| OLD | NEW |
|---|
| 1 default rel | 1 default rel |
| 2 %define XMMWORD | 2 %define XMMWORD |
| 3 %define YMMWORD | 3 %define YMMWORD |
| 4 %define ZMMWORD | 4 %define ZMMWORD |
| 5 section .text code align=64 | 5 section .text code align=64 |
| 6 | 6 |
| 7 | 7 |
| 8 global rsaz_1024_sqr_avx2 |
| 9 |
| 10 ALIGN 64 |
| 11 rsaz_1024_sqr_avx2: |
| 12 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 13 mov QWORD[16+rsp],rsi |
| 14 mov rax,rsp |
| 15 $L$SEH_begin_rsaz_1024_sqr_avx2: |
| 16 mov rdi,rcx |
| 17 mov rsi,rdx |
| 18 mov rdx,r8 |
| 19 mov rcx,r9 |
| 20 mov r8,QWORD[40+rsp] |
| 21 |
| 22 |
| 23 |
| 24 lea rax,[rsp] |
| 25 |
| 26 push rbx |
| 27 |
| 28 push rbp |
| 29 |
| 30 push r12 |
| 31 |
| 32 push r13 |
| 33 |
| 34 push r14 |
| 35 |
| 36 push r15 |
| 37 |
| 38 vzeroupper |
| 39 lea rsp,[((-168))+rsp] |
| 40 vmovaps XMMWORD[(-216)+rax],xmm6 |
| 41 vmovaps XMMWORD[(-200)+rax],xmm7 |
| 42 vmovaps XMMWORD[(-184)+rax],xmm8 |
| 43 vmovaps XMMWORD[(-168)+rax],xmm9 |
| 44 vmovaps XMMWORD[(-152)+rax],xmm10 |
| 45 vmovaps XMMWORD[(-136)+rax],xmm11 |
| 46 vmovaps XMMWORD[(-120)+rax],xmm12 |
| 47 vmovaps XMMWORD[(-104)+rax],xmm13 |
| 48 vmovaps XMMWORD[(-88)+rax],xmm14 |
| 49 vmovaps XMMWORD[(-72)+rax],xmm15 |
| 50 $L$sqr_1024_body: |
| 51 mov rbp,rax |
| 52 |
| 53 mov r13,rdx |
| 54 sub rsp,832 |
| 55 mov r15,r13 |
| 56 sub rdi,-128 |
| 57 sub rsi,-128 |
| 58 sub r13,-128 |
| 59 |
| 60 and r15,4095 |
| 61 add r15,32*10 |
| 62 shr r15,12 |
| 63 vpxor ymm9,ymm9,ymm9 |
| 64 jz NEAR $L$sqr_1024_no_n_copy |
| 65 |
| 66 |
| 67 |
| 68 |
| 69 |
| 70 sub rsp,32*10 |
| 71 vmovdqu ymm0,YMMWORD[((0-128))+r13] |
| 72 and rsp,-2048 |
| 73 vmovdqu ymm1,YMMWORD[((32-128))+r13] |
| 74 vmovdqu ymm2,YMMWORD[((64-128))+r13] |
| 75 vmovdqu ymm3,YMMWORD[((96-128))+r13] |
| 76 vmovdqu ymm4,YMMWORD[((128-128))+r13] |
| 77 vmovdqu ymm5,YMMWORD[((160-128))+r13] |
| 78 vmovdqu ymm6,YMMWORD[((192-128))+r13] |
| 79 vmovdqu ymm7,YMMWORD[((224-128))+r13] |
| 80 vmovdqu ymm8,YMMWORD[((256-128))+r13] |
| 81 lea r13,[((832+128))+rsp] |
| 82 vmovdqu YMMWORD[(0-128)+r13],ymm0 |
| 83 vmovdqu YMMWORD[(32-128)+r13],ymm1 |
| 84 vmovdqu YMMWORD[(64-128)+r13],ymm2 |
| 85 vmovdqu YMMWORD[(96-128)+r13],ymm3 |
| 86 vmovdqu YMMWORD[(128-128)+r13],ymm4 |
| 87 vmovdqu YMMWORD[(160-128)+r13],ymm5 |
| 88 vmovdqu YMMWORD[(192-128)+r13],ymm6 |
| 89 vmovdqu YMMWORD[(224-128)+r13],ymm7 |
| 90 vmovdqu YMMWORD[(256-128)+r13],ymm8 |
| 91 vmovdqu YMMWORD[(288-128)+r13],ymm9 |
| 92 |
| 93 $L$sqr_1024_no_n_copy: |
| 94 and rsp,-1024 |
| 95 |
| 96 vmovdqu ymm1,YMMWORD[((32-128))+rsi] |
| 97 vmovdqu ymm2,YMMWORD[((64-128))+rsi] |
| 98 vmovdqu ymm3,YMMWORD[((96-128))+rsi] |
| 99 vmovdqu ymm4,YMMWORD[((128-128))+rsi] |
| 100 vmovdqu ymm5,YMMWORD[((160-128))+rsi] |
| 101 vmovdqu ymm6,YMMWORD[((192-128))+rsi] |
| 102 vmovdqu ymm7,YMMWORD[((224-128))+rsi] |
| 103 vmovdqu ymm8,YMMWORD[((256-128))+rsi] |
| 104 |
| 105 lea rbx,[192+rsp] |
| 106 vpbroadcastq ymm15,QWORD[$L$and_mask] |
| 107 jmp NEAR $L$OOP_GRANDE_SQR_1024 |
| 108 |
| 109 ALIGN 32 |
| 110 $L$OOP_GRANDE_SQR_1024: |
| 111 lea r9,[((576+128))+rsp] |
| 112 lea r12,[448+rsp] |
| 113 |
| 114 |
| 115 |
| 116 |
| 117 vpaddq ymm1,ymm1,ymm1 |
| 118 vpbroadcastq ymm10,QWORD[((0-128))+rsi] |
| 119 vpaddq ymm2,ymm2,ymm2 |
| 120 vmovdqa YMMWORD[(0-128)+r9],ymm1 |
| 121 vpaddq ymm3,ymm3,ymm3 |
| 122 vmovdqa YMMWORD[(32-128)+r9],ymm2 |
| 123 vpaddq ymm4,ymm4,ymm4 |
| 124 vmovdqa YMMWORD[(64-128)+r9],ymm3 |
| 125 vpaddq ymm5,ymm5,ymm5 |
| 126 vmovdqa YMMWORD[(96-128)+r9],ymm4 |
| 127 vpaddq ymm6,ymm6,ymm6 |
| 128 vmovdqa YMMWORD[(128-128)+r9],ymm5 |
| 129 vpaddq ymm7,ymm7,ymm7 |
| 130 vmovdqa YMMWORD[(160-128)+r9],ymm6 |
| 131 vpaddq ymm8,ymm8,ymm8 |
| 132 vmovdqa YMMWORD[(192-128)+r9],ymm7 |
| 133 vpxor ymm9,ymm9,ymm9 |
| 134 vmovdqa YMMWORD[(224-128)+r9],ymm8 |
| 135 |
| 136 vpmuludq ymm0,ymm10,YMMWORD[((0-128))+rsi] |
| 137 vpbroadcastq ymm11,QWORD[((32-128))+rsi] |
| 138 vmovdqu YMMWORD[(288-192)+rbx],ymm9 |
| 139 vpmuludq ymm1,ymm1,ymm10 |
| 140 vmovdqu YMMWORD[(320-448)+r12],ymm9 |
| 141 vpmuludq ymm2,ymm2,ymm10 |
| 142 vmovdqu YMMWORD[(352-448)+r12],ymm9 |
| 143 vpmuludq ymm3,ymm3,ymm10 |
| 144 vmovdqu YMMWORD[(384-448)+r12],ymm9 |
| 145 vpmuludq ymm4,ymm4,ymm10 |
| 146 vmovdqu YMMWORD[(416-448)+r12],ymm9 |
| 147 vpmuludq ymm5,ymm5,ymm10 |
| 148 vmovdqu YMMWORD[(448-448)+r12],ymm9 |
| 149 vpmuludq ymm6,ymm6,ymm10 |
| 150 vmovdqu YMMWORD[(480-448)+r12],ymm9 |
| 151 vpmuludq ymm7,ymm7,ymm10 |
| 152 vmovdqu YMMWORD[(512-448)+r12],ymm9 |
| 153 vpmuludq ymm8,ymm8,ymm10 |
| 154 vpbroadcastq ymm10,QWORD[((64-128))+rsi] |
| 155 vmovdqu YMMWORD[(544-448)+r12],ymm9 |
| 156 |
| 157 mov r15,rsi |
| 158 mov r14d,4 |
| 159 jmp NEAR $L$sqr_entry_1024 |
| 160 ALIGN 32 |
| 161 $L$OOP_SQR_1024: |
| 162 vpbroadcastq ymm11,QWORD[((32-128))+r15] |
| 163 vpmuludq ymm0,ymm10,YMMWORD[((0-128))+rsi] |
| 164 vpaddq ymm0,ymm0,YMMWORD[((0-192))+rbx] |
| 165 vpmuludq ymm1,ymm10,YMMWORD[((0-128))+r9] |
| 166 vpaddq ymm1,ymm1,YMMWORD[((32-192))+rbx] |
| 167 vpmuludq ymm2,ymm10,YMMWORD[((32-128))+r9] |
| 168 vpaddq ymm2,ymm2,YMMWORD[((64-192))+rbx] |
| 169 vpmuludq ymm3,ymm10,YMMWORD[((64-128))+r9] |
| 170 vpaddq ymm3,ymm3,YMMWORD[((96-192))+rbx] |
| 171 vpmuludq ymm4,ymm10,YMMWORD[((96-128))+r9] |
| 172 vpaddq ymm4,ymm4,YMMWORD[((128-192))+rbx] |
| 173 vpmuludq ymm5,ymm10,YMMWORD[((128-128))+r9] |
| 174 vpaddq ymm5,ymm5,YMMWORD[((160-192))+rbx] |
| 175 vpmuludq ymm6,ymm10,YMMWORD[((160-128))+r9] |
| 176 vpaddq ymm6,ymm6,YMMWORD[((192-192))+rbx] |
| 177 vpmuludq ymm7,ymm10,YMMWORD[((192-128))+r9] |
| 178 vpaddq ymm7,ymm7,YMMWORD[((224-192))+rbx] |
| 179 vpmuludq ymm8,ymm10,YMMWORD[((224-128))+r9] |
| 180 vpbroadcastq ymm10,QWORD[((64-128))+r15] |
| 181 vpaddq ymm8,ymm8,YMMWORD[((256-192))+rbx] |
| 182 $L$sqr_entry_1024: |
| 183 vmovdqu YMMWORD[(0-192)+rbx],ymm0 |
| 184 vmovdqu YMMWORD[(32-192)+rbx],ymm1 |
| 185 |
| 186 vpmuludq ymm12,ymm11,YMMWORD[((32-128))+rsi] |
| 187 vpaddq ymm2,ymm2,ymm12 |
| 188 vpmuludq ymm14,ymm11,YMMWORD[((32-128))+r9] |
| 189 vpaddq ymm3,ymm3,ymm14 |
| 190 vpmuludq ymm13,ymm11,YMMWORD[((64-128))+r9] |
| 191 vpaddq ymm4,ymm4,ymm13 |
| 192 vpmuludq ymm12,ymm11,YMMWORD[((96-128))+r9] |
| 193 vpaddq ymm5,ymm5,ymm12 |
| 194 vpmuludq ymm14,ymm11,YMMWORD[((128-128))+r9] |
| 195 vpaddq ymm6,ymm6,ymm14 |
| 196 vpmuludq ymm13,ymm11,YMMWORD[((160-128))+r9] |
| 197 vpaddq ymm7,ymm7,ymm13 |
| 198 vpmuludq ymm12,ymm11,YMMWORD[((192-128))+r9] |
| 199 vpaddq ymm8,ymm8,ymm12 |
| 200 vpmuludq ymm0,ymm11,YMMWORD[((224-128))+r9] |
| 201 vpbroadcastq ymm11,QWORD[((96-128))+r15] |
| 202 vpaddq ymm0,ymm0,YMMWORD[((288-192))+rbx] |
| 203 |
| 204 vmovdqu YMMWORD[(64-192)+rbx],ymm2 |
| 205 vmovdqu YMMWORD[(96-192)+rbx],ymm3 |
| 206 |
| 207 vpmuludq ymm13,ymm10,YMMWORD[((64-128))+rsi] |
| 208 vpaddq ymm4,ymm4,ymm13 |
| 209 vpmuludq ymm12,ymm10,YMMWORD[((64-128))+r9] |
| 210 vpaddq ymm5,ymm5,ymm12 |
| 211 vpmuludq ymm14,ymm10,YMMWORD[((96-128))+r9] |
| 212 vpaddq ymm6,ymm6,ymm14 |
| 213 vpmuludq ymm13,ymm10,YMMWORD[((128-128))+r9] |
| 214 vpaddq ymm7,ymm7,ymm13 |
| 215 vpmuludq ymm12,ymm10,YMMWORD[((160-128))+r9] |
| 216 vpaddq ymm8,ymm8,ymm12 |
| 217 vpmuludq ymm14,ymm10,YMMWORD[((192-128))+r9] |
| 218 vpaddq ymm0,ymm0,ymm14 |
| 219 vpmuludq ymm1,ymm10,YMMWORD[((224-128))+r9] |
| 220 vpbroadcastq ymm10,QWORD[((128-128))+r15] |
| 221 vpaddq ymm1,ymm1,YMMWORD[((320-448))+r12] |
| 222 |
| 223 vmovdqu YMMWORD[(128-192)+rbx],ymm4 |
| 224 vmovdqu YMMWORD[(160-192)+rbx],ymm5 |
| 225 |
| 226 vpmuludq ymm12,ymm11,YMMWORD[((96-128))+rsi] |
| 227 vpaddq ymm6,ymm6,ymm12 |
