gcc/config/aarch64/arm_neon.h

   1 /* ARM NEON intrinsics include file.
   2
   3    Copyright (C) 2011-2018 Free Software Foundation, Inc.
   4    Contributed by ARM Ltd.
   5
   6    This file is part of GCC.
   7
   8    GCC is free software; you can redistribute it and/or modify it
   9    under the terms of the GNU General Public License as published
  10    by the Free Software Foundation; either version 3, or (at your
  11    option) any later version.
  12
  13    GCC is distributed in the hope that it will be useful, but WITHOUT
  14    ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  15    or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
  16    License for more details.
  17
  18    Under Section 7 of GPL version 3, you are granted additional
  19    permissions described in the GCC Runtime Library Exception, version
  20    3.1, as published by the Free Software Foundation.
  21
  22    You should have received a copy of the GNU General Public License and
  23    a copy of the GCC Runtime Library Exception along with this program;
  24    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  25    <http://www.gnu.org/licenses/>.  */
  26
  27 #ifndef _AARCH64_NEON_H_
  28 #define _AARCH64_NEON_H_
  29
  30 #pragma GCC push_options
  31 #pragma GCC target ("+nothing+simd")
  32
  33 #include <stdint.h>
  34
  35 #define __AARCH64_UINT64_C(__C) ((uint64_t) __C)
  36 #define __AARCH64_INT64_C(__C) ((int64_t) __C)
  37
  38 typedef __Int8x8_t int8x8_t;
  39 typedef __Int16x4_t int16x4_t;
  40 typedef __Int32x2_t int32x2_t;
  41 typedef __Int64x1_t int64x1_t;
  42 typedef __Float16x4_t float16x4_t;
  43 typedef __Float32x2_t float32x2_t;
  44 typedef __Poly8x8_t poly8x8_t;
  45 typedef __Poly16x4_t poly16x4_t;
  46 typedef __Uint8x8_t uint8x8_t;
  47 typedef __Uint16x4_t uint16x4_t;
  48 typedef __Uint32x2_t uint32x2_t;
  49 typedef __Float64x1_t float64x1_t;
  50 typedef __Uint64x1_t uint64x1_t;
  51 typedef __Int8x16_t int8x16_t;
  52 typedef __Int16x8_t int16x8_t;
  53 typedef __Int32x4_t int32x4_t;
  54 typedef __Int64x2_t int64x2_t;
  55 typedef __Float16x8_t float16x8_t;
  56 typedef __Float32x4_t float32x4_t;
  57 typedef __Float64x2_t float64x2_t;
  58 typedef __Poly8x16_t poly8x16_t;
  59 typedef __Poly16x8_t poly16x8_t;
  60 typedef __Poly64x2_t poly64x2_t;
  61 typedef __Poly64x1_t poly64x1_t;
  62 typedef __Uint8x16_t uint8x16_t;
  63 typedef __Uint16x8_t uint16x8_t;
  64 typedef __Uint32x4_t uint32x4_t;
  65 typedef __Uint64x2_t uint64x2_t;
  66
  67 typedef __Poly8_t poly8_t;
  68 typedef __Poly16_t poly16_t;
  69 typedef __Poly64_t poly64_t;
  70 typedef __Poly128_t poly128_t;
  71
  72 typedef __fp16 float16_t;
  73 typedef float float32_t;
  74 typedef double float64_t;
  75
  76 typedef struct int8x8x2_t
  77 {
  78   int8x8_t val[2];
  79 } int8x8x2_t;
  80
  81 typedef struct int8x16x2_t
  82 {
  83   int8x16_t val[2];
  84 } int8x16x2_t;
  85
  86 typedef struct int16x4x2_t
  87 {
  88   int16x4_t val[2];
  89 } int16x4x2_t;
  90
  91 typedef struct int16x8x2_t
  92 {
  93   int16x8_t val[2];
  94 } int16x8x2_t;
  95
  96 typedef struct int32x2x2_t
  97 {
  98   int32x2_t val[2];
  99 } int32x2x2_t;
 100
 101 typedef struct int32x4x2_t
 102 {
 103   int32x4_t val[2];
 104 } int32x4x2_t;
 105
 106 typedef struct int64x1x2_t
 107 {
 108   int64x1_t val[2];
 109 } int64x1x2_t;
 110
 111 typedef struct int64x2x2_t
 112 {
 113   int64x2_t val[2];
 114 } int64x2x2_t;
 115
 116 typedef struct uint8x8x2_t
 117 {
 118   uint8x8_t val[2];
 119 } uint8x8x2_t;
 120
 121 typedef struct uint8x16x2_t
 122 {
 123   uint8x16_t val[2];
 124 } uint8x16x2_t;
 125
 126 typedef struct uint16x4x2_t
 127 {
 128   uint16x4_t val[2];
 129 } uint16x4x2_t;
 130
 131 typedef struct uint16x8x2_t
 132 {
 133   uint16x8_t val[2];
 134 } uint16x8x2_t;
 135
 136 typedef struct uint32x2x2_t
 137 {
 138   uint32x2_t val[2];
 139 } uint32x2x2_t;
 140
 141 typedef struct uint32x4x2_t
 142 {
 143   uint32x4_t val[2];
 144 } uint32x4x2_t;
 145
 146 typedef struct uint64x1x2_t
 147 {
 148   uint64x1_t val[2];
 149 } uint64x1x2_t;
 150
 151 typedef struct uint64x2x2_t
 152 {
 153   uint64x2_t val[2];
 154 } uint64x2x2_t;
 155
 156 typedef struct float16x4x2_t
 157 {
 158   float16x4_t val[2];
 159 } float16x4x2_t;
 160
 161 typedef struct float16x8x2_t
 162 {
 163   float16x8_t val[2];
 164 } float16x8x2_t;
 165
 166 typedef struct float32x2x2_t
 167 {
 168   float32x2_t val[2];
 169 } float32x2x2_t;
 170
 171 typedef struct float32x4x2_t
 172 {
 173   float32x4_t val[2];
 174 } float32x4x2_t;
 175
 176 typedef struct float64x2x2_t
 177 {
 178   float64x2_t val[2];
 179 } float64x2x2_t;
 180
 181 typedef struct float64x1x2_t
 182 {
 183   float64x1_t val[2];
 184 } float64x1x2_t;
 185
 186 typedef struct poly8x8x2_t
 187 {
 188   poly8x8_t val[2];
 189 } poly8x8x2_t;
 190
 191 typedef struct poly8x16x2_t
 192 {
 193   poly8x16_t val[2];
 194 } poly8x16x2_t;
 195
 196 typedef struct poly16x4x2_t
 197 {
 198   poly16x4_t val[2];
 199 } poly16x4x2_t;
 200
 201 typedef struct poly16x8x2_t
 202 {
 203   poly16x8_t val[2];
 204 } poly16x8x2_t;
 205
 206 typedef struct poly64x1x2_t
 207 {
 208   poly64x1_t val[2];
 209 } poly64x1x2_t;
 210
 211 typedef struct poly64x1x3_t
 212 {
 213   poly64x1_t val[3];
 214 } poly64x1x3_t;
 215
 216 typedef struct poly64x1x4_t
 217 {
 218   poly64x1_t val[4];
 219 } poly64x1x4_t;
 220
 221 typedef struct poly64x2x2_t
 222 {
 223   poly64x2_t val[2];
 224 } poly64x2x2_t;
 225
 226 typedef struct poly64x2x3_t
 227 {
 228   poly64x2_t val[3];
 229 } poly64x2x3_t;
 230
 231 typedef struct poly64x2x4_t
 232 {
 233   poly64x2_t val[4];
 234 } poly64x2x4_t;
 235
 236 typedef struct int8x8x3_t
 237 {
 238   int8x8_t val[3];
 239 } int8x8x3_t;
 240
 241 typedef struct int8x16x3_t
 242 {
 243   int8x16_t val[3];
 244 } int8x16x3_t;
 245
 246 typedef struct int16x4x3_t
 247 {
 248   int16x4_t val[3];
 249 } int16x4x3_t;
 250
 251 typedef struct int16x8x3_t
 252 {
 253   int16x8_t val[3];
 254 } int16x8x3_t;
 255
 256 typedef struct int32x2x3_t
 257 {
 258   int32x2_t val[3];
 259 } int32x2x3_t;
 260
 261 typedef struct int32x4x3_t
 262 {
 263   int32x4_t val[3];
 264 } int32x4x3_t;
 265
 266 typedef struct int64x1x3_t
 267 {
 268   int64x1_t val[3];
 269 } int64x1x3_t;
 270
 271 typedef struct int64x2x3_t
 272 {
 273   int64x2_t val[3];
 274 } int64x2x3_t;
 275
 276 typedef struct uint8x8x3_t
 277 {
 278   uint8x8_t val[3];
 279 } uint8x8x3_t;
 280
 281 typedef struct uint8x16x3_t
 282 {
 283   uint8x16_t val[3];
 284 } uint8x16x3_t;
 285
 286 typedef struct uint16x4x3_t
 287 {
 288   uint16x4_t val[3];
 289 } uint16x4x3_t;
 290
 291 typedef struct uint16x8x3_t
 292 {
 293   uint16x8_t val[3];
 294 } uint16x8x3_t;
 295
 296 typedef struct uint32x2x3_t
 297 {
 298   uint32x2_t val[3];
 299 } uint32x2x3_t;
 300
 301 typedef struct uint32x4x3_t
 302 {
 303   uint32x4_t val[3];
 304 } uint32x4x3_t;
 305
 306 typedef struct uint64x1x3_t
 307 {
 308   uint64x1_t val[3];
 309 } uint64x1x3_t;
 310
 311 typedef struct uint64x2x3_t
 312 {
 313   uint64x2_t val[3];
 314 } uint64x2x3_t;
 315
 316 typedef struct float16x4x3_t
 317 {
 318   float16x4_t val[3];
 319 } float16x4x3_t;
 320
 321 typedef struct float16x8x3_t
 322 {
 323   float16x8_t val[3];
 324 } float16x8x3_t;
 325
 326 typedef struct float32x2x3_t
 327 {
 328   float32x2_t val[3];
 329 } float32x2x3_t;
 330
 331 typedef struct float32x4x3_t
 332 {
 333   float32x4_t val[3];
 334 } float32x4x3_t;
 335
 336 typedef struct float64x2x3_t
 337 {
 338   float64x2_t val[3];
 339 } float64x2x3_t;
 340
 341 typedef struct float64x1x3_t
 342 {
 343   float64x1_t val[3];
 344 } float64x1x3_t;
 345
 346 typedef struct poly8x8x3_t
 347 {
 348   poly8x8_t val[3];
 349 } poly8x8x3_t;
 350
 351 typedef struct poly8x16x3_t
 352 {
 353   poly8x16_t val[3];
 354 } poly8x16x3_t;
 355
 356 typedef struct poly16x4x3_t
 357 {
 358   poly16x4_t val[3];
 359 } poly16x4x3_t;
 360
 361 typedef struct poly16x8x3_t
 362 {
 363   poly16x8_t val[3];
 364 } poly16x8x3_t;
 365
 366 typedef struct int8x8x4_t
 367 {
 368   int8x8_t val[4];
 369 } int8x8x4_t;
 370
 371 typedef struct int8x16x4_t
 372 {
 373   int8x16_t val[4];
 374 } int8x16x4_t;
 375
 376 typedef struct int16x4x4_t
 377 {
 378   int16x4_t val[4];
 379 } int16x4x4_t;
 380
 381 typedef struct int16x8x4_t
 382 {
 383   int16x8_t val[4];
 384 } int16x8x4_t;
 385
 386 typedef struct int32x2x4_t
 387 {
 388   int32x2_t val[4];
 389 } int32x2x4_t;
 390
 391 typedef struct int32x4x4_t
 392 {
 393   int32x4_t val[4];
 394 } int32x4x4_t;
 395
 396 typedef struct int64x1x4_t
 397 {
 398   int64x1_t val[4];
 399 } int64x1x4_t;
 400
 401 typedef struct int64x2x4_t
 402 {
 403   int64x2_t val[4];
 404 } int64x2x4_t;
 405
 406 typedef struct uint8x8x4_t
 407 {
 408   uint8x8_t val[4];
 409 } uint8x8x4_t;
 410
 411 typedef struct uint8x16x4_t
 412 {
 413   uint8x16_t val[4];
 414 } uint8x16x4_t;
 415
 416 typedef struct uint16x4x4_t
 417 {
 418   uint16x4_t val[4];
 419 } uint16x4x4_t;
 420
 421 typedef struct uint16x8x4_t
 422 {
 423   uint16x8_t val[4];
 424 } uint16x8x4_t;
 425
 426 typedef struct uint32x2x4_t
 427 {
 428   uint32x2_t val[4];
 429 } uint32x2x4_t;
 430
 431 typedef struct uint32x4x4_t
 432 {
 433   uint32x4_t val[4];
 434 } uint32x4x4_t;
 435
 436 typedef struct uint64x1x4_t
 437 {
 438   uint64x1_t val[4];
 439 } uint64x1x4_t;
 440
 441 typedef struct uint64x2x4_t
 442 {
 443   uint64x2_t val[4];
 444 } uint64x2x4_t;
 445
 446 typedef struct float16x4x4_t
 447 {
 448   float16x4_t val[4];
 449 } float16x4x4_t;
 450
 451 typedef struct float16x8x4_t
 452 {
 453   float16x8_t val[4];
 454 } float16x8x4_t;
 455
 456 typedef struct float32x2x4_t
 457 {
 458   float32x2_t val[4];
 459 } float32x2x4_t;
 460
 461 typedef struct float32x4x4_t
 462 {
 463   float32x4_t val[4];
 464 } float32x4x4_t;
 465
 466 typedef struct float64x2x4_t
 467 {
 468   float64x2_t val[4];
 469 } float64x2x4_t;
 470
 471 typedef struct float64x1x4_t
 472 {
 473   float64x1_t val[4];
 474 } float64x1x4_t;
 475
 476 typedef struct poly8x8x4_t
 477 {
 478   poly8x8_t val[4];
 479 } poly8x8x4_t;
 480
 481 typedef struct poly8x16x4_t
 482 {
 483   poly8x16_t val[4];
 484 } poly8x16x4_t;
 485
 486 typedef struct poly16x4x4_t
 487 {
 488   poly16x4_t val[4];
 489 } poly16x4x4_t;
 490
 491 typedef struct poly16x8x4_t
 492 {
 493   poly16x8_t val[4];
 494 } poly16x8x4_t;
 495
 496 /* __aarch64_vdup_lane internal macros.  */
 497 #define __aarch64_vdup_lane_any(__size, __q, __a, __b) \
 498   vdup##__q##_n_##__size (__aarch64_vget_lane_any (__a, __b))
 499
 500 #define __aarch64_vdup_lane_f16(__a, __b) \
 501    __aarch64_vdup_lane_any (f16, , __a, __b)
 502 #define __aarch64_vdup_lane_f32(__a, __b) \
 503    __aarch64_vdup_lane_any (f32, , __a, __b)
 504 #define __aarch64_vdup_lane_f64(__a, __b) \
 505    __aarch64_vdup_lane_any (f64, , __a, __b)
 506 #define __aarch64_vdup_lane_p8(__a, __b) \
 507    __aarch64_vdup_lane_any (p8, , __a, __b)
 508 #define __aarch64_vdup_lane_p16(__a, __b) \
 509    __aarch64_vdup_lane_any (p16, , __a, __b)
 510 #define __aarch64_vdup_lane_p64(__a, __b) \
 511    __aarch64_vdup_lane_any (p64, , __a, __b)
 512 #define __aarch64_vdup_lane_s8(__a, __b) \
 513    __aarch64_vdup_lane_any (s8, , __a, __b)
 514 #define __aarch64_vdup_lane_s16(__a, __b) \
 515    __aarch64_vdup_lane_any (s16, , __a, __b)
 516 #define __aarch64_vdup_lane_s32(__a, __b) \
 517    __aarch64_vdup_lane_any (s32, , __a, __b)
 518 #define __aarch64_vdup_lane_s64(__a, __b) \
 519   __aarch64_vdup_lane_any (s64, , __a, __b)
 520 #define __aarch64_vdup_lane_u8(__a, __b) \
 521    __aarch64_vdup_lane_any (u8, , __a, __b)
 522 #define __aarch64_vdup_lane_u16(__a, __b) \
 523    __aarch64_vdup_lane_any (u16, , __a, __b)
 524 #define __aarch64_vdup_lane_u32(__a, __b) \
 525    __aarch64_vdup_lane_any (u32, , __a, __b)
 526 #define __aarch64_vdup_lane_u64(__a, __b) \
 527    __aarch64_vdup_lane_any (u64, , __a, __b)
 528
 529 /* __aarch64_vdup_laneq internal macros.  */
 530 #define __aarch64_vdup_laneq_f16(__a, __b) \
 531    __aarch64_vdup_lane_any (f16, , __a, __b)
 532 #define __aarch64_vdup_laneq_f32(__a, __b) \
 533    __aarch64_vdup_lane_any (f32, , __a, __b)
 534 #define __aarch64_vdup_laneq_f64(__a, __b) \
 535    __aarch64_vdup_lane_any (f64, , __a, __b)
 536 #define __aarch64_vdup_laneq_p8(__a, __b) \
 537    __aarch64_vdup_lane_any (p8, , __a, __b)
 538 #define __aarch64_vdup_laneq_p16(__a, __b) \
 539    __aarch64_vdup_lane_any (p16, , __a, __b)
 540 #define __aarch64_vdup_laneq_p64(__a, __b) \
 541    __aarch64_vdup_lane_any (p64, , __a, __b)
 542 #define __aarch64_vdup_laneq_s8(__a, __b) \
 543    __aarch64_vdup_lane_any (s8, , __a, __b)
 544 #define __aarch64_vdup_laneq_s16(__a, __b) \
 545    __aarch64_vdup_lane_any (s16, , __a, __b)
 546 #define __aarch64_vdup_laneq_s32(__a, __b) \
 547    __aarch64_vdup_lane_any (s32, , __a, __b)
 548 #define __aarch64_vdup_laneq_s64(__a, __b) \
 549    __aarch64_vdup_lane_any (s64, , __a, __b)
 550 #define __aarch64_vdup_laneq_u8(__a, __b) \
 551    __aarch64_vdup_lane_any (u8, , __a, __b)
 552 #define __aarch64_vdup_laneq_u16(__a, __b) \
 553    __aarch64_vdup_lane_any (u16, , __a, __b)
 554 #define __aarch64_vdup_laneq_u32(__a, __b) \
 555    __aarch64_vdup_lane_any (u32, , __a, __b)
 556 #define __aarch64_vdup_laneq_u64(__a, __b) \
 557    __aarch64_vdup_lane_any (u64, , __a, __b)
 558
 559 /* __aarch64_vdupq_lane internal macros.  */
 560 #define __aarch64_vdupq_lane_f16(__a, __b) \
 561    __aarch64_vdup_lane_any (f16, q, __a, __b)
 562 #define __aarch64_vdupq_lane_f32(__a, __b) \
 563    __aarch64_vdup_lane_any (f32, q, __a, __b)
 564 #define __aarch64_vdupq_lane_f64(__a, __b) \
 565    __aarch64_vdup_lane_any (f64, q, __a, __b)
 566 #define __aarch64_vdupq_lane_p8(__a, __b) \
 567    __aarch64_vdup_lane_any (p8, q, __a, __b)
 568 #define __aarch64_vdupq_lane_p16(__a, __b) \
 569    __aarch64_vdup_lane_any (p16, q, __a, __b)
 570 #define __aarch64_vdupq_lane_p64(__a, __b) \
 571    __aarch64_vdup_lane_any (p64, q, __a, __b)
 572 #define __aarch64_vdupq_lane_s8(__a, __b) \
 573    __aarch64_vdup_lane_any (s8, q, __a, __b)
 574 #define __aarch64_vdupq_lane_s16(__a, __b) \
 575    __aarch64_vdup_lane_any (s16, q, __a, __b)
 576 #define __aarch64_vdupq_lane_s32(__a, __b) \
 577    __aarch64_vdup_lane_any (s32, q, __a, __b)
 578 #define __aarch64_vdupq_lane_s64(__a, __b) \
 579    __aarch64_vdup_lane_any (s64, q, __a, __b)
 580 #define __aarch64_vdupq_lane_u8(__a, __b) \
 581    __aarch64_vdup_lane_any (u8, q, __a, __b)
 582 #define __aarch64_vdupq_lane_u16(__a, __b) \
 583    __aarch64_vdup_lane_any (u16, q, __a, __b)
 584 #define __aarch64_vdupq_lane_u32(__a, __b) \
 585    __aarch64_vdup_lane_any (u32, q, __a, __b)
 586 #define __aarch64_vdupq_lane_u64(__a, __b) \
 587    __aarch64_vdup_lane_any (u64, q, __a, __b)
 588
 589 /* __aarch64_vdupq_laneq internal macros.  */
 590 #define __aarch64_vdupq_laneq_f16(__a, __b) \
 591    __aarch64_vdup_lane_any (f16, q, __a, __b)
 592 #define __aarch64_vdupq_laneq_f32(__a, __b) \
 593    __aarch64_vdup_lane_any (f32, q, __a, __b)
 594 #define __aarch64_vdupq_laneq_f64(__a, __b) \
 595    __aarch64_vdup_lane_any (f64, q, __a, __b)
 596 #define __aarch64_vdupq_laneq_p8(__a, __b) \
 597    __aarch64_vdup_lane_any (p8, q, __a, __b)
 598 #define __aarch64_vdupq_laneq_p16(__a, __b) \
 599    __aarch64_vdup_lane_any (p16, q, __a, __b)
 600 #define __aarch64_vdupq_laneq_p64(__a, __b) \
 601    __aarch64_vdup_lane_any (p64, q, __a, __b)
 602 #define __aarch64_vdupq_laneq_s8(__a, __b) \
 603    __aarch64_vdup_lane_any (s8, q, __a, __b)
 604 #define __aarch64_vdupq_laneq_s16(__a, __b) \
 605    __aarch64_vdup_lane_any (s16, q, __a, __b)
 606 #define __aarch64_vdupq_laneq_s32(__a, __b) \
 607    __aarch64_vdup_lane_any (s32, q, __a, __b)
 608 #define __aarch64_vdupq_laneq_s64(__a, __b) \
 609    __aarch64_vdup_lane_any (s64, q, __a, __b)
 610 #define __aarch64_vdupq_laneq_u8(__a, __b) \
 611    __aarch64_vdup_lane_any (u8, q, __a, __b)
 612 #define __aarch64_vdupq_laneq_u16(__a, __b) \
 613    __aarch64_vdup_lane_any (u16, q, __a, __b)
 614 #define __aarch64_vdupq_laneq_u32(__a, __b) \
 615    __aarch64_vdup_lane_any (u32, q, __a, __b)
 616 #define __aarch64_vdupq_laneq_u64(__a, __b) \
 617    __aarch64_vdup_lane_any (u64, q, __a, __b)
 618
 619 /* Internal macro for lane indices.  */
 620
 621 #define __AARCH64_NUM_LANES(__v) (sizeof (__v) / sizeof (__v[0]))
 622 #define __AARCH64_LANE_CHECK(__vec, __idx)      \
 623         __builtin_aarch64_im_lane_boundsi (sizeof(__vec), sizeof(__vec[0]), __idx)
 624
 625 /* For big-endian, GCC's vector indices are the opposite way around
 626    to the architectural lane indices used by Neon intrinsics.  */
 627 #ifdef __AARCH64EB__
 628 #define __aarch64_lane(__vec, __idx) (__AARCH64_NUM_LANES (__vec) - 1 - __idx)
 629 #else
 630 #define __aarch64_lane(__vec, __idx) __idx
 631 #endif
 632
 633 /* vget_lane internal macro.  */
 634 #define __aarch64_vget_lane_any(__vec, __index)                         \
 635   __extension__                                                         \
 636   ({                                                                    \
 637     __AARCH64_LANE_CHECK (__vec, __index);                              \
 638     __vec[__aarch64_lane (__vec, __index)];                             \
 639   })
 640
 641 /* vset_lane and vld1_lane internal macro.  */
 642 #define __aarch64_vset_lane_any(__elem, __vec, __index)                 \
 643   __extension__                                                         \
 644   ({                                                                    \
 645     __AARCH64_LANE_CHECK (__vec, __index);                              \
 646     __vec[__aarch64_lane (__vec, __index)] = __elem;                    \
 647     __vec;                                                              \
 648   })
 649
 650 /* vadd  */
 651 __extension__ extern __inline int8x8_t
 652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 653 vadd_s8 (int8x8_t __a, int8x8_t __b)
 654 {
 655   return __a + __b;
 656 }
 657
 658 __extension__ extern __inline int16x4_t
 659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 660 vadd_s16 (int16x4_t __a, int16x4_t __b)
 661 {
 662   return __a + __b;
 663 }
 664
 665 __extension__ extern __inline int32x2_t
 666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 667 vadd_s32 (int32x2_t __a, int32x2_t __b)
 668 {
 669   return __a + __b;
 670 }
 671
 672 __extension__ extern __inline float32x2_t
 673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 674 vadd_f32 (float32x2_t __a, float32x2_t __b)
 675 {
 676   return __a + __b;
 677 }
 678
 679 __extension__ extern __inline float64x1_t
 680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 681 vadd_f64 (float64x1_t __a, float64x1_t __b)
 682 {
 683   return __a + __b;
 684 }
 685
 686 __extension__ extern __inline uint8x8_t
 687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 688 vadd_u8 (uint8x8_t __a, uint8x8_t __b)
 689 {
 690   return __a + __b;
 691 }
 692
 693 __extension__ extern __inline uint16x4_t
 694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 695 vadd_u16 (uint16x4_t __a, uint16x4_t __b)
 696 {
 697   return __a + __b;
 698 }
 699
 700 __extension__ extern __inline uint32x2_t
 701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 702 vadd_u32 (uint32x2_t __a, uint32x2_t __b)
 703 {
 704   return __a + __b;
 705 }
 706
 707 __extension__ extern __inline int64x1_t
 708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 709 vadd_s64 (int64x1_t __a, int64x1_t __b)
 710 {
 711   return __a + __b;
 712 }
 713
 714 __extension__ extern __inline uint64x1_t
 715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 716 vadd_u64 (uint64x1_t __a, uint64x1_t __b)
 717 {
 718   return __a + __b;
 719 }
 720
 721 __extension__ extern __inline int8x16_t
 722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 723 vaddq_s8 (int8x16_t __a, int8x16_t __b)
 724 {
 725   return __a + __b;
 726 }
 727
 728 __extension__ extern __inline int16x8_t
 729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 730 vaddq_s16 (int16x8_t __a, int16x8_t __b)
 731 {
 732   return __a + __b;
 733 }
 734
 735 __extension__ extern __inline int32x4_t
 736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 737 vaddq_s32 (int32x4_t __a, int32x4_t __b)
 738 {
 739   return __a + __b;
 740 }
 741
 742 __extension__ extern __inline int64x2_t
 743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 744 vaddq_s64 (int64x2_t __a, int64x2_t __b)
 745 {
 746   return __a + __b;
 747 }
 748
 749 __extension__ extern __inline float32x4_t
 750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 751 vaddq_f32 (float32x4_t __a, float32x4_t __b)
 752 {
 753   return __a + __b;
 754 }
 755
 756 __extension__ extern __inline float64x2_t
 757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 758 vaddq_f64 (float64x2_t __a, float64x2_t __b)
 759 {
 760   return __a + __b;
 761 }
 762
 763 __extension__ extern __inline uint8x16_t
 764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 765 vaddq_u8 (uint8x16_t __a, uint8x16_t __b)
 766 {
 767   return __a + __b;
 768 }
 769
 770 __extension__ extern __inline uint16x8_t
 771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 772 vaddq_u16 (uint16x8_t __a, uint16x8_t __b)
 773 {
 774   return __a + __b;
 775 }
 776
 777 __extension__ extern __inline uint32x4_t
 778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 779 vaddq_u32 (uint32x4_t __a, uint32x4_t __b)
 780 {
 781   return __a + __b;
 782 }
 783
 784 __extension__ extern __inline uint64x2_t
 785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 786 vaddq_u64 (uint64x2_t __a, uint64x2_t __b)
 787 {
 788   return __a + __b;
 789 }
 790
 791 __extension__ extern __inline int16x8_t
 792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 793 vaddl_s8 (int8x8_t __a, int8x8_t __b)
 794 {
 795   return (int16x8_t) __builtin_aarch64_saddlv8qi (__a, __b);
 796 }
 797
 798 __extension__ extern __inline int32x4_t
 799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 800 vaddl_s16 (int16x4_t __a, int16x4_t __b)
 801 {
 802   return (int32x4_t) __builtin_aarch64_saddlv4hi (__a, __b);
 803 }
 804
 805 __extension__ extern __inline int64x2_t
 806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 807 vaddl_s32 (int32x2_t __a, int32x2_t __b)
 808 {
 809   return (int64x2_t) __builtin_aarch64_saddlv2si (__a, __b);
 810 }
 811
 812 __extension__ extern __inline uint16x8_t
 813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 814 vaddl_u8 (uint8x8_t __a, uint8x8_t __b)
 815 {
 816   return (uint16x8_t) __builtin_aarch64_uaddlv8qi ((int8x8_t) __a,
 817                                                    (int8x8_t) __b);
 818 }
 819
 820 __extension__ extern __inline uint32x4_t
 821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 822 vaddl_u16 (uint16x4_t __a, uint16x4_t __b)
 823 {
 824   return (uint32x4_t) __builtin_aarch64_uaddlv4hi ((int16x4_t) __a,
 825                                                    (int16x4_t) __b);
 826 }
 827
 828 __extension__ extern __inline uint64x2_t
 829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 830 vaddl_u32 (uint32x2_t __a, uint32x2_t __b)
 831 {
 832   return (uint64x2_t) __builtin_aarch64_uaddlv2si ((int32x2_t) __a,
 833                                                    (int32x2_t) __b);
 834 }
 835
 836 __extension__ extern __inline int16x8_t
 837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 838 vaddl_high_s8 (int8x16_t __a, int8x16_t __b)
 839 {
 840   return (int16x8_t) __builtin_aarch64_saddl2v16qi (__a, __b);
 841 }
 842
 843 __extension__ extern __inline int32x4_t
 844 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 845 vaddl_high_s16 (int16x8_t __a, int16x8_t __b)
 846 {
 847   return (int32x4_t) __builtin_aarch64_saddl2v8hi (__a, __b);
 848 }
 849
 850 __extension__ extern __inline int64x2_t
 851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 852 vaddl_high_s32 (int32x4_t __a, int32x4_t __b)
 853 {
 854   return (int64x2_t) __builtin_aarch64_saddl2v4si (__a, __b);
 855 }
 856
 857 __extension__ extern __inline uint16x8_t
 858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 859 vaddl_high_u8 (uint8x16_t __a, uint8x16_t __b)
 860 {
 861   return (uint16x8_t) __builtin_aarch64_uaddl2v16qi ((int8x16_t) __a,
 862                                                      (int8x16_t) __b);
 863 }
 864
 865 __extension__ extern __inline uint32x4_t
 866 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 867 vaddl_high_u16 (uint16x8_t __a, uint16x8_t __b)
 868 {
 869   return (uint32x4_t) __builtin_aarch64_uaddl2v8hi ((int16x8_t) __a,
 870                                                     (int16x8_t) __b);
 871 }
 872
 873 __extension__ extern __inline uint64x2_t
 874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 875 vaddl_high_u32 (uint32x4_t __a, uint32x4_t __b)
 876 {
 877   return (uint64x2_t) __builtin_aarch64_uaddl2v4si ((int32x4_t) __a,
 878                                                     (int32x4_t) __b);
 879 }
 880
 881 __extension__ extern __inline int16x8_t
 882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 883 vaddw_s8 (int16x8_t __a, int8x8_t __b)
 884 {
 885   return (int16x8_t) __builtin_aarch64_saddwv8qi (__a, __b);
 886 }
 887
 888 __extension__ extern __inline int32x4_t
 889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 890 vaddw_s16 (int32x4_t __a, int16x4_t __b)
 891 {
 892   return (int32x4_t) __builtin_aarch64_saddwv4hi (__a, __b);
 893 }
 894
 895 __extension__ extern __inline int64x2_t
 896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 897 vaddw_s32 (int64x2_t __a, int32x2_t __b)
 898 {
 899   return (int64x2_t) __builtin_aarch64_saddwv2si (__a, __b);
 900 }
 901
 902 __extension__ extern __inline uint16x8_t
 903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 904 vaddw_u8 (uint16x8_t __a, uint8x8_t __b)
 905 {
 906   return (uint16x8_t) __builtin_aarch64_uaddwv8qi ((int16x8_t) __a,
 907                                                    (int8x8_t) __b);
 908 }
 909
 910 __extension__ extern __inline uint32x4_t
 911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 912 vaddw_u16 (uint32x4_t __a, uint16x4_t __b)
 913 {
 914   return (uint32x4_t) __builtin_aarch64_uaddwv4hi ((int32x4_t) __a,
 915                                                    (int16x4_t) __b);
 916 }
 917
 918 __extension__ extern __inline uint64x2_t
 919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 920 vaddw_u32 (uint64x2_t __a, uint32x2_t __b)
 921 {
 922   return (uint64x2_t) __builtin_aarch64_uaddwv2si ((int64x2_t) __a,
 923                                                    (int32x2_t) __b);
 924 }
 925
 926 __extension__ extern __inline int16x8_t
 927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 928 vaddw_high_s8 (int16x8_t __a, int8x16_t __b)
 929 {
 930   return (int16x8_t) __builtin_aarch64_saddw2v16qi (__a, __b);
 931 }
 932
 933 __extension__ extern __inline int32x4_t
 934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 935 vaddw_high_s16 (int32x4_t __a, int16x8_t __b)
 936 {
 937   return (int32x4_t) __builtin_aarch64_saddw2v8hi (__a, __b);
 938 }
 939
 940 __extension__ extern __inline int64x2_t
 941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 942 vaddw_high_s32 (int64x2_t __a, int32x4_t __b)
 943 {
 944   return (int64x2_t) __builtin_aarch64_saddw2v4si (__a, __b);
 945 }
 946
 947 __extension__ extern __inline uint16x8_t
 948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 949 vaddw_high_u8 (uint16x8_t __a, uint8x16_t __b)
 950 {
 951   return (uint16x8_t) __builtin_aarch64_uaddw2v16qi ((int16x8_t) __a,
 952                                                      (int8x16_t) __b);
 953 }
 954
 955 __extension__ extern __inline uint32x4_t
 956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 957 vaddw_high_u16 (uint32x4_t __a, uint16x8_t __b)
 958 {
 959   return (uint32x4_t) __builtin_aarch64_uaddw2v8hi ((int32x4_t) __a,
 960                                                     (int16x8_t) __b);
 961 }
 962
 963 __extension__ extern __inline uint64x2_t
 964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 965 vaddw_high_u32 (uint64x2_t __a, uint32x4_t __b)
 966 {
 967   return (uint64x2_t) __builtin_aarch64_uaddw2v4si ((int64x2_t) __a,
 968                                                     (int32x4_t) __b);
 969 }
 970
 971 __extension__ extern __inline int8x8_t
 972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 973 vhadd_s8 (int8x8_t __a, int8x8_t __b)
 974 {
 975   return (int8x8_t) __builtin_aarch64_shaddv8qi (__a, __b);
 976 }
 977
 978 __extension__ extern __inline int16x4_t
 979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 980 vhadd_s16 (int16x4_t __a, int16x4_t __b)
 981 {
 982   return (int16x4_t) __builtin_aarch64_shaddv4hi (__a, __b);
 983 }
 984
 985 __extension__ extern __inline int32x2_t
 986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 987 vhadd_s32 (int32x2_t __a, int32x2_t __b)
 988 {
 989   return (int32x2_t) __builtin_aarch64_shaddv2si (__a, __b);
 990 }
 991
 992 __extension__ extern __inline uint8x8_t
 993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 994 vhadd_u8 (uint8x8_t __a, uint8x8_t __b)
 995 {
 996   return (uint8x8_t) __builtin_aarch64_uhaddv8qi ((int8x8_t) __a,
 997                                                   (int8x8_t) __b);
 998 }
 999
1000 __extension__ extern __inline uint16x4_t
1001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1002 vhadd_u16 (uint16x4_t __a, uint16x4_t __b)
1003 {
1004   return (uint16x4_t) __builtin_aarch64_uhaddv4hi ((int16x4_t) __a,
1005                                                    (int16x4_t) __b);
1006 }
1007
1008 __extension__ extern __inline uint32x2_t
1009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1010 vhadd_u32 (uint32x2_t __a, uint32x2_t __b)
1011 {
1012   return (uint32x2_t) __builtin_aarch64_uhaddv2si ((int32x2_t) __a,
1013                                                    (int32x2_t) __b);
1014 }
1015
1016 __extension__ extern __inline int8x16_t
1017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1018 vhaddq_s8 (int8x16_t __a, int8x16_t __b)
1019 {
1020   return (int8x16_t) __builtin_aarch64_shaddv16qi (__a, __b);
1021 }
1022
1023 __extension__ extern __inline int16x8_t
1024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1025 vhaddq_s16 (int16x8_t __a, int16x8_t __b)
1026 {
1027   return (int16x8_t) __builtin_aarch64_shaddv8hi (__a, __b);
1028 }
1029
1030 __extension__ extern __inline int32x4_t
1031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1032 vhaddq_s32 (int32x4_t __a, int32x4_t __b)
1033 {
1034   return (int32x4_t) __builtin_aarch64_shaddv4si (__a, __b);
1035 }
1036
1037 __extension__ extern __inline uint8x16_t
1038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1039 vhaddq_u8 (uint8x16_t __a, uint8x16_t __b)
1040 {
1041   return (uint8x16_t) __builtin_aarch64_uhaddv16qi ((int8x16_t) __a,
1042                                                     (int8x16_t) __b);
1043 }
1044
1045 __extension__ extern __inline uint16x8_t
1046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1047 vhaddq_u16 (uint16x8_t __a, uint16x8_t __b)
1048 {
1049   return (uint16x8_t) __builtin_aarch64_uhaddv8hi ((int16x8_t) __a,
1050                                                    (int16x8_t) __b);
1051 }
1052
1053 __extension__ extern __inline uint32x4_t
1054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1055 vhaddq_u32 (uint32x4_t __a, uint32x4_t __b)
1056 {
1057   return (uint32x4_t) __builtin_aarch64_uhaddv4si ((int32x4_t) __a,
1058                                                    (int32x4_t) __b);
1059 }
1060
1061 __extension__ extern __inline int8x8_t
1062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1063 vrhadd_s8 (int8x8_t __a, int8x8_t __b)
1064 {
1065   return (int8x8_t) __builtin_aarch64_srhaddv8qi (__a, __b);
1066 }
1067
1068 __extension__ extern __inline int16x4_t
1069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1070 vrhadd_s16 (int16x4_t __a, int16x4_t __b)
1071 {
1072   return (int16x4_t) __builtin_aarch64_srhaddv4hi (__a, __b);
1073 }
1074
1075 __extension__ extern __inline int32x2_t
1076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1077 vrhadd_s32 (int32x2_t __a, int32x2_t __b)
1078 {
1079   return (int32x2_t) __builtin_aarch64_srhaddv2si (__a, __b);
1080 }
1081
1082 __extension__ extern __inline uint8x8_t
1083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1084 vrhadd_u8 (uint8x8_t __a, uint8x8_t __b)
1085 {
1086   return (uint8x8_t) __builtin_aarch64_urhaddv8qi ((int8x8_t) __a,
1087                                                    (int8x8_t) __b);
1088 }
1089
1090 __extension__ extern __inline uint16x4_t
1091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1092 vrhadd_u16 (uint16x4_t __a, uint16x4_t __b)
1093 {
1094   return (uint16x4_t) __builtin_aarch64_urhaddv4hi ((int16x4_t) __a,
1095                                                     (int16x4_t) __b);
1096 }
1097
1098 __extension__ extern __inline uint32x2_t
1099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1100 vrhadd_u32 (uint32x2_t __a, uint32x2_t __b)
1101 {
1102   return (uint32x2_t) __builtin_aarch64_urhaddv2si ((int32x2_t) __a,
1103                                                     (int32x2_t) __b);
1104 }
1105
1106 __extension__ extern __inline int8x16_t
1107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1108 vrhaddq_s8 (int8x16_t __a, int8x16_t __b)
1109 {
1110   return (int8x16_t) __builtin_aarch64_srhaddv16qi (__a, __b);
1111 }
1112
1113 __extension__ extern __inline int16x8_t
1114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1115 vrhaddq_s16 (int16x8_t __a, int16x8_t __b)
1116 {
1117   return (int16x8_t) __builtin_aarch64_srhaddv8hi (__a, __b);
1118 }
1119
1120 __extension__ extern __inline int32x4_t
1121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1122 vrhaddq_s32 (int32x4_t __a, int32x4_t __b)
1123 {
1124   return (int32x4_t) __builtin_aarch64_srhaddv4si (__a, __b);
1125 }
1126
1127 __extension__ extern __inline uint8x16_t
1128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1129 vrhaddq_u8 (uint8x16_t __a, uint8x16_t __b)
1130 {
1131   return (uint8x16_t) __builtin_aarch64_urhaddv16qi ((int8x16_t) __a,
1132                                                      (int8x16_t) __b);
1133 }
1134
1135 __extension__ extern __inline uint16x8_t
1136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1137 vrhaddq_u16 (uint16x8_t __a, uint16x8_t __b)
1138 {
1139   return (uint16x8_t) __builtin_aarch64_urhaddv8hi ((int16x8_t) __a,
1140                                                     (int16x8_t) __b);
1141 }
1142
1143 __extension__ extern __inline uint32x4_t
1144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1145 vrhaddq_u32 (uint32x4_t __a, uint32x4_t __b)
1146 {
1147   return (uint32x4_t) __builtin_aarch64_urhaddv4si ((int32x4_t) __a,
1148                                                     (int32x4_t) __b);
1149 }
1150
1151 __extension__ extern __inline int8x8_t
1152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1153 vaddhn_s16 (int16x8_t __a, int16x8_t __b)
1154 {
1155   return (int8x8_t) __builtin_aarch64_addhnv8hi (__a, __b);
1156 }
1157
1158 __extension__ extern __inline int16x4_t
1159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1160 vaddhn_s32 (int32x4_t __a, int32x4_t __b)
1161 {
1162   return (int16x4_t) __builtin_aarch64_addhnv4si (__a, __b);
1163 }
1164
1165 __extension__ extern __inline int32x2_t
1166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1167 vaddhn_s64 (int64x2_t __a, int64x2_t __b)
1168 {
1169   return (int32x2_t) __builtin_aarch64_addhnv2di (__a, __b);
1170 }
1171
1172 __extension__ extern __inline uint8x8_t
1173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1174 vaddhn_u16 (uint16x8_t __a, uint16x8_t __b)
1175 {
1176   return (uint8x8_t) __builtin_aarch64_addhnv8hi ((int16x8_t) __a,
1177                                                   (int16x8_t) __b);
1178 }
1179
1180 __extension__ extern __inline uint16x4_t
1181 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1182 vaddhn_u32 (uint32x4_t __a, uint32x4_t __b)
1183 {
1184   return (uint16x4_t) __builtin_aarch64_addhnv4si ((int32x4_t) __a,
1185                                                    (int32x4_t) __b);
1186 }
1187
1188 __extension__ extern __inline uint32x2_t
1189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1190 vaddhn_u64 (uint64x2_t __a, uint64x2_t __b)
1191 {
1192   return (uint32x2_t) __builtin_aarch64_addhnv2di ((int64x2_t) __a,
1193                                                    (int64x2_t) __b);
1194 }
1195
1196 __extension__ extern __inline int8x8_t
1197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1198 vraddhn_s16 (int16x8_t __a, int16x8_t __b)
1199 {
1200   return (int8x8_t) __builtin_aarch64_raddhnv8hi (__a, __b);
1201 }
1202
1203 __extension__ extern __inline int16x4_t
1204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1205 vraddhn_s32 (int32x4_t __a, int32x4_t __b)
1206 {
1207   return (int16x4_t) __builtin_aarch64_raddhnv4si (__a, __b);
1208 }
1209
1210 __extension__ extern __inline int32x2_t
1211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1212 vraddhn_s64 (int64x2_t __a, int64x2_t __b)
1213 {
1214   return (int32x2_t) __builtin_aarch64_raddhnv2di (__a, __b);
1215 }
1216
1217 __extension__ extern __inline uint8x8_t
1218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1219 vraddhn_u16 (uint16x8_t __a, uint16x8_t __b)
1220 {
1221   return (uint8x8_t) __builtin_aarch64_raddhnv8hi ((int16x8_t) __a,
1222                                                    (int16x8_t) __b);
1223 }
1224
1225 __extension__ extern __inline uint16x4_t
1226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1227 vraddhn_u32 (uint32x4_t __a, uint32x4_t __b)
1228 {
1229   return (uint16x4_t) __builtin_aarch64_raddhnv4si ((int32x4_t) __a,
1230                                                     (int32x4_t) __b);
1231 }
1232
1233 __extension__ extern __inline uint32x2_t
1234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1235 vraddhn_u64 (uint64x2_t __a, uint64x2_t __b)
1236 {
1237   return (uint32x2_t) __builtin_aarch64_raddhnv2di ((int64x2_t) __a,
1238                                                     (int64x2_t) __b);
1239 }
1240
1241 __extension__ extern __inline int8x16_t
1242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1243 vaddhn_high_s16 (int8x8_t __a, int16x8_t __b, int16x8_t __c)
1244 {
1245   return (int8x16_t) __builtin_aarch64_addhn2v8hi (__a, __b, __c);
1246 }
1247
1248 __extension__ extern __inline int16x8_t
1249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1250 vaddhn_high_s32 (int16x4_t __a, int32x4_t __b, int32x4_t __c)
1251 {
1252   return (int16x8_t) __builtin_aarch64_addhn2v4si (__a, __b, __c);
1253 }
1254
1255 __extension__ extern __inline int32x4_t
1256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1257 vaddhn_high_s64 (int32x2_t __a, int64x2_t __b, int64x2_t __c)
1258 {
1259   return (int32x4_t) __builtin_aarch64_addhn2v2di (__a, __b, __c);
1260 }
1261
1262 __extension__ extern __inline uint8x16_t
1263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1264 vaddhn_high_u16 (uint8x8_t __a, uint16x8_t __b, uint16x8_t __c)
1265 {
1266   return (uint8x16_t) __builtin_aarch64_addhn2v8hi ((int8x8_t) __a,
1267                                                     (int16x8_t) __b,
1268                                                     (int16x8_t) __c);
1269 }
1270
1271 __extension__ extern __inline uint16x8_t
1272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1273 vaddhn_high_u32 (uint16x4_t __a, uint32x4_t __b, uint32x4_t __c)
1274 {
1275   return (uint16x8_t) __builtin_aarch64_addhn2v4si ((int16x4_t) __a,
1276                                                     (int32x4_t) __b,
1277                                                     (int32x4_t) __c);
1278 }
1279
1280 __extension__ extern __inline uint32x4_t
1281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1282 vaddhn_high_u64 (uint32x2_t __a, uint64x2_t __b, uint64x2_t __c)
1283 {
1284   return (uint32x4_t) __builtin_aarch64_addhn2v2di ((int32x2_t) __a,
1285                                                     (int64x2_t) __b,
1286                                                     (int64x2_t) __c);
1287 }
1288
1289 __extension__ extern __inline int8x16_t
1290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1291 vraddhn_high_s16 (int8x8_t __a, int16x8_t __b, int16x8_t __c)
1292 {
1293   return (int8x16_t) __builtin_aarch64_raddhn2v8hi (__a, __b, __c);
1294 }
1295
1296 __extension__ extern __inline int16x8_t
1297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1298 vraddhn_high_s32 (int16x4_t __a, int32x4_t __b, int32x4_t __c)
1299 {
1300   return (int16x8_t) __builtin_aarch64_raddhn2v4si (__a, __b, __c);
1301 }
1302
1303 __extension__ extern __inline int32x4_t
1304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1305 vraddhn_high_s64 (int32x2_t __a, int64x2_t __b, int64x2_t __c)
1306 {
1307   return (int32x4_t) __builtin_aarch64_raddhn2v2di (__a, __b, __c);
1308 }
1309
1310 __extension__ extern __inline uint8x16_t
1311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1312 vraddhn_high_u16 (uint8x8_t __a, uint16x8_t __b, uint16x8_t __c)
1313 {
1314   return (uint8x16_t) __builtin_aarch64_raddhn2v8hi ((int8x8_t) __a,
1315                                                      (int16x8_t) __b,
1316                                                      (int16x8_t) __c);
1317 }
1318
1319 __extension__ extern __inline uint16x8_t
1320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1321 vraddhn_high_u32 (uint16x4_t __a, uint32x4_t __b, uint32x4_t __c)
1322 {
1323   return (uint16x8_t) __builtin_aarch64_raddhn2v4si ((int16x4_t) __a,
1324                                                      (int32x4_t) __b,
1325                                                      (int32x4_t) __c);
1326 }
1327
1328 __extension__ extern __inline uint32x4_t
1329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1330 vraddhn_high_u64 (uint32x2_t __a, uint64x2_t __b, uint64x2_t __c)
1331 {
1332   return (uint32x4_t) __builtin_aarch64_raddhn2v2di ((int32x2_t) __a,
1333                                                      (int64x2_t) __b,
1334                                                      (int64x2_t) __c);
1335 }
1336
1337 __extension__ extern __inline float32x2_t
1338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1339 vdiv_f32 (float32x2_t __a, float32x2_t __b)
1340 {
1341   return __a / __b;
1342 }
1343
1344 __extension__ extern __inline float64x1_t
1345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1346 vdiv_f64 (float64x1_t __a, float64x1_t __b)
1347 {
1348   return __a / __b;
1349 }
1350
1351 __extension__ extern __inline float32x4_t
1352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1353 vdivq_f32 (float32x4_t __a, float32x4_t __b)
1354 {
1355   return __a / __b;
1356 }
1357
1358 __extension__ extern __inline float64x2_t
1359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1360 vdivq_f64 (float64x2_t __a, float64x2_t __b)
1361 {
1362   return __a / __b;
1363 }
1364
1365 __extension__ extern __inline int8x8_t
1366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1367 vmul_s8 (int8x8_t __a, int8x8_t __b)
1368 {
1369   return __a * __b;
1370 }
1371
1372 __extension__ extern __inline int16x4_t
1373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1374 vmul_s16 (int16x4_t __a, int16x4_t __b)
1375 {
1376   return __a * __b;
1377 }
1378
1379 __extension__ extern __inline int32x2_t
1380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1381 vmul_s32 (int32x2_t __a, int32x2_t __b)
1382 {
1383   return __a * __b;
1384 }
1385
1386 __extension__ extern __inline float32x2_t
1387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1388 vmul_f32 (float32x2_t __a, float32x2_t __b)
1389 {
1390   return __a * __b;
1391 }
1392
1393 __extension__ extern __inline float64x1_t
1394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1395 vmul_f64 (float64x1_t __a, float64x1_t __b)
1396 {
1397   return __a * __b;
1398 }
1399
1400 __extension__ extern __inline uint8x8_t
1401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1402 vmul_u8 (uint8x8_t __a, uint8x8_t __b)
1403 {
1404   return __a * __b;
1405 }
1406
1407 __extension__ extern __inline uint16x4_t
1408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1409 vmul_u16 (uint16x4_t __a, uint16x4_t __b)
1410 {
1411   return __a * __b;
1412 }
1413
1414 __extension__ extern __inline uint32x2_t
1415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1416 vmul_u32 (uint32x2_t __a, uint32x2_t __b)
1417 {
1418   return __a * __b;
1419 }
1420
1421 __extension__ extern __inline poly8x8_t
1422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1423 vmul_p8 (poly8x8_t __a, poly8x8_t __b)
1424 {
1425   return (poly8x8_t) __builtin_aarch64_pmulv8qi ((int8x8_t) __a,
1426                                                  (int8x8_t) __b);
1427 }
1428
1429 __extension__ extern __inline int8x16_t
1430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1431 vmulq_s8 (int8x16_t __a, int8x16_t __b)
1432 {
1433   return __a * __b;
1434 }
1435
1436 __extension__ extern __inline int16x8_t
1437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1438 vmulq_s16 (int16x8_t __a, int16x8_t __b)
1439 {
1440   return __a * __b;
1441 }
1442
1443 __extension__ extern __inline int32x4_t
1444 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1445 vmulq_s32 (int32x4_t __a, int32x4_t __b)
1446 {
1447   return __a * __b;
1448 }
1449
1450 __extension__ extern __inline float32x4_t
1451 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1452 vmulq_f32 (float32x4_t __a, float32x4_t __b)
1453 {
1454   return __a * __b;
1455 }
1456
1457 __extension__ extern __inline float64x2_t
1458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1459 vmulq_f64 (float64x2_t __a, float64x2_t __b)
1460 {
1461   return __a * __b;
1462 }
1463
1464 __extension__ extern __inline uint8x16_t
1465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1466 vmulq_u8 (uint8x16_t __a, uint8x16_t __b)
1467 {
1468   return __a * __b;
1469 }
1470
1471 __extension__ extern __inline uint16x8_t
1472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1473 vmulq_u16 (uint16x8_t __a, uint16x8_t __b)
1474 {
1475   return __a * __b;
1476 }
1477
1478 __extension__ extern __inline uint32x4_t
1479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1480 vmulq_u32 (uint32x4_t __a, uint32x4_t __b)
1481 {
1482   return __a * __b;
1483 }
1484
1485 __extension__ extern __inline poly8x16_t
1486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1487 vmulq_p8 (poly8x16_t __a, poly8x16_t __b)
1488 {
1489   return (poly8x16_t) __builtin_aarch64_pmulv16qi ((int8x16_t) __a,
1490                                                    (int8x16_t) __b);
1491 }
1492
1493 __extension__ extern __inline int8x8_t
1494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1495 vand_s8 (int8x8_t __a, int8x8_t __b)
1496 {
1497   return __a & __b;
1498 }
1499
1500 __extension__ extern __inline int16x4_t
1501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1502 vand_s16 (int16x4_t __a, int16x4_t __b)
1503 {
1504   return __a & __b;
1505 }
1506
1507 __extension__ extern __inline int32x2_t
1508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1509 vand_s32 (int32x2_t __a, int32x2_t __b)
1510 {
1511   return __a & __b;
1512 }
1513
1514 __extension__ extern __inline uint8x8_t
1515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1516 vand_u8 (uint8x8_t __a, uint8x8_t __b)
1517 {
1518   return __a & __b;
1519 }
1520
1521 __extension__ extern __inline uint16x4_t
1522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1523 vand_u16 (uint16x4_t __a, uint16x4_t __b)
1524 {
1525   return __a & __b;
1526 }
1527
1528 __extension__ extern __inline uint32x2_t
1529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1530 vand_u32 (uint32x2_t __a, uint32x2_t __b)
1531 {
1532   return __a & __b;
1533 }
1534
1535 __extension__ extern __inline int64x1_t
1536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1537 vand_s64 (int64x1_t __a, int64x1_t __b)
1538 {
1539   return __a & __b;
1540 }
1541
1542 __extension__ extern __inline uint64x1_t
1543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1544 vand_u64 (uint64x1_t __a, uint64x1_t __b)
1545 {
1546   return __a & __b;
1547 }
1548
1549 __extension__ extern __inline int8x16_t
1550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1551 vandq_s8 (int8x16_t __a, int8x16_t __b)
1552 {
1553   return __a & __b;
1554 }
1555
1556 __extension__ extern __inline int16x8_t
1557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1558 vandq_s16 (int16x8_t __a, int16x8_t __b)
1559 {
1560   return __a & __b;
1561 }
1562
1563 __extension__ extern __inline int32x4_t
1564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1565 vandq_s32 (int32x4_t __a, int32x4_t __b)
1566 {
1567   return __a & __b;
1568 }
1569
1570 __extension__ extern __inline int64x2_t
1571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1572 vandq_s64 (int64x2_t __a, int64x2_t __b)
1573 {
1574   return __a & __b;
1575 }
1576
1577 __extension__ extern __inline uint8x16_t
1578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1579 vandq_u8 (uint8x16_t __a, uint8x16_t __b)
1580 {
1581   return __a & __b;
1582 }
1583
1584 __extension__ extern __inline uint16x8_t
1585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1586 vandq_u16 (uint16x8_t __a, uint16x8_t __b)
1587 {
1588   return __a & __b;
1589 }
1590
1591 __extension__ extern __inline uint32x4_t
1592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1593 vandq_u32 (uint32x4_t __a, uint32x4_t __b)
1594 {
1595   return __a & __b;
1596 }
1597
1598 __extension__ extern __inline uint64x2_t
1599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1600 vandq_u64 (uint64x2_t __a, uint64x2_t __b)
1601 {
1602   return __a & __b;
1603 }
1604
1605 __extension__ extern __inline int8x8_t
1606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1607 vorr_s8 (int8x8_t __a, int8x8_t __b)
1608 {
1609   return __a | __b;
1610 }
1611
1612 __extension__ extern __inline int16x4_t
1613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1614 vorr_s16 (int16x4_t __a, int16x4_t __b)
1615 {
1616   return __a | __b;
1617 }
1618
1619 __extension__ extern __inline int32x2_t
1620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1621 vorr_s32 (int32x2_t __a, int32x2_t __b)
1622 {
1623   return __a | __b;
1624 }
1625
1626 __extension__ extern __inline uint8x8_t
1627 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1628 vorr_u8 (uint8x8_t __a, uint8x8_t __b)
1629 {
1630   return __a | __b;
1631 }
1632
1633 __extension__ extern __inline uint16x4_t
1634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1635 vorr_u16 (uint16x4_t __a, uint16x4_t __b)
1636 {
1637   return __a | __b;
1638 }
1639
1640 __extension__ extern __inline uint32x2_t
1641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1642 vorr_u32 (uint32x2_t __a, uint32x2_t __b)
1643 {
1644   return __a | __b;
1645 }
1646
1647 __extension__ extern __inline int64x1_t
1648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1649 vorr_s64 (int64x1_t __a, int64x1_t __b)
1650 {
1651   return __a | __b;
1652 }
1653
1654 __extension__ extern __inline uint64x1_t
1655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1656 vorr_u64 (uint64x1_t __a, uint64x1_t __b)
1657 {
1658   return __a | __b;
1659 }
1660
1661 __extension__ extern __inline int8x16_t
1662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1663 vorrq_s8 (int8x16_t __a, int8x16_t __b)
1664 {
1665   return __a | __b;
1666 }
1667
1668 __extension__ extern __inline int16x8_t
1669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1670 vorrq_s16 (int16x8_t __a, int16x8_t __b)
1671 {
1672   return __a | __b;
1673 }
1674
1675 __extension__ extern __inline int32x4_t
1676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1677 vorrq_s32 (int32x4_t __a, int32x4_t __b)
1678 {
1679   return __a | __b;
1680 }
1681
1682 __extension__ extern __inline int64x2_t
1683 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1684 vorrq_s64 (int64x2_t __a, int64x2_t __b)
1685 {
1686   return __a | __b;
1687 }
1688
1689 __extension__ extern __inline uint8x16_t
1690 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1691 vorrq_u8 (uint8x16_t __a, uint8x16_t __b)
1692 {
1693   return __a | __b;
1694 }
1695
1696 __extension__ extern __inline uint16x8_t
1697 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1698 vorrq_u16 (uint16x8_t __a, uint16x8_t __b)
1699 {
1700   return __a | __b;
1701 }
1702
1703 __extension__ extern __inline uint32x4_t
1704 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1705 vorrq_u32 (uint32x4_t __a, uint32x4_t __b)
1706 {
1707   return __a | __b;
1708 }
1709
1710 __extension__ extern __inline uint64x2_t
1711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1712 vorrq_u64 (uint64x2_t __a, uint64x2_t __b)
1713 {
1714   return __a | __b;
1715 }
1716
1717 __extension__ extern __inline int8x8_t
1718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1719 veor_s8 (int8x8_t __a, int8x8_t __b)
1720 {
1721   return __a ^ __b;
1722 }
1723
1724 __extension__ extern __inline int16x4_t
1725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1726 veor_s16 (int16x4_t __a, int16x4_t __b)
1727 {
1728   return __a ^ __b;
1729 }
1730
1731 __extension__ extern __inline int32x2_t
1732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1733 veor_s32 (int32x2_t __a, int32x2_t __b)
1734 {
1735   return __a ^ __b;
1736 }
1737
1738 __extension__ extern __inline uint8x8_t
1739 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1740 veor_u8 (uint8x8_t __a, uint8x8_t __b)
1741 {
1742   return __a ^ __b;
1743 }
1744
1745 __extension__ extern __inline uint16x4_t
1746 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1747 veor_u16 (uint16x4_t __a, uint16x4_t __b)
1748 {
1749   return __a ^ __b;
1750 }
1751
1752 __extension__ extern __inline uint32x2_t
1753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1754 veor_u32 (uint32x2_t __a, uint32x2_t __b)
1755 {
1756   return __a ^ __b;
1757 }
1758
1759 __extension__ extern __inline int64x1_t
1760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1761 veor_s64 (int64x1_t __a, int64x1_t __b)
1762 {
1763   return __a ^ __b;
1764 }
1765
1766 __extension__ extern __inline uint64x1_t
1767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1768 veor_u64 (uint64x1_t __a, uint64x1_t __b)
1769 {
1770   return __a ^ __b;
1771 }
1772
1773 __extension__ extern __inline int8x16_t
1774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1775 veorq_s8 (int8x16_t __a, int8x16_t __b)
1776 {
1777   return __a ^ __b;
1778 }
1779
1780 __extension__ extern __inline int16x8_t
1781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1782 veorq_s16 (int16x8_t __a, int16x8_t __b)
1783 {
1784   return __a ^ __b;
1785 }
1786
1787 __extension__ extern __inline int32x4_t
1788 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1789 veorq_s32 (int32x4_t __a, int32x4_t __b)
1790 {
1791   return __a ^ __b;
1792 }
1793
1794 __extension__ extern __inline int64x2_t
1795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1796 veorq_s64 (int64x2_t __a, int64x2_t __b)
1797 {
1798   return __a ^ __b;
1799 }
1800
1801 __extension__ extern __inline uint8x16_t
1802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1803 veorq_u8 (uint8x16_t __a, uint8x16_t __b)
1804 {
1805   return __a ^ __b;
1806 }
1807
1808 __extension__ extern __inline uint16x8_t
1809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1810 veorq_u16 (uint16x8_t __a, uint16x8_t __b)
1811 {
1812   return __a ^ __b;
1813 }
1814
1815 __extension__ extern __inline uint32x4_t
1816 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1817 veorq_u32 (uint32x4_t __a, uint32x4_t __b)
1818 {
1819   return __a ^ __b;
1820 }
1821
1822 __extension__ extern __inline uint64x2_t
1823 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1824 veorq_u64 (uint64x2_t __a, uint64x2_t __b)
1825 {
1826   return __a ^ __b;
1827 }
1828
1829 __extension__ extern __inline int8x8_t
1830 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1831 vbic_s8 (int8x8_t __a, int8x8_t __b)
1832 {
1833   return __a & ~__b;
1834 }
1835
1836 __extension__ extern __inline int16x4_t
1837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1838 vbic_s16 (int16x4_t __a, int16x4_t __b)
1839 {
1840   return __a & ~__b;
1841 }
1842
1843 __extension__ extern __inline int32x2_t
1844 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1845 vbic_s32 (int32x2_t __a, int32x2_t __b)
1846 {
1847   return __a & ~__b;
1848 }
1849
1850 __extension__ extern __inline uint8x8_t
1851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1852 vbic_u8 (uint8x8_t __a, uint8x8_t __b)
1853 {
1854   return __a & ~__b;
1855 }
1856
