| 1 | /* |
| 2 | * Generic vectorized operation runtime |
| 3 | * |
| 4 | * Copyright (c) 2018 Linaro |
| 5 | * |
| 6 | * This library is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU Lesser General Public |
| 8 | * License as published by the Free Software Foundation; either |
| 9 | * version 2.1 of the License, or (at your option) any later version. |
| 10 | * |
| 11 | * This library is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 14 | * Lesser General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU Lesser General Public |
| 17 | * License along with this library; if not, see <http://www.gnu.org/licenses/>. |
| 18 | */ |
| 19 | |
| 20 | #include "qemu/osdep.h" |
| 21 | #include "qemu/host-utils.h" |
| 22 | #include "exec/helper-proto-common.h" |
| 23 | #include "tcg/tcg-gvec-desc.h" |
| 24 | |
| 25 | |
| 26 | static inline void clear_high(void *d, intptr_t oprsz, uint32_t desc) |
| 27 | { |
| 28 | intptr_t maxsz = simd_maxsz(desc); |
| 29 | intptr_t i; |
| 30 | |
| 31 | if (unlikely(maxsz > oprsz)) { |
| 32 | for (i = oprsz; i < maxsz; i += sizeof(uint64_t)) { |
| 33 | *(uint64_t *)(d + i) = 0; |
| 34 | } |
| 35 | } |
| 36 | } |
| 37 | |
| 38 | void HELPER(gvec_add8)(void *d, void *a, void *b, uint32_t desc) |
| 39 | { |
| 40 | intptr_t oprsz = simd_oprsz(desc); |
| 41 | intptr_t i; |
| 42 | |
| 43 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 44 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) + *(uint8_t *)(b + i); |
| 45 | } |
| 46 | clear_high(d, oprsz, desc); |
| 47 | } |
| 48 | |
| 49 | void HELPER(gvec_add16)(void *d, void *a, void *b, uint32_t desc) |
| 50 | { |
| 51 | intptr_t oprsz = simd_oprsz(desc); |
| 52 | intptr_t i; |
| 53 | |
| 54 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 55 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) + *(uint16_t *)(b + i); |
| 56 | } |
| 57 | clear_high(d, oprsz, desc); |
| 58 | } |
| 59 | |
| 60 | void HELPER(gvec_add32)(void *d, void *a, void *b, uint32_t desc) |
| 61 | { |
| 62 | intptr_t oprsz = simd_oprsz(desc); |
| 63 | intptr_t i; |
| 64 | |
| 65 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 66 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) + *(uint32_t *)(b + i); |
| 67 | } |
| 68 | clear_high(d, oprsz, desc); |
| 69 | } |
| 70 | |
| 71 | void HELPER(gvec_add64)(void *d, void *a, void *b, uint32_t desc) |
| 72 | { |
| 73 | intptr_t oprsz = simd_oprsz(desc); |
| 74 | intptr_t i; |
| 75 | |
| 76 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 77 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) + *(uint64_t *)(b + i); |
| 78 | } |
| 79 | clear_high(d, oprsz, desc); |
| 80 | } |
| 81 | |
| 82 | void HELPER(gvec_adds8)(void *d, void *a, uint64_t b, uint32_t desc) |
| 83 | { |
| 84 | intptr_t oprsz = simd_oprsz(desc); |
| 85 | intptr_t i; |
| 86 | |
| 87 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 88 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) + (uint8_t)b; |
| 89 | } |
| 90 | clear_high(d, oprsz, desc); |
| 91 | } |
| 92 | |
| 93 | void HELPER(gvec_adds16)(void *d, void *a, uint64_t b, uint32_t desc) |
| 94 | { |
| 95 | intptr_t oprsz = simd_oprsz(desc); |
| 96 | intptr_t i; |
| 97 | |
| 98 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 99 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) + (uint16_t)b; |
| 100 | } |
| 101 | clear_high(d, oprsz, desc); |
| 102 | } |
| 103 | |
| 104 | void HELPER(gvec_adds32)(void *d, void *a, uint64_t b, uint32_t desc) |
| 105 | { |
| 106 | intptr_t oprsz = simd_oprsz(desc); |
| 107 | intptr_t i; |
| 108 | |
| 109 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 110 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) + (uint32_t)b; |
| 111 | } |
| 112 | clear_high(d, oprsz, desc); |
| 113 | } |
| 114 | |
| 115 | void HELPER(gvec_adds64)(void *d, void *a, uint64_t b, uint32_t desc) |
| 116 | { |
| 117 | intptr_t oprsz = simd_oprsz(desc); |
| 118 | intptr_t i; |
| 119 | |
| 120 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 121 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) + b; |
| 122 | } |
| 123 | clear_high(d, oprsz, desc); |
| 124 | } |
| 125 | |
| 126 | void HELPER(gvec_sub8)(void *d, void *a, void *b, uint32_t desc) |
| 127 | { |
| 128 | intptr_t oprsz = simd_oprsz(desc); |
| 129 | intptr_t i; |
| 130 | |
| 131 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 132 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) - *(uint8_t *)(b + i); |
| 133 | } |
| 134 | clear_high(d, oprsz, desc); |
| 135 | } |
| 136 | |
| 137 | void HELPER(gvec_sub16)(void *d, void *a, void *b, uint32_t desc) |
| 138 | { |
| 139 | intptr_t oprsz = simd_oprsz(desc); |
| 140 | intptr_t i; |
| 141 | |
| 142 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 143 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) - *(uint16_t *)(b + i); |
| 144 | } |
| 145 | clear_high(d, oprsz, desc); |
| 146 | } |
| 147 | |
| 148 | void HELPER(gvec_sub32)(void *d, void *a, void *b, uint32_t desc) |
| 149 | { |
| 150 | intptr_t oprsz = simd_oprsz(desc); |
| 151 | intptr_t i; |
| 152 | |
| 153 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 154 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) - *(uint32_t *)(b + i); |
| 155 | } |
| 156 | clear_high(d, oprsz, desc); |
| 157 | } |
| 158 | |
| 159 | void HELPER(gvec_sub64)(void *d, void *a, void *b, uint32_t desc) |
| 160 | { |
| 161 | intptr_t oprsz = simd_oprsz(desc); |
| 162 | intptr_t i; |
| 163 | |
| 164 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 165 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) - *(uint64_t *)(b + i); |
| 166 | } |
| 167 | clear_high(d, oprsz, desc); |
| 168 | } |
| 169 | |
| 170 | void HELPER(gvec_subs8)(void *d, void *a, uint64_t b, uint32_t desc) |
| 171 | { |
| 172 | intptr_t oprsz = simd_oprsz(desc); |
| 173 | intptr_t i; |
| 174 | |
| 175 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 176 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) - (uint8_t)b; |
| 177 | } |
| 178 | clear_high(d, oprsz, desc); |
| 179 | } |
| 180 | |
| 181 | void HELPER(gvec_subs16)(void *d, void *a, uint64_t b, uint32_t desc) |
| 182 | { |
| 183 | intptr_t oprsz = simd_oprsz(desc); |
| 184 | intptr_t i; |
| 185 | |
| 186 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 187 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) - (uint16_t)b; |
| 188 | } |
| 189 | clear_high(d, oprsz, desc); |
| 190 | } |
| 191 | |
| 192 | void HELPER(gvec_subs32)(void *d, void *a, uint64_t b, uint32_t desc) |