| 228 vpmuludq ymm14,ymm11,YMMWORD[((96-128))+r9] |
| 229 vpaddq ymm7,ymm7,ymm14 |
| 230 vpmuludq ymm13,ymm11,YMMWORD[((128-128))+r9] |
| 231 vpaddq ymm8,ymm8,ymm13 |
| 232 vpmuludq ymm12,ymm11,YMMWORD[((160-128))+r9] |
| 233 vpaddq ymm0,ymm0,ymm12 |
| 234 vpmuludq ymm14,ymm11,YMMWORD[((192-128))+r9] |
| 235 vpaddq ymm1,ymm1,ymm14 |
| 236 vpmuludq ymm2,ymm11,YMMWORD[((224-128))+r9] |
| 237 vpbroadcastq ymm11,QWORD[((160-128))+r15] |
| 238 vpaddq ymm2,ymm2,YMMWORD[((352-448))+r12] |
| 239 |
| 240 vmovdqu YMMWORD[(192-192)+rbx],ymm6 |
| 241 vmovdqu YMMWORD[(224-192)+rbx],ymm7 |
| 242 |
| 243 vpmuludq ymm12,ymm10,YMMWORD[((128-128))+rsi] |
| 244 vpaddq ymm8,ymm8,ymm12 |
| 245 vpmuludq ymm14,ymm10,YMMWORD[((128-128))+r9] |
| 246 vpaddq ymm0,ymm0,ymm14 |
| 247 vpmuludq ymm13,ymm10,YMMWORD[((160-128))+r9] |
| 248 vpaddq ymm1,ymm1,ymm13 |
| 249 vpmuludq ymm12,ymm10,YMMWORD[((192-128))+r9] |
| 250 vpaddq ymm2,ymm2,ymm12 |
| 251 vpmuludq ymm3,ymm10,YMMWORD[((224-128))+r9] |
| 252 vpbroadcastq ymm10,QWORD[((192-128))+r15] |
| 253 vpaddq ymm3,ymm3,YMMWORD[((384-448))+r12] |
| 254 |
| 255 vmovdqu YMMWORD[(256-192)+rbx],ymm8 |
| 256 vmovdqu YMMWORD[(288-192)+rbx],ymm0 |
| 257 lea rbx,[8+rbx] |
| 258 |
| 259 vpmuludq ymm13,ymm11,YMMWORD[((160-128))+rsi] |
| 260 vpaddq ymm1,ymm1,ymm13 |
| 261 vpmuludq ymm12,ymm11,YMMWORD[((160-128))+r9] |
| 262 vpaddq ymm2,ymm2,ymm12 |
| 263 vpmuludq ymm14,ymm11,YMMWORD[((192-128))+r9] |
| 264 vpaddq ymm3,ymm3,ymm14 |
| 265 vpmuludq ymm4,ymm11,YMMWORD[((224-128))+r9] |
| 266 vpbroadcastq ymm11,QWORD[((224-128))+r15] |
| 267 vpaddq ymm4,ymm4,YMMWORD[((416-448))+r12] |
| 268 |
| 269 vmovdqu YMMWORD[(320-448)+r12],ymm1 |
| 270 vmovdqu YMMWORD[(352-448)+r12],ymm2 |
| 271 |
| 272 vpmuludq ymm12,ymm10,YMMWORD[((192-128))+rsi] |
| 273 vpaddq ymm3,ymm3,ymm12 |
| 274 vpmuludq ymm14,ymm10,YMMWORD[((192-128))+r9] |
| 275 vpbroadcastq ymm0,QWORD[((256-128))+r15] |
| 276 vpaddq ymm4,ymm4,ymm14 |
| 277 vpmuludq ymm5,ymm10,YMMWORD[((224-128))+r9] |
| 278 vpbroadcastq ymm10,QWORD[((0+8-128))+r15] |
| 279 vpaddq ymm5,ymm5,YMMWORD[((448-448))+r12] |
| 280 |
| 281 vmovdqu YMMWORD[(384-448)+r12],ymm3 |
| 282 vmovdqu YMMWORD[(416-448)+r12],ymm4 |
| 283 lea r15,[8+r15] |
| 284 |
| 285 vpmuludq ymm12,ymm11,YMMWORD[((224-128))+rsi] |
| 286 vpaddq ymm5,ymm5,ymm12 |
| 287 vpmuludq ymm6,ymm11,YMMWORD[((224-128))+r9] |
| 288 vpaddq ymm6,ymm6,YMMWORD[((480-448))+r12] |
| 289 |
| 290 vpmuludq ymm7,ymm0,YMMWORD[((256-128))+rsi] |
| 291 vmovdqu YMMWORD[(448-448)+r12],ymm5 |
| 292 vpaddq ymm7,ymm7,YMMWORD[((512-448))+r12] |
| 293 vmovdqu YMMWORD[(480-448)+r12],ymm6 |
| 294 vmovdqu YMMWORD[(512-448)+r12],ymm7 |
| 295 lea r12,[8+r12] |
| 296 |
| 297 dec r14d |
| 298 jnz NEAR $L$OOP_SQR_1024 |
| 299 |
| 300 vmovdqu ymm8,YMMWORD[256+rsp] |
| 301 vmovdqu ymm1,YMMWORD[288+rsp] |
| 302 vmovdqu ymm2,YMMWORD[320+rsp] |
| 303 lea rbx,[192+rsp] |
| 304 |
| 305 vpsrlq ymm14,ymm8,29 |
| 306 vpand ymm8,ymm8,ymm15 |
| 307 vpsrlq ymm11,ymm1,29 |
| 308 vpand ymm1,ymm1,ymm15 |
| 309 |
| 310 vpermq ymm14,ymm14,0x93 |
| 311 vpxor ymm9,ymm9,ymm9 |
| 312 vpermq ymm11,ymm11,0x93 |
| 313 |
| 314 vpblendd ymm10,ymm14,ymm9,3 |
| 315 vpblendd ymm14,ymm11,ymm14,3 |
| 316 vpaddq ymm8,ymm8,ymm10 |
| 317 vpblendd ymm11,ymm9,ymm11,3 |
| 318 vpaddq ymm1,ymm1,ymm14 |
| 319 vpaddq ymm2,ymm2,ymm11 |
| 320 vmovdqu YMMWORD[(288-192)+rbx],ymm1 |
| 321 vmovdqu YMMWORD[(320-192)+rbx],ymm2 |
| 322 |
| 323 mov rax,QWORD[rsp] |
| 324 mov r10,QWORD[8+rsp] |
| 325 mov r11,QWORD[16+rsp] |
| 326 mov r12,QWORD[24+rsp] |
| 327 vmovdqu ymm1,YMMWORD[32+rsp] |
| 328 vmovdqu ymm2,YMMWORD[((64-192))+rbx] |
| 329 vmovdqu ymm3,YMMWORD[((96-192))+rbx] |
| 330 vmovdqu ymm4,YMMWORD[((128-192))+rbx] |
| 331 vmovdqu ymm5,YMMWORD[((160-192))+rbx] |
| 332 vmovdqu ymm6,YMMWORD[((192-192))+rbx] |
| 333 vmovdqu ymm7,YMMWORD[((224-192))+rbx] |
| 334 |
| 335 mov r9,rax |
| 336 imul eax,ecx |
| 337 and eax,0x1fffffff |
| 338 vmovd xmm12,eax |
| 339 |
| 340 mov rdx,rax |
| 341 imul rax,QWORD[((-128))+r13] |
| 342 vpbroadcastq ymm12,xmm12 |
| 343 add r9,rax |
| 344 mov rax,rdx |
| 345 imul rax,QWORD[((8-128))+r13] |
| 346 shr r9,29 |
| 347 add r10,rax |
| 348 mov rax,rdx |
| 349 imul rax,QWORD[((16-128))+r13] |
| 350 add r10,r9 |
| 351 add r11,rax |
| 352 imul rdx,QWORD[((24-128))+r13] |
| 353 add r12,rdx |
| 354 |
| 355 mov rax,r10 |
| 356 imul eax,ecx |
| 357 and eax,0x1fffffff |
| 358 |
| 359 mov r14d,9 |
| 360 jmp NEAR $L$OOP_REDUCE_1024 |
| 361 |
| 362 ALIGN 32 |
| 363 $L$OOP_REDUCE_1024: |
| 364 vmovd xmm13,eax |
| 365 vpbroadcastq ymm13,xmm13 |
| 366 |
| 367 vpmuludq ymm10,ymm12,YMMWORD[((32-128))+r13] |
| 368 mov rdx,rax |
| 369 imul rax,QWORD[((-128))+r13] |
| 370 vpaddq ymm1,ymm1,ymm10 |
| 371 add r10,rax |
| 372 vpmuludq ymm14,ymm12,YMMWORD[((64-128))+r13] |
| 373 mov rax,rdx |
| 374 imul rax,QWORD[((8-128))+r13] |
| 375 vpaddq ymm2,ymm2,ymm14 |
| 376 vpmuludq ymm11,ymm12,YMMWORD[((96-128))+r13] |
| 377 DB 0x67 |
| 378 add r11,rax |
| 379 DB 0x67 |
| 380 mov rax,rdx |
| 381 imul rax,QWORD[((16-128))+r13] |
| 382 shr r10,29 |
| 383 vpaddq ymm3,ymm3,ymm11 |
| 384 vpmuludq ymm10,ymm12,YMMWORD[((128-128))+r13] |
| 385 add r12,rax |
| 386 add r11,r10 |
| 387 vpaddq ymm4,ymm4,ymm10 |
| 388 vpmuludq ymm14,ymm12,YMMWORD[((160-128))+r13] |
| 389 mov rax,r11 |
| 390 imul eax,ecx |
| 391 vpaddq ymm5,ymm5,ymm14 |
| 392 vpmuludq ymm11,ymm12,YMMWORD[((192-128))+r13] |
| 393 and eax,0x1fffffff |
| 394 vpaddq ymm6,ymm6,ymm11 |
| 395 vpmuludq ymm10,ymm12,YMMWORD[((224-128))+r13] |
| 396 vpaddq ymm7,ymm7,ymm10 |
| 397 vpmuludq ymm14,ymm12,YMMWORD[((256-128))+r13] |
| 398 vmovd xmm12,eax |
| 399 |
| 400 vpaddq ymm8,ymm8,ymm14 |
| 401 |
| 402 vpbroadcastq ymm12,xmm12 |
| 403 |
| 404 vpmuludq ymm11,ymm13,YMMWORD[((32-8-128))+r13] |
| 405 vmovdqu ymm14,YMMWORD[((96-8-128))+r13] |
| 406 mov rdx,rax |
| 407 imul rax,QWORD[((-128))+r13] |
| 408 vpaddq ymm1,ymm1,ymm11 |
| 409 vpmuludq ymm10,ymm13,YMMWORD[((64-8-128))+r13] |
| 410 vmovdqu ymm11,YMMWORD[((128-8-128))+r13] |
| 411 add r11,rax |
| 412 mov rax,rdx |
| 413 imul rax,QWORD[((8-128))+r13] |
| 414 vpaddq ymm2,ymm2,ymm10 |
| 415 add rax,r12 |
| 416 shr r11,29 |
| 417 vpmuludq ymm14,ymm14,ymm13 |
| 418 vmovdqu ymm10,YMMWORD[((160-8-128))+r13] |
| 419 add rax,r11 |
| 420 vpaddq ymm3,ymm3,ymm14 |
| 421 vpmuludq ymm11,ymm11,ymm13 |
| 422 vmovdqu ymm14,YMMWORD[((192-8-128))+r13] |
| 423 DB 0x67 |
| 424 mov r12,rax |
| 425 imul eax,ecx |
| 426 vpaddq ymm4,ymm4,ymm11 |
| 427 vpmuludq ymm10,ymm10,ymm13 |
| 428 DB 0xc4,0x41,0x7e,0x6f,0x9d,0x58,0x00,0x00,0x00 |
| 429 and eax,0x1fffffff |
| 430 vpaddq ymm5,ymm5,ymm10 |
| 431 vpmuludq ymm14,ymm14,ymm13 |
| 432 vmovdqu ymm10,YMMWORD[((256-8-128))+r13] |
| 433 vpaddq ymm6,ymm6,ymm14 |
| 434 vpmuludq ymm11,ymm11,ymm13 |
| 435 vmovdqu ymm9,YMMWORD[((288-8-128))+r13] |
| 436 vmovd xmm0,eax |
| 437 imul rax,QWORD[((-128))+r13] |
| 438 vpaddq ymm7,ymm7,ymm11 |
| 439 vpmuludq ymm10,ymm10,ymm13 |
| 440 vmovdqu ymm14,YMMWORD[((32-16-128))+r13] |
| 441 vpbroadcastq ymm0,xmm0 |
| 442 vpaddq ymm8,ymm8,ymm10 |
| 443 vpmuludq ymm9,ymm9,ymm13 |
| 444 vmovdqu ymm11,YMMWORD[((64-16-128))+r13] |
| 445 add r12,rax |
| 446 |
| 447 vmovdqu ymm13,YMMWORD[((32-24-128))+r13] |
| 448 vpmuludq ymm14,ymm14,ymm12 |
| 449 vmovdqu ymm10,YMMWORD[((96-16-128))+r13] |
| 450 vpaddq ymm1,ymm1,ymm14 |
| 451 vpmuludq ymm13,ymm13,ymm0 |
| 452 vpmuludq ymm11,ymm11,ymm12 |
| 453 DB 0xc4,0x41,0x7e,0x6f,0xb5,0xf0,0xff,0xff,0xff |
| 454 vpaddq ymm13,ymm13,ymm1 |
| 455 vpaddq ymm2,ymm2,ymm11 |
| 456 vpmuludq ymm10,ymm10,ymm12 |
