,pt %l0, 4f 1444 ldx [%i0 + 24], %o3 1445 1446 ldx [%i0 + 32], %o4 1447 sllx %o0, %l0, %o0 1448 srlx %o1, %l1, %g1 1449 or %g1, %o0, %o0 1450 sllx %o1, %l0, %o1 1451 srlx %o2, %l1, %g1 1452 or %g1, %o1, %o1 1453 sllx %o2, %l0, %o2 1454 srlx %o3, %l1, %g1 1455 or %g1, %o2, %o2 1456 sllx %o3, %l0, %o3 1457 srlx %o4, %l1, %o4 1458 or %o4, %o3, %o3 1459 4: 1460 .word 0x99b02302 !movxtod %g2,%f12 1461 .word 0x9db02303 !movxtod %g3,%f14 1462 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1463 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1464 1465 srax %g3, 63, %l7 ! next tweak value 1466 addcc %g2, %g2, %g2 1467 and %l7, 0x87, %l7 1468 .word 0x87b0c223 !addxc %g3,%g3,%g3 1469 xor %l7, %g2, %g2 1470 1471 .word 0x91b02302 !movxtod %g2,%f8 1472 .word 0x95b02303 !movxtod %g3,%f10 1473 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1474 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1475 1476 xor %g4, %o0, %o0 ! ^= rk[0] 1477 xor %g5, %o1, %o1 1478 xor %g4, %o2, %o2 ! ^= rk[0] 1479 xor %g5, %o3, %o3 1480 .word 0x81b02308 !movxtod %o0,%f0 1481 .word 0x85b02309 !movxtod %o1,%f2 1482 .word 0x89b0230a !movxtod %o2,%f4 1483 .word 0x8db0230b !movxtod %o3,%f6 1484 1485 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1486 .word 0x85b38d82 !fxor %f14,%f2,%f2 1487 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1488 .word 0x8db28d86 !fxor %f10,%f6,%f6 1489 1490 prefetch [%i1 + 63], 22 1491 prefetch [%i0 + 32+63], 20 1492 call _aes128_decrypt_2x 1493 add %i0, 32, %i0 1494 1495 .word 0x91b02302 !movxtod %g2,%f8 1496 .word 0x95b02303 !movxtod %g3,%f10 1497 1498 srax %g3, 63, %l7 ! next tweak value 1499 addcc %g2, %g2, %g2 1500 and %l7, 0x87, %l7 1501 .word 0x87b0c223 !addxc %g3,%g3,%g3 1502 xor %l7, %g2, %g2 1503 1504 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1505 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1506 1507 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1508 .word 0x85b38d82 !fxor %f14,%f2,%f2 1509 .word 0x89b20d84 !fxor %f8,%f4,%f4 1510 .word 0x8db28d86 !fxor %f10,%f6,%f6 1511 1512 brnz,pn %l2, 2f 1513 sub %i2, 2, %i2 1514 1515 std %f0, [%i1 + 0] 1516 std %f2, [%i1 + 8] 1517 std %f4, [%i1 + 16] 1518 std %f6, [%i1 + 24] 1519 brnz,pt %i2, .L128_xts_deloop2x 1520 add %i1, 32, %i1 1521 1522 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1523 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1524 brnz,pn %i5, .L128_xts_desteal 1525 nop 1526 1527 ret 1528 restore 1529 1530 .align 16 1531 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1532 ! and ~3x deterioration 1533 ! in inp==out case 1534 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 1535 .word 0x95b00902 !faligndata %f0,%f2,%f10 1536 .word 0x99b08904 !faligndata %f2,%f4,%f12 1537 .word 0x9db10906 !faligndata %f4,%f6,%f14 1538 .word 0x81b18906 !faligndata %f6,%f6,%f0 1539 1540 stda %f8, [%i1 + %l3]0xc0 ! partial store 1541 std %f10, [%i1 + 8] 1542 std %f12, [%i1 + 16] 1543 std %f14, [%i1 + 24] 1544 add %i1, 32, %i1 1545 orn %g0, %l3, %l3 1546 stda %f0, [%i1 + %l3]0xc0 ! partial store 1547 1548 brnz,pt %i2, .L128_xts_deloop2x+4 1549 orn %g0, %l3, %l3 1550 1551 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1552 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1553 brnz,pn %i5, .L128_xts_desteal 1554 nop 1555 1556 ret 1557 restore 1558 1559 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1560 .align 32 1561 .L128_xts_deblk: 1562 add %i1, %i2, %l5 1563 and %l5, 63, %l5 ! tail 1564 sub %i2, %l5, %i2 1565 add %l5, 15, %l5 ! round up to 16n 1566 srlx %i2, 4, %i2 1567 srl %l5, 4, %l5 1568 sub %i2, 1, %i2 1569 add %l5, 1, %l5 1570 1571 .L128_xts_deblk2x: 1572 ldx [%i0 + 0], %o0 1573 ldx [%i0 + 8], %o1 1574 ldx [%i0 + 16], %o2 1575 brz,pt %l0, 5f 1576 ldx [%i0 + 24], %o3 1577 1578 ldx [%i0 + 32], %o4 1579 sllx %o0, %l0, %o0 1580 srlx %o1, %l1, %g1 1581 or %g1, %o0, %o0 1582 sllx %o1, %l0, %o1 1583 srlx %o2, %l1, %g1 1584 or %g1, %o1, %o1 1585 sllx %o2, %l0, %o2 1586 srlx %o3, %l1, %g1 1587 or %g1, %o2, %o2 1588 sllx %o3, %l0, %o3 1589 srlx %o4, %l1, %o4 1590 or %o4, %o3, %o3 1591 5: 1592 .word 0x99b02302 !movxtod %g2,%f12 1593 .word 0x9db02303 !movxtod %g3,%f14 1594 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1595 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1596 1597 srax %g3, 63, %l7 ! next tweak value 1598 addcc %g2, %g2, %g2 1599 and %l7, 0x87, %l7 1600 .word 0x87b0c223 !addxc %g3,%g3,%g3 1601 xor %l7, %g2, %g2 1602 1603 .word 0x91b02302 !movxtod %g2,%f8 1604 .word 0x95b02303 !movxtod %g3,%f10 1605 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1606 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1607 1608 xor %g4, %o0, %o0 ! ^= rk[0] 1609 xor %g5, %o1, %o1 1610 xor %g4, %o2, %o2 ! ^= rk[0] 1611 xor %g5, %o3, %o3 1612 .word 0x81b02308 !movxtod %o0,%f0 1613 .word 0x85b02309 !movxtod %o1,%f2 1614 .word 0x89b0230a !movxtod %o2,%f4 1615 .word 0x8db0230b !movxtod %o3,%f6 1616 1617 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1618 .word 0x85b38d82 !fxor %f14,%f2,%f2 1619 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 1620 .word 0x8db28d86 !fxor %f10,%f6,%f6 1621 1622 prefetch [%i0 + 32+63], 20 1623 call _aes128_decrypt_2x 1624 add %i0, 32, %i0 1625 1626 .word 0x91b02302 !movxtod %g2,%f8 1627 .word 0x95b02303 !movxtod %g3,%f10 1628 1629 srax %g3, 63, %l7 ! next tweak value 1630 addcc %g2, %g2, %g2 1631 and %l7, 0x87, %l7 1632 .word 0x87b0c223 !addxc %g3,%g3,%g3 1633 xor %l7, %g2, %g2 1634 1635 .word 0x91b20988 !bshuffle %f8,%f8,%f8 1636 .word 0x95b2898a !bshuffle %f10,%f10,%f10 1637 1638 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1639 .word 0x85b38d82 !fxor %f14,%f2,%f2 1640 .word 0x89b20d84 !fxor %f8,%f4,%f4 1641 .word 0x8db28d86 !fxor %f10,%f6,%f6 1642 1643 subcc %i2, 2, %i2 1644 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1645 add %i1, 8, %i1 1646 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1647 add %i1, 8, %i1 1648 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1649 add %i1, 8, %i1 1650 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1651 bgu,pt SIZE_T_CC, .L128_xts_deblk2x 1652 add %i1, 8, %i1 1653 1654 add %l5, %i2, %i2 1655 andcc %i2, 1, %g0 ! is number of blocks even? 1656 membar #StoreLoad|#StoreStore 1657 bnz,pt %icc, .L128_xts_deloop 1658 srl %i2, 0, %i2 1659 brnz,pn %i2, .L128_xts_deloop2x 1660 nop 1661 1662 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 1663 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 1664 brnz,pn %i5, .L128_xts_desteal 1665 nop 1666 1667 ret 1668 restore 1669 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1670 .align 32 1671 .L128_xts_desteal: 1672 ldx [%i0 + 0], %o0 1673 brz,pt %l0, 8f 1674 ldx [%i0 + 8], %o1 1675 1676 ldx [%i0 + 16], %o2 1677 sllx %o0, %l0, %o0 1678 srlx %o1, %l1, %g1 1679 sllx %o1, %l0, %o1 1680 or %g1, %o0, %o0 1681 srlx %o2, %l1, %o2 1682 or %o2, %o1, %o1 1683 8: 1684 srax %g3, 63, %l7 ! next tweak value 1685 addcc %g2, %g2, %o2 1686 and %l7, 0x87, %l7 1687 .word 0x97b0c223 !addxc %g3,%g3,%o3 1688 xor %l7, %o2, %o2 1689 1690 .word 0x99b0230a !movxtod %o2,%f12 1691 .word 0x9db0230b !movxtod %o3,%f14 1692 .word 0x99b3098c !bshuffle %f12,%f12,%f12 1693 .word 0x9db3898e !bshuffle %f14,%f14,%f14 1694 1695 xor %g4, %o0, %o0 ! ^= rk[0] 1696 xor %g5, %o1, %o1 1697 .word 0x81b02308 !movxtod %o0,%f0 1698 .word 0x85b02309 !movxtod %o1,%f2 1699 1700 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1701 .word 0x85b38d82 !fxor %f14,%f2,%f2 1702 1703 call _aes128_decrypt_1x 1704 add %i0, 16, %i0 1705 1706 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 1707 .word 0x85b38d82 !fxor %f14,%f2,%f2 1708 1709 std %f0, [%fp + STACK_BIAS-16] 1710 std %f2, [%fp + STACK_BIAS-8] 1711 1712 srl %l0, 3, %l0 1713 add %fp, STACK_BIAS-16, %l7 1714 add %i0, %l0, %i0 ! original %i0+%i2&-15 1715 add %i1, %l2, %i1 ! original %i1+%i2&-15 1716 mov 0, %l0 1717 add %i1, 16, %i1 1718 nop ! align 1719 1720 .L128_xts_destealing: 1721 ldub [%i0 + %l0], %o0 1722 ldub [%l7 + %l0], %o1 1723 dec %i5 1724 stb %o0, [%l7 + %l0] 1725 stb %o1, [%i1 + %l0] 1726 brnz %i5, .L128_xts_destealing 1727 inc %l0 1728 1729 mov %l7, %i0 1730 sub %i1, 16, %i1 1731 mov 0, %l0 1732 sub %i1, %l2, %i1 1733 ba .L128_xts_deloop ! one more time 1734 mov 1, %i2 ! %i5 is 0 1735 ret 1736 restore 1737 .type aes128_t4_xts_decrypt,#function 1738 .size aes128_t4_xts_decrypt,.-aes128_t4_xts_decrypt 1739 .globl aes128_t4_cbc_decrypt 1740 .align 32 1741 aes128_t4_cbc_decrypt: 1742 save %sp, -STACK_FRAME, %sp 1743 cmp %i2, 0 1744 be,pn SIZE_T_CC, .L128_cbc_dec_abort 1745 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 1746 sub %i0, %i1, %l5 ! %i0!=%i1 1747 ld [%i4 + 0], %f12 ! load ivec 1748 ld [%i4 + 4], %f13 1749 ld [%i4 + 8], %f14 1750 ld [%i4 + 12], %f15 1751 prefetch [%i0], 20 1752 prefetch [%i0 + 63], 20 1753 call _aes128_load_deckey 1754 and %i0, 7, %l0 1755 andn %i0, 7, %i0 1756 sll %l0, 3, %l0 1757 mov 64, %l1 1758 mov 0xff, %l3 1759 sub %l1, %l0, %l1 1760 and %i1, 7, %l2 1761 cmp %i2, 255 1762 movrnz %l2, 0, %l5 ! if ( %i1&7 || 1763 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 1764 brnz,pn %l5, .L128cbc_dec_blk ! %i0==%i1) 1765 srl %l3, %l2, %l3 1766 1767 andcc %i2, 16, %g0 ! is number of blocks even? 1768 srlx %i2, 4, %i2 1769 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 1770 bz %icc, .L128_cbc_dec_loop2x 1771 prefetch [%i1], 22 1772 .L128_cbc_dec_loop: 1773 ldx [%i0 + 0], %o0 1774 brz,pt %l0, 4f 1775 ldx [%i0 + 8], %o1 1776 1777 ldx [%i0 + 16], %o2 1778 sllx %o0, %l0, %o0 1779 srlx %o1, %l1, %g1 1780 sllx %o1, %l0, %o1 1781 or %g1, %o0, %o0 1782 srlx %o2, %l1, %o2 1783 or %o2, %o1, %o1 1784 4: 1785 xor %g4, %o0, %o2 ! ^= rk[0] 1786 xor %g5, %o1, %o3 1787 .word 0x81b0230a !movxtod %o2,%f0 1788 .word 0x85b0230b !movxtod %o3,%f2 1789 1790 prefetch [%i1 + 63], 22 1791 prefetch [%i0 + 16+63], 20 1792 call _aes128_decrypt_1x 1793 add %i0, 16, %i0 1794 1795 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 1796 .word 0x85b38d82 !fxor %f14,%f2,%f2 1797 .word 0x99b02308 !movxtod %o0,%f12 1798 .word 0x9db02309 !movxtod %o1,%f14 1799 1800 brnz,pn %l2, 2f 1801 sub %i2, 1, %i2 1802 1803 std %f0, [%i1 + 0] 1804 std %f2, [%i1 + 8] 1805 brnz,pt %i2, .L128_cbc_dec_loop2x 1806 add %i1, 16, %i1 1807 st %f12, [%i4 + 0] 1808 st %f13, [%i4 + 4] 1809 st %f14, [%i4 + 8] 1810 st %f15, [%i4 + 12] 1811 .L128_cbc_dec_abort: 1812 ret 1813 restore 1814 1815 .align 16 1816 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1817 ! and ~3x deterioration 1818 ! in inp==out case 1819 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 1820 .word 0x8db00902 !faligndata %f0,%f2,%f6 1821 .word 0x91b08902 !faligndata %f2,%f2,%f8 1822 1823 stda %f4, [%i1 + %l3]0xc0 ! partial store 1824 std %f6, [%i1 + 8] 1825 add %i1, 16, %i1 1826 orn %g0, %l3, %l3 1827 stda %f8, [%i1 + %l3]0xc0 ! partial store 1828 1829 brnz,pt %i2, .L128_cbc_dec_loop2x+4 1830 orn %g0, %l3, %l3 1831 st %f12, [%i4 + 0] 1832 st %f13, [%i4 + 4] 1833 st %f14, [%i4 + 8] 1834 st %f15, [%i4 + 12] 1835 ret 1836 restore 1837 1838 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1839 .align 32 1840 .L128_cbc_dec_loop2x: 1841 ldx [%i0 + 0], %o0 1842 ldx [%i0 + 8], %o1 1843 ldx [%i0 + 16], %o2 1844 brz,pt %l0, 4f 1845 ldx [%i0 + 24], %o3 1846 1847 ldx [%i0 + 32], %o4 1848 sllx %o0, %l0, %o0 1849 srlx %o1, %l1, %g1 1850 or %g1, %o0, %o0 1851 sllx %o1, %l0, %o1 1852 srlx %o2, %l1, %g1 1853 or %g1, %o1, %o1 1854 sllx %o2, %l0, %o2 1855 srlx %o3, %l1, %g1 1856 or %g1, %o2, %o2 1857 sllx %o3, %l0, %o3 1858 srlx %o4, %l1, %o4 1859 or %o4, %o3, %o3 1860 4: 1861 xor %g4, %o0, %o4 ! ^= rk[0] 1862 xor %g5, %o1, %o5 1863 .word 0x81b0230c !movxtod %o4,%f0 1864 .word 0x85b0230d !movxtod %o5,%f2 1865 xor %g4, %o2, %o4 1866 xor %g5, %o3, %o5 1867 .word 0x89b0230c !movxtod %o4,%f4 1868 .word 0x8db0230d !movxtod %o5,%f6 1869 1870 prefetch [%i1 + 63], 22 1871 prefetch [%i0 + 32+63], 20 1872 call _aes128_decrypt_2x 1873 add %i0, 32, %i0 1874 1875 .word 0x91b02308 !movxtod %o0,%f8 1876 .word 0x95b02309 !movxtod %o1,%f10 1877 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 1878 .word 0x85b38d82 !fxor %f14,%f2,%f2 1879 .word 0x99b0230a !movxtod %o2,%f12 1880 .word 0x9db0230b !movxtod %o3,%f14 1881 .word 0x89b20d84 !fxor %f8,%f4,%f4 1882 .word 0x8db28d86 !fxor %f10,%f6,%f6 1883 1884 brnz,pn %l2, 2f 1885 sub %i2, 2, %i2 1886 1887 std %f0, [%i1 + 0] 1888 std %f2, [%i1 + 8] 1889 std %f4, [%i1 + 16] 1890 std %f6, [%i1 + 24] 1891 brnz,pt %i2, .L128_cbc_dec_loop2x 1892 add %i1, 32, %i1 1893 st %f12, [%i4 + 0] 1894 st %f13, [%i4 + 4] 1895 st %f14, [%i4 + 8] 1896 st %f15, [%i4 + 12] 1897 ret 1898 restore 1899 1900 .align 16 1901 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 1902 ! and ~3x deterioration 1903 ! in inp==out case 1904 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 1905 .word 0x81b00902 !faligndata %f0,%f2,%f0 1906 .word 0x85b08904 !faligndata %f2,%f4,%f2 1907 .word 0x89b10906 !faligndata %f4,%f6,%f4 1908 .word 0x8db18906 !faligndata %f6,%f6,%f6 1909 stda %f8, [%i1 + %l3]0xc0 ! partial store 1910 std %f0, [%i1 + 8] 1911 std %f2, [%i1 + 16] 1912 std %f4, [%i1 + 24] 1913 add %i1, 32, %i1 1914 orn %g0, %l3, %l3 1915 stda %f6, [%i1 + %l3]0xc0 ! partial store 1916 1917 brnz,pt %i2, .L128_cbc_dec_loop2x+4 1918 orn %g0, %l3, %l3 1919 st %f12, [%i4 + 0] 1920 st %f13, [%i4 + 4] 1921 st %f14, [%i4 + 8] 1922 st %f15, [%i4 + 12] 1923 ret 1924 restore 1925 1926 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1927 .align 32 1928 .L128cbc_dec_blk: 1929 add %i1, %i2, %l5 1930 and %l5, 63, %l5 ! tail 1931 sub %i2, %l5, %i2 1932 add %l5, 15, %l5 ! round up to 16n 1933 srlx %i2, 4, %i2 1934 srl %l5, 4, %l5 1935 sub %i2, 1, %i2 1936 add %l5, 1, %l5 1937 1938 .L128_cbc_dec_blk_loop2x: 1939 ldx [%i0 + 0], %o0 1940 ldx [%i0 + 8], %o1 1941 ldx [%i0 + 16], %o2 1942 brz,pt %l0, 5f 1943 ldx [%i0 + 24], %o3 1944 1945 ldx [%i0 + 32], %o4 1946 sllx %o0, %l0, %o0 1947 srlx %o1, %l1, %g1 1948 or %g1, %o0, %o0 1949 sllx %o1, %l0, %o1 1950 srlx %o2, %l1, %g1 1951 or %g1, %o1, %o1 1952 sllx %o2, %l0, %o2 1953 srlx %o3, %l1, %g1 1954 or %g1, %o2, %o2 1955 sllx %o3, %l0, %o3 1956 srlx %o4, %l1, %o4 1957 or %o4, %o3, %o3 1958 5: 1959 xor %g4, %o0, %o4 ! ^= rk[0] 1960 xor %g5, %o1, %o5 1961 .word 0x81b0230c !movxtod %o4,%f0 1962 .word 0x85b0230d !movxtod %o5,%f2 1963 xor %g4, %o2, %o4 1964 xor %g5, %o3, %o5 1965 .word 0x89b0230c !movxtod %o4,%f4 1966 .word 0x8db0230d !movxtod %o5,%f6 1967 1968 prefetch [%i0 + 32+63], 20 1969 call _aes128_decrypt_2x 1970 add %i0, 32, %i0 1971 subcc %i2, 2, %i2 1972 1973 .word 0x91b02308 !movxtod %o0,%f8 1974 .word 0x95b02309 !movxtod %o1,%f10 1975 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 1976 .word 0x85b38d82 !fxor %f14,%f2,%f2 1977 .word 0x99b0230a !movxtod %o2,%f12 1978 .word 0x9db0230b !movxtod %o3,%f14 1979 .word 0x89b20d84 !fxor %f8,%f4,%f4 1980 .word 0x8db28d86 !fxor %f10,%f6,%f6 1981 1982 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1983 add %i1, 8, %i1 1984 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1985 add %i1, 8, %i1 1986 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1987 add %i1, 8, %i1 1988 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 1989 bgu,pt SIZE_T_CC, .L128_cbc_dec_blk_loop2x 1990 add %i1, 8, %i1 1991 1992 add %l5, %i2, %i2 1993 andcc %i2, 1, %g0 ! is number of blocks even? 1994 membar #StoreLoad|#StoreStore 1995 bnz,pt %icc, .L128_cbc_dec_loop 1996 srl %i2, 0, %i2 1997 brnz,pn %i2, .L128_cbc_dec_loop2x 1998 nop 1999 st %f12, [%i4 + 0] ! write out ivec 2000 st %f13, [%i4 + 4] 2001 st %f14, [%i4 + 8] 2002 st %f15, [%i4 + 12] 2003 ret 2004 restore 2005 .type aes128_t4_cbc_decrypt,#function 2006 .size aes128_t4_cbc_decrypt,.-aes128_t4_cbc_decrypt 2007 .align 32 2008 _aes128_decrypt_1x: 2009 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 2010 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 2011 .word 0x80cd0444 !aes_dround01 %f20,%f4,%f2,%f0 2012 .word 0x84cd8464 !aes_dround23 %f22,%f4,%f2,%f2 2013 .word 0x88ce0440 !aes_dround01 %f24,%f0,%f2,%f4 2014 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 2015 .word 0x80cf0444 !aes_dround01 %f28,%f4,%f2,%f0 2016 .word 0x84cf8464 !aes_dround23 %f30,%f4,%f2,%f2 2017 .word 0x88c84440 !aes_dround01 %f32,%f0,%f2,%f4 2018 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 2019 .word 0x80c94444 !aes_dround01 %f36,%f4,%f2,%f0 2020 .word 0x84c9c464 !aes_dround23 %f38,%f4,%f2,%f2 2021 .word 0x88ca4440 !aes_dround01 %f40,%f0,%f2,%f4 2022 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 2023 .word 0x80cb4444 !aes_dround01 %f44,%f4,%f2,%f0 2024 .word 0x84cbc464 !aes_dround23 %f46,%f4,%f2,%f2 2025 .word 0x88cc4440 !aes_dround01 %f48,%f0,%f2,%f4 2026 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 2027 .word 0x80cd44c4 !aes_dround01_l %f52,%f4,%f2,%f0 2028 retl 2029 .word 0x84cdc4e4 !aes_dround23_l %f54,%f4,%f2,%f2 2030 .type _aes128_decrypt_1x,#function 2031 .size _aes128_decrypt_1x,.-_aes128_decrypt_1x 2032 2033 .align 32 2034 _aes128_decrypt_2x: 2035 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 2036 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 2037 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 2038 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 2039 .word 0x80cd0448 !aes_dround01 %f20,%f8,%f2,%f0 2040 .word 0x84cd8468 !aes_dround23 %f22,%f8,%f2,%f2 2041 .word 0x88cd0c4a !aes_dround01 %f20,%f10,%f6,%f4 2042 .word 0x8ccd8c6a !aes_dround23 %f22,%f10,%f6,%f6 2043 .word 0x90ce0440 !aes_dround01 %f24,%f0,%f2,%f8 2044 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 2045 .word 0x94ce0c44 !aes_dround01 %f24,%f4,%f6,%f10 2046 .word 0x8cce8c64 !aes_dround23 %f26,%f4,%f6,%f6 2047 .word 0x80cf0448 !aes_dround01 %f28,%f8,%f2,%f0 2048 .word 0x84cf8468 !aes_dround23 %f30,%f8,%f2,%f2 2049 .word 0x88cf0c4a !aes_dround01 %f28,%f10,%f6,%f4 2050 .word 0x8ccf8c6a !aes_dround23 %f30,%f10,%f6,%f6 2051 .word 0x90c84440 !aes_dround01 %f32,%f0,%f2,%f8 2052 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 2053 .word 0x94c84c44 !aes_dround01 %f32,%f4,%f6,%f10 2054 .word 0x8cc8cc64 !aes_dround23 %f34,%f4,%f6,%f6 2055 .word 0x80c94448 !aes_dround01 %f36,%f8,%f2,%f0 2056 .word 0x84c9c468 !aes_dround23 %f38,%f8,%f2,%f2 2057 .word 0x88c94c4a !aes_dround01 %f36,%f10,%f6,%f4 2058 .word 0x8cc9cc6a !aes_dround23 %f38,%f10,%f6,%f6 2059 .word 0x90ca4440 !aes_dround01 %f40,%f0,%f2,%f8 2060 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 2061 .word 0x94ca4c44 !aes_dround01 %f40,%f4,%f6,%f10 2062 .word 0x8ccacc64 !aes_dround23 %f42,%f4,%f6,%f6 2063 .word 0x80cb4448 !aes_dround01 %f44,%f8,%f2,%f0 2064 .word 0x84cbc468 !aes_dround23 %f46,%f8,%f2,%f2 2065 .word 0x88cb4c4a !aes_dround01 %f44,%f10,%f6,%f4 2066 .word 0x8ccbcc6a !aes_dround23 %f46,%f10,%f6,%f6 2067 .word 0x90cc4440 !aes_dround01 %f48,%f0,%f2,%f8 2068 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 2069 .word 0x94cc4c44 !aes_dround01 %f48,%f4,%f6,%f10 2070 .word 0x8ccccc64 !aes_dround23 %f50,%f4,%f6,%f6 2071 .word 0x80cd44c8 !aes_dround01_l %f52,%f8,%f2,%f0 2072 .word 0x84cdc4e8 !aes_dround23_l %f54,%f8,%f2,%f2 2073 .word 0x88cd4cca !aes_dround01_l %f52,%f10,%f6,%f4 2074 retl 2075 .word 0x8ccdccea !aes_dround23_l %f54,%f10,%f6,%f6 2076 .type _aes128_decrypt_2x,#function 2077 .size _aes128_decrypt_2x,.-_aes128_decrypt_2x 2078 .align 32 2079 _aes192_encrypt_1x: 2080 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 2081 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2082 .word 0x80cd0404 !aes_eround01 %f20,%f4,%f2,%f0 2083 .word 0x84cd8424 !aes_eround23 %f22,%f4,%f2,%f2 2084 .word 0x88ce0400 !aes_eround01 %f24,%f0,%f2,%f4 2085 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2086 .word 0x80cf0404 !aes_eround01 %f28,%f4,%f2,%f0 2087 .word 0x84cf8424 !aes_eround23 %f30,%f4,%f2,%f2 2088 .word 0x88c84400 !aes_eround01 %f32,%f0,%f2,%f4 2089 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2090 .word 0x80c94404 !aes_eround01 %f36,%f4,%f2,%f0 2091 .word 0x84c9c424 !aes_eround23 %f38,%f4,%f2,%f2 2092 .word 0x88ca4400 !aes_eround01 %f40,%f0,%f2,%f4 2093 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2094 .word 0x80cb4404 !aes_eround01 %f44,%f4,%f2,%f0 2095 .word 0x84cbc424 !aes_eround23 %f46,%f4,%f2,%f2 2096 .word 0x88cc4400 !aes_eround01 %f48,%f0,%f2,%f4 2097 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2098 .word 0x80cd4404 !aes_eround01 %f52,%f4,%f2,%f0 2099 .word 0x84cdc424 !aes_eround23 %f54,%f4,%f2,%f2 2100 .word 0x88ce4400 !aes_eround01 %f56,%f0,%f2,%f4 2101 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2102 .word 0x80cf4484 !aes_eround01_l %f60,%f4,%f2,%f0 2103 retl 2104 .word 0x84cfc4a4 !aes_eround23_l %f62,%f4,%f2,%f2 2105 .type _aes192_encrypt_1x,#function 2106 .size _aes192_encrypt_1x,.-_aes192_encrypt_1x 2107 2108 .align 32 2109 _aes192_encrypt_2x: 2110 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 2111 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2112 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 2113 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 2114 .word 0x80cd0408 !aes_eround01 %f20,%f8,%f2,%f0 2115 .word 0x84cd8428 !aes_eround23 %f22,%f8,%f2,%f2 2116 .word 0x88cd0c0a !aes_eround01 %f20,%f10,%f6,%f4 2117 .word 0x8ccd8c2a !aes_eround23 %f22,%f10,%f6,%f6 2118 .word 0x90ce0400 !aes_eround01 %f24,%f0,%f2,%f8 2119 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2120 .word 0x94ce0c04 !aes_eround01 %f24,%f4,%f6,%f10 2121 .word 0x8cce8c24 !aes_eround23 %f26,%f4,%f6,%f6 2122 .word 0x80cf0408 !aes_eround01 %f28,%f8,%f2,%f0 2123 .word 0x84cf8428 !aes_eround23 %f30,%f8,%f2,%f2 2124 .word 0x88cf0c0a !aes_eround01 %f28,%f10,%f6,%f4 2125 .word 0x8ccf8c2a !aes_eround23 %f30,%f10,%f6,%f6 2126 .word 0x90c84400 !aes_eround01 %f32,%f0,%f2,%f8 2127 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2128 .word 0x94c84c04 !aes_eround01 %f32,%f4,%f6,%f10 2129 .word 0x8cc8cc24 !aes_eround23 %f34,%f4,%f6,%f6 2130 .word 0x80c94408 !aes_eround01 %f36,%f8,%f2,%f0 2131 .word 0x84c9c428 !aes_eround23 %f38,%f8,%f2,%f2 2132 .word 0x88c94c0a !aes_eround01 %f36,%f10,%f6,%f4 2133 .word 0x8cc9cc2a !aes_eround23 %f38,%f10,%f6,%f6 2134 .word 0x90ca4400 !aes_eround01 %f40,%f0,%f2,%f8 2135 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2136 .word 0x94ca4c04 !aes_eround01 %f40,%f4,%f6,%f10 2137 .word 0x8ccacc24 !aes_eround23 %f42,%f4,%f6,%f6 2138 .word 0x80cb4408 !aes_eround01 %f44,%f8,%f2,%f0 2139 .word 0x84cbc428 !aes_eround23 %f46,%f8,%f2,%f2 2140 .word 0x88cb4c0a !aes_eround01 %f44,%f10,%f6,%f4 2141 .word 0x8ccbcc2a !aes_eround23 %f46,%f10,%f6,%f6 2142 .word 0x90cc4400 !aes_eround01 %f48,%f0,%f2,%f8 2143 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2144 .word 0x94cc4c04 !aes_eround01 %f48,%f4,%f6,%f10 2145 .word 0x8ccccc24 !aes_eround23 %f50,%f4,%f6,%f6 2146 .word 0x80cd4408 !aes_eround01 %f52,%f8,%f2,%f0 2147 .word 0x84cdc428 !aes_eround23 %f54,%f8,%f2,%f2 2148 .word 0x88cd4c0a !aes_eround01 %f52,%f10,%f6,%f4 2149 .word 0x8ccdcc2a !aes_eround23 %f54,%f10,%f6,%f6 2150 .word 0x90ce4400 !aes_eround01 %f56,%f0,%f2,%f8 2151 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2152 .word 0x94ce4c04 !aes_eround01 %f56,%f4,%f6,%f10 2153 .word 0x8ccecc24 !aes_eround23 %f58,%f4,%f6,%f6 2154 .word 0x80cf4488 !aes_eround01_l %f60,%f8,%f2,%f0 2155 .word 0x84cfc4a8 !aes_eround23_l %f62,%f8,%f2,%f2 2156 .word 0x88cf4c8a !aes_eround01_l %f60,%f10,%f6,%f4 2157 retl 2158 .word 0x8ccfccaa !aes_eround23_l %f62,%f10,%f6,%f6 2159 .type _aes192_encrypt_2x,#function 2160 .size _aes192_encrypt_2x,.-_aes192_encrypt_2x 2161 2162 .align 32 2163 _aes256_encrypt_1x: 2164 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 2165 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2166 ldd [%i3 + 208], %f16 2167 ldd [%i3 + 216], %f18 2168 .word 0x80cd0404 !aes_eround01 %f20,%f4,%f2,%f0 2169 .word 0x84cd8424 !aes_eround23 %f22,%f4,%f2,%f2 2170 ldd [%i3 + 224], %f20 2171 ldd [%i3 + 232], %f22 2172 .word 0x88ce0400 !aes_eround01 %f24,%f0,%f2,%f4 2173 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2174 .word 0x80cf0404 !aes_eround01 %f28,%f4,%f2,%f0 2175 .word 0x84cf8424 !aes_eround23 %f30,%f4,%f2,%f2 2176 .word 0x88c84400 !aes_eround01 %f32,%f0,%f2,%f4 2177 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2178 .word 0x80c94404 !aes_eround01 %f36,%f4,%f2,%f0 2179 .word 0x84c9c424 !aes_eround23 %f38,%f4,%f2,%f2 2180 .word 0x88ca4400 !aes_eround01 %f40,%f0,%f2,%f4 2181 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2182 .word 0x80cb4404 !aes_eround01 %f44,%f4,%f2,%f0 2183 .word 0x84cbc424 !aes_eround23 %f46,%f4,%f2,%f2 2184 .word 0x88cc4400 !aes_eround01 %f48,%f0,%f2,%f4 2185 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2186 .word 0x80cd4404 !aes_eround01 %f52,%f4,%f2,%f0 2187 .word 0x84cdc424 !aes_eround23 %f54,%f4,%f2,%f2 2188 .word 0x88ce4400 !aes_eround01 %f56,%f0,%f2,%f4 2189 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2190 .word 0x80cf4404 !aes_eround01 %f60,%f4,%f2,%f0 2191 .word 0x84cfc424 !aes_eround23 %f62,%f4,%f2,%f2 2192 .word 0x88cc0400 !aes_eround01 %f16,%f0,%f2,%f4 2193 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2194 ldd [%i3 + 16], %f16 2195 ldd [%i3 + 24], %f18 2196 .word 0x80cd0484 !aes_eround01_l %f20,%f4,%f2,%f0 2197 .word 0x84cd84a4 !aes_eround23_l %f22,%f4,%f2,%f2 2198 ldd [%i3 + 32], %f20 2199 retl 2200 ldd [%i3 + 40], %f22 2201 .type _aes256_encrypt_1x,#function 2202 .size _aes256_encrypt_1x,.