1857 __extension__ extern __inline uint16x4_t
1858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1859 vbic_u16 (uint16x4_t __a, uint16x4_t __b)
1860 {
1861   return __a & ~__b;
1862 }
1863
1864 __extension__ extern __inline uint32x2_t
1865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1866 vbic_u32 (uint32x2_t __a, uint32x2_t __b)
1867 {
1868   return __a & ~__b;
1869 }
1870
1871 __extension__ extern __inline int64x1_t
1872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1873 vbic_s64 (int64x1_t __a, int64x1_t __b)
1874 {
1875   return __a & ~__b;
1876 }
1877
1878 __extension__ extern __inline uint64x1_t
1879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1880 vbic_u64 (uint64x1_t __a, uint64x1_t __b)
1881 {
1882   return __a & ~__b;
1883 }
1884
1885 __extension__ extern __inline int8x16_t
1886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1887 vbicq_s8 (int8x16_t __a, int8x16_t __b)
1888 {
1889   return __a & ~__b;
1890 }
1891
1892 __extension__ extern __inline int16x8_t
1893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1894 vbicq_s16 (int16x8_t __a, int16x8_t __b)
1895 {
1896   return __a & ~__b;
1897 }
1898
1899 __extension__ extern __inline int32x4_t
1900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1901 vbicq_s32 (int32x4_t __a, int32x4_t __b)
1902 {
1903   return __a & ~__b;
1904 }
1905
1906 __extension__ extern __inline int64x2_t
1907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1908 vbicq_s64 (int64x2_t __a, int64x2_t __b)
1909 {
1910   return __a & ~__b;
1911 }
1912
1913 __extension__ extern __inline uint8x16_t
1914 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1915 vbicq_u8 (uint8x16_t __a, uint8x16_t __b)
1916 {
1917   return __a & ~__b;
1918 }
1919
1920 __extension__ extern __inline uint16x8_t
1921 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1922 vbicq_u16 (uint16x8_t __a, uint16x8_t __b)
1923 {
1924   return __a & ~__b;
1925 }
1926
1927 __extension__ extern __inline uint32x4_t
1928 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1929 vbicq_u32 (uint32x4_t __a, uint32x4_t __b)
1930 {
1931   return __a & ~__b;
1932 }
1933
1934 __extension__ extern __inline uint64x2_t
1935 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1936 vbicq_u64 (uint64x2_t __a, uint64x2_t __b)
1937 {
1938   return __a & ~__b;
1939 }
1940
1941 __extension__ extern __inline int8x8_t
1942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1943 vorn_s8 (int8x8_t __a, int8x8_t __b)
1944 {
1945   return __a | ~__b;
1946 }
1947
1948 __extension__ extern __inline int16x4_t
1949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1950 vorn_s16 (int16x4_t __a, int16x4_t __b)
1951 {
1952   return __a | ~__b;
1953 }
1954
1955 __extension__ extern __inline int32x2_t
1956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1957 vorn_s32 (int32x2_t __a, int32x2_t __b)
1958 {
1959   return __a | ~__b;
1960 }
1961
1962 __extension__ extern __inline uint8x8_t
1963 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1964 vorn_u8 (uint8x8_t __a, uint8x8_t __b)
1965 {
1966   return __a | ~__b;
1967 }
1968
1969 __extension__ extern __inline uint16x4_t
1970 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1971 vorn_u16 (uint16x4_t __a, uint16x4_t __b)
1972 {
1973   return __a | ~__b;
1974 }
1975
1976 __extension__ extern __inline uint32x2_t
1977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1978 vorn_u32 (uint32x2_t __a, uint32x2_t __b)
1979 {
1980   return __a | ~__b;
1981 }
1982
1983 __extension__ extern __inline int64x1_t
1984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1985 vorn_s64 (int64x1_t __a, int64x1_t __b)
1986 {
1987   return __a | ~__b;
1988 }
1989
1990 __extension__ extern __inline uint64x1_t
1991 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1992 vorn_u64 (uint64x1_t __a, uint64x1_t __b)
1993 {
1994   return __a | ~__b;
1995 }
1996
1997 __extension__ extern __inline int8x16_t
1998 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
1999 vornq_s8 (int8x16_t __a, int8x16_t __b)
2000 {
2001   return __a | ~__b;
2002 }
2003
2004 __extension__ extern __inline int16x8_t
2005 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2006 vornq_s16 (int16x8_t __a, int16x8_t __b)
2007 {
2008   return __a | ~__b;
2009 }
2010
2011 __extension__ extern __inline int32x4_t
2012 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2013 vornq_s32 (int32x4_t __a, int32x4_t __b)
2014 {
2015   return __a | ~__b;
2016 }
2017
2018 __extension__ extern __inline int64x2_t
2019 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2020 vornq_s64 (int64x2_t __a, int64x2_t __b)
2021 {
2022   return __a | ~__b;
2023 }
2024
2025 __extension__ extern __inline uint8x16_t
2026 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2027 vornq_u8 (uint8x16_t __a, uint8x16_t __b)
2028 {
2029   return __a | ~__b;
2030 }
2031
2032 __extension__ extern __inline uint16x8_t
2033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2034 vornq_u16 (uint16x8_t __a, uint16x8_t __b)
2035 {
2036   return __a | ~__b;
2037 }
2038
2039 __extension__ extern __inline uint32x4_t
2040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2041 vornq_u32 (uint32x4_t __a, uint32x4_t __b)
2042 {
2043   return __a | ~__b;
2044 }
2045
2046 __extension__ extern __inline uint64x2_t
2047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2048 vornq_u64 (uint64x2_t __a, uint64x2_t __b)
2049 {
2050   return __a | ~__b;
2051 }
2052
2053 __extension__ extern __inline int8x8_t
2054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2055 vsub_s8 (int8x8_t __a, int8x8_t __b)
2056 {
2057   return __a - __b;
2058 }
2059
2060 __extension__ extern __inline int16x4_t
2061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2062 vsub_s16 (int16x4_t __a, int16x4_t __b)
2063 {
2064   return __a - __b;
2065 }
2066
2067 __extension__ extern __inline int32x2_t
2068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2069 vsub_s32 (int32x2_t __a, int32x2_t __b)
2070 {
2071   return __a - __b;
2072 }
2073
2074 __extension__ extern __inline float32x2_t
2075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2076 vsub_f32 (float32x2_t __a, float32x2_t __b)
2077 {
2078   return __a - __b;
2079 }
2080
2081 __extension__ extern __inline float64x1_t
2082 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2083 vsub_f64 (float64x1_t __a, float64x1_t __b)
2084 {
2085   return __a - __b;
2086 }
2087
2088 __extension__ extern __inline uint8x8_t
2089 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2090 vsub_u8 (uint8x8_t __a, uint8x8_t __b)
2091 {
2092   return __a - __b;
2093 }
2094
2095 __extension__ extern __inline uint16x4_t
2096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2097 vsub_u16 (uint16x4_t __a, uint16x4_t __b)
2098 {
2099   return __a - __b;
2100 }
2101
2102 __extension__ extern __inline uint32x2_t
2103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2104 vsub_u32 (uint32x2_t __a, uint32x2_t __b)
2105 {
2106   return __a - __b;
2107 }
2108
2109 __extension__ extern __inline int64x1_t
2110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2111 vsub_s64 (int64x1_t __a, int64x1_t __b)
2112 {
2113   return __a - __b;
2114 }
2115
2116 __extension__ extern __inline uint64x1_t
2117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2118 vsub_u64 (uint64x1_t __a, uint64x1_t __b)
2119 {
2120   return __a - __b;
2121 }
2122
2123 __extension__ extern __inline int8x16_t
2124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2125 vsubq_s8 (int8x16_t __a, int8x16_t __b)
2126 {
2127   return __a - __b;
2128 }
2129
2130 __extension__ extern __inline int16x8_t
2131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2132 vsubq_s16 (int16x8_t __a, int16x8_t __b)
2133 {
2134   return __a - __b;
2135 }
2136
2137 __extension__ extern __inline int32x4_t
2138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2139 vsubq_s32 (int32x4_t __a, int32x4_t __b)
2140 {
2141   return __a - __b;
2142 }
2143
2144 __extension__ extern __inline int64x2_t
2145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2146 vsubq_s64 (int64x2_t __a, int64x2_t __b)
2147 {
2148   return __a - __b;
2149 }
2150
2151 __extension__ extern __inline float32x4_t
2152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2153 vsubq_f32 (float32x4_t __a, float32x4_t __b)
2154 {
2155   return __a - __b;
2156 }
2157
2158 __extension__ extern __inline float64x2_t
2159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2160 vsubq_f64 (float64x2_t __a, float64x2_t __b)
2161 {
2162   return __a - __b;
2163 }
2164
2165 __extension__ extern __inline uint8x16_t
2166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2167 vsubq_u8 (uint8x16_t __a, uint8x16_t __b)
2168 {
2169   return __a - __b;
2170 }
2171
2172 __extension__ extern __inline uint16x8_t
2173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2174 vsubq_u16 (uint16x8_t __a, uint16x8_t __b)
2175 {
2176   return __a - __b;
2177 }
2178
2179 __extension__ extern __inline uint32x4_t
2180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2181 vsubq_u32 (uint32x4_t __a, uint32x4_t __b)
2182 {
2183   return __a - __b;
2184 }
2185
2186 __extension__ extern __inline uint64x2_t
2187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2188 vsubq_u64 (uint64x2_t __a, uint64x2_t __b)
2189 {
2190   return __a - __b;
2191 }
2192
2193 __extension__ extern __inline int16x8_t
2194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2195 vsubl_s8 (int8x8_t __a, int8x8_t __b)
2196 {
2197   return (int16x8_t) __builtin_aarch64_ssublv8qi (__a, __b);
2198 }
2199
2200 __extension__ extern __inline int32x4_t
2201 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2202 vsubl_s16 (int16x4_t __a, int16x4_t __b)
2203 {
2204   return (int32x4_t) __builtin_aarch64_ssublv4hi (__a, __b);
2205 }
2206
2207 __extension__ extern __inline int64x2_t
2208 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2209 vsubl_s32 (int32x2_t __a, int32x2_t __b)
2210 {
2211   return (int64x2_t) __builtin_aarch64_ssublv2si (__a, __b);
2212 }
2213
2214 __extension__ extern __inline uint16x8_t
2215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2216 vsubl_u8 (uint8x8_t __a, uint8x8_t __b)
2217 {
2218   return (uint16x8_t) __builtin_aarch64_usublv8qi ((int8x8_t) __a,
2219                                                    (int8x8_t) __b);
2220 }
2221
2222 __extension__ extern __inline uint32x4_t
2223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2224 vsubl_u16 (uint16x4_t __a, uint16x4_t __b)
2225 {
2226   return (uint32x4_t) __builtin_aarch64_usublv4hi ((int16x4_t) __a,
2227                                                    (int16x4_t) __b);
2228 }
2229
2230 __extension__ extern __inline uint64x2_t
2231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2232 vsubl_u32 (uint32x2_t __a, uint32x2_t __b)
2233 {
2234   return (uint64x2_t) __builtin_aarch64_usublv2si ((int32x2_t) __a,
2235                                                    (int32x2_t) __b);
2236 }
2237
2238 __extension__ extern __inline int16x8_t
2239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2240 vsubl_high_s8 (int8x16_t __a, int8x16_t __b)
2241 {
2242   return (int16x8_t) __builtin_aarch64_ssubl2v16qi (__a, __b);
2243 }
2244
2245 __extension__ extern __inline int32x4_t
2246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2247 vsubl_high_s16 (int16x8_t __a, int16x8_t __b)
2248 {
2249   return (int32x4_t) __builtin_aarch64_ssubl2v8hi (__a, __b);
2250 }
2251
2252 __extension__ extern __inline int64x2_t
2253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2254 vsubl_high_s32 (int32x4_t __a, int32x4_t __b)
2255 {
2256   return (int64x2_t) __builtin_aarch64_ssubl2v4si (__a, __b);
2257 }
2258
2259 __extension__ extern __inline uint16x8_t
2260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2261 vsubl_high_u8 (uint8x16_t __a, uint8x16_t __b)
2262 {
2263   return (uint16x8_t) __builtin_aarch64_usubl2v16qi ((int8x16_t) __a,
2264                                                      (int8x16_t) __b);
2265 }
2266
2267 __extension__ extern __inline uint32x4_t
2268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2269 vsubl_high_u16 (uint16x8_t __a, uint16x8_t __b)
2270 {
2271   return (uint32x4_t) __builtin_aarch64_usubl2v8hi ((int16x8_t) __a,
2272                                                     (int16x8_t) __b);
2273 }
2274
2275 __extension__ extern __inline uint64x2_t
2276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2277 vsubl_high_u32 (uint32x4_t __a, uint32x4_t __b)
2278 {
2279   return (uint64x2_t) __builtin_aarch64_usubl2v4si ((int32x4_t) __a,
2280                                                     (int32x4_t) __b);
2281 }
2282
2283 __extension__ extern __inline int16x8_t
2284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2285 vsubw_s8 (int16x8_t __a, int8x8_t __b)
2286 {
2287   return (int16x8_t) __builtin_aarch64_ssubwv8qi (__a, __b);
2288 }
2289
2290 __extension__ extern __inline int32x4_t
2291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2292 vsubw_s16 (int32x4_t __a, int16x4_t __b)
2293 {
2294   return (int32x4_t) __builtin_aarch64_ssubwv4hi (__a, __b);
2295 }
2296
2297 __extension__ extern __inline int64x2_t
2298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2299 vsubw_s32 (int64x2_t __a, int32x2_t __b)
2300 {
2301   return (int64x2_t) __builtin_aarch64_ssubwv2si (__a, __b);
2302 }
2303
2304 __extension__ extern __inline uint16x8_t
2305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2306 vsubw_u8 (uint16x8_t __a, uint8x8_t __b)
2307 {
2308   return (uint16x8_t) __builtin_aarch64_usubwv8qi ((int16x8_t) __a,
2309                                                    (int8x8_t) __b);
2310 }
2311
2312 __extension__ extern __inline uint32x4_t
2313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2314 vsubw_u16 (uint32x4_t __a, uint16x4_t __b)
2315 {
2316   return (uint32x4_t) __builtin_aarch64_usubwv4hi ((int32x4_t) __a,
2317                                                    (int16x4_t) __b);
2318 }
2319
2320 __extension__ extern __inline uint64x2_t
2321 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2322 vsubw_u32 (uint64x2_t __a, uint32x2_t __b)
2323 {
2324   return (uint64x2_t) __builtin_aarch64_usubwv2si ((int64x2_t) __a,
2325                                                    (int32x2_t) __b);
2326 }
2327
2328 __extension__ extern __inline int16x8_t
2329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2330 vsubw_high_s8 (int16x8_t __a, int8x16_t __b)
2331 {
2332   return (int16x8_t) __builtin_aarch64_ssubw2v16qi (__a, __b);
2333 }
2334
2335 __extension__ extern __inline int32x4_t
2336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2337 vsubw_high_s16 (int32x4_t __a, int16x8_t __b)
2338 {
2339   return (int32x4_t) __builtin_aarch64_ssubw2v8hi (__a, __b);
2340 }
2341
2342 __extension__ extern __inline int64x2_t
2343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2344 vsubw_high_s32 (int64x2_t __a, int32x4_t __b)
2345 {
2346   return (int64x2_t) __builtin_aarch64_ssubw2v4si (__a, __b);
2347 }
2348
2349 __extension__ extern __inline uint16x8_t
2350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2351 vsubw_high_u8 (uint16x8_t __a, uint8x16_t __b)
2352 {
2353   return (uint16x8_t) __builtin_aarch64_usubw2v16qi ((int16x8_t) __a,
2354                                                      (int8x16_t) __b);
2355 }
2356
2357 __extension__ extern __inline uint32x4_t
2358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2359 vsubw_high_u16 (uint32x4_t __a, uint16x8_t __b)
2360 {
2361   return (uint32x4_t) __builtin_aarch64_usubw2v8hi ((int32x4_t) __a,
2362                                                     (int16x8_t) __b);
2363 }
2364
2365 __extension__ extern __inline uint64x2_t
2366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2367 vsubw_high_u32 (uint64x2_t __a, uint32x4_t __b)
2368 {
2369   return (uint64x2_t) __builtin_aarch64_usubw2v4si ((int64x2_t) __a,
2370                                                     (int32x4_t) __b);
2371 }
2372
2373 __extension__ extern __inline int8x8_t
2374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2375 vqadd_s8 (int8x8_t __a, int8x8_t __b)
2376 {
2377   return (int8x8_t) __builtin_aarch64_sqaddv8qi (__a, __b);
2378 }
2379
2380 __extension__ extern __inline int16x4_t
2381 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2382 vqadd_s16 (int16x4_t __a, int16x4_t __b)
2383 {
2384   return (int16x4_t) __builtin_aarch64_sqaddv4hi (__a, __b);
2385 }
2386
2387 __extension__ extern __inline int32x2_t
2388 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2389 vqadd_s32 (int32x2_t __a, int32x2_t __b)
2390 {
2391   return (int32x2_t) __builtin_aarch64_sqaddv2si (__a, __b);
2392 }
2393
2394 __extension__ extern __inline int64x1_t
2395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2396 vqadd_s64 (int64x1_t __a, int64x1_t __b)
2397 {
2398   return (int64x1_t) {__builtin_aarch64_sqadddi (__a[0], __b[0])};
2399 }
2400
2401 __extension__ extern __inline uint8x8_t
2402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2403 vqadd_u8 (uint8x8_t __a, uint8x8_t __b)
2404 {
2405   return __builtin_aarch64_uqaddv8qi_uuu (__a, __b);
2406 }
2407
2408 __extension__ extern __inline int8x8_t
2409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2410 vhsub_s8 (int8x8_t __a, int8x8_t __b)
2411 {
2412   return (int8x8_t)__builtin_aarch64_shsubv8qi (__a, __b);
2413 }
2414
2415 __extension__ extern __inline int16x4_t
2416 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2417 vhsub_s16 (int16x4_t __a, int16x4_t __b)
2418 {
2419   return (int16x4_t) __builtin_aarch64_shsubv4hi (__a, __b);
2420 }
2421
2422 __extension__ extern __inline int32x2_t
2423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2424 vhsub_s32 (int32x2_t __a, int32x2_t __b)
2425 {
2426   return (int32x2_t) __builtin_aarch64_shsubv2si (__a, __b);
2427 }
2428
2429 __extension__ extern __inline uint8x8_t
2430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2431 vhsub_u8 (uint8x8_t __a, uint8x8_t __b)
2432 {
2433   return (uint8x8_t) __builtin_aarch64_uhsubv8qi ((int8x8_t) __a,
2434                                                   (int8x8_t) __b);
2435 }
2436
2437 __extension__ extern __inline uint16x4_t
2438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2439 vhsub_u16 (uint16x4_t __a, uint16x4_t __b)
2440 {
2441   return (uint16x4_t) __builtin_aarch64_uhsubv4hi ((int16x4_t) __a,
2442                                                    (int16x4_t) __b);
2443 }
2444
2445 __extension__ extern __inline uint32x2_t
2446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2447 vhsub_u32 (uint32x2_t __a, uint32x2_t __b)
2448 {
2449   return (uint32x2_t) __builtin_aarch64_uhsubv2si ((int32x2_t) __a,
2450                                                    (int32x2_t) __b);
2451 }
2452
2453 __extension__ extern __inline int8x16_t
2454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2455 vhsubq_s8 (int8x16_t __a, int8x16_t __b)
2456 {
2457   return (int8x16_t) __builtin_aarch64_shsubv16qi (__a, __b);
2458 }
2459
2460 __extension__ extern __inline int16x8_t
2461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2462 vhsubq_s16 (int16x8_t __a, int16x8_t __b)
2463 {
2464   return (int16x8_t) __builtin_aarch64_shsubv8hi (__a, __b);
2465 }
2466
2467 __extension__ extern __inline int32x4_t
2468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2469 vhsubq_s32 (int32x4_t __a, int32x4_t __b)
2470 {
2471   return (int32x4_t) __builtin_aarch64_shsubv4si (__a, __b);
2472 }
2473
2474 __extension__ extern __inline uint8x16_t
2475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2476 vhsubq_u8 (uint8x16_t __a, uint8x16_t __b)
2477 {
2478   return (uint8x16_t) __builtin_aarch64_uhsubv16qi ((int8x16_t) __a,
2479                                                     (int8x16_t) __b);
2480 }
2481
2482 __extension__ extern __inline uint16x8_t
2483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2484 vhsubq_u16 (uint16x8_t __a, uint16x8_t __b)
2485 {
2486   return (uint16x8_t) __builtin_aarch64_uhsubv8hi ((int16x8_t) __a,
2487                                                    (int16x8_t) __b);
2488 }
2489
2490 __extension__ extern __inline uint32x4_t
2491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2492 vhsubq_u32 (uint32x4_t __a, uint32x4_t __b)
2493 {
2494   return (uint32x4_t) __builtin_aarch64_uhsubv4si ((int32x4_t) __a,
2495                                                    (int32x4_t) __b);
2496 }
2497
2498 __extension__ extern __inline int8x8_t
2499 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2500 vsubhn_s16 (int16x8_t __a, int16x8_t __b)
2501 {
2502   return (int8x8_t) __builtin_aarch64_subhnv8hi (__a, __b);
2503 }
2504
2505 __extension__ extern __inline int16x4_t
2506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2507 vsubhn_s32 (int32x4_t __a, int32x4_t __b)
2508 {
2509   return (int16x4_t) __builtin_aarch64_subhnv4si (__a, __b);
2510 }
2511
2512 __extension__ extern __inline int32x2_t
2513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2514 vsubhn_s64 (int64x2_t __a, int64x2_t __b)
2515 {
2516   return (int32x2_t) __builtin_aarch64_subhnv2di (__a, __b);
2517 }
2518
2519 __extension__ extern __inline uint8x8_t
2520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2521 vsubhn_u16 (uint16x8_t __a, uint16x8_t __b)
2522 {
2523   return (uint8x8_t) __builtin_aarch64_subhnv8hi ((int16x8_t) __a,
2524                                                   (int16x8_t) __b);
2525 }
2526
2527 __extension__ extern __inline uint16x4_t
2528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2529 vsubhn_u32 (uint32x4_t __a, uint32x4_t __b)
2530 {
2531   return (uint16x4_t) __builtin_aarch64_subhnv4si ((int32x4_t) __a,
2532                                                    (int32x4_t) __b);
2533 }
2534
2535 __extension__ extern __inline uint32x2_t
2536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2537 vsubhn_u64 (uint64x2_t __a, uint64x2_t __b)
2538 {
2539   return (uint32x2_t) __builtin_aarch64_subhnv2di ((int64x2_t) __a,
2540                                                    (int64x2_t) __b);
2541 }
2542
2543 __extension__ extern __inline int8x8_t
2544 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2545 vrsubhn_s16 (int16x8_t __a, int16x8_t __b)
2546 {
2547   return (int8x8_t) __builtin_aarch64_rsubhnv8hi (__a, __b);
2548 }
2549
2550 __extension__ extern __inline int16x4_t
2551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2552 vrsubhn_s32 (int32x4_t __a, int32x4_t __b)
2553 {
2554   return (int16x4_t) __builtin_aarch64_rsubhnv4si (__a, __b);
2555 }
2556
2557 __extension__ extern __inline int32x2_t
2558 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2559 vrsubhn_s64 (int64x2_t __a, int64x2_t __b)
2560 {
2561   return (int32x2_t) __builtin_aarch64_rsubhnv2di (__a, __b);
2562 }
2563
2564 __extension__ extern __inline uint8x8_t
2565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2566 vrsubhn_u16 (uint16x8_t __a, uint16x8_t __b)
2567 {
2568   return (uint8x8_t) __builtin_aarch64_rsubhnv8hi ((int16x8_t) __a,
2569                                                    (int16x8_t) __b);
2570 }
2571
2572 __extension__ extern __inline uint16x4_t
2573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2574 vrsubhn_u32 (uint32x4_t __a, uint32x4_t __b)
2575 {
2576   return (uint16x4_t) __builtin_aarch64_rsubhnv4si ((int32x4_t) __a,
2577                                                     (int32x4_t) __b);
2578 }
2579
2580 __extension__ extern __inline uint32x2_t
2581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2582 vrsubhn_u64 (uint64x2_t __a, uint64x2_t __b)
2583 {
2584   return (uint32x2_t) __builtin_aarch64_rsubhnv2di ((int64x2_t) __a,
2585                                                     (int64x2_t) __b);
2586 }
2587
2588 __extension__ extern __inline int8x16_t
2589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2590 vrsubhn_high_s16 (int8x8_t __a, int16x8_t __b, int16x8_t __c)
2591 {
2592   return (int8x16_t) __builtin_aarch64_rsubhn2v8hi (__a, __b, __c);
2593 }
2594
2595 __extension__ extern __inline int16x8_t
2596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2597 vrsubhn_high_s32 (int16x4_t __a, int32x4_t __b, int32x4_t __c)
2598 {
2599   return (int16x8_t) __builtin_aarch64_rsubhn2v4si (__a, __b, __c);
2600 }
2601
2602 __extension__ extern __inline int32x4_t
2603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2604 vrsubhn_high_s64 (int32x2_t __a, int64x2_t __b, int64x2_t __c)
2605 {
2606   return (int32x4_t) __builtin_aarch64_rsubhn2v2di (__a, __b, __c);
2607 }
2608
2609 __extension__ extern __inline uint8x16_t
2610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2611 vrsubhn_high_u16 (uint8x8_t __a, uint16x8_t __b, uint16x8_t __c)
2612 {
2613   return (uint8x16_t) __builtin_aarch64_rsubhn2v8hi ((int8x8_t) __a,
2614                                                      (int16x8_t) __b,
2615                                                      (int16x8_t) __c);
2616 }
2617
2618 __extension__ extern __inline uint16x8_t
2619 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2620 vrsubhn_high_u32 (uint16x4_t __a, uint32x4_t __b, uint32x4_t __c)
2621 {
2622   return (uint16x8_t) __builtin_aarch64_rsubhn2v4si ((int16x4_t) __a,
2623                                                      (int32x4_t) __b,
2624                                                      (int32x4_t) __c);
2625 }
2626
2627 __extension__ extern __inline uint32x4_t
2628 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2629 vrsubhn_high_u64 (uint32x2_t __a, uint64x2_t __b, uint64x2_t __c)
2630 {
2631   return (uint32x4_t) __builtin_aarch64_rsubhn2v2di ((int32x2_t) __a,
2632                                                      (int64x2_t) __b,
2633                                                      (int64x2_t) __c);
2634 }
2635
2636 __extension__ extern __inline int8x16_t
2637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2638 vsubhn_high_s16 (int8x8_t __a, int16x8_t __b, int16x8_t __c)
2639 {
2640   return (int8x16_t) __builtin_aarch64_subhn2v8hi (__a, __b, __c);
2641 }
2642
2643 __extension__ extern __inline int16x8_t
2644 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2645 vsubhn_high_s32 (int16x4_t __a, int32x4_t __b, int32x4_t __c)
2646 {
2647   return (int16x8_t) __builtin_aarch64_subhn2v4si (__a, __b, __c);;
2648 }
2649
2650 __extension__ extern __inline int32x4_t
2651 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2652 vsubhn_high_s64 (int32x2_t __a, int64x2_t __b, int64x2_t __c)
2653 {
2654   return (int32x4_t) __builtin_aarch64_subhn2v2di (__a, __b, __c);
2655 }
2656
2657 __extension__ extern __inline uint8x16_t
2658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2659 vsubhn_high_u16 (uint8x8_t __a, uint16x8_t __b, uint16x8_t __c)
2660 {
2661   return (uint8x16_t) __builtin_aarch64_subhn2v8hi ((int8x8_t) __a,
2662                                                     (int16x8_t) __b,
2663                                                     (int16x8_t) __c);
2664 }
2665
2666 __extension__ extern __inline uint16x8_t
2667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2668 vsubhn_high_u32 (uint16x4_t __a, uint32x4_t __b, uint32x4_t __c)
2669 {
2670   return (uint16x8_t) __builtin_aarch64_subhn2v4si ((int16x4_t) __a,
2671                                                     (int32x4_t) __b,
2672                                                     (int32x4_t) __c);
2673 }
2674
2675 __extension__ extern __inline uint32x4_t
2676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2677 vsubhn_high_u64 (uint32x2_t __a, uint64x2_t __b, uint64x2_t __c)
2678 {
2679   return (uint32x4_t) __builtin_aarch64_subhn2v2di ((int32x2_t) __a,
2680                                                     (int64x2_t) __b,
2681                                                     (int64x2_t) __c);
2682 }
2683
2684 __extension__ extern __inline uint16x4_t
2685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2686 vqadd_u16 (uint16x4_t __a, uint16x4_t __b)
2687 {
2688   return __builtin_aarch64_uqaddv4hi_uuu (__a, __b);
2689 }
2690
2691 __extension__ extern __inline uint32x2_t
2692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2693 vqadd_u32 (uint32x2_t __a, uint32x2_t __b)
2694 {
2695   return __builtin_aarch64_uqaddv2si_uuu (__a, __b);
2696 }
2697
2698 __extension__ extern __inline uint64x1_t
2699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2700 vqadd_u64 (uint64x1_t __a, uint64x1_t __b)
2701 {
2702   return (uint64x1_t) {__builtin_aarch64_uqadddi_uuu (__a[0], __b[0])};
2703 }
2704
2705 __extension__ extern __inline int8x16_t
2706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2707 vqaddq_s8 (int8x16_t __a, int8x16_t __b)
2708 {
2709   return (int8x16_t) __builtin_aarch64_sqaddv16qi (__a, __b);
2710 }
2711
2712 __extension__ extern __inline int16x8_t
2713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2714 vqaddq_s16 (int16x8_t __a, int16x8_t __b)
2715 {
2716   return (int16x8_t) __builtin_aarch64_sqaddv8hi (__a, __b);
2717 }
2718
2719 __extension__ extern __inline int32x4_t
2720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2721 vqaddq_s32 (int32x4_t __a, int32x4_t __b)
2722 {
2723   return (int32x4_t) __builtin_aarch64_sqaddv4si (__a, __b);
2724 }
2725
2726 __extension__ extern __inline int64x2_t
2727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2728 vqaddq_s64 (int64x2_t __a, int64x2_t __b)
2729 {
2730   return (int64x2_t) __builtin_aarch64_sqaddv2di (__a, __b);
2731 }
2732
2733 __extension__ extern __inline uint8x16_t
2734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2735 vqaddq_u8 (uint8x16_t __a, uint8x16_t __b)
2736 {
2737   return __builtin_aarch64_uqaddv16qi_uuu (__a, __b);
2738 }
2739
2740 __extension__ extern __inline uint16x8_t
2741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2742 vqaddq_u16 (uint16x8_t __a, uint16x8_t __b)
2743 {
2744   return __builtin_aarch64_uqaddv8hi_uuu (__a, __b);
2745 }
2746
2747 __extension__ extern __inline uint32x4_t
2748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2749 vqaddq_u32 (uint32x4_t __a, uint32x4_t __b)
2750 {
2751   return __builtin_aarch64_uqaddv4si_uuu (__a, __b);
2752 }
2753
2754 __extension__ extern __inline uint64x2_t
2755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2756 vqaddq_u64 (uint64x2_t __a, uint64x2_t __b)
2757 {
2758   return __builtin_aarch64_uqaddv2di_uuu (__a, __b);
2759 }
2760
2761 __extension__ extern __inline int8x8_t
2762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2763 vqsub_s8 (int8x8_t __a, int8x8_t __b)
2764 {
2765   return (int8x8_t) __builtin_aarch64_sqsubv8qi (__a, __b);
2766 }
2767
2768 __extension__ extern __inline int16x4_t
2769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2770 vqsub_s16 (int16x4_t __a, int16x4_t __b)
2771 {
2772   return (int16x4_t) __builtin_aarch64_sqsubv4hi (__a, __b);
2773 }
2774
2775 __extension__ extern __inline int32x2_t
2776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2777 vqsub_s32 (int32x2_t __a, int32x2_t __b)
2778 {
2779   return (int32x2_t) __builtin_aarch64_sqsubv2si (__a, __b);
2780 }
2781
2782 __extension__ extern __inline int64x1_t
2783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2784 vqsub_s64 (int64x1_t __a, int64x1_t __b)
2785 {
2786   return (int64x1_t) {__builtin_aarch64_sqsubdi (__a[0], __b[0])};
2787 }
2788
2789 __extension__ extern __inline uint8x8_t
2790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2791 vqsub_u8 (uint8x8_t __a, uint8x8_t __b)
2792 {
2793   return __builtin_aarch64_uqsubv8qi_uuu (__a, __b);
2794 }
2795
2796 __extension__ extern __inline uint16x4_t
2797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2798 vqsub_u16 (uint16x4_t __a, uint16x4_t __b)
2799 {
2800   return __builtin_aarch64_uqsubv4hi_uuu (__a, __b);
2801 }
2802
2803 __extension__ extern __inline uint32x2_t
2804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2805 vqsub_u32 (uint32x2_t __a, uint32x2_t __b)
2806 {
2807   return __builtin_aarch64_uqsubv2si_uuu (__a, __b);
2808 }
2809
2810 __extension__ extern __inline uint64x1_t
2811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2812 vqsub_u64 (uint64x1_t __a, uint64x1_t __b)
2813 {
2814   return (uint64x1_t) {__builtin_aarch64_uqsubdi_uuu (__a[0], __b[0])};
2815 }
2816
2817 __extension__ extern __inline int8x16_t
2818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2819 vqsubq_s8 (int8x16_t __a, int8x16_t __b)
2820 {
2821   return (int8x16_t) __builtin_aarch64_sqsubv16qi (__a, __b);
2822 }
2823
2824 __extension__ extern __inline int16x8_t
2825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2826 vqsubq_s16 (int16x8_t __a, int16x8_t __b)
2827 {
2828   return (int16x8_t) __builtin_aarch64_sqsubv8hi (__a, __b);
2829 }
2830
2831 __extension__ extern __inline int32x4_t
2832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2833 vqsubq_s32 (int32x4_t __a, int32x4_t __b)
2834 {
2835   return (int32x4_t) __builtin_aarch64_sqsubv4si (__a, __b);
2836 }
2837
2838 __extension__ extern __inline int64x2_t
2839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2840 vqsubq_s64 (int64x2_t __a, int64x2_t __b)
2841 {
2842   return (int64x2_t) __builtin_aarch64_sqsubv2di (__a, __b);
2843 }
2844
2845 __extension__ extern __inline uint8x16_t
2846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2847 vqsubq_u8 (uint8x16_t __a, uint8x16_t __b)
2848 {
2849   return __builtin_aarch64_uqsubv16qi_uuu (__a, __b);
2850 }
2851
2852 __extension__ extern __inline uint16x8_t
2853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2854 vqsubq_u16 (uint16x8_t __a, uint16x8_t __b)
2855 {
2856   return __builtin_aarch64_uqsubv8hi_uuu (__a, __b);
2857 }
2858
2859 __extension__ extern __inline uint32x4_t
2860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2861 vqsubq_u32 (uint32x4_t __a, uint32x4_t __b)
2862 {
2863   return __builtin_aarch64_uqsubv4si_uuu (__a, __b);
2864 }
2865
2866 __extension__ extern __inline uint64x2_t
2867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2868 vqsubq_u64 (uint64x2_t __a, uint64x2_t __b)
2869 {
2870   return __builtin_aarch64_uqsubv2di_uuu (__a, __b);
2871 }
2872
2873 __extension__ extern __inline int8x8_t
2874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2875 vqneg_s8 (int8x8_t __a)
2876 {
2877   return (int8x8_t) __builtin_aarch64_sqnegv8qi (__a);
2878 }
2879
2880 __extension__ extern __inline int16x4_t
2881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2882 vqneg_s16 (int16x4_t __a)
2883 {
2884   return (int16x4_t) __builtin_aarch64_sqnegv4hi (__a);
2885 }
2886
2887 __extension__ extern __inline int32x2_t
2888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2889 vqneg_s32 (int32x2_t __a)
2890 {
2891   return (int32x2_t) __builtin_aarch64_sqnegv2si (__a);
2892 }
2893
2894 __extension__ extern __inline int64x1_t
2895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2896 vqneg_s64 (int64x1_t __a)
2897 {
2898   return (int64x1_t) {__builtin_aarch64_sqnegdi (__a[0])};
2899 }
2900
2901 __extension__ extern __inline int8x16_t
2902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2903 vqnegq_s8 (int8x16_t __a)
2904 {
2905   return (int8x16_t) __builtin_aarch64_sqnegv16qi (__a);
2906 }
2907
2908 __extension__ extern __inline int16x8_t
2909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2910 vqnegq_s16 (int16x8_t __a)
2911 {
2912   return (int16x8_t) __builtin_aarch64_sqnegv8hi (__a);
2913 }
2914
2915 __extension__ extern __inline int32x4_t
2916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2917 vqnegq_s32 (int32x4_t __a)
2918 {
2919   return (int32x4_t) __builtin_aarch64_sqnegv4si (__a);
2920 }
2921
2922 __extension__ extern __inline int8x8_t
2923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2924 vqabs_s8 (int8x8_t __a)
2925 {
2926   return (int8x8_t) __builtin_aarch64_sqabsv8qi (__a);
2927 }
2928
2929 __extension__ extern __inline int16x4_t
2930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2931 vqabs_s16 (int16x4_t __a)
2932 {
2933   return (int16x4_t) __builtin_aarch64_sqabsv4hi (__a);
2934 }
2935
2936 __extension__ extern __inline int32x2_t
2937 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2938 vqabs_s32 (int32x2_t __a)
2939 {
2940   return (int32x2_t) __builtin_aarch64_sqabsv2si (__a);
2941 }
2942
2943 __extension__ extern __inline int64x1_t
2944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2945 vqabs_s64 (int64x1_t __a)
2946 {
2947   return (int64x1_t) {__builtin_aarch64_sqabsdi (__a[0])};
2948 }
2949
2950 __extension__ extern __inline int8x16_t
2951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2952 vqabsq_s8 (int8x16_t __a)
2953 {
2954   return (int8x16_t) __builtin_aarch64_sqabsv16qi (__a);
2955 }
2956
2957 __extension__ extern __inline int16x8_t
2958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2959 vqabsq_s16 (int16x8_t __a)
2960 {
2961   return (int16x8_t) __builtin_aarch64_sqabsv8hi (__a);
2962 }
2963
2964 __extension__ extern __inline int32x4_t
2965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2966 vqabsq_s32 (int32x4_t __a)
2967 {
2968   return (int32x4_t) __builtin_aarch64_sqabsv4si (__a);
2969 }
2970
2971 __extension__ extern __inline int16x4_t
2972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2973 vqdmulh_s16 (int16x4_t __a, int16x4_t __b)
2974 {
2975   return (int16x4_t) __builtin_aarch64_sqdmulhv4hi (__a, __b);
2976 }
2977
2978 __extension__ extern __inline int32x2_t
2979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2980 vqdmulh_s32 (int32x2_t __a, int32x2_t __b)
2981 {
2982   return (int32x2_t) __builtin_aarch64_sqdmulhv2si (__a, __b);
2983 }
2984
2985 __extension__ extern __inline int16x8_t
2986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2987 vqdmulhq_s16 (int16x8_t __a, int16x8_t __b)
2988 {
2989   return (int16x8_t) __builtin_aarch64_sqdmulhv8hi (__a, __b);
2990 }
2991
2992 __extension__ extern __inline int32x4_t
2993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
2994 vqdmulhq_s32 (int32x4_t __a, int32x4_t __b)
2995 {
2996   return (int32x4_t) __builtin_aarch64_sqdmulhv4si (__a, __b);
2997 }
2998
2999 __extension__ extern __inline int16x4_t
3000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3001 vqrdmulh_s16 (int16x4_t __a, int16x4_t __b)
3002 {
3003   return (int16x4_t) __builtin_aarch64_sqrdmulhv4hi (__a, __b);
3004 }
3005
3006 __extension__ extern __inline int32x2_t
3007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3008 vqrdmulh_s32 (int32x2_t __a, int32x2_t __b)
3009 {
3010   return (int32x2_t) __builtin_aarch64_sqrdmulhv2si (__a, __b);
3011 }
3012
3013 __extension__ extern __inline int16x8_t
3014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3015 vqrdmulhq_s16 (int16x8_t __a, int16x8_t __b)
3016 {
3017   return (int16x8_t) __builtin_aarch64_sqrdmulhv8hi (__a, __b);
3018 }
3019
3020 __extension__ extern __inline int32x4_t
3021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3022 vqrdmulhq_s32 (int32x4_t __a, int32x4_t __b)
3023 {
3024   return (int32x4_t) __builtin_aarch64_sqrdmulhv4si (__a, __b);
3025 }
3026
3027 __extension__ extern __inline int8x8_t
3028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3029 vcreate_s8 (uint64_t __a)
3030 {
3031   return (int8x8_t) __a;
3032 }
3033
3034 __extension__ extern __inline int16x4_t
3035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3036 vcreate_s16 (uint64_t __a)
3037 {
3038   return (int16x4_t) __a;
3039 }
3040
3041 __extension__ extern __inline int32x2_t
3042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3043 vcreate_s32 (uint64_t __a)
3044 {
3045   return (int32x2_t) __a;
3046 }
3047
3048 __extension__ extern __inline int64x1_t
3049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3050 vcreate_s64 (uint64_t __a)
3051 {
3052   return (int64x1_t) {__a};
3053 }
3054
3055 __extension__ extern __inline float16x4_t
3056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3057 vcreate_f16 (uint64_t __a)
3058 {
3059   return (float16x4_t) __a;
3060 }
3061
3062 __extension__ extern __inline float32x2_t
3063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3064 vcreate_f32 (uint64_t __a)
3065 {
3066   return (float32x2_t) __a;
3067 }
3068
3069 __extension__ extern __inline uint8x8_t
3070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3071 vcreate_u8 (uint64_t __a)
3072 {
3073   return (uint8x8_t) __a;
3074 }
3075
3076 __extension__ extern __inline uint16x4_t
3077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3078 vcreate_u16 (uint64_t __a)
3079 {
3080   return (uint16x4_t) __a;
3081 }
3082
3083 __extension__ extern __inline uint32x2_t
3084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3085 vcreate_u32 (uint64_t __a)
3086 {
3087   return (uint32x2_t) __a;
3088 }
3089
3090 __extension__ extern __inline uint64x1_t
3091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3092 vcreate_u64 (uint64_t __a)
3093 {
3094   return (uint64x1_t) {__a};
3095 }
3096
3097 __extension__ extern __inline float64x1_t
3098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3099 vcreate_f64 (uint64_t __a)
3100 {
3101   return (float64x1_t) __a;
3102 }
3103
3104 __extension__ extern __inline poly8x8_t
3105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3106 vcreate_p8 (uint64_t __a)
3107 {
3108   return (poly8x8_t) __a;
3109 }
3110
3111 __extension__ extern __inline poly16x4_t
3112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3113 vcreate_p16 (uint64_t __a)
3114 {
3115   return (poly16x4_t) __a;
3116 }
3117
3118 __extension__ extern __inline poly64x1_t
3119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3120 vcreate_p64 (uint64_t __a)
3121 {
3122   return (poly64x1_t) __a;
3123 }
3124
3125 /* vget_lane  */
3126
3127 __extension__ extern __inline float16_t
3128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3129 vget_lane_f16 (float16x4_t __a, const int __b)
3130 {
3131   return __aarch64_vget_lane_any (__a, __b);
3132 }
3133
3134 __extension__ extern __inline float32_t
3135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3136 vget_lane_f32 (float32x2_t __a, const int __b)
3137 {
3138   return __aarch64_vget_lane_any (__a, __b);
3139 }
3140
3141 __extension__ extern __inline float64_t
3142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3143 vget_lane_f64 (float64x1_t __a, const int __b)
3144 {
3145   return __aarch64_vget_lane_any (__a, __b);
3146 }
3147
3148 __extension__ extern __inline poly8_t
3149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3150 vget_lane_p8 (poly8x8_t __a, const int __b)
3151 {
3152   return __aarch64_vget_lane_any (__a, __b);
3153 }
3154
3155 __extension__ extern __inline poly16_t
3156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3157 vget_lane_p16 (poly16x4_t __a, const int __b)
3158 {
3159   return __aarch64_vget_lane_any (__a, __b);
3160 }
3161
3162 __extension__ extern __inline poly64_t
3163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3164 vget_lane_p64 (poly64x1_t __a, const int __b)
3165 {
3166   return __aarch64_vget_lane_any (__a, __b);
3167 }
3168
3169 __extension__ extern __inline int8_t
3170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3171 vget_lane_s8 (int8x8_t __a, const int __b)
3172 {
3173   return __aarch64_vget_lane_any (__a, __b);
3174 }
3175
3176 __extension__ extern __inline int16_t
3177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3178 vget_lane_s16 (int16x4_t __a, const int __b)
3179 {
3180   return __aarch64_vget_lane_any (__a, __b);
3181 }
3182
3183 __extension__ extern __inline int32_t
3184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3185 vget_lane_s32 (int32x2_t __a, const int __b)
3186 {
3187   return __aarch64_vget_lane_any (__a, __b);
3188 }
3189
3190 __extension__ extern __inline int64_t
3191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3192 vget_lane_s64 (int64x1_t __a, const int __b)
3193 {
3194   return __aarch64_vget_lane_any (__a, __b);
3195 }
3196
3197 __extension__ extern __inline uint8_t
3198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3199 vget_lane_u8 (uint8x8_t __a, const int __b)
3200 {
3201   return __aarch64_vget_lane_any (__a, __b);
3202 }
3203
3204 __extension__ extern __inline uint16_t
3205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3206 vget_lane_u16 (uint16x4_t __a, const int __b)
3207 {
3208   return __aarch64_vget_lane_any (__a, __b);
3209 }
3210
3211 __extension__ extern __inline uint32_t
3212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3213 vget_lane_u32 (uint32x2_t __a, const int __b)
3214 {
3215   return __aarch64_vget_lane_any (__a, __b);
3216 }
3217
3218 __extension__ extern __inline uint64_t
3219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3220 vget_lane_u64 (uint64x1_t __a, const int __b)
3221 {
3222   return __aarch64_vget_lane_any (__a, __b);
3223 }
3224
3225 /* vgetq_lane  */
3226
3227 __extension__ extern __inline float16_t
3228 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3229 vgetq_lane_f16 (float16x8_t __a, const int __b)
3230 {
3231   return __aarch64_vget_lane_any (__a, __b);
3232 }
3233
3234 __extension__ extern __inline float32_t
3235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3236 vgetq_lane_f32 (float32x4_t __a, const int __b)
3237 {
3238   return __aarch64_vget_lane_any (__a, __b);
3239 }
3240
3241 __extension__ extern __inline float64_t
3242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3243 vgetq_lane_f64 (float64x2_t __a, const int __b)
3244 {
3245   return __aarch64_vget_lane_any (__a, __b);
3246 }
3247
3248 __extension__ extern __inline poly8_t
3249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3250 vgetq_lane_p8 (poly8x16_t __a, const int __b)
3251 {
3252   return __aarch64_vget_lane_any (__a, __b);
3253 }
3254
3255 __extension__ extern __inline poly16_t
3256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3257 vgetq_lane_p16 (poly16x8_t __a, const int __b)
3258 {
3259   return __aarch64_vget_lane_any (__a, __b);
3260 }
3261
3262 __extension__ extern __inline poly64_t
3263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3264 vgetq_lane_p64 (poly64x2_t __a, const int __b)
3265 {
3266   return __aarch64_vget_lane_any (__a, __b);
3267 }
3268
3269 __extension__ extern __inline int8_t
3270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3271 vgetq_lane_s8 (int8x16_t __a, const int __b)
3272 {
3273   return __aarch64_vget_lane_any (__a, __b);
3274 }
3275
3276 __extension__ extern __inline int16_t
3277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3278 vgetq_lane_s16 (int16x8_t __a, const int __b)
3279 {
3280   return __aarch64_vget_lane_any (__a, __b);
3281 }
3282
3283 __extension__ extern __inline int32_t
3284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3285 vgetq_lane_s32 (int32x4_t __a, const int __b)
3286 {
3287   return __aarch64_vget_lane_any (__a, __b);
3288 }
3289
3290 __extension__ extern __inline int64_t
3291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3292 vgetq_lane_s64 (int64x2_t __a, const int __b)
3293 {
3294   return __aarch64_vget_lane_any (__a, __b);
3295 }
3296
3297 __extension__ extern __inline uint8_t
3298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3299 vgetq_lane_u8 (uint8x16_t __a, const int __b)
3300 {
3301   return __aarch64_vget_lane_any (__a, __b);
3302 }
3303
3304 __extension__ extern __inline uint16_t
3305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3306 vgetq_lane_u16 (uint16x8_t __a, const int __b)
3307 {
3308   return __aarch64_vget_lane_any (__a, __b);
3309 }
3310
3311 __extension__ extern __inline uint32_t
3312 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3313 vgetq_lane_u32 (uint32x4_t __a, const int __b)
3314 {
3315   return __aarch64_vget_lane_any (__a, __b);
3316 }
3317
3318 __extension__ extern __inline uint64_t
3319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3320 vgetq_lane_u64 (uint64x2_t __a, const int __b)
3321 {
3322   return __aarch64_vget_lane_any (__a, __b);
3323 }
3324
3325 /* vreinterpret  */
3326
3327 __extension__ extern __inline poly8x8_t
3328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3329 vreinterpret_p8_f16 (float16x4_t __a)
3330 {
3331   return (poly8x8_t) __a;
3332 }
3333
3334 __extension__ extern __inline poly8x8_t
3335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3336 vreinterpret_p8_f64 (float64x1_t __a)
3337 {
3338   return (poly8x8_t) __a;
3339 }
3340
3341 __extension__ extern __inline poly8x8_t
3342 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3343 vreinterpret_p8_s8 (int8x8_t __a)
3344 {
3345   return (poly8x8_t) __a;
3346 }
3347
3348 __extension__ extern __inline poly8x8_t
3349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3350 vreinterpret_p8_s16 (int16x4_t __a)
3351 {
3352   return (poly8x8_t) __a;
3353 }
3354
3355 __extension__ extern __inline poly8x8_t
3356 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3357 vreinterpret_p8_s32 (int32x2_t __a)
3358 {
3359   return (poly8x8_t) __a;
3360 }
3361
3362 __extension__ extern __inline poly8x8_t
3363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3364 vreinterpret_p8_s64 (int64x1_t __a)
3365 {
3366   return (poly8x8_t) __a;
3367 }
3368
3369 __extension__ extern __inline poly8x8_t
3370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3371 vreinterpret_p8_f32 (float32x2_t __a)
3372 {
3373   return (poly8x8_t) __a;
3374 }
3375
3376 __extension__ extern __inline poly8x8_t
3377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3378 vreinterpret_p8_u8 (uint8x8_t __a)
3379 {
3380   return (poly8x8_t) __a;
3381 }
3382
3383 __extension__ extern __inline poly8x8_t
3384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3385 vreinterpret_p8_u16 (uint16x4_t __a)
3386 {
3387   return (poly8x8_t) __a;
3388 }
3389
3390 __extension__ extern __inline poly8x8_t
3391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3392 vreinterpret_p8_u32 (uint32x2_t __a)
3393 {
3394   return (poly8x8_t) __a;
3395 }
3396
3397 __extension__ extern __inline poly8x8_t
3398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3399 vreinterpret_p8_u64 (uint64x1_t __a)
3400 {
3401   return (poly8x8_t) __a;
3402 }
3403
3404 __extension__ extern __inline poly8x8_t
3405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3406 vreinterpret_p8_p16 (poly16x4_t __a)
3407 {
3408   return (poly8x8_t) __a;
3409 }
3410
3411 __extension__ extern __inline poly8x8_t
3412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3413 vreinterpret_p8_p64 (poly64x1_t __a)
3414 {
3415   return (poly8x8_t) __a;
3416 }
3417
3418 __extension__ extern __inline poly8x16_t
3419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3420 vreinterpretq_p8_f64 (float64x2_t __a)
3421 {
3422   return (poly8x16_t) __a;
3423 }
3424
3425 __extension__ extern __inline poly8x16_t
3426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3427 vreinterpretq_p8_s8 (int8x16_t __a)
3428 {
3429   return (poly8x16_t) __a;
3430 }
3431
3432 __extension__ extern __inline poly8x16_t
3433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3434 vreinterpretq_p8_s16 (int16x8_t __a)
3435 {
3436   return (poly8x16_t) __a;
3437 }
3438
3439 __extension__ extern __inline poly8x16_t
3440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3441 vreinterpretq_p8_s32 (int32x4_t __a)
3442 {
3443   return (poly8x16_t) __a;
3444 }
3445
3446 __extension__ extern __inline poly8x16_t
3447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3448 vreinterpretq_p8_s64 (int64x2_t __a)
3449 {
3450   return (poly8x16_t) __a;
3451 }
3452
3453 __extension__ extern __inline poly8x16_t
3454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3455 vreinterpretq_p8_f16 (float16x8_t __a)
3456 {
3457   return (poly8x16_t) __a;
3458 }
3459
3460 __extension__ extern __inline poly8x16_t
3461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3462 vreinterpretq_p8_f32 (float32x4_t __a)
3463 {
3464   return (poly8x16_t) __a;
3465 }
3466
3467 __extension__ extern __inline poly8x16_t
3468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3469 vreinterpretq_p8_u8 (uint8x16_t __a)
3470 {
3471   return (poly8x16_t) __a;
3472 }
3473
3474 __extension__ extern __inline poly8x16_t
3475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3476 vreinterpretq_p8_u16 (uint16x8_t __a)
3477 {
3478   return (poly8x16_t) __a;
3479 }
3480
3481 __extension__ extern __inline poly8x16_t
3482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3483 vreinterpretq_p8_u32 (uint32x4_t __a)
3484 {
3485   return (poly8x16_t) __a;
3486 }
3487
3488 __extension__ extern __inline poly8x16_t
3489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3490 vreinterpretq_p8_u64 (uint64x2_t __a)
3491 {
3492   return (poly8x16_t) __a;
3493 }
3494
3495 __extension__ extern __inline poly8x16_t
3496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3497 vreinterpretq_p8_p16 (poly16x8_t __a)
3498 {
3499   return (poly8x16_t) __a;
3500 }
3501
3502 __extension__ extern __inline poly8x16_t
3503 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3504 vreinterpretq_p8_p64 (poly64x2_t __a)
3505 {
3506   return (poly8x16_t) __a;
3507 }
3508
3509 __extension__ extern __inline poly8x16_t
3510 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3511 vreinterpretq_p8_p128 (poly128_t __a)
3512 {
3513   return (poly8x16_t)__a;
3514 }
3515
3516 __extension__ extern __inline poly16x4_t
3517 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3518 vreinterpret_p16_f16 (float16x4_t __a)
3519 {
3520   return (poly16x4_t) __a;
3521 }
3522
3523 __extension__ extern __inline poly16x4_t
3524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3525 vreinterpret_p16_f64 (float64x1_t __a)
3526 {
3527   return (poly16x4_t) __a;
3528 }
3529
3530 __extension__ extern __inline poly16x4_t
3531 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3532 vreinterpret_p16_s8 (int8x8_t __a)
3533 {
3534   return (poly16x4_t) __a;
3535 }
3536
3537 __extension__ extern __inline poly16x4_t
3538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3539 vreinterpret_p16_s16 (int16x4_t __a)
3540 {
3541   return (poly16x4_t) __a;
3542 }
3543
3544 __extension__ extern __inline poly16x4_t
3545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3546 vreinterpret_p16_s32 (int32x2_t __a)
3547 {
3548   return (poly16x4_t) __a;
3549 }
3550
3551 __extension__ extern __inline poly16x4_t
3552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3553 vreinterpret_p16_s64 (int64x1_t __a)
3554 {
3555   return (poly16x4_t) __a;
3556 }
3557
3558 __extension__ extern __inline poly16x4_t
3559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3560 vreinterpret_p16_f32 (float32x2_t __a)
3561 {
3562   return (poly16x4_t) __a;
3563 }
3564
3565 __extension__ extern __inline poly16x4_t
3566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3567 vreinterpret_p16_u8 (uint8x8_t __a)
3568 {
3569   return (poly16x4_t) __a;
3570 }
3571
3572 __extension__ extern __inline poly16x4_t
3573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3574 vreinterpret_p16_u16 (uint16x4_t __a)
3575 {
3576   return (poly16x4_t) __a;
3577 }
3578
3579 __extension__ extern __inline poly16x4_t
3580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3581 vreinterpret_p16_u32 (uint32x2_t __a)
3582 {
3583   return (poly16x4_t) __a;
3584 }
3585
3586 __extension__ extern __inline poly16x4_t
3587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3588 vreinterpret_p16_u64 (uint64x1_t __a)
3589 {
3590   return (poly16x4_t) __a;
3591 }
3592
3593 __extension__ extern __inline poly16x4_t
3594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3595 vreinterpret_p16_p8 (poly8x8_t __a)
3596 {
3597   return (poly16x4_t) __a;
3598 }
3599
3600 __extension__ extern __inline poly16x4_t
3601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3602 vreinterpret_p16_p64 (poly64x1_t __a)
3603 {
3604   return (poly16x4_t) __a;
3605 }
3606
3607 __extension__ extern __inline poly16x8_t
3608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3609 vreinterpretq_p16_f64 (float64x2_t __a)
3610 {
3611   return (poly16x8_t) __a;
3612 }
3613
3614 __extension__ extern __inline poly16x8_t
3615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3616 vreinterpretq_p16_s8 (int8x16_t __a)
3617 {
3618   return (poly16x8_t) __a;
3619 }
3620
3621 __extension__ extern __inline poly16x8_t
3622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3623 vreinterpretq_p16_s16 (int16x8_t __a)
3624 {
3625   return (poly16x8_t) __a;
3626 }
3627
3628 __extension__ extern __inline poly16x8_t
3629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3630 vreinterpretq_p16_s32 (int32x4_t __a)
3631 {
3632   return (poly16x8_t) __a;
3633 }
3634
3635 __extension__ extern __inline poly16x8_t
3636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3637 vreinterpretq_p16_s64 (int64x2_t __a)
3638 {
3639   return (poly16x8_t) __a;
3640 }
3641
3642 __extension__ extern __inline poly16x8_t
3643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3644 vreinterpretq_p16_f16 (float16x8_t __a)
3645 {
3646   return (poly16x8_t) __a;
3647 }
3648
3649 __extension__ extern __inline poly16x8_t
3650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3651 vreinterpretq_p16_f32 (float32x4_t __a)
3652 {
3653   return (poly16x8_t) __a;
3654 }
3655
3656 __extension__ extern __inline poly16x8_t