| 457 vmovdqu ymm11,YMMWORD[((160-16-128))+r13] |
| 458 DB 0x67 |
| 459 vmovq rax,xmm13 |
| 460 vmovdqu YMMWORD[rsp],ymm13 |
| 461 vpaddq ymm3,ymm3,ymm10 |
| 462 vpmuludq ymm14,ymm14,ymm12 |
| 463 vmovdqu ymm10,YMMWORD[((192-16-128))+r13] |
| 464 vpaddq ymm4,ymm4,ymm14 |
| 465 vpmuludq ymm11,ymm11,ymm12 |
| 466 vmovdqu ymm14,YMMWORD[((224-16-128))+r13] |
| 467 vpaddq ymm5,ymm5,ymm11 |
| 468 vpmuludq ymm10,ymm10,ymm12 |
| 469 vmovdqu ymm11,YMMWORD[((256-16-128))+r13] |
| 470 vpaddq ymm6,ymm6,ymm10 |
| 471 vpmuludq ymm14,ymm14,ymm12 |
| 472 shr r12,29 |
| 473 vmovdqu ymm10,YMMWORD[((288-16-128))+r13] |
| 474 add rax,r12 |
| 475 vpaddq ymm7,ymm7,ymm14 |
| 476 vpmuludq ymm11,ymm11,ymm12 |
| 477 |
| 478 mov r9,rax |
| 479 imul eax,ecx |
| 480 vpaddq ymm8,ymm8,ymm11 |
| 481 vpmuludq ymm10,ymm10,ymm12 |
| 482 and eax,0x1fffffff |
| 483 vmovd xmm12,eax |
| 484 vmovdqu ymm11,YMMWORD[((96-24-128))+r13] |
| 485 DB 0x67 |
| 486 vpaddq ymm9,ymm9,ymm10 |
| 487 vpbroadcastq ymm12,xmm12 |
| 488 |
| 489 vpmuludq ymm14,ymm0,YMMWORD[((64-24-128))+r13] |
| 490 vmovdqu ymm10,YMMWORD[((128-24-128))+r13] |
| 491 mov rdx,rax |
| 492 imul rax,QWORD[((-128))+r13] |
| 493 mov r10,QWORD[8+rsp] |
| 494 vpaddq ymm1,ymm2,ymm14 |
| 495 vpmuludq ymm11,ymm11,ymm0 |
| 496 vmovdqu ymm14,YMMWORD[((160-24-128))+r13] |
| 497 add r9,rax |
| 498 mov rax,rdx |
| 499 imul rax,QWORD[((8-128))+r13] |
| 500 DB 0x67 |
| 501 shr r9,29 |
| 502 mov r11,QWORD[16+rsp] |
| 503 vpaddq ymm2,ymm3,ymm11 |
| 504 vpmuludq ymm10,ymm10,ymm0 |
| 505 vmovdqu ymm11,YMMWORD[((192-24-128))+r13] |
| 506 add r10,rax |
| 507 mov rax,rdx |
| 508 imul rax,QWORD[((16-128))+r13] |
| 509 vpaddq ymm3,ymm4,ymm10 |
| 510 vpmuludq ymm14,ymm14,ymm0 |
| 511 vmovdqu ymm10,YMMWORD[((224-24-128))+r13] |
| 512 imul rdx,QWORD[((24-128))+r13] |
| 513 add r11,rax |
| 514 lea rax,[r10*1+r9] |
| 515 vpaddq ymm4,ymm5,ymm14 |
| 516 vpmuludq ymm11,ymm11,ymm0 |
| 517 vmovdqu ymm14,YMMWORD[((256-24-128))+r13] |
| 518 mov r10,rax |
| 519 imul eax,ecx |
| 520 vpmuludq ymm10,ymm10,ymm0 |
| 521 vpaddq ymm5,ymm6,ymm11 |
| 522 vmovdqu ymm11,YMMWORD[((288-24-128))+r13] |
| 523 and eax,0x1fffffff |
| 524 vpaddq ymm6,ymm7,ymm10 |
| 525 vpmuludq ymm14,ymm14,ymm0 |
| 526 add rdx,QWORD[24+rsp] |
| 527 vpaddq ymm7,ymm8,ymm14 |
| 528 vpmuludq ymm11,ymm11,ymm0 |
| 529 vpaddq ymm8,ymm9,ymm11 |
| 530 vmovq xmm9,r12 |
| 531 mov r12,rdx |
| 532 |
| 533 dec r14d |
| 534 jnz NEAR $L$OOP_REDUCE_1024 |
| 535 lea r12,[448+rsp] |
| 536 vpaddq ymm0,ymm13,ymm9 |
| 537 vpxor ymm9,ymm9,ymm9 |
| 538 |
| 539 vpaddq ymm0,ymm0,YMMWORD[((288-192))+rbx] |
| 540 vpaddq ymm1,ymm1,YMMWORD[((320-448))+r12] |
| 541 vpaddq ymm2,ymm2,YMMWORD[((352-448))+r12] |
| 542 vpaddq ymm3,ymm3,YMMWORD[((384-448))+r12] |
| 543 vpaddq ymm4,ymm4,YMMWORD[((416-448))+r12] |
| 544 vpaddq ymm5,ymm5,YMMWORD[((448-448))+r12] |
| 545 vpaddq ymm6,ymm6,YMMWORD[((480-448))+r12] |
| 546 vpaddq ymm7,ymm7,YMMWORD[((512-448))+r12] |
| 547 vpaddq ymm8,ymm8,YMMWORD[((544-448))+r12] |
| 548 |
| 549 vpsrlq ymm14,ymm0,29 |
| 550 vpand ymm0,ymm0,ymm15 |
| 551 vpsrlq ymm11,ymm1,29 |
| 552 vpand ymm1,ymm1,ymm15 |
| 553 vpsrlq ymm12,ymm2,29 |
| 554 vpermq ymm14,ymm14,0x93 |
| 555 vpand ymm2,ymm2,ymm15 |
| 556 vpsrlq ymm13,ymm3,29 |
| 557 vpermq ymm11,ymm11,0x93 |
| 558 vpand ymm3,ymm3,ymm15 |
| 559 vpermq ymm12,ymm12,0x93 |
| 560 |
| 561 vpblendd ymm10,ymm14,ymm9,3 |
| 562 vpermq ymm13,ymm13,0x93 |
| 563 vpblendd ymm14,ymm11,ymm14,3 |
| 564 vpaddq ymm0,ymm0,ymm10 |
| 565 vpblendd ymm11,ymm12,ymm11,3 |
| 566 vpaddq ymm1,ymm1,ymm14 |
| 567 vpblendd ymm12,ymm13,ymm12,3 |
| 568 vpaddq ymm2,ymm2,ymm11 |
| 569 vpblendd ymm13,ymm9,ymm13,3 |
| 570 vpaddq ymm3,ymm3,ymm12 |
| 571 vpaddq ymm4,ymm4,ymm13 |
| 572 |
| 573 vpsrlq ymm14,ymm0,29 |
| 574 vpand ymm0,ymm0,ymm15 |
| 575 vpsrlq ymm11,ymm1,29 |
| 576 vpand ymm1,ymm1,ymm15 |
| 577 vpsrlq ymm12,ymm2,29 |
| 578 vpermq ymm14,ymm14,0x93 |
| 579 vpand ymm2,ymm2,ymm15 |
| 580 vpsrlq ymm13,ymm3,29 |
| 581 vpermq ymm11,ymm11,0x93 |
| 582 vpand ymm3,ymm3,ymm15 |
| 583 vpermq ymm12,ymm12,0x93 |
| 584 |
| 585 vpblendd ymm10,ymm14,ymm9,3 |
| 586 vpermq ymm13,ymm13,0x93 |
| 587 vpblendd ymm14,ymm11,ymm14,3 |
| 588 vpaddq ymm0,ymm0,ymm10 |
| 589 vpblendd ymm11,ymm12,ymm11,3 |
| 590 vpaddq ymm1,ymm1,ymm14 |
| 591 vmovdqu YMMWORD[(0-128)+rdi],ymm0 |
| 592 vpblendd ymm12,ymm13,ymm12,3 |
| 593 vpaddq ymm2,ymm2,ymm11 |
| 594 vmovdqu YMMWORD[(32-128)+rdi],ymm1 |
| 595 vpblendd ymm13,ymm9,ymm13,3 |
| 596 vpaddq ymm3,ymm3,ymm12 |
| 597 vmovdqu YMMWORD[(64-128)+rdi],ymm2 |
| 598 vpaddq ymm4,ymm4,ymm13 |
| 599 vmovdqu YMMWORD[(96-128)+rdi],ymm3 |
| 600 vpsrlq ymm14,ymm4,29 |
| 601 vpand ymm4,ymm4,ymm15 |
| 602 vpsrlq ymm11,ymm5,29 |
| 603 vpand ymm5,ymm5,ymm15 |
| 604 vpsrlq ymm12,ymm6,29 |
| 605 vpermq ymm14,ymm14,0x93 |
| 606 vpand ymm6,ymm6,ymm15 |
| 607 vpsrlq ymm13,ymm7,29 |
| 608 vpermq ymm11,ymm11,0x93 |
| 609 vpand ymm7,ymm7,ymm15 |
| 610 vpsrlq ymm0,ymm8,29 |
| 611 vpermq ymm12,ymm12,0x93 |
| 612 vpand ymm8,ymm8,ymm15 |
| 613 vpermq ymm13,ymm13,0x93 |
| 614 |
| 615 vpblendd ymm10,ymm14,ymm9,3 |
| 616 vpermq ymm0,ymm0,0x93 |
| 617 vpblendd ymm14,ymm11,ymm14,3 |
| 618 vpaddq ymm4,ymm4,ymm10 |
| 619 vpblendd ymm11,ymm12,ymm11,3 |
| 620 vpaddq ymm5,ymm5,ymm14 |
| 621 vpblendd ymm12,ymm13,ymm12,3 |
| 622 vpaddq ymm6,ymm6,ymm11 |
| 623 vpblendd ymm13,ymm0,ymm13,3 |
| 624 vpaddq ymm7,ymm7,ymm12 |
| 625 vpaddq ymm8,ymm8,ymm13 |
| 626 |
| 627 vpsrlq ymm14,ymm4,29 |
| 628 vpand ymm4,ymm4,ymm15 |
| 629 vpsrlq ymm11,ymm5,29 |
| 630 vpand ymm5,ymm5,ymm15 |
| 631 vpsrlq ymm12,ymm6,29 |
| 632 vpermq ymm14,ymm14,0x93 |
| 633 vpand ymm6,ymm6,ymm15 |
| 634 vpsrlq ymm13,ymm7,29 |
| 635 vpermq ymm11,ymm11,0x93 |
| 636 vpand ymm7,ymm7,ymm15 |
| 637 vpsrlq ymm0,ymm8,29 |
| 638 vpermq ymm12,ymm12,0x93 |
| 639 vpand ymm8,ymm8,ymm15 |
| 640 vpermq ymm13,ymm13,0x93 |
| 641 |
| 642 vpblendd ymm10,ymm14,ymm9,3 |
| 643 vpermq ymm0,ymm0,0x93 |
| 644 vpblendd ymm14,ymm11,ymm14,3 |
| 645 vpaddq ymm4,ymm4,ymm10 |
| 646 vpblendd ymm11,ymm12,ymm11,3 |
| 647 vpaddq ymm5,ymm5,ymm14 |
| 648 vmovdqu YMMWORD[(128-128)+rdi],ymm4 |
| 649 vpblendd ymm12,ymm13,ymm12,3 |
| 650 vpaddq ymm6,ymm6,ymm11 |
| 651 vmovdqu YMMWORD[(160-128)+rdi],ymm5 |
| 652 vpblendd ymm13,ymm0,ymm13,3 |
| 653 vpaddq ymm7,ymm7,ymm12 |
| 654 vmovdqu YMMWORD[(192-128)+rdi],ymm6 |
| 655 vpaddq ymm8,ymm8,ymm13 |
| 656 vmovdqu YMMWORD[(224-128)+rdi],ymm7 |
| 657 vmovdqu YMMWORD[(256-128)+rdi],ymm8 |
| 658 |
| 659 mov rsi,rdi |
| 660 dec r8d |
| 661 jne NEAR $L$OOP_GRANDE_SQR_1024 |
| 662 |
| 663 vzeroall |
| 664 mov rax,rbp |
| 665 |
| 666 $L$sqr_1024_in_tail: |
| 667 movaps xmm6,XMMWORD[((-216))+rax] |
| 668 movaps xmm7,XMMWORD[((-200))+rax] |
| 669 movaps xmm8,XMMWORD[((-184))+rax] |
| 670 movaps xmm9,XMMWORD[((-168))+rax] |
| 671 movaps xmm10,XMMWORD[((-152))+rax] |
| 672 movaps xmm11,XMMWORD[((-136))+rax] |
| 673 movaps xmm12,XMMWORD[((-120))+rax] |
| 674 movaps xmm13,XMMWORD[((-104))+rax] |
| 675 movaps xmm14,XMMWORD[((-88))+rax] |
| 676 movaps xmm15,XMMWORD[((-72))+rax] |
| 677 mov r15,QWORD[((-48))+rax] |
| 678 |
| 679 mov r14,QWORD[((-40))+rax] |
| 680 |
| 681 mov r13,QWORD[((-32))+rax] |
| 682 |
| 683 mov r12,QWORD[((-24))+rax] |
| 684 |
| 685 mov rbp,QWORD[((-16))+rax] |
| 686 |
| 687 mov rbx,QWORD[((-8))+rax] |
| 688 |
| 689 lea rsp,[rax] |
| 690 |
| 691 $L$sqr_1024_epilogue: |
| 692 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 693 mov rsi,QWORD[16+rsp] |
| 694 DB 0F3h,0C3h ;repret |
| 695 |
| 696 $L$SEH_end_rsaz_1024_sqr_avx2: |
| 697 global rsaz_1024_mul_avx2 |
| 698 |
| 699 ALIGN 64 |
| 700 rsaz_1024_mul_avx2: |