-_aes256_encrypt_1x 2203 2204 .align 32 2205 _aes256_encrypt_2x: 2206 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 2207 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2208 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 2209 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 2210 ldd [%i3 + 208], %f16 2211 ldd [%i3 + 216], %f18 2212 .word 0x80cd0408 !aes_eround01 %f20,%f8,%f2,%f0 2213 .word 0x84cd8428 !aes_eround23 %f22,%f8,%f2,%f2 2214 .word 0x88cd0c0a !aes_eround01 %f20,%f10,%f6,%f4 2215 .word 0x8ccd8c2a !aes_eround23 %f22,%f10,%f6,%f6 2216 ldd [%i3 + 224], %f20 2217 ldd [%i3 + 232], %f22 2218 .word 0x90ce0400 !aes_eround01 %f24,%f0,%f2,%f8 2219 .word 0x84ce8420 !aes_eround23 %f26,%f0,%f2,%f2 2220 .word 0x94ce0c04 !aes_eround01 %f24,%f4,%f6,%f10 2221 .word 0x8cce8c24 !aes_eround23 %f26,%f4,%f6,%f6 2222 .word 0x80cf0408 !aes_eround01 %f28,%f8,%f2,%f0 2223 .word 0x84cf8428 !aes_eround23 %f30,%f8,%f2,%f2 2224 .word 0x88cf0c0a !aes_eround01 %f28,%f10,%f6,%f4 2225 .word 0x8ccf8c2a !aes_eround23 %f30,%f10,%f6,%f6 2226 .word 0x90c84400 !aes_eround01 %f32,%f0,%f2,%f8 2227 .word 0x84c8c420 !aes_eround23 %f34,%f0,%f2,%f2 2228 .word 0x94c84c04 !aes_eround01 %f32,%f4,%f6,%f10 2229 .word 0x8cc8cc24 !aes_eround23 %f34,%f4,%f6,%f6 2230 .word 0x80c94408 !aes_eround01 %f36,%f8,%f2,%f0 2231 .word 0x84c9c428 !aes_eround23 %f38,%f8,%f2,%f2 2232 .word 0x88c94c0a !aes_eround01 %f36,%f10,%f6,%f4 2233 .word 0x8cc9cc2a !aes_eround23 %f38,%f10,%f6,%f6 2234 .word 0x90ca4400 !aes_eround01 %f40,%f0,%f2,%f8 2235 .word 0x84cac420 !aes_eround23 %f42,%f0,%f2,%f2 2236 .word 0x94ca4c04 !aes_eround01 %f40,%f4,%f6,%f10 2237 .word 0x8ccacc24 !aes_eround23 %f42,%f4,%f6,%f6 2238 .word 0x80cb4408 !aes_eround01 %f44,%f8,%f2,%f0 2239 .word 0x84cbc428 !aes_eround23 %f46,%f8,%f2,%f2 2240 .word 0x88cb4c0a !aes_eround01 %f44,%f10,%f6,%f4 2241 .word 0x8ccbcc2a !aes_eround23 %f46,%f10,%f6,%f6 2242 .word 0x90cc4400 !aes_eround01 %f48,%f0,%f2,%f8 2243 .word 0x84ccc420 !aes_eround23 %f50,%f0,%f2,%f2 2244 .word 0x94cc4c04 !aes_eround01 %f48,%f4,%f6,%f10 2245 .word 0x8ccccc24 !aes_eround23 %f50,%f4,%f6,%f6 2246 .word 0x80cd4408 !aes_eround01 %f52,%f8,%f2,%f0 2247 .word 0x84cdc428 !aes_eround23 %f54,%f8,%f2,%f2 2248 .word 0x88cd4c0a !aes_eround01 %f52,%f10,%f6,%f4 2249 .word 0x8ccdcc2a !aes_eround23 %f54,%f10,%f6,%f6 2250 .word 0x90ce4400 !aes_eround01 %f56,%f0,%f2,%f8 2251 .word 0x84cec420 !aes_eround23 %f58,%f0,%f2,%f2 2252 .word 0x94ce4c04 !aes_eround01 %f56,%f4,%f6,%f10 2253 .word 0x8ccecc24 !aes_eround23 %f58,%f4,%f6,%f6 2254 .word 0x80cf4408 !aes_eround01 %f60,%f8,%f2,%f0 2255 .word 0x84cfc428 !aes_eround23 %f62,%f8,%f2,%f2 2256 .word 0x88cf4c0a !aes_eround01 %f60,%f10,%f6,%f4 2257 .word 0x8ccfcc2a !aes_eround23 %f62,%f10,%f6,%f6 2258 .word 0x90cc0400 !aes_eround01 %f16,%f0,%f2,%f8 2259 .word 0x84cc8420 !aes_eround23 %f18,%f0,%f2,%f2 2260 .word 0x94cc0c04 !aes_eround01 %f16,%f4,%f6,%f10 2261 .word 0x8ccc8c24 !aes_eround23 %f18,%f4,%f6,%f6 2262 ldd [%i3 + 16], %f16 2263 ldd [%i3 + 24], %f18 2264 .word 0x80cd0488 !aes_eround01_l %f20,%f8,%f2,%f0 2265 .word 0x84cd84a8 !aes_eround23_l %f22,%f8,%f2,%f2 2266 .word 0x88cd0c8a !aes_eround01_l %f20,%f10,%f6,%f4 2267 .word 0x8ccd8caa !aes_eround23_l %f22,%f10,%f6,%f6 2268 ldd [%i3 + 32], %f20 2269 retl 2270 ldd [%i3 + 40], %f22 2271 .type _aes256_encrypt_2x,#function 2272 .size _aes256_encrypt_2x,.-_aes256_encrypt_2x 2273 2274 .align 32 2275 _aes192_loadkey: 2276 ldx [%i3 + 0], %g4 2277 ldx [%i3 + 8], %g5 2278 ldd [%i3 + 16], %f16 2279 ldd [%i3 + 24], %f18 2280 ldd [%i3 + 32], %f20 2281 ldd [%i3 + 40], %f22 2282 ldd [%i3 + 48], %f24 2283 ldd [%i3 + 56], %f26 2284 ldd [%i3 + 64], %f28 2285 ldd [%i3 + 72], %f30 2286 ldd [%i3 + 80], %f32 2287 ldd [%i3 + 88], %f34 2288 ldd [%i3 + 96], %f36 2289 ldd [%i3 + 104], %f38 2290 ldd [%i3 + 112], %f40 2291 ldd [%i3 + 120], %f42 2292 ldd [%i3 + 128], %f44 2293 ldd [%i3 + 136], %f46 2294 ldd [%i3 + 144], %f48 2295 ldd [%i3 + 152], %f50 2296 ldd [%i3 + 160], %f52 2297 ldd [%i3 + 168], %f54 2298 ldd [%i3 + 176], %f56 2299 ldd [%i3 + 184], %f58 2300 ldd [%i3 + 192], %f60 2301 ldd [%i3 + 200], %f62 2302 retl 2303 nop 2304 .type _aes192_loadkey,#function 2305 .size _aes192_loadkey,.-_aes192_loadkey 2306 _aes256_loadkey=_aes192_loadkey 2307 _aes192_load_enckey=_aes192_loadkey 2308 _aes192_load_deckey=_aes192_loadkey 2309 _aes256_load_enckey=_aes192_loadkey 2310 _aes256_load_deckey=_aes192_loadkey 2311 .globl aes256_t4_cbc_encrypt 2312 .align 32 2313 aes256_t4_cbc_encrypt: 2314 save %sp, -STACK_FRAME, %sp 2315 cmp %i2, 0 2316 be,pn SIZE_T_CC, .L256_cbc_enc_abort 2317 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2318 sub %i0, %i1, %l5 ! %i0!=%i1 2319 ld [%i4 + 0], %f0 2320 ld [%i4 + 4], %f1 2321 ld [%i4 + 8], %f2 2322 ld [%i4 + 12], %f3 2323 prefetch [%i0], 20 2324 prefetch [%i0 + 63], 20 2325 call _aes256_load_enckey 2326 and %i0, 7, %l0 2327 andn %i0, 7, %i0 2328 sll %l0, 3, %l0 2329 mov 64, %l1 2330 mov 0xff, %l3 2331 sub %l1, %l0, %l1 2332 and %i1, 7, %l2 2333 cmp %i2, 127 2334 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2335 movleu SIZE_T_CC, 0, %l5 ! %i2<128 || 2336 brnz,pn %l5, .L256cbc_enc_blk ! %i0==%i1) 2337 srl %l3, %l2, %l3 2338 2339 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2340 srlx %i2, 4, %i2 2341 prefetch [%i1], 22 2342 2343 .L256_cbc_enc_loop: 2344 ldx [%i0 + 0], %o0 2345 brz,pt %l0, 4f 2346 ldx [%i0 + 8], %o1 2347 2348 ldx [%i0 + 16], %o2 2349 sllx %o0, %l0, %o0 2350 srlx %o1, %l1, %g1 2351 sllx %o1, %l0, %o1 2352 or %g1, %o0, %o0 2353 srlx %o2, %l1, %o2 2354 or %o2, %o1, %o1 2355 4: 2356 xor %g4, %o0, %o0 ! ^= rk[0] 2357 xor %g5, %o1, %o1 2358 .word 0x99b02308 !movxtod %o0,%f12 2359 .word 0x9db02309 !movxtod %o1,%f14 2360 2361 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2362 .word 0x85b38d82 !fxor %f14,%f2,%f2 2363 prefetch [%i1 + 63], 22 2364 prefetch [%i0 + 16+63], 20 2365 call _aes256_encrypt_1x 2366 add %i0, 16, %i0 2367 2368 brnz,pn %l2, 2f 2369 sub %i2, 1, %i2 2370 2371 std %f0, [%i1 + 0] 2372 std %f2, [%i1 + 8] 2373 brnz,pt %i2, .L256_cbc_enc_loop 2374 add %i1, 16, %i1 2375 st %f0, [%i4 + 0] 2376 st %f1, [%i4 + 4] 2377 st %f2, [%i4 + 8] 2378 st %f3, [%i4 + 12] 2379 .L256_cbc_enc_abort: 2380 ret 2381 restore 2382 2383 .align 16 2384 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2385 ! and ~3x deterioration 2386 ! in inp==out case 2387 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2388 .word 0x8db00902 !faligndata %f0,%f2,%f6 2389 .word 0x91b08902 !faligndata %f2,%f2,%f8 2390 2391 stda %f4, [%i1 + %l3]0xc0 ! partial store 2392 std %f6, [%i1 + 8] 2393 add %i1, 16, %i1 2394 orn %g0, %l3, %l3 2395 stda %f8, [%i1 + %l3]0xc0 ! partial store 2396 2397 brnz,pt %i2, .L256_cbc_enc_loop+4 2398 orn %g0, %l3, %l3 2399 st %f0, [%i4 + 0] 2400 st %f1, [%i4 + 4] 2401 st %f2, [%i4 + 8] 2402 st %f3, [%i4 + 12] 2403 ret 2404 restore 2405 2406 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2407 .align 32 2408 .L256cbc_enc_blk: 2409 add %i1, %i2, %l5 2410 and %l5, 63, %l5 ! tail 2411 sub %i2, %l5, %i2 2412 add %l5, 15, %l5 ! round up to 16n 2413 srlx %i2, 4, %i2 2414 srl %l5, 4, %l5 2415 2416 .L256_cbc_enc_blk_loop: 2417 ldx [%i0 + 0], %o0 2418 brz,pt %l0, 5f 2419 ldx [%i0 + 8], %o1 2420 2421 ldx [%i0 + 16], %o2 2422 sllx %o0, %l0, %o0 2423 srlx %o1, %l1, %g1 2424 sllx %o1, %l0, %o1 2425 or %g1, %o0, %o0 2426 srlx %o2, %l1, %o2 2427 or %o2, %o1, %o1 2428 5: 2429 xor %g4, %o0, %o0 ! ^= rk[0] 2430 xor %g5, %o1, %o1 2431 .word 0x99b02308 !movxtod %o0,%f12 2432 .word 0x9db02309 !movxtod %o1,%f14 2433 2434 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2435 .word 0x85b38d82 !fxor %f14,%f2,%f2 2436 prefetch [%i0 + 16+63], 20 2437 call _aes256_encrypt_1x 2438 add %i0, 16, %i0 2439 sub %i2, 1, %i2 2440 2441 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2442 add %i1, 8, %i1 2443 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2444 brnz,pt %i2, .L256_cbc_enc_blk_loop 2445 add %i1, 8, %i1 2446 2447 membar #StoreLoad|#StoreStore 2448 brnz,pt %l5, .L256_cbc_enc_loop 2449 mov %l5, %i2 2450 st %f0, [%i4 + 0] 2451 st %f1, [%i4 + 4] 2452 st %f2, [%i4 + 8] 2453 st %f3, [%i4 + 12] 2454 ret 2455 restore 2456 .type aes256_t4_cbc_encrypt,#function 2457 .size aes256_t4_cbc_encrypt,.-aes256_t4_cbc_encrypt 2458 .globl aes192_t4_cbc_encrypt 2459 .align 32 2460 aes192_t4_cbc_encrypt: 2461 save %sp, -STACK_FRAME, %sp 2462 cmp %i2, 0 2463 be,pn SIZE_T_CC, .L192_cbc_enc_abort 2464 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2465 sub %i0, %i1, %l5 ! %i0!=%i1 2466 ld [%i4 + 0], %f0 2467 ld [%i4 + 4], %f1 2468 ld [%i4 + 8], %f2 2469 ld [%i4 + 12], %f3 2470 prefetch [%i0], 20 2471 prefetch [%i0 + 63], 20 2472 call _aes192_load_enckey 2473 and %i0, 7, %l0 2474 andn %i0, 7, %i0 2475 sll %l0, 3, %l0 2476 mov 64, %l1 2477 mov 0xff, %l3 2478 sub %l1, %l0, %l1 2479 and %i1, 7, %l2 2480 cmp %i2, 127 2481 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2482 movleu SIZE_T_CC, 0, %l5 ! %i2<128 || 2483 brnz,pn %l5, .L192cbc_enc_blk ! %i0==%i1) 2484 srl %l3, %l2, %l3 2485 2486 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2487 srlx %i2, 4, %i2 2488 prefetch [%i1], 22 2489 2490 .L192_cbc_enc_loop: 2491 ldx [%i0 + 0], %o0 2492 brz,pt %l0, 4f 2493 ldx [%i0 + 8], %o1 2494 2495 ldx [%i0 + 16], %o2 2496 sllx %o0, %l0, %o0 2497 srlx %o1, %l1, %g1 2498 sllx %o1, %l0, %o1 2499 or %g1, %o0, %o0 2500 srlx %o2, %l1, %o2 2501 or %o2, %o1, %o1 2502 4: 2503 xor %g4, %o0, %o0 ! ^= rk[0] 2504 xor %g5, %o1, %o1 2505 .word 0x99b02308 !movxtod %o0,%f12 2506 .word 0x9db02309 !movxtod %o1,%f14 2507 2508 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2509 .word 0x85b38d82 !fxor %f14,%f2,%f2 2510 prefetch [%i1 + 63], 22 2511 prefetch [%i0 + 16+63], 20 2512 call _aes192_encrypt_1x 2513 add %i0, 16, %i0 2514 2515 brnz,pn %l2, 2f 2516 sub %i2, 1, %i2 2517 2518 std %f0, [%i1 + 0] 2519 std %f2, [%i1 + 8] 2520 brnz,pt %i2, .L192_cbc_enc_loop 2521 add %i1, 16, %i1 2522 st %f0, [%i4 + 0] 2523 st %f1, [%i4 + 4] 2524 st %f2, [%i4 + 8] 2525 st %f3, [%i4 + 12] 2526 .L192_cbc_enc_abort: 2527 ret 2528 restore 2529 2530 .align 16 2531 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2532 ! and ~3x deterioration 2533 ! in inp==out case 2534 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2535 .word 0x8db00902 !faligndata %f0,%f2,%f6 2536 .word 0x91b08902 !faligndata %f2,%f2,%f8 2537 2538 stda %f4, [%i1 + %l3]0xc0 ! partial store 2539 std %f6, [%i1 + 8] 2540 add %i1, 16, %i1 2541 orn %g0, %l3, %l3 2542 stda %f8, [%i1 + %l3]0xc0 ! partial store 2543 2544 brnz,pt %i2, .L192_cbc_enc_loop+4 2545 orn %g0, %l3, %l3 2546 st %f0, [%i4 + 0] 2547 st %f1, [%i4 + 4] 2548 st %f2, [%i4 + 8] 2549 st %f3, [%i4 + 12] 2550 ret 2551 restore 2552 2553 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2554 .align 32 2555 .L192cbc_enc_blk: 2556 add %i1, %i2, %l5 2557 and %l5, 63, %l5 ! tail 2558 sub %i2, %l5, %i2 2559 add %l5, 15, %l5 ! round up to 16n 2560 srlx %i2, 4, %i2 2561 srl %l5, 4, %l5 2562 2563 .L192_cbc_enc_blk_loop: 2564 ldx [%i0 + 0], %o0 2565 brz,pt %l0, 5f 2566 ldx [%i0 + 8], %o1 2567 2568 ldx [%i0 + 16], %o2 2569 sllx %o0, %l0, %o0 2570 srlx %o1, %l1, %g1 2571 sllx %o1, %l0, %o1 2572 or %g1, %o0, %o0 2573 srlx %o2, %l1, %o2 2574 or %o2, %o1, %o1 2575 5: 2576 xor %g4, %o0, %o0 ! ^= rk[0] 2577 xor %g5, %o1, %o1 2578 .word 0x99b02308 !movxtod %o0,%f12 2579 .word 0x9db02309 !movxtod %o1,%f14 2580 2581 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 2582 .word 0x85b38d82 !fxor %f14,%f2,%f2 2583 prefetch [%i0 + 16+63], 20 2584 call _aes192_encrypt_1x 2585 add %i0, 16, %i0 2586 sub %i2, 1, %i2 2587 2588 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2589 add %i1, 8, %i1 2590 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2591 brnz,pt %i2, .L192_cbc_enc_blk_loop 2592 add %i1, 8, %i1 2593 2594 membar #StoreLoad|#StoreStore 2595 brnz,pt %l5, .L192_cbc_enc_loop 2596 mov %l5, %i2 2597 st %f0, [%i4 + 0] 2598 st %f1, [%i4 + 4] 2599 st %f2, [%i4 + 8] 2600 st %f3, [%i4 + 12] 2601 ret 2602 restore 2603 .type aes192_t4_cbc_encrypt,#function 2604 .size aes192_t4_cbc_encrypt,.