3657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3658 vreinterpretq_p16_u8 (uint8x16_t __a)
3659 {
3660   return (poly16x8_t) __a;
3661 }
3662
3663 __extension__ extern __inline poly16x8_t
3664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3665 vreinterpretq_p16_u16 (uint16x8_t __a)
3666 {
3667   return (poly16x8_t) __a;
3668 }
3669
3670 __extension__ extern __inline poly16x8_t
3671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3672 vreinterpretq_p16_u32 (uint32x4_t __a)
3673 {
3674   return (poly16x8_t) __a;
3675 }
3676
3677 __extension__ extern __inline poly16x8_t
3678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3679 vreinterpretq_p16_u64 (uint64x2_t __a)
3680 {
3681   return (poly16x8_t) __a;
3682 }
3683
3684 __extension__ extern __inline poly16x8_t
3685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3686 vreinterpretq_p16_p8 (poly8x16_t __a)
3687 {
3688   return (poly16x8_t) __a;
3689 }
3690
3691 __extension__ extern __inline poly16x8_t
3692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3693 vreinterpretq_p16_p64 (poly64x2_t __a)
3694 {
3695   return (poly16x8_t) __a;
3696 }
3697
3698 __extension__ extern __inline poly16x8_t
3699 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3700 vreinterpretq_p16_p128 (poly128_t __a)
3701 {
3702   return (poly16x8_t)__a;
3703 }
3704
3705 __extension__ extern __inline poly64x1_t
3706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3707 vreinterpret_p64_f16 (float16x4_t __a)
3708 {
3709   return (poly64x1_t) __a;
3710 }
3711
3712 __extension__ extern __inline poly64x1_t
3713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3714 vreinterpret_p64_f64 (float64x1_t __a)
3715 {
3716   return (poly64x1_t) __a;
3717 }
3718
3719 __extension__ extern __inline poly64x1_t
3720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3721 vreinterpret_p64_s8 (int8x8_t __a)
3722 {
3723   return (poly64x1_t) __a;
3724 }
3725
3726 __extension__ extern __inline poly64x1_t
3727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3728 vreinterpret_p64_s16 (int16x4_t __a)
3729 {
3730   return (poly64x1_t) __a;
3731 }
3732
3733 __extension__ extern __inline poly64x1_t
3734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3735 vreinterpret_p64_s32 (int32x2_t __a)
3736 {
3737   return (poly64x1_t) __a;
3738 }
3739
3740 __extension__ extern __inline poly64x1_t
3741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3742 vreinterpret_p64_s64 (int64x1_t __a)
3743 {
3744   return (poly64x1_t) __a;
3745 }
3746
3747 __extension__ extern __inline poly64x1_t
3748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3749 vreinterpret_p64_f32 (float32x2_t __a)
3750 {
3751   return (poly64x1_t) __a;
3752 }
3753
3754 __extension__ extern __inline poly64x1_t
3755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3756 vreinterpret_p64_u8 (uint8x8_t __a)
3757 {
3758   return (poly64x1_t) __a;
3759 }
3760
3761 __extension__ extern __inline poly64x1_t
3762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3763 vreinterpret_p64_u16 (uint16x4_t __a)
3764 {
3765   return (poly64x1_t) __a;
3766 }
3767
3768 __extension__ extern __inline poly64x1_t
3769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3770 vreinterpret_p64_u32 (uint32x2_t __a)
3771 {
3772   return (poly64x1_t) __a;
3773 }
3774
3775 __extension__ extern __inline poly64x1_t
3776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3777 vreinterpret_p64_u64 (uint64x1_t __a)
3778 {
3779   return (poly64x1_t) __a;
3780 }
3781
3782 __extension__ extern __inline poly64x1_t
3783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3784 vreinterpret_p64_p8 (poly8x8_t __a)
3785 {
3786   return (poly64x1_t) __a;
3787 }
3788
3789 __extension__ extern __inline poly64x1_t
3790 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3791 vreinterpret_p64_p16 (poly16x4_t __a)
3792 {
3793   return (poly64x1_t)__a;
3794 }
3795
3796 __extension__ extern __inline poly64x2_t
3797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3798 vreinterpretq_p64_f64 (float64x2_t __a)
3799 {
3800   return (poly64x2_t) __a;
3801 }
3802
3803 __extension__ extern __inline poly64x2_t
3804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3805 vreinterpretq_p64_s8 (int8x16_t __a)
3806 {
3807   return (poly64x2_t) __a;
3808 }
3809
3810 __extension__ extern __inline poly64x2_t
3811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3812 vreinterpretq_p64_s16 (int16x8_t __a)
3813 {
3814   return (poly64x2_t) __a;
3815 }
3816
3817 __extension__ extern __inline poly64x2_t
3818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3819 vreinterpretq_p64_s32 (int32x4_t __a)
3820 {
3821   return (poly64x2_t) __a;
3822 }
3823
3824 __extension__ extern __inline poly64x2_t
3825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3826 vreinterpretq_p64_s64 (int64x2_t __a)
3827 {
3828   return (poly64x2_t) __a;
3829 }
3830
3831 __extension__ extern __inline poly64x2_t
3832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3833 vreinterpretq_p64_f16 (float16x8_t __a)
3834 {
3835   return (poly64x2_t) __a;
3836 }
3837
3838 __extension__ extern __inline poly64x2_t
3839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3840 vreinterpretq_p64_f32 (float32x4_t __a)
3841 {
3842   return (poly64x2_t) __a;
3843 }
3844
3845 __extension__ extern __inline poly64x2_t
3846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3847 vreinterpretq_p64_p128 (poly128_t __a)
3848 {
3849   return (poly64x2_t)__a;
3850 }
3851
3852 __extension__ extern __inline poly64x2_t
3853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3854 vreinterpretq_p64_u8 (uint8x16_t __a)
3855 {
3856   return (poly64x2_t) __a;
3857 }
3858
3859 __extension__ extern __inline poly64x2_t
3860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3861 vreinterpretq_p64_u16 (uint16x8_t __a)
3862 {
3863   return (poly64x2_t) __a;
3864 }
3865
3866 __extension__ extern __inline poly64x2_t
3867 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3868 vreinterpretq_p64_p16 (poly16x8_t __a)
3869 {
3870   return (poly64x2_t)__a;
3871 }
3872
3873 __extension__ extern __inline poly64x2_t
3874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3875 vreinterpretq_p64_u32 (uint32x4_t __a)
3876 {
3877   return (poly64x2_t) __a;
3878 }
3879
3880 __extension__ extern __inline poly64x2_t
3881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3882 vreinterpretq_p64_u64 (uint64x2_t __a)
3883 {
3884   return (poly64x2_t) __a;
3885 }
3886
3887 __extension__ extern __inline poly64x2_t
3888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3889 vreinterpretq_p64_p8 (poly8x16_t __a)
3890 {
3891   return (poly64x2_t) __a;
3892 }
3893
3894 __extension__ extern __inline poly128_t
3895 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3896 vreinterpretq_p128_p8 (poly8x16_t __a)
3897 {
3898   return (poly128_t)__a;
3899 }
3900
3901 __extension__ extern __inline poly128_t
3902 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3903 vreinterpretq_p128_p16 (poly16x8_t __a)
3904 {
3905   return (poly128_t)__a;
3906 }
3907
3908 __extension__ extern __inline poly128_t
3909 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3910 vreinterpretq_p128_f16 (float16x8_t __a)
3911 {
3912   return (poly128_t) __a;
3913 }
3914
3915 __extension__ extern __inline poly128_t
3916 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3917 vreinterpretq_p128_f32 (float32x4_t __a)
3918 {
3919   return (poly128_t)__a;
3920 }
3921
3922 __extension__ extern __inline poly128_t
3923 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3924 vreinterpretq_p128_p64 (poly64x2_t __a)
3925 {
3926   return (poly128_t)__a;
3927 }
3928
3929 __extension__ extern __inline poly128_t
3930 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3931 vreinterpretq_p128_s64 (int64x2_t __a)
3932 {
3933   return (poly128_t)__a;
3934 }
3935
3936 __extension__ extern __inline poly128_t
3937 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3938 vreinterpretq_p128_u64 (uint64x2_t __a)
3939 {
3940   return (poly128_t)__a;
3941 }
3942
3943 __extension__ extern __inline poly128_t
3944 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3945 vreinterpretq_p128_s8 (int8x16_t __a)
3946 {
3947   return (poly128_t)__a;
3948 }
3949
3950 __extension__ extern __inline poly128_t
3951 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3952 vreinterpretq_p128_s16 (int16x8_t __a)
3953 {
3954   return (poly128_t)__a;
3955 }
3956
3957 __extension__ extern __inline poly128_t
3958 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3959 vreinterpretq_p128_s32 (int32x4_t __a)
3960 {
3961   return (poly128_t)__a;
3962 }
3963
3964 __extension__ extern __inline poly128_t
3965 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3966 vreinterpretq_p128_u8 (uint8x16_t __a)
3967 {
3968   return (poly128_t)__a;
3969 }
3970
3971 __extension__ extern __inline poly128_t
3972 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3973 vreinterpretq_p128_u16 (uint16x8_t __a)
3974 {
3975   return (poly128_t)__a;
3976 }
3977
3978 __extension__ extern __inline poly128_t
3979 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
3980 vreinterpretq_p128_u32 (uint32x4_t __a)
3981 {
3982   return (poly128_t)__a;
3983 }
3984
3985 __extension__ extern __inline float16x4_t
3986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3987 vreinterpret_f16_f64 (float64x1_t __a)
3988 {
3989   return (float16x4_t) __a;
3990 }
3991
3992 __extension__ extern __inline float16x4_t
3993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3994 vreinterpret_f16_s8 (int8x8_t __a)
3995 {
3996   return (float16x4_t) __a;
3997 }
3998
3999 __extension__ extern __inline float16x4_t
4000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4001 vreinterpret_f16_s16 (int16x4_t __a)
4002 {
4003   return (float16x4_t) __a;
4004 }
4005
4006 __extension__ extern __inline float16x4_t
4007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4008 vreinterpret_f16_s32 (int32x2_t __a)
4009 {
4010   return (float16x4_t) __a;
4011 }
4012
4013 __extension__ extern __inline float16x4_t
4014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4015 vreinterpret_f16_s64 (int64x1_t __a)
4016 {
4017   return (float16x4_t) __a;
4018 }
4019
4020 __extension__ extern __inline float16x4_t
4021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4022 vreinterpret_f16_f32 (float32x2_t __a)
4023 {
4024   return (float16x4_t) __a;
4025 }
4026
4027 __extension__ extern __inline float16x4_t
4028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4029 vreinterpret_f16_u8 (uint8x8_t __a)
4030 {
4031   return (float16x4_t) __a;
4032 }
4033
4034 __extension__ extern __inline float16x4_t
4035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4036 vreinterpret_f16_u16 (uint16x4_t __a)
4037 {
4038   return (float16x4_t) __a;
4039 }
4040
4041 __extension__ extern __inline float16x4_t
4042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4043 vreinterpret_f16_u32 (uint32x2_t __a)
4044 {
4045   return (float16x4_t) __a;
4046 }
4047
4048 __extension__ extern __inline float16x4_t
4049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4050 vreinterpret_f16_u64 (uint64x1_t __a)
4051 {
4052   return (float16x4_t) __a;
4053 }
4054
4055 __extension__ extern __inline float16x4_t
4056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4057 vreinterpret_f16_p8 (poly8x8_t __a)
4058 {
4059   return (float16x4_t) __a;
4060 }
4061
4062 __extension__ extern __inline float16x4_t
4063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4064 vreinterpret_f16_p16 (poly16x4_t __a)
4065 {
4066   return (float16x4_t) __a;
4067 }
4068
4069 __extension__ extern __inline float16x4_t
4070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4071 vreinterpret_f16_p64 (poly64x1_t __a)
4072 {
4073   return (float16x4_t) __a;
4074 }
4075
4076 __extension__ extern __inline float16x8_t
4077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4078 vreinterpretq_f16_f64 (float64x2_t __a)
4079 {
4080   return (float16x8_t) __a;
4081 }
4082
4083 __extension__ extern __inline float16x8_t
4084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4085 vreinterpretq_f16_s8 (int8x16_t __a)
4086 {
4087   return (float16x8_t) __a;
4088 }
4089
4090 __extension__ extern __inline float16x8_t
4091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4092 vreinterpretq_f16_s16 (int16x8_t __a)
4093 {
4094   return (float16x8_t) __a;
4095 }
4096
4097 __extension__ extern __inline float16x8_t
4098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4099 vreinterpretq_f16_s32 (int32x4_t __a)
4100 {
4101   return (float16x8_t) __a;
4102 }
4103
4104 __extension__ extern __inline float16x8_t
4105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4106 vreinterpretq_f16_s64 (int64x2_t __a)
4107 {
4108   return (float16x8_t) __a;
4109 }
4110
4111 __extension__ extern __inline float16x8_t
4112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4113 vreinterpretq_f16_f32 (float32x4_t __a)
4114 {
4115   return (float16x8_t) __a;
4116 }
4117
4118 __extension__ extern __inline float16x8_t
4119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4120 vreinterpretq_f16_u8 (uint8x16_t __a)
4121 {
4122   return (float16x8_t) __a;
4123 }
4124
4125 __extension__ extern __inline float16x8_t
4126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4127 vreinterpretq_f16_u16 (uint16x8_t __a)
4128 {
4129   return (float16x8_t) __a;
4130 }
4131
4132 __extension__ extern __inline float16x8_t
4133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4134 vreinterpretq_f16_u32 (uint32x4_t __a)
4135 {
4136   return (float16x8_t) __a;
4137 }
4138
4139 __extension__ extern __inline float16x8_t
4140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4141 vreinterpretq_f16_u64 (uint64x2_t __a)
4142 {
4143   return (float16x8_t) __a;
4144 }
4145
4146 __extension__ extern __inline float16x8_t
4147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4148 vreinterpretq_f16_p8 (poly8x16_t __a)
4149 {
4150   return (float16x8_t) __a;
4151 }
4152
4153 __extension__ extern __inline float16x8_t
4154 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
4155 vreinterpretq_f16_p128 (poly128_t __a)
4156 {
4157   return (float16x8_t) __a;
4158 }
4159
4160 __extension__ extern __inline float16x8_t
4161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4162 vreinterpretq_f16_p16 (poly16x8_t __a)
4163 {
4164   return (float16x8_t) __a;
4165 }
4166
4167 __extension__ extern __inline float16x8_t
4168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4169 vreinterpretq_f16_p64 (poly64x2_t __a)
4170 {
4171   return (float16x8_t) __a;
4172 }
4173
4174 __extension__ extern __inline float32x2_t
4175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4176 vreinterpret_f32_f16 (float16x4_t __a)
4177 {
4178   return (float32x2_t) __a;
4179 }
4180
4181 __extension__ extern __inline float32x2_t
4182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4183 vreinterpret_f32_f64 (float64x1_t __a)
4184 {
4185   return (float32x2_t) __a;
4186 }
4187
4188 __extension__ extern __inline float32x2_t
4189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4190 vreinterpret_f32_s8 (int8x8_t __a)
4191 {
4192   return (float32x2_t) __a;
4193 }
4194
4195 __extension__ extern __inline float32x2_t
4196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4197 vreinterpret_f32_s16 (int16x4_t __a)
4198 {
4199   return (float32x2_t) __a;
4200 }
4201
4202 __extension__ extern __inline float32x2_t
4203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4204 vreinterpret_f32_s32 (int32x2_t __a)
4205 {
4206   return (float32x2_t) __a;
4207 }
4208
4209 __extension__ extern __inline float32x2_t
4210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4211 vreinterpret_f32_s64 (int64x1_t __a)
4212 {
4213   return (float32x2_t) __a;
4214 }
4215
4216 __extension__ extern __inline float32x2_t
4217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4218 vreinterpret_f32_u8 (uint8x8_t __a)
4219 {
4220   return (float32x2_t) __a;
4221 }
4222
4223 __extension__ extern __inline float32x2_t
4224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4225 vreinterpret_f32_u16 (uint16x4_t __a)
4226 {
4227   return (float32x2_t) __a;
4228 }
4229
4230 __extension__ extern __inline float32x2_t
4231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4232 vreinterpret_f32_u32 (uint32x2_t __a)
4233 {
4234   return (float32x2_t) __a;
4235 }
4236
4237 __extension__ extern __inline float32x2_t
4238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4239 vreinterpret_f32_u64 (uint64x1_t __a)
4240 {
4241   return (float32x2_t) __a;
4242 }
4243
4244 __extension__ extern __inline float32x2_t
4245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4246 vreinterpret_f32_p8 (poly8x8_t __a)
4247 {
4248   return (float32x2_t) __a;
4249 }
4250
4251 __extension__ extern __inline float32x2_t
4252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4253 vreinterpret_f32_p16 (poly16x4_t __a)
4254 {
4255   return (float32x2_t) __a;
4256 }
4257
4258 __extension__ extern __inline float32x2_t
4259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4260 vreinterpret_f32_p64 (poly64x1_t __a)
4261 {
4262   return (float32x2_t) __a;
4263 }
4264
4265 __extension__ extern __inline float32x4_t
4266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4267 vreinterpretq_f32_f16 (float16x8_t __a)
4268 {
4269   return (float32x4_t) __a;
4270 }
4271
4272 __extension__ extern __inline float32x4_t
4273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4274 vreinterpretq_f32_f64 (float64x2_t __a)
4275 {
4276   return (float32x4_t) __a;
4277 }
4278
4279 __extension__ extern __inline float32x4_t
4280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4281 vreinterpretq_f32_s8 (int8x16_t __a)
4282 {
4283   return (float32x4_t) __a;
4284 }
4285
4286 __extension__ extern __inline float32x4_t
4287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4288 vreinterpretq_f32_s16 (int16x8_t __a)
4289 {
4290   return (float32x4_t) __a;
4291 }
4292
4293 __extension__ extern __inline float32x4_t
4294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4295 vreinterpretq_f32_s32 (int32x4_t __a)
4296 {
4297   return (float32x4_t) __a;
4298 }
4299
4300 __extension__ extern __inline float32x4_t
4301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4302 vreinterpretq_f32_s64 (int64x2_t __a)
4303 {
4304   return (float32x4_t) __a;
4305 }
4306
4307 __extension__ extern __inline float32x4_t
4308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4309 vreinterpretq_f32_u8 (uint8x16_t __a)
4310 {
4311   return (float32x4_t) __a;
4312 }
4313
4314 __extension__ extern __inline float32x4_t
4315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4316 vreinterpretq_f32_u16 (uint16x8_t __a)
4317 {
4318   return (float32x4_t) __a;
4319 }
4320
4321 __extension__ extern __inline float32x4_t
4322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4323 vreinterpretq_f32_u32 (uint32x4_t __a)
4324 {
4325   return (float32x4_t) __a;
4326 }
4327
4328 __extension__ extern __inline float32x4_t
4329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4330 vreinterpretq_f32_u64 (uint64x2_t __a)
4331 {
4332   return (float32x4_t) __a;
4333 }
4334
4335 __extension__ extern __inline float32x4_t
4336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4337 vreinterpretq_f32_p8 (poly8x16_t __a)
4338 {
4339   return (float32x4_t) __a;
4340 }
4341
4342 __extension__ extern __inline float32x4_t
4343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4344 vreinterpretq_f32_p16 (poly16x8_t __a)
4345 {
4346   return (float32x4_t) __a;
4347 }
4348
4349 __extension__ extern __inline float32x4_t
4350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4351 vreinterpretq_f32_p64 (poly64x2_t __a)
4352 {
4353   return (float32x4_t) __a;
4354 }
4355
4356 __extension__ extern __inline float32x4_t
4357 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
4358 vreinterpretq_f32_p128 (poly128_t __a)
4359 {
4360   return (float32x4_t)__a;
4361 }
4362
4363
4364 __extension__ extern __inline float64x1_t
4365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4366 vreinterpret_f64_f16 (float16x4_t __a)
4367 {
4368   return (float64x1_t) __a;
4369 }
4370
4371 __extension__ extern __inline float64x1_t
4372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4373 vreinterpret_f64_f32 (float32x2_t __a)
4374 {
4375   return (float64x1_t) __a;
4376 }
4377
4378 __extension__ extern __inline float64x1_t
4379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4380 vreinterpret_f64_p8 (poly8x8_t __a)
4381 {
4382   return (float64x1_t) __a;
4383 }
4384
4385 __extension__ extern __inline float64x1_t
4386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4387 vreinterpret_f64_p16 (poly16x4_t __a)
4388 {
4389   return (float64x1_t) __a;
4390 }
4391
4392 __extension__ extern __inline float64x1_t
4393 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4394 vreinterpret_f64_p64 (poly64x1_t __a)
4395 {
4396   return (float64x1_t) __a;
4397 }
4398
4399 __extension__ extern __inline float64x1_t
4400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4401 vreinterpret_f64_s8 (int8x8_t __a)
4402 {
4403   return (float64x1_t) __a;
4404 }
4405
4406 __extension__ extern __inline float64x1_t
4407 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4408 vreinterpret_f64_s16 (int16x4_t __a)
4409 {
4410   return (float64x1_t) __a;
4411 }
4412
4413 __extension__ extern __inline float64x1_t
4414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4415 vreinterpret_f64_s32 (int32x2_t __a)
4416 {
4417   return (float64x1_t) __a;
4418 }
4419
4420 __extension__ extern __inline float64x1_t
4421 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4422 vreinterpret_f64_s64 (int64x1_t __a)
4423 {
4424   return (float64x1_t) __a;
4425 }
4426
4427 __extension__ extern __inline float64x1_t
4428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4429 vreinterpret_f64_u8 (uint8x8_t __a)
4430 {
4431   return (float64x1_t) __a;
4432 }
4433
4434 __extension__ extern __inline float64x1_t
4435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4436 vreinterpret_f64_u16 (uint16x4_t __a)
4437 {
4438   return (float64x1_t) __a;
4439 }
4440
4441 __extension__ extern __inline float64x1_t
4442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4443 vreinterpret_f64_u32 (uint32x2_t __a)
4444 {
4445   return (float64x1_t) __a;
4446 }
4447
4448 __extension__ extern __inline float64x1_t
4449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4450 vreinterpret_f64_u64 (uint64x1_t __a)
4451 {
4452   return (float64x1_t) __a;
4453 }
4454
4455 __extension__ extern __inline float64x2_t
4456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4457 vreinterpretq_f64_f16 (float16x8_t __a)
4458 {
4459   return (float64x2_t) __a;
4460 }
4461
4462 __extension__ extern __inline float64x2_t
4463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4464 vreinterpretq_f64_f32 (float32x4_t __a)
4465 {
4466   return (float64x2_t) __a;
4467 }
4468
4469 __extension__ extern __inline float64x2_t
4470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4471 vreinterpretq_f64_p8 (poly8x16_t __a)
4472 {
4473   return (float64x2_t) __a;
4474 }
4475
4476 __extension__ extern __inline float64x2_t
4477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4478 vreinterpretq_f64_p16 (poly16x8_t __a)
4479 {
4480   return (float64x2_t) __a;
4481 }
4482
4483 __extension__ extern __inline float64x2_t
4484 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4485 vreinterpretq_f64_p64 (poly64x2_t __a)
4486 {
4487   return (float64x2_t) __a;
4488 }
4489
4490 __extension__ extern __inline float64x2_t
4491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4492 vreinterpretq_f64_s8 (int8x16_t __a)
4493 {
4494   return (float64x2_t) __a;
4495 }
4496
4497 __extension__ extern __inline float64x2_t
4498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4499 vreinterpretq_f64_s16 (int16x8_t __a)
4500 {
4501   return (float64x2_t) __a;
4502 }
4503
4504 __extension__ extern __inline float64x2_t
4505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4506 vreinterpretq_f64_s32 (int32x4_t __a)
4507 {
4508   return (float64x2_t) __a;
4509 }
4510
4511 __extension__ extern __inline float64x2_t
4512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4513 vreinterpretq_f64_s64 (int64x2_t __a)
4514 {
4515   return (float64x2_t) __a;
4516 }
4517
4518 __extension__ extern __inline float64x2_t
4519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4520 vreinterpretq_f64_u8 (uint8x16_t __a)
4521 {
4522   return (float64x2_t) __a;
4523 }
4524
4525 __extension__ extern __inline float64x2_t
4526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4527 vreinterpretq_f64_u16 (uint16x8_t __a)
4528 {
4529   return (float64x2_t) __a;
4530 }
4531
4532 __extension__ extern __inline float64x2_t
4533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4534 vreinterpretq_f64_u32 (uint32x4_t __a)
4535 {
4536   return (float64x2_t) __a;
4537 }
4538
4539 __extension__ extern __inline float64x2_t
4540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4541 vreinterpretq_f64_u64 (uint64x2_t __a)
4542 {
4543   return (float64x2_t) __a;
4544 }
4545
4546 __extension__ extern __inline int64x1_t
4547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4548 vreinterpret_s64_f16 (float16x4_t __a)
4549 {
4550   return (int64x1_t) __a;
4551 }
4552
4553 __extension__ extern __inline int64x1_t
4554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4555 vreinterpret_s64_f64 (float64x1_t __a)
4556 {
4557   return (int64x1_t) __a;
4558 }
4559
4560 __extension__ extern __inline int64x1_t
4561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4562 vreinterpret_s64_s8 (int8x8_t __a)
4563 {
4564   return (int64x1_t) __a;
4565 }
4566
4567 __extension__ extern __inline int64x1_t
4568 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4569 vreinterpret_s64_s16 (int16x4_t __a)
4570 {
4571   return (int64x1_t) __a;
4572 }
4573
4574 __extension__ extern __inline int64x1_t
4575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4576 vreinterpret_s64_s32 (int32x2_t __a)
4577 {
4578   return (int64x1_t) __a;
4579 }
4580
4581 __extension__ extern __inline int64x1_t
4582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4583 vreinterpret_s64_f32 (float32x2_t __a)
4584 {
4585   return (int64x1_t) __a;
4586 }
4587
4588 __extension__ extern __inline int64x1_t
4589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4590 vreinterpret_s64_u8 (uint8x8_t __a)
4591 {
4592   return (int64x1_t) __a;
4593 }
4594
4595 __extension__ extern __inline int64x1_t
4596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4597 vreinterpret_s64_u16 (uint16x4_t __a)
4598 {
4599   return (int64x1_t) __a;
4600 }
4601
4602 __extension__ extern __inline int64x1_t
4603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4604 vreinterpret_s64_u32 (uint32x2_t __a)
4605 {
4606   return (int64x1_t) __a;
4607 }
4608
4609 __extension__ extern __inline int64x1_t
4610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4611 vreinterpret_s64_u64 (uint64x1_t __a)
4612 {
4613   return (int64x1_t) __a;
4614 }
4615
4616 __extension__ extern __inline int64x1_t
4617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4618 vreinterpret_s64_p8 (poly8x8_t __a)
4619 {
4620   return (int64x1_t) __a;
4621 }
4622
4623 __extension__ extern __inline int64x1_t
4624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4625 vreinterpret_s64_p16 (poly16x4_t __a)
4626 {
4627   return (int64x1_t) __a;
4628 }
4629
4630 __extension__ extern __inline int64x1_t
4631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4632 vreinterpret_s64_p64 (poly64x1_t __a)
4633 {
4634   return (int64x1_t) __a;
4635 }
4636
4637 __extension__ extern __inline int64x2_t
4638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4639 vreinterpretq_s64_f64 (float64x2_t __a)
4640 {
4641   return (int64x2_t) __a;
4642 }
4643
4644 __extension__ extern __inline int64x2_t
4645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4646 vreinterpretq_s64_s8 (int8x16_t __a)
4647 {
4648   return (int64x2_t) __a;
4649 }
4650
4651 __extension__ extern __inline int64x2_t
4652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4653 vreinterpretq_s64_s16 (int16x8_t __a)
4654 {
4655   return (int64x2_t) __a;
4656 }
4657
4658 __extension__ extern __inline int64x2_t
4659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4660 vreinterpretq_s64_s32 (int32x4_t __a)
4661 {
4662   return (int64x2_t) __a;
4663 }
4664
4665 __extension__ extern __inline int64x2_t
4666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4667 vreinterpretq_s64_f16 (float16x8_t __a)
4668 {
4669   return (int64x2_t) __a;
4670 }
4671
4672 __extension__ extern __inline int64x2_t
4673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4674 vreinterpretq_s64_f32 (float32x4_t __a)
4675 {
4676   return (int64x2_t) __a;
4677 }
4678
4679 __extension__ extern __inline int64x2_t
4680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4681 vreinterpretq_s64_u8 (uint8x16_t __a)
4682 {
4683   return (int64x2_t) __a;
4684 }
4685
4686 __extension__ extern __inline int64x2_t
4687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4688 vreinterpretq_s64_u16 (uint16x8_t __a)
4689 {
4690   return (int64x2_t) __a;
4691 }
4692
4693 __extension__ extern __inline int64x2_t
4694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4695 vreinterpretq_s64_u32 (uint32x4_t __a)
4696 {
4697   return (int64x2_t) __a;
4698 }
4699
4700 __extension__ extern __inline int64x2_t
4701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4702 vreinterpretq_s64_u64 (uint64x2_t __a)
4703 {
4704   return (int64x2_t) __a;
4705 }
4706
4707 __extension__ extern __inline int64x2_t
4708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4709 vreinterpretq_s64_p8 (poly8x16_t __a)
4710 {
4711   return (int64x2_t) __a;
4712 }
4713
4714 __extension__ extern __inline int64x2_t
4715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4716 vreinterpretq_s64_p16 (poly16x8_t __a)
4717 {
4718   return (int64x2_t) __a;
4719 }
4720
4721 __extension__ extern __inline int64x2_t
4722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4723 vreinterpretq_s64_p64 (poly64x2_t __a)
4724 {
4725   return (int64x2_t) __a;
4726 }
4727
4728 __extension__ extern __inline int64x2_t
4729 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
4730 vreinterpretq_s64_p128 (poly128_t __a)
4731 {
4732   return (int64x2_t)__a;
4733 }
4734
4735 __extension__ extern __inline uint64x1_t
4736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4737 vreinterpret_u64_f16 (float16x4_t __a)
4738 {
4739   return (uint64x1_t) __a;
4740 }
4741
4742 __extension__ extern __inline uint64x1_t
4743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4744 vreinterpret_u64_f64 (float64x1_t __a)
4745 {
4746   return (uint64x1_t) __a;
4747 }
4748
4749 __extension__ extern __inline uint64x1_t
4750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4751 vreinterpret_u64_s8 (int8x8_t __a)
4752 {
4753   return (uint64x1_t) __a;
4754 }
4755
4756 __extension__ extern __inline uint64x1_t
4757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4758 vreinterpret_u64_s16 (int16x4_t __a)
4759 {
4760   return (uint64x1_t) __a;
4761 }
4762
4763 __extension__ extern __inline uint64x1_t
4764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4765 vreinterpret_u64_s32 (int32x2_t __a)
4766 {
4767   return (uint64x1_t) __a;
4768 }
4769
4770 __extension__ extern __inline uint64x1_t
4771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4772 vreinterpret_u64_s64 (int64x1_t __a)
4773 {
4774   return (uint64x1_t) __a;
4775 }
4776
4777 __extension__ extern __inline uint64x1_t
4778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4779 vreinterpret_u64_f32 (float32x2_t __a)
4780 {
4781   return (uint64x1_t) __a;
4782 }
4783
4784 __extension__ extern __inline uint64x1_t
4785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4786 vreinterpret_u64_u8 (uint8x8_t __a)
4787 {
4788   return (uint64x1_t) __a;
4789 }
4790
4791 __extension__ extern __inline uint64x1_t
4792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4793 vreinterpret_u64_u16 (uint16x4_t __a)
4794 {
4795   return (uint64x1_t) __a;
4796 }
4797
4798 __extension__ extern __inline uint64x1_t
4799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4800 vreinterpret_u64_u32 (uint32x2_t __a)
4801 {
4802   return (uint64x1_t) __a;
4803 }
4804
4805 __extension__ extern __inline uint64x1_t
4806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4807 vreinterpret_u64_p8 (poly8x8_t __a)
4808 {
4809   return (uint64x1_t) __a;
4810 }
4811
4812 __extension__ extern __inline uint64x1_t
4813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4814 vreinterpret_u64_p16 (poly16x4_t __a)
4815 {
4816   return (uint64x1_t) __a;
4817 }
4818
4819 __extension__ extern __inline uint64x1_t
4820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4821 vreinterpret_u64_p64 (poly64x1_t __a)
4822 {
4823   return (uint64x1_t) __a;
4824 }
4825
4826 __extension__ extern __inline uint64x2_t
4827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4828 vreinterpretq_u64_f64 (float64x2_t __a)
4829 {
4830   return (uint64x2_t) __a;
4831 }
4832
4833 __extension__ extern __inline uint64x2_t
4834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4835 vreinterpretq_u64_s8 (int8x16_t __a)
4836 {
4837   return (uint64x2_t) __a;
4838 }
4839
4840 __extension__ extern __inline uint64x2_t
4841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4842 vreinterpretq_u64_s16 (int16x8_t __a)
4843 {
4844   return (uint64x2_t) __a;
4845 }
4846
4847 __extension__ extern __inline uint64x2_t
4848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4849 vreinterpretq_u64_s32 (int32x4_t __a)
4850 {
4851   return (uint64x2_t) __a;
4852 }
4853
4854 __extension__ extern __inline uint64x2_t
4855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4856 vreinterpretq_u64_s64 (int64x2_t __a)
4857 {
4858   return (uint64x2_t) __a;
4859 }
4860
4861 __extension__ extern __inline uint64x2_t
4862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4863 vreinterpretq_u64_f16 (float16x8_t __a)
4864 {
4865   return (uint64x2_t) __a;
4866 }
4867
4868 __extension__ extern __inline uint64x2_t
4869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4870 vreinterpretq_u64_f32 (float32x4_t __a)
4871 {
4872   return (uint64x2_t) __a;
4873 }
4874
4875 __extension__ extern __inline uint64x2_t
4876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4877 vreinterpretq_u64_u8 (uint8x16_t __a)
4878 {
4879   return (uint64x2_t) __a;
4880 }
4881
4882 __extension__ extern __inline uint64x2_t
4883 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4884 vreinterpretq_u64_u16 (uint16x8_t __a)
4885 {
4886   return (uint64x2_t) __a;
4887 }
4888
4889 __extension__ extern __inline uint64x2_t
4890 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4891 vreinterpretq_u64_u32 (uint32x4_t __a)
4892 {
4893   return (uint64x2_t) __a;
4894 }
4895
4896 __extension__ extern __inline uint64x2_t
4897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4898 vreinterpretq_u64_p8 (poly8x16_t __a)
4899 {
4900   return (uint64x2_t) __a;
4901 }
4902
4903 __extension__ extern __inline uint64x2_t
4904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4905 vreinterpretq_u64_p16 (poly16x8_t __a)
4906 {
4907   return (uint64x2_t) __a;
4908 }
4909
4910 __extension__ extern __inline uint64x2_t
4911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4912 vreinterpretq_u64_p64 (poly64x2_t __a)
4913 {
4914   return (uint64x2_t) __a;
4915 }
4916
4917 __extension__ extern __inline uint64x2_t
4918 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
4919 vreinterpretq_u64_p128 (poly128_t __a)
4920 {
4921   return (uint64x2_t)__a;
4922 }
4923
4924 __extension__ extern __inline int8x8_t
4925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4926 vreinterpret_s8_f16 (float16x4_t __a)
4927 {
4928   return (int8x8_t) __a;
4929 }
4930
4931 __extension__ extern __inline int8x8_t
4932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4933 vreinterpret_s8_f64 (float64x1_t __a)
4934 {
4935   return (int8x8_t) __a;
4936 }
4937
4938 __extension__ extern __inline int8x8_t
4939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4940 vreinterpret_s8_s16 (int16x4_t __a)
4941 {
4942   return (int8x8_t) __a;
4943 }
4944
4945 __extension__ extern __inline int8x8_t
4946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4947 vreinterpret_s8_s32 (int32x2_t __a)
4948 {
4949   return (int8x8_t) __a;
4950 }
4951
4952 __extension__ extern __inline int8x8_t
4953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4954 vreinterpret_s8_s64 (int64x1_t __a)
4955 {
4956   return (int8x8_t) __a;
4957 }
4958
4959 __extension__ extern __inline int8x8_t
4960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4961 vreinterpret_s8_f32 (float32x2_t __a)
4962 {
4963   return (int8x8_t) __a;
4964 }
4965
4966 __extension__ extern __inline int8x8_t
4967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4968 vreinterpret_s8_u8 (uint8x8_t __a)
4969 {
4970   return (int8x8_t) __a;
4971 }
4972
4973 __extension__ extern __inline int8x8_t
4974 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4975 vreinterpret_s8_u16 (uint16x4_t __a)
4976 {
4977   return (int8x8_t) __a;
4978 }
4979
4980 __extension__ extern __inline int8x8_t
4981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4982 vreinterpret_s8_u32 (uint32x2_t __a)
4983 {
4984   return (int8x8_t) __a;
4985 }
4986
4987 __extension__ extern __inline int8x8_t
4988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4989 vreinterpret_s8_u64 (uint64x1_t __a)
4990 {
4991   return (int8x8_t) __a;
4992 }
4993
4994 __extension__ extern __inline int8x8_t
4995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4996 vreinterpret_s8_p8 (poly8x8_t __a)
4997 {
4998   return (int8x8_t) __a;
4999 }
5000
5001 __extension__ extern __inline int8x8_t
5002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5003 vreinterpret_s8_p16 (poly16x4_t __a)
5004 {
5005   return (int8x8_t) __a;
5006 }
5007
5008 __extension__ extern __inline int8x8_t
5009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5010 vreinterpret_s8_p64 (poly64x1_t __a)
5011 {
5012   return (int8x8_t) __a;
5013 }
5014
5015 __extension__ extern __inline int8x16_t
5016 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5017 vreinterpretq_s8_f64 (float64x2_t __a)
5018 {
5019   return (int8x16_t) __a;
5020 }
5021
5022 __extension__ extern __inline int8x16_t
5023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5024 vreinterpretq_s8_s16 (int16x8_t __a)
5025 {
5026   return (int8x16_t) __a;
5027 }
5028
5029 __extension__ extern __inline int8x16_t
5030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5031 vreinterpretq_s8_s32 (int32x4_t __a)
5032 {
5033   return (int8x16_t) __a;
5034 }
5035
5036 __extension__ extern __inline int8x16_t
5037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5038 vreinterpretq_s8_s64 (int64x2_t __a)
5039 {
5040   return (int8x16_t) __a;
5041 }
5042
5043 __extension__ extern __inline int8x16_t
5044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5045 vreinterpretq_s8_f16 (float16x8_t __a)
5046 {
5047   return (int8x16_t) __a;
5048 }
5049
5050 __extension__ extern __inline int8x16_t
5051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5052 vreinterpretq_s8_f32 (float32x4_t __a)
5053 {
5054   return (int8x16_t) __a;
5055 }
5056
5057 __extension__ extern __inline int8x16_t
5058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5059 vreinterpretq_s8_u8 (uint8x16_t __a)
5060 {
5061   return (int8x16_t) __a;
5062 }
5063
5064 __extension__ extern __inline int8x16_t
5065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5066 vreinterpretq_s8_u16 (uint16x8_t __a)
5067 {
5068   return (int8x16_t) __a;
5069 }
5070
5071 __extension__ extern __inline int8x16_t
5072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5073 vreinterpretq_s8_u32 (uint32x4_t __a)
5074 {
5075   return (int8x16_t) __a;
5076 }
5077
5078 __extension__ extern __inline int8x16_t
5079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5080 vreinterpretq_s8_u64 (uint64x2_t __a)
5081 {
5082   return (int8x16_t) __a;
5083 }
5084
5085 __extension__ extern __inline int8x16_t
5086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5087 vreinterpretq_s8_p8 (poly8x16_t __a)
5088 {
5089   return (int8x16_t) __a;
5090 }
5091
5092 __extension__ extern __inline int8x16_t
5093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5094 vreinterpretq_s8_p16 (poly16x8_t __a)
5095 {
5096   return (int8x16_t) __a;
5097 }
5098
5099 __extension__ extern __inline int8x16_t
5100 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5101 vreinterpretq_s8_p64 (poly64x2_t __a)
5102 {
5103   return (int8x16_t) __a;
5104 }
5105
5106 __extension__ extern __inline int8x16_t
5107 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
5108 vreinterpretq_s8_p128 (poly128_t __a)
5109 {
5110   return (int8x16_t)__a;
5111 }
5112
5113 __extension__ extern __inline int16x4_t
5114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5115 vreinterpret_s16_f16 (float16x4_t __a)
5116 {
5117   return (int16x4_t) __a;
5118 }
5119
5120 __extension__ extern __inline int16x4_t
5121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5122 vreinterpret_s16_f64 (float64x1_t __a)
5123 {
5124   return (int16x4_t) __a;
5125 }
5126
5127 __extension__ extern __inline int16x4_t
5128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5129 vreinterpret_s16_s8 (int8x8_t __a)
5130 {
5131   return (int16x4_t) __a;
5132 }
5133
5134 __extension__ extern __inline int16x4_t
5135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5136 vreinterpret_s16_s32 (int32x2_t __a)
5137 {
5138   return (int16x4_t) __a;
5139 }
5140
5141 __extension__ extern __inline int16x4_t
5142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5143 vreinterpret_s16_s64 (int64x1_t __a)
5144 {
5145   return (int16x4_t) __a;
5146 }
5147
5148 __extension__ extern __inline int16x4_t
5149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5150 vreinterpret_s16_f32 (float32x2_t __a)
5151 {
5152   return (int16x4_t) __a;
5153 }
5154
5155 __extension__ extern __inline int16x4_t
5156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5157 vreinterpret_s16_u8 (uint8x8_t __a)
5158 {
5159   return (int16x4_t) __a;
5160 }
5161
5162 __extension__ extern __inline int16x4_t
5163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5164 vreinterpret_s16_u16 (uint16x4_t __a)
5165 {
5166   return (int16x4_t) __a;
5167 }
5168
5169 __extension__ extern __inline int16x4_t
5170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5171 vreinterpret_s16_u32 (uint32x2_t __a)
5172 {
5173   return (int16x4_t) __a;
5174 }
5175
5176 __extension__ extern __inline int16x4_t
5177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5178 vreinterpret_s16_u64 (uint64x1_t __a)
5179 {
5180   return (int16x4_t) __a;
5181 }
5182
5183 __extension__ extern __inline int16x4_t
5184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5185 vreinterpret_s16_p8 (poly8x8_t __a)
5186 {
5187   return (int16x4_t) __a;
5188 }
5189
5190 __extension__ extern __inline int16x4_t
5191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5192 vreinterpret_s16_p16 (poly16x4_t __a)
5193 {
5194   return (int16x4_t) __a;
5195 }
5196
5197 __extension__ extern __inline int16x4_t
5198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5199 vreinterpret_s16_p64 (poly64x1_t __a)
5200 {
5201   return (int16x4_t) __a;
5202 }
5203
5204 __extension__ extern __inline int16x8_t
5205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5206 vreinterpretq_s16_f64 (float64x2_t __a)
5207 {
5208   return (int16x8_t) __a;
5209 }
5210
5211 __extension__ extern __inline int16x8_t
5212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5213 vreinterpretq_s16_s8 (int8x16_t __a)
5214 {
5215   return (int16x8_t) __a;
5216 }
5217
5218 __extension__ extern __inline int16x8_t
5219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5220 vreinterpretq_s16_s32 (int32x4_t __a)
5221 {
5222   return (int16x8_t) __a;
5223 }
5224
5225 __extension__ extern __inline int16x8_t
5226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5227 vreinterpretq_s16_s64 (int64x2_t __a)
5228 {
5229   return (int16x8_t) __a;
5230 }
5231
5232 __extension__ extern __inline int16x8_t
5233 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5234 vreinterpretq_s16_f16 (float16x8_t __a)
5235 {
5236   return (int16x8_t) __a;
5237 }
5238
5239 __extension__ extern __inline int16x8_t
5240 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5241 vreinterpretq_s16_f32 (float32x4_t __a)
5242 {
5243   return (int16x8_t) __a;
5244 }
5245
5246 __extension__ extern __inline int16x8_t
5247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5248 vreinterpretq_s16_u8 (uint8x16_t __a)
5249 {
5250   return (int16x8_t) __a;
5251 }
5252
5253 __extension__ extern __inline int16x8_t
5254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5255 vreinterpretq_s16_u16 (uint16x8_t __a)
5256 {
5257   return (int16x8_t) __a;
5258 }
5259
5260 __extension__ extern __inline int16x8_t
5261 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5262 vreinterpretq_s16_u32 (uint32x4_t __a)
5263 {
5264   return (int16x8_t) __a;
5265 }
5266
5267 __extension__ extern __inline int16x8_t
5268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5269 vreinterpretq_s16_u64 (uint64x2_t __a)
5270 {
5271   return (int16x8_t) __a;
5272 }
5273
5274 __extension__ extern __inline int16x8_t
5275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5276 vreinterpretq_s16_p8 (poly8x16_t __a)
5277 {
5278   return (int16x8_t) __a;
5279 }
5280
5281 __extension__ extern __inline int16x8_t
5282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5283 vreinterpretq_s16_p16 (poly16x8_t __a)
5284 {
5285   return (int16x8_t) __a;
5286 }
5287
5288 __extension__ extern __inline int16x8_t
5289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5290 vreinterpretq_s16_p64 (poly64x2_t __a)
5291 {
5292   return (int16x8_t) __a;
5293 }
5294
5295 __extension__ extern __inline int16x8_t
5296 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
5297 vreinterpretq_s16_p128 (poly128_t __a)
5298 {
5299   return (int16x8_t)__a;
5300 }
5301
5302 __extension__ extern __inline int32x2_t
5303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5304 vreinterpret_s32_f16 (float16x4_t __a)
5305 {
5306   return (int32x2_t) __a;
5307 }
5308
5309 __extension__ extern __inline int32x2_t
5310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5311 vreinterpret_s32_f64 (float64x1_t __a)
5312 {
5313   return (int32x2_t) __a;
5314 }
5315
5316 __extension__ extern __inline int32x2_t
5317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5318 vreinterpret_s32_s8 (int8x8_t __a)
5319 {
5320   return (int32x2_t) __a;
5321 }
5322
5323 __extension__ extern __inline int32x2_t
5324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5325 vreinterpret_s32_s16 (int16x4_t __a)
5326 {
5327   return (int32x2_t) __a;
5328 }
5329
5330 __extension__ extern __inline int32x2_t
5331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5332 vreinterpret_s32_s64 (int64x1_t __a)
5333 {
5334   return (int32x2_t) __a;
5335 }
5336
5337 __extension__ extern __inline int32x2_t
5338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5339 vreinterpret_s32_f32 (float32x2_t __a)
5340 {
5341   return (int32x2_t) __a;
5342 }
5343
5344 __extension__ extern __inline int32x2_t
5345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5346 vreinterpret_s32_u8 (uint8x8_t __a)
5347 {
5348   return (int32x2_t) __a;
5349 }
5350
5351 __extension__ extern __inline int32x2_t
5352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5353 vreinterpret_s32_u16 (uint16x4_t __a)
5354 {
5355   return (int32x2_t) __a;
5356 }
5357
5358 __extension__ extern __inline int32x2_t
5359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5360 vreinterpret_s32_u32 (uint32x2_t __a)
5361 {
5362   return (int32x2_t) __a;
5363 }
5364
5365 __extension__ extern __inline int32x2_t
5366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5367 vreinterpret_s32_u64 (uint64x1_t __a)
5368 {
5369   return (int32x2_t) __a;
5370 }
5371
5372 __extension__ extern __inline int32x2_t
5373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5374 vreinterpret_s32_p8 (poly8x8_t __a)
5375 {
5376   return (int32x2_t) __a;
5377 }
5378
5379 __extension__ extern __inline int32x2_t
5380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5381 vreinterpret_s32_p16 (poly16x4_t __a)
5382 {
5383   return (int32x2_t) __a;
5384 }
5385
5386 __extension__ extern __inline int32x2_t
5387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5388 vreinterpret_s32_p64 (poly64x1_t __a)
5389 {
5390   return (int32x2_t) __a;
5391 }
5392
5393 __extension__ extern __inline int32x4_t
5394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5395 vreinterpretq_s32_f64 (float64x2_t __a)
5396 {
5397   return (int32x4_t) __a;
5398 }
5399
5400 __extension__ extern __inline int32x4_t
5401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5402 vreinterpretq_s32_s8 (int8x16_t __a)
5403 {
5404   return (int32x4_t) __a;
5405 }
5406
5407 __extension__ extern __inline int32x4_t
5408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5409 vreinterpretq_s32_s16 (int16x8_t __a)
5410 {
5411   return (int32x4_t) __a;
5412 }
5413
5414 __extension__ extern __inline int32x4_t
5415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5416 vreinterpretq_s32_s64 (int64x2_t __a)
5417 {
5418   return (int32x4_t) __a;
5419 }
5420
5421 __extension__ extern __inline int32x4_t
5422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5423 vreinterpretq_s32_f16 (float16x8_t __a)
5424 {
5425   return (int32x4_t) __a;
5426 }
5427
5428 __extension__ extern __inline int32x4_t
5429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5430 vreinterpretq_s32_f32 (float32x4_t __a)
5431 {
5432   return (int32x4_t) __a;
5433 }
5434
5435 __extension__ extern __inline int32x4_t
5436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5437 vreinterpretq_s32_u8 (uint8x16_t __a)
5438 {
5439   return (int32x4_t) __a;
5440 }
5441
5442 __extension__ extern __inline int32x4_t
5443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5444 vreinterpretq_s32_u16 (uint16x8_t __a)
5445 {
5446   return (int32x4_t) __a;
5447 }
5448
5449 __extension__ extern __inline int32x4_t
5450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5451 vreinterpretq_s32_u32 (uint32x4_t __a)
5452 {
5453   return (int32x4_t) __a;
5454 }
5455
5456 __extension__ extern __inline int32x4_t
5457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5458 vreinterpretq_s32_u64 (uint64x2_t __a)
5459 {
5460   return (int32x4_t) __a;
5461 }
5462
5463 __extension__ extern __inline int32x4_t
5464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5465 vreinterpretq_s32_p8 (poly8x16_t __a)
5466 {
5467   return (int32x4_t) __a;
5468 }
5469
5470 __extension__ extern __inline int32x4_t
5471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5472 vreinterpretq_s32_p16 (poly16x8_t __a)
5473 {
5474   return (int32x4_t) __a;
5475 }
5476
5477 __extension__ extern __inline int32x4_t
5478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5479 vreinterpretq_s32_p64 (poly64x2_t __a)
5480 {
5481   return (int32x4_t) __a;
5482 }
5483
5484 __extension__ extern __inline int32x4_t
5485 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
5486 vreinterpretq_s32_p128 (poly128_t __a)
5487 {
5488   return (int32x4_t)__a;
5489 }
5490
5491 __extension__ extern __inline uint8x8_t
5492 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5493 vreinterpret_u8_f16 (float16x4_t __a)
5494 {
5495   return (uint8x8_t) __a;
5496 }
5497
5498 __extension__ extern __inline uint8x8_t
5499 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5500 vreinterpret_u8_f64 (float64x1_t __a)
5501 {
5502   return (uint8x8_t) __a;
5503 }
5504
5505 __extension__ extern __inline uint8x8_t
5506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5507 vreinterpret_u8_s8 (int8x8_t __a)
5508 {
5509   return (uint8x8_t) __a;
5510 }
5511
5512 __extension__ extern __inline uint8x8_t
5513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5514 vreinterpret_u8_s16 (int16x4_t __a)
5515 {
5516   return (uint8x8_t) __a;
5517 }
5518
5519 __extension__ extern __inline uint8x8_t
5520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5521 vreinterpret_u8_s32 (int32x2_t __a)
5522 {
5523   return (uint8x8_t) __a;
5524 }
5525
5526 __extension__ extern __inline uint8x8_t
5527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5528 vreinterpret_u8_s64 (int64x1_t __a)
5529 {
5530   return (uint8x8_t) __a;
5531 }
5532
5533 __extension__ extern __inline uint8x8_t
5534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5535 vreinterpret_u8_f32 (float32x2_t __a)
5536 {
5537   return (uint8x8_t) __a;
5538 }
5539
5540 __extension__ extern __inline uint8x8_t
5541 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5542 vreinterpret_u8_u16 (uint16x4_t __a)
5543 {
5544   return (uint8x8_t) __a;
5545 }
5546
5547 __extension__ extern __inline uint8x8_t
5548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5549 vreinterpret_u8_u32 (uint32x2_t __a)
5550 {
5551   return (uint8x8_t) __a;
5552 }
5553
5554 __extension__ extern __inline uint8x8_t
5555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5556 vreinterpret_u8_u64 (uint64x1_t __a)
5557 {
5558   return (uint8x8_t) __a;
5559 }
5560
5561 __extension__ extern __inline uint8x8_t
5562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5563 vreinterpret_u8_p8 (poly8x8_t __a)
5564 {
5565   return (uint8x8_t) __a;
5566 }
5567
5568 __extension__ extern __inline uint8x8_t
5569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5570 vreinterpret_u8_p16 (poly16x4_t __a)
5571 {
5572   return (uint8x8_t) __a;
5573 }
5574
5575 __extension__ extern __inline uint8x8_t
5576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5577 vreinterpret_u8_p64 (poly64x1_t __a)
5578 {
5579   return (uint8x8_t) __a;
5580 }
5581
5582 __extension__ extern __inline uint8x16_t
5583 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5584 vreinterpretq_u8_f64 (float64x2_t __a)
5585 {
5586   return (uint8x16_t) __a;
5587 }
5588
5589 __extension__ extern __inline uint8x16_t
5590 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5591 vreinterpretq_u8_s8 (int8x16_t __a)
5592 {
5593   return (uint8x16_t) __a;
5594 }
5595
5596 __extension__ extern __inline uint8x16_t
5597 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5598 vreinterpretq_u8_s16 (int16x8_t __a)
5599 {
5600   return (uint8x16_t) __a;
5601 }
5602
5603 __extension__ extern __inline uint8x16_t
5604 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5605 vreinterpretq_u8_s32 (int32x4_t __a)
5606 {
5607   return (uint8x16_t) __a;
5608 }
5609
5610 __extension__ extern __inline uint8x16_t
5611 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5612 vreinterpretq_u8_s64 (int64x2_t __a)
5613 {
5614   return (uint8x16_t) __a;
5615 }
5616
5617 __extension__ extern __inline uint8x16_t
5618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5619 vreinterpretq_u8_f16 (float16x8_t __a)
5620 {
5621   return (uint8x16_t) __a;
5622 }
5623
5624 __extension__ extern __inline uint8x16_t
5625 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5626 vreinterpretq_u8_f32 (float32x4_t __a)
5627 {
5628   return (uint8x16_t) __a;
5629 }
5630
5631 __extension__ extern __inline uint8x16_t
5632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5633 vreinterpretq_u8_u16 (uint16x8_t __a)
5634 {
5635   return (uint8x16_t) __a;
5636 }
5637
5638 __extension__ extern __inline uint8x16_t
5639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5640 vreinterpretq_u8_u32 (uint32x4_t __a)
5641 {
5642   return (uint8x16_t) __a;
5643 }
5644
5645 __extension__ extern __inline uint8x16_t
5646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5647 vreinterpretq_u8_u64 (uint64x2_t __a)
5648 {
5649   return (uint8x16_t) __a;
5650 }
5651
5652 __extension__ extern __inline uint8x16_t
5653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5654 vreinterpretq_u8_p8 (poly8x16_t __a)
5655 {
5656   return (uint8x16_t) __a;
5657 }
5658
5659 __extension__ extern __inline uint8x16_t
5660 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5661 vreinterpretq_u8_p16 (poly16x8_t __a)
5662 {
5663   return (uint8x16_t) __a;
5664 }
5665
5666 __extension__ extern __inline uint8x16_t
5667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5668 vreinterpretq_u8_p64 (poly64x2_t __a)
5669 {
5670   return (uint8x16_t) __a;
5671 }
5672
5673 __extension__ extern __inline uint8x16_t
5674 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
5675 vreinterpretq_u8_p128 (poly128_t __a)
5676 {
5677   return (uint8x16_t)__a;
5678 }
5679
5680 __extension__ extern __inline uint16x4_t
5681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5682 vreinterpret_u16_f16 (float16x4_t __a)
5683 {
5684   return (uint16x4_t) __a;
5685 }
5686
5687 __extension__ extern __inline uint16x4_t