| 701 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 702 mov QWORD[16+rsp],rsi |
| 703 mov rax,rsp |
| 704 $L$SEH_begin_rsaz_1024_mul_avx2: |
| 705 mov rdi,rcx |
| 706 mov rsi,rdx |
| 707 mov rdx,r8 |
| 708 mov rcx,r9 |
| 709 mov r8,QWORD[40+rsp] |
| 710 |
| 711 |
| 712 |
| 713 lea rax,[rsp] |
| 714 |
| 715 push rbx |
| 716 |
| 717 push rbp |
| 718 |
| 719 push r12 |
| 720 |
| 721 push r13 |
| 722 |
| 723 push r14 |
| 724 |
| 725 push r15 |
| 726 |
| 727 vzeroupper |
| 728 lea rsp,[((-168))+rsp] |
| 729 vmovaps XMMWORD[(-216)+rax],xmm6 |
| 730 vmovaps XMMWORD[(-200)+rax],xmm7 |
| 731 vmovaps XMMWORD[(-184)+rax],xmm8 |
| 732 vmovaps XMMWORD[(-168)+rax],xmm9 |
| 733 vmovaps XMMWORD[(-152)+rax],xmm10 |
| 734 vmovaps XMMWORD[(-136)+rax],xmm11 |
| 735 vmovaps XMMWORD[(-120)+rax],xmm12 |
| 736 vmovaps XMMWORD[(-104)+rax],xmm13 |
| 737 vmovaps XMMWORD[(-88)+rax],xmm14 |
| 738 vmovaps XMMWORD[(-72)+rax],xmm15 |
| 739 $L$mul_1024_body: |
| 740 mov rbp,rax |
| 741 |
| 742 vzeroall |
| 743 mov r13,rdx |
| 744 sub rsp,64 |
| 745 |
| 746 |
| 747 |
| 748 |
| 749 |
| 750 |
| 751 DB 0x67,0x67 |
| 752 mov r15,rsi |
| 753 and r15,4095 |
| 754 add r15,32*10 |
| 755 shr r15,12 |
| 756 mov r15,rsi |
| 757 cmovnz rsi,r13 |
| 758 cmovnz r13,r15 |
| 759 |
| 760 mov r15,rcx |
| 761 sub rsi,-128 |
| 762 sub rcx,-128 |
| 763 sub rdi,-128 |
| 764 |
| 765 and r15,4095 |
| 766 add r15,32*10 |
| 767 DB 0x67,0x67 |
| 768 shr r15,12 |
| 769 jz NEAR $L$mul_1024_no_n_copy |
| 770 |
| 771 |
| 772 |
| 773 |
| 774 |
| 775 sub rsp,32*10 |
| 776 vmovdqu ymm0,YMMWORD[((0-128))+rcx] |
| 777 and rsp,-512 |
| 778 vmovdqu ymm1,YMMWORD[((32-128))+rcx] |
| 779 vmovdqu ymm2,YMMWORD[((64-128))+rcx] |
| 780 vmovdqu ymm3,YMMWORD[((96-128))+rcx] |
| 781 vmovdqu ymm4,YMMWORD[((128-128))+rcx] |
| 782 vmovdqu ymm5,YMMWORD[((160-128))+rcx] |
| 783 vmovdqu ymm6,YMMWORD[((192-128))+rcx] |
| 784 vmovdqu ymm7,YMMWORD[((224-128))+rcx] |
| 785 vmovdqu ymm8,YMMWORD[((256-128))+rcx] |
| 786 lea rcx,[((64+128))+rsp] |
| 787 vmovdqu YMMWORD[(0-128)+rcx],ymm0 |
| 788 vpxor ymm0,ymm0,ymm0 |
| 789 vmovdqu YMMWORD[(32-128)+rcx],ymm1 |
| 790 vpxor ymm1,ymm1,ymm1 |
| 791 vmovdqu YMMWORD[(64-128)+rcx],ymm2 |
| 792 vpxor ymm2,ymm2,ymm2 |
| 793 vmovdqu YMMWORD[(96-128)+rcx],ymm3 |
| 794 vpxor ymm3,ymm3,ymm3 |
| 795 vmovdqu YMMWORD[(128-128)+rcx],ymm4 |
| 796 vpxor ymm4,ymm4,ymm4 |
| 797 vmovdqu YMMWORD[(160-128)+rcx],ymm5 |
| 798 vpxor ymm5,ymm5,ymm5 |
| 799 vmovdqu YMMWORD[(192-128)+rcx],ymm6 |
| 800 vpxor ymm6,ymm6,ymm6 |
| 801 vmovdqu YMMWORD[(224-128)+rcx],ymm7 |
| 802 vpxor ymm7,ymm7,ymm7 |
| 803 vmovdqu YMMWORD[(256-128)+rcx],ymm8 |
| 804 vmovdqa ymm8,ymm0 |
| 805 vmovdqu YMMWORD[(288-128)+rcx],ymm9 |
| 806 $L$mul_1024_no_n_copy: |
| 807 and rsp,-64 |
| 808 |
| 809 mov rbx,QWORD[r13] |
| 810 vpbroadcastq ymm10,QWORD[r13] |
| 811 vmovdqu YMMWORD[rsp],ymm0 |
| 812 xor r9,r9 |
| 813 DB 0x67 |
| 814 xor r10,r10 |
| 815 xor r11,r11 |
| 816 xor r12,r12 |
| 817 |
| 818 vmovdqu ymm15,YMMWORD[$L$and_mask] |
| 819 mov r14d,9 |
| 820 vmovdqu YMMWORD[(288-128)+rdi],ymm9 |
| 821 jmp NEAR $L$oop_mul_1024 |
| 822 |
| 823 ALIGN 32 |
| 824 $L$oop_mul_1024: |
| 825 vpsrlq ymm9,ymm3,29 |
| 826 mov rax,rbx |
| 827 imul rax,QWORD[((-128))+rsi] |
| 828 add rax,r9 |
| 829 mov r10,rbx |
| 830 imul r10,QWORD[((8-128))+rsi] |
| 831 add r10,QWORD[8+rsp] |
| 832 |
| 833 mov r9,rax |
| 834 imul eax,r8d |
| 835 and eax,0x1fffffff |
| 836 |
| 837 mov r11,rbx |
| 838 imul r11,QWORD[((16-128))+rsi] |
| 839 add r11,QWORD[16+rsp] |
| 840 |
| 841 mov r12,rbx |
| 842 imul r12,QWORD[((24-128))+rsi] |
| 843 add r12,QWORD[24+rsp] |
| 844 vpmuludq ymm0,ymm10,YMMWORD[((32-128))+rsi] |
| 845 vmovd xmm11,eax |
| 846 vpaddq ymm1,ymm1,ymm0 |
| 847 vpmuludq ymm12,ymm10,YMMWORD[((64-128))+rsi] |
| 848 vpbroadcastq ymm11,xmm11 |
| 849 vpaddq ymm2,ymm2,ymm12 |
| 850 vpmuludq ymm13,ymm10,YMMWORD[((96-128))+rsi] |
| 851 vpand ymm3,ymm3,ymm15 |
| 852 vpaddq ymm3,ymm3,ymm13 |
| 853 vpmuludq ymm0,ymm10,YMMWORD[((128-128))+rsi] |
| 854 vpaddq ymm4,ymm4,ymm0 |
| 855 vpmuludq ymm12,ymm10,YMMWORD[((160-128))+rsi] |
| 856 vpaddq ymm5,ymm5,ymm12 |
| 857 vpmuludq ymm13,ymm10,YMMWORD[((192-128))+rsi] |
| 858 vpaddq ymm6,ymm6,ymm13 |
| 859 vpmuludq ymm0,ymm10,YMMWORD[((224-128))+rsi] |
| 860 vpermq ymm9,ymm9,0x93 |
| 861 vpaddq ymm7,ymm7,ymm0 |
| 862 vpmuludq ymm12,ymm10,YMMWORD[((256-128))+rsi] |
| 863 vpbroadcastq ymm10,QWORD[8+r13] |
| 864 vpaddq ymm8,ymm8,ymm12 |
| 865 |
| 866 mov rdx,rax |
| 867 imul rax,QWORD[((-128))+rcx] |
| 868 add r9,rax |
| 869 mov rax,rdx |
| 870 imul rax,QWORD[((8-128))+rcx] |
| 871 add r10,rax |
| 872 mov rax,rdx |
| 873 imul rax,QWORD[((16-128))+rcx] |
| 874 add r11,rax |
| 875 shr r9,29 |
| 876 imul rdx,QWORD[((24-128))+rcx] |
| 877 add r12,rdx |
| 878 add r10,r9 |
| 879 |
| 880 vpmuludq ymm13,ymm11,YMMWORD[((32-128))+rcx] |
| 881 vmovq rbx,xmm10 |
| 882 vpaddq ymm1,ymm1,ymm13 |
| 883 vpmuludq ymm0,ymm11,YMMWORD[((64-128))+rcx] |
| 884 vpaddq ymm2,ymm2,ymm0 |
| 885 vpmuludq ymm12,ymm11,YMMWORD[((96-128))+rcx] |
| 886 vpaddq ymm3,ymm3,ymm12 |
| 887 vpmuludq ymm13,ymm11,YMMWORD[((128-128))+rcx] |
| 888 vpaddq ymm4,ymm4,ymm13 |
| 889 vpmuludq ymm0,ymm11,YMMWORD[((160-128))+rcx] |
| 890 vpaddq ymm5,ymm5,ymm0 |
| 891 vpmuludq ymm12,ymm11,YMMWORD[((192-128))+rcx] |
| 892 vpaddq ymm6,ymm6,ymm12 |
| 893 vpmuludq ymm13,ymm11,YMMWORD[((224-128))+rcx] |
| 894 vpblendd ymm9,ymm9,ymm14,3 |
| 895 vpaddq ymm7,ymm7,ymm13 |
| 896 vpmuludq ymm0,ymm11,YMMWORD[((256-128))+rcx] |
| 897 vpaddq ymm3,ymm3,ymm9 |
| 898 vpaddq ymm8,ymm8,ymm0 |
| 899 |
| 900 mov rax,rbx |
| 901 imul rax,QWORD[((-128))+rsi] |
| 902 add r10,rax |
| 903 vmovdqu ymm12,YMMWORD[((-8+32-128))+rsi] |
| 904 mov rax,rbx |
| 905 imul rax,QWORD[((8-128))+rsi] |
| 906 add r11,rax |
| 907 vmovdqu ymm13,YMMWORD[((-8+64-128))+rsi] |
| 908 |
| 909 mov rax,r10 |
| 910 imul eax,r8d |
| 911 and eax,0x1fffffff |
| 912 |
| 913 imul rbx,QWORD[((16-128))+rsi] |
| 914 add r12,rbx |
| 915 vpmuludq ymm12,ymm12,ymm10 |
| 916 vmovd xmm11,eax |
| 917 vmovdqu ymm0,YMMWORD[((-8+96-128))+rsi] |
| 918 vpaddq ymm1,ymm1,ymm12 |
| 919 vpmuludq ymm13,ymm13,ymm10 |
| 920 vpbroadcastq ymm11,xmm11 |
| 921 vmovdqu ymm12,YMMWORD[((-8+128-128))+rsi] |
| 922 vpaddq ymm2,ymm2,ymm13 |
| 923 vpmuludq ymm0,ymm0,ymm10 |
| 924 vmovdqu ymm13,YMMWORD[((-8+160-128))+rsi] |
| 925 vpaddq ymm3,ymm3,ymm0 |
| 926 vpmuludq ymm12,ymm12,ymm10 |
| 927 vmovdqu ymm0,YMMWORD[((-8+192-128))+rsi] |
| 928 vpaddq ymm4,ymm4,ymm12 |
| 929 vpmuludq ymm13,ymm13,ymm10 |
| 930 vmovdqu ymm12,YMMWORD[((-8+224-128))+rsi] |
| 931 vpaddq ymm5,ymm5,ymm13 |
| 932 vpmuludq ymm0,ymm0,ymm10 |
| 933 vmovdqu ymm13,YMMWORD[((-8+256-128))+rsi] |
| 934 vpaddq ymm6,ymm6,ymm0 |
| 935 vpmuludq ymm12,ymm12,ymm10 |
| 936 vmovdqu ymm9,YMMWORD[((-8+288-128))+rsi] |
| 937 vpaddq ymm7,ymm7,ymm12 |
| 938 vpmuludq ymm13,ymm13,ymm10 |
| 939 vpaddq ymm8,ymm8,ymm13 |
| 940 vpmuludq ymm9,ymm9,ymm10 |
| 941 vpbroadcastq ymm10,QWORD[16+r13] |
| 942 |
| 943 mov rdx,rax |
| 944 imul rax,QWORD[((-128))+rcx] |
| 945 add r10,rax |
| 946 vmovdqu ymm0,YMMWORD[((-8+32-128))+rcx] |
| 947 mov rax,rdx |
| 948 imul rax,QWORD[((8-128))+rcx] |
| 949 add r11,rax |
| 950 vmovdqu ymm12,YMMWORD[((-8+64-128))+rcx] |
| 951 shr r10,29 |
| 952 imul rdx,QWORD[((16-128))+rcx] |
| 953 add r12,rdx |
| 954 add r11,r10 |
| 955 |
| 956 vpmuludq ymm0,ymm0,ymm11 |
| 957 vmovq rbx,xmm10 |
| 958 vmovdqu ymm13,YMMWORD[((-8+96-128))+rcx] |
| 959 vpaddq ymm1,ymm1,ymm0 |
| 960 vpmuludq ymm12,ymm12,ymm11 |
| 961 vmovdqu ymm0,YMMWORD[((-8+128-128))+rcx] |
| 962 vpaddq ymm2,ymm2,ymm12 |