| 193 | { |
| 194 | intptr_t oprsz = simd_oprsz(desc); |
| 195 | intptr_t i; |
| 196 | |
| 197 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 198 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) - (uint32_t)b; |
| 199 | } |
| 200 | clear_high(d, oprsz, desc); |
| 201 | } |
| 202 | |
| 203 | void HELPER(gvec_subs64)(void *d, void *a, uint64_t b, uint32_t desc) |
| 204 | { |
| 205 | intptr_t oprsz = simd_oprsz(desc); |
| 206 | intptr_t i; |
| 207 | |
| 208 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 209 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) - b; |
| 210 | } |
| 211 | clear_high(d, oprsz, desc); |
| 212 | } |
| 213 | |
| 214 | void HELPER(gvec_mul8)(void *d, void *a, void *b, uint32_t desc) |
| 215 | { |
| 216 | intptr_t oprsz = simd_oprsz(desc); |
| 217 | intptr_t i; |
| 218 | |
| 219 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 220 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) * *(uint8_t *)(b + i); |
| 221 | } |
| 222 | clear_high(d, oprsz, desc); |
| 223 | } |
| 224 | |
| 225 | void HELPER(gvec_mul16)(void *d, void *a, void *b, uint32_t desc) |
| 226 | { |
| 227 | intptr_t oprsz = simd_oprsz(desc); |
| 228 | intptr_t i; |
| 229 | |
| 230 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 231 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) * *(uint16_t *)(b + i); |
| 232 | } |
| 233 | clear_high(d, oprsz, desc); |
| 234 | } |
| 235 | |
| 236 | void HELPER(gvec_mul32)(void *d, void *a, void *b, uint32_t desc) |
| 237 | { |
| 238 | intptr_t oprsz = simd_oprsz(desc); |
| 239 | intptr_t i; |
| 240 | |
| 241 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 242 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) * *(uint32_t *)(b + i); |
| 243 | } |
| 244 | clear_high(d, oprsz, desc); |
| 245 | } |
| 246 | |
| 247 | void HELPER(gvec_mul64)(void *d, void *a, void *b, uint32_t desc) |
| 248 | { |
| 249 | intptr_t oprsz = simd_oprsz(desc); |
| 250 | intptr_t i; |
| 251 | |
| 252 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 253 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) * *(uint64_t *)(b + i); |
| 254 | } |
| 255 | clear_high(d, oprsz, desc); |
| 256 | } |
| 257 | |
| 258 | void HELPER(gvec_muls8)(void *d, void *a, uint64_t b, uint32_t desc) |
| 259 | { |
| 260 | intptr_t oprsz = simd_oprsz(desc); |
| 261 | intptr_t i; |
| 262 | |
| 263 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 264 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) * (uint8_t)b; |
| 265 | } |
| 266 | clear_high(d, oprsz, desc); |
| 267 | } |
| 268 | |
| 269 | void HELPER(gvec_muls16)(void *d, void *a, uint64_t b, uint32_t desc) |
| 270 | { |
| 271 | intptr_t oprsz = simd_oprsz(desc); |
| 272 | intptr_t i; |
| 273 | |
| 274 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 275 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) * (uint16_t)b; |
| 276 | } |
| 277 | clear_high(d, oprsz, desc); |
| 278 | } |
| 279 | |
| 280 | void HELPER(gvec_muls32)(void *d, void *a, uint64_t b, uint32_t desc) |
| 281 | { |
| 282 | intptr_t oprsz = simd_oprsz(desc); |
| 283 | intptr_t i; |
| 284 | |
| 285 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 286 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) * (uint32_t)b; |
| 287 | } |
| 288 | clear_high(d, oprsz, desc); |
| 289 | } |
| 290 | |
| 291 | void HELPER(gvec_muls64)(void *d, void *a, uint64_t b, uint32_t desc) |
| 292 | { |
| 293 | intptr_t oprsz = simd_oprsz(desc); |
| 294 | intptr_t i; |
| 295 | |
| 296 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 297 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) * b; |
| 298 | } |
| 299 | clear_high(d, oprsz, desc); |
| 300 | } |
| 301 | |
| 302 | void HELPER(gvec_neg8)(void *d, void *a, uint32_t desc) |
| 303 | { |
| 304 | intptr_t oprsz = simd_oprsz(desc); |
| 305 | intptr_t i; |
| 306 | |
| 307 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 308 | *(uint8_t *)(d + i) = -*(uint8_t *)(a + i); |
| 309 | } |
| 310 | clear_high(d, oprsz, desc); |
| 311 | } |
| 312 | |
| 313 | void HELPER(gvec_neg16)(void *d, void *a, uint32_t desc) |
| 314 | { |
| 315 | intptr_t oprsz = simd_oprsz(desc); |
| 316 | intptr_t i; |
| 317 | |
| 318 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 319 | *(uint16_t *)(d + i) = -*(uint16_t *)(a + i); |
| 320 | } |
| 321 | clear_high(d, oprsz, desc); |
| 322 | } |
| 323 | |
| 324 | void HELPER(gvec_neg32)(void *d, void *a, uint32_t desc) |
| 325 | { |
| 326 | intptr_t oprsz = simd_oprsz(desc); |
| 327 | intptr_t i; |
| 328 | |
| 329 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 330 | *(uint32_t *)(d + i) = -*(uint32_t *)(a + i); |
| 331 | } |
| 332 | clear_high(d, oprsz, desc); |
| 333 | } |
| 334 | |
| 335 | void HELPER(gvec_neg64)(void *d, void *a, uint32_t desc) |
| 336 | { |
| 337 | intptr_t oprsz = simd_oprsz(desc); |
| 338 | intptr_t i; |
| 339 | |
| 340 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 341 | *(uint64_t *)(d + i) = -*(uint64_t *)(a + i); |
| 342 | } |
| 343 | clear_high(d, oprsz, desc); |
| 344 | } |
| 345 | |
| 346 | void HELPER(gvec_abs8)(void *d, void *a, uint32_t desc) |
| 347 | { |
| 348 | intptr_t oprsz = simd_oprsz(desc); |
| 349 | intptr_t i; |
| 350 | |
| 351 | for (i = 0; i < oprsz; i += sizeof(int8_t)) { |
| 352 | int8_t aa = *(int8_t *)(a + i); |
| 353 | *(int8_t *)(d + i) = aa < 0 ? -aa : aa; |
| 354 | } |
| 355 | clear_high(d, oprsz, desc); |
| 356 | } |
| 357 | |
| 358 | void HELPER(gvec_abs16)(void *d, void *a, uint32_t desc) |
| 359 | { |
| 360 | intptr_t oprsz = simd_oprsz(desc); |
| 361 | intptr_t i; |
| 362 | |
| 363 | for (i = 0; i < oprsz; i += sizeof(int16_t)) { |
| 364 | int16_t aa = *(int16_t *)(a + i); |
| 365 | *(int16_t *)(d + i) = aa < 0 ? -aa : aa; |
| 366 | } |
| 367 | clear_high(d, oprsz, desc); |
| 368 | } |
| 369 | |
| 370 | void HELPER(gvec_abs32)(void *d, void *a, uint32_t desc) |
| 371 | { |
| 372 | intptr_t oprsz = simd_oprsz(desc); |
| 373 | intptr_t i; |
| 374 | |
| 375 | for (i = 0; i < oprsz; i += sizeof(int32_t)) { |
| 376 | int32_t aa = *(int32_t *)(a + i); |
| 377 | *(int32_t *)(d + i) = aa < 0 ? -aa : aa; |
| 378 | } |
| 379 | clear_high(d, oprsz, desc); |
| 380 | } |
| 381 | |
| 382 | void HELPER(gvec_abs64)(void *d, void *a, uint32_t desc) |
| 383 | { |
| 384 | intptr_t oprsz = simd_oprsz(desc); |
| 385 | intptr_t i; |
| 386 | |
| 387 | for (i = 0; i < oprsz; i += sizeof(int64_t)) { |
| 388 | int64_t aa = *(int64_t *)(a + i); |
| 389 | *(int64_t *)(d + i) = aa < 0 ? -aa : aa; |