5688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5689 vreinterpret_u16_f64 (float64x1_t __a)
5690 {
5691   return (uint16x4_t) __a;
5692 }
5693
5694 __extension__ extern __inline uint16x4_t
5695 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5696 vreinterpret_u16_s8 (int8x8_t __a)
5697 {
5698   return (uint16x4_t) __a;
5699 }
5700
5701 __extension__ extern __inline uint16x4_t
5702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5703 vreinterpret_u16_s16 (int16x4_t __a)
5704 {
5705   return (uint16x4_t) __a;
5706 }
5707
5708 __extension__ extern __inline uint16x4_t
5709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5710 vreinterpret_u16_s32 (int32x2_t __a)
5711 {
5712   return (uint16x4_t) __a;
5713 }
5714
5715 __extension__ extern __inline uint16x4_t
5716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5717 vreinterpret_u16_s64 (int64x1_t __a)
5718 {
5719   return (uint16x4_t) __a;
5720 }
5721
5722 __extension__ extern __inline uint16x4_t
5723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5724 vreinterpret_u16_f32 (float32x2_t __a)
5725 {
5726   return (uint16x4_t) __a;
5727 }
5728
5729 __extension__ extern __inline uint16x4_t
5730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5731 vreinterpret_u16_u8 (uint8x8_t __a)
5732 {
5733   return (uint16x4_t) __a;
5734 }
5735
5736 __extension__ extern __inline uint16x4_t
5737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5738 vreinterpret_u16_u32 (uint32x2_t __a)
5739 {
5740   return (uint16x4_t) __a;
5741 }
5742
5743 __extension__ extern __inline uint16x4_t
5744 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5745 vreinterpret_u16_u64 (uint64x1_t __a)
5746 {
5747   return (uint16x4_t) __a;
5748 }
5749
5750 __extension__ extern __inline uint16x4_t
5751 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5752 vreinterpret_u16_p8 (poly8x8_t __a)
5753 {
5754   return (uint16x4_t) __a;
5755 }
5756
5757 __extension__ extern __inline uint16x4_t
5758 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5759 vreinterpret_u16_p16 (poly16x4_t __a)
5760 {
5761   return (uint16x4_t) __a;
5762 }
5763
5764 __extension__ extern __inline uint16x4_t
5765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5766 vreinterpret_u16_p64 (poly64x1_t __a)
5767 {
5768   return (uint16x4_t) __a;
5769 }
5770
5771 __extension__ extern __inline uint16x8_t
5772 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5773 vreinterpretq_u16_f64 (float64x2_t __a)
5774 {
5775   return (uint16x8_t) __a;
5776 }
5777
5778 __extension__ extern __inline uint16x8_t
5779 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5780 vreinterpretq_u16_s8 (int8x16_t __a)
5781 {
5782   return (uint16x8_t) __a;
5783 }
5784
5785 __extension__ extern __inline uint16x8_t
5786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5787 vreinterpretq_u16_s16 (int16x8_t __a)
5788 {
5789   return (uint16x8_t) __a;
5790 }
5791
5792 __extension__ extern __inline uint16x8_t
5793 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5794 vreinterpretq_u16_s32 (int32x4_t __a)
5795 {
5796   return (uint16x8_t) __a;
5797 }
5798
5799 __extension__ extern __inline uint16x8_t
5800 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5801 vreinterpretq_u16_s64 (int64x2_t __a)
5802 {
5803   return (uint16x8_t) __a;
5804 }
5805
5806 __extension__ extern __inline uint16x8_t
5807 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5808 vreinterpretq_u16_f16 (float16x8_t __a)
5809 {
5810   return (uint16x8_t) __a;
5811 }
5812
5813 __extension__ extern __inline uint16x8_t
5814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5815 vreinterpretq_u16_f32 (float32x4_t __a)
5816 {
5817   return (uint16x8_t) __a;
5818 }
5819
5820 __extension__ extern __inline uint16x8_t
5821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5822 vreinterpretq_u16_u8 (uint8x16_t __a)
5823 {
5824   return (uint16x8_t) __a;
5825 }
5826
5827 __extension__ extern __inline uint16x8_t
5828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5829 vreinterpretq_u16_u32 (uint32x4_t __a)
5830 {
5831   return (uint16x8_t) __a;
5832 }
5833
5834 __extension__ extern __inline uint16x8_t
5835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5836 vreinterpretq_u16_u64 (uint64x2_t __a)
5837 {
5838   return (uint16x8_t) __a;
5839 }
5840
5841 __extension__ extern __inline uint16x8_t
5842 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5843 vreinterpretq_u16_p8 (poly8x16_t __a)
5844 {
5845   return (uint16x8_t) __a;
5846 }
5847
5848 __extension__ extern __inline uint16x8_t
5849 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5850 vreinterpretq_u16_p16 (poly16x8_t __a)
5851 {
5852   return (uint16x8_t) __a;
5853 }
5854
5855 __extension__ extern __inline uint16x8_t
5856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5857 vreinterpretq_u16_p64 (poly64x2_t __a)
5858 {
5859   return (uint16x8_t) __a;
5860 }
5861
5862 __extension__ extern __inline uint16x8_t
5863 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
5864 vreinterpretq_u16_p128 (poly128_t __a)
5865 {
5866   return (uint16x8_t)__a;
5867 }
5868
5869 __extension__ extern __inline uint32x2_t
5870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5871 vreinterpret_u32_f16 (float16x4_t __a)
5872 {
5873   return (uint32x2_t) __a;
5874 }
5875
5876 __extension__ extern __inline uint32x2_t
5877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5878 vreinterpret_u32_f64 (float64x1_t __a)
5879 {
5880   return (uint32x2_t) __a;
5881 }
5882
5883 __extension__ extern __inline uint32x2_t
5884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5885 vreinterpret_u32_s8 (int8x8_t __a)
5886 {
5887   return (uint32x2_t) __a;
5888 }
5889
5890 __extension__ extern __inline uint32x2_t
5891 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5892 vreinterpret_u32_s16 (int16x4_t __a)
5893 {
5894   return (uint32x2_t) __a;
5895 }
5896
5897 __extension__ extern __inline uint32x2_t
5898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5899 vreinterpret_u32_s32 (int32x2_t __a)
5900 {
5901   return (uint32x2_t) __a;
5902 }
5903
5904 __extension__ extern __inline uint32x2_t
5905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5906 vreinterpret_u32_s64 (int64x1_t __a)
5907 {
5908   return (uint32x2_t) __a;
5909 }
5910
5911 __extension__ extern __inline uint32x2_t
5912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5913 vreinterpret_u32_f32 (float32x2_t __a)
5914 {
5915   return (uint32x2_t) __a;
5916 }
5917
5918 __extension__ extern __inline uint32x2_t
5919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5920 vreinterpret_u32_u8 (uint8x8_t __a)
5921 {
5922   return (uint32x2_t) __a;
5923 }
5924
5925 __extension__ extern __inline uint32x2_t
5926 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5927 vreinterpret_u32_u16 (uint16x4_t __a)
5928 {
5929   return (uint32x2_t) __a;
5930 }
5931
5932 __extension__ extern __inline uint32x2_t
5933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5934 vreinterpret_u32_u64 (uint64x1_t __a)
5935 {
5936   return (uint32x2_t) __a;
5937 }
5938
5939 __extension__ extern __inline uint32x2_t
5940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5941 vreinterpret_u32_p8 (poly8x8_t __a)
5942 {
5943   return (uint32x2_t) __a;
5944 }
5945
5946 __extension__ extern __inline uint32x2_t
5947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5948 vreinterpret_u32_p16 (poly16x4_t __a)
5949 {
5950   return (uint32x2_t) __a;
5951 }
5952
5953 __extension__ extern __inline uint32x2_t
5954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5955 vreinterpret_u32_p64 (poly64x1_t __a)
5956 {
5957   return (uint32x2_t) __a;
5958 }
5959
5960 __extension__ extern __inline uint32x4_t
5961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5962 vreinterpretq_u32_f64 (float64x2_t __a)
5963 {
5964   return (uint32x4_t) __a;
5965 }
5966
5967 __extension__ extern __inline uint32x4_t
5968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5969 vreinterpretq_u32_s8 (int8x16_t __a)
5970 {
5971   return (uint32x4_t) __a;
5972 }
5973
5974 __extension__ extern __inline uint32x4_t
5975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5976 vreinterpretq_u32_s16 (int16x8_t __a)
5977 {
5978   return (uint32x4_t) __a;
5979 }
5980
5981 __extension__ extern __inline uint32x4_t
5982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5983 vreinterpretq_u32_s32 (int32x4_t __a)
5984 {
5985   return (uint32x4_t) __a;
5986 }
5987
5988 __extension__ extern __inline uint32x4_t
5989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5990 vreinterpretq_u32_s64 (int64x2_t __a)
5991 {
5992   return (uint32x4_t) __a;
5993 }
5994
5995 __extension__ extern __inline uint32x4_t
5996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5997 vreinterpretq_u32_f16 (float16x8_t __a)
5998 {
5999   return (uint32x4_t) __a;
6000 }
6001
6002 __extension__ extern __inline uint32x4_t
6003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6004 vreinterpretq_u32_f32 (float32x4_t __a)
6005 {
6006   return (uint32x4_t) __a;
6007 }
6008
6009 __extension__ extern __inline uint32x4_t
6010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6011 vreinterpretq_u32_u8 (uint8x16_t __a)
6012 {
6013   return (uint32x4_t) __a;
6014 }
6015
6016 __extension__ extern __inline uint32x4_t
6017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6018 vreinterpretq_u32_u16 (uint16x8_t __a)
6019 {
6020   return (uint32x4_t) __a;
6021 }
6022
6023 __extension__ extern __inline uint32x4_t
6024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6025 vreinterpretq_u32_u64 (uint64x2_t __a)
6026 {
6027   return (uint32x4_t) __a;
6028 }
6029
6030 __extension__ extern __inline uint32x4_t
6031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6032 vreinterpretq_u32_p8 (poly8x16_t __a)
6033 {
6034   return (uint32x4_t) __a;
6035 }
6036
6037 __extension__ extern __inline uint32x4_t
6038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6039 vreinterpretq_u32_p16 (poly16x8_t __a)
6040 {
6041   return (uint32x4_t) __a;
6042 }
6043
6044 __extension__ extern __inline uint32x4_t
6045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6046 vreinterpretq_u32_p64 (poly64x2_t __a)
6047 {
6048   return (uint32x4_t) __a;
6049 }
6050
6051 __extension__ extern __inline uint32x4_t
6052 __attribute__  ((__always_inline__, __gnu_inline__, __artificial__))
6053 vreinterpretq_u32_p128 (poly128_t __a)
6054 {
6055   return (uint32x4_t)__a;
6056 }
6057
6058 /* vset_lane  */
6059
6060 __extension__ extern __inline float16x4_t
6061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6062 vset_lane_f16 (float16_t __elem, float16x4_t __vec, const int __index)
6063 {
6064   return __aarch64_vset_lane_any (__elem, __vec, __index);
6065 }
6066
6067 __extension__ extern __inline float32x2_t
6068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6069 vset_lane_f32 (float32_t __elem, float32x2_t __vec, const int __index)
6070 {
6071   return __aarch64_vset_lane_any (__elem, __vec, __index);
6072 }
6073
6074 __extension__ extern __inline float64x1_t
6075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6076 vset_lane_f64 (float64_t __elem, float64x1_t __vec, const int __index)
6077 {
6078   return __aarch64_vset_lane_any (__elem, __vec, __index);
6079 }
6080
6081 __extension__ extern __inline poly8x8_t
6082 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6083 vset_lane_p8 (poly8_t __elem, poly8x8_t __vec, const int __index)
6084 {
6085   return __aarch64_vset_lane_any (__elem, __vec, __index);
6086 }
6087
6088 __extension__ extern __inline poly16x4_t
6089 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6090 vset_lane_p16 (poly16_t __elem, poly16x4_t __vec, const int __index)
6091 {
6092   return __aarch64_vset_lane_any (__elem, __vec, __index);
6093 }
6094
6095 __extension__ extern __inline poly64x1_t
6096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6097 vset_lane_p64 (poly64_t __elem, poly64x1_t __vec, const int __index)
6098 {
6099   return __aarch64_vset_lane_any (__elem, __vec, __index);
6100 }
6101
6102 __extension__ extern __inline int8x8_t
6103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6104 vset_lane_s8 (int8_t __elem, int8x8_t __vec, const int __index)
6105 {
6106   return __aarch64_vset_lane_any (__elem, __vec, __index);
6107 }
6108
6109 __extension__ extern __inline int16x4_t
6110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6111 vset_lane_s16 (int16_t __elem, int16x4_t __vec, const int __index)
6112 {
6113   return __aarch64_vset_lane_any (__elem, __vec, __index);
6114 }
6115
6116 __extension__ extern __inline int32x2_t
6117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6118 vset_lane_s32 (int32_t __elem, int32x2_t __vec, const int __index)
6119 {
6120   return __aarch64_vset_lane_any (__elem, __vec, __index);
6121 }
6122
6123 __extension__ extern __inline int64x1_t
6124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6125 vset_lane_s64 (int64_t __elem, int64x1_t __vec, const int __index)
6126 {
6127   return __aarch64_vset_lane_any (__elem, __vec, __index);
6128 }
6129
6130 __extension__ extern __inline uint8x8_t
6131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6132 vset_lane_u8 (uint8_t __elem, uint8x8_t __vec, const int __index)
6133 {
6134   return __aarch64_vset_lane_any (__elem, __vec, __index);
6135 }
6136
6137 __extension__ extern __inline uint16x4_t
6138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6139 vset_lane_u16 (uint16_t __elem, uint16x4_t __vec, const int __index)
6140 {
6141   return __aarch64_vset_lane_any (__elem, __vec, __index);
6142 }
6143
6144 __extension__ extern __inline uint32x2_t
6145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6146 vset_lane_u32 (uint32_t __elem, uint32x2_t __vec, const int __index)
6147 {
6148   return __aarch64_vset_lane_any (__elem, __vec, __index);
6149 }
6150
6151 __extension__ extern __inline uint64x1_t
6152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6153 vset_lane_u64 (uint64_t __elem, uint64x1_t __vec, const int __index)
6154 {
6155   return __aarch64_vset_lane_any (__elem, __vec, __index);
6156 }
6157
6158 /* vsetq_lane  */
6159
6160 __extension__ extern __inline float16x8_t
6161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6162 vsetq_lane_f16 (float16_t __elem, float16x8_t __vec, const int __index)
6163 {
6164   return __aarch64_vset_lane_any (__elem, __vec, __index);
6165 }
6166
6167 __extension__ extern __inline float32x4_t
6168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6169 vsetq_lane_f32 (float32_t __elem, float32x4_t __vec, const int __index)
6170 {
6171   return __aarch64_vset_lane_any (__elem, __vec, __index);
6172 }
6173
6174 __extension__ extern __inline float64x2_t
6175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6176 vsetq_lane_f64 (float64_t __elem, float64x2_t __vec, const int __index)
6177 {
6178   return __aarch64_vset_lane_any (__elem, __vec, __index);
6179 }
6180
6181 __extension__ extern __inline poly8x16_t
6182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6183 vsetq_lane_p8 (poly8_t __elem, poly8x16_t __vec, const int __index)
6184 {
6185   return __aarch64_vset_lane_any (__elem, __vec, __index);
6186 }
6187
6188 __extension__ extern __inline poly16x8_t
6189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6190 vsetq_lane_p16 (poly16_t __elem, poly16x8_t __vec, const int __index)
6191 {
6192   return __aarch64_vset_lane_any (__elem, __vec, __index);
6193 }
6194
6195 __extension__ extern __inline poly64x2_t
6196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6197 vsetq_lane_p64 (poly64_t __elem, poly64x2_t __vec, const int __index)
6198 {
6199   return __aarch64_vset_lane_any (__elem, __vec, __index);
6200 }
6201
6202 __extension__ extern __inline int8x16_t
6203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6204 vsetq_lane_s8 (int8_t __elem, int8x16_t __vec, const int __index)
6205 {
6206   return __aarch64_vset_lane_any (__elem, __vec, __index);
6207 }
6208
6209 __extension__ extern __inline int16x8_t
6210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6211 vsetq_lane_s16 (int16_t __elem, int16x8_t __vec, const int __index)
6212 {
6213   return __aarch64_vset_lane_any (__elem, __vec, __index);
6214 }
6215
6216 __extension__ extern __inline int32x4_t
6217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6218 vsetq_lane_s32 (int32_t __elem, int32x4_t __vec, const int __index)
6219 {
6220   return __aarch64_vset_lane_any (__elem, __vec, __index);
6221 }
6222
6223 __extension__ extern __inline int64x2_t
6224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6225 vsetq_lane_s64 (int64_t __elem, int64x2_t __vec, const int __index)
6226 {
6227   return __aarch64_vset_lane_any (__elem, __vec, __index);
6228 }
6229
6230 __extension__ extern __inline uint8x16_t
6231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6232 vsetq_lane_u8 (uint8_t __elem, uint8x16_t __vec, const int __index)
6233 {
6234   return __aarch64_vset_lane_any (__elem, __vec, __index);
6235 }
6236
6237 __extension__ extern __inline uint16x8_t
6238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6239 vsetq_lane_u16 (uint16_t __elem, uint16x8_t __vec, const int __index)
6240 {
6241   return __aarch64_vset_lane_any (__elem, __vec, __index);
6242 }
6243
6244 __extension__ extern __inline uint32x4_t
6245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6246 vsetq_lane_u32 (uint32_t __elem, uint32x4_t __vec, const int __index)
6247 {
6248   return __aarch64_vset_lane_any (__elem, __vec, __index);
6249 }
6250
6251 __extension__ extern __inline uint64x2_t
6252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6253 vsetq_lane_u64 (uint64_t __elem, uint64x2_t __vec, const int __index)
6254 {
6255   return __aarch64_vset_lane_any (__elem, __vec, __index);
6256 }
6257
6258 #define __GET_LOW(__TYPE) \
6259   uint64x2_t tmp = vreinterpretq_u64_##__TYPE (__a);  \
6260   uint64x1_t lo = vcreate_u64 (vgetq_lane_u64 (tmp, 0));  \
6261   return vreinterpret_##__TYPE##_u64 (lo);
6262
6263 __extension__ extern __inline float16x4_t
6264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6265 vget_low_f16 (float16x8_t __a)
6266 {
6267   __GET_LOW (f16);
6268 }
6269
6270 __extension__ extern __inline float32x2_t
6271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6272 vget_low_f32 (float32x4_t __a)
6273 {
6274   __GET_LOW (f32);
6275 }
6276
6277 __extension__ extern __inline float64x1_t
6278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6279 vget_low_f64 (float64x2_t __a)
6280 {
6281   return (float64x1_t) {vgetq_lane_f64 (__a, 0)};
6282 }
6283
6284 __extension__ extern __inline poly8x8_t
6285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6286 vget_low_p8 (poly8x16_t __a)
6287 {
6288   __GET_LOW (p8);
6289 }
6290
6291 __extension__ extern __inline poly16x4_t
6292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6293 vget_low_p16 (poly16x8_t __a)
6294 {
6295   __GET_LOW (p16);
6296 }
6297
6298 __extension__ extern __inline poly64x1_t
6299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6300 vget_low_p64 (poly64x2_t __a)
6301 {
6302   __GET_LOW (p64);
6303 }
6304
6305 __extension__ extern __inline int8x8_t
6306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6307 vget_low_s8 (int8x16_t __a)
6308 {
6309   __GET_LOW (s8);
6310 }
6311
6312 __extension__ extern __inline int16x4_t
6313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6314 vget_low_s16 (int16x8_t __a)
6315 {
6316   __GET_LOW (s16);
6317 }
6318
6319 __extension__ extern __inline int32x2_t
6320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6321 vget_low_s32 (int32x4_t __a)
6322 {
6323   __GET_LOW (s32);
6324 }
6325
6326 __extension__ extern __inline int64x1_t
6327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6328 vget_low_s64 (int64x2_t __a)
6329 {
6330   __GET_LOW (s64);
6331 }
6332
6333 __extension__ extern __inline uint8x8_t
6334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6335 vget_low_u8 (uint8x16_t __a)
6336 {
6337   __GET_LOW (u8);
6338 }
6339
6340 __extension__ extern __inline uint16x4_t
6341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6342 vget_low_u16 (uint16x8_t __a)
6343 {
6344   __GET_LOW (u16);
6345 }
6346
6347 __extension__ extern __inline uint32x2_t
6348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6349 vget_low_u32 (uint32x4_t __a)
6350 {
6351   __GET_LOW (u32);
6352 }
6353
6354 __extension__ extern __inline uint64x1_t
6355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6356 vget_low_u64 (uint64x2_t __a)
6357 {
6358   return vcreate_u64 (vgetq_lane_u64 (__a, 0));
6359 }
6360
6361 #undef __GET_LOW
6362
6363 #define __GET_HIGH(__TYPE)                                      \
6364   uint64x2_t tmp = vreinterpretq_u64_##__TYPE (__a);            \
6365   uint64x1_t hi = vcreate_u64 (vgetq_lane_u64 (tmp, 1));        \
6366   return vreinterpret_##__TYPE##_u64 (hi);
6367
6368 __extension__ extern __inline float16x4_t
6369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6370 vget_high_f16 (float16x8_t __a)
6371 {
6372   __GET_HIGH (f16);
6373 }
6374
6375 __extension__ extern __inline float32x2_t
6376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6377 vget_high_f32 (float32x4_t __a)
6378 {
6379   __GET_HIGH (f32);
6380 }
6381
6382 __extension__ extern __inline float64x1_t
6383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6384 vget_high_f64 (float64x2_t __a)
6385 {
6386   __GET_HIGH (f64);
6387 }
6388
6389 __extension__ extern __inline poly8x8_t
6390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6391 vget_high_p8 (poly8x16_t __a)
6392 {
6393   __GET_HIGH (p8);
6394 }
6395
6396 __extension__ extern __inline poly16x4_t
6397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6398 vget_high_p16 (poly16x8_t __a)
6399 {
6400   __GET_HIGH (p16);
6401 }
6402
6403 __extension__ extern __inline poly64x1_t
6404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6405 vget_high_p64 (poly64x2_t __a)
6406 {
6407   __GET_HIGH (p64);
6408 }
6409
6410 __extension__ extern __inline int8x8_t
6411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6412 vget_high_s8 (int8x16_t __a)
6413 {
6414   __GET_HIGH (s8);
6415 }
6416
6417 __extension__ extern __inline int16x4_t
6418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6419 vget_high_s16 (int16x8_t __a)
6420 {
6421   __GET_HIGH (s16);
6422 }
6423
6424 __extension__ extern __inline int32x2_t
6425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6426 vget_high_s32 (int32x4_t __a)
6427 {
6428   __GET_HIGH (s32);
6429 }
6430
6431 __extension__ extern __inline int64x1_t
6432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6433 vget_high_s64 (int64x2_t __a)
6434 {
6435   __GET_HIGH (s64);
6436 }
6437
6438 __extension__ extern __inline uint8x8_t
6439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6440 vget_high_u8 (uint8x16_t __a)
6441 {
6442   __GET_HIGH (u8);
6443 }
6444
6445 __extension__ extern __inline uint16x4_t
6446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6447 vget_high_u16 (uint16x8_t __a)
6448 {
6449   __GET_HIGH (u16);
6450 }
6451
6452 __extension__ extern __inline uint32x2_t
6453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6454 vget_high_u32 (uint32x4_t __a)
6455 {
6456   __GET_HIGH (u32);
6457 }
6458
6459 #undef __GET_HIGH
6460
6461 __extension__ extern __inline uint64x1_t
6462 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6463 vget_high_u64 (uint64x2_t __a)
6464 {
6465   return vcreate_u64 (vgetq_lane_u64 (__a, 1));
6466 }
6467
6468 __extension__ extern __inline int8x16_t
6469 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6470 vcombine_s8 (int8x8_t __a, int8x8_t __b)
6471 {
6472   return (int8x16_t) __builtin_aarch64_combinev8qi (__a, __b);
6473 }
6474
6475 __extension__ extern __inline int16x8_t
6476 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6477 vcombine_s16 (int16x4_t __a, int16x4_t __b)
6478 {
6479   return (int16x8_t) __builtin_aarch64_combinev4hi (__a, __b);
6480 }
6481
6482 __extension__ extern __inline int32x4_t
6483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6484 vcombine_s32 (int32x2_t __a, int32x2_t __b)
6485 {
6486   return (int32x4_t) __builtin_aarch64_combinev2si (__a, __b);
6487 }
6488
6489 __extension__ extern __inline int64x2_t
6490 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6491 vcombine_s64 (int64x1_t __a, int64x1_t __b)
6492 {
6493   return __builtin_aarch64_combinedi (__a[0], __b[0]);
6494 }
6495
6496 __extension__ extern __inline float16x8_t
6497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6498 vcombine_f16 (float16x4_t __a, float16x4_t __b)
6499 {
6500   return __builtin_aarch64_combinev4hf (__a, __b);
6501 }
6502
6503 __extension__ extern __inline float32x4_t
6504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6505 vcombine_f32 (float32x2_t __a, float32x2_t __b)
6506 {
6507   return (float32x4_t) __builtin_aarch64_combinev2sf (__a, __b);
6508 }
6509
6510 __extension__ extern __inline uint8x16_t
6511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6512 vcombine_u8 (uint8x8_t __a, uint8x8_t __b)
6513 {
6514   return (uint8x16_t) __builtin_aarch64_combinev8qi ((int8x8_t) __a,
6515                                                      (int8x8_t) __b);
6516 }
6517
6518 __extension__ extern __inline uint16x8_t
6519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6520 vcombine_u16 (uint16x4_t __a, uint16x4_t __b)
6521 {
6522   return (uint16x8_t) __builtin_aarch64_combinev4hi ((int16x4_t) __a,
6523                                                      (int16x4_t) __b);
6524 }
6525
6526 __extension__ extern __inline uint32x4_t
6527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6528 vcombine_u32 (uint32x2_t __a, uint32x2_t __b)
6529 {
6530   return (uint32x4_t) __builtin_aarch64_combinev2si ((int32x2_t) __a,
6531                                                      (int32x2_t) __b);
6532 }
6533
6534 __extension__ extern __inline uint64x2_t
6535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6536 vcombine_u64 (uint64x1_t __a, uint64x1_t __b)
6537 {
6538   return (uint64x2_t) __builtin_aarch64_combinedi (__a[0], __b[0]);
6539 }
6540
6541 __extension__ extern __inline float64x2_t
6542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6543 vcombine_f64 (float64x1_t __a, float64x1_t __b)
6544 {
6545   return __builtin_aarch64_combinedf (__a[0], __b[0]);
6546 }
6547
6548 __extension__ extern __inline poly8x16_t
6549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6550 vcombine_p8 (poly8x8_t __a, poly8x8_t __b)
6551 {
6552   return (poly8x16_t) __builtin_aarch64_combinev8qi ((int8x8_t) __a,
6553                                                      (int8x8_t) __b);
6554 }
6555
6556 __extension__ extern __inline poly16x8_t
6557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6558 vcombine_p16 (poly16x4_t __a, poly16x4_t __b)
6559 {
6560   return (poly16x8_t) __builtin_aarch64_combinev4hi ((int16x4_t) __a,
6561                                                      (int16x4_t) __b);
6562 }
6563
6564 __extension__ extern __inline poly64x2_t
6565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6566 vcombine_p64 (poly64x1_t __a, poly64x1_t __b)
6567 {
6568   return (poly64x2_t) __builtin_aarch64_combinedi_ppp (__a[0], __b[0]);
6569 }
6570
6571 /* Start of temporary inline asm implementations.  */
6572
6573 __extension__ extern __inline int8x8_t
6574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6575 vaba_s8 (int8x8_t a, int8x8_t b, int8x8_t c)
6576 {
6577   int8x8_t result;
6578   __asm__ ("saba %0.8b,%2.8b,%3.8b"
6579            : "=w"(result)
6580            : "0"(a), "w"(b), "w"(c)
6581            : /* No clobbers */);
6582   return result;
6583 }
6584
6585 __extension__ extern __inline int16x4_t
6586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6587 vaba_s16 (int16x4_t a, int16x4_t b, int16x4_t c)
6588 {
6589   int16x4_t result;
6590   __asm__ ("saba %0.4h,%2.4h,%3.4h"
6591            : "=w"(result)
6592            : "0"(a), "w"(b), "w"(c)
6593            : /* No clobbers */);
6594   return result;
6595 }
6596
6597 __extension__ extern __inline int32x2_t
6598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6599 vaba_s32 (int32x2_t a, int32x2_t b, int32x2_t c)
6600 {
6601   int32x2_t result;
6602   __asm__ ("saba %0.2s,%2.2s,%3.2s"
6603            : "=w"(result)
6604            : "0"(a), "w"(b), "w"(c)
6605            : /* No clobbers */);
6606   return result;
6607 }
6608
6609 __extension__ extern __inline uint8x8_t
6610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6611 vaba_u8 (uint8x8_t a, uint8x8_t b, uint8x8_t c)
6612 {
6613   uint8x8_t result;
6614   __asm__ ("uaba %0.8b,%2.8b,%3.8b"
6615            : "=w"(result)
6616            : "0"(a), "w"(b), "w"(c)
6617            : /* No clobbers */);
6618   return result;
6619 }
6620
6621 __extension__ extern __inline uint16x4_t
6622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6623 vaba_u16 (uint16x4_t a, uint16x4_t b, uint16x4_t c)
6624 {
6625   uint16x4_t result;
6626   __asm__ ("uaba %0.4h,%2.4h,%3.4h"
6627            : "=w"(result)
6628            : "0"(a), "w"(b), "w"(c)
6629            : /* No clobbers */);
6630   return result;
6631 }
6632
6633 __extension__ extern __inline uint32x2_t
6634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6635 vaba_u32 (uint32x2_t a, uint32x2_t b, uint32x2_t c)
6636 {
6637   uint32x2_t result;
6638   __asm__ ("uaba %0.2s,%2.2s,%3.2s"
6639            : "=w"(result)
6640            : "0"(a), "w"(b), "w"(c)
6641            : /* No clobbers */);
6642   return result;
6643 }
6644
6645 __extension__ extern __inline int16x8_t
6646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6647 vabal_high_s8 (int16x8_t a, int8x16_t b, int8x16_t c)
6648 {
6649   int16x8_t result;
6650   __asm__ ("sabal2 %0.8h,%2.16b,%3.16b"
6651            : "=w"(result)
6652            : "0"(a), "w"(b), "w"(c)
6653            : /* No clobbers */);
6654   return result;
6655 }
6656
6657 __extension__ extern __inline int32x4_t
6658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6659 vabal_high_s16 (int32x4_t a, int16x8_t b, int16x8_t c)
6660 {
6661   int32x4_t result;
6662   __asm__ ("sabal2 %0.4s,%2.8h,%3.8h"
6663            : "=w"(result)
6664            : "0"(a), "w"(b), "w"(c)
6665            : /* No clobbers */);
6666   return result;
6667 }
6668
6669 __extension__ extern __inline int64x2_t
6670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6671 vabal_high_s32 (int64x2_t a, int32x4_t b, int32x4_t c)
6672 {
6673   int64x2_t result;
6674   __asm__ ("sabal2 %0.2d,%2.4s,%3.4s"
6675            : "=w"(result)
6676            : "0"(a), "w"(b), "w"(c)
6677            : /* No clobbers */);
6678   return result;
6679 }
6680
6681 __extension__ extern __inline uint16x8_t
6682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6683 vabal_high_u8 (uint16x8_t a, uint8x16_t b, uint8x16_t c)
6684 {
6685   uint16x8_t result;
6686   __asm__ ("uabal2 %0.8h,%2.16b,%3.16b"
6687            : "=w"(result)
6688            : "0"(a), "w"(b), "w"(c)
6689            : /* No clobbers */);
6690   return result;
6691 }
6692
6693 __extension__ extern __inline uint32x4_t
6694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6695 vabal_high_u16 (uint32x4_t a, uint16x8_t b, uint16x8_t c)
6696 {
6697   uint32x4_t result;
6698   __asm__ ("uabal2 %0.4s,%2.8h,%3.8h"
6699            : "=w"(result)
6700            : "0"(a), "w"(b), "w"(c)
6701            : /* No clobbers */);
6702   return result;
6703 }
6704
6705 __extension__ extern __inline uint64x2_t
6706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6707 vabal_high_u32 (uint64x2_t a, uint32x4_t b, uint32x4_t c)
6708 {
6709   uint64x2_t result;
6710   __asm__ ("uabal2 %0.2d,%2.4s,%3.4s"
6711            : "=w"(result)
6712            : "0"(a), "w"(b), "w"(c)
6713            : /* No clobbers */);
6714   return result;
6715 }
6716
6717 __extension__ extern __inline int16x8_t
6718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6719 vabal_s8 (int16x8_t a, int8x8_t b, int8x8_t c)
6720 {
6721   int16x8_t result;
6722   __asm__ ("sabal %0.8h,%2.8b,%3.8b"
6723            : "=w"(result)
6724            : "0"(a), "w"(b), "w"(c)
6725            : /* No clobbers */);
6726   return result;
6727 }
6728
6729 __extension__ extern __inline int32x4_t
6730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6731 vabal_s16 (int32x4_t a, int16x4_t b, int16x4_t c)
6732 {
6733   int32x4_t result;
6734   __asm__ ("sabal %0.4s,%2.4h,%3.4h"
6735            : "=w"(result)
6736            : "0"(a), "w"(b), "w"(c)
6737            : /* No clobbers */);
6738   return result;
6739 }
6740
6741 __extension__ extern __inline int64x2_t
6742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6743 vabal_s32 (int64x2_t a, int32x2_t b, int32x2_t c)
6744 {
6745   int64x2_t result;
6746   __asm__ ("sabal %0.2d,%2.2s,%3.2s"
6747            : "=w"(result)
6748            : "0"(a), "w"(b), "w"(c)
6749            : /* No clobbers */);
6750   return result;
6751 }
6752
6753 __extension__ extern __inline uint16x8_t
6754 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6755 vabal_u8 (uint16x8_t a, uint8x8_t b, uint8x8_t c)
6756 {
6757   uint16x8_t result;
6758   __asm__ ("uabal %0.8h,%2.8b,%3.8b"
6759            : "=w"(result)
6760            : "0"(a), "w"(b), "w"(c)
6761            : /* No clobbers */);
6762   return result;
6763 }
6764
6765 __extension__ extern __inline uint32x4_t
6766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6767 vabal_u16 (uint32x4_t a, uint16x4_t b, uint16x4_t c)
6768 {
6769   uint32x4_t result;
6770   __asm__ ("uabal %0.4s,%2.4h,%3.4h"
6771            : "=w"(result)
6772            : "0"(a), "w"(b), "w"(c)
6773            : /* No clobbers */);
6774   return result;
6775 }
6776
6777 __extension__ extern __inline uint64x2_t
6778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6779 vabal_u32 (uint64x2_t a, uint32x2_t b, uint32x2_t c)
6780 {
6781   uint64x2_t result;
6782   __asm__ ("uabal %0.2d,%2.2s,%3.2s"
6783            : "=w"(result)
6784            : "0"(a), "w"(b), "w"(c)
6785            : /* No clobbers */);
6786   return result;
6787 }
6788
6789 __extension__ extern __inline int8x16_t
6790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6791 vabaq_s8 (int8x16_t a, int8x16_t b, int8x16_t c)
6792 {
6793   int8x16_t result;
6794   __asm__ ("saba %0.16b,%2.16b,%3.16b"
6795            : "=w"(result)
6796            : "0"(a), "w"(b), "w"(c)
6797            : /* No clobbers */);
6798   return result;
6799 }
6800
6801 __extension__ extern __inline int16x8_t
6802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6803 vabaq_s16 (int16x8_t a, int16x8_t b, int16x8_t c)
6804 {
6805   int16x8_t result;
6806   __asm__ ("saba %0.8h,%2.8h,%3.8h"
6807            : "=w"(result)
6808            : "0"(a), "w"(b), "w"(c)
6809            : /* No clobbers */);
6810   return result;
6811 }
6812
6813 __extension__ extern __inline int32x4_t
6814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6815 vabaq_s32 (int32x4_t a, int32x4_t b, int32x4_t c)
6816 {
6817   int32x4_t result;
6818   __asm__ ("saba %0.4s,%2.4s,%3.4s"
6819            : "=w"(result)
6820            : "0"(a), "w"(b), "w"(c)
6821            : /* No clobbers */);
6822   return result;
6823 }
6824
6825 __extension__ extern __inline uint8x16_t
6826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6827 vabaq_u8 (uint8x16_t a, uint8x16_t b, uint8x16_t c)
6828 {
6829   uint8x16_t result;
6830   __asm__ ("uaba %0.16b,%2.16b,%3.16b"
6831            : "=w"(result)
6832            : "0"(a), "w"(b), "w"(c)
6833            : /* No clobbers */);
6834   return result;
6835 }
6836
6837 __extension__ extern __inline uint16x8_t
6838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6839 vabaq_u16 (uint16x8_t a, uint16x8_t b, uint16x8_t c)
6840 {
6841   uint16x8_t result;
6842   __asm__ ("uaba %0.8h,%2.8h,%3.8h"
6843            : "=w"(result)
6844            : "0"(a), "w"(b), "w"(c)
6845            : /* No clobbers */);
6846   return result;
6847 }
6848
6849 __extension__ extern __inline uint32x4_t
6850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6851 vabaq_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t c)
6852 {
6853   uint32x4_t result;
6854   __asm__ ("uaba %0.4s,%2.4s,%3.4s"
6855            : "=w"(result)
6856            : "0"(a), "w"(b), "w"(c)
6857            : /* No clobbers */);
6858   return result;
6859 }
6860
6861 __extension__ extern __inline int8x8_t
6862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6863 vabd_s8 (int8x8_t a, int8x8_t b)
6864 {
6865   int8x8_t result;
6866   __asm__ ("sabd %0.8b, %1.8b, %2.8b"
6867            : "=w"(result)
6868            : "w"(a), "w"(b)
6869            : /* No clobbers */);
6870   return result;
6871 }
6872
6873 __extension__ extern __inline int16x4_t
6874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6875 vabd_s16 (int16x4_t a, int16x4_t b)
6876 {
6877   int16x4_t result;
6878   __asm__ ("sabd %0.4h, %1.4h, %2.4h"
6879            : "=w"(result)
6880            : "w"(a), "w"(b)
6881            : /* No clobbers */);
6882   return result;
6883 }
6884
6885 __extension__ extern __inline int32x2_t
6886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6887 vabd_s32 (int32x2_t a, int32x2_t b)
6888 {
6889   int32x2_t result;
6890   __asm__ ("sabd %0.2s, %1.2s, %2.2s"
6891            : "=w"(result)
6892            : "w"(a), "w"(b)
6893            : /* No clobbers */);
6894   return result;
6895 }
6896
6897 __extension__ extern __inline uint8x8_t
6898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6899 vabd_u8 (uint8x8_t a, uint8x8_t b)
6900 {
6901   uint8x8_t result;
6902   __asm__ ("uabd %0.8b, %1.8b, %2.8b"
6903            : "=w"(result)
6904            : "w"(a), "w"(b)
6905            : /* No clobbers */);
6906   return result;
6907 }
6908
6909 __extension__ extern __inline uint16x4_t
6910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6911 vabd_u16 (uint16x4_t a, uint16x4_t b)
6912 {
6913   uint16x4_t result;
6914   __asm__ ("uabd %0.4h, %1.4h, %2.4h"
6915            : "=w"(result)
6916            : "w"(a), "w"(b)
6917            : /* No clobbers */);
6918   return result;
6919 }
6920
6921 __extension__ extern __inline uint32x2_t
6922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6923 vabd_u32 (uint32x2_t a, uint32x2_t b)
6924 {
6925   uint32x2_t result;
6926   __asm__ ("uabd %0.2s, %1.2s, %2.2s"
6927            : "=w"(result)
6928            : "w"(a), "w"(b)
6929            : /* No clobbers */);
6930   return result;
6931 }
6932
6933 __extension__ extern __inline int16x8_t
6934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6935 vabdl_high_s8 (int8x16_t a, int8x16_t b)
6936 {
6937   int16x8_t result;
6938   __asm__ ("sabdl2 %0.8h,%1.16b,%2.16b"
6939            : "=w"(result)
6940            : "w"(a), "w"(b)
6941            : /* No clobbers */);
6942   return result;
6943 }
6944
6945 __extension__ extern __inline int32x4_t
6946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6947 vabdl_high_s16 (int16x8_t a, int16x8_t b)
6948 {
6949   int32x4_t result;
6950   __asm__ ("sabdl2 %0.4s,%1.8h,%2.8h"
6951            : "=w"(result)
6952            : "w"(a), "w"(b)
6953            : /* No clobbers */);
6954   return result;
6955 }
6956
6957 __extension__ extern __inline int64x2_t
6958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6959 vabdl_high_s32 (int32x4_t a, int32x4_t b)
6960 {
6961   int64x2_t result;
6962   __asm__ ("sabdl2 %0.2d,%1.4s,%2.4s"
6963            : "=w"(result)
6964            : "w"(a), "w"(b)
6965            : /* No clobbers */);
6966   return result;
6967 }
6968
6969 __extension__ extern __inline uint16x8_t
6970 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6971 vabdl_high_u8 (uint8x16_t a, uint8x16_t b)
6972 {
6973   uint16x8_t result;
6974   __asm__ ("uabdl2 %0.8h,%1.16b,%2.16b"
6975            : "=w"(result)
6976            : "w"(a), "w"(b)
6977            : /* No clobbers */);
6978   return result;
6979 }
6980
6981 __extension__ extern __inline uint32x4_t
6982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6983 vabdl_high_u16 (uint16x8_t a, uint16x8_t b)
6984 {
6985   uint32x4_t result;
6986   __asm__ ("uabdl2 %0.4s,%1.8h,%2.8h"
6987            : "=w"(result)
6988            : "w"(a), "w"(b)
6989            : /* No clobbers */);
6990   return result;
6991 }
6992
6993 __extension__ extern __inline uint64x2_t
6994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6995 vabdl_high_u32 (uint32x4_t a, uint32x4_t b)
6996 {
6997   uint64x2_t result;
6998   __asm__ ("uabdl2 %0.2d,%1.4s,%2.4s"
6999            : "=w"(result)
7000            : "w"(a), "w"(b)
7001            : /* No clobbers */);
7002   return result;
7003 }
7004
7005 __extension__ extern __inline int16x8_t
7006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7007 vabdl_s8 (int8x8_t a, int8x8_t b)
7008 {
7009   int16x8_t result;
7010   __asm__ ("sabdl %0.8h, %1.8b, %2.8b"
7011            : "=w"(result)
7012            : "w"(a), "w"(b)
7013            : /* No clobbers */);
7014   return result;
7015 }
7016
7017 __extension__ extern __inline int32x4_t
7018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7019 vabdl_s16 (int16x4_t a, int16x4_t b)
7020 {
7021   int32x4_t result;
7022   __asm__ ("sabdl %0.4s, %1.4h, %2.4h"
7023            : "=w"(result)
7024            : "w"(a), "w"(b)
7025            : /* No clobbers */);
7026   return result;
7027 }
7028
7029 __extension__ extern __inline int64x2_t
7030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7031 vabdl_s32 (int32x2_t a, int32x2_t b)
7032 {
7033   int64x2_t result;
7034   __asm__ ("sabdl %0.2d, %1.2s, %2.2s"
7035            : "=w"(result)
7036            : "w"(a), "w"(b)
7037            : /* No clobbers */);
7038   return result;
7039 }
7040
7041 __extension__ extern __inline uint16x8_t
7042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7043 vabdl_u8 (uint8x8_t a, uint8x8_t b)
7044 {
7045   uint16x8_t result;
7046   __asm__ ("uabdl %0.8h, %1.8b, %2.8b"
7047            : "=w"(result)
7048            : "w"(a), "w"(b)
7049            : /* No clobbers */);
7050   return result;
7051 }
7052
7053 __extension__ extern __inline uint32x4_t
7054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7055 vabdl_u16 (uint16x4_t a, uint16x4_t b)
7056 {
7057   uint32x4_t result;
7058   __asm__ ("uabdl %0.4s, %1.4h, %2.4h"
7059            : "=w"(result)
7060            : "w"(a), "w"(b)
7061            : /* No clobbers */);
7062   return result;
7063 }
7064
7065 __extension__ extern __inline uint64x2_t
7066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7067 vabdl_u32 (uint32x2_t a, uint32x2_t b)
7068 {
7069   uint64x2_t result;
7070   __asm__ ("uabdl %0.2d, %1.2s, %2.2s"
7071            : "=w"(result)
7072            : "w"(a), "w"(b)
7073            : /* No clobbers */);
7074   return result;
7075 }
7076
7077 __extension__ extern __inline int8x16_t
7078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7079 vabdq_s8 (int8x16_t a, int8x16_t b)
7080 {
7081   int8x16_t result;
7082   __asm__ ("sabd %0.16b, %1.16b, %2.16b"
7083            : "=w"(result)
7084            : "w"(a), "w"(b)
7085            : /* No clobbers */);
7086   return result;
7087 }
7088
7089 __extension__ extern __inline int16x8_t
7090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7091 vabdq_s16 (int16x8_t a, int16x8_t b)
7092 {
7093   int16x8_t result;
7094   __asm__ ("sabd %0.8h, %1.8h, %2.8h"
7095            : "=w"(result)
7096            : "w"(a), "w"(b)
7097            : /* No clobbers */);
7098   return result;
7099 }
7100
7101 __extension__ extern __inline int32x4_t
7102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7103 vabdq_s32 (int32x4_t a, int32x4_t b)
7104 {
7105   int32x4_t result;
7106   __asm__ ("sabd %0.4s, %1.4s, %2.4s"
7107            : "=w"(result)
7108            : "w"(a), "w"(b)
7109            : /* No clobbers */);
7110   return result;
7111 }
7112
7113 __extension__ extern __inline uint8x16_t
7114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7115 vabdq_u8 (uint8x16_t a, uint8x16_t b)
7116 {
7117   uint8x16_t result;
7118   __asm__ ("uabd %0.16b, %1.16b, %2.16b"
7119            : "=w"(result)
7120            : "w"(a), "w"(b)
7121            : /* No clobbers */);
7122   return result;
7123 }
7124
7125 __extension__ extern __inline uint16x8_t
7126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7127 vabdq_u16 (uint16x8_t a, uint16x8_t b)
7128 {
7129   uint16x8_t result;
7130   __asm__ ("uabd %0.8h, %1.8h, %2.8h"
7131            : "=w"(result)
7132            : "w"(a), "w"(b)
7133            : /* No clobbers */);
7134   return result;
7135 }
7136
7137 __extension__ extern __inline uint32x4_t
7138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7139 vabdq_u32 (uint32x4_t a, uint32x4_t b)
7140 {
7141   uint32x4_t result;
7142   __asm__ ("uabd %0.4s, %1.4s, %2.4s"
7143            : "=w"(result)
7144            : "w"(a), "w"(b)
7145            : /* No clobbers */);
7146   return result;
7147 }
7148
7149 __extension__ extern __inline int16_t
7150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7151 vaddlv_s8 (int8x8_t a)
7152 {
7153   int16_t result;
7154   __asm__ ("saddlv %h0,%1.8b"
7155            : "=w"(result)
7156            : "w"(a)
7157            : /* No clobbers */);
7158   return result;
7159 }
7160
7161 __extension__ extern __inline int32_t
7162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7163 vaddlv_s16 (int16x4_t a)
7164 {
7165   int32_t result;
7166   __asm__ ("saddlv %s0,%1.4h"
7167            : "=w"(result)
7168            : "w"(a)
7169            : /* No clobbers */);
7170   return result;
7171 }
7172
7173 __extension__ extern __inline uint16_t
7174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7175 vaddlv_u8 (uint8x8_t a)
7176 {
7177   uint16_t result;
7178   __asm__ ("uaddlv %h0,%1.8b"
7179            : "=w"(result)
7180            : "w"(a)
7181            : /* No clobbers */);
7182   return result;
7183 }
7184
7185 __extension__ extern __inline uint32_t
7186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7187 vaddlv_u16 (uint16x4_t a)
7188 {
7189   uint32_t result;
7190   __asm__ ("uaddlv %s0,%1.4h"
7191            : "=w"(result)
7192            : "w"(a)
7193            : /* No clobbers */);
7194   return result;
7195 }
7196
7197 __extension__ extern __inline int16_t
7198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7199 vaddlvq_s8 (int8x16_t a)
7200 {
7201   int16_t result;
7202   __asm__ ("saddlv %h0,%1.16b"
7203            : "=w"(result)
7204            : "w"(a)
7205            : /* No clobbers */);
7206   return result;
7207 }
7208
7209 __extension__ extern __inline int32_t
7210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7211 vaddlvq_s16 (int16x8_t a)
7212 {
7213   int32_t result;
7214   __asm__ ("saddlv %s0,%1.8h"
7215            : "=w"(result)
7216            : "w"(a)
7217            : /* No clobbers */);
7218   return result;
7219 }
7220
7221 __extension__ extern __inline int64_t
7222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7223 vaddlvq_s32 (int32x4_t a)
7224 {
7225   int64_t result;
7226   __asm__ ("saddlv %d0,%1.4s"
7227            : "=w"(result)
7228            : "w"(a)
7229            : /* No clobbers */);
7230   return result;
7231 }
7232
7233 __extension__ extern __inline uint16_t
7234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7235 vaddlvq_u8 (uint8x16_t a)
7236 {
7237   uint16_t result;
7238   __asm__ ("uaddlv %h0,%1.16b"
7239            : "=w"(result)
7240            : "w"(a)
7241            : /* No clobbers */);
7242   return result;
7243 }
7244
7245 __extension__ extern __inline uint32_t
7246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7247 vaddlvq_u16 (uint16x8_t a)
7248 {
7249   uint32_t result;
7250   __asm__ ("uaddlv %s0,%1.8h"
7251            : "=w"(result)
7252            : "w"(a)
7253            : /* No clobbers */);
7254   return result;
7255 }
7256
7257 __extension__ extern __inline uint64_t
7258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7259 vaddlvq_u32 (uint32x4_t a)
7260 {
7261   uint64_t result;
7262   __asm__ ("uaddlv %d0,%1.4s"
7263            : "=w"(result)
7264            : "w"(a)
7265            : /* No clobbers */);
7266   return result;
7267 }
7268
7269 __extension__ extern __inline float32x2_t
7270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7271 vcvtx_f32_f64 (float64x2_t a)
7272 {
7273   float32x2_t result;
7274   __asm__ ("fcvtxn %0.2s,%1.2d"
7275            : "=w"(result)
7276            : "w"(a)
7277            : /* No clobbers */);
7278   return result;
7279 }
7280
7281 __extension__ extern __inline float32x4_t
7282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7283 vcvtx_high_f32_f64 (float32x2_t a, float64x2_t b)
7284 {
7285   float32x4_t result;
7286   __asm__ ("fcvtxn2 %0.4s,%1.2d"
7287            : "=w"(result)
7288            : "w" (b), "0"(a)
7289            : /* No clobbers */);
7290   return result;
7291 }
7292
7293 __extension__ extern __inline float32_t
7294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7295 vcvtxd_f32_f64 (float64_t a)
7296 {
7297   float32_t result;
7298   __asm__ ("fcvtxn %s0,%d1"
7299            : "=w"(result)
7300            : "w"(a)
7301            : /* No clobbers */);
7302   return result;
7303 }
7304
7305 __extension__ extern __inline float32x2_t
7306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7307 vmla_n_f32 (float32x2_t a, float32x2_t b, float32_t c)
7308 {
7309   float32x2_t result;
7310   float32x2_t t1;
7311   __asm__ ("fmul %1.2s, %3.2s, %4.s[0]; fadd %0.2s, %0.2s, %1.2s"
7312            : "=w"(result), "=w"(t1)
7313            : "0"(a), "w"(b), "w"(c)
7314            : /* No clobbers */);
7315   return result;
7316 }
7317
7318 __extension__ extern __inline int16x4_t
7319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7320 vmla_n_s16 (int16x4_t a, int16x4_t b, int16_t c)
7321 {
7322   int16x4_t result;
7323   __asm__ ("mla %0.4h,%2.4h,%3.h[0]"
7324            : "=w"(result)
7325            : "0"(a), "w"(b), "x"(c)
7326            : /* No clobbers */);
7327   return result;
7328 }
7329
7330 __extension__ extern __inline int32x2_t
7331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7332 vmla_n_s32 (int32x2_t a, int32x2_t b, int32_t c)
7333 {
7334   int32x2_t result;
7335   __asm__ ("mla %0.2s,%2.2s,%3.s[0]"
7336            : "=w"(result)
7337            : "0"(a), "w"(b), "w"(c)
7338            : /* No clobbers */);
7339   return result;
7340 }
7341
7342 __extension__ extern __inline uint16x4_t
7343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7344 vmla_n_u16 (uint16x4_t a, uint16x4_t b, uint16_t c)
7345 {
7346   uint16x4_t result;
7347   __asm__ ("mla %0.4h,%2.4h,%3.h[0]"
7348            : "=w"(result)
7349            : "0"(a), "w"(b), "x"(c)
7350            : /* No clobbers */);
7351   return result;
7352 }
7353
7354 __extension__ extern __inline uint32x2_t
7355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7356 vmla_n_u32 (uint32x2_t a, uint32x2_t b, uint32_t c)
7357 {
7358   uint32x2_t result;
7359   __asm__ ("mla %0.2s,%2.2s,%3.s[0]"
7360            : "=w"(result)
7361            : "0"(a), "w"(b), "w"(c)
7362            : /* No clobbers */);
7363   return result;
7364 }
7365
7366 __extension__ extern __inline int8x8_t
7367 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7368 vmla_s8 (int8x8_t a, int8x8_t b, int8x8_t c)
7369 {
7370   int8x8_t result;
7371   __asm__ ("mla %0.8b, %2.8b, %3.8b"
7372            : "=w"(result)
7373            : "0"(a), "w"(b), "w"(c)
7374            : /* No clobbers */);
7375   return result;
7376 }
7377
7378 __extension__ extern __inline int16x4_t
7379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7380 vmla_s16 (int16x4_t a, int16x4_t b, int16x4_t c)
7381 {
7382   int16x4_t result;
7383   __asm__ ("mla %0.4h, %2.4h, %3.4h"
7384            : "=w"(result)
7385            : "0"(a), "w"(b), "w"(c)
7386            : /* No clobbers */);
7387   return result;
7388 }
7389
7390 __extension__ extern __inline int32x2_t
7391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7392 vmla_s32 (int32x2_t a, int32x2_t b, int32x2_t c)
7393 {
7394   int32x2_t result;
7395   __asm__ ("mla %0.2s, %2.2s, %3.2s"
7396            : "=w"(result)
7397            : "0"(a), "w"(b), "w"(c)
7398            : /* No clobbers */);
7399   return result;
7400 }
7401
7402 __extension__ extern __inline uint8x8_t
7403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7404 vmla_u8 (uint8x8_t a, uint8x8_t b, uint8x8_t c)
7405 {
7406   uint8x8_t result;
7407   __asm__ ("mla %0.8b, %2.8b, %3.8b"
7408            : "=w"(result)
7409            : "0"(a), "w"(b), "w"(c)
7410            : /* No clobbers */);
7411   return result;
7412 }
7413
7414 __extension__ extern __inline uint16x4_t
7415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7416 vmla_u16 (uint16x4_t a, uint16x4_t b, uint16x4_t c)
7417 {
7418   uint16x4_t result;
7419   __asm__ ("mla %0.4h, %2.4h, %3.4h"
7420            : "=w"(result)
7421            : "0"(a), "w"(b), "w"(c)
7422            : /* No clobbers */);
7423   return result;
7424 }
7425
7426 __extension__ extern __inline uint32x2_t
7427 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7428 vmla_u32 (uint32x2_t a, uint32x2_t b, uint32x2_t c)
7429 {
7430   uint32x2_t result;
7431   __asm__ ("mla %0.2s, %2.2s, %3.2s"
7432            : "=w"(result)
7433            : "0"(a), "w"(b), "w"(c)
7434            : /* No clobbers */);
7435   return result;
7436 }
7437
7438 #define vmlal_high_lane_s16(a, b, c, d)                                 \
7439   __extension__                                                         \
7440     ({                                                                  \
7441        int16x4_t c_ = (c);                                              \
7442        int16x8_t b_ = (b);                                              \
7443        int32x4_t a_ = (a);                                              \
7444        int32x4_t result;                                                \
7445        __asm__ ("smlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7446                 : "=w"(result)                                          \
7447                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7448                 : /* No clobbers */);                                   \
7449        result;                                                          \
7450      })
7451
7452 #define vmlal_high_lane_s32(a, b, c, d)                                 \
7453   __extension__                                                         \
7454     ({                                                                  \
7455        int32x2_t c_ = (c);                                              \
7456        int32x4_t b_ = (b);                                              \
7457        int64x2_t a_ = (a);                                              \
7458        int64x2_t result;                                                \
7459        __asm__ ("smlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7460                 : "=w"(result)                                          \
7461                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7462                 : /* No clobbers */);                                   \
7463        result;                                                          \
7464      })
7465
7466 #define vmlal_high_lane_u16(a, b, c, d)                                 \
7467   __extension__                                                         \
7468     ({                                                                  \
7469        uint16x4_t c_ = (c);                                             \
7470        uint16x8_t b_ = (b);                                             \
7471        uint32x4_t a_ = (a);                                             \
7472        uint32x4_t result;                                               \
7473        __asm__ ("umlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7474                 : "=w"(result)                                          \
7475                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7476                 : /* No clobbers */);                                   \
7477        result;                                                          \
7478      })
7479
7480 #define vmlal_high_lane_u32(a, b, c, d)                                 \
7481   __extension__                                                         \
7482     ({                                                                  \
7483        uint32x2_t c_ = (c);                                             \
7484        uint32x4_t b_ = (b);                                             \
7485        uint64x2_t a_ = (a);                                             \
7486        uint64x2_t result;                                               \
7487        __asm__ ("umlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7488                 : "=w"(result)                                          \
7489                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7490                 : /* No clobbers */);                                   \