| 963 vpmuludq ymm13,ymm13,ymm11 |
| 964 vmovdqu ymm12,YMMWORD[((-8+160-128))+rcx] |
| 965 vpaddq ymm3,ymm3,ymm13 |
| 966 vpmuludq ymm0,ymm0,ymm11 |
| 967 vmovdqu ymm13,YMMWORD[((-8+192-128))+rcx] |
| 968 vpaddq ymm4,ymm4,ymm0 |
| 969 vpmuludq ymm12,ymm12,ymm11 |
| 970 vmovdqu ymm0,YMMWORD[((-8+224-128))+rcx] |
| 971 vpaddq ymm5,ymm5,ymm12 |
| 972 vpmuludq ymm13,ymm13,ymm11 |
| 973 vmovdqu ymm12,YMMWORD[((-8+256-128))+rcx] |
| 974 vpaddq ymm6,ymm6,ymm13 |
| 975 vpmuludq ymm0,ymm0,ymm11 |
| 976 vmovdqu ymm13,YMMWORD[((-8+288-128))+rcx] |
| 977 vpaddq ymm7,ymm7,ymm0 |
| 978 vpmuludq ymm12,ymm12,ymm11 |
| 979 vpaddq ymm8,ymm8,ymm12 |
| 980 vpmuludq ymm13,ymm13,ymm11 |
| 981 vpaddq ymm9,ymm9,ymm13 |
| 982 |
| 983 vmovdqu ymm0,YMMWORD[((-16+32-128))+rsi] |
| 984 mov rax,rbx |
| 985 imul rax,QWORD[((-128))+rsi] |
| 986 add rax,r11 |
| 987 |
| 988 vmovdqu ymm12,YMMWORD[((-16+64-128))+rsi] |
| 989 mov r11,rax |
| 990 imul eax,r8d |
| 991 and eax,0x1fffffff |
| 992 |
| 993 imul rbx,QWORD[((8-128))+rsi] |
| 994 add r12,rbx |
| 995 vpmuludq ymm0,ymm0,ymm10 |
| 996 vmovd xmm11,eax |
| 997 vmovdqu ymm13,YMMWORD[((-16+96-128))+rsi] |
| 998 vpaddq ymm1,ymm1,ymm0 |
| 999 vpmuludq ymm12,ymm12,ymm10 |
| 1000 vpbroadcastq ymm11,xmm11 |
| 1001 vmovdqu ymm0,YMMWORD[((-16+128-128))+rsi] |
| 1002 vpaddq ymm2,ymm2,ymm12 |
| 1003 vpmuludq ymm13,ymm13,ymm10 |
| 1004 vmovdqu ymm12,YMMWORD[((-16+160-128))+rsi] |
| 1005 vpaddq ymm3,ymm3,ymm13 |
| 1006 vpmuludq ymm0,ymm0,ymm10 |
| 1007 vmovdqu ymm13,YMMWORD[((-16+192-128))+rsi] |
| 1008 vpaddq ymm4,ymm4,ymm0 |
| 1009 vpmuludq ymm12,ymm12,ymm10 |
| 1010 vmovdqu ymm0,YMMWORD[((-16+224-128))+rsi] |
| 1011 vpaddq ymm5,ymm5,ymm12 |
| 1012 vpmuludq ymm13,ymm13,ymm10 |
| 1013 vmovdqu ymm12,YMMWORD[((-16+256-128))+rsi] |
| 1014 vpaddq ymm6,ymm6,ymm13 |
| 1015 vpmuludq ymm0,ymm0,ymm10 |
| 1016 vmovdqu ymm13,YMMWORD[((-16+288-128))+rsi] |
| 1017 vpaddq ymm7,ymm7,ymm0 |
| 1018 vpmuludq ymm12,ymm12,ymm10 |
| 1019 vpaddq ymm8,ymm8,ymm12 |
| 1020 vpmuludq ymm13,ymm13,ymm10 |
| 1021 vpbroadcastq ymm10,QWORD[24+r13] |
| 1022 vpaddq ymm9,ymm9,ymm13 |
| 1023 |
| 1024 vmovdqu ymm0,YMMWORD[((-16+32-128))+rcx] |
| 1025 mov rdx,rax |
| 1026 imul rax,QWORD[((-128))+rcx] |
| 1027 add r11,rax |
| 1028 vmovdqu ymm12,YMMWORD[((-16+64-128))+rcx] |
| 1029 imul rdx,QWORD[((8-128))+rcx] |
| 1030 add r12,rdx |
| 1031 shr r11,29 |
| 1032 |
| 1033 vpmuludq ymm0,ymm0,ymm11 |
| 1034 vmovq rbx,xmm10 |
| 1035 vmovdqu ymm13,YMMWORD[((-16+96-128))+rcx] |
| 1036 vpaddq ymm1,ymm1,ymm0 |
| 1037 vpmuludq ymm12,ymm12,ymm11 |
| 1038 vmovdqu ymm0,YMMWORD[((-16+128-128))+rcx] |
| 1039 vpaddq ymm2,ymm2,ymm12 |
| 1040 vpmuludq ymm13,ymm13,ymm11 |
| 1041 vmovdqu ymm12,YMMWORD[((-16+160-128))+rcx] |
| 1042 vpaddq ymm3,ymm3,ymm13 |
| 1043 vpmuludq ymm0,ymm0,ymm11 |
| 1044 vmovdqu ymm13,YMMWORD[((-16+192-128))+rcx] |
| 1045 vpaddq ymm4,ymm4,ymm0 |
| 1046 vpmuludq ymm12,ymm12,ymm11 |
| 1047 vmovdqu ymm0,YMMWORD[((-16+224-128))+rcx] |
| 1048 vpaddq ymm5,ymm5,ymm12 |
| 1049 vpmuludq ymm13,ymm13,ymm11 |
| 1050 vmovdqu ymm12,YMMWORD[((-16+256-128))+rcx] |
| 1051 vpaddq ymm6,ymm6,ymm13 |
| 1052 vpmuludq ymm0,ymm0,ymm11 |
| 1053 vmovdqu ymm13,YMMWORD[((-16+288-128))+rcx] |
| 1054 vpaddq ymm7,ymm7,ymm0 |
| 1055 vpmuludq ymm12,ymm12,ymm11 |
| 1056 vmovdqu ymm0,YMMWORD[((-24+32-128))+rsi] |
| 1057 vpaddq ymm8,ymm8,ymm12 |
| 1058 vpmuludq ymm13,ymm13,ymm11 |
| 1059 vmovdqu ymm12,YMMWORD[((-24+64-128))+rsi] |
| 1060 vpaddq ymm9,ymm9,ymm13 |
| 1061 |
| 1062 add r12,r11 |
| 1063 imul rbx,QWORD[((-128))+rsi] |
| 1064 add r12,rbx |
| 1065 |
| 1066 mov rax,r12 |
| 1067 imul eax,r8d |
| 1068 and eax,0x1fffffff |
| 1069 |
| 1070 vpmuludq ymm0,ymm0,ymm10 |
| 1071 vmovd xmm11,eax |
| 1072 vmovdqu ymm13,YMMWORD[((-24+96-128))+rsi] |
| 1073 vpaddq ymm1,ymm1,ymm0 |
| 1074 vpmuludq ymm12,ymm12,ymm10 |
| 1075 vpbroadcastq ymm11,xmm11 |
| 1076 vmovdqu ymm0,YMMWORD[((-24+128-128))+rsi] |
| 1077 vpaddq ymm2,ymm2,ymm12 |
| 1078 vpmuludq ymm13,ymm13,ymm10 |
| 1079 vmovdqu ymm12,YMMWORD[((-24+160-128))+rsi] |
| 1080 vpaddq ymm3,ymm3,ymm13 |
| 1081 vpmuludq ymm0,ymm0,ymm10 |
| 1082 vmovdqu ymm13,YMMWORD[((-24+192-128))+rsi] |
| 1083 vpaddq ymm4,ymm4,ymm0 |
| 1084 vpmuludq ymm12,ymm12,ymm10 |
| 1085 vmovdqu ymm0,YMMWORD[((-24+224-128))+rsi] |
| 1086 vpaddq ymm5,ymm5,ymm12 |
| 1087 vpmuludq ymm13,ymm13,ymm10 |
| 1088 vmovdqu ymm12,YMMWORD[((-24+256-128))+rsi] |
| 1089 vpaddq ymm6,ymm6,ymm13 |
| 1090 vpmuludq ymm0,ymm0,ymm10 |
| 1091 vmovdqu ymm13,YMMWORD[((-24+288-128))+rsi] |
| 1092 vpaddq ymm7,ymm7,ymm0 |
| 1093 vpmuludq ymm12,ymm12,ymm10 |
| 1094 vpaddq ymm8,ymm8,ymm12 |
| 1095 vpmuludq ymm13,ymm13,ymm10 |
| 1096 vpbroadcastq ymm10,QWORD[32+r13] |
| 1097 vpaddq ymm9,ymm9,ymm13 |
| 1098 add r13,32 |
| 1099 |
| 1100 vmovdqu ymm0,YMMWORD[((-24+32-128))+rcx] |
| 1101 imul rax,QWORD[((-128))+rcx] |
| 1102 add r12,rax |
| 1103 shr r12,29 |
| 1104 |
| 1105 vmovdqu ymm12,YMMWORD[((-24+64-128))+rcx] |
| 1106 vpmuludq ymm0,ymm0,ymm11 |
| 1107 vmovq rbx,xmm10 |
| 1108 vmovdqu ymm13,YMMWORD[((-24+96-128))+rcx] |
| 1109 vpaddq ymm0,ymm1,ymm0 |
| 1110 vpmuludq ymm12,ymm12,ymm11 |
| 1111 vmovdqu YMMWORD[rsp],ymm0 |
| 1112 vpaddq ymm1,ymm2,ymm12 |
| 1113 vmovdqu ymm0,YMMWORD[((-24+128-128))+rcx] |
| 1114 vpmuludq ymm13,ymm13,ymm11 |
| 1115 vmovdqu ymm12,YMMWORD[((-24+160-128))+rcx] |
| 1116 vpaddq ymm2,ymm3,ymm13 |
| 1117 vpmuludq ymm0,ymm0,ymm11 |
| 1118 vmovdqu ymm13,YMMWORD[((-24+192-128))+rcx] |
| 1119 vpaddq ymm3,ymm4,ymm0 |
| 1120 vpmuludq ymm12,ymm12,ymm11 |
| 1121 vmovdqu ymm0,YMMWORD[((-24+224-128))+rcx] |
| 1122 vpaddq ymm4,ymm5,ymm12 |
| 1123 vpmuludq ymm13,ymm13,ymm11 |
| 1124 vmovdqu ymm12,YMMWORD[((-24+256-128))+rcx] |
| 1125 vpaddq ymm5,ymm6,ymm13 |
| 1126 vpmuludq ymm0,ymm0,ymm11 |
| 1127 vmovdqu ymm13,YMMWORD[((-24+288-128))+rcx] |
| 1128 mov r9,r12 |
| 1129 vpaddq ymm6,ymm7,ymm0 |
| 1130 vpmuludq ymm12,ymm12,ymm11 |
| 1131 add r9,QWORD[rsp] |
| 1132 vpaddq ymm7,ymm8,ymm12 |
| 1133 vpmuludq ymm13,ymm13,ymm11 |
| 1134 vmovq xmm12,r12 |
| 1135 vpaddq ymm8,ymm9,ymm13 |
| 1136 |
| 1137 dec r14d |
| 1138 jnz NEAR $L$oop_mul_1024 |
| 1139 vpermq ymm15,ymm15,0 |
| 1140 vpaddq ymm0,ymm12,YMMWORD[rsp] |
| 1141 |
| 1142 vpsrlq ymm12,ymm0,29 |
| 1143 vpand ymm0,ymm0,ymm15 |
| 1144 vpsrlq ymm13,ymm1,29 |
| 1145 vpand ymm1,ymm1,ymm15 |
| 1146 vpsrlq ymm10,ymm2,29 |
| 1147 vpermq ymm12,ymm12,0x93 |
| 1148 vpand ymm2,ymm2,ymm15 |
| 1149 vpsrlq ymm11,ymm3,29 |
| 1150 vpermq ymm13,ymm13,0x93 |
| 1151 vpand ymm3,ymm3,ymm15 |
| 1152 |
| 1153 vpblendd ymm9,ymm12,ymm14,3 |
| 1154 vpermq ymm10,ymm10,0x93 |
| 1155 vpblendd ymm12,ymm13,ymm12,3 |
| 1156 vpermq ymm11,ymm11,0x93 |
| 1157 vpaddq ymm0,ymm0,ymm9 |
| 1158 vpblendd ymm13,ymm10,ymm13,3 |
| 1159 vpaddq ymm1,ymm1,ymm12 |
| 1160 vpblendd ymm10,ymm11,ymm10,3 |
| 1161 vpaddq ymm2,ymm2,ymm13 |
| 1162 vpblendd ymm11,ymm14,ymm11,3 |
| 1163 vpaddq ymm3,ymm3,ymm10 |
| 1164 vpaddq ymm4,ymm4,ymm11 |
| 1165 |
| 1166 vpsrlq ymm12,ymm0,29 |
| 1167 vpand ymm0,ymm0,ymm15 |
| 1168 vpsrlq ymm13,ymm1,29 |
| 1169 vpand ymm1,ymm1,ymm15 |
| 1170 vpsrlq ymm10,ymm2,29 |
| 1171 vpermq ymm12,ymm12,0x93 |
| 1172 vpand ymm2,ymm2,ymm15 |
| 1173 vpsrlq ymm11,ymm3,29 |
| 1174 vpermq ymm13,ymm13,0x93 |
| 1175 vpand ymm3,ymm3,ymm15 |
| 1176 vpermq ymm10,ymm10,0x93 |
| 1177 |
| 1178 vpblendd ymm9,ymm12,ymm14,3 |
| 1179 vpermq ymm11,ymm11,0x93 |
| 1180 vpblendd ymm12,ymm13,ymm12,3 |
| 1181 vpaddq ymm0,ymm0,ymm9 |
| 1182 vpblendd ymm13,ymm10,ymm13,3 |
| 1183 vpaddq ymm1,ymm1,ymm12 |
| 1184 vpblendd ymm10,ymm11,ymm10,3 |
| 1185 vpaddq ymm2,ymm2,ymm13 |
| 1186 vpblendd ymm11,ymm14,ymm11,3 |