-aes192_t4_cbc_encrypt 2605 .globl aes256_t4_ctr32_encrypt 2606 .align 32 2607 aes256_t4_ctr32_encrypt: 2608 save %sp, -STACK_FRAME, %sp 2609 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2610 2611 prefetch [%i0], 20 2612 prefetch [%i0 + 63], 20 2613 call _aes256_load_enckey 2614 sllx %i2, 4, %i2 2615 2616 ld [%i4 + 0], %l4 ! counter 2617 ld [%i4 + 4], %l5 2618 ld [%i4 + 8], %l6 2619 ld [%i4 + 12], %l7 2620 2621 sllx %l4, 32, %o5 2622 or %l5, %o5, %o5 2623 sllx %l6, 32, %g1 2624 xor %o5, %g4, %g4 ! ^= rk[0] 2625 xor %g1, %g5, %g5 2626 .word 0x9db02304 !movxtod %g4,%f14 ! most significant 64 bits 2627 2628 sub %i0, %i1, %l5 ! %i0!=%i1 2629 and %i0, 7, %l0 2630 andn %i0, 7, %i0 2631 sll %l0, 3, %l0 2632 mov 64, %l1 2633 mov 0xff, %l3 2634 sub %l1, %l0, %l1 2635 and %i1, 7, %l2 2636 cmp %i2, 255 2637 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2638 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 2639 brnz,pn %l5, .L256_ctr32_blk ! %i0==%i1) 2640 srl %l3, %l2, %l3 2641 2642 andcc %i2, 16, %g0 ! is number of blocks even? 2643 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2644 bz %icc, .L256_ctr32_loop2x 2645 srlx %i2, 4, %i2 2646 .L256_ctr32_loop: 2647 ldx [%i0 + 0], %o0 2648 brz,pt %l0, 4f 2649 ldx [%i0 + 8], %o1 2650 2651 ldx [%i0 + 16], %o2 2652 sllx %o0, %l0, %o0 2653 srlx %o1, %l1, %g1 2654 sllx %o1, %l0, %o1 2655 or %g1, %o0, %o0 2656 srlx %o2, %l1, %o2 2657 or %o2, %o1, %o1 2658 4: 2659 xor %g5, %l7, %g1 ! ^= rk[0] 2660 add %l7, 1, %l7 2661 .word 0x85b02301 !movxtod %g1,%f2 2662 srl %l7, 0, %l7 ! clruw 2663 prefetch [%i1 + 63], 22 2664 prefetch [%i0 + 16+63], 20 2665 .word 0x88cc040e !aes_eround01 %f16,%f14,%f2,%f4 2666 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 2667 call _aes256_encrypt_1x+8 2668 add %i0, 16, %i0 2669 2670 .word 0x95b02308 !movxtod %o0,%f10 2671 .word 0x99b02309 !movxtod %o1,%f12 2672 .word 0x81b28d80 !fxor %f10,%f0,%f0 ! ^= inp 2673 .word 0x85b30d82 !fxor %f12,%f2,%f2 2674 2675 brnz,pn %l2, 2f 2676 sub %i2, 1, %i2 2677 2678 std %f0, [%i1 + 0] 2679 std %f2, [%i1 + 8] 2680 brnz,pt %i2, .L256_ctr32_loop2x 2681 add %i1, 16, %i1 2682 2683 ret 2684 restore 2685 2686 .align 16 2687 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2688 ! and ~3x deterioration 2689 ! in inp==out case 2690 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2691 .word 0x8db00902 !faligndata %f0,%f2,%f6 2692 .word 0x91b08902 !faligndata %f2,%f2,%f8 2693 stda %f4, [%i1 + %l3]0xc0 ! partial store 2694 std %f6, [%i1 + 8] 2695 add %i1, 16, %i1 2696 orn %g0, %l3, %l3 2697 stda %f8, [%i1 + %l3]0xc0 ! partial store 2698 2699 brnz,pt %i2, .L256_ctr32_loop2x+4 2700 orn %g0, %l3, %l3 2701 2702 ret 2703 restore 2704 2705 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2706 .align 32 2707 .L256_ctr32_loop2x: 2708 ldx [%i0 + 0], %o0 2709 ldx [%i0 + 8], %o1 2710 ldx [%i0 + 16], %o2 2711 brz,pt %l0, 4f 2712 ldx [%i0 + 24], %o3 2713 2714 ldx [%i0 + 32], %o4 2715 sllx %o0, %l0, %o0 2716 srlx %o1, %l1, %g1 2717 or %g1, %o0, %o0 2718 sllx %o1, %l0, %o1 2719 srlx %o2, %l1, %g1 2720 or %g1, %o1, %o1 2721 sllx %o2, %l0, %o2 2722 srlx %o3, %l1, %g1 2723 or %g1, %o2, %o2 2724 sllx %o3, %l0, %o3 2725 srlx %o4, %l1, %o4 2726 or %o4, %o3, %o3 2727 4: 2728 xor %g5, %l7, %g1 ! ^= rk[0] 2729 add %l7, 1, %l7 2730 .word 0x85b02301 !movxtod %g1,%f2 2731 srl %l7, 0, %l7 ! clruw 2732 xor %g5, %l7, %g1 2733 add %l7, 1, %l7 2734 .word 0x8db02301 !movxtod %g1,%f6 2735 srl %l7, 0, %l7 ! clruw 2736 prefetch [%i1 + 63], 22 2737 prefetch [%i0 + 32+63], 20 2738 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 2739 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 2740 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 2741 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 2742 call _aes256_encrypt_2x+16 2743 add %i0, 32, %i0 2744 2745 .word 0x91b02308 !movxtod %o0,%f8 2746 .word 0x95b02309 !movxtod %o1,%f10 2747 .word 0x99b0230a !movxtod %o2,%f12 2748 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 2749 .word 0x91b0230b !movxtod %o3,%f8 2750 .word 0x85b28d82 !fxor %f10,%f2,%f2 2751 .word 0x89b30d84 !fxor %f12,%f4,%f4 2752 .word 0x8db20d86 !fxor %f8,%f6,%f6 2753 2754 brnz,pn %l2, 2f 2755 sub %i2, 2, %i2 2756 2757 std %f0, [%i1 + 0] 2758 std %f2, [%i1 + 8] 2759 std %f4, [%i1 + 16] 2760 std %f6, [%i1 + 24] 2761 brnz,pt %i2, .L256_ctr32_loop2x 2762 add %i1, 32, %i1 2763 2764 ret 2765 restore 2766 2767 .align 16 2768 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2769 ! and ~3x deterioration 2770 ! in inp==out case 2771 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 2772 .word 0x81b00902 !faligndata %f0,%f2,%f0 2773 .word 0x85b08904 !faligndata %f2,%f4,%f2 2774 .word 0x89b10906 !faligndata %f4,%f6,%f4 2775 .word 0x8db18906 !faligndata %f6,%f6,%f6 2776 2777 stda %f8, [%i1 + %l3]0xc0 ! partial store 2778 std %f0, [%i1 + 8] 2779 std %f2, [%i1 + 16] 2780 std %f4, [%i1 + 24] 2781 add %i1, 32, %i1 2782 orn %g0, %l3, %l3 2783 stda %f6, [%i1 + %l3]0xc0 ! partial store 2784 2785 brnz,pt %i2, .L256_ctr32_loop2x+4 2786 orn %g0, %l3, %l3 2787 2788 ret 2789 restore 2790 2791 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2792 .align 32 2793 .L256_ctr32_blk: 2794 add %i1, %i2, %l5 2795 and %l5, 63, %l5 ! tail 2796 sub %i2, %l5, %i2 2797 add %l5, 15, %l5 ! round up to 16n 2798 srlx %i2, 4, %i2 2799 srl %l5, 4, %l5 2800 sub %i2, 1, %i2 2801 add %l5, 1, %l5 2802 2803 .L256_ctr32_blk_loop2x: 2804 ldx [%i0 + 0], %o0 2805 ldx [%i0 + 8], %o1 2806 ldx [%i0 + 16], %o2 2807 brz,pt %l0, 5f 2808 ldx [%i0 + 24], %o3 2809 2810 ldx [%i0 + 32], %o4 2811 sllx %o0, %l0, %o0 2812 srlx %o1, %l1, %g1 2813 or %g1, %o0, %o0 2814 sllx %o1, %l0, %o1 2815 srlx %o2, %l1, %g1 2816 or %g1, %o1, %o1 2817 sllx %o2, %l0, %o2 2818 srlx %o3, %l1, %g1 2819 or %g1, %o2, %o2 2820 sllx %o3, %l0, %o3 2821 srlx %o4, %l1, %o4 2822 or %o4, %o3, %o3 2823 5: 2824 xor %g5, %l7, %g1 ! ^= rk[0] 2825 add %l7, 1, %l7 2826 .word 0x85b02301 !movxtod %g1,%f2 2827 srl %l7, 0, %l7 ! clruw 2828 xor %g5, %l7, %g1 2829 add %l7, 1, %l7 2830 .word 0x8db02301 !movxtod %g1,%f6 2831 srl %l7, 0, %l7 ! clruw 2832 prefetch [%i0 + 32+63], 20 2833 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 2834 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 2835 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 2836 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 2837 call _aes256_encrypt_2x+16 2838 add %i0, 32, %i0 2839 subcc %i2, 2, %i2 2840 2841 .word 0x91b02308 !movxtod %o0,%f8 2842 .word 0x95b02309 !movxtod %o1,%f10 2843 .word 0x99b0230a !movxtod %o2,%f12 2844 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 2845 .word 0x91b0230b !movxtod %o3,%f8 2846 .word 0x85b28d82 !fxor %f10,%f2,%f2 2847 .word 0x89b30d84 !fxor %f12,%f4,%f4 2848 .word 0x8db20d86 !fxor %f8,%f6,%f6 2849 2850 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2851 add %i1, 8, %i1 2852 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2853 add %i1, 8, %i1 2854 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2855 add %i1, 8, %i1 2856 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 2857 bgu,pt SIZE_T_CC, .L256_ctr32_blk_loop2x 2858 add %i1, 8, %i1 2859 2860 add %l5, %i2, %i2 2861 andcc %i2, 1, %g0 ! is number of blocks even? 2862 membar #StoreLoad|#StoreStore 2863 bnz,pt %icc, .L256_ctr32_loop 2864 srl %i2, 0, %i2 2865 brnz,pn %i2, .L256_ctr32_loop2x 2866 nop 2867 2868 ret 2869 restore 2870 .type aes256_t4_ctr32_encrypt,#function 2871 .size aes256_t4_ctr32_encrypt,.-aes256_t4_ctr32_encrypt 2872 .globl aes256_t4_xts_encrypt 2873 .align 32 2874 aes256_t4_xts_encrypt: 2875 save %sp, -STACK_FRAME-16, %sp 2876 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 2877 2878 mov %i5, %o0 2879 add %fp, STACK_BIAS-16, %o1 2880 call aes_t4_encrypt 2881 mov %i4, %o2 2882 2883 add %fp, STACK_BIAS-16, %l7 2884 ldxa [%l7]0x88, %g2 2885 add %fp, STACK_BIAS-8, %l7 2886 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 2887 2888 sethi %hi(0x76543210), %l7 2889 or %l7, %lo(0x76543210), %l7 2890 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 2891 2892 prefetch [%i0], 20 2893 prefetch [%i0 + 63], 20 2894 call _aes256_load_enckey 2895 and %i2, 15, %i5 2896 and %i2, -16, %i2 2897 2898 sub %i0, %i1, %l5 ! %i0!=%i1 2899 and %i0, 7, %l0 2900 andn %i0, 7, %i0 2901 sll %l0, 3, %l0 2902 mov 64, %l1 2903 mov 0xff, %l3 2904 sub %l1, %l0, %l1 2905 and %i1, 7, %l2 2906 cmp %i2, 255 2907 movrnz %l2, 0, %l5 ! if ( %i1&7 || 2908 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 2909 brnz,pn %l5, .L256_xts_enblk ! %i0==%i1) 2910 srl %l3, %l2, %l3 2911 2912 andcc %i2, 16, %g0 ! is number of blocks even? 2913 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 2914 bz %icc, .L256_xts_enloop2x 2915 srlx %i2, 4, %i2 2916 .L256_xts_enloop: 2917 ldx [%i0 + 0], %o0 2918 brz,pt %l0, 4f 2919 ldx [%i0 + 8], %o1 2920 2921 ldx [%i0 + 16], %o2 2922 sllx %o0, %l0, %o0 2923 srlx %o1, %l1, %g1 2924 sllx %o1, %l0, %o1 2925 or %g1, %o0, %o0 2926 srlx %o2, %l1, %o2 2927 or %o2, %o1, %o1 2928 4: 2929 .word 0x99b02302 !movxtod %g2,%f12 2930 .word 0x9db02303 !movxtod %g3,%f14 2931 .word 0x99b3098c !bshuffle %f12,%f12,%f12 2932 .word 0x9db3898e !bshuffle %f14,%f14,%f14 2933 2934 xor %g4, %o0, %o0 ! ^= rk[0] 2935 xor %g5, %o1, %o1 2936 .word 0x81b02308 !movxtod %o0,%f0 2937 .word 0x85b02309 !movxtod %o1,%f2 2938 2939 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 2940 .word 0x85b38d82 !fxor %f14,%f2,%f2 2941 2942 prefetch [%i1 + 63], 22 2943 prefetch [%i0 + 16+63], 20 2944 call _aes256_encrypt_1x 2945 add %i0, 16, %i0 2946 2947 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 2948 .word 0x85b38d82 !fxor %f14,%f2,%f2 2949 2950 srax %g3, 63, %l7 ! next tweak value 2951 addcc %g2, %g2, %g2 2952 and %l7, 0x87, %l7 2953 .word 0x87b0c223 !addxc %g3,%g3,%g3 2954 xor %l7, %g2, %g2 2955 2956 brnz,pn %l2, 2f 2957 sub %i2, 1, %i2 2958 2959 std %f0, [%i1 + 0] 2960 std %f2, [%i1 + 8] 2961 brnz,pt %i2, .L256_xts_enloop2x 2962 add %i1, 16, %i1 2963 2964 brnz,pn %i5, .L256_xts_ensteal 2965 nop 2966 2967 ret 2968 restore 2969 2970 .align 16 2971 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 2972 ! and ~3x deterioration 2973 ! in inp==out case 2974 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 2975 .word 0x8db00902 !faligndata %f0,%f2,%f6 2976 .word 0x91b08902 !faligndata %f2,%f2,%f8 2977 stda %f4, [%i1 + %l3]0xc0 ! partial store 2978 std %f6, [%i1 + 8] 2979 add %i1, 16, %i1 2980 orn %g0, %l3, %l3 2981 stda %f8, [%i1 + %l3]0xc0 ! partial store 2982 2983 brnz,pt %i2, .L256_xts_enloop2x+4 2984 orn %g0, %l3, %l3 2985 2986 brnz,pn %i5, .L256_xts_ensteal 2987 nop 2988 2989 ret 2990 restore 2991 2992 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2993 .align 32 2994 .L256_xts_enloop2x: 2995 ldx [%i0 + 0], %o0 2996 ldx [%i0 + 8], %o1 2997 ldx [%i0 + 16], %o2 2998 brz,pt %l0, 4f 2999 ldx [%i0 + 24], %o3 3000 3001 ldx [%i0 + 32], %o4 3002 sllx %o0, %l0, %o0 3003 srlx %o1, %l1, %g1 3004 or %g1, %o0, %o0 3005 sllx %o1, %l0, %o1 3006 srlx %o2, %l1, %g1 3007 or %g1, %o1, %o1 3008 sllx %o2, %l0, %o2 3009 srlx %o3, %l1, %g1 3010 or %g1, %o2, %o2 3011 sllx %o3, %l0, %o3 3012 srlx %o4, %l1, %o4 3013 or %o4, %o3, %o3 3014 4: 3015 .word 0x99b02302 !movxtod %g2,%f12 3016 .word 0x9db02303 !movxtod %g3,%f14 3017 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3018 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3019 3020 srax %g3, 63, %l7 ! next tweak value 3021 addcc %g2, %g2, %g2 3022 and %l7, 0x87, %l7 3023 .word 0x87b0c223 !addxc %g3,%g3,%g3 3024 xor %l7, %g2, %g2 3025 3026 .word 0x91b02302 !movxtod %g2,%f8 3027 .word 0x95b02303 !movxtod %g3,%f10 3028 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3029 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3030 3031 xor %g4, %o0, %o0 ! ^= rk[0] 3032 xor %g5, %o1, %o1 3033 xor %g4, %o2, %o2 ! ^= rk[0] 3034 xor %g5, %o3, %o3 3035 .word 0x81b02308 !movxtod %o0,%f0 3036 .word 0x85b02309 !movxtod %o1,%f2 3037 .word 0x89b0230a !movxtod %o2,%f4 3038 .word 0x8db0230b !movxtod %o3,%f6 3039 3040 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3041 .word 0x85b38d82 !fxor %f14,%f2,%f2 