7491        result;                                                          \
7492      })
7493
7494 #define vmlal_high_laneq_s16(a, b, c, d)                                \
7495   __extension__                                                         \
7496     ({                                                                  \
7497        int16x8_t c_ = (c);                                              \
7498        int16x8_t b_ = (b);                                              \
7499        int32x4_t a_ = (a);                                              \
7500        int32x4_t result;                                                \
7501        __asm__ ("smlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7502                 : "=w"(result)                                          \
7503                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7504                 : /* No clobbers */);                                   \
7505        result;                                                          \
7506      })
7507
7508 #define vmlal_high_laneq_s32(a, b, c, d)                                \
7509   __extension__                                                         \
7510     ({                                                                  \
7511        int32x4_t c_ = (c);                                              \
7512        int32x4_t b_ = (b);                                              \
7513        int64x2_t a_ = (a);                                              \
7514        int64x2_t result;                                                \
7515        __asm__ ("smlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7516                 : "=w"(result)                                          \
7517                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7518                 : /* No clobbers */);                                   \
7519        result;                                                          \
7520      })
7521
7522 #define vmlal_high_laneq_u16(a, b, c, d)                                \
7523   __extension__                                                         \
7524     ({                                                                  \
7525        uint16x8_t c_ = (c);                                             \
7526        uint16x8_t b_ = (b);                                             \
7527        uint32x4_t a_ = (a);                                             \
7528        uint32x4_t result;                                               \
7529        __asm__ ("umlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7530                 : "=w"(result)                                          \
7531                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7532                 : /* No clobbers */);                                   \
7533        result;                                                          \
7534      })
7535
7536 #define vmlal_high_laneq_u32(a, b, c, d)                                \
7537   __extension__                                                         \
7538     ({                                                                  \
7539        uint32x4_t c_ = (c);                                             \
7540        uint32x4_t b_ = (b);                                             \
7541        uint64x2_t a_ = (a);                                             \
7542        uint64x2_t result;                                               \
7543        __asm__ ("umlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7544                 : "=w"(result)                                          \
7545                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7546                 : /* No clobbers */);                                   \
7547        result;                                                          \
7548      })
7549
7550 __extension__ extern __inline int32x4_t
7551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7552 vmlal_high_n_s16 (int32x4_t a, int16x8_t b, int16_t c)
7553 {
7554   int32x4_t result;
7555   __asm__ ("smlal2 %0.4s,%2.8h,%3.h[0]"
7556            : "=w"(result)
7557            : "0"(a), "w"(b), "x"(c)
7558            : /* No clobbers */);
7559   return result;
7560 }
7561
7562 __extension__ extern __inline int64x2_t
7563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7564 vmlal_high_n_s32 (int64x2_t a, int32x4_t b, int32_t c)
7565 {
7566   int64x2_t result;
7567   __asm__ ("smlal2 %0.2d,%2.4s,%3.s[0]"
7568            : "=w"(result)
7569            : "0"(a), "w"(b), "w"(c)
7570            : /* No clobbers */);
7571   return result;
7572 }
7573
7574 __extension__ extern __inline uint32x4_t
7575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7576 vmlal_high_n_u16 (uint32x4_t a, uint16x8_t b, uint16_t c)
7577 {
7578   uint32x4_t result;
7579   __asm__ ("umlal2 %0.4s,%2.8h,%3.h[0]"
7580            : "=w"(result)
7581            : "0"(a), "w"(b), "x"(c)
7582            : /* No clobbers */);
7583   return result;
7584 }
7585
7586 __extension__ extern __inline uint64x2_t
7587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7588 vmlal_high_n_u32 (uint64x2_t a, uint32x4_t b, uint32_t c)
7589 {
7590   uint64x2_t result;
7591   __asm__ ("umlal2 %0.2d,%2.4s,%3.s[0]"
7592            : "=w"(result)
7593            : "0"(a), "w"(b), "w"(c)
7594            : /* No clobbers */);
7595   return result;
7596 }
7597
7598 __extension__ extern __inline int16x8_t
7599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7600 vmlal_high_s8 (int16x8_t a, int8x16_t b, int8x16_t c)
7601 {
7602   int16x8_t result;
7603   __asm__ ("smlal2 %0.8h,%2.16b,%3.16b"
7604            : "=w"(result)
7605            : "0"(a), "w"(b), "w"(c)
7606            : /* No clobbers */);
7607   return result;
7608 }
7609
7610 __extension__ extern __inline int32x4_t
7611 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7612 vmlal_high_s16 (int32x4_t a, int16x8_t b, int16x8_t c)
7613 {
7614   int32x4_t result;
7615   __asm__ ("smlal2 %0.4s,%2.8h,%3.8h"
7616            : "=w"(result)
7617            : "0"(a), "w"(b), "w"(c)
7618            : /* No clobbers */);
7619   return result;
7620 }
7621
7622 __extension__ extern __inline int64x2_t
7623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7624 vmlal_high_s32 (int64x2_t a, int32x4_t b, int32x4_t c)
7625 {
7626   int64x2_t result;
7627   __asm__ ("smlal2 %0.2d,%2.4s,%3.4s"
7628            : "=w"(result)
7629            : "0"(a), "w"(b), "w"(c)
7630            : /* No clobbers */);
7631   return result;
7632 }
7633
7634 __extension__ extern __inline uint16x8_t
7635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7636 vmlal_high_u8 (uint16x8_t a, uint8x16_t b, uint8x16_t c)
7637 {
7638   uint16x8_t result;
7639   __asm__ ("umlal2 %0.8h,%2.16b,%3.16b"
7640            : "=w"(result)
7641            : "0"(a), "w"(b), "w"(c)
7642            : /* No clobbers */);
7643   return result;
7644 }
7645
7646 __extension__ extern __inline uint32x4_t
7647 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7648 vmlal_high_u16 (uint32x4_t a, uint16x8_t b, uint16x8_t c)
7649 {
7650   uint32x4_t result;
7651   __asm__ ("umlal2 %0.4s,%2.8h,%3.8h"
7652            : "=w"(result)
7653            : "0"(a), "w"(b), "w"(c)
7654            : /* No clobbers */);
7655   return result;
7656 }
7657
7658 __extension__ extern __inline uint64x2_t
7659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7660 vmlal_high_u32 (uint64x2_t a, uint32x4_t b, uint32x4_t c)
7661 {
7662   uint64x2_t result;
7663   __asm__ ("umlal2 %0.2d,%2.4s,%3.4s"
7664            : "=w"(result)
7665            : "0"(a), "w"(b), "w"(c)
7666            : /* No clobbers */);
7667   return result;
7668 }
7669
7670 #define vmlal_lane_s16(a, b, c, d)                                      \
7671   __extension__                                                         \
7672     ({                                                                  \
7673        int16x4_t c_ = (c);                                              \
7674        int16x4_t b_ = (b);                                              \
7675        int32x4_t a_ = (a);                                              \
7676        int32x4_t result;                                                \
7677        __asm__ ("smlal %0.4s,%2.4h,%3.h[%4]"                            \
7678                 : "=w"(result)                                          \
7679                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7680                 : /* No clobbers */);                                   \
7681        result;                                                          \
7682      })
7683
7684 #define vmlal_lane_s32(a, b, c, d)                                      \
7685   __extension__                                                         \
7686     ({                                                                  \
7687        int32x2_t c_ = (c);                                              \
7688        int32x2_t b_ = (b);                                              \
7689        int64x2_t a_ = (a);                                              \
7690        int64x2_t result;                                                \
7691        __asm__ ("smlal %0.2d,%2.2s,%3.s[%4]"                            \
7692                 : "=w"(result)                                          \
7693                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7694                 : /* No clobbers */);                                   \
7695        result;                                                          \
7696      })
7697
7698 #define vmlal_lane_u16(a, b, c, d)                                      \
7699   __extension__                                                         \
7700     ({                                                                  \
7701        uint16x4_t c_ = (c);                                             \
7702        uint16x4_t b_ = (b);                                             \
7703        uint32x4_t a_ = (a);                                             \
7704        uint32x4_t result;                                               \
7705        __asm__ ("umlal %0.4s,%2.4h,%3.h[%4]"                            \
7706                 : "=w"(result)                                          \
7707                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7708                 : /* No clobbers */);                                   \
7709        result;                                                          \
7710      })
7711
7712 #define vmlal_lane_u32(a, b, c, d)                                      \
7713   __extension__                                                         \
7714     ({                                                                  \
7715        uint32x2_t c_ = (c);                                             \
7716        uint32x2_t b_ = (b);                                             \
7717        uint64x2_t a_ = (a);                                             \
7718        uint64x2_t result;                                               \
7719        __asm__ ("umlal %0.2d, %2.2s, %3.s[%4]"                          \
7720                 : "=w"(result)                                          \
7721                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7722                 : /* No clobbers */);                                   \
7723        result;                                                          \
7724      })
7725
7726 #define vmlal_laneq_s16(a, b, c, d)                                     \
7727   __extension__                                                         \
7728     ({                                                                  \
7729        int16x8_t c_ = (c);                                              \
7730        int16x4_t b_ = (b);                                              \
7731        int32x4_t a_ = (a);                                              \
7732        int32x4_t result;                                                \
7733        __asm__ ("smlal %0.4s, %2.4h, %3.h[%4]"                          \
7734                 : "=w"(result)                                          \
7735                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7736                 : /* No clobbers */);                                   \
7737        result;                                                          \
7738      })
7739
7740 #define vmlal_laneq_s32(a, b, c, d)                                     \
7741   __extension__                                                         \
7742     ({                                                                  \
7743        int32x4_t c_ = (c);                                              \
7744        int32x2_t b_ = (b);                                              \
7745        int64x2_t a_ = (a);                                              \
7746        int64x2_t result;                                                \
7747        __asm__ ("smlal %0.2d, %2.2s, %3.s[%4]"                          \
7748                 : "=w"(result)                                          \
7749                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7750                 : /* No clobbers */);                                   \
7751        result;                                                          \
7752      })
7753
7754 #define vmlal_laneq_u16(a, b, c, d)                                     \
7755   __extension__                                                         \
7756     ({                                                                  \
7757        uint16x8_t c_ = (c);                                             \
7758        uint16x4_t b_ = (b);                                             \
7759        uint32x4_t a_ = (a);                                             \
7760        uint32x4_t result;                                               \
7761        __asm__ ("umlal %0.4s, %2.4h, %3.h[%4]"                          \
7762                 : "=w"(result)                                          \
7763                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7764                 : /* No clobbers */);                                   \
7765        result;                                                          \
7766      })
7767
7768 #define vmlal_laneq_u32(a, b, c, d)                                     \
7769   __extension__                                                         \
7770     ({                                                                  \
7771        uint32x4_t c_ = (c);                                             \
7772        uint32x2_t b_ = (b);                                             \
7773        uint64x2_t a_ = (a);                                             \
7774        uint64x2_t result;                                               \
7775        __asm__ ("umlal %0.2d, %2.2s, %3.s[%4]"                          \
7776                 : "=w"(result)                                          \
7777                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7778                 : /* No clobbers */);                                   \
7779        result;                                                          \
7780      })
7781
7782 __extension__ extern __inline int32x4_t
7783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7784 vmlal_n_s16 (int32x4_t a, int16x4_t b, int16_t c)
7785 {
7786   int32x4_t result;
7787   __asm__ ("smlal %0.4s,%2.4h,%3.h[0]"
7788            : "=w"(result)
7789            : "0"(a), "w"(b), "x"(c)
7790            : /* No clobbers */);
7791   return result;
7792 }
7793
7794 __extension__ extern __inline int64x2_t
7795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7796 vmlal_n_s32 (int64x2_t a, int32x2_t b, int32_t c)
7797 {
7798   int64x2_t result;
7799   __asm__ ("smlal %0.2d,%2.2s,%3.s[0]"
7800            : "=w"(result)
7801            : "0"(a), "w"(b), "w"(c)
7802            : /* No clobbers */);
7803   return result;
7804 }
7805
7806 __extension__ extern __inline uint32x4_t
7807 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7808 vmlal_n_u16 (uint32x4_t a, uint16x4_t b, uint16_t c)
7809 {
7810   uint32x4_t result;
7811   __asm__ ("umlal %0.4s,%2.4h,%3.h[0]"
7812            : "=w"(result)
7813            : "0"(a), "w"(b), "x"(c)
7814            : /* No clobbers */);
7815   return result;
7816 }
7817
7818 __extension__ extern __inline uint64x2_t
7819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7820 vmlal_n_u32 (uint64x2_t a, uint32x2_t b, uint32_t c)
7821 {
7822   uint64x2_t result;
7823   __asm__ ("umlal %0.2d,%2.2s,%3.s[0]"
7824            : "=w"(result)
7825            : "0"(a), "w"(b), "w"(c)
7826            : /* No clobbers */);
7827   return result;
7828 }
7829
7830 __extension__ extern __inline int16x8_t
7831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7832 vmlal_s8 (int16x8_t a, int8x8_t b, int8x8_t c)
7833 {
7834   int16x8_t result;
7835   __asm__ ("smlal %0.8h,%2.8b,%3.8b"
7836            : "=w"(result)
7837            : "0"(a), "w"(b), "w"(c)
7838            : /* No clobbers */);
7839   return result;
7840 }
7841
7842 __extension__ extern __inline int32x4_t
7843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7844 vmlal_s16 (int32x4_t a, int16x4_t b, int16x4_t c)
7845 {
7846   int32x4_t result;
7847   __asm__ ("smlal %0.4s,%2.4h,%3.4h"
7848            : "=w"(result)
7849            : "0"(a), "w"(b), "w"(c)
7850            : /* No clobbers */);
7851   return result;
7852 }
7853
7854 __extension__ extern __inline int64x2_t
7855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7856 vmlal_s32 (int64x2_t a, int32x2_t b, int32x2_t c)
7857 {
7858   int64x2_t result;
7859   __asm__ ("smlal %0.2d,%2.2s,%3.2s"
7860            : "=w"(result)
7861            : "0"(a), "w"(b), "w"(c)
7862            : /* No clobbers */);
7863   return result;
7864 }
7865
7866 __extension__ extern __inline uint16x8_t
7867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7868 vmlal_u8 (uint16x8_t a, uint8x8_t b, uint8x8_t c)
7869 {
7870   uint16x8_t result;
7871   __asm__ ("umlal %0.8h,%2.8b,%3.8b"
7872            : "=w"(result)
7873            : "0"(a), "w"(b), "w"(c)
7874            : /* No clobbers */);
7875   return result;
7876 }
7877
7878 __extension__ extern __inline uint32x4_t
7879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7880 vmlal_u16 (uint32x4_t a, uint16x4_t b, uint16x4_t c)
7881 {
7882   uint32x4_t result;
7883   __asm__ ("umlal %0.4s,%2.4h,%3.4h"
7884            : "=w"(result)
7885            : "0"(a), "w"(b), "w"(c)
7886            : /* No clobbers */);
7887   return result;
7888 }
7889
7890 __extension__ extern __inline uint64x2_t
7891 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7892 vmlal_u32 (uint64x2_t a, uint32x2_t b, uint32x2_t c)
7893 {
7894   uint64x2_t result;
7895   __asm__ ("umlal %0.2d,%2.2s,%3.2s"
7896            : "=w"(result)
7897            : "0"(a), "w"(b), "w"(c)
7898            : /* No clobbers */);
7899   return result;
7900 }
7901
7902 __extension__ extern __inline float32x4_t
7903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7904 vmlaq_n_f32 (float32x4_t a, float32x4_t b, float32_t c)
7905 {
7906   float32x4_t result;
7907   float32x4_t t1;
7908   __asm__ ("fmul %1.4s, %3.4s, %4.s[0]; fadd %0.4s, %0.4s, %1.4s"
7909            : "=w"(result), "=w"(t1)
7910            : "0"(a), "w"(b), "w"(c)
7911            : /* No clobbers */);
7912   return result;
7913 }
7914
7915 __extension__ extern __inline int16x8_t
7916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7917 vmlaq_n_s16 (int16x8_t a, int16x8_t b, int16_t c)
7918 {
7919   int16x8_t result;
7920   __asm__ ("mla %0.8h,%2.8h,%3.h[0]"
7921            : "=w"(result)
7922            : "0"(a), "w"(b), "x"(c)
7923            : /* No clobbers */);
7924   return result;
7925 }
7926
7927 __extension__ extern __inline int32x4_t
7928 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7929 vmlaq_n_s32 (int32x4_t a, int32x4_t b, int32_t c)
7930 {
7931   int32x4_t result;
7932   __asm__ ("mla %0.4s,%2.4s,%3.s[0]"
7933            : "=w"(result)
7934            : "0"(a), "w"(b), "w"(c)
7935            : /* No clobbers */);
7936   return result;
7937 }
7938
7939 __extension__ extern __inline uint16x8_t
7940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7941 vmlaq_n_u16 (uint16x8_t a, uint16x8_t b, uint16_t c)
7942 {
7943   uint16x8_t result;
7944   __asm__ ("mla %0.8h,%2.8h,%3.h[0]"
7945            : "=w"(result)
7946            : "0"(a), "w"(b), "x"(c)
7947            : /* No clobbers */);
7948   return result;
7949 }
7950
7951 __extension__ extern __inline uint32x4_t
7952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7953 vmlaq_n_u32 (uint32x4_t a, uint32x4_t b, uint32_t c)
7954 {
7955   uint32x4_t result;
7956   __asm__ ("mla %0.4s,%2.4s,%3.s[0]"
7957            : "=w"(result)
7958            : "0"(a), "w"(b), "w"(c)
7959            : /* No clobbers */);
7960   return result;
7961 }
7962
7963 __extension__ extern __inline int8x16_t
7964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7965 vmlaq_s8 (int8x16_t a, int8x16_t b, int8x16_t c)
7966 {
7967   int8x16_t result;
7968   __asm__ ("mla %0.16b, %2.16b, %3.16b"
7969            : "=w"(result)
7970            : "0"(a), "w"(b), "w"(c)
7971            : /* No clobbers */);
7972   return result;
7973 }
7974
7975 __extension__ extern __inline int16x8_t
7976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7977 vmlaq_s16 (int16x8_t a, int16x8_t b, int16x8_t c)
7978 {
7979   int16x8_t result;
7980   __asm__ ("mla %0.8h, %2.8h, %3.8h"
7981            : "=w"(result)
7982            : "0"(a), "w"(b), "w"(c)
7983            : /* No clobbers */);
7984   return result;
7985 }
7986
7987 __extension__ extern __inline int32x4_t
7988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7989 vmlaq_s32 (int32x4_t a, int32x4_t b, int32x4_t c)
7990 {
7991   int32x4_t result;
7992   __asm__ ("mla %0.4s, %2.4s, %3.4s"
7993            : "=w"(result)
7994            : "0"(a), "w"(b), "w"(c)
7995            : /* No clobbers */);
7996   return result;
7997 }
7998
7999 __extension__ extern __inline uint8x16_t
8000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8001 vmlaq_u8 (uint8x16_t a, uint8x16_t b, uint8x16_t c)
8002 {
8003   uint8x16_t result;
8004   __asm__ ("mla %0.16b, %2.16b, %3.16b"
8005            : "=w"(result)
8006            : "0"(a), "w"(b), "w"(c)
8007            : /* No clobbers */);
8008   return result;
8009 }
8010
8011 __extension__ extern __inline uint16x8_t
8012 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8013 vmlaq_u16 (uint16x8_t a, uint16x8_t b, uint16x8_t c)
8014 {
8015   uint16x8_t result;
8016   __asm__ ("mla %0.8h, %2.8h, %3.8h"
8017            : "=w"(result)
8018            : "0"(a), "w"(b), "w"(c)
8019            : /* No clobbers */);
8020   return result;
8021 }
8022
8023 __extension__ extern __inline uint32x4_t
8024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8025 vmlaq_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t c)
8026 {
8027   uint32x4_t result;
8028   __asm__ ("mla %0.4s, %2.4s, %3.4s"
8029            : "=w"(result)
8030            : "0"(a), "w"(b), "w"(c)
8031            : /* No clobbers */);
8032   return result;
8033 }
8034
8035 __extension__ extern __inline float32x2_t
8036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8037 vmls_n_f32 (float32x2_t a, float32x2_t b, float32_t c)
8038 {
8039   float32x2_t result;
8040   float32x2_t t1;
8041   __asm__ ("fmul %1.2s, %3.2s, %4.s[0]; fsub %0.2s, %0.2s, %1.2s"
8042            : "=w"(result), "=w"(t1)
8043            : "0"(a), "w"(b), "w"(c)
8044            : /* No clobbers */);
8045   return result;
8046 }
8047
8048 __extension__ extern __inline int16x4_t
8049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8050 vmls_n_s16 (int16x4_t a, int16x4_t b, int16_t c)
8051 {
8052   int16x4_t result;
8053   __asm__ ("mls %0.4h, %2.4h, %3.h[0]"
8054            : "=w"(result)
8055            : "0"(a), "w"(b), "x"(c)
8056            : /* No clobbers */);
8057   return result;
8058 }
8059
8060 __extension__ extern __inline int32x2_t
8061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8062 vmls_n_s32 (int32x2_t a, int32x2_t b, int32_t c)
8063 {
8064   int32x2_t result;
8065   __asm__ ("mls %0.2s, %2.2s, %3.s[0]"
8066            : "=w"(result)
8067            : "0"(a), "w"(b), "w"(c)
8068            : /* No clobbers */);
8069   return result;
8070 }
8071
8072 __extension__ extern __inline uint16x4_t
8073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8074 vmls_n_u16 (uint16x4_t a, uint16x4_t b, uint16_t c)
8075 {
8076   uint16x4_t result;
8077   __asm__ ("mls %0.4h, %2.4h, %3.h[0]"
8078            : "=w"(result)
8079            : "0"(a), "w"(b), "x"(c)
8080            : /* No clobbers */);
8081   return result;
8082 }
8083
8084 __extension__ extern __inline uint32x2_t
8085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8086 vmls_n_u32 (uint32x2_t a, uint32x2_t b, uint32_t c)
8087 {
8088   uint32x2_t result;
8089   __asm__ ("mls %0.2s, %2.2s, %3.s[0]"
8090            : "=w"(result)
8091            : "0"(a), "w"(b), "w"(c)
8092            : /* No clobbers */);
8093   return result;
8094 }
8095
8096 __extension__ extern __inline int8x8_t
8097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8098 vmls_s8 (int8x8_t a, int8x8_t b, int8x8_t c)
8099 {
8100   int8x8_t result;
8101   __asm__ ("mls %0.8b,%2.8b,%3.8b"
8102            : "=w"(result)
8103            : "0"(a), "w"(b), "w"(c)
8104            : /* No clobbers */);
8105   return result;
8106 }
8107
8108 __extension__ extern __inline int16x4_t
8109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8110 vmls_s16 (int16x4_t a, int16x4_t b, int16x4_t c)
8111 {
8112   int16x4_t result;
8113   __asm__ ("mls %0.4h,%2.4h,%3.4h"
8114            : "=w"(result)
8115            : "0"(a), "w"(b), "w"(c)
8116            : /* No clobbers */);
8117   return result;
8118 }
8119
8120 __extension__ extern __inline int32x2_t
8121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8122 vmls_s32 (int32x2_t a, int32x2_t b, int32x2_t c)
8123 {
8124   int32x2_t result;
8125   __asm__ ("mls %0.2s,%2.2s,%3.2s"
8126            : "=w"(result)
8127            : "0"(a), "w"(b), "w"(c)
8128            : /* No clobbers */);
8129   return result;
8130 }
8131
8132 __extension__ extern __inline uint8x8_t
8133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8134 vmls_u8 (uint8x8_t a, uint8x8_t b, uint8x8_t c)
8135 {
8136   uint8x8_t result;
8137   __asm__ ("mls %0.8b,%2.8b,%3.8b"
8138            : "=w"(result)
8139            : "0"(a), "w"(b), "w"(c)
8140            : /* No clobbers */);
8141   return result;
8142 }
8143
8144 __extension__ extern __inline uint16x4_t
8145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8146 vmls_u16 (uint16x4_t a, uint16x4_t b, uint16x4_t c)
8147 {
8148   uint16x4_t result;
8149   __asm__ ("mls %0.4h,%2.4h,%3.4h"
8150            : "=w"(result)
8151            : "0"(a), "w"(b), "w"(c)
8152            : /* No clobbers */);
8153   return result;
8154 }
8155
8156 __extension__ extern __inline uint32x2_t
8157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8158 vmls_u32 (uint32x2_t a, uint32x2_t b, uint32x2_t c)
8159 {
8160   uint32x2_t result;
8161   __asm__ ("mls %0.2s,%2.2s,%3.2s"
8162            : "=w"(result)
8163            : "0"(a), "w"(b), "w"(c)
8164            : /* No clobbers */);
8165   return result;
8166 }
8167
8168 #define vmlsl_high_lane_s16(a, b, c, d)                                 \
8169   __extension__                                                         \
8170     ({                                                                  \
8171        int16x4_t c_ = (c);                                              \
8172        int16x8_t b_ = (b);                                              \
8173        int32x4_t a_ = (a);                                              \
8174        int32x4_t result;                                                \
8175        __asm__ ("smlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8176                 : "=w"(result)                                          \
8177                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8178                 : /* No clobbers */);                                   \
8179        result;                                                          \
8180      })
8181
8182 #define vmlsl_high_lane_s32(a, b, c, d)                                 \
8183   __extension__                                                         \
8184     ({                                                                  \
8185        int32x2_t c_ = (c);                                              \
8186        int32x4_t b_ = (b);                                              \
8187        int64x2_t a_ = (a);                                              \
8188        int64x2_t result;                                                \
8189        __asm__ ("smlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8190                 : "=w"(result)                                          \
8191                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8192                 : /* No clobbers */);                                   \
8193        result;                                                          \
8194      })
8195
8196 #define vmlsl_high_lane_u16(a, b, c, d)                                 \
8197   __extension__                                                         \
8198     ({                                                                  \
8199        uint16x4_t c_ = (c);                                             \
8200        uint16x8_t b_ = (b);                                             \
8201        uint32x4_t a_ = (a);                                             \
8202        uint32x4_t result;                                               \
8203        __asm__ ("umlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8204                 : "=w"(result)                                          \
8205                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8206                 : /* No clobbers */);                                   \
8207        result;                                                          \
8208      })
8209
8210 #define vmlsl_high_lane_u32(a, b, c, d)                                 \
8211   __extension__                                                         \
8212     ({                                                                  \
8213        uint32x2_t c_ = (c);                                             \
8214        uint32x4_t b_ = (b);                                             \
8215        uint64x2_t a_ = (a);                                             \
8216        uint64x2_t result;                                               \
8217        __asm__ ("umlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8218                 : "=w"(result)                                          \
8219                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8220                 : /* No clobbers */);                                   \
8221        result;                                                          \
8222      })
8223
8224 #define vmlsl_high_laneq_s16(a, b, c, d)                                \
8225   __extension__                                                         \
8226     ({                                                                  \
8227        int16x8_t c_ = (c);                                              \
8228        int16x8_t b_ = (b);                                              \
8229        int32x4_t a_ = (a);                                              \
8230        int32x4_t result;                                                \
8231        __asm__ ("smlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8232                 : "=w"(result)                                          \
8233                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8234                 : /* No clobbers */);                                   \
8235        result;                                                          \
8236      })
8237
8238 #define vmlsl_high_laneq_s32(a, b, c, d)                                \
8239   __extension__                                                         \
8240     ({                                                                  \
8241        int32x4_t c_ = (c);                                              \
8242        int32x4_t b_ = (b);                                              \
8243        int64x2_t a_ = (a);                                              \
8244        int64x2_t result;                                                \
8245        __asm__ ("smlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8246                 : "=w"(result)                                          \
8247                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8248                 : /* No clobbers */);                                   \
8249        result;                                                          \
8250      })
8251
8252 #define vmlsl_high_laneq_u16(a, b, c, d)                                \
8253   __extension__                                                         \
8254     ({                                                                  \
8255        uint16x8_t c_ = (c);                                             \
8256        uint16x8_t b_ = (b);                                             \
8257        uint32x4_t a_ = (a);                                             \
8258        uint32x4_t result;                                               \
8259        __asm__ ("umlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8260                 : "=w"(result)                                          \
8261                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8262                 : /* No clobbers */);                                   \
8263        result;                                                          \
8264      })
8265
8266 #define vmlsl_high_laneq_u32(a, b, c, d)                                \
8267   __extension__                                                         \
8268     ({                                                                  \
8269        uint32x4_t c_ = (c);                                             \
8270        uint32x4_t b_ = (b);                                             \
8271        uint64x2_t a_ = (a);                                             \
8272        uint64x2_t result;                                               \
8273        __asm__ ("umlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8274                 : "=w"(result)                                          \
8275                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8276                 : /* No clobbers */);                                   \
8277        result;                                                          \
8278      })
8279
8280 __extension__ extern __inline int32x4_t
8281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8282 vmlsl_high_n_s16 (int32x4_t a, int16x8_t b, int16_t c)
8283 {
8284   int32x4_t result;
8285   __asm__ ("smlsl2 %0.4s, %2.8h, %3.h[0]"
8286            : "=w"(result)
8287            : "0"(a), "w"(b), "x"(c)
8288            : /* No clobbers */);
8289   return result;
8290 }
8291
8292 __extension__ extern __inline int64x2_t
8293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8294 vmlsl_high_n_s32 (int64x2_t a, int32x4_t b, int32_t c)
8295 {
8296   int64x2_t result;
8297   __asm__ ("smlsl2 %0.2d, %2.4s, %3.s[0]"
8298            : "=w"(result)
8299            : "0"(a), "w"(b), "w"(c)
8300            : /* No clobbers */);
8301   return result;
8302 }
8303
8304 __extension__ extern __inline uint32x4_t
8305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8306 vmlsl_high_n_u16 (uint32x4_t a, uint16x8_t b, uint16_t c)
8307 {
8308   uint32x4_t result;
8309   __asm__ ("umlsl2 %0.4s, %2.8h, %3.h[0]"
8310            : "=w"(result)
8311            : "0"(a), "w"(b), "x"(c)
8312            : /* No clobbers */);
8313   return result;
8314 }
8315
8316 __extension__ extern __inline uint64x2_t
8317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8318 vmlsl_high_n_u32 (uint64x2_t a, uint32x4_t b, uint32_t c)
8319 {
8320   uint64x2_t result;
8321   __asm__ ("umlsl2 %0.2d, %2.4s, %3.s[0]"
8322            : "=w"(result)
8323            : "0"(a), "w"(b), "w"(c)
8324            : /* No clobbers */);
8325   return result;
8326 }
8327
8328 __extension__ extern __inline int16x8_t
8329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8330 vmlsl_high_s8 (int16x8_t a, int8x16_t b, int8x16_t c)
8331 {
8332   int16x8_t result;
8333   __asm__ ("smlsl2 %0.8h,%2.16b,%3.16b"
8334            : "=w"(result)
8335            : "0"(a), "w"(b), "w"(c)
8336            : /* No clobbers */);
8337   return result;
8338 }
8339
8340 __extension__ extern __inline int32x4_t
8341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8342 vmlsl_high_s16 (int32x4_t a, int16x8_t b, int16x8_t c)
8343 {
8344   int32x4_t result;
8345   __asm__ ("smlsl2 %0.4s,%2.8h,%3.8h"
8346            : "=w"(result)
8347            : "0"(a), "w"(b), "w"(c)
8348            : /* No clobbers */);
8349   return result;
8350 }
8351
8352 __extension__ extern __inline int64x2_t
8353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8354 vmlsl_high_s32 (int64x2_t a, int32x4_t b, int32x4_t c)
8355 {
8356   int64x2_t result;
8357   __asm__ ("smlsl2 %0.2d,%2.4s,%3.4s"
8358            : "=w"(result)
8359            : "0"(a), "w"(b), "w"(c)
8360            : /* No clobbers */);
8361   return result;
8362 }
8363
8364 __extension__ extern __inline uint16x8_t
8365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8366 vmlsl_high_u8 (uint16x8_t a, uint8x16_t b, uint8x16_t c)
8367 {
8368   uint16x8_t result;
8369   __asm__ ("umlsl2 %0.8h,%2.16b,%3.16b"
8370            : "=w"(result)
8371            : "0"(a), "w"(b), "w"(c)
8372            : /* No clobbers */);
8373   return result;
8374 }
8375
8376 __extension__ extern __inline uint32x4_t
8377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8378 vmlsl_high_u16 (uint32x4_t a, uint16x8_t b, uint16x8_t c)
8379 {
8380   uint32x4_t result;
8381   __asm__ ("umlsl2 %0.4s,%2.8h,%3.8h"
8382            : "=w"(result)
8383            : "0"(a), "w"(b), "w"(c)
8384            : /* No clobbers */);
8385   return result;
8386 }
8387
8388 __extension__ extern __inline uint64x2_t
8389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8390 vmlsl_high_u32 (uint64x2_t a, uint32x4_t b, uint32x4_t c)
8391 {
8392   uint64x2_t result;
8393   __asm__ ("umlsl2 %0.2d,%2.4s,%3.4s"
8394            : "=w"(result)
8395            : "0"(a), "w"(b), "w"(c)
8396            : /* No clobbers */);
8397   return result;
8398 }
8399
8400 #define vmlsl_lane_s16(a, b, c, d)                                      \
8401   __extension__                                                         \
8402     ({                                                                  \
8403        int16x4_t c_ = (c);                                              \
8404        int16x4_t b_ = (b);                                              \
8405        int32x4_t a_ = (a);                                              \
8406        int32x4_t result;                                                \
8407        __asm__ ("smlsl %0.4s, %2.4h, %3.h[%4]"                          \
8408                 : "=w"(result)                                          \
8409                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8410                 : /* No clobbers */);                                   \
8411        result;                                                          \
8412      })
8413
8414 #define vmlsl_lane_s32(a, b, c, d)                                      \
8415   __extension__                                                         \
8416     ({                                                                  \
8417        int32x2_t c_ = (c);                                              \
8418        int32x2_t b_ = (b);                                              \
8419        int64x2_t a_ = (a);                                              \
8420        int64x2_t result;                                                \
8421        __asm__ ("smlsl %0.2d, %2.2s, %3.s[%4]"                          \
8422                 : "=w"(result)                                          \
8423                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8424                 : /* No clobbers */);                                   \
8425        result;                                                          \
8426      })
8427
8428 #define vmlsl_lane_u16(a, b, c, d)                                      \
8429   __extension__                                                         \
8430     ({                                                                  \
8431        uint16x4_t c_ = (c);                                             \
8432        uint16x4_t b_ = (b);                                             \
8433        uint32x4_t a_ = (a);                                             \
8434        uint32x4_t result;                                               \
8435        __asm__ ("umlsl %0.4s, %2.4h, %3.h[%4]"                          \
8436                 : "=w"(result)                                          \
8437                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8438                 : /* No clobbers */);                                   \
8439        result;                                                          \
8440      })
8441
8442 #define vmlsl_lane_u32(a, b, c, d)                                      \
8443   __extension__                                                         \
8444     ({                                                                  \
8445        uint32x2_t c_ = (c);                                             \
8446        uint32x2_t b_ = (b);                                             \
8447        uint64x2_t a_ = (a);                                             \
8448        uint64x2_t result;                                               \
8449        __asm__ ("umlsl %0.2d, %2.2s, %3.s[%4]"                          \
8450                 : "=w"(result)                                          \
8451                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8452                 : /* No clobbers */);                                   \
8453        result;                                                          \
8454      })
8455
8456 #define vmlsl_laneq_s16(a, b, c, d)                                     \
8457   __extension__                                                         \
8458     ({                                                                  \
8459        int16x8_t c_ = (c);                                              \
8460        int16x4_t b_ = (b);                                              \
8461        int32x4_t a_ = (a);                                              \
8462        int32x4_t result;                                                \
8463        __asm__ ("smlsl %0.4s, %2.4h, %3.h[%4]"                          \
8464                 : "=w"(result)                                          \
8465                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8466                 : /* No clobbers */);                                   \
8467        result;                                                          \
8468      })
8469
8470 #define vmlsl_laneq_s32(a, b, c, d)                                     \
8471   __extension__                                                         \
8472     ({                                                                  \
8473        int32x4_t c_ = (c);                                              \
8474        int32x2_t b_ = (b);                                              \
8475        int64x2_t a_ = (a);                                              \
8476        int64x2_t result;                                                \
8477        __asm__ ("smlsl %0.2d, %2.2s, %3.s[%4]"                          \
8478                 : "=w"(result)                                          \
8479                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8480                 : /* No clobbers */);                                   \
8481        result;                                                          \
8482      })
8483
8484 #define vmlsl_laneq_u16(a, b, c, d)                                     \
8485   __extension__                                                         \
8486     ({                                                                  \
8487        uint16x8_t c_ = (c);                                             \
8488        uint16x4_t b_ = (b);                                             \
8489        uint32x4_t a_ = (a);                                             \
8490        uint32x4_t result;                                               \
8491        __asm__ ("umlsl %0.4s, %2.4h, %3.h[%4]"                          \
8492                 : "=w"(result)                                          \
8493                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8494                 : /* No clobbers */);                                   \
8495        result;                                                          \
8496      })
8497
8498 #define vmlsl_laneq_u32(a, b, c, d)                                     \
8499   __extension__                                                         \
8500     ({                                                                  \
8501        uint32x4_t c_ = (c);                                             \
8502        uint32x2_t b_ = (b);                                             \
8503        uint64x2_t a_ = (a);                                             \
8504        uint64x2_t result;                                               \
8505        __asm__ ("umlsl %0.2d, %2.2s, %3.s[%4]"                          \
8506                 : "=w"(result)                                          \
8507                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8508                 : /* No clobbers */);                                   \
8509        result;                                                          \
8510      })
8511
8512 __extension__ extern __inline int32x4_t
8513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8514 vmlsl_n_s16 (int32x4_t a, int16x4_t b, int16_t c)
8515 {
8516   int32x4_t result;
8517   __asm__ ("smlsl %0.4s, %2.4h, %3.h[0]"
8518            : "=w"(result)
8519            : "0"(a), "w"(b), "x"(c)
8520            : /* No clobbers */);
8521   return result;
8522 }
8523
8524 __extension__ extern __inline int64x2_t
8525 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8526 vmlsl_n_s32 (int64x2_t a, int32x2_t b, int32_t c)
8527 {
8528   int64x2_t result;
8529   __asm__ ("smlsl %0.2d, %2.2s, %3.s[0]"
8530            : "=w"(result)
8531            : "0"(a), "w"(b), "w"(c)
8532            : /* No clobbers */);
8533   return result;
8534 }
8535
8536 __extension__ extern __inline uint32x4_t
8537 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8538 vmlsl_n_u16 (uint32x4_t a, uint16x4_t b, uint16_t c)
8539 {
8540   uint32x4_t result;
8541   __asm__ ("umlsl %0.4s, %2.4h, %3.h[0]"
8542            : "=w"(result)
8543            : "0"(a), "w"(b), "x"(c)
8544            : /* No clobbers */);
8545   return result;
8546 }
8547
8548 __extension__ extern __inline uint64x2_t
8549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8550 vmlsl_n_u32 (uint64x2_t a, uint32x2_t b, uint32_t c)
8551 {
8552   uint64x2_t result;
8553   __asm__ ("umlsl %0.2d, %2.2s, %3.s[0]"
8554            : "=w"(result)
8555            : "0"(a), "w"(b), "w"(c)
8556            : /* No clobbers */);
8557   return result;
8558 }
8559
8560 __extension__ extern __inline int16x8_t
8561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8562 vmlsl_s8 (int16x8_t a, int8x8_t b, int8x8_t c)
8563 {
8564   int16x8_t result;
8565   __asm__ ("smlsl %0.8h, %2.8b, %3.8b"
8566            : "=w"(result)
8567            : "0"(a), "w"(b), "w"(c)
8568            : /* No clobbers */);
8569   return result;
8570 }
8571
8572 __extension__ extern __inline int32x4_t
8573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8574 vmlsl_s16 (int32x4_t a, int16x4_t b, int16x4_t c)
8575 {
8576   int32x4_t result;
8577   __asm__ ("smlsl %0.4s, %2.4h, %3.4h"
8578            : "=w"(result)
8579            : "0"(a), "w"(b), "w"(c)
8580            : /* No clobbers */);
8581   return result;
8582 }
8583
8584 __extension__ extern __inline int64x2_t
8585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8586 vmlsl_s32 (int64x2_t a, int32x2_t b, int32x2_t c)
8587 {
8588   int64x2_t result;
8589   __asm__ ("smlsl %0.2d, %2.2s, %3.2s"
8590            : "=w"(result)
8591            : "0"(a), "w"(b), "w"(c)
8592            : /* No clobbers */);
8593   return result;
8594 }
8595
8596 __extension__ extern __inline uint16x8_t
8597 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8598 vmlsl_u8 (uint16x8_t a, uint8x8_t b, uint8x8_t c)
8599 {
8600   uint16x8_t result;
8601   __asm__ ("umlsl %0.8h, %2.8b, %3.8b"
8602            : "=w"(result)
8603            : "0"(a), "w"(b), "w"(c)
8604            : /* No clobbers */);
8605   return result;
8606 }
8607
8608 __extension__ extern __inline uint32x4_t
8609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8610 vmlsl_u16 (uint32x4_t a, uint16x4_t b, uint16x4_t c)
8611 {
8612   uint32x4_t result;
8613   __asm__ ("umlsl %0.4s, %2.4h, %3.4h"
8614            : "=w"(result)
8615            : "0"(a), "w"(b), "w"(c)
8616            : /* No clobbers */);
8617   return result;
8618 }
8619
8620 __extension__ extern __inline uint64x2_t
8621 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8622 vmlsl_u32 (uint64x2_t a, uint32x2_t b, uint32x2_t c)
8623 {
8624   uint64x2_t result;
8625   __asm__ ("umlsl %0.2d, %2.2s, %3.2s"
8626            : "=w"(result)
8627            : "0"(a), "w"(b), "w"(c)
8628            : /* No clobbers */);
8629   return result;
8630 }
8631
8632 __extension__ extern __inline float32x4_t
8633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8634 vmlsq_n_f32 (float32x4_t a, float32x4_t b, float32_t c)
8635 {
8636   float32x4_t result;
8637   float32x4_t t1;
8638   __asm__ ("fmul %1.4s, %3.4s, %4.s[0]; fsub %0.4s, %0.4s, %1.4s"
8639            : "=w"(result), "=w"(t1)
8640            : "0"(a), "w"(b), "w"(c)
8641            : /* No clobbers */);
8642   return result;
8643 }
8644
8645 __extension__ extern __inline int16x8_t
8646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8647 vmlsq_n_s16 (int16x8_t a, int16x8_t b, int16_t c)
8648 {
8649   int16x8_t result;
8650   __asm__ ("mls %0.8h, %2.8h, %3.h[0]"
8651            : "=w"(result)
8652            : "0"(a), "w"(b), "x"(c)
8653            : /* No clobbers */);
8654   return result;
8655 }
8656
8657 __extension__ extern __inline int32x4_t
8658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8659 vmlsq_n_s32 (int32x4_t a, int32x4_t b, int32_t c)
8660 {
8661   int32x4_t result;
8662   __asm__ ("mls %0.4s, %2.4s, %3.s[0]"
8663            : "=w"(result)
8664            : "0"(a), "w"(b), "w"(c)
8665            : /* No clobbers */);
8666   return result;
8667 }
8668
8669 __extension__ extern __inline uint16x8_t
8670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8671 vmlsq_n_u16 (uint16x8_t a, uint16x8_t b, uint16_t c)
8672 {
8673   uint16x8_t result;
8674   __asm__ ("mls %0.8h, %2.8h, %3.h[0]"
8675            : "=w"(result)
8676            : "0"(a), "w"(b), "x"(c)
8677            : /* No clobbers */);
8678   return result;
8679 }
8680
8681 __extension__ extern __inline uint32x4_t
8682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8683 vmlsq_n_u32 (uint32x4_t a, uint32x4_t b, uint32_t c)
8684 {
8685   uint32x4_t result;
8686   __asm__ ("mls %0.4s, %2.4s, %3.s[0]"
8687            : "=w"(result)
8688            : "0"(a), "w"(b), "w"(c)
8689            : /* No clobbers */);
8690   return result;
8691 }
8692
8693 __extension__ extern __inline int8x16_t
8694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8695 vmlsq_s8 (int8x16_t a, int8x16_t b, int8x16_t c)
8696 {
8697   int8x16_t result;
8698   __asm__ ("mls %0.16b,%2.16b,%3.16b"
8699            : "=w"(result)
8700            : "0"(a), "w"(b), "w"(c)
8701            : /* No clobbers */);
8702   return result;
8703 }
8704
8705 __extension__ extern __inline int16x8_t
8706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8707 vmlsq_s16 (int16x8_t a, int16x8_t b, int16x8_t c)
8708 {
8709   int16x8_t result;
8710   __asm__ ("mls %0.8h,%2.8h,%3.8h"
8711            : "=w"(result)
8712            : "0"(a), "w"(b), "w"(c)
8713            : /* No clobbers */);
8714   return result;
8715 }
8716
8717 __extension__ extern __inline int32x4_t
8718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8719 vmlsq_s32 (int32x4_t a, int32x4_t b, int32x4_t c)
8720 {
8721   int32x4_t result;
8722   __asm__ ("mls %0.4s,%2.4s,%3.4s"
8723            : "=w"(result)
8724            : "0"(a), "w"(b), "w"(c)
8725            : /* No clobbers */);
8726   return result;
8727 }
8728
8729 __extension__ extern __inline uint8x16_t
8730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8731 vmlsq_u8 (uint8x16_t a, uint8x16_t b, uint8x16_t c)
8732 {
8733   uint8x16_t result;
8734   __asm__ ("mls %0.16b,%2.16b,%3.16b"
8735            : "=w"(result)
8736            : "0"(a), "w"(b), "w"(c)
8737            : /* No clobbers */);
8738   return result;
8739 }
8740
8741 __extension__ extern __inline uint16x8_t
8742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8743 vmlsq_u16 (uint16x8_t a, uint16x8_t b, uint16x8_t c)
8744 {
8745   uint16x8_t result;
8746   __asm__ ("mls %0.8h,%2.8h,%3.8h"
8747            : "=w"(result)
8748            : "0"(a), "w"(b), "w"(c)
8749            : /* No clobbers */);
8750   return result;
8751 }
8752
8753 __extension__ extern __inline uint32x4_t
8754 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8755 vmlsq_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t c)
8756 {
8757   uint32x4_t result;
8758   __asm__ ("mls %0.4s,%2.4s,%3.4s"
8759            : "=w"(result)
8760            : "0"(a), "w"(b), "w"(c)
8761            : /* No clobbers */);
8762   return result;
8763 }
8764
8765 __extension__ extern __inline int16x8_t
8766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8767 vmovl_high_s8 (int8x16_t a)
8768 {
8769   int16x8_t result;
8770   __asm__ ("sshll2 %0.8h,%1.16b,#0"
8771            : "=w"(result)
8772            : "w"(a)
8773            : /* No clobbers */);
8774   return result;
8775 }
8776
8777 __extension__ extern __inline int32x4_t
8778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8779 vmovl_high_s16 (int16x8_t a)
8780 {
8781   int32x4_t result;
8782   __asm__ ("sshll2 %0.4s,%1.8h,#0"
8783            : "=w"(result)
8784            : "w"(a)
8785            : /* No clobbers */);
8786   return result;
8787 }
8788
8789 __extension__ extern __inline int64x2_t
8790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8791 vmovl_high_s32 (int32x4_t a)
8792 {
8793   int64x2_t result;
8794   __asm__ ("sshll2 %0.2d,%1.4s,#0"
8795            : "=w"(result)
8796            : "w"(a)
8797            : /* No clobbers */);
8798   return result;
8799 }
8800
8801 __extension__ extern __inline uint16x8_t
8802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8803 vmovl_high_u8 (uint8x16_t a)
8804 {
8805   uint16x8_t result;
8806   __asm__ ("ushll2 %0.8h,%1.16b,#0"
8807            : "=w"(result)
8808            : "w"(a)
8809            : /* No clobbers */);
8810   return result;
8811 }
8812
8813 __extension__ extern __inline uint32x4_t
8814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8815 vmovl_high_u16 (uint16x8_t a)
8816 {
8817   uint32x4_t result;
8818   __asm__ ("ushll2 %0.4s,%1.8h,#0"
8819            : "=w"(result)
8820            : "w"(a)
8821            : /* No clobbers */);
8822   return result;
8823 }
8824
8825 __extension__ extern __inline uint64x2_t
8826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8827 vmovl_high_u32 (uint32x4_t a)
8828 {
8829   uint64x2_t result;
8830   __asm__ ("ushll2 %0.2d,%1.4s,#0"
8831            : "=w"(result)
8832            : "w"(a)
8833            : /* No clobbers */);
8834   return result;
8835 }
8836
8837 __extension__ extern __inline int16x8_t