| 1187 vpaddq ymm3,ymm3,ymm10 |
| 1188 vpaddq ymm4,ymm4,ymm11 |
| 1189 |
| 1190 vmovdqu YMMWORD[(0-128)+rdi],ymm0 |
| 1191 vmovdqu YMMWORD[(32-128)+rdi],ymm1 |
| 1192 vmovdqu YMMWORD[(64-128)+rdi],ymm2 |
| 1193 vmovdqu YMMWORD[(96-128)+rdi],ymm3 |
| 1194 vpsrlq ymm12,ymm4,29 |
| 1195 vpand ymm4,ymm4,ymm15 |
| 1196 vpsrlq ymm13,ymm5,29 |
| 1197 vpand ymm5,ymm5,ymm15 |
| 1198 vpsrlq ymm10,ymm6,29 |
| 1199 vpermq ymm12,ymm12,0x93 |
| 1200 vpand ymm6,ymm6,ymm15 |
| 1201 vpsrlq ymm11,ymm7,29 |
| 1202 vpermq ymm13,ymm13,0x93 |
| 1203 vpand ymm7,ymm7,ymm15 |
| 1204 vpsrlq ymm0,ymm8,29 |
| 1205 vpermq ymm10,ymm10,0x93 |
| 1206 vpand ymm8,ymm8,ymm15 |
| 1207 vpermq ymm11,ymm11,0x93 |
| 1208 |
| 1209 vpblendd ymm9,ymm12,ymm14,3 |
| 1210 vpermq ymm0,ymm0,0x93 |
| 1211 vpblendd ymm12,ymm13,ymm12,3 |
| 1212 vpaddq ymm4,ymm4,ymm9 |
| 1213 vpblendd ymm13,ymm10,ymm13,3 |
| 1214 vpaddq ymm5,ymm5,ymm12 |
| 1215 vpblendd ymm10,ymm11,ymm10,3 |
| 1216 vpaddq ymm6,ymm6,ymm13 |
| 1217 vpblendd ymm11,ymm0,ymm11,3 |
| 1218 vpaddq ymm7,ymm7,ymm10 |
| 1219 vpaddq ymm8,ymm8,ymm11 |
| 1220 |
| 1221 vpsrlq ymm12,ymm4,29 |
| 1222 vpand ymm4,ymm4,ymm15 |
| 1223 vpsrlq ymm13,ymm5,29 |
| 1224 vpand ymm5,ymm5,ymm15 |
| 1225 vpsrlq ymm10,ymm6,29 |
| 1226 vpermq ymm12,ymm12,0x93 |
| 1227 vpand ymm6,ymm6,ymm15 |
| 1228 vpsrlq ymm11,ymm7,29 |
| 1229 vpermq ymm13,ymm13,0x93 |
| 1230 vpand ymm7,ymm7,ymm15 |
| 1231 vpsrlq ymm0,ymm8,29 |
| 1232 vpermq ymm10,ymm10,0x93 |
| 1233 vpand ymm8,ymm8,ymm15 |
| 1234 vpermq ymm11,ymm11,0x93 |
| 1235 |
| 1236 vpblendd ymm9,ymm12,ymm14,3 |
| 1237 vpermq ymm0,ymm0,0x93 |
| 1238 vpblendd ymm12,ymm13,ymm12,3 |
| 1239 vpaddq ymm4,ymm4,ymm9 |
| 1240 vpblendd ymm13,ymm10,ymm13,3 |
| 1241 vpaddq ymm5,ymm5,ymm12 |
| 1242 vpblendd ymm10,ymm11,ymm10,3 |
| 1243 vpaddq ymm6,ymm6,ymm13 |
| 1244 vpblendd ymm11,ymm0,ymm11,3 |
| 1245 vpaddq ymm7,ymm7,ymm10 |
| 1246 vpaddq ymm8,ymm8,ymm11 |
| 1247 |
| 1248 vmovdqu YMMWORD[(128-128)+rdi],ymm4 |
| 1249 vmovdqu YMMWORD[(160-128)+rdi],ymm5 |
| 1250 vmovdqu YMMWORD[(192-128)+rdi],ymm6 |
| 1251 vmovdqu YMMWORD[(224-128)+rdi],ymm7 |
| 1252 vmovdqu YMMWORD[(256-128)+rdi],ymm8 |
| 1253 vzeroupper |
| 1254 |
| 1255 mov rax,rbp |
| 1256 |
| 1257 $L$mul_1024_in_tail: |
| 1258 movaps xmm6,XMMWORD[((-216))+rax] |
| 1259 movaps xmm7,XMMWORD[((-200))+rax] |
| 1260 movaps xmm8,XMMWORD[((-184))+rax] |
| 1261 movaps xmm9,XMMWORD[((-168))+rax] |
| 1262 movaps xmm10,XMMWORD[((-152))+rax] |
| 1263 movaps xmm11,XMMWORD[((-136))+rax] |
| 1264 movaps xmm12,XMMWORD[((-120))+rax] |
| 1265 movaps xmm13,XMMWORD[((-104))+rax] |
| 1266 movaps xmm14,XMMWORD[((-88))+rax] |
| 1267 movaps xmm15,XMMWORD[((-72))+rax] |
| 1268 mov r15,QWORD[((-48))+rax] |
| 1269 |
| 1270 mov r14,QWORD[((-40))+rax] |
| 1271 |
| 1272 mov r13,QWORD[((-32))+rax] |
| 1273 |
| 1274 mov r12,QWORD[((-24))+rax] |
| 1275 |
| 1276 mov rbp,QWORD[((-16))+rax] |
| 1277 |
| 1278 mov rbx,QWORD[((-8))+rax] |
| 1279 |
| 1280 lea rsp,[rax] |
| 1281 |
| 1282 $L$mul_1024_epilogue: |
| 1283 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1284 mov rsi,QWORD[16+rsp] |
| 1285 DB 0F3h,0C3h ;repret |
| 1286 |
| 1287 $L$SEH_end_rsaz_1024_mul_avx2: |
| 1288 global rsaz_1024_red2norm_avx2 |
| 1289 |
| 1290 ALIGN 32 |
| 1291 rsaz_1024_red2norm_avx2: |
| 1292 sub rdx,-128 |
| 1293 xor rax,rax |
| 1294 mov r8,QWORD[((-128))+rdx] |
| 1295 mov r9,QWORD[((-120))+rdx] |
| 1296 mov r10,QWORD[((-112))+rdx] |
| 1297 shl r8,0 |
| 1298 shl r9,29 |
| 1299 mov r11,r10 |
| 1300 shl r10,58 |
| 1301 shr r11,6 |
| 1302 add rax,r8 |
| 1303 add rax,r9 |
| 1304 add rax,r10 |
| 1305 adc r11,0 |
| 1306 mov QWORD[rcx],rax |
| 1307 mov rax,r11 |
| 1308 mov r8,QWORD[((-104))+rdx] |
| 1309 mov r9,QWORD[((-96))+rdx] |
| 1310 shl r8,23 |
| 1311 mov r10,r9 |
| 1312 shl r9,52 |
| 1313 shr r10,12 |
| 1314 add rax,r8 |
| 1315 add rax,r9 |
| 1316 adc r10,0 |
| 1317 mov QWORD[8+rcx],rax |
| 1318 mov rax,r10 |
| 1319 mov r11,QWORD[((-88))+rdx] |
| 1320 mov r8,QWORD[((-80))+rdx] |
| 1321 shl r11,17 |
| 1322 mov r9,r8 |
| 1323 shl r8,46 |
| 1324 shr r9,18 |
| 1325 add rax,r11 |
| 1326 add rax,r8 |
| 1327 adc r9,0 |
| 1328 mov QWORD[16+rcx],rax |
| 1329 mov rax,r9 |
| 1330 mov r10,QWORD[((-72))+rdx] |
| 1331 mov r11,QWORD[((-64))+rdx] |
| 1332 shl r10,11 |
| 1333 mov r8,r11 |
| 1334 shl r11,40 |
| 1335 shr r8,24 |
| 1336 add rax,r10 |
| 1337 add rax,r11 |
| 1338 adc r8,0 |
| 1339 mov QWORD[24+rcx],rax |
| 1340 mov rax,r8 |
| 1341 mov r9,QWORD[((-56))+rdx] |
| 1342 mov r10,QWORD[((-48))+rdx] |
| 1343 mov r11,QWORD[((-40))+rdx] |
| 1344 shl r9,5 |
| 1345 shl r10,34 |
| 1346 mov r8,r11 |
| 1347 shl r11,63 |
| 1348 shr r8,1 |
| 1349 add rax,r9 |
| 1350 add rax,r10 |
| 1351 add rax,r11 |
| 1352 adc r8,0 |
| 1353 mov QWORD[32+rcx],rax |
| 1354 mov rax,r8 |
| 1355 mov r9,QWORD[((-32))+rdx] |
| 1356 mov r10,QWORD[((-24))+rdx] |
| 1357 shl r9,28 |
| 1358 mov r11,r10 |
| 1359 shl r10,57 |
| 1360 shr r11,7 |
| 1361 add rax,r9 |
| 1362 add rax,r10 |
| 1363 adc r11,0 |
| 1364 mov QWORD[40+rcx],rax |
| 1365 mov rax,r11 |
| 1366 mov r8,QWORD[((-16))+rdx] |
| 1367 mov r9,QWORD[((-8))+rdx] |
| 1368 shl r8,22 |
| 1369 mov r10,r9 |
| 1370 shl r9,51 |
| 1371 shr r10,13 |
| 1372 add rax,r8 |
| 1373 add rax,r9 |
| 1374 adc r10,0 |
| 1375 mov QWORD[48+rcx],rax |
| 1376 mov rax,r10 |
| 1377 mov r11,QWORD[rdx] |
| 1378 mov r8,QWORD[8+rdx] |
| 1379 shl r11,16 |
| 1380 mov r9,r8 |
| 1381 shl r8,45 |
| 1382 shr r9,19 |
| 1383 add rax,r11 |
| 1384 add rax,r8 |
| 1385 adc r9,0 |
| 1386 mov QWORD[56+rcx],rax |
| 1387 mov rax,r9 |
| 1388 mov r10,QWORD[16+rdx] |
| 1389 mov r11,QWORD[24+rdx] |
| 1390 shl r10,10 |
| 1391 mov r8,r11 |
| 1392 shl r11,39 |
| 1393 shr r8,25 |
| 1394 add rax,r10 |
| 1395 add rax,r11 |
| 1396 adc r8,0 |
| 1397 mov QWORD[64+rcx],rax |
| 1398 mov rax,r8 |
| 1399 mov r9,QWORD[32+rdx] |
| 1400 mov r10,QWORD[40+rdx] |
| 1401 mov r11,QWORD[48+rdx] |
| 1402 shl r9,4 |
| 1403 shl r10,33 |
| 1404 mov r8,r11 |
| 1405 shl r11,62 |
| 1406 shr r8,2 |
| 1407 add rax,r9 |
| 1408 add rax,r10 |
| 1409 add rax,r11 |
| 1410 adc r8,0 |
| 1411 mov QWORD[72+rcx],rax |
| 1412 mov rax,r8 |
| 1413 mov r9,QWORD[56+rdx] |
| 1414 mov r10,QWORD[64+rdx] |
| 1415 shl r9,27 |
| 1416 mov r11,r10 |
| 1417 shl r10,56 |
| 1418 shr r11,8 |
| 1419 add rax,r9 |
| 1420 add rax,r10 |
| 1421 adc r11,0 |
| 1422 mov QWORD[80+rcx],rax |
| 1423 mov rax,r11 |
| 1424 mov r8,QWORD[72+rdx] |
| 1425 mov r9,QWORD[80+rdx] |
| 1426 shl r8,21 |
| 1427 mov r10,r9 |
| 1428 shl r9,50 |
| 1429 shr r10,14 |
| 1430 add rax,r8 |
| 1431 add rax,r9 |
| 1432 adc r10,0 |
| 1433 mov QWORD[88+rcx],rax |
| 1434 mov rax,r10 |
| 1435 mov r11,QWORD[88+rdx] |
| 1436 mov r8,QWORD[96+rdx] |
| 1437 shl r11,15 |
| 1438 mov r9,r8 |
| 1439 shl r8,44 |
| 1440 shr r9,20 |
| 1441 add rax,r11 |
| 1442 add rax,r8 |
| 1443 adc r9,0 |
| 1444 mov QWORD[96+rcx],rax |
| 1445 mov rax,r9 |
| 1446 mov r10,QWORD[104+rdx] |
| 1447 mov r11,QWORD[112+rdx] |
| 1448 shl r10,9 |
| 1449 mov r8,r11 |
| 1450 shl r11,38 |
| 1451 shr r8,26 |
| 1452 add rax,r10 |
| 1453 add rax,r11 |
| 1454 adc r8,0 |
| 1455 mov QWORD[104+rcx],rax |
| 1456 mov rax,r8 |
| 1457 mov r9,QWORD[120+rdx] |
| 1458 mov r10,QWORD[128+rdx] |
| 1459 mov r11,QWORD[136+rdx] |
| 1460 shl r9,3 |
| 1461 shl r10,32 |
| 1462 mov r8,r11 |
| 1463 shl r11,61 |
| 1464 shr r8,3 |
| 1465 add rax,r9 |
| 1466 add rax,r10 |
| 1467 add rax,r11 |
| 1468 adc r8,0 |
| 1469 mov QWORD[112+rcx],rax |
| 1470 mov rax,r8 |
| 1471 mov r9,QWORD[144+rdx] |
| 1472 mov r10,QWORD[152+rdx] |
| 1473 shl r9,26 |
| 1474 mov r11,r10 |
| 1475 shl r10,55 |