| 390 | } |
| 391 | clear_high(d, oprsz, desc); |
| 392 | } |
| 393 | |
| 394 | void HELPER(gvec_mov)(void *d, void *a, uint32_t desc) |
| 395 | { |
| 396 | intptr_t oprsz = simd_oprsz(desc); |
| 397 | |
| 398 | memcpy(d, a, oprsz); |
| 399 | clear_high(d, oprsz, desc); |
| 400 | } |
| 401 | |
| 402 | void HELPER(gvec_dup64)(void *d, uint32_t desc, uint64_t c) |
| 403 | { |
| 404 | intptr_t oprsz = simd_oprsz(desc); |
| 405 | intptr_t i; |
| 406 | |
| 407 | if (c == 0) { |
| 408 | oprsz = 0; |
| 409 | } else { |
| 410 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 411 | *(uint64_t *)(d + i) = c; |
| 412 | } |
| 413 | } |
| 414 | clear_high(d, oprsz, desc); |
| 415 | } |
| 416 | |
| 417 | void HELPER(gvec_dup32)(void *d, uint32_t desc, uint32_t c) |
| 418 | { |
| 419 | intptr_t oprsz = simd_oprsz(desc); |
| 420 | intptr_t i; |
| 421 | |
| 422 | if (c == 0) { |
| 423 | oprsz = 0; |
| 424 | } else { |
| 425 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 426 | *(uint32_t *)(d + i) = c; |
| 427 | } |
| 428 | } |
| 429 | clear_high(d, oprsz, desc); |
| 430 | } |
| 431 | |
| 432 | void HELPER(gvec_dup16)(void *d, uint32_t desc, uint32_t c) |
| 433 | { |
| 434 | HELPER(gvec_dup32)(d, desc, 0x00010001 * (c & 0xffff)); |
| 435 | } |
| 436 | |
| 437 | void HELPER(gvec_dup8)(void *d, uint32_t desc, uint32_t c) |
| 438 | { |
| 439 | HELPER(gvec_dup32)(d, desc, 0x01010101 * (c & 0xff)); |
| 440 | } |
| 441 | |
| 442 | void HELPER(gvec_not)(void *d, void *a, uint32_t desc) |
| 443 | { |
| 444 | intptr_t oprsz = simd_oprsz(desc); |
| 445 | intptr_t i; |
| 446 | |
| 447 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 448 | *(uint64_t *)(d + i) = ~*(uint64_t *)(a + i); |
| 449 | } |
| 450 | clear_high(d, oprsz, desc); |
| 451 | } |
| 452 | |
| 453 | void HELPER(gvec_and)(void *d, void *a, void *b, uint32_t desc) |
| 454 | { |
| 455 | intptr_t oprsz = simd_oprsz(desc); |
| 456 | intptr_t i; |
| 457 | |
| 458 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 459 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) & *(uint64_t *)(b + i); |
| 460 | } |
| 461 | clear_high(d, oprsz, desc); |
| 462 | } |
| 463 | |
| 464 | void HELPER(gvec_or)(void *d, void *a, void *b, uint32_t desc) |
| 465 | { |
| 466 | intptr_t oprsz = simd_oprsz(desc); |
| 467 | intptr_t i; |
| 468 | |
| 469 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 470 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) | *(uint64_t *)(b + i); |
| 471 | } |
| 472 | clear_high(d, oprsz, desc); |
| 473 | } |
| 474 | |
| 475 | void HELPER(gvec_xor)(void *d, void *a, void *b, uint32_t desc) |
| 476 | { |
| 477 | intptr_t oprsz = simd_oprsz(desc); |
| 478 | intptr_t i; |
| 479 | |
| 480 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 481 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) ^ *(uint64_t *)(b + i); |
| 482 | } |
| 483 | clear_high(d, oprsz, desc); |
| 484 | } |
| 485 | |
| 486 | void HELPER(gvec_andc)(void *d, void *a, void *b, uint32_t desc) |
| 487 | { |
| 488 | intptr_t oprsz = simd_oprsz(desc); |
| 489 | intptr_t i; |
| 490 | |
| 491 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 492 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) &~ *(uint64_t *)(b + i); |
| 493 | } |
| 494 | clear_high(d, oprsz, desc); |
| 495 | } |
| 496 | |
| 497 | void HELPER(gvec_orc)(void *d, void *a, void *b, uint32_t desc) |
| 498 | { |
| 499 | intptr_t oprsz = simd_oprsz(desc); |
| 500 | intptr_t i; |
| 501 | |
| 502 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 503 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) |~ *(uint64_t *)(b + i); |
| 504 | } |
| 505 | clear_high(d, oprsz, desc); |
| 506 | } |
| 507 | |
| 508 | void HELPER(gvec_nand)(void *d, void *a, void *b, uint32_t desc) |
| 509 | { |
| 510 | intptr_t oprsz = simd_oprsz(desc); |
| 511 | intptr_t i; |
| 512 | |
| 513 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 514 | *(uint64_t *)(d + i) = ~(*(uint64_t *)(a + i) & *(uint64_t *)(b + i)); |
| 515 | } |
| 516 | clear_high(d, oprsz, desc); |
| 517 | } |
| 518 | |
| 519 | void HELPER(gvec_nor)(void *d, void *a, void *b, uint32_t desc) |
| 520 | { |
| 521 | intptr_t oprsz = simd_oprsz(desc); |
| 522 | intptr_t i; |
| 523 | |
| 524 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 525 | *(uint64_t *)(d + i) = ~(*(uint64_t *)(a + i) | *(uint64_t *)(b + i)); |
| 526 | } |
| 527 | clear_high(d, oprsz, desc); |
| 528 | } |
| 529 | |
| 530 | void HELPER(gvec_eqv)(void *d, void *a, void *b, uint32_t desc) |
| 531 | { |
| 532 | intptr_t oprsz = simd_oprsz(desc); |
| 533 | intptr_t i; |
| 534 | |
| 535 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 536 | *(uint64_t *)(d + i) = ~(*(uint64_t *)(a + i) ^ *(uint64_t *)(b + i)); |
| 537 | } |
| 538 | clear_high(d, oprsz, desc); |
| 539 | } |
| 540 | |
| 541 | void HELPER(gvec_ands)(void *d, void *a, uint64_t b, uint32_t desc) |
| 542 | { |
| 543 | intptr_t oprsz = simd_oprsz(desc); |
| 544 | intptr_t i; |
| 545 | |
| 546 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 547 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) & b; |
| 548 | } |
| 549 | clear_high(d, oprsz, desc); |
| 550 | } |
| 551 | |
| 552 | void HELPER(gvec_andcs)(void *d, void *a, uint64_t b, uint32_t desc) |
| 553 | { |
| 554 | intptr_t oprsz = simd_oprsz(desc); |
| 555 | intptr_t i; |
| 556 | |
| 557 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 558 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) & ~b; |
| 559 | } |
| 560 | clear_high(d, oprsz, desc); |
| 561 | } |
| 562 | |
| 563 | void HELPER(gvec_xors)(void *d, void *a, uint64_t b, uint32_t desc) |
| 564 | { |
| 565 | intptr_t oprsz = simd_oprsz(desc); |
| 566 | intptr_t i; |
| 567 | |
| 568 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 569 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) ^ b; |
| 570 | } |
| 571 | clear_high(d, oprsz, desc); |
| 572 | } |
| 573 | |
| 574 | void HELPER(gvec_ors)(void *d, void *a, uint64_t b, uint32_t desc) |
| 575 | { |
| 576 | intptr_t oprsz = simd_oprsz(desc); |
| 577 | intptr_t i; |
| 578 | |
| 579 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 580 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) | b; |
| 581 | } |
| 582 | clear_high(d, oprsz, desc); |
| 583 | } |
| 584 | |
| 585 | void HELPER(gvec_shl8i)(void *d, void *a, uint32_t desc) |
| 586 | { |
| 587 | intptr_t oprsz = simd_oprsz(desc); |
| 588 | int shift = simd_data(desc); |
| 589 | intptr_t i; |
| 590 | |