8838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8839 vmovl_s8 (int8x8_t a)
8840 {
8841   int16x8_t result;
8842   __asm__ ("sshll %0.8h,%1.8b,#0"
8843            : "=w"(result)
8844            : "w"(a)
8845            : /* No clobbers */);
8846   return result;
8847 }
8848
8849 __extension__ extern __inline int32x4_t
8850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8851 vmovl_s16 (int16x4_t a)
8852 {
8853   int32x4_t result;
8854   __asm__ ("sshll %0.4s,%1.4h,#0"
8855            : "=w"(result)
8856            : "w"(a)
8857            : /* No clobbers */);
8858   return result;
8859 }
8860
8861 __extension__ extern __inline int64x2_t
8862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8863 vmovl_s32 (int32x2_t a)
8864 {
8865   int64x2_t result;
8866   __asm__ ("sshll %0.2d,%1.2s,#0"
8867            : "=w"(result)
8868            : "w"(a)
8869            : /* No clobbers */);
8870   return result;
8871 }
8872
8873 __extension__ extern __inline uint16x8_t
8874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8875 vmovl_u8 (uint8x8_t a)
8876 {
8877   uint16x8_t result;
8878   __asm__ ("ushll %0.8h,%1.8b,#0"
8879            : "=w"(result)
8880            : "w"(a)
8881            : /* No clobbers */);
8882   return result;
8883 }
8884
8885 __extension__ extern __inline uint32x4_t
8886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8887 vmovl_u16 (uint16x4_t a)
8888 {
8889   uint32x4_t result;
8890   __asm__ ("ushll %0.4s,%1.4h,#0"
8891            : "=w"(result)
8892            : "w"(a)
8893            : /* No clobbers */);
8894   return result;
8895 }
8896
8897 __extension__ extern __inline uint64x2_t
8898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8899 vmovl_u32 (uint32x2_t a)
8900 {
8901   uint64x2_t result;
8902   __asm__ ("ushll %0.2d,%1.2s,#0"
8903            : "=w"(result)
8904            : "w"(a)
8905            : /* No clobbers */);
8906   return result;
8907 }
8908
8909 __extension__ extern __inline int8x16_t
8910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8911 vmovn_high_s16 (int8x8_t a, int16x8_t b)
8912 {
8913   int8x16_t result = vcombine_s8 (a, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
8914   __asm__ ("xtn2 %0.16b,%1.8h"
8915            : "+w"(result)
8916            : "w"(b)
8917            : /* No clobbers */);
8918   return result;
8919 }
8920
8921 __extension__ extern __inline int16x8_t
8922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8923 vmovn_high_s32 (int16x4_t a, int32x4_t b)
8924 {
8925   int16x8_t result = vcombine_s16 (a, vcreate_s16 (__AARCH64_UINT64_C (0x0)));
8926   __asm__ ("xtn2 %0.8h,%1.4s"
8927            : "+w"(result)
8928            : "w"(b)
8929            : /* No clobbers */);
8930   return result;
8931 }
8932
8933 __extension__ extern __inline int32x4_t
8934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8935 vmovn_high_s64 (int32x2_t a, int64x2_t b)
8936 {
8937   int32x4_t result = vcombine_s32 (a, vcreate_s32 (__AARCH64_UINT64_C (0x0)));
8938   __asm__ ("xtn2 %0.4s,%1.2d"
8939            : "+w"(result)
8940            : "w"(b)
8941            : /* No clobbers */);
8942   return result;
8943 }
8944
8945 __extension__ extern __inline uint8x16_t
8946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8947 vmovn_high_u16 (uint8x8_t a, uint16x8_t b)
8948 {
8949   uint8x16_t result = vcombine_u8 (a, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
8950   __asm__ ("xtn2 %0.16b,%1.8h"
8951            : "+w"(result)
8952            : "w"(b)
8953            : /* No clobbers */);
8954   return result;
8955 }
8956
8957 __extension__ extern __inline uint16x8_t
8958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8959 vmovn_high_u32 (uint16x4_t a, uint32x4_t b)
8960 {
8961   uint16x8_t result = vcombine_u16 (a, vcreate_u16 (__AARCH64_UINT64_C (0x0)));
8962   __asm__ ("xtn2 %0.8h,%1.4s"
8963            : "+w"(result)
8964            : "w"(b)
8965            : /* No clobbers */);
8966   return result;
8967 }
8968
8969 __extension__ extern __inline uint32x4_t
8970 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8971 vmovn_high_u64 (uint32x2_t a, uint64x2_t b)
8972 {
8973   uint32x4_t result = vcombine_u32 (a, vcreate_u32 (__AARCH64_UINT64_C (0x0)));
8974   __asm__ ("xtn2 %0.4s,%1.2d"
8975            : "+w"(result)
8976            : "w"(b)
8977            : /* No clobbers */);
8978   return result;
8979 }
8980
8981 __extension__ extern __inline int8x8_t
8982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8983 vmovn_s16 (int16x8_t a)
8984 {
8985   int8x8_t result;
8986   __asm__ ("xtn %0.8b,%1.8h"
8987            : "=w"(result)
8988            : "w"(a)
8989            : /* No clobbers */);
8990   return result;
8991 }
8992
8993 __extension__ extern __inline int16x4_t
8994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8995 vmovn_s32 (int32x4_t a)
8996 {
8997   int16x4_t result;
8998   __asm__ ("xtn %0.4h,%1.4s"
8999            : "=w"(result)
9000            : "w"(a)
9001            : /* No clobbers */);
9002   return result;
9003 }
9004
9005 __extension__ extern __inline int32x2_t
9006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9007 vmovn_s64 (int64x2_t a)
9008 {
9009   int32x2_t result;
9010   __asm__ ("xtn %0.2s,%1.2d"
9011            : "=w"(result)
9012            : "w"(a)
9013            : /* No clobbers */);
9014   return result;
9015 }
9016
9017 __extension__ extern __inline uint8x8_t
9018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9019 vmovn_u16 (uint16x8_t a)
9020 {
9021   uint8x8_t result;
9022   __asm__ ("xtn %0.8b,%1.8h"
9023            : "=w"(result)
9024            : "w"(a)
9025            : /* No clobbers */);
9026   return result;
9027 }
9028
9029 __extension__ extern __inline uint16x4_t
9030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9031 vmovn_u32 (uint32x4_t a)
9032 {
9033   uint16x4_t result;
9034   __asm__ ("xtn %0.4h,%1.4s"
9035            : "=w"(result)
9036            : "w"(a)
9037            : /* No clobbers */);
9038   return result;
9039 }
9040
9041 __extension__ extern __inline uint32x2_t
9042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9043 vmovn_u64 (uint64x2_t a)
9044 {
9045   uint32x2_t result;
9046   __asm__ ("xtn %0.2s,%1.2d"
9047            : "=w"(result)
9048            : "w"(a)
9049            : /* No clobbers */);
9050   return result;
9051 }
9052
9053 #define vmull_high_lane_s16(a, b, c)                                    \
9054   __extension__                                                         \
9055     ({                                                                  \
9056        int16x4_t b_ = (b);                                              \
9057        int16x8_t a_ = (a);                                              \
9058        int32x4_t result;                                                \
9059        __asm__ ("smull2 %0.4s, %1.8h, %2.h[%3]"                         \
9060                 : "=w"(result)                                          \
9061                 : "w"(a_), "x"(b_), "i"(c)                              \
9062                 : /* No clobbers */);                                   \
9063        result;                                                          \
9064      })
9065
9066 #define vmull_high_lane_s32(a, b, c)                                    \
9067   __extension__                                                         \
9068     ({                                                                  \
9069        int32x2_t b_ = (b);                                              \
9070        int32x4_t a_ = (a);                                              \
9071        int64x2_t result;                                                \
9072        __asm__ ("smull2 %0.2d, %1.4s, %2.s[%3]"                         \
9073                 : "=w"(result)                                          \
9074                 : "w"(a_), "w"(b_), "i"(c)                              \
9075                 : /* No clobbers */);                                   \
9076        result;                                                          \
9077      })
9078
9079 #define vmull_high_lane_u16(a, b, c)                                    \
9080   __extension__                                                         \
9081     ({                                                                  \
9082        uint16x4_t b_ = (b);                                             \
9083        uint16x8_t a_ = (a);                                             \
9084        uint32x4_t result;                                               \
9085        __asm__ ("umull2 %0.4s, %1.8h, %2.h[%3]"                         \
9086                 : "=w"(result)                                          \
9087                 : "w"(a_), "x"(b_), "i"(c)                              \
9088                 : /* No clobbers */);                                   \
9089        result;                                                          \
9090      })
9091
9092 #define vmull_high_lane_u32(a, b, c)                                    \
9093   __extension__                                                         \
9094     ({                                                                  \
9095        uint32x2_t b_ = (b);                                             \
9096        uint32x4_t a_ = (a);                                             \
9097        uint64x2_t result;                                               \
9098        __asm__ ("umull2 %0.2d, %1.4s, %2.s[%3]"                         \
9099                 : "=w"(result)                                          \
9100                 : "w"(a_), "w"(b_), "i"(c)                              \
9101                 : /* No clobbers */);                                   \
9102        result;                                                          \
9103      })
9104
9105 #define vmull_high_laneq_s16(a, b, c)                                   \
9106   __extension__                                                         \
9107     ({                                                                  \
9108        int16x8_t b_ = (b);                                              \
9109        int16x8_t a_ = (a);                                              \
9110        int32x4_t result;                                                \
9111        __asm__ ("smull2 %0.4s, %1.8h, %2.h[%3]"                         \
9112                 : "=w"(result)                                          \
9113                 : "w"(a_), "x"(b_), "i"(c)                              \
9114                 : /* No clobbers */);                                   \
9115        result;                                                          \
9116      })
9117
9118 #define vmull_high_laneq_s32(a, b, c)                                   \
9119   __extension__                                                         \
9120     ({                                                                  \
9121        int32x4_t b_ = (b);                                              \
9122        int32x4_t a_ = (a);                                              \
9123        int64x2_t result;                                                \
9124        __asm__ ("smull2 %0.2d, %1.4s, %2.s[%3]"                         \
9125                 : "=w"(result)                                          \
9126                 : "w"(a_), "w"(b_), "i"(c)                              \
9127                 : /* No clobbers */);                                   \
9128        result;                                                          \
9129      })
9130
9131 #define vmull_high_laneq_u16(a, b, c)                                   \
9132   __extension__                                                         \
9133     ({                                                                  \
9134        uint16x8_t b_ = (b);                                             \
9135        uint16x8_t a_ = (a);                                             \
9136        uint32x4_t result;                                               \
9137        __asm__ ("umull2 %0.4s, %1.8h, %2.h[%3]"                         \
9138                 : "=w"(result)                                          \
9139                 : "w"(a_), "x"(b_), "i"(c)                              \
9140                 : /* No clobbers */);                                   \
9141        result;                                                          \
9142      })
9143
9144 #define vmull_high_laneq_u32(a, b, c)                                   \
9145   __extension__                                                         \
9146     ({                                                                  \
9147        uint32x4_t b_ = (b);                                             \
9148        uint32x4_t a_ = (a);                                             \
9149        uint64x2_t result;                                               \
9150        __asm__ ("umull2 %0.2d, %1.4s, %2.s[%3]"                         \
9151                 : "=w"(result)                                          \
9152                 : "w"(a_), "w"(b_), "i"(c)                              \
9153                 : /* No clobbers */);                                   \
9154        result;                                                          \
9155      })
9156
9157 __extension__ extern __inline int32x4_t
9158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9159 vmull_high_n_s16 (int16x8_t a, int16_t b)
9160 {
9161   int32x4_t result;
9162   __asm__ ("smull2 %0.4s,%1.8h,%2.h[0]"
9163            : "=w"(result)
9164            : "w"(a), "x"(b)
9165            : /* No clobbers */);
9166   return result;
9167 }
9168
9169 __extension__ extern __inline int64x2_t
9170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9171 vmull_high_n_s32 (int32x4_t a, int32_t b)
9172 {
9173   int64x2_t result;
9174   __asm__ ("smull2 %0.2d,%1.4s,%2.s[0]"
9175            : "=w"(result)
9176            : "w"(a), "w"(b)
9177            : /* No clobbers */);
9178   return result;
9179 }
9180
9181 __extension__ extern __inline uint32x4_t
9182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9183 vmull_high_n_u16 (uint16x8_t a, uint16_t b)
9184 {
9185   uint32x4_t result;
9186   __asm__ ("umull2 %0.4s,%1.8h,%2.h[0]"
9187            : "=w"(result)
9188            : "w"(a), "x"(b)
9189            : /* No clobbers */);
9190   return result;
9191 }
9192
9193 __extension__ extern __inline uint64x2_t
9194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9195 vmull_high_n_u32 (uint32x4_t a, uint32_t b)
9196 {
9197   uint64x2_t result;
9198   __asm__ ("umull2 %0.2d,%1.4s,%2.s[0]"
9199            : "=w"(result)
9200            : "w"(a), "w"(b)
9201            : /* No clobbers */);
9202   return result;
9203 }
9204
9205 __extension__ extern __inline poly16x8_t
9206 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9207 vmull_high_p8 (poly8x16_t a, poly8x16_t b)
9208 {
9209   poly16x8_t result;
9210   __asm__ ("pmull2 %0.8h,%1.16b,%2.16b"
9211            : "=w"(result)
9212            : "w"(a), "w"(b)
9213            : /* No clobbers */);
9214   return result;
9215 }
9216
9217 __extension__ extern __inline int16x8_t
9218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9219 vmull_high_s8 (int8x16_t a, int8x16_t b)
9220 {
9221   int16x8_t result;
9222   __asm__ ("smull2 %0.8h,%1.16b,%2.16b"
9223            : "=w"(result)
9224            : "w"(a), "w"(b)
9225            : /* No clobbers */);
9226   return result;
9227 }
9228
9229 __extension__ extern __inline int32x4_t
9230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9231 vmull_high_s16 (int16x8_t a, int16x8_t b)
9232 {
9233   int32x4_t result;
9234   __asm__ ("smull2 %0.4s,%1.8h,%2.8h"
9235            : "=w"(result)
9236            : "w"(a), "w"(b)
9237            : /* No clobbers */);
9238   return result;
9239 }
9240
9241 __extension__ extern __inline int64x2_t
9242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9243 vmull_high_s32 (int32x4_t a, int32x4_t b)
9244 {
9245   int64x2_t result;
9246   __asm__ ("smull2 %0.2d,%1.4s,%2.4s"
9247            : "=w"(result)
9248            : "w"(a), "w"(b)
9249            : /* No clobbers */);
9250   return result;
9251 }
9252
9253 __extension__ extern __inline uint16x8_t
9254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9255 vmull_high_u8 (uint8x16_t a, uint8x16_t b)
9256 {
9257   uint16x8_t result;
9258   __asm__ ("umull2 %0.8h,%1.16b,%2.16b"
9259            : "=w"(result)
9260            : "w"(a), "w"(b)
9261            : /* No clobbers */);
9262   return result;
9263 }
9264
9265 __extension__ extern __inline uint32x4_t
9266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9267 vmull_high_u16 (uint16x8_t a, uint16x8_t b)
9268 {
9269   uint32x4_t result;
9270   __asm__ ("umull2 %0.4s,%1.8h,%2.8h"
9271            : "=w"(result)
9272            : "w"(a), "w"(b)
9273            : /* No clobbers */);
9274   return result;
9275 }
9276
9277 __extension__ extern __inline uint64x2_t
9278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9279 vmull_high_u32 (uint32x4_t a, uint32x4_t b)
9280 {
9281   uint64x2_t result;
9282   __asm__ ("umull2 %0.2d,%1.4s,%2.4s"
9283            : "=w"(result)
9284            : "w"(a), "w"(b)
9285            : /* No clobbers */);
9286   return result;
9287 }
9288
9289 #define vmull_lane_s16(a, b, c)                                         \
9290   __extension__                                                         \
9291     ({                                                                  \
9292        int16x4_t b_ = (b);                                              \
9293        int16x4_t a_ = (a);                                              \
9294        int32x4_t result;                                                \
9295        __asm__ ("smull %0.4s,%1.4h,%2.h[%3]"                            \
9296                 : "=w"(result)                                          \
9297                 : "w"(a_), "x"(b_), "i"(c)                              \
9298                 : /* No clobbers */);                                   \
9299        result;                                                          \
9300      })
9301
9302 #define vmull_lane_s32(a, b, c)                                         \
9303   __extension__                                                         \
9304     ({                                                                  \
9305        int32x2_t b_ = (b);                                              \
9306        int32x2_t a_ = (a);                                              \
9307        int64x2_t result;                                                \
9308        __asm__ ("smull %0.2d,%1.2s,%2.s[%3]"                            \
9309                 : "=w"(result)                                          \
9310                 : "w"(a_), "w"(b_), "i"(c)                              \
9311                 : /* No clobbers */);                                   \
9312        result;                                                          \
9313      })
9314
9315 #define vmull_lane_u16(a, b, c)                                         \
9316   __extension__                                                         \
9317     ({                                                                  \
9318        uint16x4_t b_ = (b);                                             \
9319        uint16x4_t a_ = (a);                                             \
9320        uint32x4_t result;                                               \
9321        __asm__ ("umull %0.4s,%1.4h,%2.h[%3]"                            \
9322                 : "=w"(result)                                          \
9323                 : "w"(a_), "x"(b_), "i"(c)                              \
9324                 : /* No clobbers */);                                   \
9325        result;                                                          \
9326      })
9327
9328 #define vmull_lane_u32(a, b, c)                                         \
9329   __extension__                                                         \
9330     ({                                                                  \
9331        uint32x2_t b_ = (b);                                             \
9332        uint32x2_t a_ = (a);                                             \
9333        uint64x2_t result;                                               \
9334        __asm__ ("umull %0.2d, %1.2s, %2.s[%3]"                          \
9335                 : "=w"(result)                                          \
9336                 : "w"(a_), "w"(b_), "i"(c)                              \
9337                 : /* No clobbers */);                                   \
9338        result;                                                          \
9339      })
9340
9341 #define vmull_laneq_s16(a, b, c)                                        \
9342   __extension__                                                         \
9343     ({                                                                  \
9344        int16x8_t b_ = (b);                                              \
9345        int16x4_t a_ = (a);                                              \
9346        int32x4_t result;                                                \
9347        __asm__ ("smull %0.4s, %1.4h, %2.h[%3]"                          \
9348                 : "=w"(result)                                          \
9349                 : "w"(a_), "x"(b_), "i"(c)                              \
9350                 : /* No clobbers */);                                   \
9351        result;                                                          \
9352      })
9353
9354 #define vmull_laneq_s32(a, b, c)                                        \
9355   __extension__                                                         \
9356     ({                                                                  \
9357        int32x4_t b_ = (b);                                              \
9358        int32x2_t a_ = (a);                                              \
9359        int64x2_t result;                                                \
9360        __asm__ ("smull %0.2d, %1.2s, %2.s[%3]"                          \
9361                 : "=w"(result)                                          \
9362                 : "w"(a_), "w"(b_), "i"(c)                              \
9363                 : /* No clobbers */);                                   \
9364        result;                                                          \
9365      })
9366
9367 #define vmull_laneq_u16(a, b, c)                                        \
9368   __extension__                                                         \
9369     ({                                                                  \
9370        uint16x8_t b_ = (b);                                             \
9371        uint16x4_t a_ = (a);                                             \
9372        uint32x4_t result;                                               \
9373        __asm__ ("umull %0.4s, %1.4h, %2.h[%3]"                          \
9374                 : "=w"(result)                                          \
9375                 : "w"(a_), "x"(b_), "i"(c)                              \
9376                 : /* No clobbers */);                                   \
9377        result;                                                          \
9378      })
9379
9380 #define vmull_laneq_u32(a, b, c)                                        \
9381   __extension__                                                         \
9382     ({                                                                  \
9383        uint32x4_t b_ = (b);                                             \
9384        uint32x2_t a_ = (a);                                             \
9385        uint64x2_t result;                                               \
9386        __asm__ ("umull %0.2d, %1.2s, %2.s[%3]"                          \
9387                 : "=w"(result)                                          \
9388                 : "w"(a_), "w"(b_), "i"(c)                              \
9389                 : /* No clobbers */);                                   \
9390        result;                                                          \
9391      })
9392
9393 __extension__ extern __inline int32x4_t
9394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9395 vmull_n_s16 (int16x4_t a, int16_t b)
9396 {
9397   int32x4_t result;
9398   __asm__ ("smull %0.4s,%1.4h,%2.h[0]"
9399            : "=w"(result)
9400            : "w"(a), "x"(b)
9401            : /* No clobbers */);
9402   return result;
9403 }
9404
9405 __extension__ extern __inline int64x2_t
9406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9407 vmull_n_s32 (int32x2_t a, int32_t b)
9408 {
9409   int64x2_t result;
9410   __asm__ ("smull %0.2d,%1.2s,%2.s[0]"
9411            : "=w"(result)
9412            : "w"(a), "w"(b)
9413            : /* No clobbers */);
9414   return result;
9415 }
9416
9417 __extension__ extern __inline uint32x4_t
9418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9419 vmull_n_u16 (uint16x4_t a, uint16_t b)
9420 {
9421   uint32x4_t result;
9422   __asm__ ("umull %0.4s,%1.4h,%2.h[0]"
9423            : "=w"(result)
9424            : "w"(a), "x"(b)
9425            : /* No clobbers */);
9426   return result;
9427 }
9428
9429 __extension__ extern __inline uint64x2_t
9430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9431 vmull_n_u32 (uint32x2_t a, uint32_t b)
9432 {
9433   uint64x2_t result;
9434   __asm__ ("umull %0.2d,%1.2s,%2.s[0]"
9435            : "=w"(result)
9436            : "w"(a), "w"(b)
9437            : /* No clobbers */);
9438   return result;
9439 }
9440
9441 __extension__ extern __inline poly16x8_t
9442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9443 vmull_p8 (poly8x8_t a, poly8x8_t b)
9444 {
9445   poly16x8_t result;
9446   __asm__ ("pmull %0.8h, %1.8b, %2.8b"
9447            : "=w"(result)
9448            : "w"(a), "w"(b)
9449            : /* No clobbers */);
9450   return result;
9451 }
9452
9453 __extension__ extern __inline int16x8_t
9454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9455 vmull_s8 (int8x8_t a, int8x8_t b)
9456 {
9457   int16x8_t result;
9458   __asm__ ("smull %0.8h, %1.8b, %2.8b"
9459            : "=w"(result)
9460            : "w"(a), "w"(b)
9461            : /* No clobbers */);
9462   return result;
9463 }
9464
9465 __extension__ extern __inline int32x4_t
9466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9467 vmull_s16 (int16x4_t a, int16x4_t b)
9468 {
9469   int32x4_t result;
9470   __asm__ ("smull %0.4s, %1.4h, %2.4h"
9471            : "=w"(result)
9472            : "w"(a), "w"(b)
9473            : /* No clobbers */);
9474   return result;
9475 }
9476
9477 __extension__ extern __inline int64x2_t
9478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9479 vmull_s32 (int32x2_t a, int32x2_t b)
9480 {
9481   int64x2_t result;
9482   __asm__ ("smull %0.2d, %1.2s, %2.2s"
9483            : "=w"(result)
9484            : "w"(a), "w"(b)
9485            : /* No clobbers */);
9486   return result;
9487 }
9488
9489 __extension__ extern __inline uint16x8_t
9490 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9491 vmull_u8 (uint8x8_t a, uint8x8_t b)
9492 {
9493   uint16x8_t result;
9494   __asm__ ("umull %0.8h, %1.8b, %2.8b"
9495            : "=w"(result)
9496            : "w"(a), "w"(b)
9497            : /* No clobbers */);
9498   return result;
9499 }
9500
9501 __extension__ extern __inline uint32x4_t
9502 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9503 vmull_u16 (uint16x4_t a, uint16x4_t b)
9504 {
9505   uint32x4_t result;
9506   __asm__ ("umull %0.4s, %1.4h, %2.4h"
9507            : "=w"(result)
9508            : "w"(a), "w"(b)
9509            : /* No clobbers */);
9510   return result;
9511 }
9512
9513 __extension__ extern __inline uint64x2_t
9514 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9515 vmull_u32 (uint32x2_t a, uint32x2_t b)
9516 {
9517   uint64x2_t result;
9518   __asm__ ("umull %0.2d, %1.2s, %2.2s"
9519            : "=w"(result)
9520            : "w"(a), "w"(b)
9521            : /* No clobbers */);
9522   return result;
9523 }
9524
9525 __extension__ extern __inline int16x4_t
9526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9527 vpadal_s8 (int16x4_t a, int8x8_t b)
9528 {
9529   int16x4_t result;
9530   __asm__ ("sadalp %0.4h,%2.8b"
9531            : "=w"(result)
9532            : "0"(a), "w"(b)
9533            : /* No clobbers */);
9534   return result;
9535 }
9536
9537 __extension__ extern __inline int32x2_t
9538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9539 vpadal_s16 (int32x2_t a, int16x4_t b)
9540 {
9541   int32x2_t result;
9542   __asm__ ("sadalp %0.2s,%2.4h"
9543            : "=w"(result)
9544            : "0"(a), "w"(b)
9545            : /* No clobbers */);
9546   return result;
9547 }
9548
9549 __extension__ extern __inline int64x1_t
9550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9551 vpadal_s32 (int64x1_t a, int32x2_t b)
9552 {
9553   int64x1_t result;
9554   __asm__ ("sadalp %0.1d,%2.2s"
9555            : "=w"(result)
9556            : "0"(a), "w"(b)
9557            : /* No clobbers */);
9558   return result;
9559 }
9560
9561 __extension__ extern __inline uint16x4_t
9562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9563 vpadal_u8 (uint16x4_t a, uint8x8_t b)
9564 {
9565   uint16x4_t result;
9566   __asm__ ("uadalp %0.4h,%2.8b"
9567            : "=w"(result)
9568            : "0"(a), "w"(b)
9569            : /* No clobbers */);
9570   return result;
9571 }
9572
9573 __extension__ extern __inline uint32x2_t
9574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9575 vpadal_u16 (uint32x2_t a, uint16x4_t b)
9576 {
9577   uint32x2_t result;
9578   __asm__ ("uadalp %0.2s,%2.4h"
9579            : "=w"(result)
9580            : "0"(a), "w"(b)
9581            : /* No clobbers */);
9582   return result;
9583 }
9584
9585 __extension__ extern __inline uint64x1_t
9586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9587 vpadal_u32 (uint64x1_t a, uint32x2_t b)
9588 {
9589   uint64x1_t result;
9590   __asm__ ("uadalp %0.1d,%2.2s"
9591            : "=w"(result)
9592            : "0"(a), "w"(b)
9593            : /* No clobbers */);
9594   return result;
9595 }
9596
9597 __extension__ extern __inline int16x8_t
9598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9599 vpadalq_s8 (int16x8_t a, int8x16_t b)
9600 {
9601   int16x8_t result;
9602   __asm__ ("sadalp %0.8h,%2.16b"
9603            : "=w"(result)
9604            : "0"(a), "w"(b)
9605            : /* No clobbers */);
9606   return result;
9607 }
9608
9609 __extension__ extern __inline int32x4_t
9610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9611 vpadalq_s16 (int32x4_t a, int16x8_t b)
9612 {
9613   int32x4_t result;
9614   __asm__ ("sadalp %0.4s,%2.8h"
9615            : "=w"(result)
9616            : "0"(a), "w"(b)
9617            : /* No clobbers */);
9618   return result;
9619 }
9620
9621 __extension__ extern __inline int64x2_t
9622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9623 vpadalq_s32 (int64x2_t a, int32x4_t b)
9624 {
9625   int64x2_t result;
9626   __asm__ ("sadalp %0.2d,%2.4s"
9627            : "=w"(result)
9628            : "0"(a), "w"(b)
9629            : /* No clobbers */);
9630   return result;
9631 }
9632
9633 __extension__ extern __inline uint16x8_t
9634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9635 vpadalq_u8 (uint16x8_t a, uint8x16_t b)
9636 {
9637   uint16x8_t result;
9638   __asm__ ("uadalp %0.8h,%2.16b"
9639            : "=w"(result)
9640            : "0"(a), "w"(b)
9641            : /* No clobbers */);
9642   return result;
9643 }
9644
9645 __extension__ extern __inline uint32x4_t
9646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9647 vpadalq_u16 (uint32x4_t a, uint16x8_t b)
9648 {
9649   uint32x4_t result;
9650   __asm__ ("uadalp %0.4s,%2.8h"
9651            : "=w"(result)
9652            : "0"(a), "w"(b)
9653            : /* No clobbers */);
9654   return result;
9655 }
9656
9657 __extension__ extern __inline uint64x2_t
9658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9659 vpadalq_u32 (uint64x2_t a, uint32x4_t b)
9660 {
9661   uint64x2_t result;
9662   __asm__ ("uadalp %0.2d,%2.4s"
9663            : "=w"(result)
9664            : "0"(a), "w"(b)
9665            : /* No clobbers */);
9666   return result;
9667 }
9668
9669 __extension__ extern __inline int16x4_t
9670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9671 vpaddl_s8 (int8x8_t a)
9672 {
9673   int16x4_t result;
9674   __asm__ ("saddlp %0.4h,%1.8b"
9675            : "=w"(result)
9676            : "w"(a)
9677            : /* No clobbers */);
9678   return result;
9679 }
9680
9681 __extension__ extern __inline int32x2_t
9682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9683 vpaddl_s16 (int16x4_t a)
9684 {
9685   int32x2_t result;
9686   __asm__ ("saddlp %0.2s,%1.4h"
9687            : "=w"(result)
9688            : "w"(a)
9689            : /* No clobbers */);
9690   return result;
9691 }
9692
9693 __extension__ extern __inline int64x1_t
9694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9695 vpaddl_s32 (int32x2_t a)
9696 {
9697   int64x1_t result;
9698   __asm__ ("saddlp %0.1d,%1.2s"
9699            : "=w"(result)
9700            : "w"(a)
9701            : /* No clobbers */);
9702   return result;
9703 }
9704
9705 __extension__ extern __inline uint16x4_t
9706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9707 vpaddl_u8 (uint8x8_t a)
9708 {
9709   uint16x4_t result;
9710   __asm__ ("uaddlp %0.4h,%1.8b"
9711            : "=w"(result)
9712            : "w"(a)
9713            : /* No clobbers */);
9714   return result;
9715 }
9716
9717 __extension__ extern __inline uint32x2_t
9718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9719 vpaddl_u16 (uint16x4_t a)
9720 {
9721   uint32x2_t result;
9722   __asm__ ("uaddlp %0.2s,%1.4h"
9723            : "=w"(result)
9724            : "w"(a)
9725            : /* No clobbers */);
9726   return result;
9727 }
9728
9729 __extension__ extern __inline uint64x1_t
9730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9731 vpaddl_u32 (uint32x2_t a)
9732 {
9733   uint64x1_t result;
9734   __asm__ ("uaddlp %0.1d,%1.2s"
9735            : "=w"(result)
9736            : "w"(a)
9737            : /* No clobbers */);
9738   return result;
9739 }
9740
9741 __extension__ extern __inline int16x8_t
9742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9743 vpaddlq_s8 (int8x16_t a)
9744 {
9745   int16x8_t result;
9746   __asm__ ("saddlp %0.8h,%1.16b"
9747            : "=w"(result)
9748            : "w"(a)
9749            : /* No clobbers */);
9750   return result;
9751 }
9752
9753 __extension__ extern __inline int32x4_t
9754 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9755 vpaddlq_s16 (int16x8_t a)
9756 {
9757   int32x4_t result;
9758   __asm__ ("saddlp %0.4s,%1.8h"
9759            : "=w"(result)
9760            : "w"(a)
9761            : /* No clobbers */);
9762   return result;
9763 }
9764
9765 __extension__ extern __inline int64x2_t
9766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9767 vpaddlq_s32 (int32x4_t a)
9768 {
9769   int64x2_t result;
9770   __asm__ ("saddlp %0.2d,%1.4s"
9771            : "=w"(result)
9772            : "w"(a)
9773            : /* No clobbers */);
9774   return result;
9775 }
9776
9777 __extension__ extern __inline uint16x8_t
9778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9779 vpaddlq_u8 (uint8x16_t a)
9780 {
9781   uint16x8_t result;
9782   __asm__ ("uaddlp %0.8h,%1.16b"
9783            : "=w"(result)
9784            : "w"(a)
9785            : /* No clobbers */);
9786   return result;
9787 }
9788
9789 __extension__ extern __inline uint32x4_t
9790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9791 vpaddlq_u16 (uint16x8_t a)
9792 {
9793   uint32x4_t result;
9794   __asm__ ("uaddlp %0.4s,%1.8h"
9795            : "=w"(result)
9796            : "w"(a)
9797            : /* No clobbers */);
9798   return result;
9799 }
9800
9801 __extension__ extern __inline uint64x2_t
9802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9803 vpaddlq_u32 (uint32x4_t a)
9804 {
9805   uint64x2_t result;
9806   __asm__ ("uaddlp %0.2d,%1.4s"
9807            : "=w"(result)
9808            : "w"(a)
9809            : /* No clobbers */);
9810   return result;
9811 }
9812
9813 __extension__ extern __inline int8x16_t
9814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9815 vpaddq_s8 (int8x16_t a, int8x16_t b)
9816 {
9817   int8x16_t result;
9818   __asm__ ("addp %0.16b,%1.16b,%2.16b"
9819            : "=w"(result)
9820            : "w"(a), "w"(b)
9821            : /* No clobbers */);
9822   return result;
9823 }
9824
9825 __extension__ extern __inline int16x8_t
9826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9827 vpaddq_s16 (int16x8_t a, int16x8_t b)
9828 {
9829   int16x8_t result;
9830   __asm__ ("addp %0.8h,%1.8h,%2.8h"
9831            : "=w"(result)
9832            : "w"(a), "w"(b)
9833            : /* No clobbers */);
9834   return result;
9835 }
9836
9837 __extension__ extern __inline int32x4_t
9838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9839 vpaddq_s32 (int32x4_t a, int32x4_t b)
9840 {
9841   int32x4_t result;
9842   __asm__ ("addp %0.4s,%1.4s,%2.4s"
9843            : "=w"(result)
9844            : "w"(a), "w"(b)
9845            : /* No clobbers */);
9846   return result;
9847 }
9848
9849 __extension__ extern __inline int64x2_t
9850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9851 vpaddq_s64 (int64x2_t a, int64x2_t b)
9852 {
9853   int64x2_t result;
9854   __asm__ ("addp %0.2d,%1.2d,%2.2d"
9855            : "=w"(result)
9856            : "w"(a), "w"(b)
9857            : /* No clobbers */);
9858   return result;
9859 }
9860
9861 __extension__ extern __inline uint8x16_t
9862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9863 vpaddq_u8 (uint8x16_t a, uint8x16_t b)
9864 {
9865   uint8x16_t result;
9866   __asm__ ("addp %0.16b,%1.16b,%2.16b"
9867            : "=w"(result)
9868            : "w"(a), "w"(b)
9869            : /* No clobbers */);
9870   return result;
9871 }
9872
9873 __extension__ extern __inline uint16x8_t
9874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9875 vpaddq_u16 (uint16x8_t a, uint16x8_t b)
9876 {
9877   uint16x8_t result;
9878   __asm__ ("addp %0.8h,%1.8h,%2.8h"
9879            : "=w"(result)
9880            : "w"(a), "w"(b)
9881            : /* No clobbers */);
9882   return result;
9883 }
9884
9885 __extension__ extern __inline uint32x4_t
9886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9887 vpaddq_u32 (uint32x4_t a, uint32x4_t b)
9888 {
9889   uint32x4_t result;
9890   __asm__ ("addp %0.4s,%1.4s,%2.4s"
9891            : "=w"(result)
9892            : "w"(a), "w"(b)
9893            : /* No clobbers */);
9894   return result;
9895 }
9896
9897 __extension__ extern __inline uint64x2_t
9898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9899 vpaddq_u64 (uint64x2_t a, uint64x2_t b)
9900 {
9901   uint64x2_t result;
9902   __asm__ ("addp %0.2d,%1.2d,%2.2d"
9903            : "=w"(result)
9904            : "w"(a), "w"(b)
9905            : /* No clobbers */);
9906   return result;
9907 }
9908
9909 __extension__ extern __inline int16x4_t
9910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9911 vqdmulh_n_s16 (int16x4_t a, int16_t b)
9912 {
9913   int16x4_t result;
9914   __asm__ ("sqdmulh %0.4h,%1.4h,%2.h[0]"
9915            : "=w"(result)
9916            : "w"(a), "x"(b)
9917            : /* No clobbers */);
9918   return result;
9919 }
9920
9921 __extension__ extern __inline int32x2_t
9922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9923 vqdmulh_n_s32 (int32x2_t a, int32_t b)
9924 {
9925   int32x2_t result;
9926   __asm__ ("sqdmulh %0.2s,%1.2s,%2.s[0]"
9927            : "=w"(result)
9928            : "w"(a), "w"(b)
9929            : /* No clobbers */);
9930   return result;
9931 }
9932
9933 __extension__ extern __inline int16x8_t
9934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9935 vqdmulhq_n_s16 (int16x8_t a, int16_t b)
9936 {
9937   int16x8_t result;
9938   __asm__ ("sqdmulh %0.8h,%1.8h,%2.h[0]"
9939            : "=w"(result)
9940            : "w"(a), "x"(b)
9941            : /* No clobbers */);
9942   return result;
9943 }
9944
9945 __extension__ extern __inline int32x4_t
9946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9947 vqdmulhq_n_s32 (int32x4_t a, int32_t b)
9948 {
9949   int32x4_t result;
9950   __asm__ ("sqdmulh %0.4s,%1.4s,%2.s[0]"
9951            : "=w"(result)
9952            : "w"(a), "w"(b)
9953            : /* No clobbers */);
9954   return result;
9955 }
9956
9957 __extension__ extern __inline int8x16_t
9958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9959 vqmovn_high_s16 (int8x8_t a, int16x8_t b)
9960 {
9961   int8x16_t result = vcombine_s8 (a, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
9962   __asm__ ("sqxtn2 %0.16b, %1.8h"
9963            : "+w"(result)
9964            : "w"(b)
9965            : /* No clobbers */);
9966   return result;
9967 }
9968
9969 __extension__ extern __inline int16x8_t
9970 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9971 vqmovn_high_s32 (int16x4_t a, int32x4_t b)
9972 {
9973   int16x8_t result = vcombine_s16 (a, vcreate_s16 (__AARCH64_UINT64_C (0x0)));
9974   __asm__ ("sqxtn2 %0.8h, %1.4s"
9975            : "+w"(result)
9976            : "w"(b)
9977            : /* No clobbers */);
9978   return result;
9979 }
9980
9981 __extension__ extern __inline int32x4_t
9982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9983 vqmovn_high_s64 (int32x2_t a, int64x2_t b)
9984 {
9985   int32x4_t result = vcombine_s32 (a, vcreate_s32 (__AARCH64_UINT64_C (0x0)));
9986   __asm__ ("sqxtn2 %0.4s, %1.2d"
9987            : "+w"(result)
9988            : "w"(b)
9989            : /* No clobbers */);
9990   return result;
9991 }
9992
9993 __extension__ extern __inline uint8x16_t
9994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9995 vqmovn_high_u16 (uint8x8_t a, uint16x8_t b)
9996 {
9997   uint8x16_t result = vcombine_u8 (a, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
9998   __asm__ ("uqxtn2 %0.16b, %1.8h"
9999            : "+w"(result)
10000            : "w"(b)
10001            : /* No clobbers */);
10002   return result;
10003 }
10004
10005 __extension__ extern __inline uint16x8_t
10006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10007 vqmovn_high_u32 (uint16x4_t a, uint32x4_t b)
10008 {
10009   uint16x8_t result = vcombine_u16 (a, vcreate_u16 (__AARCH64_UINT64_C (0x0)));
10010   __asm__ ("uqxtn2 %0.8h, %1.4s"
10011            : "+w"(result)
10012            : "w"(b)
10013            : /* No clobbers */);
10014   return result;
10015 }
10016
10017 __extension__ extern __inline uint32x4_t
10018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10019 vqmovn_high_u64 (uint32x2_t a, uint64x2_t b)
10020 {
10021   uint32x4_t result = vcombine_u32 (a, vcreate_u32 (__AARCH64_UINT64_C (0x0)));
10022   __asm__ ("uqxtn2 %0.4s, %1.2d"
10023            : "+w"(result)
10024            : "w"(b)
10025            : /* No clobbers */);
10026   return result;
10027 }
10028
10029 __extension__ extern __inline uint8x16_t
10030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10031 vqmovun_high_s16 (uint8x8_t a, int16x8_t b)
10032 {
10033   uint8x16_t result = vcombine_u8 (a, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
10034   __asm__ ("sqxtun2 %0.16b, %1.8h"
10035            : "+w"(result)
10036            : "w"(b)
10037            : /* No clobbers */);
10038   return result;
10039 }
10040
10041 __extension__ extern __inline uint16x8_t
10042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10043 vqmovun_high_s32 (uint16x4_t a, int32x4_t b)
10044 {
10045   uint16x8_t result = vcombine_u16 (a, vcreate_u16 (__AARCH64_UINT64_C (0x0)));
10046   __asm__ ("sqxtun2 %0.8h, %1.4s"
10047            : "+w"(result)
10048            : "w"(b)
10049            : /* No clobbers */);
10050   return result;
10051 }
10052
10053 __extension__ extern __inline uint32x4_t
10054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10055 vqmovun_high_s64 (uint32x2_t a, int64x2_t b)
10056 {
10057   uint32x4_t result = vcombine_u32 (a, vcreate_u32 (__AARCH64_UINT64_C (0x0)));
10058   __asm__ ("sqxtun2 %0.4s, %1.2d"
10059            : "+w"(result)
10060            : "w"(b)
10061            : /* No clobbers */);
10062   return result;
10063 }
10064
10065 __extension__ extern __inline int16x4_t
10066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10067 vqrdmulh_n_s16 (int16x4_t a, int16_t b)
10068 {
10069   int16x4_t result;
10070   __asm__ ("sqrdmulh %0.4h,%1.4h,%2.h[0]"
10071            : "=w"(result)
10072            : "w"(a), "x"(b)
10073            : /* No clobbers */);
10074   return result;
10075 }
10076
10077 __extension__ extern __inline int32x2_t
10078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10079 vqrdmulh_n_s32 (int32x2_t a, int32_t b)
10080 {
10081   int32x2_t result;
10082   __asm__ ("sqrdmulh %0.2s,%1.2s,%2.s[0]"
10083            : "=w"(result)
10084            : "w"(a), "w"(b)
10085            : /* No clobbers */);
10086   return result;
10087 }
10088
10089 __extension__ extern __inline int16x8_t
10090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10091 vqrdmulhq_n_s16 (int16x8_t a, int16_t b)
10092 {
10093   int16x8_t result;
10094   __asm__ ("sqrdmulh %0.8h,%1.8h,%2.h[0]"
10095            : "=w"(result)
10096            : "w"(a), "x"(b)
10097            : /* No clobbers */);
10098   return result;
10099 }
10100
10101 __extension__ extern __inline int32x4_t
10102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10103 vqrdmulhq_n_s32 (int32x4_t a, int32_t b)
10104 {
10105   int32x4_t result;
10106   __asm__ ("sqrdmulh %0.4s,%1.4s,%2.s[0]"
10107            : "=w"(result)
10108            : "w"(a), "w"(b)
10109            : /* No clobbers */);
10110   return result;
10111 }
10112
10113 #define vqrshrn_high_n_s16(a, b, c)                                     \
10114   __extension__                                                         \
10115     ({                                                                  \
10116        int16x8_t b_ = (b);                                              \
10117        int8x8_t a_ = (a);                                               \
10118        int8x16_t result = vcombine_s8                                   \
10119                             (a_, vcreate_s8                             \
10120                                    (__AARCH64_UINT64_C (0x0)));         \
10121        __asm__ ("sqrshrn2 %0.16b, %1.8h, #%2"                           \
10122                 : "+w"(result)                                          \
10123                 : "w"(b_), "i"(c)                                       \
10124                 : /* No clobbers */);                                   \
10125        result;                                                          \
10126      })
10127
10128 #define vqrshrn_high_n_s32(a, b, c)                                     \
10129   __extension__                                                         \
10130     ({                                                                  \
10131        int32x4_t b_ = (b);                                              \
10132        int16x4_t a_ = (a);                                              \
10133        int16x8_t result = vcombine_s16                                  \
10134                             (a_, vcreate_s16                            \
10135                                    (__AARCH64_UINT64_C (0x0)));         \
10136        __asm__ ("sqrshrn2 %0.8h, %1.4s, #%2"                            \
10137                 : "+w"(result)                                          \
10138                 : "w"(b_), "i"(c)                                       \
10139                 : /* No clobbers */);                                   \
10140        result;                                                          \
10141      })
10142
10143 #define vqrshrn_high_n_s64(a, b, c)                                     \
10144   __extension__                                                         \
10145     ({                                                                  \
10146        int64x2_t b_ = (b);                                              \
10147        int32x2_t a_ = (a);                                              \
10148        int32x4_t result = vcombine_s32                                  \
10149                             (a_, vcreate_s32                            \
10150                                    (__AARCH64_UINT64_C (0x0)));         \
10151        __asm__ ("sqrshrn2 %0.4s, %1.2d, #%2"                            \
10152                 : "+w"(result)                                          \
10153                 : "w"(b_), "i"(c)                                       \
10154                 : /* No clobbers */);                                   \
10155        result;                                                          \
10156      })
10157
10158 #define vqrshrn_high_n_u16(a, b, c)                                     \
10159   __extension__                                                         \
10160     ({                                                                  \
10161        uint16x8_t b_ = (b);                                             \
10162        uint8x8_t a_ = (a);                                              \
10163        uint8x16_t result = vcombine_u8                                  \
10164                              (a_, vcreate_u8                            \
10165                                     (__AARCH64_UINT64_C (0x0)));        \
10166        __asm__ ("uqrshrn2 %0.16b, %1.8h, #%2"                           \
10167                 : "+w"(result)                                          \
10168                 : "w"(b_), "i"(c)                                       \
10169                 : /* No clobbers */);                                   \
10170        result;                                                          \
10171      })
10172
10173 #define vqrshrn_high_n_u32(a, b, c)                                     \
10174   __extension__                                                         \
10175     ({                                                                  \
10176        uint32x4_t b_ = (b);                                             \
10177        uint16x4_t a_ = (a);                                             \
10178        uint16x8_t result = vcombine_u16                                 \
10179                              (a_, vcreate_u16                           \
10180                                     (__AARCH64_UINT64_C (0x0)));        \
10181        __asm__ ("uqrshrn2 %0.8h, %1.4s, #%2"                            \
10182                 : "+w"(result)                                          \
10183                 : "w"(b_), "i"(c)                                       \
10184                 : /* No clobbers */);                                   \
10185        result;                                                          \
10186      })
10187
10188 #define vqrshrn_high_n_u64(a, b, c)                                     \
10189   __extension__                                                         \
10190     ({                                                                  \
10191        uint64x2_t b_ = (b);                                             \
10192        uint32x2_t a_ = (a);                                             \
10193        uint32x4_t result = vcombine_u32                                 \
10194                              (a_, vcreate_u32                           \
10195                                     (__AARCH64_UINT64_C (0x0)));        \
10196        __asm__ ("uqrshrn2 %0.4s, %1.2d, #%2"                            \
10197                 : "+w"(result)                                          \
10198                 : "w"(b_), "i"(c)                                       \
10199                 : /* No clobbers */);                                   \
10200        result;                                                          \
10201      })
10202
10203 #define vqrshrun_high_n_s16(a, b, c)                                    \
10204   __extension__                                                         \
10205     ({                                                                  \
10206        int16x8_t b_ = (b);                                              \
10207        uint8x8_t a_ = (a);                                              \
10208        uint8x16_t result = vcombine_u8                                  \
10209                              (a_, vcreate_u8                            \
10210                                     (__AARCH64_UINT64_C (0x0)));        \
10211        __asm__ ("sqrshrun2 %0.16b, %1.8h, #%2"                          \
10212                 : "+w"(result)                                          \
10213                 : "w"(b_), "i"(c)                                       \
10214                 : /* No clobbers */);                                   \
10215        result;                                                          \
10216      })
10217
10218 #define vqrshrun_high_n_s32(a, b, c)                                    \
10219   __extension__                                                         \
10220     ({                                                                  \
10221        int32x4_t b_ = (b);                                              \
10222        uint16x4_t a_ = (a);                                             \
10223        uint16x8_t result = vcombine_u16                                 \
10224                              (a_, vcreate_u16                           \
10225                                     (__AARCH64_UINT64_C (0x0)));        \
10226        __asm__ ("sqrshrun2 %0.8h, %1.4s, #%2"                           \
10227                 : "+w"(result)                                          \
10228                 : "w"(b_), "i"(c)                                       \
10229                 : /* No clobbers */);                                   \
10230        result;                                                          \
10231      })
10232
10233 #define vqrshrun_high_n_s64(a, b, c)                                    \
10234   __extension__                                                         \
10235     ({                                                                  \
10236        int64x2_t b_ = (b);                                              \
10237        uint32x2_t a_ = (a);                                             \
10238        uint32x4_t result = vcombine_u32                                 \
10239                              (a_, vcreate_u32                           \
10240                                     (__AARCH64_UINT64_C (0x0)));        \
10241        __asm__ ("sqrshrun2 %0.4s, %1.2d, #%2"                           \
10242                 : "+w"(result)                                          \
10243                 : "w"(b_), "i"(c)                                       \
10244                 : /* No clobbers */);                                   \
10245        result;                                                          \
10246      })
10247
10248 #define vqshrn_high_n_s16(a, b, c)                                      \
10249   __extension__                                                         \
10250     ({                                                                  \
10251        int16x8_t b_ = (b);                                              \
10252        int8x8_t a_ = (a);                                               \
10253        int8x16_t result = vcombine_s8                                   \
10254                             (a_, vcreate_s8                             \
10255                                    (__AARCH64_UINT64_C (0x0)));         \
10256        __asm__ ("sqshrn2 %0.16b, %1.8h, #%2"                            \
10257                 : "+w"(result)                                          \
10258                 : "w"(b_), "i"(c)                                       \
10259                 : /* No clobbers */);                                   \
10260        result;                                                          \
10261      })
10262
10263 #define vqshrn_high_n_s32(a, b, c)                                      \
10264   __extension__                                                         \
10265     ({                                                                  \