| 1476 shr r11,9 |
| 1477 add rax,r9 |
| 1478 add rax,r10 |
| 1479 adc r11,0 |
| 1480 mov QWORD[120+rcx],rax |
| 1481 mov rax,r11 |
| 1482 DB 0F3h,0C3h ;repret |
| 1483 |
| 1484 |
| 1485 global rsaz_1024_norm2red_avx2 |
| 1486 |
| 1487 ALIGN 32 |
| 1488 rsaz_1024_norm2red_avx2: |
| 1489 sub rcx,-128 |
| 1490 mov r8,QWORD[rdx] |
| 1491 mov eax,0x1fffffff |
| 1492 mov r9,QWORD[8+rdx] |
| 1493 mov r11,r8 |
| 1494 shr r11,0 |
| 1495 and r11,rax |
| 1496 mov QWORD[((-128))+rcx],r11 |
| 1497 mov r10,r8 |
| 1498 shr r10,29 |
| 1499 and r10,rax |
| 1500 mov QWORD[((-120))+rcx],r10 |
| 1501 shrd r8,r9,58 |
| 1502 and r8,rax |
| 1503 mov QWORD[((-112))+rcx],r8 |
| 1504 mov r10,QWORD[16+rdx] |
| 1505 mov r8,r9 |
| 1506 shr r8,23 |
| 1507 and r8,rax |
| 1508 mov QWORD[((-104))+rcx],r8 |
| 1509 shrd r9,r10,52 |
| 1510 and r9,rax |
| 1511 mov QWORD[((-96))+rcx],r9 |
| 1512 mov r11,QWORD[24+rdx] |
| 1513 mov r9,r10 |
| 1514 shr r9,17 |
| 1515 and r9,rax |
| 1516 mov QWORD[((-88))+rcx],r9 |
| 1517 shrd r10,r11,46 |
| 1518 and r10,rax |
| 1519 mov QWORD[((-80))+rcx],r10 |
| 1520 mov r8,QWORD[32+rdx] |
| 1521 mov r10,r11 |
| 1522 shr r10,11 |
| 1523 and r10,rax |
| 1524 mov QWORD[((-72))+rcx],r10 |
| 1525 shrd r11,r8,40 |
| 1526 and r11,rax |
| 1527 mov QWORD[((-64))+rcx],r11 |
| 1528 mov r9,QWORD[40+rdx] |
| 1529 mov r11,r8 |
| 1530 shr r11,5 |
| 1531 and r11,rax |
| 1532 mov QWORD[((-56))+rcx],r11 |
| 1533 mov r10,r8 |
| 1534 shr r10,34 |
| 1535 and r10,rax |
| 1536 mov QWORD[((-48))+rcx],r10 |
| 1537 shrd r8,r9,63 |
| 1538 and r8,rax |
| 1539 mov QWORD[((-40))+rcx],r8 |
| 1540 mov r10,QWORD[48+rdx] |
| 1541 mov r8,r9 |
| 1542 shr r8,28 |
| 1543 and r8,rax |
| 1544 mov QWORD[((-32))+rcx],r8 |
| 1545 shrd r9,r10,57 |
| 1546 and r9,rax |
| 1547 mov QWORD[((-24))+rcx],r9 |
| 1548 mov r11,QWORD[56+rdx] |
| 1549 mov r9,r10 |
| 1550 shr r9,22 |
| 1551 and r9,rax |
| 1552 mov QWORD[((-16))+rcx],r9 |
| 1553 shrd r10,r11,51 |
| 1554 and r10,rax |
| 1555 mov QWORD[((-8))+rcx],r10 |
| 1556 mov r8,QWORD[64+rdx] |
| 1557 mov r10,r11 |
| 1558 shr r10,16 |
| 1559 and r10,rax |
| 1560 mov QWORD[rcx],r10 |
| 1561 shrd r11,r8,45 |
| 1562 and r11,rax |
| 1563 mov QWORD[8+rcx],r11 |
| 1564 mov r9,QWORD[72+rdx] |
| 1565 mov r11,r8 |
| 1566 shr r11,10 |
| 1567 and r11,rax |
| 1568 mov QWORD[16+rcx],r11 |
| 1569 shrd r8,r9,39 |
| 1570 and r8,rax |
| 1571 mov QWORD[24+rcx],r8 |
| 1572 mov r10,QWORD[80+rdx] |
| 1573 mov r8,r9 |
| 1574 shr r8,4 |
| 1575 and r8,rax |
| 1576 mov QWORD[32+rcx],r8 |
| 1577 mov r11,r9 |
| 1578 shr r11,33 |
| 1579 and r11,rax |
| 1580 mov QWORD[40+rcx],r11 |
| 1581 shrd r9,r10,62 |
| 1582 and r9,rax |
| 1583 mov QWORD[48+rcx],r9 |
| 1584 mov r11,QWORD[88+rdx] |
| 1585 mov r9,r10 |
| 1586 shr r9,27 |
| 1587 and r9,rax |
| 1588 mov QWORD[56+rcx],r9 |
| 1589 shrd r10,r11,56 |
| 1590 and r10,rax |
| 1591 mov QWORD[64+rcx],r10 |
| 1592 mov r8,QWORD[96+rdx] |
| 1593 mov r10,r11 |
| 1594 shr r10,21 |
| 1595 and r10,rax |
| 1596 mov QWORD[72+rcx],r10 |
| 1597 shrd r11,r8,50 |
| 1598 and r11,rax |
| 1599 mov QWORD[80+rcx],r11 |
| 1600 mov r9,QWORD[104+rdx] |
| 1601 mov r11,r8 |
| 1602 shr r11,15 |
| 1603 and r11,rax |
| 1604 mov QWORD[88+rcx],r11 |
| 1605 shrd r8,r9,44 |
| 1606 and r8,rax |
| 1607 mov QWORD[96+rcx],r8 |
| 1608 mov r10,QWORD[112+rdx] |
| 1609 mov r8,r9 |
| 1610 shr r8,9 |
| 1611 and r8,rax |
| 1612 mov QWORD[104+rcx],r8 |
| 1613 shrd r9,r10,38 |
| 1614 and r9,rax |
| 1615 mov QWORD[112+rcx],r9 |
| 1616 mov r11,QWORD[120+rdx] |
| 1617 mov r9,r10 |
| 1618 shr r9,3 |
| 1619 and r9,rax |
| 1620 mov QWORD[120+rcx],r9 |
| 1621 mov r8,r10 |
| 1622 shr r8,32 |
| 1623 and r8,rax |
| 1624 mov QWORD[128+rcx],r8 |
| 1625 shrd r10,r11,61 |
| 1626 and r10,rax |
| 1627 mov QWORD[136+rcx],r10 |
| 1628 xor r8,r8 |
| 1629 mov r10,r11 |
| 1630 shr r10,26 |
| 1631 and r10,rax |
| 1632 mov QWORD[144+rcx],r10 |
| 1633 shrd r11,r8,55 |
| 1634 and r11,rax |
| 1635 mov QWORD[152+rcx],r11 |
| 1636 mov QWORD[160+rcx],r8 |
| 1637 mov QWORD[168+rcx],r8 |
| 1638 mov QWORD[176+rcx],r8 |
| 1639 mov QWORD[184+rcx],r8 |
| 1640 DB 0F3h,0C3h ;repret |
| 1641 |
| 1642 global rsaz_1024_scatter5_avx2 |
| 1643 |
| 1644 ALIGN 32 |
| 1645 rsaz_1024_scatter5_avx2: |
| 1646 vzeroupper |
| 1647 vmovdqu ymm5,YMMWORD[$L$scatter_permd] |
| 1648 shl r8d,4 |
| 1649 lea rcx,[r8*1+rcx] |
| 1650 mov eax,9 |
| 1651 jmp NEAR $L$oop_scatter_1024 |
| 1652 |
| 1653 ALIGN 32 |
| 1654 $L$oop_scatter_1024: |
| 1655 vmovdqu ymm0,YMMWORD[rdx] |
| 1656 lea rdx,[32+rdx] |
| 1657 vpermd ymm0,ymm5,ymm0 |
| 1658 vmovdqu XMMWORD[rcx],xmm0 |
| 1659 lea rcx,[512+rcx] |
| 1660 dec eax |
| 1661 jnz NEAR $L$oop_scatter_1024 |
| 1662 |
| 1663 vzeroupper |
| 1664 DB 0F3h,0C3h ;repret |
| 1665 |
| 1666 |
| 1667 global rsaz_1024_gather5_avx2 |
| 1668 |
| 1669 ALIGN 32 |
| 1670 rsaz_1024_gather5_avx2: |
| 1671 |
| 1672 vzeroupper |
| 1673 mov r11,rsp |
| 1674 |
| 1675 lea rax,[((-136))+rsp] |
| 1676 $L$SEH_begin_rsaz_1024_gather5: |
| 1677 |
| 1678 DB 0x48,0x8d,0x60,0xe0 |
| 1679 DB 0xc5,0xf8,0x29,0x70,0xe0 |
| 1680 DB 0xc5,0xf8,0x29,0x78,0xf0 |
| 1681 DB 0xc5,0x78,0x29,0x40,0x00 |
| 1682 DB 0xc5,0x78,0x29,0x48,0x10 |
| 1683 DB 0xc5,0x78,0x29,0x50,0x20 |
| 1684 DB 0xc5,0x78,0x29,0x58,0x30 |
| 1685 DB 0xc5,0x78,0x29,0x60,0x40 |
| 1686 DB 0xc5,0x78,0x29,0x68,0x50 |
| 1687 DB 0xc5,0x78,0x29,0x70,0x60 |
| 1688 DB 0xc5,0x78,0x29,0x78,0x70 |
| 1689 lea rsp,[((-256))+rsp] |
| 1690 and rsp,-32 |
| 1691 lea r10,[$L$inc] |
| 1692 lea rax,[((-128))+rsp] |
| 1693 |
| 1694 vmovd xmm4,r8d |
| 1695 vmovdqa ymm0,YMMWORD[r10] |
| 1696 vmovdqa ymm1,YMMWORD[32+r10] |
| 1697 vmovdqa ymm5,YMMWORD[64+r10] |
| 1698 vpbroadcastd ymm4,xmm4 |
| 1699 |
| 1700 vpaddd ymm2,ymm0,ymm5 |
| 1701 vpcmpeqd ymm0,ymm0,ymm4 |
| 1702 vpaddd ymm3,ymm1,ymm5 |
| 1703 vpcmpeqd ymm1,ymm1,ymm4 |
| 1704 vmovdqa YMMWORD[(0+128)+rax],ymm0 |
| 1705 vpaddd ymm0,ymm2,ymm5 |
| 1706 vpcmpeqd ymm2,ymm2,ymm4 |
| 1707 vmovdqa YMMWORD[(32+128)+rax],ymm1 |
| 1708 vpaddd ymm1,ymm3,ymm5 |
| 1709 vpcmpeqd ymm3,ymm3,ymm4 |
| 1710 vmovdqa YMMWORD[(64+128)+rax],ymm2 |
| 1711 vpaddd ymm2,ymm0,ymm5 |
| 1712 vpcmpeqd ymm0,ymm0,ymm4 |
| 1713 vmovdqa YMMWORD[(96+128)+rax],ymm3 |
| 1714 vpaddd ymm3,ymm1,ymm5 |
| 1715 vpcmpeqd ymm1,ymm1,ymm4 |
| 1716 vmovdqa YMMWORD[(128+128)+rax],ymm0 |
| 1717 vpaddd ymm8,ymm2,ymm5 |
| 1718 vpcmpeqd ymm2,ymm2,ymm4 |
| 1719 vmovdqa YMMWORD[(160+128)+rax],ymm1 |
| 1720 vpaddd ymm9,ymm3,ymm5 |
| 1721 vpcmpeqd ymm3,ymm3,ymm4 |
| 1722 vmovdqa YMMWORD[(192+128)+rax],ymm2 |
| 1723 vpaddd ymm10,ymm8,ymm5 |
| 1724 vpcmpeqd ymm8,ymm8,ymm4 |
| 1725 vmovdqa YMMWORD[(224+128)+rax],ymm3 |
| 1726 vpaddd ymm11,ymm9,ymm5 |
| 1727 vpcmpeqd ymm9,ymm9,ymm4 |
| 1728 vpaddd ymm12,ymm10,ymm5 |
| 1729 vpcmpeqd ymm10,ymm10,ymm4 |
| 1730 vpaddd ymm13,ymm11,ymm5 |
| 1731 vpcmpeqd ymm11,ymm11,ymm4 |
| 1732 vpaddd ymm14,ymm12,ymm5 |
| 1733 vpcmpeqd ymm12,ymm12,ymm4 |
| 1734 vpaddd ymm15,ymm13,ymm5 |
| 1735 vpcmpeqd ymm13,ymm13,ymm4 |
| 1736 vpcmpeqd ymm14,ymm14,ymm4 |
| 1737 vpcmpeqd ymm15,ymm15,ymm4 |
| 1738 |
| 1739 vmovdqa ymm7,YMMWORD[((-32))+r10] |
| 1740 lea rdx,[128+rdx] |
| 1741 mov r8d,9 |
| 1742 |
| 1743 $L$oop_gather_1024: |
| 1744 vmovdqa ymm0,YMMWORD[((0-128))+rdx] |
| 1745 vmovdqa ymm1,YMMWORD[((32-128))+rdx] |
| 1746 vmovdqa ymm2,YMMWORD[((64-128))+rdx] |
| 1747 vmovdqa ymm3,YMMWORD[((96-128))+rdx] |
| 1748 vpand ymm0,ymm0,YMMWORD[((0+128))+rax] |
| 1749 vpand ymm1,ymm1,YMMWORD[((32+128))+rax] |
| 1750 vpand ymm2,ymm2,YMMWORD[((64+128))+rax] |
| 1751 vpor ymm4,ymm1,ymm0 |
| 1752 vpand ymm3,ymm3,YMMWORD[((96+128))+rax] |
| 1753 vmovdqa ymm0,YMMWORD[((128-128))+rdx] |