| 591 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 592 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) << shift; |
| 593 | } |
| 594 | clear_high(d, oprsz, desc); |
| 595 | } |
| 596 | |
| 597 | void HELPER(gvec_shl16i)(void *d, void *a, uint32_t desc) |
| 598 | { |
| 599 | intptr_t oprsz = simd_oprsz(desc); |
| 600 | int shift = simd_data(desc); |
| 601 | intptr_t i; |
| 602 | |
| 603 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 604 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) << shift; |
| 605 | } |
| 606 | clear_high(d, oprsz, desc); |
| 607 | } |
| 608 | |
| 609 | void HELPER(gvec_shl32i)(void *d, void *a, uint32_t desc) |
| 610 | { |
| 611 | intptr_t oprsz = simd_oprsz(desc); |
| 612 | int shift = simd_data(desc); |
| 613 | intptr_t i; |
| 614 | |
| 615 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 616 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) << shift; |
| 617 | } |
| 618 | clear_high(d, oprsz, desc); |
| 619 | } |
| 620 | |
| 621 | void HELPER(gvec_shl64i)(void *d, void *a, uint32_t desc) |
| 622 | { |
| 623 | intptr_t oprsz = simd_oprsz(desc); |
| 624 | int shift = simd_data(desc); |
| 625 | intptr_t i; |
| 626 | |
| 627 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 628 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) << shift; |
| 629 | } |
| 630 | clear_high(d, oprsz, desc); |
| 631 | } |
| 632 | |
| 633 | void HELPER(gvec_shr8i)(void *d, void *a, uint32_t desc) |
| 634 | { |
| 635 | intptr_t oprsz = simd_oprsz(desc); |
| 636 | int shift = simd_data(desc); |
| 637 | intptr_t i; |
| 638 | |
| 639 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 640 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) >> shift; |
| 641 | } |
| 642 | clear_high(d, oprsz, desc); |
| 643 | } |
| 644 | |
| 645 | void HELPER(gvec_shr16i)(void *d, void *a, uint32_t desc) |
| 646 | { |
| 647 | intptr_t oprsz = simd_oprsz(desc); |
| 648 | int shift = simd_data(desc); |
| 649 | intptr_t i; |
| 650 | |
| 651 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 652 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) >> shift; |
| 653 | } |
| 654 | clear_high(d, oprsz, desc); |
| 655 | } |
| 656 | |
| 657 | void HELPER(gvec_shr32i)(void *d, void *a, uint32_t desc) |
| 658 | { |
| 659 | intptr_t oprsz = simd_oprsz(desc); |
| 660 | int shift = simd_data(desc); |
| 661 | intptr_t i; |
| 662 | |
| 663 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 664 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) >> shift; |
| 665 | } |
| 666 | clear_high(d, oprsz, desc); |
| 667 | } |
| 668 | |
| 669 | void HELPER(gvec_shr64i)(void *d, void *a, uint32_t desc) |
| 670 | { |
| 671 | intptr_t oprsz = simd_oprsz(desc); |
| 672 | int shift = simd_data(desc); |
| 673 | intptr_t i; |
| 674 | |
| 675 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 676 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) >> shift; |
| 677 | } |
| 678 | clear_high(d, oprsz, desc); |
| 679 | } |
| 680 | |
| 681 | void HELPER(gvec_sar8i)(void *d, void *a, uint32_t desc) |
| 682 | { |
| 683 | intptr_t oprsz = simd_oprsz(desc); |
| 684 | int shift = simd_data(desc); |
| 685 | intptr_t i; |
| 686 | |
| 687 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 688 | *(int8_t *)(d + i) = *(int8_t *)(a + i) >> shift; |
| 689 | } |
| 690 | clear_high(d, oprsz, desc); |
| 691 | } |
| 692 | |
| 693 | void HELPER(gvec_sar16i)(void *d, void *a, uint32_t desc) |
| 694 | { |
| 695 | intptr_t oprsz = simd_oprsz(desc); |
| 696 | int shift = simd_data(desc); |
| 697 | intptr_t i; |
| 698 | |
| 699 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 700 | *(int16_t *)(d + i) = *(int16_t *)(a + i) >> shift; |
| 701 | } |
| 702 | clear_high(d, oprsz, desc); |
| 703 | } |
| 704 | |
| 705 | void HELPER(gvec_sar32i)(void *d, void *a, uint32_t desc) |
| 706 | { |
| 707 | intptr_t oprsz = simd_oprsz(desc); |
| 708 | int shift = simd_data(desc); |
| 709 | intptr_t i; |
| 710 | |
| 711 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 712 | *(int32_t *)(d + i) = *(int32_t *)(a + i) >> shift; |
| 713 | } |
| 714 | clear_high(d, oprsz, desc); |
| 715 | } |
| 716 | |
| 717 | void HELPER(gvec_sar64i)(void *d, void *a, uint32_t desc) |
| 718 | { |
| 719 | intptr_t oprsz = simd_oprsz(desc); |
| 720 | int shift = simd_data(desc); |
| 721 | intptr_t i; |
| 722 | |
| 723 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 724 | *(int64_t *)(d + i) = *(int64_t *)(a + i) >> shift; |
| 725 | } |
| 726 | clear_high(d, oprsz, desc); |
| 727 | } |
| 728 | |
| 729 | void HELPER(gvec_rotl8i)(void *d, void *a, uint32_t desc) |
| 730 | { |
| 731 | intptr_t oprsz = simd_oprsz(desc); |
| 732 | int shift = simd_data(desc); |
| 733 | intptr_t i; |
| 734 | |
| 735 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 736 | *(uint8_t *)(d + i) = rol8(*(uint8_t *)(a + i), shift); |
| 737 | } |
| 738 | clear_high(d, oprsz, desc); |
| 739 | } |
| 740 | |
| 741 | void HELPER(gvec_rotl16i)(void *d, void *a, uint32_t desc) |
| 742 | { |
| 743 | intptr_t oprsz = simd_oprsz(desc); |
| 744 | int shift = simd_data(desc); |
| 745 | intptr_t i; |
| 746 | |
| 747 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 748 | *(uint16_t *)(d + i) = rol16(*(uint16_t *)(a + i), shift); |
| 749 | } |
| 750 | clear_high(d, oprsz, desc); |
| 751 | } |
| 752 | |
| 753 | void HELPER(gvec_rotl32i)(void *d, void *a, uint32_t desc) |
| 754 | { |
| 755 | intptr_t oprsz = simd_oprsz(desc); |
| 756 | int shift = simd_data(desc); |
| 757 | intptr_t i; |
| 758 | |
| 759 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 760 | *(uint32_t *)(d + i) = rol32(*(uint32_t *)(a + i), shift); |
| 761 | } |
| 762 | clear_high(d, oprsz, desc); |
| 763 | } |
| 764 | |
| 765 | void HELPER(gvec_rotl64i)(void *d, void *a, uint32_t desc) |
| 766 | { |
| 767 | intptr_t oprsz = simd_oprsz(desc); |
| 768 | int shift = simd_data(desc); |
| 769 | intptr_t i; |
| 770 | |
| 771 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 772 | *(uint64_t *)(d + i) = rol64(*(uint64_t *)(a + i), shift); |
| 773 | } |
| 774 | clear_high(d, oprsz, desc); |
| 775 | } |
| 776 | |
| 777 | void HELPER(gvec_shl8v)(void *d, void *a, void *b, uint32_t desc) |
| 778 | { |
| 779 | intptr_t oprsz = simd_oprsz(desc); |
| 780 | intptr_t i; |
| 781 | |
| 782 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 783 | uint8_t sh = *(uint8_t *)(b + i) & 7; |
| 784 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) << sh; |
| 785 | } |
| 786 | clear_high(d, oprsz, desc); |