10266        int32x4_t b_ = (b);                                              \
10267        int16x4_t a_ = (a);                                              \
10268        int16x8_t result = vcombine_s16                                  \
10269                             (a_, vcreate_s16                            \
10270                                    (__AARCH64_UINT64_C (0x0)));         \
10271        __asm__ ("sqshrn2 %0.8h, %1.4s, #%2"                             \
10272                 : "+w"(result)                                          \
10273                 : "w"(b_), "i"(c)                                       \
10274                 : /* No clobbers */);                                   \
10275        result;                                                          \
10276      })
10277
10278 #define vqshrn_high_n_s64(a, b, c)                                      \
10279   __extension__                                                         \
10280     ({                                                                  \
10281        int64x2_t b_ = (b);                                              \
10282        int32x2_t a_ = (a);                                              \
10283        int32x4_t result = vcombine_s32                                  \
10284                             (a_, vcreate_s32                            \
10285                                    (__AARCH64_UINT64_C (0x0)));         \
10286        __asm__ ("sqshrn2 %0.4s, %1.2d, #%2"                             \
10287                 : "+w"(result)                                          \
10288                 : "w"(b_), "i"(c)                                       \
10289                 : /* No clobbers */);                                   \
10290        result;                                                          \
10291      })
10292
10293 #define vqshrn_high_n_u16(a, b, c)                                      \
10294   __extension__                                                         \
10295     ({                                                                  \
10296        uint16x8_t b_ = (b);                                             \
10297        uint8x8_t a_ = (a);                                              \
10298        uint8x16_t result = vcombine_u8                                  \
10299                              (a_, vcreate_u8                            \
10300                                     (__AARCH64_UINT64_C (0x0)));        \
10301        __asm__ ("uqshrn2 %0.16b, %1.8h, #%2"                            \
10302                 : "+w"(result)                                          \
10303                 : "w"(b_), "i"(c)                                       \
10304                 : /* No clobbers */);                                   \
10305        result;                                                          \
10306      })
10307
10308 #define vqshrn_high_n_u32(a, b, c)                                      \
10309   __extension__                                                         \
10310     ({                                                                  \
10311        uint32x4_t b_ = (b);                                             \
10312        uint16x4_t a_ = (a);                                             \
10313        uint16x8_t result = vcombine_u16                                 \
10314                              (a_, vcreate_u16                           \
10315                                     (__AARCH64_UINT64_C (0x0)));        \
10316        __asm__ ("uqshrn2 %0.8h, %1.4s, #%2"                             \
10317                 : "+w"(result)                                          \
10318                 : "w"(b_), "i"(c)                                       \
10319                 : /* No clobbers */);                                   \
10320        result;                                                          \
10321      })
10322
10323 #define vqshrn_high_n_u64(a, b, c)                                      \
10324   __extension__                                                         \
10325     ({                                                                  \
10326        uint64x2_t b_ = (b);                                             \
10327        uint32x2_t a_ = (a);                                             \
10328        uint32x4_t result = vcombine_u32                                 \
10329                              (a_, vcreate_u32                           \
10330                                     (__AARCH64_UINT64_C (0x0)));        \
10331        __asm__ ("uqshrn2 %0.4s, %1.2d, #%2"                             \
10332                 : "+w"(result)                                          \
10333                 : "w"(b_), "i"(c)                                       \
10334                 : /* No clobbers */);                                   \
10335        result;                                                          \
10336      })
10337
10338 #define vqshrun_high_n_s16(a, b, c)                                     \
10339   __extension__                                                         \
10340     ({                                                                  \
10341        int16x8_t b_ = (b);                                              \
10342        uint8x8_t a_ = (a);                                              \
10343        uint8x16_t result = vcombine_u8                                  \
10344                              (a_, vcreate_u8                            \
10345                                     (__AARCH64_UINT64_C (0x0)));        \
10346        __asm__ ("sqshrun2 %0.16b, %1.8h, #%2"                           \
10347                 : "+w"(result)                                          \
10348                 : "w"(b_), "i"(c)                                       \
10349                 : /* No clobbers */);                                   \
10350        result;                                                          \
10351      })
10352
10353 #define vqshrun_high_n_s32(a, b, c)                                     \
10354   __extension__                                                         \
10355     ({                                                                  \
10356        int32x4_t b_ = (b);                                              \
10357        uint16x4_t a_ = (a);                                             \
10358        uint16x8_t result = vcombine_u16                                 \
10359                              (a_, vcreate_u16                           \
10360                                     (__AARCH64_UINT64_C (0x0)));        \
10361        __asm__ ("sqshrun2 %0.8h, %1.4s, #%2"                            \
10362                 : "+w"(result)                                          \
10363                 : "w"(b_), "i"(c)                                       \
10364                 : /* No clobbers */);                                   \
10365        result;                                                          \
10366      })
10367
10368 #define vqshrun_high_n_s64(a, b, c)                                     \
10369   __extension__                                                         \
10370     ({                                                                  \
10371        int64x2_t b_ = (b);                                              \
10372        uint32x2_t a_ = (a);                                             \
10373        uint32x4_t result = vcombine_u32                                 \
10374                              (a_, vcreate_u32                           \
10375                                     (__AARCH64_UINT64_C (0x0)));        \
10376        __asm__ ("sqshrun2 %0.4s, %1.2d, #%2"                            \
10377                 : "+w"(result)                                          \
10378                 : "w"(b_), "i"(c)                                       \
10379                 : /* No clobbers */);                                   \
10380        result;                                                          \
10381      })
10382
10383 #define vrshrn_high_n_s16(a, b, c)                                      \
10384   __extension__                                                         \
10385     ({                                                                  \
10386        int16x8_t b_ = (b);                                              \
10387        int8x8_t a_ = (a);                                               \
10388        int8x16_t result = vcombine_s8                                   \
10389                             (a_, vcreate_s8                             \
10390                                    (__AARCH64_UINT64_C (0x0)));         \
10391        __asm__ ("rshrn2 %0.16b,%1.8h,#%2"                               \
10392                 : "+w"(result)                                          \
10393                 : "w"(b_), "i"(c)                                       \
10394                 : /* No clobbers */);                                   \
10395        result;                                                          \
10396      })
10397
10398 #define vrshrn_high_n_s32(a, b, c)                                      \
10399   __extension__                                                         \
10400     ({                                                                  \
10401        int32x4_t b_ = (b);                                              \
10402        int16x4_t a_ = (a);                                              \
10403        int16x8_t result = vcombine_s16                                  \
10404                             (a_, vcreate_s16                            \
10405                                    (__AARCH64_UINT64_C (0x0)));         \
10406        __asm__ ("rshrn2 %0.8h,%1.4s,#%2"                                \
10407                 : "+w"(result)                                          \
10408                 : "w"(b_), "i"(c)                                       \
10409                 : /* No clobbers */);                                   \
10410        result;                                                          \
10411      })
10412
10413 #define vrshrn_high_n_s64(a, b, c)                                      \
10414   __extension__                                                         \
10415     ({                                                                  \
10416        int64x2_t b_ = (b);                                              \
10417        int32x2_t a_ = (a);                                              \
10418        int32x4_t result = vcombine_s32                                  \
10419                             (a_, vcreate_s32                            \
10420                                    (__AARCH64_UINT64_C (0x0)));         \
10421        __asm__ ("rshrn2 %0.4s,%1.2d,#%2"                                \
10422                 : "+w"(result)                                          \
10423                 : "w"(b_), "i"(c)                                       \
10424                 : /* No clobbers */);                                   \
10425        result;                                                          \
10426      })
10427
10428 #define vrshrn_high_n_u16(a, b, c)                                      \
10429   __extension__                                                         \
10430     ({                                                                  \
10431        uint16x8_t b_ = (b);                                             \
10432        uint8x8_t a_ = (a);                                              \
10433        uint8x16_t result = vcombine_u8                                  \
10434                             (a_, vcreate_u8                             \
10435                                    (__AARCH64_UINT64_C (0x0)));         \
10436        __asm__ ("rshrn2 %0.16b,%1.8h,#%2"                               \
10437                 : "+w"(result)                                          \
10438                 : "w"(b_), "i"(c)                                       \
10439                 : /* No clobbers */);                                   \
10440        result;                                                          \
10441      })
10442
10443 #define vrshrn_high_n_u32(a, b, c)                                      \
10444   __extension__                                                         \
10445     ({                                                                  \
10446        uint32x4_t b_ = (b);                                             \
10447        uint16x4_t a_ = (a);                                             \
10448        uint16x8_t result = vcombine_u16                                 \
10449                             (a_, vcreate_u16                            \
10450                                    (__AARCH64_UINT64_C (0x0)));         \
10451        __asm__ ("rshrn2 %0.8h,%1.4s,#%2"                                \
10452                 : "+w"(result)                                          \
10453                 : "w"(b_), "i"(c)                                       \
10454                 : /* No clobbers */);                                   \
10455        result;                                                          \
10456      })
10457
10458 #define vrshrn_high_n_u64(a, b, c)                                      \
10459   __extension__                                                         \
10460     ({                                                                  \
10461        uint64x2_t b_ = (b);                                             \
10462        uint32x2_t a_ = (a);                                             \
10463        uint32x4_t result = vcombine_u32                                 \
10464                             (a_, vcreate_u32                            \
10465                                    (__AARCH64_UINT64_C (0x0)));         \
10466        __asm__ ("rshrn2 %0.4s,%1.2d,#%2"                                \
10467                 : "+w"(result)                                          \
10468                 : "w"(b_), "i"(c)                                       \
10469                 : /* No clobbers */);                                   \
10470        result;                                                          \
10471      })
10472
10473 #define vrshrn_n_s16(a, b)                                              \
10474   __extension__                                                         \
10475     ({                                                                  \
10476        int16x8_t a_ = (a);                                              \
10477        int8x8_t result;                                                 \
10478        __asm__ ("rshrn %0.8b,%1.8h,%2"                                  \
10479                 : "=w"(result)                                          \
10480                 : "w"(a_), "i"(b)                                       \
10481                 : /* No clobbers */);                                   \
10482        result;                                                          \
10483      })
10484
10485 #define vrshrn_n_s32(a, b)                                              \
10486   __extension__                                                         \
10487     ({                                                                  \
10488        int32x4_t a_ = (a);                                              \
10489        int16x4_t result;                                                \
10490        __asm__ ("rshrn %0.4h,%1.4s,%2"                                  \
10491                 : "=w"(result)                                          \
10492                 : "w"(a_), "i"(b)                                       \
10493                 : /* No clobbers */);                                   \
10494        result;                                                          \
10495      })
10496
10497 #define vrshrn_n_s64(a, b)                                              \
10498   __extension__                                                         \
10499     ({                                                                  \
10500        int64x2_t a_ = (a);                                              \
10501        int32x2_t result;                                                \
10502        __asm__ ("rshrn %0.2s,%1.2d,%2"                                  \
10503                 : "=w"(result)                                          \
10504                 : "w"(a_), "i"(b)                                       \
10505                 : /* No clobbers */);                                   \
10506        result;                                                          \
10507      })
10508
10509 #define vrshrn_n_u16(a, b)                                              \
10510   __extension__                                                         \
10511     ({                                                                  \
10512        uint16x8_t a_ = (a);                                             \
10513        uint8x8_t result;                                                \
10514        __asm__ ("rshrn %0.8b,%1.8h,%2"                                  \
10515                 : "=w"(result)                                          \
10516                 : "w"(a_), "i"(b)                                       \
10517                 : /* No clobbers */);                                   \
10518        result;                                                          \
10519      })
10520
10521 #define vrshrn_n_u32(a, b)                                              \
10522   __extension__                                                         \
10523     ({                                                                  \
10524        uint32x4_t a_ = (a);                                             \
10525        uint16x4_t result;                                               \
10526        __asm__ ("rshrn %0.4h,%1.4s,%2"                                  \
10527                 : "=w"(result)                                          \
10528                 : "w"(a_), "i"(b)                                       \
10529                 : /* No clobbers */);                                   \
10530        result;                                                          \
10531      })
10532
10533 #define vrshrn_n_u64(a, b)                                              \
10534   __extension__                                                         \
10535     ({                                                                  \
10536        uint64x2_t a_ = (a);                                             \
10537        uint32x2_t result;                                               \
10538        __asm__ ("rshrn %0.2s,%1.2d,%2"                                  \
10539                 : "=w"(result)                                          \
10540                 : "w"(a_), "i"(b)                                       \
10541                 : /* No clobbers */);                                   \
10542        result;                                                          \
10543      })
10544
10545 __extension__ extern __inline uint32x2_t
10546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10547 vrsqrte_u32 (uint32x2_t a)
10548 {
10549   uint32x2_t result;
10550   __asm__ ("ursqrte %0.2s,%1.2s"
10551            : "=w"(result)
10552            : "w"(a)
10553            : /* No clobbers */);
10554   return result;
10555 }
10556
10557 __extension__ extern __inline uint32x4_t
10558 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10559 vrsqrteq_u32 (uint32x4_t a)
10560 {
10561   uint32x4_t result;
10562   __asm__ ("ursqrte %0.4s,%1.4s"
10563            : "=w"(result)
10564            : "w"(a)
10565            : /* No clobbers */);
10566   return result;
10567 }
10568
10569 #define vshrn_high_n_s16(a, b, c)                                       \
10570   __extension__                                                         \
10571     ({                                                                  \
10572        int16x8_t b_ = (b);                                              \
10573        int8x8_t a_ = (a);                                               \
10574        int8x16_t result = vcombine_s8                                   \
10575                             (a_, vcreate_s8                             \
10576                                    (__AARCH64_UINT64_C (0x0)));         \
10577        __asm__ ("shrn2 %0.16b,%1.8h,#%2"                                \
10578                 : "+w"(result)                                          \
10579                 : "w"(b_), "i"(c)                                       \
10580                 : /* No clobbers */);                                   \
10581        result;                                                          \
10582      })
10583
10584 #define vshrn_high_n_s32(a, b, c)                                       \
10585   __extension__                                                         \
10586     ({                                                                  \
10587        int32x4_t b_ = (b);                                              \
10588        int16x4_t a_ = (a);                                              \
10589        int16x8_t result = vcombine_s16                                  \
10590                             (a_, vcreate_s16                            \
10591                                    (__AARCH64_UINT64_C (0x0)));         \
10592        __asm__ ("shrn2 %0.8h,%1.4s,#%2"                                 \
10593                 : "+w"(result)                                          \
10594                 : "w"(b_), "i"(c)                                       \
10595                 : /* No clobbers */);                                   \
10596        result;                                                          \
10597      })
10598
10599 #define vshrn_high_n_s64(a, b, c)                                       \
10600   __extension__                                                         \
10601     ({                                                                  \
10602        int64x2_t b_ = (b);                                              \
10603        int32x2_t a_ = (a);                                              \
10604        int32x4_t result = vcombine_s32                                  \
10605                             (a_, vcreate_s32                            \
10606                                    (__AARCH64_UINT64_C (0x0)));         \
10607        __asm__ ("shrn2 %0.4s,%1.2d,#%2"                                 \
10608                 : "+w"(result)                                          \
10609                 : "w"(b_), "i"(c)                                       \
10610                 : /* No clobbers */);                                   \
10611        result;                                                          \
10612      })
10613
10614 #define vshrn_high_n_u16(a, b, c)                                       \
10615   __extension__                                                         \
10616     ({                                                                  \
10617        uint16x8_t b_ = (b);                                             \
10618        uint8x8_t a_ = (a);                                              \
10619        uint8x16_t result = vcombine_u8                                  \
10620                             (a_, vcreate_u8                             \
10621                                    (__AARCH64_UINT64_C (0x0)));         \
10622        __asm__ ("shrn2 %0.16b,%1.8h,#%2"                                \
10623                 : "+w"(result)                                          \
10624                 : "w"(b_), "i"(c)                                       \
10625                 : /* No clobbers */);                                   \
10626        result;                                                          \
10627      })
10628
10629 #define vshrn_high_n_u32(a, b, c)                                       \
10630   __extension__                                                         \
10631     ({                                                                  \
10632        uint32x4_t b_ = (b);                                             \
10633        uint16x4_t a_ = (a);                                             \
10634        uint16x8_t result = vcombine_u16                                 \
10635                             (a_, vcreate_u16                            \
10636                                    (__AARCH64_UINT64_C (0x0)));         \
10637        __asm__ ("shrn2 %0.8h,%1.4s,#%2"                                 \
10638                 : "+w"(result)                                          \
10639                 : "w"(b_), "i"(c)                                       \
10640                 : /* No clobbers */);                                   \
10641        result;                                                          \
10642      })
10643
10644 #define vshrn_high_n_u64(a, b, c)                                       \
10645   __extension__                                                         \
10646     ({                                                                  \
10647        uint64x2_t b_ = (b);                                             \
10648        uint32x2_t a_ = (a);                                             \
10649        uint32x4_t result = vcombine_u32                                 \
10650                             (a_, vcreate_u32                            \
10651                                    (__AARCH64_UINT64_C (0x0)));         \
10652        __asm__ ("shrn2 %0.4s,%1.2d,#%2"                                 \
10653                 : "+w"(result)                                          \
10654                 : "w"(b_), "i"(c)                                       \
10655                 : /* No clobbers */);                                   \
10656        result;                                                          \
10657      })
10658
10659 #define vshrn_n_s16(a, b)                                               \
10660   __extension__                                                         \
10661     ({                                                                  \
10662        int16x8_t a_ = (a);                                              \
10663        int8x8_t result;                                                 \
10664        __asm__ ("shrn %0.8b,%1.8h,%2"                                   \
10665                 : "=w"(result)                                          \
10666                 : "w"(a_), "i"(b)                                       \
10667                 : /* No clobbers */);                                   \
10668        result;                                                          \
10669      })
10670
10671 #define vshrn_n_s32(a, b)                                               \
10672   __extension__                                                         \
10673     ({                                                                  \
10674        int32x4_t a_ = (a);                                              \
10675        int16x4_t result;                                                \
10676        __asm__ ("shrn %0.4h,%1.4s,%2"                                   \
10677                 : "=w"(result)                                          \
10678                 : "w"(a_), "i"(b)                                       \
10679                 : /* No clobbers */);                                   \
10680        result;                                                          \
10681      })
10682
10683 #define vshrn_n_s64(a, b)                                               \
10684   __extension__                                                         \
10685     ({                                                                  \
10686        int64x2_t a_ = (a);                                              \
10687        int32x2_t result;                                                \
10688        __asm__ ("shrn %0.2s,%1.2d,%2"                                   \
10689                 : "=w"(result)                                          \
10690                 : "w"(a_), "i"(b)                                       \
10691                 : /* No clobbers */);                                   \
10692        result;                                                          \
10693      })
10694
10695 #define vshrn_n_u16(a, b)                                               \
10696   __extension__                                                         \
10697     ({                                                                  \
10698        uint16x8_t a_ = (a);                                             \
10699        uint8x8_t result;                                                \
10700        __asm__ ("shrn %0.8b,%1.8h,%2"                                   \
10701                 : "=w"(result)                                          \
10702                 : "w"(a_), "i"(b)                                       \
10703                 : /* No clobbers */);                                   \
10704        result;                                                          \
10705      })
10706
10707 #define vshrn_n_u32(a, b)                                               \
10708   __extension__                                                         \
10709     ({                                                                  \
10710        uint32x4_t a_ = (a);                                             \
10711        uint16x4_t result;                                               \
10712        __asm__ ("shrn %0.4h,%1.4s,%2"                                   \
10713                 : "=w"(result)                                          \
10714                 : "w"(a_), "i"(b)                                       \
10715                 : /* No clobbers */);                                   \
10716        result;                                                          \
10717      })
10718
10719 #define vshrn_n_u64(a, b)                                               \
10720   __extension__                                                         \
10721     ({                                                                  \
10722        uint64x2_t a_ = (a);                                             \
10723        uint32x2_t result;                                               \
10724        __asm__ ("shrn %0.2s,%1.2d,%2"                                   \
10725                 : "=w"(result)                                          \
10726                 : "w"(a_), "i"(b)                                       \
10727                 : /* No clobbers */);                                   \
10728        result;                                                          \
10729      })
10730
10731 #define vsli_n_p8(a, b, c)                                              \
10732   __extension__                                                         \
10733     ({                                                                  \
10734        poly8x8_t b_ = (b);                                              \
10735        poly8x8_t a_ = (a);                                              \
10736        poly8x8_t result;                                                \
10737        __asm__ ("sli %0.8b,%2.8b,%3"                                    \
10738                 : "=w"(result)                                          \
10739                 : "0"(a_), "w"(b_), "i"(c)                              \
10740                 : /* No clobbers */);                                   \
10741        result;                                                          \
10742      })
10743
10744 #define vsli_n_p16(a, b, c)                                             \
10745   __extension__                                                         \
10746     ({                                                                  \
10747        poly16x4_t b_ = (b);                                             \
10748        poly16x4_t a_ = (a);                                             \
10749        poly16x4_t result;                                               \
10750        __asm__ ("sli %0.4h,%2.4h,%3"                                    \
10751                 : "=w"(result)                                          \
10752                 : "0"(a_), "w"(b_), "i"(c)                              \
10753                 : /* No clobbers */);                                   \
10754        result;                                                          \
10755      })
10756
10757 #define vsliq_n_p8(a, b, c)                                             \
10758   __extension__                                                         \
10759     ({                                                                  \
10760        poly8x16_t b_ = (b);                                             \
10761        poly8x16_t a_ = (a);                                             \
10762        poly8x16_t result;                                               \
10763        __asm__ ("sli %0.16b,%2.16b,%3"                                  \
10764                 : "=w"(result)                                          \
10765                 : "0"(a_), "w"(b_), "i"(c)                              \
10766                 : /* No clobbers */);                                   \
10767        result;                                                          \
10768      })
10769
10770 #define vsliq_n_p16(a, b, c)                                            \
10771   __extension__                                                         \
10772     ({                                                                  \
10773        poly16x8_t b_ = (b);                                             \
10774        poly16x8_t a_ = (a);                                             \
10775        poly16x8_t result;                                               \
10776        __asm__ ("sli %0.8h,%2.8h,%3"                                    \
10777                 : "=w"(result)                                          \
10778                 : "0"(a_), "w"(b_), "i"(c)                              \
10779                 : /* No clobbers */);                                   \
10780        result;                                                          \
10781      })
10782
10783 #define vsri_n_p8(a, b, c)                                              \
10784   __extension__                                                         \
10785     ({                                                                  \
10786        poly8x8_t b_ = (b);                                              \
10787        poly8x8_t a_ = (a);                                              \
10788        poly8x8_t result;                                                \
10789        __asm__ ("sri %0.8b,%2.8b,%3"                                    \
10790                 : "=w"(result)                                          \
10791                 : "0"(a_), "w"(b_), "i"(c)                              \
10792                 : /* No clobbers */);                                   \
10793        result;                                                          \
10794      })
10795
10796 #define vsri_n_p16(a, b, c)                                             \
10797   __extension__                                                         \
10798     ({                                                                  \
10799        poly16x4_t b_ = (b);                                             \
10800        poly16x4_t a_ = (a);                                             \
10801        poly16x4_t result;                                               \
10802        __asm__ ("sri %0.4h,%2.4h,%3"                                    \
10803                 : "=w"(result)                                          \
10804                 : "0"(a_), "w"(b_), "i"(c)                              \
10805                 : /* No clobbers */);                                   \
10806        result;                                                          \
10807      })
10808
10809 #define vsri_n_p64(a, b, c)                                             \
10810   __extension__                                                         \
10811     ({                                                                  \
10812        poly64x1_t b_ = (b);                                             \
10813        poly64x1_t a_ = (a);                                             \
10814        poly64x1_t result;                                               \
10815        __asm__ ("sri %d0,%d2,%3"                                        \
10816                 : "=w"(result)                                          \
10817                 : "0"(a_), "w"(b_), "i"(c)                              \
10818                 : /* No clobbers.  */);                                 \
10819        result;                                                          \
10820      })
10821
10822 #define vsriq_n_p8(a, b, c)                                             \
10823   __extension__                                                         \
10824     ({                                                                  \
10825        poly8x16_t b_ = (b);                                             \
10826        poly8x16_t a_ = (a);                                             \
10827        poly8x16_t result;                                               \
10828        __asm__ ("sri %0.16b,%2.16b,%3"                                  \
10829                 : "=w"(result)                                          \
10830                 : "0"(a_), "w"(b_), "i"(c)                              \
10831                 : /* No clobbers */);                                   \
10832        result;                                                          \
10833      })
10834
10835 #define vsriq_n_p16(a, b, c)                                            \
10836   __extension__                                                         \
10837     ({                                                                  \
10838        poly16x8_t b_ = (b);                                             \
10839        poly16x8_t a_ = (a);                                             \
10840        poly16x8_t result;                                               \
10841        __asm__ ("sri %0.8h,%2.8h,%3"                                    \
10842                 : "=w"(result)                                          \
10843                 : "0"(a_), "w"(b_), "i"(c)                              \
10844                 : /* No clobbers */);                                   \
10845        result;                                                          \
10846      })
10847
10848 #define vsriq_n_p64(a, b, c)                                            \
10849   __extension__                                                         \
10850     ({                                                                  \
10851        poly64x2_t b_ = (b);                                             \
10852        poly64x2_t a_ = (a);                                             \
10853        poly64x2_t result;                                               \
10854        __asm__ ("sri %0.2d,%2.2d,%3"                                    \
10855                 : "=w"(result)                                          \
10856                 : "0"(a_), "w"(b_), "i"(c)                              \
10857                 : /* No clobbers.  */);                                 \
10858        result;                                                          \
10859      })
10860
10861 __extension__ extern __inline uint8x8_t
10862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10863 vtst_p8 (poly8x8_t a, poly8x8_t b)
10864 {
10865   return (uint8x8_t) ((((uint8x8_t) a) & ((uint8x8_t) b))
10866                        != 0);
10867 }
10868
10869 __extension__ extern __inline uint16x4_t
10870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10871 vtst_p16 (poly16x4_t a, poly16x4_t b)
10872 {
10873   return (uint16x4_t) ((((uint16x4_t) a) & ((uint16x4_t) b))
10874                        != 0);
10875 }
10876
10877 __extension__ extern __inline uint64x1_t
10878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10879 vtst_p64 (poly64x1_t a, poly64x1_t b)
10880 {
10881   return (uint64x1_t) ((a & b) != __AARCH64_INT64_C (0));
10882 }
10883
10884 __extension__ extern __inline uint8x16_t
10885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10886 vtstq_p8 (poly8x16_t a, poly8x16_t b)
10887 {
10888   return (uint8x16_t) ((((uint8x16_t) a) & ((uint8x16_t) b))
10889                        != 0);
10890 }
10891
10892 __extension__ extern __inline uint16x8_t
10893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10894 vtstq_p16 (poly16x8_t a, poly16x8_t b)
10895 {
10896   return (uint16x8_t) ((((uint16x8_t) a) & ((uint16x8_t) b))
10897                        != 0);
10898 }
10899
10900 __extension__ extern __inline uint64x2_t
10901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10902 vtstq_p64 (poly64x2_t a, poly64x2_t b)
10903 {
10904   return (uint64x2_t) ((((uint64x2_t) a) & ((uint64x2_t) b))
10905                        != __AARCH64_INT64_C (0));
10906 }
10907
10908 /* End of temporary inline asm implementations.  */
10909
10910 /* Start of temporary inline asm for vldn, vstn and friends.  */
10911
10912 /* Create struct element types for duplicating loads.
10913
10914    Create 2 element structures of:
10915
10916    +------+----+----+----+----+
10917    |      | 8  | 16 | 32 | 64 |
10918    +------+----+----+----+----+
10919    |int   | Y  | Y  | N  | N  |
10920    +------+----+----+----+----+
10921    |uint  | Y  | Y  | N  | N  |
10922    +------+----+----+----+----+
10923    |float | -  | Y  | N  | N  |
10924    +------+----+----+----+----+
10925    |poly  | Y  | Y  | -  | -  |
10926    +------+----+----+----+----+
10927
10928    Create 3 element structures of:
10929
10930    +------+----+----+----+----+
10931    |      | 8  | 16 | 32 | 64 |
10932    +------+----+----+----+----+
10933    |int   | Y  | Y  | Y  | Y  |
10934    +------+----+----+----+----+
10935    |uint  | Y  | Y  | Y  | Y  |
10936    +------+----+----+----+----+
10937    |float | -  | Y  | Y  | Y  |
10938    +------+----+----+----+----+
10939    |poly  | Y  | Y  | -  | -  |
10940    +------+----+----+----+----+
10941
10942    Create 4 element structures of:
10943
10944    +------+----+----+----+----+
10945    |      | 8  | 16 | 32 | 64 |
10946    +------+----+----+----+----+
10947    |int   | Y  | N  | N  | Y  |
10948    +------+----+----+----+----+
10949    |uint  | Y  | N  | N  | Y  |
10950    +------+----+----+----+----+
10951    |float | -  | N  | N  | Y  |
10952    +------+----+----+----+----+
10953    |poly  | Y  | N  | -  | -  |
10954    +------+----+----+----+----+
10955
10956   This is required for casting memory reference.  */
10957 #define __STRUCTN(t, sz, nelem)                 \
10958   typedef struct t ## sz ## x ## nelem ## _t {  \
10959     t ## sz ## _t val[nelem];                   \
10960   }  t ## sz ## x ## nelem ## _t;
10961
10962 /* 2-element structs.  */
10963 __STRUCTN (int, 8, 2)
10964 __STRUCTN (int, 16, 2)
10965 __STRUCTN (uint, 8, 2)
10966 __STRUCTN (uint, 16, 2)
10967 __STRUCTN (float, 16, 2)
10968 __STRUCTN (poly, 8, 2)
10969 __STRUCTN (poly, 16, 2)
10970 /* 3-element structs.  */
10971 __STRUCTN (int, 8, 3)
10972 __STRUCTN (int, 16, 3)
10973 __STRUCTN (int, 32, 3)
10974 __STRUCTN (int, 64, 3)
10975 __STRUCTN (uint, 8, 3)
10976 __STRUCTN (uint, 16, 3)
10977 __STRUCTN (uint, 32, 3)
10978 __STRUCTN (uint, 64, 3)
10979 __STRUCTN (float, 16, 3)
10980 __STRUCTN (float, 32, 3)
10981 __STRUCTN (float, 64, 3)
10982 __STRUCTN (poly, 8, 3)
10983 __STRUCTN (poly, 16, 3)
10984 /* 4-element structs.  */
10985 __STRUCTN (int, 8, 4)
10986 __STRUCTN (int, 64, 4)
10987 __STRUCTN (uint, 8, 4)
10988 __STRUCTN (uint, 64, 4)
10989 __STRUCTN (poly, 8, 4)
10990 __STRUCTN (float, 64, 4)
10991 #undef __STRUCTN
10992
10993
10994 #define __ST2_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10995                         qmode, ptr_mode, funcsuffix, signedtype)             \
10996 __extension__ extern __inline void                                           \
10997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10998 vst2_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10999                           intype __b, const int __c)                         \
11000 {                                                                            \
11001   __builtin_aarch64_simd_oi __o;                                             \
11002   largetype __temp;                                                          \
11003   __temp.val[0]                                                              \
11004     = vcombine_##funcsuffix (__b.val[0],                                     \
11005                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11006   __temp.val[1]                                                              \
11007     = vcombine_##funcsuffix (__b.val[1],                                     \
11008                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11009   __o = __builtin_aarch64_set_qregoi##qmode (__o,                            \
11010                                              (signedtype) __temp.val[0], 0); \
11011   __o = __builtin_aarch64_set_qregoi##qmode (__o,                            \
11012                                              (signedtype) __temp.val[1], 1); \
11013   __builtin_aarch64_st2_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
11014                                      __ptr, __o, __c);                       \
11015 }
11016
11017 __ST2_LANE_FUNC (float16x4x2_t, float16x8x2_t, float16_t, v4hf, v8hf, hf, f16,
11018                  float16x8_t)
11019 __ST2_LANE_FUNC (float32x2x2_t, float32x4x2_t, float32_t, v2sf, v4sf, sf, f32,
11020                  float32x4_t)
11021 __ST2_LANE_FUNC (float64x1x2_t, float64x2x2_t, float64_t, df, v2df, df, f64,
11022                  float64x2_t)
11023 __ST2_LANE_FUNC (poly8x8x2_t, poly8x16x2_t, poly8_t, v8qi, v16qi, qi, p8,
11024                  int8x16_t)
11025 __ST2_LANE_FUNC (poly16x4x2_t, poly16x8x2_t, poly16_t, v4hi, v8hi, hi, p16,
11026                  int16x8_t)
11027 __ST2_LANE_FUNC (poly64x1x2_t, poly64x2x2_t, poly64_t, di, v2di_ssps, di, p64,
11028                  poly64x2_t)
11029 __ST2_LANE_FUNC (int8x8x2_t, int8x16x2_t, int8_t, v8qi, v16qi, qi, s8,
11030                  int8x16_t)
11031 __ST2_LANE_FUNC (int16x4x2_t, int16x8x2_t, int16_t, v4hi, v8hi, hi, s16,
11032                  int16x8_t)
11033 __ST2_LANE_FUNC (int32x2x2_t, int32x4x2_t, int32_t, v2si, v4si, si, s32,
11034                  int32x4_t)
11035 __ST2_LANE_FUNC (int64x1x2_t, int64x2x2_t, int64_t, di, v2di, di, s64,
11036                  int64x2_t)
11037 __ST2_LANE_FUNC (uint8x8x2_t, uint8x16x2_t, uint8_t, v8qi, v16qi, qi, u8,
11038                  int8x16_t)
11039 __ST2_LANE_FUNC (uint16x4x2_t, uint16x8x2_t, uint16_t, v4hi, v8hi, hi, u16,
11040                  int16x8_t)
11041 __ST2_LANE_FUNC (uint32x2x2_t, uint32x4x2_t, uint32_t, v2si, v4si, si, u32,
11042                  int32x4_t)
11043 __ST2_LANE_FUNC (uint64x1x2_t, uint64x2x2_t, uint64_t, di, v2di, di, u64,
11044                  int64x2_t)
11045
11046 #undef __ST2_LANE_FUNC
11047 #define __ST2_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
11048 __extension__ extern __inline void                                          \
11049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11050 vst2q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
11051                            intype __b, const int __c)                       \
11052 {                                                                           \
11053   union { intype __i;                                                       \
11054           __builtin_aarch64_simd_oi __o; } __temp = { __b };                \
11055   __builtin_aarch64_st2_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
11056                                     __ptr, __temp.__o, __c);                \
11057 }
11058
11059 __ST2_LANE_FUNC (float16x8x2_t, float16_t, v8hf, hf, f16)
11060 __ST2_LANE_FUNC (float32x4x2_t, float32_t, v4sf, sf, f32)
11061 __ST2_LANE_FUNC (float64x2x2_t, float64_t, v2df, df, f64)
11062 __ST2_LANE_FUNC (poly8x16x2_t, poly8_t, v16qi, qi, p8)
11063 __ST2_LANE_FUNC (poly16x8x2_t, poly16_t, v8hi, hi, p16)
11064 __ST2_LANE_FUNC (poly64x2x2_t, poly64_t, v2di, di, p64)
11065 __ST2_LANE_FUNC (int8x16x2_t, int8_t, v16qi, qi, s8)
11066 __ST2_LANE_FUNC (int16x8x2_t, int16_t, v8hi, hi, s16)
11067 __ST2_LANE_FUNC (int32x4x2_t, int32_t, v4si, si, s32)
11068 __ST2_LANE_FUNC (int64x2x2_t, int64_t, v2di, di, s64)
11069 __ST2_LANE_FUNC (uint8x16x2_t, uint8_t, v16qi, qi, u8)
11070 __ST2_LANE_FUNC (uint16x8x2_t, uint16_t, v8hi, hi, u16)
11071 __ST2_LANE_FUNC (uint32x4x2_t, uint32_t, v4si, si, u32)
11072 __ST2_LANE_FUNC (uint64x2x2_t, uint64_t, v2di, di, u64)
11073
11074 #define __ST3_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
11075                         qmode, ptr_mode, funcsuffix, signedtype)             \
11076 __extension__ extern __inline void                                           \
11077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11078 vst3_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
11079                           intype __b, const int __c)                         \
11080 {                                                                            \
11081   __builtin_aarch64_simd_ci __o;                                             \
11082   largetype __temp;                                                          \
11083   __temp.val[0]                                                              \
11084     = vcombine_##funcsuffix (__b.val[0],                                     \
11085                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11086   __temp.val[1]                                                              \
11087     = vcombine_##funcsuffix (__b.val[1],                                     \
11088                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11089   __temp.val[2]                                                              \
11090     = vcombine_##funcsuffix (__b.val[2],                                     \
11091                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11092   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
11093                                              (signedtype) __temp.val[0], 0); \
11094   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
11095                                              (signedtype) __temp.val[1], 1); \
11096   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
11097                                              (signedtype) __temp.val[2], 2); \
11098   __builtin_aarch64_st3_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
11099                                      __ptr, __o, __c);                       \
11100 }
11101
11102 __ST3_LANE_FUNC (float16x4x3_t, float16x8x3_t, float16_t, v4hf, v8hf, hf, f16,
11103                  float16x8_t)
11104 __ST3_LANE_FUNC (float32x2x3_t, float32x4x3_t, float32_t, v2sf, v4sf, sf, f32,
11105                  float32x4_t)
11106 __ST3_LANE_FUNC (float64x1x3_t, float64x2x3_t, float64_t, df, v2df, df, f64,
11107                  float64x2_t)
11108 __ST3_LANE_FUNC (poly8x8x3_t, poly8x16x3_t, poly8_t, v8qi, v16qi, qi, p8,
11109                  int8x16_t)
11110 __ST3_LANE_FUNC (poly16x4x3_t, poly16x8x3_t, poly16_t, v4hi, v8hi, hi, p16,
11111                  int16x8_t)
11112 __ST3_LANE_FUNC (poly64x1x3_t, poly64x2x3_t, poly64_t, di, v2di_ssps, di, p64,
11113                  poly64x2_t)
11114 __ST3_LANE_FUNC (int8x8x3_t, int8x16x3_t, int8_t, v8qi, v16qi, qi, s8,
11115                  int8x16_t)
11116 __ST3_LANE_FUNC (int16x4x3_t, int16x8x3_t, int16_t, v4hi, v8hi, hi, s16,
11117                  int16x8_t)
11118 __ST3_LANE_FUNC (int32x2x3_t, int32x4x3_t, int32_t, v2si, v4si, si, s32,
11119                  int32x4_t)
11120 __ST3_LANE_FUNC (int64x1x3_t, int64x2x3_t, int64_t, di, v2di, di, s64,
11121                  int64x2_t)
11122 __ST3_LANE_FUNC (uint8x8x3_t, uint8x16x3_t, uint8_t, v8qi, v16qi, qi, u8,
11123                  int8x16_t)
11124 __ST3_LANE_FUNC (uint16x4x3_t, uint16x8x3_t, uint16_t, v4hi, v8hi, hi, u16,
11125                  int16x8_t)
11126 __ST3_LANE_FUNC (uint32x2x3_t, uint32x4x3_t, uint32_t, v2si, v4si, si, u32,
11127                  int32x4_t)
11128 __ST3_LANE_FUNC (uint64x1x3_t, uint64x2x3_t, uint64_t, di, v2di, di, u64,
11129                  int64x2_t)
11130
11131 #undef __ST3_LANE_FUNC
11132 #define __ST3_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
11133 __extension__ extern __inline void                                          \
11134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11135 vst3q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
11136                            intype __b, const int __c)                       \
11137 {                                                                           \
11138   union { intype __i;                                                       \
11139           __builtin_aarch64_simd_ci __o; } __temp = { __b };                \
11140   __builtin_aarch64_st3_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
11141                                     __ptr, __temp.__o, __c);                \
11142 }
11143
11144 __ST3_LANE_FUNC (float16x8x3_t, float16_t, v8hf, hf, f16)
11145 __ST3_LANE_FUNC (float32x4x3_t, float32_t, v4sf, sf, f32)
11146 __ST3_LANE_FUNC (float64x2x3_t, float64_t, v2df, df, f64)
11147 __ST3_LANE_FUNC (poly8x16x3_t, poly8_t, v16qi, qi, p8)
11148 __ST3_LANE_FUNC (poly16x8x3_t, poly16_t, v8hi, hi, p16)
11149 __ST3_LANE_FUNC (poly64x2x3_t, poly64_t, v2di, di, p64)
11150 __ST3_LANE_FUNC (int8x16x3_t, int8_t, v16qi, qi, s8)
11151 __ST3_LANE_FUNC (int16x8x3_t, int16_t, v8hi, hi, s16)
11152 __ST3_LANE_FUNC (int32x4x3_t, int32_t, v4si, si, s32)
11153 __ST3_LANE_FUNC (int64x2x3_t, int64_t, v2di, di, s64)
11154 __ST3_LANE_FUNC (uint8x16x3_t, uint8_t, v16qi, qi, u8)
11155 __ST3_LANE_FUNC (uint16x8x3_t, uint16_t, v8hi, hi, u16)
11156 __ST3_LANE_FUNC (uint32x4x3_t, uint32_t, v4si, si, u32)
11157 __ST3_LANE_FUNC (uint64x2x3_t, uint64_t, v2di, di, u64)
11158
11159 #define __ST4_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
11160                         qmode, ptr_mode, funcsuffix, signedtype)             \
11161 __extension__ extern __inline void                                           \
11162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11163 vst4_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
11164                           intype __b, const int __c)                         \
11165 {                                                                            \
11166   __builtin_aarch64_simd_xi __o;                                             \
11167   largetype __temp;                                                          \
11168   __temp.val[0]                                                              \
11169     = vcombine_##funcsuffix (__b.val[0],                                     \
11170                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11171   __temp.val[1]                                                              \
11172     = vcombine_##funcsuffix (__b.val[1],                                     \
11173                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11174   __temp.val[2]                                                              \
11175     = vcombine_##funcsuffix (__b.val[2],                                     \
11176                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11177   __temp.val[3]                                                              \
11178     = vcombine_##funcsuffix (__b.val[3],                                     \
11179                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
11180   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
11181                                              (signedtype) __temp.val[0], 0); \