3042 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3043 .word 0x8db28d86 !fxor %f10,%f6,%f6 3044 3045 prefetch [%i1 + 63], 22 3046 prefetch [%i0 + 32+63], 20 3047 call _aes256_encrypt_2x 3048 add %i0, 32, %i0 3049 3050 .word 0x91b02302 !movxtod %g2,%f8 3051 .word 0x95b02303 !movxtod %g3,%f10 3052 3053 srax %g3, 63, %l7 ! next tweak value 3054 addcc %g2, %g2, %g2 3055 and %l7, 0x87, %l7 3056 .word 0x87b0c223 !addxc %g3,%g3,%g3 3057 xor %l7, %g2, %g2 3058 3059 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3060 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3061 3062 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3063 .word 0x85b38d82 !fxor %f14,%f2,%f2 3064 .word 0x89b20d84 !fxor %f8,%f4,%f4 3065 .word 0x8db28d86 !fxor %f10,%f6,%f6 3066 3067 brnz,pn %l2, 2f 3068 sub %i2, 2, %i2 3069 3070 std %f0, [%i1 + 0] 3071 std %f2, [%i1 + 8] 3072 std %f4, [%i1 + 16] 3073 std %f6, [%i1 + 24] 3074 brnz,pt %i2, .L256_xts_enloop2x 3075 add %i1, 32, %i1 3076 3077 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3078 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3079 brnz,pn %i5, .L256_xts_ensteal 3080 nop 3081 3082 ret 3083 restore 3084 3085 .align 16 3086 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3087 ! and ~3x deterioration 3088 ! in inp==out case 3089 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 3090 .word 0x95b00902 !faligndata %f0,%f2,%f10 3091 .word 0x99b08904 !faligndata %f2,%f4,%f12 3092 .word 0x9db10906 !faligndata %f4,%f6,%f14 3093 .word 0x81b18906 !faligndata %f6,%f6,%f0 3094 3095 stda %f8, [%i1 + %l3]0xc0 ! partial store 3096 std %f10, [%i1 + 8] 3097 std %f12, [%i1 + 16] 3098 std %f14, [%i1 + 24] 3099 add %i1, 32, %i1 3100 orn %g0, %l3, %l3 3101 stda %f0, [%i1 + %l3]0xc0 ! partial store 3102 3103 brnz,pt %i2, .L256_xts_enloop2x+4 3104 orn %g0, %l3, %l3 3105 3106 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3107 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3108 brnz,pn %i5, .L256_xts_ensteal 3109 nop 3110 3111 ret 3112 restore 3113 3114 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3115 .align 32 3116 .L256_xts_enblk: 3117 add %i1, %i2, %l5 3118 and %l5, 63, %l5 ! tail 3119 sub %i2, %l5, %i2 3120 add %l5, 15, %l5 ! round up to 16n 3121 srlx %i2, 4, %i2 3122 srl %l5, 4, %l5 3123 sub %i2, 1, %i2 3124 add %l5, 1, %l5 3125 3126 .L256_xts_enblk2x: 3127 ldx [%i0 + 0], %o0 3128 ldx [%i0 + 8], %o1 3129 ldx [%i0 + 16], %o2 3130 brz,pt %l0, 5f 3131 ldx [%i0 + 24], %o3 3132 3133 ldx [%i0 + 32], %o4 3134 sllx %o0, %l0, %o0 3135 srlx %o1, %l1, %g1 3136 or %g1, %o0, %o0 3137 sllx %o1, %l0, %o1 3138 srlx %o2, %l1, %g1 3139 or %g1, %o1, %o1 3140 sllx %o2, %l0, %o2 3141 srlx %o3, %l1, %g1 3142 or %g1, %o2, %o2 3143 sllx %o3, %l0, %o3 3144 srlx %o4, %l1, %o4 3145 or %o4, %o3, %o3 3146 5: 3147 .word 0x99b02302 !movxtod %g2,%f12 3148 .word 0x9db02303 !movxtod %g3,%f14 3149 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3150 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3151 3152 srax %g3, 63, %l7 ! next tweak value 3153 addcc %g2, %g2, %g2 3154 and %l7, 0x87, %l7 3155 .word 0x87b0c223 !addxc %g3,%g3,%g3 3156 xor %l7, %g2, %g2 3157 3158 .word 0x91b02302 !movxtod %g2,%f8 3159 .word 0x95b02303 !movxtod %g3,%f10 3160 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3161 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3162 3163 xor %g4, %o0, %o0 ! ^= rk[0] 3164 xor %g5, %o1, %o1 3165 xor %g4, %o2, %o2 ! ^= rk[0] 3166 xor %g5, %o3, %o3 3167 .word 0x81b02308 !movxtod %o0,%f0 3168 .word 0x85b02309 !movxtod %o1,%f2 3169 .word 0x89b0230a !movxtod %o2,%f4 3170 .word 0x8db0230b !movxtod %o3,%f6 3171 3172 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3173 .word 0x85b38d82 !fxor %f14,%f2,%f2 3174 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3175 .word 0x8db28d86 !fxor %f10,%f6,%f6 3176 3177 prefetch [%i0 + 32+63], 20 3178 call _aes256_encrypt_2x 3179 add %i0, 32, %i0 3180 3181 .word 0x91b02302 !movxtod %g2,%f8 3182 .word 0x95b02303 !movxtod %g3,%f10 3183 3184 srax %g3, 63, %l7 ! next tweak value 3185 addcc %g2, %g2, %g2 3186 and %l7, 0x87, %l7 3187 .word 0x87b0c223 !addxc %g3,%g3,%g3 3188 xor %l7, %g2, %g2 3189 3190 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3191 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3192 3193 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3194 .word 0x85b38d82 !fxor %f14,%f2,%f2 3195 .word 0x89b20d84 !fxor %f8,%f4,%f4 3196 .word 0x8db28d86 !fxor %f10,%f6,%f6 3197 3198 subcc %i2, 2, %i2 3199 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3200 add %i1, 8, %i1 3201 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3202 add %i1, 8, %i1 3203 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3204 add %i1, 8, %i1 3205 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3206 bgu,pt SIZE_T_CC, .L256_xts_enblk2x 3207 add %i1, 8, %i1 3208 3209 add %l5, %i2, %i2 3210 andcc %i2, 1, %g0 ! is number of blocks even? 3211 membar #StoreLoad|#StoreStore 3212 bnz,pt %icc, .L256_xts_enloop 3213 srl %i2, 0, %i2 3214 brnz,pn %i2, .L256_xts_enloop2x 3215 nop 3216 3217 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3218 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3219 brnz,pn %i5, .L256_xts_ensteal 3220 nop 3221 3222 ret 3223 restore 3224 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3225 .align 32 3226 .L256_xts_ensteal: 3227 std %f0, [%fp + STACK_BIAS-16] ! copy of output 3228 std %f2, [%fp + STACK_BIAS-8] 3229 3230 srl %l0, 3, %l0 3231 add %fp, STACK_BIAS-16, %l7 3232 add %i0, %l0, %i0 ! original %i0+%i2&-15 3233 add %i1, %l2, %i1 ! original %i1+%i2&-15 3234 mov 0, %l0 3235 nop ! align 3236 3237 .L256_xts_enstealing: 3238 ldub [%i0 + %l0], %o0 3239 ldub [%l7 + %l0], %o1 3240 dec %i5 3241 stb %o0, [%l7 + %l0] 3242 stb %o1, [%i1 + %l0] 3243 brnz %i5, .L256_xts_enstealing 3244 inc %l0 3245 3246 mov %l7, %i0 3247 sub %i1, 16, %i1 3248 mov 0, %l0 3249 sub %i1, %l2, %i1 3250 ba .L256_xts_enloop ! one more time 3251 mov 1, %i2 ! %i5 is 0 3252 ret 3253 restore 3254 .type aes256_t4_xts_encrypt,#function 3255 .size aes256_t4_xts_encrypt,.-aes256_t4_xts_encrypt 3256 .globl aes256_t4_xts_decrypt 3257 .align 32 3258 aes256_t4_xts_decrypt: 3259 save %sp, -STACK_FRAME-16, %sp 3260 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 3261 3262 mov %i5, %o0 3263 add %fp, STACK_BIAS-16, %o1 3264 call aes_t4_encrypt 3265 mov %i4, %o2 3266 3267 add %fp, STACK_BIAS-16, %l7 3268 ldxa [%l7]0x88, %g2 3269 add %fp, STACK_BIAS-8, %l7 3270 ldxa [%l7]0x88, %g3 ! %g3:%g2 is tweak 3271 3272 sethi %hi(0x76543210), %l7 3273 or %l7, %lo(0x76543210), %l7 3274 .word 0x81b5c320 !bmask %l7,%g0,%g0 ! byte swap mask 3275 3276 prefetch [%i0], 20 3277 prefetch [%i0 + 63], 20 3278 call _aes256_load_deckey 3279 and %i2, 15, %i5 3280 and %i2, -16, %i2 3281 mov 0, %l7 3282 movrnz %i5, 16, %l7 3283 sub %i2, %l7, %i2 3284 3285 sub %i0, %i1, %l5 ! %i0!=%i1 3286 and %i0, 7, %l0 3287 andn %i0, 7, %i0 3288 sll %l0, 3, %l0 3289 mov 64, %l1 3290 mov 0xff, %l3 3291 sub %l1, %l0, %l1 3292 and %i1, 7, %l2 3293 cmp %i2, 255 3294 movrnz %l2, 0, %l5 ! if ( %i1&7 || 3295 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 3296 brnz,pn %l5, .L256_xts_deblk ! %i0==%i1) 3297 srl %l3, %l2, %l3 3298 3299 andcc %i2, 16, %g0 ! is number of blocks even? 3300 brz,pn %i2, .L256_xts_desteal 3301 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 3302 bz %icc, .L256_xts_deloop2x 3303 srlx %i2, 4, %i2 3304 .L256_xts_deloop: 3305 ldx [%i0 + 0], %o0 3306 brz,pt %l0, 4f 3307 ldx [%i0 + 8], %o1 3308 3309 ldx [%i0 + 16], %o2 3310 sllx %o0, %l0, %o0 3311 srlx %o1, %l1, %g1 3312 sllx %o1, %l0, %o1 3313 or %g1, %o0, %o0 3314 srlx %o2, %l1, %o2 3315 or %o2, %o1, %o1 3316 4: 3317 .word 0x99b02302 !movxtod %g2,%f12 3318 .word 0x9db02303 !movxtod %g3,%f14 3319 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3320 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3321 3322 xor %g4, %o0, %o0 ! ^= rk[0] 3323 xor %g5, %o1, %o1 3324 .word 0x81b02308 !movxtod %o0,%f0 3325 .word 0x85b02309 !movxtod %o1,%f2 3326 3327 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3328 .word 0x85b38d82 !fxor %f14,%f2,%f2 3329 3330 prefetch [%i1 + 63], 22 3331 prefetch [%i0 + 16+63], 20 3332 call _aes256_decrypt_1x 3333 add %i0, 16, %i0 3334 3335 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3336 .word 0x85b38d82 !fxor %f14,%f2,%f2 3337 3338 srax %g3, 63, %l7 ! next tweak value 3339 addcc %g2, %g2, %g2 3340 and %l7, 0x87, %l7 3341 .word 0x87b0c223 !addxc %g3,%g3,%g3 3342 xor %l7, %g2, %g2 3343 3344 brnz,pn %l2, 2f 3345 sub %i2, 1, %i2 3346 3347 std %f0, [%i1 + 0] 3348 std %f2, [%i1 + 8] 3349 brnz,pt %i2, .L256_xts_deloop2x 3350 add %i1, 16, %i1 3351 3352 brnz,pn %i5, .L256_xts_desteal 3353 nop 3354 3355 ret 3356 restore 3357 3358 .align 16 3359 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3360 ! and ~3x deterioration 3361 ! in inp==out case 3362 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 3363 .word 0x8db00902 !faligndata %f0,%f2,%f6 3364 .word 0x91b08902 !faligndata %f2,%f2,%f8 3365 stda %f4, [%i1 + %l3]0xc0 ! partial store 3366 std %f6, [%i1 + 8] 3367 add %i1, 16, %i1 3368 orn %g0, %l3, %l3 3369 stda %f8, [%i1 + %l3]0xc0 ! partial store 3370 3371 brnz,pt %i2, .L256_xts_deloop2x+4 3372 orn %g0, %l3, %l3 3373 3374 brnz,pn %i5, .L256_xts_desteal 3375 nop 3376 3377 ret 3378 restore 3379 3380 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3381 .align 32 3382 .L256_xts_deloop2x: 3383 ldx [%i0 + 0], %o0 3384 ldx [%i0 + 8], %o1 3385 ldx [%i0 + 16], %o2 3386 brz,pt %l0, 4f 3387 ldx [%i0 + 24], %o3 3388 3389 ldx [%i0 + 32], %o4 3390 sllx %o0, %l0, %o0 3391 srlx %o1, %l1, %g1 3392 or %g1, %o0, %o0 3393 sllx %o1, %l0, %o1 3394 srlx %o2, %l1, %g1 3395 or %g1, %o1, %o1 3396 sllx %o2, %l0, %o2 3397 srlx %o3, %l1, %g1 3398 or %g1, %o2, %o2 3399 sllx %o3, %l0, %o3 3400 srlx %o4, %l1, %o4 3401 or %o4, %o3, %o3 3402 4: 3403 .word 0x99b02302 !movxtod %g2,%f12 3404 .word 0x9db02303 !movxtod %g3,%f14 3405 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3406 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3407 3408 srax %g3, 63, %l7 ! next tweak value 3409 addcc %g2, %g2, %g2 3410 and %l7, 0x87, %l7 3411 .word 0x87b0c223 !addxc %g3,%g3,%g3 3412 xor %l7, %g2, %g2 3413 3414 .word 0x91b02302 !movxtod %g2,%f8 3415 .word 0x95b02303 !movxtod %g3,%f10 3416 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3417 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3418 3419 xor %g4, %o0, %o0 ! ^= rk[0] 3420 xor %g5, %o1, %o1 3421 xor %g4, %o2, %o2 ! ^= rk[0] 3422 xor %g5, %o3, %o3 3423 .word 0x81b02308 !movxtod %o0,%f0 3424 .word 0x85b02309 !movxtod %o1,%f2 3425 .word 0x89b0230a !movxtod %o2,%f4 3426 .word 0x8db0230b !movxtod %o3,%f6 3427 3428 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3429 .word 0x85b38d82 !fxor %f14,%f2,%f2 3430 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3431 .word 0x8db28d86 !fxor %f10,%f6,%f6 3432 3433 prefetch [%i1 + 63], 22 3434 prefetch [%i0 + 32+63], 20 3435 call _aes256_decrypt_2x 3436 add %i0, 32, %i0 3437 3438 .word 0x91b02302 !movxtod %g2,%f8 3439 .word 0x95b02303 !movxtod %g3,%f10 3440 3441 srax %g3, 63, %l7 ! next tweak value 3442 addcc %g2, %g2, %g2 3443 and %l7, 0x87, %l7 3444 .word 0x87b0c223 !addxc %g3,%g3,%g3 3445 xor %l7, %g2, %g2 3446 3447 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3448 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3449 3450 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3451 .word 0x85b38d82 !fxor %f14,%f2,%f2 3452 .word 0x89b20d84 !fxor %f8,%f4,%f4 3453 .word 0x8db28d86 !fxor %f10,%f6,%f6 3454 3455 brnz,pn %l2, 2f 3456 sub %i2, 2, %i2 3457 3458 std %f0, [%i1 + 0] 3459 std %f2, [%i1 + 8] 3460 std %f4, [%i1 + 16] 3461 std %f6, [%i1 + 24] 3462 brnz,pt %i2, .L256_xts_deloop2x 3463 add %i1, 32, %i1 3464 3465 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3466 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3467 brnz,pn %i5, .L256_xts_desteal 3468 nop 3469 3470 ret 3471 restore 3472 3473 .align 16 3474 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3475 ! and ~3x deterioration 3476 ! in inp==out case 3477 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 3478 .word 0x95b00902 !faligndata %f0,%f2,%f10 3479 .word 0x99b08904 !faligndata %f2,%f4,%f12 3480 .word 0x9db10906 !faligndata %f4,%f6,%f14 3481 .word 0x81b18906 !faligndata %f6,%f6,%f0 3482 3483 stda %f8, [%i1 + %l3]0xc0 ! partial store 3484 std %f10, [%i1 + 8] 3485 std %f12, [%i1 + 16] 3486 std %f14, [%i1 + 24] 3487 add %i1, 32, %i1 3488 orn %g0, %l3, %l3 3489 stda %f0, [%i1 + %l3]0xc0 ! partial store 3490 3491 brnz,pt %i2, .L256_xts_deloop2x+4 3492 orn %g0, %l3, %l3 3493 3494 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3495 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3496 brnz,pn %i5, .L256_xts_desteal 3497 nop 3498 3499 ret 3500 restore 3501 3502 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3503 .align 32 3504 .L256_xts_deblk: 3505 add %i1, %i2, %l5 3506 and %l5, 63, %l5 ! tail 3507 sub %i2, %l5, %i2 3508 add %l5, 15, %l5 ! round up to 16n 3509 srlx %i2, 4, %i2 3510 srl %l5, 4, %l5 3511 sub %i2, 1, %i2 3512 add %l5, 1, %l5 3513 3514 .L256_xts_deblk2x: 3515 ldx [%i0 + 0], %o0 3516 ldx [%i0 + 8], %o1 3517 ldx [%i0 + 16], %o2 3518 brz,pt %l0, 5f 3519 ldx [%i0 + 24], %o3 3520 3521 ldx [%i0 + 32], %o4 3522 sllx %o0, %l0, %o0 3523 srlx %o1, %l1, %g1 3524 or %g1, %o0, %o0 3525 sllx %o1, %l0, %o1 3526 srlx %o2, %l1, %g1 3527 or %g1, %o1, %o1 3528 sllx %o2, %l0, %o2 3529 srlx %o3, %l1, %g1 3530 or %g1, %o2, %o2 3531 sllx %o3, %l0, %o3 3532 srlx %o4, %l1, %o4 3533 or %o4, %o3, %o3 3534 5: 3535 .word 0x99b02302 !movxtod %g2,%f12 3536 .word 0x9db02303 !movxtod %g3,%f14 3537 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3538 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3539 3540 srax %g3, 63, %l7 ! next tweak value 3541 addcc %g2, %g2, %g2 3542 and %l7, 0x87, %l7 3543 .word 0x87b0c223 !addxc %g3,%g3,%g3 3544 xor %l7, %g2, %g2 3545 3546 .word 0x91b02302 !movxtod %g2,%f8 3547 .word 0x95b02303 !movxtod %g3,%f10 3548 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3549 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3550 3551 xor %g4, %o0, %o0 ! ^= rk[0] 3552 xor %g5, %o1, %o1 3553 xor %g4, %o2, %o2 ! ^= rk[0] 3554 xor %g5, %o3, %o3 3555 .word 0x81b02308 !movxtod %o0,%f0 3556 .word 0x85b02309 !movxtod %o1,%f2 3557 .word 0x89b0230a !movxtod %o2,%f4 3558 .word 0x8db0230b !movxtod %o3,%f6 3559 3560 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3561 .word 0x85b38d82 !fxor %f14,%f2,%f2 3562 .word 0x89b20d84 !fxor %f8,%f4,%f4 ! ^= tweak[0] 3563 .word 0x8db28d86 !fxor %f10,%f6,%f6 3564 3565 prefetch [%i0 + 32+63], 20 3566 call _aes256_decrypt_2x 3567 add %i0, 32, %i0 3568 3569 .word 0x91b02302 !movxtod %g2,%f8 3570 .word 0x95b02303 !movxtod %g3,%f10 3571 3572 srax %g3, 63, %l7 ! next tweak value 3573 addcc %g2, %g2, %g2 3574 and %l7, 0x87, %l7 3575 .word 0x87b0c223 !addxc %g3,%g3,%g3 3576 xor %l7, %g2, %g2 3577 3578 .word 0x91b20988 !bshuffle %f8,%f8,%f8 3579 .word 0x95b2898a !bshuffle %f10,%f10,%f10 3580 3581 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3582 .word 0x85b38d82 !fxor %f14,%f2,%f2 3583 .word 0x89b20d84 !fxor %f8,%f4,%f4 3584 .word 0x8db28d86 !fxor %f10,%f6,%f6 3585 3586 subcc %i2, 2, %i2 3587 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3588 add %i1, 8, %i1 3589 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3590 add %i1, 8, %i1 3591 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3592 add %i1, 8, %i1 3593 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3594 bgu,pt SIZE_T_CC, .L256_xts_deblk2x 3595 add %i1, 8, %i1 3596 3597 add %l5, %i2, %i2 3598 andcc %i2, 1, %g0 ! is number of blocks even? 3599 membar #StoreLoad|#StoreStore 3600 bnz,pt %icc, .L256_xts_deloop 3601 srl %i2, 0, %i2 3602 brnz,pn %i2, .L256_xts_deloop2x 3603 nop 3604 3605 .word 0x81b00f04 !fsrc2 %f0,%f4,%f0 3606 .word 0x85b00f06 !fsrc2 %f0,%f6,%f2 3607 brnz,pn %i5, .L256_xts_desteal 3608 nop 3609 3610 ret 3611 restore 3612 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3613 .align 32 3614 .L256_xts_desteal: 3615 ldx [%i0 + 0], %o0 3616 brz,pt %l0, 8f 3617 ldx [%i0 + 8], %o1 3618 3619 ldx [%i0 + 16], %o2 3620 sllx %o0, %l0, %o0 3621 srlx %o1, %l1, %g1 3622 sllx %o1, %l0, %o1 3623 or %g1, %o0, %o0 3624 srlx %o2, %l1, %o2 3625 or %o2, %o1, %o1 3626 8: 3627 srax %g3, 63, %l7 ! next tweak value 3628 addcc %g2, %g2, %o2 3629 and %l7, 0x87, %l7 3630 .word 0x97b0c223 !addxc %g3,%g3,%o3 3631 xor %l7, %o2, %o2 3632 3633 .word 0x99b0230a !movxtod %o2,%f12 3634 .word 0x9db0230b !movxtod %o3,%f14 3635 .word 0x99b3098c !bshuffle %f12,%f12,%f12 3636 .word 0x9db3898e !bshuffle %f14,%f14,%f14 3637 3638 xor %g4, %o0, %o0 ! ^= rk[0] 3639 xor %g5, %o1, %o1 3640 .word 0x81b02308 !movxtod %o0,%f0 3641 .word 0x85b02309 !movxtod %o1,%f2 3642 3643 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3644 .word 0x85b38d82 !fxor %f14,%f2,%f2 3645 3646 call _aes256_decrypt_1x 3647 add %i0, 16, %i0 3648 3649 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= tweak[0] 3650 .word 0x85b38d82 !fxor %f14,%f2,%f2 3651 3652 std %f0, [%fp + STACK_BIAS-16] 3653 std %f2, [%fp + STACK_BIAS-8] 3654 3655 srl %l0, 3, %l0 3656 add %fp, STACK_BIAS-16, %l7 3657 add %i0, %l0, %i0 ! original %i0+%i2&-15 3658 add %i1, %l2, %i1 ! original %i1+%i2&-15 3659 mov 0, %l0 3660 add %i1, 16, %i1 3661 nop ! align 3662 3663 .L256_xts_destealing: 3664 ldub [%i0 + %l0], %o0 3665 ldub [%l7 + %l0], %o1 3666 dec %i5 3667 stb %o0, [%l7 + %l0] 3668 stb %o1, [%i1 + %l0] 3669 brnz %i5, .L256_xts_destealing 3670 inc %l0 3671 3672 mov %l7, %i0 3673 sub %i1, 16, %i1 3674 mov 0, %l0 3675 sub %i1, %l2, %i1 3676 ba .L256_xts_deloop ! one more time 3677 mov 1, %i2 ! %i5 is 0 3678 ret 3679 restore 3680 .type aes256_t4_xts_decrypt,#function 3681 .size aes256_t4_xts_decrypt,.-aes256_t4_xts_decrypt 3682 .globl aes192_t4_ctr32_encrypt 3683 .align 32 3684 aes192_t4_ctr32_encrypt: 3685 save %sp, -STACK_FRAME, %sp 3686 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 3687 3688 prefetch [%i0], 20 3689 prefetch [%i0 + 63], 20 3690 call _aes192_load_enckey 3691 sllx %i2, 4, %i2 3692 3693 ld [%i4 + 0], %l4 ! counter 3694 ld [%i4 + 4], %l5 3695 ld [%i4 + 8], %l6 3696 ld [%i4 + 12], %l7 3697 3698 sllx %l4, 32, %o5 3699 or %l5, %o5, %o5 3700 sllx %l6, 32, %g1 3701 xor %o5, %g4, %g4 ! ^= rk[0] 3702 xor %g1, %g5, %g5 3703 .word 0x9db02304 !movxtod %g4,%f14 ! most significant 64 bits 3704 3705 sub %i0, %i1, %l5 ! %i0!=%i1 3706 and %i0, 7, %l0 3707 andn %i0, 7, %i0 3708 sll %l0, 3, %l0 3709 mov 64, %l1 3710 mov 0xff, %l3 3711 sub %l1, %l0, %l1 3712 and %i1, 7, %l2 3713 cmp %i2, 255 3714 movrnz %l2, 0, %l5 ! if ( %i1&7 || 3715 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 3716 brnz,pn %l5, .L192_ctr32_blk ! %i0==%i1) 3717 srl %l3, %l2, %l3 3718 3719 andcc %i2, 16, %g0 ! is number of blocks even? 3720 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 3721 bz %icc, .L192_ctr32_loop2x 3722 srlx %i2, 4, %i2 3723 .L192_ctr32_loop: 3724 ldx [%i0 + 0], %o0 3725 brz,pt %l0, 4f 3726 ldx [%i0 + 8], %o1 3727 3728 ldx [%i0 + 16], %o2 3729 sllx %o0, %l0, %o0 3730 srlx %o1, %l1, %g1 3731 sllx %o1, %l0, %o1 3732 or %g1, %o0, %o0 3733 srlx %o2, %l1, %o2 3734 or %o2, %o1, %o1 3735 4: 3736 xor %g5, %l7, %g1 ! ^= rk[0] 3737 add %l7, 1, %l7 3738 .word 0x85b02301 !movxtod %g1,%f2 3739 srl %l7, 0, %l7 ! clruw 3740 prefetch [%i1 + 63], 22 3741 prefetch [%i0 + 16+63], 20 3742 .word 0x88cc040e !aes_eround01 %f16,%f14,%f2,%f4 3743 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 3744 call _aes192_encrypt_1x+8 3745 add %i0, 16, %i0 3746 3747 .word 0x95b02308 !movxtod %o0,%f10 3748 .word 0x99b02309 !movxtod %o1,%f12 3749 .word 0x81b28d80 !fxor %f10,%f0,%f0 ! ^= inp 3750 .word 0x85b30d82 !fxor %f12,%f2,%f2 3751 3752 brnz,pn %l2, 2f 3753 sub %i2, 1, %i2 3754 3755 std %f0, [%i1 + 0] 3756 std %f2, [%i1 + 8] 3757 brnz,pt %i2, .L192_ctr32_loop2x 3758 add %i1, 16, %i1 3759 3760 ret 3761 restore 3762 3763 .align 16 3764 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3765 ! and ~3x deterioration 3766 ! in inp==out case 3767 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 3768 .word 0x8db00902 !faligndata %f0,%f2,%f6 3769 .word 0x91b08902 !faligndata %f2,%f2,%f8 3770 stda %f4, [%i1 + %l3]0xc0 ! partial store 3771 std %f6, [%i1 + 8] 3772 add %i1, 16, %i1 3773 orn %g0, %l3, %l3 3774 stda %f8, [%i1 + %l3]0xc0 ! partial store 3775 3776 brnz,pt %i2, .L192_ctr32_loop2x+4 3777 orn %g0, %l3, %l3 3778 3779 ret 3780 restore 3781 3782 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3783 .align 32 3784 .L192_ctr32_loop2x: 3785 ldx [%i0 + 0], %o0 3786 ldx [%i0 + 8], %o1 3787 ldx [%i0 + 16], %o2 3788 brz,pt %l0, 4f 3789 ldx [%i0 + 24], %o3 3790 3791 ldx [%i0 + 32], %o4 3792 sllx %o0, %l0, %o0 3793 srlx %o1, %l1, %g1 3794 or %g1, %o0, %o0 3795 sllx %o1, %l0, %o1 3796 srlx %o2, %l1, %g1 3797 or %g1, %o1, %o1 3798 sllx %o2, %l0, %o2 3799 srlx %o3, %l1, %g1 3800 or %g1, %o2, %o2 3801 sllx %o3, %l0, %o3 3802 srlx %o4, %l1, %o4 3803 or %o4, %o3, %o3 3804 4: 3805 xor %g5, %l7, %g1 ! ^= rk[0] 3806 add %l7, 1, %l7 3807 .word 0x85b02301 !movxtod %g1,%f2 3808 srl %l7, 0, %l7 ! clruw 3809 xor %g5, %l7, %g1 3810 add %l7, 1, %l7 3811 .word 0x8db02301 !movxtod %g1,%f6 3812 srl %l7, 0, %l7 ! clruw 3813 prefetch [%i1 + 63], 22 3814 prefetch [%i0 + 32+63], 20 3815 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 3816 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 3817 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 3818 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 3819 call _aes192_encrypt_2x+16 3820 add %i0, 32, %i0 3821 3822 .word 0x91b02308 !movxtod %o0,%f8 3823 .word 0x95b02309 !movxtod %o1,%f10 3824 .word 0x99b0230a !movxtod %o2,%f12 3825 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 3826 .word 0x91b0230b !movxtod %o3,%f8 3827 .word 0x85b28d82 !fxor %f10,%f2,%f2 3828 .word 0x89b30d84 !fxor %f12,%f4,%f4 3829 .word 0x8db20d86 !fxor %f8,%f6,%f6 3830 3831 brnz,pn %l2, 2f 3832 sub %i2, 2, %i2 3833 3834 std %f0, [%i1 + 0] 3835 std %f2, [%i1 + 8] 3836 std %f4, [%i1 + 16] 3837 std %f6, [%i1 + 24] 3838 brnz,pt %i2, .L192_ctr32_loop2x 3839 add %i1, 32, %i1 3840 3841 ret 3842 restore 3843 3844 .align 16 3845 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 3846 ! and ~3x deterioration 3847 ! in inp==out case 3848 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 3849 .word 0x81b00902 !faligndata %f0,%f2,%f0 3850 .word 0x85b08904 !faligndata %f2,%f4,%f2 3851 .word 0x89b10906 !faligndata %f4,%f6,%f4 3852 .word 0x8db18906 !faligndata %f6,%f6,%f6 3853 3854 stda %f8, [%i1 + %l3]0xc0 ! partial store 3855 std %f0, [%i1 + 8] 3856 std %f2, [%i1 + 16] 3857 std %f4, [%i1 + 24] 3858 add %i1, 32, %i1 3859 orn %g0, %l3, %l3 3860 stda %f6, [%i1 + %l3]0xc0 ! partial store 3861 3862 brnz,pt %i2, .L192_ctr32_loop2x+4 3863 orn %g0, %l3, %l3 3864 3865 ret 3866 restore 3867 3868 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3869 .align 32 3870 .L192_ctr32_blk: 3871 add %i1, %i2, %l5 3872 and %l5, 63, %l5 ! tail 3873 sub %i2, %l5, %i2 3874 add %l5, 15, %l5 ! round up to 16n 3875 srlx %i2, 4, %i2 3876 srl %l5, 4, %l5 3877 sub %i2, 1, %i2 3878 add %l5, 1, %l5 3879 3880 .L192_ctr32_blk_loop2x: 3881 ldx [%i0 + 0], %o0 3882 ldx [%i0 + 8], %o1 3883 ldx [%i0 + 16], %o2 3884 brz,pt %l0, 5f 3885 ldx [%i0 + 24], %o3 3886 3887 ldx [%i0 + 32], %o4 3888 sllx %o0, %l0, %o0 3889 srlx %o1, %l1, %g1 3890 or %g1, %o0, %o0 3891 sllx %o1, %l0, %o1 3892 srlx %o2, %l1, %g1 3893 or %g1, %o1, %o1 3894 sllx %o2, %l0, %o2 3895 srlx %o3, %l1, %g1 3896 or %g1, %o2, %o2 3897 sllx %o3, %l0, %o3 3898 srlx %o4, %l1, %o4 3899 or %o4, %o3, %o3 3900 5: 3901 xor %g5, %l7, %g1 ! ^= rk[0] 3902 add %l7, 1, %l7 3903 .word 0x85b02301 !movxtod %g1,%f2 3904 srl %l7, 0, %l7 ! clruw 3905 xor %g5, %l7, %g1 3906 add %l7, 1, %l7 3907 .word 0x8db02301 !movxtod %g1,%f6 3908 srl %l7, 0, %l7 ! clruw 3909 prefetch [%i0 + 32+63], 20 3910 .word 0x90cc040e !aes_eround01 %f16,%f14,%f2,%f8 3911 .word 0x84cc842e !aes_eround23 %f18,%f14,%f2,%f2 3912 .word 0x94cc0c0e !aes_eround01 %f16,%f14,%f6,%f10 3913 .word 0x8ccc8c2e !aes_eround23 %f18,%f14,%f6,%f6 3914 call _aes192_encrypt_2x+16 3915 add %i0, 32, %i0 3916 subcc %i2, 2, %i2 3917 3918 .word 0x91b02308 !movxtod %o0,%f8 3919 .word 0x95b02309 !movxtod %o1,%f10 3920 .word 0x99b0230a !movxtod %o2,%f12 3921 .word 0x81b20d80 !fxor %f8,%f0,%f0 ! ^= inp 3922 .word 0x91b0230b !movxtod %o3,%f8 3923 .word 0x85b28d82 !fxor %f10,%f2,%f2 3924 .word 0x89b30d84 !fxor %f12,%f4,%f4 3925 .word 0x8db20d86 !fxor %f8,%f6,%f6 3926 3927 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3928 add %i1, 8, %i1 3929 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3930 add %i1, 8, %i1 3931 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3932 add %i1, 8, %i1 3933 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 3934 bgu,pt SIZE_T_CC, .L192_ctr32_blk_loop2x 3935 add %i1, 8, %i1 3936 3937 add %l5, %i2, %i2 3938 andcc %i2, 1, %g0 ! is number of blocks even? 3939 membar #StoreLoad|#StoreStore 3940 bnz,pt %icc, .L192_ctr32_loop 3941 srl %i2, 0, %i2 3942 brnz,pn %i2, .L192_ctr32_loop2x 3943 nop 3944 3945 ret 3946 restore 3947 .type aes192_t4_ctr32_encrypt,#function 3948 .size aes192_t4_ctr32_encrypt,.