| 787 | } |
| 788 | |
| 789 | void HELPER(gvec_shl16v)(void *d, void *a, void *b, uint32_t desc) |
| 790 | { |
| 791 | intptr_t oprsz = simd_oprsz(desc); |
| 792 | intptr_t i; |
| 793 | |
| 794 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 795 | uint8_t sh = *(uint16_t *)(b + i) & 15; |
| 796 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) << sh; |
| 797 | } |
| 798 | clear_high(d, oprsz, desc); |
| 799 | } |
| 800 | |
| 801 | void HELPER(gvec_shl32v)(void *d, void *a, void *b, uint32_t desc) |
| 802 | { |
| 803 | intptr_t oprsz = simd_oprsz(desc); |
| 804 | intptr_t i; |
| 805 | |
| 806 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 807 | uint8_t sh = *(uint32_t *)(b + i) & 31; |
| 808 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) << sh; |
| 809 | } |
| 810 | clear_high(d, oprsz, desc); |
| 811 | } |
| 812 | |
| 813 | void HELPER(gvec_shl64v)(void *d, void *a, void *b, uint32_t desc) |
| 814 | { |
| 815 | intptr_t oprsz = simd_oprsz(desc); |
| 816 | intptr_t i; |
| 817 | |
| 818 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 819 | uint8_t sh = *(uint64_t *)(b + i) & 63; |
| 820 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) << sh; |
| 821 | } |
| 822 | clear_high(d, oprsz, desc); |
| 823 | } |
| 824 | |
| 825 | void HELPER(gvec_shr8v)(void *d, void *a, void *b, uint32_t desc) |
| 826 | { |
| 827 | intptr_t oprsz = simd_oprsz(desc); |
| 828 | intptr_t i; |
| 829 | |
| 830 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 831 | uint8_t sh = *(uint8_t *)(b + i) & 7; |
| 832 | *(uint8_t *)(d + i) = *(uint8_t *)(a + i) >> sh; |
| 833 | } |
| 834 | clear_high(d, oprsz, desc); |
| 835 | } |
| 836 | |
| 837 | void HELPER(gvec_shr16v)(void *d, void *a, void *b, uint32_t desc) |
| 838 | { |
| 839 | intptr_t oprsz = simd_oprsz(desc); |
| 840 | intptr_t i; |
| 841 | |
| 842 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 843 | uint8_t sh = *(uint16_t *)(b + i) & 15; |
| 844 | *(uint16_t *)(d + i) = *(uint16_t *)(a + i) >> sh; |
| 845 | } |
| 846 | clear_high(d, oprsz, desc); |
| 847 | } |
| 848 | |
| 849 | void HELPER(gvec_shr32v)(void *d, void *a, void *b, uint32_t desc) |
| 850 | { |
| 851 | intptr_t oprsz = simd_oprsz(desc); |
| 852 | intptr_t i; |
| 853 | |
| 854 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 855 | uint8_t sh = *(uint32_t *)(b + i) & 31; |
| 856 | *(uint32_t *)(d + i) = *(uint32_t *)(a + i) >> sh; |
| 857 | } |
| 858 | clear_high(d, oprsz, desc); |
| 859 | } |
| 860 | |
| 861 | void HELPER(gvec_shr64v)(void *d, void *a, void *b, uint32_t desc) |
| 862 | { |
| 863 | intptr_t oprsz = simd_oprsz(desc); |
| 864 | intptr_t i; |
| 865 | |
| 866 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 867 | uint8_t sh = *(uint64_t *)(b + i) & 63; |
| 868 | *(uint64_t *)(d + i) = *(uint64_t *)(a + i) >> sh; |
| 869 | } |
| 870 | clear_high(d, oprsz, desc); |
| 871 | } |
| 872 | |
| 873 | void HELPER(gvec_sar8v)(void *d, void *a, void *b, uint32_t desc) |
| 874 | { |
| 875 | intptr_t oprsz = simd_oprsz(desc); |
| 876 | intptr_t i; |
| 877 | |
| 878 | for (i = 0; i < oprsz; i += sizeof(int8_t)) { |
| 879 | uint8_t sh = *(uint8_t *)(b + i) & 7; |
| 880 | *(int8_t *)(d + i) = *(int8_t *)(a + i) >> sh; |
| 881 | } |
| 882 | clear_high(d, oprsz, desc); |
| 883 | } |
| 884 | |
| 885 | void HELPER(gvec_sar16v)(void *d, void *a, void *b, uint32_t desc) |
| 886 | { |
| 887 | intptr_t oprsz = simd_oprsz(desc); |
| 888 | intptr_t i; |
| 889 | |
| 890 | for (i = 0; i < oprsz; i += sizeof(int16_t)) { |
| 891 | uint8_t sh = *(uint16_t *)(b + i) & 15; |
| 892 | *(int16_t *)(d + i) = *(int16_t *)(a + i) >> sh; |
| 893 | } |
| 894 | clear_high(d, oprsz, desc); |
| 895 | } |
| 896 | |
| 897 | void HELPER(gvec_sar32v)(void *d, void *a, void *b, uint32_t desc) |
| 898 | { |
| 899 | intptr_t oprsz = simd_oprsz(desc); |
| 900 | intptr_t i; |
| 901 | |
| 902 | for (i = 0; i < oprsz; i += sizeof(int32_t)) { |
| 903 | uint8_t sh = *(uint32_t *)(b + i) & 31; |
| 904 | *(int32_t *)(d + i) = *(int32_t *)(a + i) >> sh; |
| 905 | } |
| 906 | clear_high(d, oprsz, desc); |
| 907 | } |
| 908 | |
| 909 | void HELPER(gvec_sar64v)(void *d, void *a, void *b, uint32_t desc) |
| 910 | { |
| 911 | intptr_t oprsz = simd_oprsz(desc); |
| 912 | intptr_t i; |
| 913 | |
| 914 | for (i = 0; i < oprsz; i += sizeof(int64_t)) { |
| 915 | uint8_t sh = *(uint64_t *)(b + i) & 63; |
| 916 | *(int64_t *)(d + i) = *(int64_t *)(a + i) >> sh; |
| 917 | } |
| 918 | clear_high(d, oprsz, desc); |
| 919 | } |
| 920 | |
| 921 | void HELPER(gvec_rotl8v)(void *d, void *a, void *b, uint32_t desc) |
| 922 | { |
| 923 | intptr_t oprsz = simd_oprsz(desc); |
| 924 | intptr_t i; |
| 925 | |
| 926 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 927 | uint8_t sh = *(uint8_t *)(b + i) & 7; |
| 928 | *(uint8_t *)(d + i) = rol8(*(uint8_t *)(a + i), sh); |
| 929 | } |
| 930 | clear_high(d, oprsz, desc); |
| 931 | } |
| 932 | |
| 933 | void HELPER(gvec_rotl16v)(void *d, void *a, void *b, uint32_t desc) |
| 934 | { |
| 935 | intptr_t oprsz = simd_oprsz(desc); |
| 936 | intptr_t i; |
| 937 | |
| 938 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 939 | uint8_t sh = *(uint16_t *)(b + i) & 15; |
| 940 | *(uint16_t *)(d + i) = rol16(*(uint16_t *)(a + i), sh); |
| 941 | } |
| 942 | clear_high(d, oprsz, desc); |
| 943 | } |
| 944 | |
| 945 | void HELPER(gvec_rotl32v)(void *d, void *a, void *b, uint32_t desc) |
| 946 | { |
| 947 | intptr_t oprsz = simd_oprsz(desc); |
| 948 | intptr_t i; |
| 949 | |
| 950 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 951 | uint8_t sh = *(uint32_t *)(b + i) & 31; |
| 952 | *(uint32_t *)(d + i) = rol32(*(uint32_t *)(a + i), sh); |
| 953 | } |
| 954 | clear_high(d, oprsz, desc); |
| 955 | } |
| 956 | |
| 957 | void HELPER(gvec_rotl64v)(void *d, void *a, void *b, uint32_t desc) |
| 958 | { |
| 959 | intptr_t oprsz = simd_oprsz(desc); |
| 960 | intptr_t i; |
| 961 | |
| 962 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 963 | uint8_t sh = *(uint64_t *)(b + i) & 63; |
| 964 | *(uint64_t *)(d + i) = rol64(*(uint64_t *)(a + i), sh); |
| 965 | } |
| 966 | clear_high(d, oprsz, desc); |
| 967 | } |
| 968 | |
| 969 | void HELPER(gvec_rotr8v)(void *d, void *a, void *b, uint32_t desc) |
| 970 | { |
| 971 | intptr_t oprsz = simd_oprsz(desc); |
| 972 | intptr_t i; |
| 973 | |
| 974 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 975 | uint8_t sh = *(uint8_t *)(b + i) & 7; |
| 976 | *(uint8_t *)(d + i) = ror8(*(uint8_t *)(a + i), sh); |