| 1754 vmovdqa ymm1,YMMWORD[((160-128))+rdx] |
| 1755 vpor ymm5,ymm3,ymm2 |
| 1756 vmovdqa ymm2,YMMWORD[((192-128))+rdx] |
| 1757 vmovdqa ymm3,YMMWORD[((224-128))+rdx] |
| 1758 vpand ymm0,ymm0,YMMWORD[((128+128))+rax] |
| 1759 vpand ymm1,ymm1,YMMWORD[((160+128))+rax] |
| 1760 vpand ymm2,ymm2,YMMWORD[((192+128))+rax] |
| 1761 vpor ymm4,ymm4,ymm0 |
| 1762 vpand ymm3,ymm3,YMMWORD[((224+128))+rax] |
| 1763 vpand ymm0,ymm8,YMMWORD[((256-128))+rdx] |
| 1764 vpor ymm5,ymm5,ymm1 |
| 1765 vpand ymm1,ymm9,YMMWORD[((288-128))+rdx] |
| 1766 vpor ymm4,ymm4,ymm2 |
| 1767 vpand ymm2,ymm10,YMMWORD[((320-128))+rdx] |
| 1768 vpor ymm5,ymm5,ymm3 |
| 1769 vpand ymm3,ymm11,YMMWORD[((352-128))+rdx] |
| 1770 vpor ymm4,ymm4,ymm0 |
| 1771 vpand ymm0,ymm12,YMMWORD[((384-128))+rdx] |
| 1772 vpor ymm5,ymm5,ymm1 |
| 1773 vpand ymm1,ymm13,YMMWORD[((416-128))+rdx] |
| 1774 vpor ymm4,ymm4,ymm2 |
| 1775 vpand ymm2,ymm14,YMMWORD[((448-128))+rdx] |
| 1776 vpor ymm5,ymm5,ymm3 |
| 1777 vpand ymm3,ymm15,YMMWORD[((480-128))+rdx] |
| 1778 lea rdx,[512+rdx] |
| 1779 vpor ymm4,ymm4,ymm0 |
| 1780 vpor ymm5,ymm5,ymm1 |
| 1781 vpor ymm4,ymm4,ymm2 |
| 1782 vpor ymm5,ymm5,ymm3 |
| 1783 |
| 1784 vpor ymm4,ymm4,ymm5 |
| 1785 vextracti128 xmm5,ymm4,1 |
| 1786 vpor xmm5,xmm5,xmm4 |
| 1787 vpermd ymm5,ymm7,ymm5 |
| 1788 vmovdqu YMMWORD[rcx],ymm5 |
| 1789 lea rcx,[32+rcx] |
| 1790 dec r8d |
| 1791 jnz NEAR $L$oop_gather_1024 |
| 1792 |
| 1793 vpxor ymm0,ymm0,ymm0 |
| 1794 vmovdqu YMMWORD[rcx],ymm0 |
| 1795 vzeroupper |
| 1796 movaps xmm6,XMMWORD[((-168))+r11] |
| 1797 movaps xmm7,XMMWORD[((-152))+r11] |
| 1798 movaps xmm8,XMMWORD[((-136))+r11] |
| 1799 movaps xmm9,XMMWORD[((-120))+r11] |
| 1800 movaps xmm10,XMMWORD[((-104))+r11] |
| 1801 movaps xmm11,XMMWORD[((-88))+r11] |
| 1802 movaps xmm12,XMMWORD[((-72))+r11] |
| 1803 movaps xmm13,XMMWORD[((-56))+r11] |
| 1804 movaps xmm14,XMMWORD[((-40))+r11] |
| 1805 movaps xmm15,XMMWORD[((-24))+r11] |
| 1806 lea rsp,[r11] |
| 1807 |
| 1808 DB 0F3h,0C3h ;repret |
| 1809 |
| 1810 $L$SEH_end_rsaz_1024_gather5: |
| 1811 |
| 1812 EXTERN OPENSSL_ia32cap_P |
| 8 global rsaz_avx2_eligible | 1813 global rsaz_avx2_eligible |
| 9 | 1814 |
| 1815 ALIGN 32 |
| 10 rsaz_avx2_eligible: | 1816 rsaz_avx2_eligible: |
| 11 » xor» eax,eax | 1817 » mov» eax,DWORD[((OPENSSL_ia32cap_P+8))] |
| 1818 » mov» ecx,524544 |
| 1819 » mov» edx,0 |
| 1820 » and» ecx,eax |
| 1821 » cmp» ecx,524544 |
| 1822 » cmove» eax,edx |
| 1823 » and» eax,32 |
| 1824 » shr» eax,5 |
| 12 DB 0F3h,0C3h ;repret | 1825 DB 0F3h,0C3h ;repret |
| 13 | 1826 |
| 14 | 1827 |
| 15 global» rsaz_1024_sqr_avx2 | 1828 ALIGN» 64 |
| 16 global» rsaz_1024_mul_avx2 | 1829 $L$and_mask: |
| 17 global» rsaz_1024_norm2red_avx2 | 1830 » DQ» 0x1fffffff,0x1fffffff,0x1fffffff,-1 |
| 18 global» rsaz_1024_red2norm_avx2 | 1831 $L$scatter_permd: |
| 19 global» rsaz_1024_scatter5_avx2 | 1832 » DD» 0,2,4,6,7,7,7,7 |
| 20 global» rsaz_1024_gather5_avx2 | 1833 $L$gather_permd: |
| 21 | 1834 » DD» 0,7,1,7,2,7,3,7 |
| 22 rsaz_1024_sqr_avx2: | 1835 $L$inc: |
| 23 rsaz_1024_mul_avx2: | 1836 » DD» 0,0,0,0,1,1,1,1 |
| 24 rsaz_1024_norm2red_avx2: | 1837 » DD» 2,2,2,2,3,3,3,3 |
| 25 rsaz_1024_red2norm_avx2: | 1838 » DD» 4,4,4,4,4,4,4,4 |
| 26 rsaz_1024_scatter5_avx2: | 1839 ALIGN» 64 |
| 27 rsaz_1024_gather5_avx2: | 1840 EXTERN» __imp_RtlVirtualUnwind |
| 28 DB» 0x0f,0x0b | 1841 |
| 1842 ALIGN» 16 |
| 1843 rsaz_se_handler: |
| 1844 » push» rsi |
| 1845 » push» rdi |
| 1846 » push» rbx |
| 1847 » push» rbp |
| 1848 » push» r12 |
| 1849 » push» r13 |
| 1850 » push» r14 |
| 1851 » push» r15 |
| 1852 » pushfq |
| 1853 » sub» rsp,64 |
| 1854 |
| 1855 » mov» rax,QWORD[120+r8] |
| 1856 » mov» rbx,QWORD[248+r8] |
| 1857 |
| 1858 » mov» rsi,QWORD[8+r9] |
| 1859 » mov» r11,QWORD[56+r9] |
| 1860 |
| 1861 » mov» r10d,DWORD[r11] |
| 1862 » lea» r10,[r10*1+rsi] |
| 1863 » cmp» rbx,r10 |
| 1864 » jb» NEAR $L$common_seh_tail |
| 1865 |
| 1866 » mov» r10d,DWORD[4+r11] |
| 1867 » lea» r10,[r10*1+rsi] |
| 1868 » cmp» rbx,r10 |
| 1869 » jae» NEAR $L$common_seh_tail |
| 1870 |
| 1871 » mov» rbp,QWORD[160+r8] |
| 1872 |
| 1873 » mov» r10d,DWORD[8+r11] |
| 1874 » lea» r10,[r10*1+rsi] |
| 1875 » cmp» rbx,r10 |
| 1876 » cmovc» rax,rbp |
| 1877 |
| 1878 » mov» r15,QWORD[((-48))+rax] |
| 1879 » mov» r14,QWORD[((-40))+rax] |
| 1880 » mov» r13,QWORD[((-32))+rax] |
| 1881 » mov» r12,QWORD[((-24))+rax] |
| 1882 » mov» rbp,QWORD[((-16))+rax] |
| 1883 » mov» rbx,QWORD[((-8))+rax] |
| 1884 » mov» QWORD[240+r8],r15 |
| 1885 » mov» QWORD[232+r8],r14 |
| 1886 » mov» QWORD[224+r8],r13 |
| 1887 » mov» QWORD[216+r8],r12 |
| 1888 » mov» QWORD[160+r8],rbp |
| 1889 » mov» QWORD[144+r8],rbx |
| 1890 |
| 1891 » lea» rsi,[((-216))+rax] |
| 1892 » lea» rdi,[512+r8] |
| 1893 » mov» ecx,20 |
| 1894 » DD» 0xa548f3fc |
| 1895 |
| 1896 $L$common_seh_tail: |
| 1897 » mov» rdi,QWORD[8+rax] |
| 1898 » mov» rsi,QWORD[16+rax] |
| 1899 » mov» QWORD[152+r8],rax |
| 1900 » mov» QWORD[168+r8],rsi |
| 1901 » mov» QWORD[176+r8],rdi |
| 1902 |
| 1903 » mov» rdi,QWORD[40+r9] |
| 1904 » mov» rsi,r8 |
| 1905 » mov» ecx,154 |
| 1906 » DD» 0xa548f3fc |
| 1907 |
| 1908 » mov» rsi,r9 |
| 1909 » xor» rcx,rcx |
| 1910 » mov» rdx,QWORD[8+rsi] |
| 1911 » mov» r8,QWORD[rsi] |
| 1912 » mov» r9,QWORD[16+rsi] |
| 1913 » mov» r10,QWORD[40+rsi] |
| 1914 » lea» r11,[56+rsi] |
| 1915 » lea» r12,[24+rsi] |
| 1916 » mov» QWORD[32+rsp],r10 |
| 1917 » mov» QWORD[40+rsp],r11 |
| 1918 » mov» QWORD[48+rsp],r12 |
| 1919 » mov» QWORD[56+rsp],rcx |
| 1920 » call» QWORD[__imp_RtlVirtualUnwind] |
| 1921 |
| 1922 » mov» eax,1 |
| 1923 » add» rsp,64 |
| 1924 » popfq |
| 1925 » pop» r15 |
| 1926 » pop» r14 |
| 1927 » pop» r13 |
| 1928 » pop» r12 |
| 1929 » pop» rbp |
| 1930 » pop» rbx |
| 1931 » pop» rdi |
| 1932 » pop» rsi |
| 29 DB 0F3h,0C3h ;repret | 1933 DB 0F3h,0C3h ;repret |
| 30 | 1934 |
| 1935 |
| 1936 section .pdata rdata align=4 |
| 1937 ALIGN 4 |
| 1938 DD $L$SEH_begin_rsaz_1024_sqr_avx2 wrt ..imagebase |
| 1939 DD $L$SEH_end_rsaz_1024_sqr_avx2 wrt ..imagebase |
| 1940 DD $L$SEH_info_rsaz_1024_sqr_avx2 wrt ..imagebase |
| 1941 |
| 1942 DD $L$SEH_begin_rsaz_1024_mul_avx2 wrt ..imagebase |
| 1943 DD $L$SEH_end_rsaz_1024_mul_avx2 wrt ..imagebase |
| 1944 DD $L$SEH_info_rsaz_1024_mul_avx2 wrt ..imagebase |
| 1945 |
| 1946 DD $L$SEH_begin_rsaz_1024_gather5 wrt ..imagebase |
| 1947 DD $L$SEH_end_rsaz_1024_gather5 wrt ..imagebase |
| 1948 DD $L$SEH_info_rsaz_1024_gather5 wrt ..imagebase |
| 1949 section .xdata rdata align=8 |
| 1950 ALIGN 8 |
| 1951 $L$SEH_info_rsaz_1024_sqr_avx2: |
| 1952 DB 9,0,0,0 |
| 1953 DD rsaz_se_handler wrt ..imagebase |
| 1954 DD $L$sqr_1024_body wrt ..imagebase,$L$sqr_1024_epilogue wrt ..imag
ebase,$L$sqr_1024_in_tail wrt ..imagebase |
| 1955 DD 0 |
| 1956 $L$SEH_info_rsaz_1024_mul_avx2: |
| 1957 DB 9,0,0,0 |
| 1958 DD rsaz_se_handler wrt ..imagebase |
| 1959 DD $L$mul_1024_body wrt ..imagebase,$L$mul_1024_epilogue wrt ..imag
ebase,$L$mul_1024_in_tail wrt ..imagebase |
| 1960 DD 0 |
| 1961 $L$SEH_info_rsaz_1024_gather5: |
| 1962 DB 0x01,0x36,0x17,0x0b |
| 1963 DB 0x36,0xf8,0x09,0x00 |
| 1964 DB 0x31,0xe8,0x08,0x00 |
| 1965 DB 0x2c,0xd8,0x07,0x00 |
| 1966 DB 0x27,0xc8,0x06,0x00 |
| 1967 DB 0x22,0xb8,0x05,0x00 |
| 1968 DB 0x1d,0xa8,0x04,0x00 |
| 1969 DB 0x18,0x98,0x03,0x00 |
| 1970 DB 0x13,0x88,0x02,0x00 |
| 1971 DB 0x0e,0x78,0x01,0x00 |
| 1972 DB 0x09,0x68,0x00,0x00 |
| 1973 DB 0x04,0x01,0x15,0x00 |
| 1974 DB 0x00,0xb3,0x00,0x00 |
| OLD | NEW |
|---|
Chromium Code Reviews
Issue 2734373002:
Roll src/third_party/boringssl/src d6c22ee93..d04ca9535 (Closed)