-aes192_t4_ctr32_encrypt 3949 .globl aes192_t4_cbc_decrypt 3950 .align 32 3951 aes192_t4_cbc_decrypt: 3952 save %sp, -STACK_FRAME, %sp 3953 cmp %i2, 0 3954 be,pn SIZE_T_CC, .L192_cbc_dec_abort 3955 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 3956 sub %i0, %i1, %l5 ! %i0!=%i1 3957 ld [%i4 + 0], %f12 ! load ivec 3958 ld [%i4 + 4], %f13 3959 ld [%i4 + 8], %f14 3960 ld [%i4 + 12], %f15 3961 prefetch [%i0], 20 3962 prefetch [%i0 + 63], 20 3963 call _aes192_load_deckey 3964 and %i0, 7, %l0 3965 andn %i0, 7, %i0 3966 sll %l0, 3, %l0 3967 mov 64, %l1 3968 mov 0xff, %l3 3969 sub %l1, %l0, %l1 3970 and %i1, 7, %l2 3971 cmp %i2, 255 3972 movrnz %l2, 0, %l5 ! if ( %i1&7 || 3973 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 3974 brnz,pn %l5, .L192cbc_dec_blk ! %i0==%i1) 3975 srl %l3, %l2, %l3 3976 3977 andcc %i2, 16, %g0 ! is number of blocks even? 3978 srlx %i2, 4, %i2 3979 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 3980 bz %icc, .L192_cbc_dec_loop2x 3981 prefetch [%i1], 22 3982 .L192_cbc_dec_loop: 3983 ldx [%i0 + 0], %o0 3984 brz,pt %l0, 4f 3985 ldx [%i0 + 8], %o1 3986 3987 ldx [%i0 + 16], %o2 3988 sllx %o0, %l0, %o0 3989 srlx %o1, %l1, %g1 3990 sllx %o1, %l0, %o1 3991 or %g1, %o0, %o0 3992 srlx %o2, %l1, %o2 3993 or %o2, %o1, %o1 3994 4: 3995 xor %g4, %o0, %o2 ! ^= rk[0] 3996 xor %g5, %o1, %o3 3997 .word 0x81b0230a !movxtod %o2,%f0 3998 .word 0x85b0230b !movxtod %o3,%f2 3999 4000 prefetch [%i1 + 63], 22 4001 prefetch [%i0 + 16+63], 20 4002 call _aes192_decrypt_1x 4003 add %i0, 16, %i0 4004 4005 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4006 .word 0x85b38d82 !fxor %f14,%f2,%f2 4007 .word 0x99b02308 !movxtod %o0,%f12 4008 .word 0x9db02309 !movxtod %o1,%f14 4009 4010 brnz,pn %l2, 2f 4011 sub %i2, 1, %i2 4012 4013 std %f0, [%i1 + 0] 4014 std %f2, [%i1 + 8] 4015 brnz,pt %i2, .L192_cbc_dec_loop2x 4016 add %i1, 16, %i1 4017 st %f12, [%i4 + 0] 4018 st %f13, [%i4 + 4] 4019 st %f14, [%i4 + 8] 4020 st %f15, [%i4 + 12] 4021 .L192_cbc_dec_abort: 4022 ret 4023 restore 4024 4025 .align 16 4026 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4027 ! and ~3x deterioration 4028 ! in inp==out case 4029 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 4030 .word 0x8db00902 !faligndata %f0,%f2,%f6 4031 .word 0x91b08902 !faligndata %f2,%f2,%f8 4032 4033 stda %f4, [%i1 + %l3]0xc0 ! partial store 4034 std %f6, [%i1 + 8] 4035 add %i1, 16, %i1 4036 orn %g0, %l3, %l3 4037 stda %f8, [%i1 + %l3]0xc0 ! partial store 4038 4039 brnz,pt %i2, .L192_cbc_dec_loop2x+4 4040 orn %g0, %l3, %l3 4041 st %f12, [%i4 + 0] 4042 st %f13, [%i4 + 4] 4043 st %f14, [%i4 + 8] 4044 st %f15, [%i4 + 12] 4045 ret 4046 restore 4047 4048 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4049 .align 32 4050 .L192_cbc_dec_loop2x: 4051 ldx [%i0 + 0], %o0 4052 ldx [%i0 + 8], %o1 4053 ldx [%i0 + 16], %o2 4054 brz,pt %l0, 4f 4055 ldx [%i0 + 24], %o3 4056 4057 ldx [%i0 + 32], %o4 4058 sllx %o0, %l0, %o0 4059 srlx %o1, %l1, %g1 4060 or %g1, %o0, %o0 4061 sllx %o1, %l0, %o1 4062 srlx %o2, %l1, %g1 4063 or %g1, %o1, %o1 4064 sllx %o2, %l0, %o2 4065 srlx %o3, %l1, %g1 4066 or %g1, %o2, %o2 4067 sllx %o3, %l0, %o3 4068 srlx %o4, %l1, %o4 4069 or %o4, %o3, %o3 4070 4: 4071 xor %g4, %o0, %o4 ! ^= rk[0] 4072 xor %g5, %o1, %o5 4073 .word 0x81b0230c !movxtod %o4,%f0 4074 .word 0x85b0230d !movxtod %o5,%f2 4075 xor %g4, %o2, %o4 4076 xor %g5, %o3, %o5 4077 .word 0x89b0230c !movxtod %o4,%f4 4078 .word 0x8db0230d !movxtod %o5,%f6 4079 4080 prefetch [%i1 + 63], 22 4081 prefetch [%i0 + 32+63], 20 4082 call _aes192_decrypt_2x 4083 add %i0, 32, %i0 4084 4085 .word 0x91b02308 !movxtod %o0,%f8 4086 .word 0x95b02309 !movxtod %o1,%f10 4087 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4088 .word 0x85b38d82 !fxor %f14,%f2,%f2 4089 .word 0x99b0230a !movxtod %o2,%f12 4090 .word 0x9db0230b !movxtod %o3,%f14 4091 .word 0x89b20d84 !fxor %f8,%f4,%f4 4092 .word 0x8db28d86 !fxor %f10,%f6,%f6 4093 4094 brnz,pn %l2, 2f 4095 sub %i2, 2, %i2 4096 4097 std %f0, [%i1 + 0] 4098 std %f2, [%i1 + 8] 4099 std %f4, [%i1 + 16] 4100 std %f6, [%i1 + 24] 4101 brnz,pt %i2, .L192_cbc_dec_loop2x 4102 add %i1, 32, %i1 4103 st %f12, [%i4 + 0] 4104 st %f13, [%i4 + 4] 4105 st %f14, [%i4 + 8] 4106 st %f15, [%i4 + 12] 4107 ret 4108 restore 4109 4110 .align 16 4111 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4112 ! and ~3x deterioration 4113 ! in inp==out case 4114 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 4115 .word 0x81b00902 !faligndata %f0,%f2,%f0 4116 .word 0x85b08904 !faligndata %f2,%f4,%f2 4117 .word 0x89b10906 !faligndata %f4,%f6,%f4 4118 .word 0x8db18906 !faligndata %f6,%f6,%f6 4119 stda %f8, [%i1 + %l3]0xc0 ! partial store 4120 std %f0, [%i1 + 8] 4121 std %f2, [%i1 + 16] 4122 std %f4, [%i1 + 24] 4123 add %i1, 32, %i1 4124 orn %g0, %l3, %l3 4125 stda %f6, [%i1 + %l3]0xc0 ! partial store 4126 4127 brnz,pt %i2, .L192_cbc_dec_loop2x+4 4128 orn %g0, %l3, %l3 4129 st %f12, [%i4 + 0] 4130 st %f13, [%i4 + 4] 4131 st %f14, [%i4 + 8] 4132 st %f15, [%i4 + 12] 4133 ret 4134 restore 4135 4136 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4137 .align 32 4138 .L192cbc_dec_blk: 4139 add %i1, %i2, %l5 4140 and %l5, 63, %l5 ! tail 4141 sub %i2, %l5, %i2 4142 add %l5, 15, %l5 ! round up to 16n 4143 srlx %i2, 4, %i2 4144 srl %l5, 4, %l5 4145 sub %i2, 1, %i2 4146 add %l5, 1, %l5 4147 4148 .L192_cbc_dec_blk_loop2x: 4149 ldx [%i0 + 0], %o0 4150 ldx [%i0 + 8], %o1 4151 ldx [%i0 + 16], %o2 4152 brz,pt %l0, 5f 4153 ldx [%i0 + 24], %o3 4154 4155 ldx [%i0 + 32], %o4 4156 sllx %o0, %l0, %o0 4157 srlx %o1, %l1, %g1 4158 or %g1, %o0, %o0 4159 sllx %o1, %l0, %o1 4160 srlx %o2, %l1, %g1 4161 or %g1, %o1, %o1 4162 sllx %o2, %l0, %o2 4163 srlx %o3, %l1, %g1 4164 or %g1, %o2, %o2 4165 sllx %o3, %l0, %o3 4166 srlx %o4, %l1, %o4 4167 or %o4, %o3, %o3 4168 5: 4169 xor %g4, %o0, %o4 ! ^= rk[0] 4170 xor %g5, %o1, %o5 4171 .word 0x81b0230c !movxtod %o4,%f0 4172 .word 0x85b0230d !movxtod %o5,%f2 4173 xor %g4, %o2, %o4 4174 xor %g5, %o3, %o5 4175 .word 0x89b0230c !movxtod %o4,%f4 4176 .word 0x8db0230d !movxtod %o5,%f6 4177 4178 prefetch [%i0 + 32+63], 20 4179 call _aes192_decrypt_2x 4180 add %i0, 32, %i0 4181 subcc %i2, 2, %i2 4182 4183 .word 0x91b02308 !movxtod %o0,%f8 4184 .word 0x95b02309 !movxtod %o1,%f10 4185 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4186 .word 0x85b38d82 !fxor %f14,%f2,%f2 4187 .word 0x99b0230a !movxtod %o2,%f12 4188 .word 0x9db0230b !movxtod %o3,%f14 4189 .word 0x89b20d84 !fxor %f8,%f4,%f4 4190 .word 0x8db28d86 !fxor %f10,%f6,%f6 4191 4192 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4193 add %i1, 8, %i1 4194 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4195 add %i1, 8, %i1 4196 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4197 add %i1, 8, %i1 4198 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4199 bgu,pt SIZE_T_CC, .L192_cbc_dec_blk_loop2x 4200 add %i1, 8, %i1 4201 4202 add %l5, %i2, %i2 4203 andcc %i2, 1, %g0 ! is number of blocks even? 4204 membar #StoreLoad|#StoreStore 4205 bnz,pt %icc, .L192_cbc_dec_loop 4206 srl %i2, 0, %i2 4207 brnz,pn %i2, .L192_cbc_dec_loop2x 4208 nop 4209 st %f12, [%i4 + 0] ! write out ivec 4210 st %f13, [%i4 + 4] 4211 st %f14, [%i4 + 8] 4212 st %f15, [%i4 + 12] 4213 ret 4214 restore 4215 .type aes192_t4_cbc_decrypt,#function 4216 .size aes192_t4_cbc_decrypt,.-aes192_t4_cbc_decrypt 4217 .globl aes256_t4_cbc_decrypt 4218 .align 32 4219 aes256_t4_cbc_decrypt: 4220 save %sp, -STACK_FRAME, %sp 4221 cmp %i2, 0 4222 be,pn SIZE_T_CC, .L256_cbc_dec_abort 4223 srln %i2, 0, %i2 ! needed on v8+, "nop" on v9 4224 sub %i0, %i1, %l5 ! %i0!=%i1 4225 ld [%i4 + 0], %f12 ! load ivec 4226 ld [%i4 + 4], %f13 4227 ld [%i4 + 8], %f14 4228 ld [%i4 + 12], %f15 4229 prefetch [%i0], 20 4230 prefetch [%i0 + 63], 20 4231 call _aes256_load_deckey 4232 and %i0, 7, %l0 4233 andn %i0, 7, %i0 4234 sll %l0, 3, %l0 4235 mov 64, %l1 4236 mov 0xff, %l3 4237 sub %l1, %l0, %l1 4238 and %i1, 7, %l2 4239 cmp %i2, 255 4240 movrnz %l2, 0, %l5 ! if ( %i1&7 || 4241 movleu SIZE_T_CC, 0, %l5 ! %i2<256 || 4242 brnz,pn %l5, .L256cbc_dec_blk ! %i0==%i1) 4243 srl %l3, %l2, %l3 4244 4245 andcc %i2, 16, %g0 ! is number of blocks even? 4246 srlx %i2, 4, %i2 4247 .word 0xb3b64340 !alignaddrl %i1,%g0,%i1 4248 bz %icc, .L256_cbc_dec_loop2x 4249 prefetch [%i1], 22 4250 .L256_cbc_dec_loop: 4251 ldx [%i0 + 0], %o0 4252 brz,pt %l0, 4f 4253 ldx [%i0 + 8], %o1 4254 4255 ldx [%i0 + 16], %o2 4256 sllx %o0, %l0, %o0 4257 srlx %o1, %l1, %g1 4258 sllx %o1, %l0, %o1 4259 or %g1, %o0, %o0 4260 srlx %o2, %l1, %o2 4261 or %o2, %o1, %o1 4262 4: 4263 xor %g4, %o0, %o2 ! ^= rk[0] 4264 xor %g5, %o1, %o3 4265 .word 0x81b0230a !movxtod %o2,%f0 4266 .word 0x85b0230b !movxtod %o3,%f2 4267 4268 prefetch [%i1 + 63], 22 4269 prefetch [%i0 + 16+63], 20 4270 call _aes256_decrypt_1x 4271 add %i0, 16, %i0 4272 4273 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4274 .word 0x85b38d82 !fxor %f14,%f2,%f2 4275 .word 0x99b02308 !movxtod %o0,%f12 4276 .word 0x9db02309 !movxtod %o1,%f14 4277 4278 brnz,pn %l2, 2f 4279 sub %i2, 1, %i2 4280 4281 std %f0, [%i1 + 0] 4282 std %f2, [%i1 + 8] 4283 brnz,pt %i2, .L256_cbc_dec_loop2x 4284 add %i1, 16, %i1 4285 st %f12, [%i4 + 0] 4286 st %f13, [%i4 + 4] 4287 st %f14, [%i4 + 8] 4288 st %f15, [%i4 + 12] 4289 .L256_cbc_dec_abort: 4290 ret 4291 restore 4292 4293 .align 16 4294 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4295 ! and ~3x deterioration 4296 ! in inp==out case 4297 .word 0x89b00900 !faligndata %f0,%f0,%f4 ! handle unaligned output 4298 .word 0x8db00902 !faligndata %f0,%f2,%f6 4299 .word 0x91b08902 !faligndata %f2,%f2,%f8 4300 4301 stda %f4, [%i1 + %l3]0xc0 ! partial store 4302 std %f6, [%i1 + 8] 4303 add %i1, 16, %i1 4304 orn %g0, %l3, %l3 4305 stda %f8, [%i1 + %l3]0xc0 ! partial store 4306 4307 brnz,pt %i2, .L256_cbc_dec_loop2x+4 4308 orn %g0, %l3, %l3 4309 st %f12, [%i4 + 0] 4310 st %f13, [%i4 + 4] 4311 st %f14, [%i4 + 8] 4312 st %f15, [%i4 + 12] 4313 ret 4314 restore 4315 4316 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4317 .align 32 4318 .L256_cbc_dec_loop2x: 4319 ldx [%i0 + 0], %o0 4320 ldx [%i0 + 8], %o1 4321 ldx [%i0 + 16], %o2 4322 brz,pt %l0, 4f 4323 ldx [%i0 + 24], %o3 4324 4325 ldx [%i0 + 32], %o4 4326 sllx %o0, %l0, %o0 4327 srlx %o1, %l1, %g1 4328 or %g1, %o0, %o0 4329 sllx %o1, %l0, %o1 4330 srlx %o2, %l1, %g1 4331 or %g1, %o1, %o1 4332 sllx %o2, %l0, %o2 4333 srlx %o3, %l1, %g1 4334 or %g1, %o2, %o2 4335 sllx %o3, %l0, %o3 4336 srlx %o4, %l1, %o4 4337 or %o4, %o3, %o3 4338 4: 4339 xor %g4, %o0, %o4 ! ^= rk[0] 4340 xor %g5, %o1, %o5 4341 .word 0x81b0230c !movxtod %o4,%f0 4342 .word 0x85b0230d !movxtod %o5,%f2 4343 xor %g4, %o2, %o4 4344 xor %g5, %o3, %o5 4345 .word 0x89b0230c !movxtod %o4,%f4 4346 .word 0x8db0230d !movxtod %o5,%f6 4347 4348 prefetch [%i1 + 63], 22 4349 prefetch [%i0 + 32+63], 20 4350 call _aes256_decrypt_2x 4351 add %i0, 32, %i0 4352 4353 .word 0x91b02308 !movxtod %o0,%f8 4354 .word 0x95b02309 !movxtod %o1,%f10 4355 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4356 .word 0x85b38d82 !fxor %f14,%f2,%f2 4357 .word 0x99b0230a !movxtod %o2,%f12 4358 .word 0x9db0230b !movxtod %o3,%f14 4359 .word 0x89b20d84 !fxor %f8,%f4,%f4 4360 .word 0x8db28d86 !fxor %f10,%f6,%f6 4361 4362 brnz,pn %l2, 2f 4363 sub %i2, 2, %i2 4364 4365 std %f0, [%i1 + 0] 4366 std %f2, [%i1 + 8] 4367 std %f4, [%i1 + 16] 4368 std %f6, [%i1 + 24] 4369 brnz,pt %i2, .L256_cbc_dec_loop2x 4370 add %i1, 32, %i1 4371 st %f12, [%i4 + 0] 4372 st %f13, [%i4 + 4] 4373 st %f14, [%i4 + 8] 4374 st %f15, [%i4 + 12] 4375 ret 4376 restore 4377 4378 .align 16 4379 2: ldxa [%i0]0x82, %o0 ! avoid read-after-write hazard 4380 ! and ~3x deterioration 4381 ! in inp==out case 4382 .word 0x91b00900 !faligndata %f0,%f0,%f8 ! handle unaligned output 4383 .word 0x81b00902 !faligndata %f0,%f2,%f0 4384 .word 0x85b08904 !faligndata %f2,%f4,%f2 4385 .word 0x89b10906 !faligndata %f4,%f6,%f4 4386 .word 0x8db18906 !faligndata %f6,%f6,%f6 4387 stda %f8, [%i1 + %l3]0xc0 ! partial store 4388 std %f0, [%i1 + 8] 4389 std %f2, [%i1 + 16] 4390 std %f4, [%i1 + 24] 4391 add %i1, 32, %i1 4392 orn %g0, %l3, %l3 4393 stda %f6, [%i1 + %l3]0xc0 ! partial store 4394 4395 brnz,pt %i2, .L256_cbc_dec_loop2x+4 4396 orn %g0, %l3, %l3 4397 st %f12, [%i4 + 0] 4398 st %f13, [%i4 + 4] 4399 st %f14, [%i4 + 8] 4400 st %f15, [%i4 + 12] 4401 ret 4402 restore 4403 4404 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4405 .align 32 4406 .L256cbc_dec_blk: 4407 add %i1, %i2, %l5 4408 and %l5, 63, %l5 ! tail 4409 sub %i2, %l5, %i2 4410 add %l5, 15, %l5 ! round up to 16n 4411 srlx %i2, 4, %i2 4412 srl %l5, 4, %l5 4413 sub %i2, 1, %i2 4414 add %l5, 1, %l5 4415 4416 .L256_cbc_dec_blk_loop2x: 4417 ldx [%i0 + 0], %o0 4418 ldx [%i0 + 8], %o1 4419 ldx [%i0 + 16], %o2 4420 brz,pt %l0, 5f 4421 ldx [%i0 + 24], %o3 4422 4423 ldx [%i0 + 32], %o4 4424 sllx %o0, %l0, %o0 4425 srlx %o1, %l1, %g1 4426 or %g1, %o0, %o0 4427 sllx %o1, %l0, %o1 4428 srlx %o2, %l1, %g1 4429 or %g1, %o1, %o1 4430 sllx %o2, %l0, %o2 4431 srlx %o3, %l1, %g1 4432 or %g1, %o2, %o2 4433 sllx %o3, %l0, %o3 4434 srlx %o4, %l1, %o4 4435 or %o4, %o3, %o3 4436 5: 4437 xor %g4, %o0, %o4 ! ^= rk[0] 4438 xor %g5, %o1, %o5 4439 .word 0x81b0230c !movxtod %o4,%f0 4440 .word 0x85b0230d !movxtod %o5,%f2 4441 xor %g4, %o2, %o4 4442 xor %g5, %o3, %o5 4443 .word 0x89b0230c !movxtod %o4,%f4 4444 .word 0x8db0230d !movxtod %o5,%f6 4445 4446 prefetch [%i0 + 32+63], 20 4447 call _aes256_decrypt_2x 4448 add %i0, 32, %i0 4449 subcc %i2, 2, %i2 4450 4451 .word 0x91b02308 !movxtod %o0,%f8 4452 .word 0x95b02309 !movxtod %o1,%f10 4453 .word 0x81b30d80 !fxor %f12,%f0,%f0 ! ^= ivec 4454 .word 0x85b38d82 !fxor %f14,%f2,%f2 4455 .word 0x99b0230a !movxtod %o2,%f12 4456 .word 0x9db0230b !movxtod %o3,%f14 4457 .word 0x89b20d84 !fxor %f8,%f4,%f4 4458 .word 0x8db28d86 !fxor %f10,%f6,%f6 4459 4460 stda %f0, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4461 add %i1, 8, %i1 4462 stda %f2, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4463 add %i1, 8, %i1 4464 stda %f4, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4465 add %i1, 8, %i1 4466 stda %f6, [%i1]0xe2 ! ASI_BLK_INIT, T4-specific 4467 bgu,pt SIZE_T_CC, .L256_cbc_dec_blk_loop2x 4468 add %i1, 8, %i1 4469 4470 add %l5, %i2, %i2 4471 andcc %i2, 1, %g0 ! is number of blocks even? 4472 membar #StoreLoad|#StoreStore 4473 bnz,pt %icc, .L256_cbc_dec_loop 4474 srl %i2, 0, %i2 4475 brnz,pn %i2, .L256_cbc_dec_loop2x 4476 nop 4477 st %f12, [%i4 + 0] ! write out ivec 4478 st %f13, [%i4 + 4] 4479 st %f14, [%i4 + 8] 4480 st %f15, [%i4 + 12] 4481 ret 4482 restore 4483 .type aes256_t4_cbc_decrypt,#function 4484 .size aes256_t4_cbc_decrypt,.-aes256_t4_cbc_decrypt 4485 .align 32 4486 _aes256_decrypt_1x: 4487 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 4488 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4489 ldd [%i3 + 208], %f16 4490 ldd [%i3 + 216], %f18 4491 .word 0x80cd0444 !aes_dround01 %f20,%f4,%f2,%f0 4492 .word 0x84cd8464 !aes_dround23 %f22,%f4,%f2,%f2 4493 ldd [%i3 + 224], %f20 4494 ldd [%i3 + 232], %f22 4495 .word 0x88ce0440 !aes_dround01 %f24,%f0,%f2,%f4 4496 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4497 .word 0x80cf0444 !aes_dround01 %f28,%f4,%f2,%f0 4498 .word 0x84cf8464 !aes_dround23 %f30,%f4,%f2,%f2 4499 .word 0x88c84440 !aes_dround01 %f32,%f0,%f2,%f4 4500 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4501 .word 0x80c94444 !aes_dround01 %f36,%f4,%f2,%f0 4502 .word 0x84c9c464 !aes_dround23 %f38,%f4,%f2,%f2 4503 .word 0x88ca4440 !aes_dround01 %f40,%f0,%f2,%f4 4504 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4505 .word 0x80cb4444 !aes_dround01 %f44,%f4,%f2,%f0 4506 .word 0x84cbc464 !aes_dround23 %f46,%f4,%f2,%f2 4507 .word 0x88cc4440 !aes_dround01 %f48,%f0,%f2,%f4 4508 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4509 .word 0x80cd4444 !aes_dround01 %f52,%f4,%f2,%f0 4510 .word 0x84cdc464 !aes_dround23 %f54,%f4,%f2,%f2 4511 .word 0x88ce4440 !aes_dround01 %f56,%f0,%f2,%f4 4512 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4513 .word 0x80cf4444 !aes_dround01 %f60,%f4,%f2,%f0 4514 .word 0x84cfc464 !aes_dround23 %f62,%f4,%f2,%f2 4515 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 4516 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4517 ldd [%i3 + 16], %f16 4518 ldd [%i3 + 24], %f18 4519 .word 0x80cd04c4 !aes_dround01_l %f20,%f4,%f2,%f0 4520 .word 0x84cd84e4 !aes_dround23_l %f22,%f4,%f2,%f2 4521 ldd [%i3 + 32], %f20 4522 retl 4523 ldd [%i3 + 40], %f22 4524 .type _aes256_decrypt_1x,#function 4525 .size _aes256_decrypt_1x,.-_aes256_decrypt_1x 4526 4527 .align 32 4528 _aes256_decrypt_2x: 4529 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 4530 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4531 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 4532 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 4533 ldd [%i3 + 208], %f16 4534 ldd [%i3 + 216], %f18 4535 .word 0x80cd0448 !aes_dround01 %f20,%f8,%f2,%f0 4536 .word 0x84cd8468 !aes_dround23 %f22,%f8,%f2,%f2 4537 .word 0x88cd0c4a !aes_dround01 %f20,%f10,%f6,%f4 4538 .word 0x8ccd8c6a !aes_dround23 %f22,%f10,%f6,%f6 4539 ldd [%i3 + 224], %f20 4540 ldd [%i3 + 232], %f22 4541 .word 0x90ce0440 !aes_dround01 %f24,%f0,%f2,%f8 4542 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4543 .word 0x94ce0c44 !aes_dround01 %f24,%f4,%f6,%f10 4544 .word 0x8cce8c64 !aes_dround23 %f26,%f4,%f6,%f6 4545 .word 0x80cf0448 !aes_dround01 %f28,%f8,%f2,%f0 4546 .word 0x84cf8468 !aes_dround23 %f30,%f8,%f2,%f2 4547 .word 0x88cf0c4a !aes_dround01 %f28,%f10,%f6,%f4 4548 .word 0x8ccf8c6a !aes_dround23 %f30,%f10,%f6,%f6 4549 .word 0x90c84440 !aes_dround01 %f32,%f0,%f2,%f8 4550 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4551 .word 0x94c84c44 !aes_dround01 %f32,%f4,%f6,%f10 4552 .word 0x8cc8cc64 !aes_dround23 %f34,%f4,%f6,%f6 4553 .word 0x80c94448 !aes_dround01 %f36,%f8,%f2,%f0 4554 .word 0x84c9c468 !aes_dround23 %f38,%f8,%f2,%f2 4555 .word 0x88c94c4a !aes_dround01 %f36,%f10,%f6,%f4 4556 .word 0x8cc9cc6a !aes_dround23 %f38,%f10,%f6,%f6 4557 .word 0x90ca4440 !aes_dround01 %f40,%f0,%f2,%f8 4558 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4559 .word 0x94ca4c44 !aes_dround01 %f40,%f4,%f6,%f10 4560 .word 0x8ccacc64 !aes_dround23 %f42,%f4,%f6,%f6 4561 .word 0x80cb4448 !aes_dround01 %f44,%f8,%f2,%f0 4562 .word 0x84cbc468 !aes_dround23 %f46,%f8,%f2,%f2 4563 .word 0x88cb4c4a !aes_dround01 %f44,%f10,%f6,%f4 4564 .word 0x8ccbcc6a !aes_dround23 %f46,%f10,%f6,%f6 4565 .word 0x90cc4440 !aes_dround01 %f48,%f0,%f2,%f8 4566 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4567 .word 0x94cc4c44 !aes_dround01 %f48,%f4,%f6,%f10 4568 .word 0x8ccccc64 !aes_dround23 %f50,%f4,%f6,%f6 4569 .word 0x80cd4448 !aes_dround01 %f52,%f8,%f2,%f0 4570 .word 0x84cdc468 !aes_dround23 %f54,%f8,%f2,%f2 4571 .word 0x88cd4c4a !aes_dround01 %f52,%f10,%f6,%f4 4572 .word 0x8ccdcc6a !aes_dround23 %f54,%f10,%f6,%f6 4573 .word 0x90ce4440 !aes_dround01 %f56,%f0,%f2,%f8 4574 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4575 .word 0x94ce4c44 !aes_dround01 %f56,%f4,%f6,%f10 4576 .word 0x8ccecc64 !aes_dround23 %f58,%f4,%f6,%f6 4577 .word 0x80cf4448 !aes_dround01 %f60,%f8,%f2,%f0 4578 .word 0x84cfc468 !aes_dround23 %f62,%f8,%f2,%f2 4579 .word 0x88cf4c4a !aes_dround01 %f60,%f10,%f6,%f4 4580 .word 0x8ccfcc6a !aes_dround23 %f62,%f10,%f6,%f6 4581 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 4582 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4583 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 4584 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 4585 ldd [%i3 + 16], %f16 4586 ldd [%i3 + 24], %f18 4587 .word 0x80cd04c8 !aes_dround01_l %f20,%f8,%f2,%f0 4588 .word 0x84cd84e8 !aes_dround23_l %f22,%f8,%f2,%f2 4589 .word 0x88cd0cca !aes_dround01_l %f20,%f10,%f6,%f4 4590 .word 0x8ccd8cea !aes_dround23_l %f22,%f10,%f6,%f6 4591 ldd [%i3 + 32], %f20 4592 retl 4593 ldd [%i3 + 40], %f22 4594 .type _aes256_decrypt_2x,#function 4595 .size _aes256_decrypt_2x,.-_aes256_decrypt_2x 4596 4597 .align 32 4598 _aes192_decrypt_1x: 4599 .word 0x88cc0440 !aes_dround01 %f16,%f0,%f2,%f4 4600 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4601 .word 0x80cd0444 !aes_dround01 %f20,%f4,%f2,%f0 4602 .word 0x84cd8464 !aes_dround23 %f22,%f4,%f2,%f2 4603 .word 0x88ce0440 !aes_dround01 %f24,%f0,%f2,%f4 4604 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4605 .word 0x80cf0444 !aes_dround01 %f28,%f4,%f2,%f0 4606 .word 0x84cf8464 !aes_dround23 %f30,%f4,%f2,%f2 4607 .word 0x88c84440 !aes_dround01 %f32,%f0,%f2,%f4 4608 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4609 .word 0x80c94444 !aes_dround01 %f36,%f4,%f2,%f0 4610 .word 0x84c9c464 !aes_dround23 %f38,%f4,%f2,%f2 4611 .word 0x88ca4440 !aes_dround01 %f40,%f0,%f2,%f4 4612 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4613 .word 0x80cb4444 !aes_dround01 %f44,%f4,%f2,%f0 4614 .word 0x84cbc464 !aes_dround23 %f46,%f4,%f2,%f2 4615 .word 0x88cc4440 !aes_dround01 %f48,%f0,%f2,%f4 4616 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4617 .word 0x80cd4444 !aes_dround01 %f52,%f4,%f2,%f0 4618 .word 0x84cdc464 !aes_dround23 %f54,%f4,%f2,%f2 4619 .word 0x88ce4440 !aes_dround01 %f56,%f0,%f2,%f4 4620 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4621 .word 0x80cf44c4 !aes_dround01_l %f60,%f4,%f2,%f0 4622 retl 4623 .word 0x84cfc4e4 !aes_dround23_l %f62,%f4,%f2,%f2 4624 .type _aes192_decrypt_1x,#function 4625 .size _aes192_decrypt_1x,.-_aes192_decrypt_1x 4626 4627 .align 32 4628 _aes192_decrypt_2x: 4629 .word 0x90cc0440 !aes_dround01 %f16,%f0,%f2,%f8 4630 .word 0x84cc8460 !aes_dround23 %f18,%f0,%f2,%f2 4631 .word 0x94cc0c44 !aes_dround01 %f16,%f4,%f6,%f10 4632 .word 0x8ccc8c64 !aes_dround23 %f18,%f4,%f6,%f6 4633 .word 0x80cd0448 !aes_dround01 %f20,%f8,%f2,%f0 4634 .word 0x84cd8468 !aes_dround23 %f22,%f8,%f2,%f2 4635 .word 0x88cd0c4a !aes_dround01 %f20,%f10,%f6,%f4 4636 .word 0x8ccd8c6a !aes_dround23 %f22,%f10,%f6,%f6 4637 .word 0x90ce0440 !aes_dround01 %f24,%f0,%f2,%f8 4638 .word 0x84ce8460 !aes_dround23 %f26,%f0,%f2,%f2 4639 .word 0x94ce0c44 !aes_dround01 %f24,%f4,%f6,%f10 4640 .word 0x8cce8c64 !aes_dround23 %f26,%f4,%f6,%f6 4641 .word 0x80cf0448 !aes_dround01 %f28,%f8,%f2,%f0 4642 .word 0x84cf8468 !aes_dround23 %f30,%f8,%f2,%f2 4643 .word 0x88cf0c4a !aes_dround01 %f28,%f10,%f6,%f4 4644 .word 0x8ccf8c6a !aes_dround23 %f30,%f10,%f6,%f6 4645 .word 0x90c84440 !aes_dround01 %f32,%f0,%f2,%f8 4646 .word 0x84c8c460 !aes_dround23 %f34,%f0,%f2,%f2 4647 .word 0x94c84c44 !aes_dround01 %f32,%f4,%f6,%f10 4648 .word 0x8cc8cc64 !aes_dround23 %f34,%f4,%f6,%f6 4649 .word 0x80c94448 !aes_dround01 %f36,%f8,%f2,%f0 4650 .word 0x84c9c468 !aes_dround23 %f38,%f8,%f2,%f2 4651 .word 0x88c94c4a !aes_dround01 %f36,%f10,%f6,%f4 4652 .word 0x8cc9cc6a !aes_dround23 %f38,%f10,%f6,%f6 4653 .word 0x90ca4440 !aes_dround01 %f40,%f0,%f2,%f8 4654 .word 0x84cac460 !aes_dround23 %f42,%f0,%f2,%f2 4655 .word 0x94ca4c44 !aes_dround01 %f40,%f4,%f6,%f10 4656 .word 0x8ccacc64 !aes_dround23 %f42,%f4,%f6,%f6 4657 .word 0x80cb4448 !aes_dround01 %f44,%f8,%f2,%f0 4658 .word 0x84cbc468 !aes_dround23 %f46,%f8,%f2,%f2 4659 .word 0x88cb4c4a !aes_dround01 %f44,%f10,%f6,%f4 4660 .word 0x8ccbcc6a !aes_dround23 %f46,%f10,%f6,%f6 4661 .word 0x90cc4440 !aes_dround01 %f48,%f0,%f2,%f8 4662 .word 0x84ccc460 !aes_dround23 %f50,%f0,%f2,%f2 4663 .word 0x94cc4c44 !aes_dround01 %f48,%f4,%f6,%f10 4664 .word 0x8ccccc64 !aes_dround23 %f50,%f4,%f6,%f6 4665 .word 0x80cd4448 !aes_dround01 %f52,%f8,%f2,%f0 4666 .word 0x84cdc468 !aes_dround23 %f54,%f8,%f2,%f2 4667 .word 0x88cd4c4a !aes_dround01 %f52,%f10,%f6,%f4 4668 .word 0x8ccdcc6a !aes_dround23 %f54,%f10,%f6,%f6 4669 .word 0x90ce4440 !aes_dround01 %f56,%f0,%f2,%f8 4670 .word 0x84cec460 !aes_dround23 %f58,%f0,%f2,%f2 4671 .word 0x94ce4c44 !aes_dround01 %f56,%f4,%f6,%f10 4672 .word 0x8ccecc64 !aes_dround23 %f58,%f4,%f6,%f6 4673 .word 0x80cf44c8 !aes_dround01_l %f60,%f8,%f2,%f0 4674 .word 0x84cfc4e8 !aes_dround23_l %f62,%f8,%f2,%f2 4675 .word 0x88cf4cca !aes_dround01_l %f60,%f10,%f6,%f4 4676 retl 4677 .word 0x8ccfccea !aes_dround23_l %f62,%f10,%f6,%f6 4678 .type _aes192_decrypt_2x,#function 4679 .size _aes192_decrypt_2x,.-_aes192_decrypt_2x 4680 .asciz "AES for SPARC T4, David S. Miller, Andy Polyakov" 4681 .align 4 4682
Indexes created Fri Mar 06 19:51:54 UTC 2026