| 977 | } |
| 978 | clear_high(d, oprsz, desc); |
| 979 | } |
| 980 | |
| 981 | void HELPER(gvec_rotr16v)(void *d, void *a, void *b, uint32_t desc) |
| 982 | { |
| 983 | intptr_t oprsz = simd_oprsz(desc); |
| 984 | intptr_t i; |
| 985 | |
| 986 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 987 | uint8_t sh = *(uint16_t *)(b + i) & 15; |
| 988 | *(uint16_t *)(d + i) = ror16(*(uint16_t *)(a + i), sh); |
| 989 | } |
| 990 | clear_high(d, oprsz, desc); |
| 991 | } |
| 992 | |
| 993 | void HELPER(gvec_rotr32v)(void *d, void *a, void *b, uint32_t desc) |
| 994 | { |
| 995 | intptr_t oprsz = simd_oprsz(desc); |
| 996 | intptr_t i; |
| 997 | |
| 998 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 999 | uint8_t sh = *(uint32_t *)(b + i) & 31; |
| 1000 | *(uint32_t *)(d + i) = ror32(*(uint32_t *)(a + i), sh); |
| 1001 | } |
| 1002 | clear_high(d, oprsz, desc); |
| 1003 | } |
| 1004 | |
| 1005 | void HELPER(gvec_rotr64v)(void *d, void *a, void *b, uint32_t desc) |
| 1006 | { |
| 1007 | intptr_t oprsz = simd_oprsz(desc); |
| 1008 | intptr_t i; |
| 1009 | |
| 1010 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 1011 | uint8_t sh = *(uint64_t *)(b + i) & 63; |
| 1012 | *(uint64_t *)(d + i) = ror64(*(uint64_t *)(a + i), sh); |
| 1013 | } |
| 1014 | clear_high(d, oprsz, desc); |
| 1015 | } |
| 1016 | |
| 1017 | #define DO_CMP1(NAME, TYPE, OP) \ |
| 1018 | void HELPER(NAME)(void *d, void *a, void *b, uint32_t desc) \ |
| 1019 | { \ |
| 1020 | intptr_t oprsz = simd_oprsz(desc); \ |
| 1021 | intptr_t i; \ |
| 1022 | for (i = 0; i < oprsz; i += sizeof(TYPE)) { \ |
| 1023 | *(TYPE *)(d + i) = -(*(TYPE *)(a + i) OP *(TYPE *)(b + i)); \ |
| 1024 | } \ |
| 1025 | clear_high(d, oprsz, desc); \ |
| 1026 | } |
| 1027 | |
| 1028 | #define DO_CMP2(SZ) \ |
| 1029 | DO_CMP1(gvec_eq##SZ, uint##SZ##_t, ==) \ |
| 1030 | DO_CMP1(gvec_ne##SZ, uint##SZ##_t, !=) \ |
| 1031 | DO_CMP1(gvec_lt##SZ, int##SZ##_t, <) \ |
| 1032 | DO_CMP1(gvec_le##SZ, int##SZ##_t, <=) \ |
| 1033 | DO_CMP1(gvec_ltu##SZ, uint##SZ##_t, <) \ |
| 1034 | DO_CMP1(gvec_leu##SZ, uint##SZ##_t, <=) |
| 1035 | |
| 1036 | DO_CMP2(8) |
| 1037 | DO_CMP2(16) |
| 1038 | DO_CMP2(32) |
| 1039 | DO_CMP2(64) |
| 1040 | |
| 1041 | #undef DO_CMP1 |
| 1042 | #undef DO_CMP2 |
| 1043 | |
| 1044 | #define DO_CMP1(NAME, TYPE, OP) \ |
| 1045 | void HELPER(NAME)(void *d, void *a, uint64_t b64, uint32_t desc) \ |
| 1046 | { \ |
| 1047 | intptr_t oprsz = simd_oprsz(desc); \ |
| 1048 | TYPE inv = simd_data(desc), b = b64; \ |
| 1049 | for (intptr_t i = 0; i < oprsz; i += sizeof(TYPE)) { \ |
| 1050 | *(TYPE *)(d + i) = -((*(TYPE *)(a + i) OP b) ^ inv); \ |
| 1051 | } \ |
| 1052 | clear_high(d, oprsz, desc); \ |
| 1053 | } |
| 1054 | |
| 1055 | #define DO_CMP2(SZ) \ |
| 1056 | DO_CMP1(gvec_eqs##SZ, uint##SZ##_t, ==) \ |
| 1057 | DO_CMP1(gvec_lts##SZ, int##SZ##_t, <) \ |
| 1058 | DO_CMP1(gvec_les##SZ, int##SZ##_t, <=) \ |
| 1059 | DO_CMP1(gvec_ltus##SZ, uint##SZ##_t, <) \ |
| 1060 | DO_CMP1(gvec_leus##SZ, uint##SZ##_t, <=) |
| 1061 | |
| 1062 | DO_CMP2(8) |
| 1063 | DO_CMP2(16) |
| 1064 | DO_CMP2(32) |
| 1065 | DO_CMP2(64) |
| 1066 | |
| 1067 | #undef DO_CMP1 |
| 1068 | #undef DO_CMP2 |
| 1069 | |
| 1070 | void HELPER(gvec_ssadd8)(void *d, void *a, void *b, uint32_t desc) |
| 1071 | { |
| 1072 | intptr_t oprsz = simd_oprsz(desc); |
| 1073 | intptr_t i; |
| 1074 | |
| 1075 | for (i = 0; i < oprsz; i += sizeof(int8_t)) { |
| 1076 | int r = *(int8_t *)(a + i) + *(int8_t *)(b + i); |
| 1077 | if (r > INT8_MAX) { |
| 1078 | r = INT8_MAX; |
| 1079 | } else if (r < INT8_MIN) { |
| 1080 | r = INT8_MIN; |
| 1081 | } |
| 1082 | *(int8_t *)(d + i) = r; |
| 1083 | } |
| 1084 | clear_high(d, oprsz, desc); |
| 1085 | } |
| 1086 | |
| 1087 | void HELPER(gvec_ssadd16)(void *d, void *a, void *b, uint32_t desc) |
| 1088 | { |
| 1089 | intptr_t oprsz = simd_oprsz(desc); |
| 1090 | intptr_t i; |
| 1091 | |
| 1092 | for (i = 0; i < oprsz; i += sizeof(int16_t)) { |
| 1093 | int r = *(int16_t *)(a + i) + *(int16_t *)(b + i); |
| 1094 | if (r > INT16_MAX) { |
| 1095 | r = INT16_MAX; |
| 1096 | } else if (r < INT16_MIN) { |
| 1097 | r = INT16_MIN; |
| 1098 | } |
| 1099 | *(int16_t *)(d + i) = r; |
| 1100 | } |
| 1101 | clear_high(d, oprsz, desc); |
| 1102 | } |
| 1103 | |
| 1104 | void HELPER(gvec_ssadd32)(void *d, void *a, void *b, uint32_t desc) |
| 1105 | { |
| 1106 | intptr_t oprsz = simd_oprsz(desc); |
| 1107 | intptr_t i; |
| 1108 | |
| 1109 | for (i = 0; i < oprsz; i += sizeof(int32_t)) { |
| 1110 | int32_t ai = *(int32_t *)(a + i); |
| 1111 | int32_t bi = *(int32_t *)(b + i); |
| 1112 | int32_t di = sadd32_saturate(ai, bi); |
| 1113 | *(int32_t *)(d + i) = di; |
| 1114 | } |
| 1115 | clear_high(d, oprsz, desc); |
| 1116 | } |
| 1117 | |
| 1118 | void HELPER(gvec_ssadd64)(void *d, void *a, void *b, uint32_t desc) |
| 1119 | { |
| 1120 | intptr_t oprsz = simd_oprsz(desc); |
| 1121 | intptr_t i; |
| 1122 | |
| 1123 | for (i = 0; i < oprsz; i += sizeof(int64_t)) { |
| 1124 | int64_t ai = *(int64_t *)(a + i); |
| 1125 | int64_t bi = *(int64_t *)(b + i); |
| 1126 | int64_t di = sadd64_saturate(ai, bi); |
| 1127 | *(int64_t *)(d + i) = di; |
| 1128 | } |
| 1129 | clear_high(d, oprsz, desc); |
| 1130 | } |
| 1131 | |
| 1132 | void HELPER(gvec_sssub8)(void *d, void *a, void *b, uint32_t desc) |
| 1133 | { |
| 1134 | intptr_t oprsz = simd_oprsz(desc); |
| 1135 | intptr_t i; |
| 1136 | |
| 1137 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 1138 | int r = *(int8_t *)(a + i) - *(int8_t *)(b + i); |
| 1139 | if (r > INT8_MAX) { |
| 1140 | r = INT8_MAX; |
| 1141 | } else if (r < INT8_MIN) { |
| 1142 | r = INT8_MIN; |
| 1143 | } |
| 1144 | *(uint8_t *)(d + i) = r; |
| 1145 | } |
| 1146 | clear_high(d, oprsz, desc); |
| 1147 | } |
| 1148 | |
| 1149 | void HELPER(gvec_sssub16)(void *d, void *a, void *b, uint32_t desc) |
| 1150 | { |
| 1151 | intptr_t oprsz = simd_oprsz(desc); |
| 1152 | intptr_t i; |
| 1153 | |
| 1154 | for (i = 0; i < oprsz; i += sizeof(int16_t)) { |
| 1155 | int r = *(int16_t *)(a + i) - *(int16_t *)(b + i); |
| 1156 | if (r > INT16_MAX) { |
| 1157 | r = INT16_MAX; |
| 1158 | } else if (r < INT16_MIN) { |
| 1159 | r = INT16_MIN; |
| 1160 | } |
| 1161 | *(int16_t *)(d + i) = r; |
| 1162 | } |
| 1163 | clear_high(d, oprsz, desc); |
| 1164 | } |
| 1165 | |
| 1166 | void HELPER(gvec_sssub32)(void *d, void *a, void *b, uint32_t desc) |
| 1167 | { |
| 1168 | intptr_t oprsz = simd_oprsz(desc); |
| 1169 | intptr_t i; |
| 1170 | |
| 1171 | for (i = 0; i < oprsz; i += sizeof(int32_t)) { |
| 1172 | int32_t ai = *(int32_t *)(a + i); |
| 1173 | int32_t bi = *(int32_t *)(b + i); |
| 1174 | int32_t di = ssub32_saturate(ai, bi); |
| 1175 | *(int32_t *)(d + i) = di; |
| 1176 | } |
| 1177 | clear_high(d, oprsz, desc); |
| 1178 | } |
| 1179 | |
| 1180 | void HELPER(gvec_sssub64)(void *d, void *a, void *b, uint32_t desc) |
| 1181 | { |
| 1182 | intptr_t oprsz = simd_oprsz(desc); |
| 1183 | intptr_t i; |
| 1184 | |
| 1185 | for (i = 0; i < oprsz; i += sizeof(int64_t)) { |
| 1186 | int64_t ai = *(int64_t *)(a + i); |
| 1187 | int64_t bi = *(int64_t *)(b + i); |
| 1188 | int64_t di = ssub64_saturate(ai, bi); |
| 1189 | *(int64_t *)(d + i) = di; |
| 1190 | } |
| 1191 | clear_high(d, oprsz, desc); |
| 1192 | } |
| 1193 | |
| 1194 | void HELPER(gvec_usadd8)(void *d, void *a, void *b, uint32_t desc) |
| 1195 | { |
| 1196 | intptr_t oprsz = simd_oprsz(desc); |
| 1197 | intptr_t i; |
| 1198 | |
| 1199 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 1200 | unsigned r = *(uint8_t *)(a + i) + *(uint8_t *)(b + i); |
| 1201 | if (r > UINT8_MAX) { |
| 1202 | r = UINT8_MAX; |
| 1203 | } |
| 1204 | *(uint8_t *)(d + i) = r; |
| 1205 | } |
| 1206 | clear_high(d, oprsz, desc); |
| 1207 | } |
| 1208 | |
| 1209 | void HELPER(gvec_usadd16)(void *d, void *a, void *b, uint32_t desc) |
| 1210 | { |
| 1211 | intptr_t oprsz = simd_oprsz(desc); |
| 1212 | intptr_t i; |
| 1213 | |
| 1214 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 1215 | unsigned r = *(uint16_t *)(a + i) + *(uint16_t *)(b + i); |
| 1216 | if (r > UINT16_MAX) { |
| 1217 | r = UINT16_MAX; |
| 1218 | } |
| 1219 | *(uint16_t *)(d + i) = r; |
| 1220 | } |
| 1221 | clear_high(d, oprsz, desc); |
| 1222 | } |
| 1223 | |
| 1224 | void HELPER(gvec_usadd32)(void *d, void *a, void *b, uint32_t desc) |
| 1225 | { |
| 1226 | intptr_t oprsz = simd_oprsz(desc); |
| 1227 | intptr_t i; |
| 1228 | |
| 1229 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 1230 | uint32_t ai = *(uint32_t *)(a + i); |
| 1231 | uint32_t bi = *(uint32_t *)(b + i); |
| 1232 | uint32_t di; |
| 1233 | if (uadd32_overflow(ai, bi, &di)) { |
| 1234 | di = UINT32_MAX; |
| 1235 | } |
| 1236 | *(uint32_t *)(d + i) = di; |
| 1237 | } |
| 1238 | clear_high(d, oprsz, desc); |
| 1239 | } |
| 1240 | |
| 1241 | void HELPER(gvec_usadd64)(void *d, void *a, void *b, uint32_t desc) |
| 1242 | { |
| 1243 | intptr_t oprsz = simd_oprsz(desc); |
| 1244 | intptr_t i; |
| 1245 | |
| 1246 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 1247 | uint64_t ai = *(uint64_t *)(a + i); |
| 1248 | uint64_t bi = *(uint64_t *)(b + i); |
| 1249 | uint64_t di; |
| 1250 | if (uadd64_overflow(ai, bi, &di)) { |
| 1251 | di = UINT64_MAX; |
| 1252 | } |
| 1253 | *(uint64_t *)(d + i) = di; |
| 1254 | } |
| 1255 | clear_high(d, oprsz, desc); |
| 1256 | } |
| 1257 | |
| 1258 | void HELPER(gvec_ussub8)(void *d, void *a, void *b, uint32_t desc) |
| 1259 | { |
| 1260 | intptr_t oprsz = simd_oprsz(desc); |
| 1261 | intptr_t i; |
| 1262 | |
| 1263 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 1264 | int r = *(uint8_t *)(a + i) - *(uint8_t *)(b + i); |
| 1265 | if (r < 0) { |
| 1266 | r = 0; |
| 1267 | } |
| 1268 | *(uint8_t *)(d + i) = r; |
| 1269 | } |
| 1270 | clear_high(d, oprsz, desc); |
| 1271 | } |
| 1272 | |
| 1273 | void HELPER(gvec_ussub16)(void *d, void *a, void *b, uint32_t desc) |
| 1274 | { |
| 1275 | intptr_t oprsz = simd_oprsz(desc); |
| 1276 | intptr_t i; |
| 1277 | |
| 1278 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 1279 | int r = *(uint16_t *)(a + i) - *(uint16_t *)(b + i); |
| 1280 | if (r < 0) { |
| 1281 | r = 0; |
| 1282 | } |
| 1283 | *(uint16_t *)(d + i) = r; |
| 1284 | } |
| 1285 | clear_high(d, oprsz, desc); |
| 1286 | } |
| 1287 | |
| 1288 | void HELPER(gvec_ussub32)(void *d, void *a, void *b, uint32_t desc) |
| 1289 | { |
| 1290 | intptr_t oprsz = simd_oprsz(desc); |
| 1291 | intptr_t i; |
| 1292 | |
| 1293 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 1294 | uint32_t ai = *(uint32_t *)(a + i); |
| 1295 | uint32_t bi = *(uint32_t *)(b + i); |
| 1296 | uint32_t di; |
| 1297 | if (usub32_overflow(ai, bi, &di)) { |
| 1298 | di = 0; |
| 1299 | } |
| 1300 | *(uint32_t *)(d + i) = di; |
| 1301 | } |
| 1302 | clear_high(d, oprsz, desc); |
| 1303 | } |
| 1304 | |
| 1305 | void HELPER(gvec_ussub64)(void *d, void *a, void *b, uint32_t desc) |
| 1306 | { |
| 1307 | intptr_t oprsz = simd_oprsz(desc); |
| 1308 | intptr_t i; |
| 1309 | |
| 1310 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 1311 | uint64_t ai = *(uint64_t *)(a + i); |
| 1312 | uint64_t bi = *(uint64_t *)(b + i); |
| 1313 | uint64_t di; |
| 1314 | if (usub64_overflow(ai, bi, &di)) { |
| 1315 | di = 0; |
| 1316 | } |
| 1317 | *(uint64_t *)(d + i) = di; |
| 1318 | } |
| 1319 | clear_high(d, oprsz, desc); |
| 1320 | } |
| 1321 | |
| 1322 | void HELPER(gvec_smin8)(void *d, void *a, void *b, uint32_t desc) |
| 1323 | { |
| 1324 | intptr_t oprsz = simd_oprsz(desc); |
| 1325 | intptr_t i; |
| 1326 | |
| 1327 | for (i = 0; i < oprsz; i += sizeof(int8_t)) { |
| 1328 | int8_t aa = *(int8_t *)(a + i); |
| 1329 | int8_t bb = *(int8_t *)(b + i); |
| 1330 | int8_t dd = aa < bb ? aa : bb; |
| 1331 | *(int8_t *)(d + i) = dd; |
| 1332 | } |
| 1333 | clear_high(d, oprsz, desc); |
| 1334 | } |
| 1335 | |
| 1336 | void HELPER(gvec_smin16)(void *d, void *a, void *b, uint32_t desc) |
| 1337 | { |
| 1338 | intptr_t oprsz = simd_oprsz(desc); |
| 1339 | intptr_t i; |
| 1340 | |
| 1341 | for (i = 0; i < oprsz; i += sizeof(int16_t)) { |
| 1342 | int16_t aa = *(int16_t *)(a + i); |
| 1343 | int16_t bb = *(int16_t *)(b + i); |
| 1344 | int16_t dd = aa < bb ? aa : bb; |
| 1345 | *(int16_t *)(d + i) = dd; |
| 1346 | } |
| 1347 | clear_high(d, oprsz, desc); |
| 1348 | } |
| 1349 | |
| 1350 | void HELPER(gvec_smin32)(void *d, void *a, void *b, uint32_t desc) |
| 1351 | { |
| 1352 | intptr_t oprsz = simd_oprsz(desc); |
| 1353 | intptr_t i; |
| 1354 | |
| 1355 | for (i = 0; i < oprsz; i += sizeof(int32_t)) { |
| 1356 | int32_t aa = *(int32_t *)(a + i); |
| 1357 | int32_t bb = *(int32_t *)(b + i); |
| 1358 | int32_t dd = aa < bb ? aa : bb; |
| 1359 | *(int32_t *)(d + i) = dd; |
| 1360 | } |
| 1361 | clear_high(d, oprsz, desc); |
| 1362 | } |
| 1363 | |
| 1364 | void HELPER(gvec_smin64)(void *d, void *a, void *b, uint32_t desc) |
| 1365 | { |
| 1366 | intptr_t oprsz = simd_oprsz(desc); |
| 1367 | intptr_t i; |
| 1368 | |
| 1369 | for (i = 0; i < oprsz; i += sizeof(int64_t)) { |
| 1370 | int64_t aa = *(int64_t *)(a + i); |
| 1371 | int64_t bb = *(int64_t *)(b + i); |
| 1372 | int64_t dd = aa < bb ? aa : bb; |
| 1373 | *(int64_t *)(d + i) = dd; |
| 1374 | } |
| 1375 | clear_high(d, oprsz, desc); |
| 1376 | } |
| 1377 | |
| 1378 | void HELPER(gvec_smax8)(void *d, void *a, void *b, uint32_t desc) |
| 1379 | { |
| 1380 | intptr_t oprsz = simd_oprsz(desc); |
| 1381 | intptr_t i; |
| 1382 | |
| 1383 | for (i = 0; i < oprsz; i += sizeof(int8_t)) { |
| 1384 | int8_t aa = *(int8_t *)(a + i); |
| 1385 | int8_t bb = *(int8_t *)(b + i); |
| 1386 | int8_t dd = aa > bb ? aa : bb; |
| 1387 | *(int8_t *)(d + i) = dd; |
| 1388 | } |
| 1389 | clear_high(d, oprsz, desc); |
| 1390 | } |
| 1391 | |
| 1392 | void HELPER(gvec_smax16)(void *d, void *a, void *b, uint32_t desc) |
| 1393 | { |
| 1394 | intptr_t oprsz = simd_oprsz(desc); |
| 1395 | intptr_t i; |
| 1396 | |
| 1397 | for (i = 0; i < oprsz; i += sizeof(int16_t)) { |
| 1398 | int16_t aa = *(int16_t *)(a + i); |
| 1399 | int16_t bb = *(int16_t *)(b + i); |
| 1400 | int16_t dd = aa > bb ? aa : bb; |
| 1401 | *(int16_t *)(d + i) = dd; |
| 1402 | } |
| 1403 | clear_high(d, oprsz, desc); |
| 1404 | } |
| 1405 | |
| 1406 | void HELPER(gvec_smax32)(void *d, void *a, void *b, uint32_t desc) |
| 1407 | { |
| 1408 | intptr_t oprsz = simd_oprsz(desc); |
| 1409 | intptr_t i; |
| 1410 | |
| 1411 | for (i = 0; i < oprsz; i += sizeof(int32_t)) { |
| 1412 | int32_t aa = *(int32_t *)(a + i); |
| 1413 | int32_t bb = *(int32_t *)(b + i); |
| 1414 | int32_t dd = aa > bb ? aa : bb; |
| 1415 | *(int32_t *)(d + i) = dd; |
| 1416 | } |
| 1417 | clear_high(d, oprsz, desc); |
| 1418 | } |
| 1419 | |
| 1420 | void HELPER(gvec_smax64)(void *d, void *a, void *b, uint32_t desc) |
| 1421 | { |
| 1422 | intptr_t oprsz = simd_oprsz(desc); |
| 1423 | intptr_t i; |
| 1424 | |
| 1425 | for (i = 0; i < oprsz; i += sizeof(int64_t)) { |
| 1426 | int64_t aa = *(int64_t *)(a + i); |
| 1427 | int64_t bb = *(int64_t *)(b + i); |
| 1428 | int64_t dd = aa > bb ? aa : bb; |
| 1429 | *(int64_t *)(d + i) = dd; |
| 1430 | } |
| 1431 | clear_high(d, oprsz, desc); |
| 1432 | } |
| 1433 | |
| 1434 | void HELPER(gvec_umin8)(void *d, void *a, void *b, uint32_t desc) |
| 1435 | { |
| 1436 | intptr_t oprsz = simd_oprsz(desc); |
| 1437 | intptr_t i; |
| 1438 | |
| 1439 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 1440 | uint8_t aa = *(uint8_t *)(a + i); |
| 1441 | uint8_t bb = *(uint8_t *)(b + i); |
| 1442 | uint8_t dd = aa < bb ? aa : bb; |
| 1443 | *(uint8_t *)(d + i) = dd; |
| 1444 | } |
| 1445 | clear_high(d, oprsz, desc); |
| 1446 | } |
| 1447 | |
| 1448 | void HELPER(gvec_umin16)(void *d, void *a, void *b, uint32_t desc) |
| 1449 | { |
| 1450 | intptr_t oprsz = simd_oprsz(desc); |
| 1451 | intptr_t i; |
| 1452 | |
| 1453 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 1454 | uint16_t aa = *(uint16_t *)(a + i); |
| 1455 | uint16_t bb = *(uint16_t *)(b + i); |
| 1456 | uint16_t dd = aa < bb ? aa : bb; |
| 1457 | *(uint16_t *)(d + i) = dd; |
| 1458 | } |
| 1459 | clear_high(d, oprsz, desc); |
| 1460 | } |
| 1461 | |
| 1462 | void HELPER(gvec_umin32)(void *d, void *a, void *b, uint32_t desc) |
| 1463 | { |
| 1464 | intptr_t oprsz = simd_oprsz(desc); |
| 1465 | intptr_t i; |
| 1466 | |
| 1467 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 1468 | uint32_t aa = *(uint32_t *)(a + i); |
| 1469 | uint32_t bb = *(uint32_t *)(b + i); |
| 1470 | uint32_t dd = aa < bb ? aa : bb; |
| 1471 | *(uint32_t *)(d + i) = dd; |
| 1472 | } |
| 1473 | clear_high(d, oprsz, desc); |
| 1474 | } |
| 1475 | |
| 1476 | void HELPER(gvec_umin64)(void *d, void *a, void *b, uint32_t desc) |
| 1477 | { |
| 1478 | intptr_t oprsz = simd_oprsz(desc); |
| 1479 | intptr_t i; |
| 1480 | |
| 1481 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 1482 | uint64_t aa = *(uint64_t *)(a + i); |
| 1483 | uint64_t bb = *(uint64_t *)(b + i); |
| 1484 | uint64_t dd = aa < bb ? aa : bb; |
| 1485 | *(uint64_t *)(d + i) = dd; |
| 1486 | } |
| 1487 | clear_high(d, oprsz, desc); |
| 1488 | } |
| 1489 | |
| 1490 | void HELPER(gvec_umax8)(void *d, void *a, void *b, uint32_t desc) |
| 1491 | { |
| 1492 | intptr_t oprsz = simd_oprsz(desc); |
| 1493 | intptr_t i; |
| 1494 | |
| 1495 | for (i = 0; i < oprsz; i += sizeof(uint8_t)) { |
| 1496 | uint8_t aa = *(uint8_t *)(a + i); |
| 1497 | uint8_t bb = *(uint8_t *)(b + i); |
| 1498 | uint8_t dd = aa > bb ? aa : bb; |
| 1499 | *(uint8_t *)(d + i) = dd; |
| 1500 | } |
| 1501 | clear_high(d, oprsz, desc); |
| 1502 | } |
| 1503 | |
| 1504 | void HELPER(gvec_umax16)(void *d, void *a, void *b, uint32_t desc) |
| 1505 | { |
| 1506 | intptr_t oprsz = simd_oprsz(desc); |
| 1507 | intptr_t i; |
| 1508 | |
| 1509 | for (i = 0; i < oprsz; i += sizeof(uint16_t)) { |
| 1510 | uint16_t aa = *(uint16_t *)(a + i); |
| 1511 | uint16_t bb = *(uint16_t *)(b + i); |
| 1512 | uint16_t dd = aa > bb ? aa : bb; |
| 1513 | *(uint16_t *)(d + i) = dd; |
| 1514 | } |
| 1515 | clear_high(d, oprsz, desc); |
| 1516 | } |
| 1517 | |
| 1518 | void HELPER(gvec_umax32)(void *d, void *a, void *b, uint32_t desc) |
| 1519 | { |
| 1520 | intptr_t oprsz = simd_oprsz(desc); |
| 1521 | intptr_t i; |
| 1522 | |
| 1523 | for (i = 0; i < oprsz; i += sizeof(uint32_t)) { |
| 1524 | uint32_t aa = *(uint32_t *)(a + i); |
| 1525 | uint32_t bb = *(uint32_t *)(b + i); |
| 1526 | uint32_t dd = aa > bb ? aa : bb; |
| 1527 | *(uint32_t *)(d + i) = dd; |
| 1528 | } |
| 1529 | clear_high(d, oprsz, desc); |
| 1530 | } |
| 1531 | |
| 1532 | void HELPER(gvec_umax64)(void *d, void *a, void *b, uint32_t desc) |
| 1533 | { |
| 1534 | intptr_t oprsz = simd_oprsz(desc); |
| 1535 | intptr_t i; |
| 1536 | |
| 1537 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 1538 | uint64_t aa = *(uint64_t *)(a + i); |
| 1539 | uint64_t bb = *(uint64_t *)(b + i); |
| 1540 | uint64_t dd = aa > bb ? aa : bb; |
| 1541 | *(uint64_t *)(d + i) = dd; |
| 1542 | } |
| 1543 | clear_high(d, oprsz, desc); |
| 1544 | } |
| 1545 | |
| 1546 | void HELPER(gvec_bitsel)(void *d, void *a, void *b, void *c, uint32_t desc) |
| 1547 | { |
| 1548 | intptr_t oprsz = simd_oprsz(desc); |
| 1549 | intptr_t i; |
| 1550 | |
| 1551 | for (i = 0; i < oprsz; i += sizeof(uint64_t)) { |
| 1552 | uint64_t aa = *(uint64_t *)(a + i); |
| 1553 | uint64_t bb = *(uint64_t *)(b + i); |
| 1554 | uint64_t cc = *(uint64_t *)(c + i); |
| 1555 | *(uint64_t *)(d + i) = (bb & aa) | (cc & ~aa); |
| 1556 | } |
| 1557 | clear_high(d, oprsz, desc); |
| 1558 | } |