11182   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
11183                                              (signedtype) __temp.val[1], 1); \
11184   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
11185                                              (signedtype) __temp.val[2], 2); \
11186   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
11187                                              (signedtype) __temp.val[3], 3); \
11188   __builtin_aarch64_st4_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
11189                                      __ptr, __o, __c);                       \
11190 }
11191
11192 __ST4_LANE_FUNC (float16x4x4_t, float16x8x4_t, float16_t, v4hf, v8hf, hf, f16,
11193                  float16x8_t)
11194 __ST4_LANE_FUNC (float32x2x4_t, float32x4x4_t, float32_t, v2sf, v4sf, sf, f32,
11195                  float32x4_t)
11196 __ST4_LANE_FUNC (float64x1x4_t, float64x2x4_t, float64_t, df, v2df, df, f64,
11197                  float64x2_t)
11198 __ST4_LANE_FUNC (poly8x8x4_t, poly8x16x4_t, poly8_t, v8qi, v16qi, qi, p8,
11199                  int8x16_t)
11200 __ST4_LANE_FUNC (poly16x4x4_t, poly16x8x4_t, poly16_t, v4hi, v8hi, hi, p16,
11201                  int16x8_t)
11202 __ST4_LANE_FUNC (poly64x1x4_t, poly64x2x4_t, poly64_t, di, v2di_ssps, di, p64,
11203                  poly64x2_t)
11204 __ST4_LANE_FUNC (int8x8x4_t, int8x16x4_t, int8_t, v8qi, v16qi, qi, s8,
11205                  int8x16_t)
11206 __ST4_LANE_FUNC (int16x4x4_t, int16x8x4_t, int16_t, v4hi, v8hi, hi, s16,
11207                  int16x8_t)
11208 __ST4_LANE_FUNC (int32x2x4_t, int32x4x4_t, int32_t, v2si, v4si, si, s32,
11209                  int32x4_t)
11210 __ST4_LANE_FUNC (int64x1x4_t, int64x2x4_t, int64_t, di, v2di, di, s64,
11211                  int64x2_t)
11212 __ST4_LANE_FUNC (uint8x8x4_t, uint8x16x4_t, uint8_t, v8qi, v16qi, qi, u8,
11213                  int8x16_t)
11214 __ST4_LANE_FUNC (uint16x4x4_t, uint16x8x4_t, uint16_t, v4hi, v8hi, hi, u16,
11215                  int16x8_t)
11216 __ST4_LANE_FUNC (uint32x2x4_t, uint32x4x4_t, uint32_t, v2si, v4si, si, u32,
11217                  int32x4_t)
11218 __ST4_LANE_FUNC (uint64x1x4_t, uint64x2x4_t, uint64_t, di, v2di, di, u64,
11219                  int64x2_t)
11220
11221 #undef __ST4_LANE_FUNC
11222 #define __ST4_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
11223 __extension__ extern __inline void                                          \
11224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11225 vst4q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
11226                            intype __b, const int __c)                       \
11227 {                                                                           \
11228   union { intype __i;                                                       \
11229           __builtin_aarch64_simd_xi __o; } __temp = { __b };                \
11230   __builtin_aarch64_st4_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
11231                                     __ptr, __temp.__o, __c);                \
11232 }
11233
11234 __ST4_LANE_FUNC (float16x8x4_t, float16_t, v8hf, hf, f16)
11235 __ST4_LANE_FUNC (float32x4x4_t, float32_t, v4sf, sf, f32)
11236 __ST4_LANE_FUNC (float64x2x4_t, float64_t, v2df, df, f64)
11237 __ST4_LANE_FUNC (poly8x16x4_t, poly8_t, v16qi, qi, p8)
11238 __ST4_LANE_FUNC (poly16x8x4_t, poly16_t, v8hi, hi, p16)
11239 __ST4_LANE_FUNC (poly64x2x4_t, poly64_t, v2di, di, p64)
11240 __ST4_LANE_FUNC (int8x16x4_t, int8_t, v16qi, qi, s8)
11241 __ST4_LANE_FUNC (int16x8x4_t, int16_t, v8hi, hi, s16)
11242 __ST4_LANE_FUNC (int32x4x4_t, int32_t, v4si, si, s32)
11243 __ST4_LANE_FUNC (int64x2x4_t, int64_t, v2di, di, s64)
11244 __ST4_LANE_FUNC (uint8x16x4_t, uint8_t, v16qi, qi, u8)
11245 __ST4_LANE_FUNC (uint16x8x4_t, uint16_t, v8hi, hi, u16)
11246 __ST4_LANE_FUNC (uint32x4x4_t, uint32_t, v4si, si, u32)
11247 __ST4_LANE_FUNC (uint64x2x4_t, uint64_t, v2di, di, u64)
11248
11249 __extension__ extern __inline int64_t
11250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11251 vaddlv_s32 (int32x2_t a)
11252 {
11253   int64_t result;
11254   __asm__ ("saddlp %0.1d, %1.2s" : "=w"(result) : "w"(a) : );
11255   return result;
11256 }
11257
11258 __extension__ extern __inline uint64_t
11259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11260 vaddlv_u32 (uint32x2_t a)
11261 {
11262   uint64_t result;
11263   __asm__ ("uaddlp %0.1d, %1.2s" : "=w"(result) : "w"(a) : );
11264   return result;
11265 }
11266
11267 __extension__ extern __inline int16x4_t
11268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11269 vqdmulh_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __c)
11270 {
11271   return __builtin_aarch64_sqdmulh_laneqv4hi (__a, __b, __c);
11272 }
11273
11274 __extension__ extern __inline int32x2_t
11275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11276 vqdmulh_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __c)
11277 {
11278   return __builtin_aarch64_sqdmulh_laneqv2si (__a, __b, __c);
11279 }
11280
11281 __extension__ extern __inline int16x8_t
11282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11283 vqdmulhq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __c)
11284 {
11285   return __builtin_aarch64_sqdmulh_laneqv8hi (__a, __b, __c);
11286 }
11287
11288 __extension__ extern __inline int32x4_t
11289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11290 vqdmulhq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __c)
11291 {
11292   return __builtin_aarch64_sqdmulh_laneqv4si (__a, __b, __c);
11293 }
11294
11295 __extension__ extern __inline int16x4_t
11296 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11297 vqrdmulh_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __c)
11298 {
11299   return  __builtin_aarch64_sqrdmulh_laneqv4hi (__a, __b, __c);
11300 }
11301
11302 __extension__ extern __inline int32x2_t
11303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11304 vqrdmulh_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __c)
11305 {
11306   return __builtin_aarch64_sqrdmulh_laneqv2si (__a, __b, __c);
11307 }
11308
11309 __extension__ extern __inline int16x8_t
11310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11311 vqrdmulhq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __c)
11312 {
11313   return __builtin_aarch64_sqrdmulh_laneqv8hi (__a, __b, __c);
11314 }
11315
11316 __extension__ extern __inline int32x4_t
11317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11318 vqrdmulhq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __c)
11319 {
11320   return __builtin_aarch64_sqrdmulh_laneqv4si (__a, __b, __c);
11321 }
11322
11323 /* Table intrinsics.  */
11324
11325 __extension__ extern __inline poly8x8_t
11326 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11327 vqtbl1_p8 (poly8x16_t a, uint8x8_t b)
11328 {
11329   poly8x8_t result;
11330   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11331            : "=w"(result)
11332            : "w"(a), "w"(b)
11333            : /* No clobbers */);
11334   return result;
11335 }
11336
11337 __extension__ extern __inline int8x8_t
11338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11339 vqtbl1_s8 (int8x16_t a, uint8x8_t b)
11340 {
11341   int8x8_t result;
11342   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11343            : "=w"(result)
11344            : "w"(a), "w"(b)
11345            : /* No clobbers */);
11346   return result;
11347 }
11348
11349 __extension__ extern __inline uint8x8_t
11350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11351 vqtbl1_u8 (uint8x16_t a, uint8x8_t b)
11352 {
11353   uint8x8_t result;
11354   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11355            : "=w"(result)
11356            : "w"(a), "w"(b)
11357            : /* No clobbers */);
11358   return result;
11359 }
11360
11361 __extension__ extern __inline poly8x16_t
11362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11363 vqtbl1q_p8 (poly8x16_t a, uint8x16_t b)
11364 {
11365   poly8x16_t result;
11366   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11367            : "=w"(result)
11368            : "w"(a), "w"(b)
11369            : /* No clobbers */);
11370   return result;
11371 }
11372
11373 __extension__ extern __inline int8x16_t
11374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11375 vqtbl1q_s8 (int8x16_t a, uint8x16_t b)
11376 {
11377   int8x16_t result;
11378   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11379            : "=w"(result)
11380            : "w"(a), "w"(b)
11381            : /* No clobbers */);
11382   return result;
11383 }
11384
11385 __extension__ extern __inline uint8x16_t
11386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11387 vqtbl1q_u8 (uint8x16_t a, uint8x16_t b)
11388 {
11389   uint8x16_t result;
11390   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11391            : "=w"(result)
11392            : "w"(a), "w"(b)
11393            : /* No clobbers */);
11394   return result;
11395 }
11396
11397 __extension__ extern __inline int8x8_t
11398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11399 vqtbx1_s8 (int8x8_t r, int8x16_t tab, uint8x8_t idx)
11400 {
11401   int8x8_t result = r;
11402   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11403            : "+w"(result)
11404            : "w"(tab), "w"(idx)
11405            : /* No clobbers */);
11406   return result;
11407 }
11408
11409 __extension__ extern __inline uint8x8_t
11410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11411 vqtbx1_u8 (uint8x8_t r, uint8x16_t tab, uint8x8_t idx)
11412 {
11413   uint8x8_t result = r;
11414   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11415            : "+w"(result)
11416            : "w"(tab), "w"(idx)
11417            : /* No clobbers */);
11418   return result;
11419 }
11420
11421 __extension__ extern __inline poly8x8_t
11422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11423 vqtbx1_p8 (poly8x8_t r, poly8x16_t tab, uint8x8_t idx)
11424 {
11425   poly8x8_t result = r;
11426   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11427            : "+w"(result)
11428            : "w"(tab), "w"(idx)
11429            : /* No clobbers */);
11430   return result;
11431 }
11432
11433 __extension__ extern __inline int8x16_t
11434 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11435 vqtbx1q_s8 (int8x16_t r, int8x16_t tab, uint8x16_t idx)
11436 {
11437   int8x16_t result = r;
11438   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11439            : "+w"(result)
11440            : "w"(tab), "w"(idx)
11441            : /* No clobbers */);
11442   return result;
11443 }
11444
11445 __extension__ extern __inline uint8x16_t
11446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11447 vqtbx1q_u8 (uint8x16_t r, uint8x16_t tab, uint8x16_t idx)
11448 {
11449   uint8x16_t result = r;
11450   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11451            : "+w"(result)
11452            : "w"(tab), "w"(idx)
11453            : /* No clobbers */);
11454   return result;
11455 }
11456
11457 __extension__ extern __inline poly8x16_t
11458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11459 vqtbx1q_p8 (poly8x16_t r, poly8x16_t tab, uint8x16_t idx)
11460 {
11461   poly8x16_t result = r;
11462   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11463            : "+w"(result)
11464            : "w"(tab), "w"(idx)
11465            : /* No clobbers */);
11466   return result;
11467 }
11468
11469 /* V7 legacy table intrinsics.  */
11470
11471 __extension__ extern __inline int8x8_t
11472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11473 vtbl1_s8 (int8x8_t tab, int8x8_t idx)
11474 {
11475   int8x8_t result;
11476   int8x16_t temp = vcombine_s8 (tab, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
11477   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11478            : "=w"(result)
11479            : "w"(temp), "w"(idx)
11480            : /* No clobbers */);
11481   return result;
11482 }
11483
11484 __extension__ extern __inline uint8x8_t
11485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11486 vtbl1_u8 (uint8x8_t tab, uint8x8_t idx)
11487 {
11488   uint8x8_t result;
11489   uint8x16_t temp = vcombine_u8 (tab, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
11490   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11491            : "=w"(result)
11492            : "w"(temp), "w"(idx)
11493            : /* No clobbers */);
11494   return result;
11495 }
11496
11497 __extension__ extern __inline poly8x8_t
11498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11499 vtbl1_p8 (poly8x8_t tab, uint8x8_t idx)
11500 {
11501   poly8x8_t result;
11502   poly8x16_t temp = vcombine_p8 (tab, vcreate_p8 (__AARCH64_UINT64_C (0x0)));
11503   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11504            : "=w"(result)
11505            : "w"(temp), "w"(idx)
11506            : /* No clobbers */);
11507   return result;
11508 }
11509
11510 __extension__ extern __inline int8x8_t
11511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11512 vtbl2_s8 (int8x8x2_t tab, int8x8_t idx)
11513 {
11514   int8x8_t result;
11515   int8x16_t temp = vcombine_s8 (tab.val[0], tab.val[1]);
11516   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11517            : "=w"(result)
11518            : "w"(temp), "w"(idx)
11519            : /* No clobbers */);
11520   return result;
11521 }
11522
11523 __extension__ extern __inline uint8x8_t
11524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11525 vtbl2_u8 (uint8x8x2_t tab, uint8x8_t idx)
11526 {
11527   uint8x8_t result;
11528   uint8x16_t temp = vcombine_u8 (tab.val[0], tab.val[1]);
11529   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11530            : "=w"(result)
11531            : "w"(temp), "w"(idx)
11532            : /* No clobbers */);
11533   return result;
11534 }
11535
11536 __extension__ extern __inline poly8x8_t
11537 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11538 vtbl2_p8 (poly8x8x2_t tab, uint8x8_t idx)
11539 {
11540   poly8x8_t result;
11541   poly8x16_t temp = vcombine_p8 (tab.val[0], tab.val[1]);
11542   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11543            : "=w"(result)
11544            : "w"(temp), "w"(idx)
11545            : /* No clobbers */);
11546   return result;
11547 }
11548
11549 __extension__ extern __inline int8x8_t
11550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11551 vtbl3_s8 (int8x8x3_t tab, int8x8_t idx)
11552 {
11553   int8x8_t result;
11554   int8x16x2_t temp;
11555   __builtin_aarch64_simd_oi __o;
11556   temp.val[0] = vcombine_s8 (tab.val[0], tab.val[1]);
11557   temp.val[1] = vcombine_s8 (tab.val[2], vcreate_s8 (__AARCH64_UINT64_C (0x0)));
11558   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11559                                            (int8x16_t) temp.val[0], 0);
11560   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11561                                            (int8x16_t) temp.val[1], 1);
11562   result = __builtin_aarch64_tbl3v8qi (__o, idx);
11563   return result;
11564 }
11565
11566 __extension__ extern __inline uint8x8_t
11567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11568 vtbl3_u8 (uint8x8x3_t tab, uint8x8_t idx)
11569 {
11570   uint8x8_t result;
11571   uint8x16x2_t temp;
11572   __builtin_aarch64_simd_oi __o;
11573   temp.val[0] = vcombine_u8 (tab.val[0], tab.val[1]);
11574   temp.val[1] = vcombine_u8 (tab.val[2], vcreate_u8 (__AARCH64_UINT64_C (0x0)));
11575   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11576                                            (int8x16_t) temp.val[0], 0);
11577   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11578                                            (int8x16_t) temp.val[1], 1);
11579   result = (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11580   return result;
11581 }
11582
11583 __extension__ extern __inline poly8x8_t
11584 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11585 vtbl3_p8 (poly8x8x3_t tab, uint8x8_t idx)
11586 {
11587   poly8x8_t result;
11588   poly8x16x2_t temp;
11589   __builtin_aarch64_simd_oi __o;
11590   temp.val[0] = vcombine_p8 (tab.val[0], tab.val[1]);
11591   temp.val[1] = vcombine_p8 (tab.val[2], vcreate_p8 (__AARCH64_UINT64_C (0x0)));
11592   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11593                                            (int8x16_t) temp.val[0], 0);
11594   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11595                                            (int8x16_t) temp.val[1], 1);
11596   result = (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11597   return result;
11598 }
11599
11600 __extension__ extern __inline int8x8_t
11601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11602 vtbl4_s8 (int8x8x4_t tab, int8x8_t idx)
11603 {
11604   int8x8_t result;
11605   int8x16x2_t temp;
11606   __builtin_aarch64_simd_oi __o;
11607   temp.val[0] = vcombine_s8 (tab.val[0], tab.val[1]);
11608   temp.val[1] = vcombine_s8 (tab.val[2], tab.val[3]);
11609   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11610                                            (int8x16_t) temp.val[0], 0);
11611   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11612                                            (int8x16_t) temp.val[1], 1);
11613   result = __builtin_aarch64_tbl3v8qi (__o, idx);
11614   return result;
11615 }
11616
11617 __extension__ extern __inline uint8x8_t
11618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11619 vtbl4_u8 (uint8x8x4_t tab, uint8x8_t idx)
11620 {
11621   uint8x8_t result;
11622   uint8x16x2_t temp;
11623   __builtin_aarch64_simd_oi __o;
11624   temp.val[0] = vcombine_u8 (tab.val[0], tab.val[1]);
11625   temp.val[1] = vcombine_u8 (tab.val[2], tab.val[3]);
11626   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11627                                            (int8x16_t) temp.val[0], 0);
11628   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11629                                            (int8x16_t) temp.val[1], 1);
11630   result = (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11631   return result;
11632 }
11633
11634 __extension__ extern __inline poly8x8_t
11635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11636 vtbl4_p8 (poly8x8x4_t tab, uint8x8_t idx)
11637 {
11638   poly8x8_t result;
11639   poly8x16x2_t temp;
11640   __builtin_aarch64_simd_oi __o;
11641   temp.val[0] = vcombine_p8 (tab.val[0], tab.val[1]);
11642   temp.val[1] = vcombine_p8 (tab.val[2], tab.val[3]);
11643   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11644                                            (int8x16_t) temp.val[0], 0);
11645   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11646                                            (int8x16_t) temp.val[1], 1);
11647   result = (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11648   return result;
11649 }
11650
11651 __extension__ extern __inline int8x8_t
11652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11653 vtbx2_s8 (int8x8_t r, int8x8x2_t tab, int8x8_t idx)
11654 {
11655   int8x8_t result = r;
11656   int8x16_t temp = vcombine_s8 (tab.val[0], tab.val[1]);
11657   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11658            : "+w"(result)
11659            : "w"(temp), "w"(idx)
11660            : /* No clobbers */);
11661   return result;
11662 }
11663
11664 __extension__ extern __inline uint8x8_t
11665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11666 vtbx2_u8 (uint8x8_t r, uint8x8x2_t tab, uint8x8_t idx)
11667 {
11668   uint8x8_t result = r;
11669   uint8x16_t temp = vcombine_u8 (tab.val[0], tab.val[1]);
11670   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11671            : "+w"(result)
11672            : "w"(temp), "w"(idx)
11673            : /* No clobbers */);
11674   return result;
11675 }
11676
11677 __extension__ extern __inline poly8x8_t
11678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11679 vtbx2_p8 (poly8x8_t r, poly8x8x2_t tab, uint8x8_t idx)
11680 {
11681   poly8x8_t result = r;
11682   poly8x16_t temp = vcombine_p8 (tab.val[0], tab.val[1]);
11683   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11684            : "+w"(result)
11685            : "w"(temp), "w"(idx)
11686            : /* No clobbers */);
11687   return result;
11688 }
11689
11690 /* End of temporary inline asm.  */
11691
11692 /* Start of optimal implementations in approved order.  */
11693
11694 /* vabd.  */
11695
11696 __extension__ extern __inline float32_t
11697 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11698 vabds_f32 (float32_t __a, float32_t __b)
11699 {
11700   return __builtin_aarch64_fabdsf (__a, __b);
11701 }
11702
11703 __extension__ extern __inline float64_t
11704 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11705 vabdd_f64 (float64_t __a, float64_t __b)
11706 {
11707   return __builtin_aarch64_fabddf (__a, __b);
11708 }
11709
11710 __extension__ extern __inline float32x2_t
11711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11712 vabd_f32 (float32x2_t __a, float32x2_t __b)
11713 {
11714   return __builtin_aarch64_fabdv2sf (__a, __b);
11715 }
11716
11717 __extension__ extern __inline float64x1_t
11718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11719 vabd_f64 (float64x1_t __a, float64x1_t __b)
11720 {
11721   return (float64x1_t) {vabdd_f64 (vget_lane_f64 (__a, 0),
11722                                    vget_lane_f64 (__b, 0))};
11723 }
11724
11725 __extension__ extern __inline float32x4_t
11726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11727 vabdq_f32 (float32x4_t __a, float32x4_t __b)
11728 {
11729   return __builtin_aarch64_fabdv4sf (__a, __b);
11730 }
11731
11732 __extension__ extern __inline float64x2_t
11733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11734 vabdq_f64 (float64x2_t __a, float64x2_t __b)
11735 {
11736   return __builtin_aarch64_fabdv2df (__a, __b);
11737 }
11738
11739 /* vabs  */
11740
11741 __extension__ extern __inline float32x2_t
11742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11743 vabs_f32 (float32x2_t __a)
11744 {
11745   return __builtin_aarch64_absv2sf (__a);
11746 }
11747
11748 __extension__ extern __inline float64x1_t
11749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11750 vabs_f64 (float64x1_t __a)
11751 {
11752   return (float64x1_t) {__builtin_fabs (__a[0])};
11753 }
11754
11755 __extension__ extern __inline int8x8_t
11756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11757 vabs_s8 (int8x8_t __a)
11758 {
11759   return __builtin_aarch64_absv8qi (__a);
11760 }
11761
11762 __extension__ extern __inline int16x4_t
11763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11764 vabs_s16 (int16x4_t __a)
11765 {
11766   return __builtin_aarch64_absv4hi (__a);
11767 }
11768
11769 __extension__ extern __inline int32x2_t
11770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11771 vabs_s32 (int32x2_t __a)
11772 {
11773   return __builtin_aarch64_absv2si (__a);
11774 }
11775
11776 __extension__ extern __inline int64x1_t
11777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11778 vabs_s64 (int64x1_t __a)
11779 {
11780   return (int64x1_t) {__builtin_aarch64_absdi (__a[0])};
11781 }
11782
11783 __extension__ extern __inline float32x4_t
11784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11785 vabsq_f32 (float32x4_t __a)
11786 {
11787   return __builtin_aarch64_absv4sf (__a);
11788 }
11789
11790 __extension__ extern __inline float64x2_t
11791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11792 vabsq_f64 (float64x2_t __a)
11793 {
11794   return __builtin_aarch64_absv2df (__a);
11795 }
11796
11797 __extension__ extern __inline int8x16_t
11798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11799 vabsq_s8 (int8x16_t __a)
11800 {
11801   return __builtin_aarch64_absv16qi (__a);
11802 }
11803
11804 __extension__ extern __inline int16x8_t
11805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11806 vabsq_s16 (int16x8_t __a)
11807 {
11808   return __builtin_aarch64_absv8hi (__a);
11809 }
11810
11811 __extension__ extern __inline int32x4_t
11812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11813 vabsq_s32 (int32x4_t __a)
11814 {
11815   return __builtin_aarch64_absv4si (__a);
11816 }
11817
11818 __extension__ extern __inline int64x2_t
11819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11820 vabsq_s64 (int64x2_t __a)
11821 {
11822   return __builtin_aarch64_absv2di (__a);
11823 }
11824
11825 /* vadd */
11826
11827 __extension__ extern __inline int64_t
11828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11829 vaddd_s64 (int64_t __a, int64_t __b)
11830 {
11831   return __a + __b;
11832 }
11833
11834 __extension__ extern __inline uint64_t
11835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11836 vaddd_u64 (uint64_t __a, uint64_t __b)
11837 {
11838   return __a + __b;
11839 }
11840
11841 /* vaddv */
11842
11843 __extension__ extern __inline int8_t
11844 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11845 vaddv_s8 (int8x8_t __a)
11846 {
11847   return __builtin_aarch64_reduc_plus_scal_v8qi (__a);
11848 }
11849
11850 __extension__ extern __inline int16_t
11851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11852 vaddv_s16 (int16x4_t __a)
11853 {
11854   return __builtin_aarch64_reduc_plus_scal_v4hi (__a);
11855 }
11856
11857 __extension__ extern __inline int32_t
11858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11859 vaddv_s32 (int32x2_t __a)
11860 {
11861   return __builtin_aarch64_reduc_plus_scal_v2si (__a);
11862 }
11863
11864 __extension__ extern __inline uint8_t
11865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11866 vaddv_u8 (uint8x8_t __a)
11867 {
11868   return (uint8_t) __builtin_aarch64_reduc_plus_scal_v8qi ((int8x8_t) __a);
11869 }
11870
11871 __extension__ extern __inline uint16_t
11872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11873 vaddv_u16 (uint16x4_t __a)
11874 {
11875   return (uint16_t) __builtin_aarch64_reduc_plus_scal_v4hi ((int16x4_t) __a);
11876 }
11877
11878 __extension__ extern __inline uint32_t
11879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11880 vaddv_u32 (uint32x2_t __a)
11881 {
11882   return (int32_t) __builtin_aarch64_reduc_plus_scal_v2si ((int32x2_t) __a);
11883 }
11884
11885 __extension__ extern __inline int8_t
11886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11887 vaddvq_s8 (int8x16_t __a)
11888 {
11889   return __builtin_aarch64_reduc_plus_scal_v16qi (__a);
11890 }
11891
11892 __extension__ extern __inline int16_t
11893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11894 vaddvq_s16 (int16x8_t __a)
11895 {
11896   return __builtin_aarch64_reduc_plus_scal_v8hi (__a);
11897 }
11898
11899 __extension__ extern __inline int32_t
11900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11901 vaddvq_s32 (int32x4_t __a)
11902 {
11903   return __builtin_aarch64_reduc_plus_scal_v4si (__a);
11904 }
11905
11906 __extension__ extern __inline int64_t
11907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11908 vaddvq_s64 (int64x2_t __a)
11909 {
11910   return __builtin_aarch64_reduc_plus_scal_v2di (__a);
11911 }
11912
11913 __extension__ extern __inline uint8_t
11914 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11915 vaddvq_u8 (uint8x16_t __a)
11916 {
11917   return (uint8_t) __builtin_aarch64_reduc_plus_scal_v16qi ((int8x16_t) __a);
11918 }
11919
11920 __extension__ extern __inline uint16_t
11921 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11922 vaddvq_u16 (uint16x8_t __a)
11923 {
11924   return (uint16_t) __builtin_aarch64_reduc_plus_scal_v8hi ((int16x8_t) __a);
11925 }
11926
11927 __extension__ extern __inline uint32_t
11928 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11929 vaddvq_u32 (uint32x4_t __a)
11930 {
11931   return (uint32_t) __builtin_aarch64_reduc_plus_scal_v4si ((int32x4_t) __a);
11932 }
11933
11934 __extension__ extern __inline uint64_t
11935 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11936 vaddvq_u64 (uint64x2_t __a)
11937 {
11938   return (uint64_t) __builtin_aarch64_reduc_plus_scal_v2di ((int64x2_t) __a);
11939 }
11940
11941 __extension__ extern __inline float32_t
11942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11943 vaddv_f32 (float32x2_t __a)
11944 {
11945   return __builtin_aarch64_reduc_plus_scal_v2sf (__a);
11946 }
11947
11948 __extension__ extern __inline float32_t
11949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11950 vaddvq_f32 (float32x4_t __a)
11951 {
11952   return __builtin_aarch64_reduc_plus_scal_v4sf (__a);
11953 }
11954
11955 __extension__ extern __inline float64_t
11956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11957 vaddvq_f64 (float64x2_t __a)
11958 {
11959   return __builtin_aarch64_reduc_plus_scal_v2df (__a);
11960 }
11961
11962 /* vbsl  */
11963
11964 __extension__ extern __inline float16x4_t
11965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11966 vbsl_f16 (uint16x4_t __a, float16x4_t __b, float16x4_t __c)
11967 {
11968   return __builtin_aarch64_simd_bslv4hf_suss (__a, __b, __c);
11969 }
11970
11971 __extension__ extern __inline float32x2_t
11972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11973 vbsl_f32 (uint32x2_t __a, float32x2_t __b, float32x2_t __c)
11974 {
11975   return __builtin_aarch64_simd_bslv2sf_suss (__a, __b, __c);
11976 }
11977
11978 __extension__ extern __inline float64x1_t
11979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11980 vbsl_f64 (uint64x1_t __a, float64x1_t __b, float64x1_t __c)
11981 {
11982   return (float64x1_t)
11983     { __builtin_aarch64_simd_bsldf_suss (__a[0], __b[0], __c[0]) };
11984 }
11985
11986 __extension__ extern __inline poly8x8_t
11987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11988 vbsl_p8 (uint8x8_t __a, poly8x8_t __b, poly8x8_t __c)
11989 {
11990   return __builtin_aarch64_simd_bslv8qi_pupp (__a, __b, __c);
11991 }
11992
11993 __extension__ extern __inline poly16x4_t
11994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11995 vbsl_p16 (uint16x4_t __a, poly16x4_t __b, poly16x4_t __c)
11996 {
11997   return __builtin_aarch64_simd_bslv4hi_pupp (__a, __b, __c);
11998 }
11999 __extension__ extern __inline poly64x1_t
12000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12001 vbsl_p64 (uint64x1_t __a, poly64x1_t __b, poly64x1_t __c)
12002 {
12003   return (poly64x1_t)
12004       {__builtin_aarch64_simd_bsldi_pupp (__a[0], __b[0], __c[0])};
12005 }
12006
12007 __extension__ extern __inline int8x8_t
12008 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12009 vbsl_s8 (uint8x8_t __a, int8x8_t __b, int8x8_t __c)
12010 {
12011   return __builtin_aarch64_simd_bslv8qi_suss (__a, __b, __c);
12012 }
12013
12014 __extension__ extern __inline int16x4_t
12015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12016 vbsl_s16 (uint16x4_t __a, int16x4_t __b, int16x4_t __c)
12017 {
12018   return __builtin_aarch64_simd_bslv4hi_suss (__a, __b, __c);
12019 }
12020
12021 __extension__ extern __inline int32x2_t
12022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12023 vbsl_s32 (uint32x2_t __a, int32x2_t __b, int32x2_t __c)
12024 {
12025   return __builtin_aarch64_simd_bslv2si_suss (__a, __b, __c);
12026 }
12027
12028 __extension__ extern __inline int64x1_t
12029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12030 vbsl_s64 (uint64x1_t __a, int64x1_t __b, int64x1_t __c)
12031 {
12032   return (int64x1_t)
12033       {__builtin_aarch64_simd_bsldi_suss (__a[0], __b[0], __c[0])};
12034 }
12035
12036 __extension__ extern __inline uint8x8_t
12037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12038 vbsl_u8 (uint8x8_t __a, uint8x8_t __b, uint8x8_t __c)
12039 {
12040   return __builtin_aarch64_simd_bslv8qi_uuuu (__a, __b, __c);
12041 }
12042
12043 __extension__ extern __inline uint16x4_t
12044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12045 vbsl_u16 (uint16x4_t __a, uint16x4_t __b, uint16x4_t __c)
12046 {
12047   return __builtin_aarch64_simd_bslv4hi_uuuu (__a, __b, __c);
12048 }
12049
12050 __extension__ extern __inline uint32x2_t
12051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12052 vbsl_u32 (uint32x2_t __a, uint32x2_t __b, uint32x2_t __c)
12053 {
12054   return __builtin_aarch64_simd_bslv2si_uuuu (__a, __b, __c);
12055 }
12056
12057 __extension__ extern __inline uint64x1_t
12058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12059 vbsl_u64 (uint64x1_t __a, uint64x1_t __b, uint64x1_t __c)
12060 {
12061   return (uint64x1_t)
12062       {__builtin_aarch64_simd_bsldi_uuuu (__a[0], __b[0], __c[0])};
12063 }
12064
12065 __extension__ extern __inline float16x8_t
12066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12067 vbslq_f16 (uint16x8_t __a, float16x8_t __b, float16x8_t __c)
12068 {
12069   return __builtin_aarch64_simd_bslv8hf_suss (__a, __b, __c);
12070 }
12071
12072 __extension__ extern __inline float32x4_t
12073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12074 vbslq_f32 (uint32x4_t __a, float32x4_t __b, float32x4_t __c)
12075 {
12076   return __builtin_aarch64_simd_bslv4sf_suss (__a, __b, __c);
12077 }
12078
12079 __extension__ extern __inline float64x2_t
12080 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12081 vbslq_f64 (uint64x2_t __a, float64x2_t __b, float64x2_t __c)
12082 {
12083   return __builtin_aarch64_simd_bslv2df_suss (__a, __b, __c);
12084 }
12085
12086 __extension__ extern __inline poly8x16_t
12087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12088 vbslq_p8 (uint8x16_t __a, poly8x16_t __b, poly8x16_t __c)
12089 {
12090   return __builtin_aarch64_simd_bslv16qi_pupp (__a, __b, __c);
12091 }
12092
12093 __extension__ extern __inline poly16x8_t
12094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12095 vbslq_p16 (uint16x8_t __a, poly16x8_t __b, poly16x8_t __c)
12096 {
12097   return __builtin_aarch64_simd_bslv8hi_pupp (__a, __b, __c);
12098 }
12099
12100 __extension__ extern __inline int8x16_t
12101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12102 vbslq_s8 (uint8x16_t __a, int8x16_t __b, int8x16_t __c)
12103 {
12104   return __builtin_aarch64_simd_bslv16qi_suss (__a, __b, __c);
12105 }
12106
12107 __extension__ extern __inline int16x8_t
12108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12109 vbslq_s16 (uint16x8_t __a, int16x8_t __b, int16x8_t __c)
12110 {
12111   return __builtin_aarch64_simd_bslv8hi_suss (__a, __b, __c);
12112 }
12113
12114 __extension__ extern __inline poly64x2_t
12115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12116 vbslq_p64 (uint64x2_t __a, poly64x2_t __b, poly64x2_t __c)
12117 {
12118   return __builtin_aarch64_simd_bslv2di_pupp (__a, __b, __c);
12119 }
12120
12121 __extension__ extern __inline int32x4_t
12122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12123 vbslq_s32 (uint32x4_t __a, int32x4_t __b, int32x4_t __c)
12124 {
12125   return __builtin_aarch64_simd_bslv4si_suss (__a, __b, __c);
12126 }
12127
12128 __extension__ extern __inline int64x2_t
12129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12130 vbslq_s64 (uint64x2_t __a, int64x2_t __b, int64x2_t __c)
12131 {
12132   return __builtin_aarch64_simd_bslv2di_suss (__a, __b, __c);
12133 }
12134
12135 __extension__ extern __inline uint8x16_t
12136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12137 vbslq_u8 (uint8x16_t __a, uint8x16_t __b, uint8x16_t __c)
12138 {
12139   return __builtin_aarch64_simd_bslv16qi_uuuu (__a, __b, __c);
12140 }
12141
12142 __extension__ extern __inline uint16x8_t
12143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12144 vbslq_u16 (uint16x8_t __a, uint16x8_t __b, uint16x8_t __c)
12145 {
12146   return __builtin_aarch64_simd_bslv8hi_uuuu (__a, __b, __c);
12147 }
12148
12149 __extension__ extern __inline uint32x4_t
12150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12151 vbslq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
12152 {
12153   return __builtin_aarch64_simd_bslv4si_uuuu (__a, __b, __c);
12154 }
12155
12156 __extension__ extern __inline uint64x2_t
12157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12158 vbslq_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
12159 {
12160   return __builtin_aarch64_simd_bslv2di_uuuu (__a, __b, __c);
12161 }
12162
12163 /* ARMv8.1-A instrinsics.  */
12164 #pragma GCC push_options
12165 #pragma GCC target ("+nothing+rdma")
12166
12167 __extension__ extern __inline int16x4_t
12168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12169 vqrdmlah_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c)
12170 {
12171   return __builtin_aarch64_sqrdmlahv4hi (__a, __b, __c);
12172 }
12173
12174 __extension__ extern __inline int32x2_t
12175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12176 vqrdmlah_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c)
12177 {
12178   return __builtin_aarch64_sqrdmlahv2si (__a, __b, __c);
12179 }
12180
12181 __extension__ extern __inline int16x8_t
12182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12183 vqrdmlahq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
12184 {
12185   return __builtin_aarch64_sqrdmlahv8hi (__a, __b, __c);
12186 }
12187
12188 __extension__ extern __inline int32x4_t
12189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12190 vqrdmlahq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
12191 {
12192   return __builtin_aarch64_sqrdmlahv4si (__a, __b, __c);
12193 }
12194
12195 __extension__ extern __inline int16x4_t
12196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12197 vqrdmlsh_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c)
12198 {
12199   return __builtin_aarch64_sqrdmlshv4hi (__a, __b, __c);
12200 }
12201
12202 __extension__ extern __inline int32x2_t
12203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12204 vqrdmlsh_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c)
12205 {
12206   return __builtin_aarch64_sqrdmlshv2si (__a, __b, __c);
12207 }
12208
12209 __extension__ extern __inline int16x8_t
12210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12211 vqrdmlshq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
12212 {
12213   return __builtin_aarch64_sqrdmlshv8hi (__a, __b, __c);
12214 }
12215
12216 __extension__ extern __inline int32x4_t
12217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12218 vqrdmlshq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
12219 {
12220   return __builtin_aarch64_sqrdmlshv4si (__a, __b, __c);
12221 }
12222
12223 __extension__ extern __inline int16x4_t
12224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12225 vqrdmlah_laneq_s16 (int16x4_t __a, int16x4_t __b, int16x8_t __c, const int __d)
12226 {
12227   return  __builtin_aarch64_sqrdmlah_laneqv4hi (__a, __b, __c, __d);
12228 }
12229
12230 __extension__ extern __inline int32x2_t
12231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12232 vqrdmlah_laneq_s32 (int32x2_t __a, int32x2_t __b, int32x4_t __c, const int __d)
12233 {
12234   return __builtin_aarch64_sqrdmlah_laneqv2si (__a, __b, __c, __d);
12235 }
12236
12237 __extension__ extern __inline int16x8_t
12238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12239 vqrdmlahq_laneq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c, const int __d)
12240 {
12241   return __builtin_aarch64_sqrdmlah_laneqv8hi (__a, __b, __c, __d);
12242 }
12243
12244 __extension__ extern __inline int32x4_t
12245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12246 vqrdmlahq_laneq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c, const int __d)
12247 {
12248   return __builtin_aarch64_sqrdmlah_laneqv4si (__a, __b, __c, __d);
12249 }
12250
12251 __extension__ extern __inline int16x4_t
12252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12253 vqrdmlsh_laneq_s16 (int16x4_t __a, int16x4_t __b, int16x8_t __c, const int __d)
12254 {
12255   return  __builtin_aarch64_sqrdmlsh_laneqv4hi (__a, __b, __c, __d);
12256 }
12257
12258 __extension__ extern __inline int32x2_t
12259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12260 vqrdmlsh_laneq_s32 (int32x2_t __a, int32x2_t __b, int32x4_t __c, const int __d)
12261 {
12262   return __builtin_aarch64_sqrdmlsh_laneqv2si (__a, __b, __c, __d);
12263 }
12264
12265 __extension__ extern __inline int16x8_t
12266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12267 vqrdmlshq_laneq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c, const int __d)
12268 {
12269   return __builtin_aarch64_sqrdmlsh_laneqv8hi (__a, __b, __c, __d);
12270 }
12271
12272 __extension__ extern __inline int32x4_t
12273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12274 vqrdmlshq_laneq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c, const int __d)
12275 {
12276   return __builtin_aarch64_sqrdmlsh_laneqv4si (__a, __b, __c, __d);
12277 }
12278
12279 __extension__ extern __inline int16x4_t
12280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12281 vqrdmlah_lane_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c, const int __d)
12282 {
12283   return  __builtin_aarch64_sqrdmlah_lanev4hi (__a, __b, __c, __d);
12284 }
12285
12286 __extension__ extern __inline int32x2_t
12287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12288 vqrdmlah_lane_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c, const int __d)
12289 {
12290   return __builtin_aarch64_sqrdmlah_lanev2si (__a, __b, __c, __d);
12291 }
12292
12293 __extension__ extern __inline int16x8_t
12294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12295 vqrdmlahq_lane_s16 (int16x8_t __a, int16x8_t __b, int16x4_t __c, const int __d)
12296 {
12297   return __builtin_aarch64_sqrdmlah_lanev8hi (__a, __b, __c, __d);
12298 }
12299
12300 __extension__ extern __inline int32x4_t
12301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12302 vqrdmlahq_lane_s32 (int32x4_t __a, int32x4_t __b, int32x2_t __c, const int __d)
12303 {
12304   return __builtin_aarch64_sqrdmlah_lanev4si (__a, __b, __c, __d);
12305 }
12306
12307 __extension__ extern __inline int16_t
12308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12309 vqrdmlahh_s16 (int16_t __a, int16_t __b, int16_t __c)
12310 {
12311   return (int16_t) __builtin_aarch64_sqrdmlahhi (__a, __b, __c);
12312 }
12313
12314 __extension__ extern __inline int16_t
12315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12316 vqrdmlahh_lane_s16 (int16_t __a, int16_t __b, int16x4_t __c, const int __d)
12317 {
12318   return __builtin_aarch64_sqrdmlah_lanehi (__a, __b, __c, __d);
12319 }
12320
12321 __extension__ extern __inline int16_t
12322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12323 vqrdmlahh_laneq_s16 (int16_t __a, int16_t __b, int16x8_t __c, const int __d)
12324 {
12325   return __builtin_aarch64_sqrdmlah_laneqhi (__a, __b, __c, __d);
12326 }
12327
12328 __extension__ extern __inline int32_t
12329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12330 vqrdmlahs_s32 (int32_t __a, int32_t __b, int32_t __c)
12331 {
12332   return (int32_t) __builtin_aarch64_sqrdmlahsi (__a, __b, __c);
12333 }
12334
12335 __extension__ extern __inline int32_t
12336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12337 vqrdmlahs_lane_s32 (int32_t __a, int32_t __b, int32x2_t __c, const int __d)
12338 {
12339   return __builtin_aarch64_sqrdmlah_lanesi (__a, __b, __c, __d);
12340 }
12341
12342 __extension__ extern __inline int32_t
12343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12344 vqrdmlahs_laneq_s32 (int32_t __a, int32_t __b, int32x4_t __c, const int __d)
12345 {
12346   return __builtin_aarch64_sqrdmlah_laneqsi (__a, __b, __c, __d);
12347 }
12348
12349 __extension__ extern __inline int16x4_t
12350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12351 vqrdmlsh_lane_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c, const int __d)
12352 {
12353   return  __builtin_aarch64_sqrdmlsh_lanev4hi (__a, __b, __c, __d);
12354 }
12355
12356 __extension__ extern __inline int32x2_t
12357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12358 vqrdmlsh_lane_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c, const int __d)
12359 {
12360   return __builtin_aarch64_sqrdmlsh_lanev2si (__a, __b, __c, __d);
12361 }
12362
12363 __extension__ extern __inline int16x8_t
12364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12365 vqrdmlshq_lane_s16 (int16x8_t __a, int16x8_t __b, int16x4_t __c, const int __d)
12366 {
12367   return __builtin_aarch64_sqrdmlsh_lanev8hi (__a, __b, __c, __d);
12368 }
12369
12370 __extension__ extern __inline int32x4_t
12371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12372 vqrdmlshq_lane_s32 (int32x4_t __a, int32x4_t __b, int32x2_t __c, const int __d)
12373 {
12374   return __builtin_aarch64_sqrdmlsh_lanev4si (__a, __b, __c, __d);
12375 }
12376
12377 __extension__ extern __inline int16_t
12378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12379 vqrdmlshh_s16 (int16_t __a, int16_t __b, int16_t __c)
12380 {
12381   return (int16_t) __builtin_aarch64_sqrdmlshhi (__a, __b, __c);
12382 }
12383
12384 __extension__ extern __inline int16_t
12385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12386 vqrdmlshh_lane_s16 (int16_t __a, int16_t __b, int16x4_t __c, const int __d)
12387 {
12388   return __builtin_aarch64_sqrdmlsh_lanehi (__a, __b, __c, __d);
12389 }
12390
12391 __extension__ extern __inline int16_t
12392 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12393 vqrdmlshh_laneq_s16 (int16_t __a, int16_t __b, int16x8_t __c, const int __d)
12394 {
12395   return __builtin_aarch64_sqrdmlsh_laneqhi (__a, __b, __c, __d);
12396 }
12397
12398 __extension__ extern __inline int32_t
12399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12400 vqrdmlshs_s32 (int32_t __a, int32_t __b, int32_t __c)
12401 {
12402   return (int32_t) __builtin_aarch64_sqrdmlshsi (__a, __b, __c);
12403 }
12404
12405 __extension__ extern __inline int32_t
12406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12407 vqrdmlshs_lane_s32 (int32_t __a, int32_t __b, int32x2_t __c, const int __d)
12408 {
12409   return __builtin_aarch64_sqrdmlsh_lanesi (__a, __b, __c, __d);
12410 }
12411
12412 __extension__ extern __inline int32_t
12413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12414 vqrdmlshs_laneq_s32 (int32_t __a, int32_t __b, int32x4_t __c, const int __d)
12415 {
12416   return __builtin_aarch64_sqrdmlsh_laneqsi (__a, __b, __c, __d);
12417 }
12418 #pragma GCC pop_options
12419
12420 #pragma GCC push_options
12421 #pragma GCC target ("+nothing+crypto")
12422 /* vaes  */
12423
12424 __extension__ extern __inline uint8x16_t
12425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12426 vaeseq_u8 (uint8x16_t data, uint8x16_t key)
12427 {
12428   return __builtin_aarch64_crypto_aesev16qi_uuu (data, key);
12429 }
12430
12431 __extension__ extern __inline uint8x16_t
12432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12433 vaesdq_u8 (uint8x16_t data, uint8x16_t key)
12434 {
12435   return __builtin_aarch64_crypto_aesdv16qi_uuu (data, key);
12436 }
12437
12438 __extension__ extern __inline uint8x16_t
12439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12440 vaesmcq_u8 (uint8x16_t data)
12441 {
12442   return __builtin_aarch64_crypto_aesmcv16qi_uu (data);
12443 }
12444
12445 __extension__ extern __inline uint8x16_t
12446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12447 vaesimcq_u8 (uint8x16_t data)
12448 {
12449   return __builtin_aarch64_crypto_aesimcv16qi_uu (data);
12450 }
12451 #pragma GCC pop_options
12452
12453 /* vcage  */
12454
12455 __extension__ extern __inline uint64x1_t
12456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12457 vcage_f64 (float64x1_t __a, float64x1_t __b)
12458 {
12459   return vabs_f64 (__a) >= vabs_f64 (__b);
12460 }
12461
12462 __extension__ extern __inline uint32_t
12463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12464 vcages_f32 (float32_t __a, float32_t __b)
12465 {
12466   return __builtin_fabsf (__a) >= __builtin_fabsf (__b) ? -1 : 0;
12467 }
12468
12469 __extension__ extern __inline uint32x2_t
12470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12471 vcage_f32 (float32x2_t __a, float32x2_t __b)
12472 {
12473   return vabs_f32 (__a) >= vabs_f32 (__b);
12474 }
12475
12476 __extension__ extern __inline uint32x4_t
12477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12478 vcageq_f32 (float32x4_t __a, float32x4_t __b)
12479 {
12480   return vabsq_f32 (__a) >= vabsq_f32 (__b);
12481 }
12482
12483 __extension__ extern __inline uint64_t
12484 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12485 vcaged_f64 (float64_t __a, float64_t __b)
12486 {
12487   return __builtin_fabs (__a) >= __builtin_fabs (__b) ? -1 : 0;
12488 }
12489
12490 __extension__ extern __inline uint64x2_t
12491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12492 vcageq_f64 (float64x2_t __a, float64x2_t __b)
12493 {
12494   return vabsq_f64 (__a) >= vabsq_f64 (__b);
12495 }
12496
12497 /* vcagt  */
12498
12499 __extension__ extern __inline uint32_t
12500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12501 vcagts_f32 (float32_t __a, float32_t __b)
12502 {
12503   return __builtin_fabsf (__a) > __builtin_fabsf (__b) ? -1 : 0;
12504 }
12505
12506 __extension__ extern __inline uint32x2_t
12507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12508 vcagt_f32 (float32x2_t __a, float32x2_t __b)
12509 {
12510   return vabs_f32 (__a) > vabs_f32 (__b);
12511 }
12512
12513 __extension__ extern __inline uint64x1_t
12514 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12515 vcagt_f64 (float64x1_t __a, float64x1_t __b)
12516 {
12517   return vabs_f64 (__a) > vabs_f64 (__b);
12518 }
12519
12520 __extension__ extern __inline uint32x4_t
12521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12522 vcagtq_f32 (float32x4_t __a, float32x4_t __b)
12523 {
12524   return vabsq_f32 (__a) > vabsq_f32 (__b);
12525 }
12526
12527 __extension__ extern __inline uint64_t
12528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12529 vcagtd_f64 (float64_t __a, float64_t __b)
12530 {
12531   return __builtin_fabs (__a) > __builtin_fabs (__b) ? -1 : 0;
12532 }
12533
12534 __extension__ extern __inline uint64x2_t
12535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12536 vcagtq_f64 (float64x2_t __a, float64x2_t __b)
12537 {
12538   return vabsq_f64 (__a) > vabsq_f64 (__b);
12539 }
12540
12541 /* vcale  */
12542
12543 __extension__ extern __inline uint32x2_t
12544 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12545 vcale_f32 (float32x2_t __a, float32x2_t __b)
12546 {
12547   return vabs_f32 (__a) <= vabs_f32 (__b);
12548 }
12549
12550 __extension__ extern __inline uint64x1_t
12551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12552 vcale_f64 (float64x1_t __a, float64x1_t __b)
12553 {
12554   return vabs_f64 (__a) <= vabs_f64 (__b);
12555 }
12556
12557 __extension__ extern __inline uint64_t
12558 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12559 vcaled_f64 (float64_t __a, float64_t __b)
12560 {
12561   return __builtin_fabs (__a) <= __builtin_fabs (__b) ? -1 : 0;
12562 }
12563
12564 __extension__ extern __inline uint32_t
12565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12566 vcales_f32 (float32_t __a, float32_t __b)
12567 {
12568   return __builtin_fabsf (__a) <= __builtin_fabsf (__b) ? -1 : 0;
12569 }
12570
12571 __extension__ extern __inline uint32x4_t
12572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12573 vcaleq_f32 (float32x4_t __a, float32x4_t __b)
12574 {
12575   return vabsq_f32 (__a) <= vabsq_f32 (__b);
12576 }
12577
12578 __extension__ extern __inline uint64x2_t
12579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12580 vcaleq_f64 (float64x2_t __a, float64x2_t __b)
12581 {
12582   return vabsq_f64 (__a) <= vabsq_f64 (__b);
12583 }
12584
12585 /* vcalt  */
12586
12587 __extension__ extern __inline uint32x2_t
12588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12589 vcalt_f32 (float32x2_t __a, float32x2_t __b)
12590 {
12591   return vabs_f32 (__a) < vabs_f32 (__b);
12592 }
12593
12594 __extension__ extern __inline uint64x1_t
12595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12596 vcalt_f64 (float64x1_t __a, float64x1_t __b)
12597 {
12598   return vabs_f64 (__a) < vabs_f64 (__b);
12599 }
12600
12601 __extension__ extern __inline uint64_t
12602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12603 vcaltd_f64 (float64_t __a, float64_t __b)
12604 {
12605   return __builtin_fabs (__a) < __builtin_fabs (__b) ? -1 : 0;
12606 }
12607
12608 __extension__ extern __inline uint32x4_t
12609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12610 vcaltq_f32 (float32x4_t __a, float32x4_t __b)
12611 {
12612   return vabsq_f32 (__a) < vabsq_f32 (__b);
12613 }
12614
12615 __extension__ extern __inline uint64x2_t
12616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12617 vcaltq_f64 (float64x2_t __a, float64x2_t __b)
12618 {
12619   return vabsq_f64 (__a) < vabsq_f64 (__b);
12620 }
12621
12622 __extension__ extern __inline uint32_t
12623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12624 vcalts_f32 (float32_t __a, float32_t __b)
12625 {
12626   return __builtin_fabsf (__a) < __builtin_fabsf (__b) ? -1 : 0;
12627 }
12628
12629 /* vceq - vector.  */
12630
12631 __extension__ extern __inline uint32x2_t
12632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12633 vceq_f32 (float32x2_t __a, float32x2_t __b)
12634 {
12635   return (uint32x2_t) (__a == __b);
12636 }
12637
12638 __extension__ extern __inline uint64x1_t
12639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12640 vceq_f64 (float64x1_t __a, float64x1_t __b)
12641 {
12642   return (uint64x1_t) (__a == __b);
12643 }
12644
12645 __extension__ extern __inline uint8x8_t
12646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12647 vceq_p8 (poly8x8_t __a, poly8x8_t __b)
12648 {
12649   return (uint8x8_t) (__a == __b);
12650 }
12651
12652 __extension__ extern __inline uint64x1_t
12653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12654 vceq_p64 (poly64x1_t __a, poly64x1_t __b)
12655 {
12656   return (uint64x1_t) (__a == __b);
12657 }
12658
12659 __extension__ extern __inline uint8x8_t
12660 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12661 vceq_s8 (int8x8_t __a, int8x8_t __b)
12662 {
12663   return (uint8x8_t) (__a == __b);
12664 }
12665
12666 __extension__ extern __inline uint16x4_t
12667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12668 vceq_s16 (int16x4_t __a, int16x4_t __b)
12669 {
12670   return (uint16x4_t) (__a == __b);
12671 }
12672
12673 __extension__ extern __inline uint32x2_t
12674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12675 vceq_s32 (int32x2_t __a, int32x2_t __b)
12676 {
12677   return (uint32x2_t) (__a == __b);
12678 }
12679
12680 __extension__ extern __inline uint64x1_t
12681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12682 vceq_s64 (int64x1_t __a, int64x1_t __b)
12683 {
12684   return (uint64x1_t) (__a == __b);
12685 }
12686
12687 __extension__ extern __inline uint8x8_t
12688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12689 vceq_u8 (uint8x8_t __a, uint8x8_t __b)
12690 {
12691   return (__a == __b);
12692 }
12693
12694 __extension__ extern __inline uint16x4_t
12695 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12696 vceq_u16 (uint16x4_t __a, uint16x4_t __b)
12697 {
12698   return (__a == __b);
12699 }
12700
12701 __extension__ extern __inline uint32x2_t
12702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12703 vceq_u32 (uint32x2_t __a, uint32x2_t __b)
12704 {
12705   return (__a == __b);
12706 }
12707
12708 __extension__ extern __inline uint64x1_t
12709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12710 vceq_u64 (uint64x1_t __a, uint64x1_t __b)
12711 {
12712   return (__a == __b);
12713 }
12714
12715 __extension__ extern __inline uint32x4_t
12716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12717 vceqq_f32 (float32x4_t __a, float32x4_t __b)
12718 {
12719   return (uint32x4_t) (__a == __b);
12720 }
12721
12722 __extension__ extern __inline uint64x2_t
12723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12724 vceqq_f64 (float64x2_t __a, float64x2_t __b)
12725 {
12726   return (uint64x2_t) (__a == __b);
12727 }
12728
12729 __extension__ extern __inline uint8x16_t
12730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12731 vceqq_p8 (poly8x16_t __a, poly8x16_t __b)
12732 {
12733   return (uint8x16_t) (__a == __b);
12734 }
12735
12736 __extension__ extern __inline uint8x16_t
12737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12738 vceqq_s8 (int8x16_t __a, int8x16_t __b)
12739 {
12740   return (uint8x16_t) (__a == __b);
12741 }
12742
12743 __extension__ extern __inline uint16x8_t
12744 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12745 vceqq_s16 (int16x8_t __a, int16x8_t __b)
12746 {
12747   return (uint16x8_t) (__a == __b);
12748 }
12749
12750 __extension__ extern __inline uint32x4_t
12751 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12752 vceqq_s32 (int32x4_t __a, int32x4_t __b)
12753 {
12754   return (uint32x4_t) (__a == __b);
12755 }
12756
12757 __extension__ extern __inline uint64x2_t
12758 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12759 vceqq_s64 (int64x2_t __a, int64x2_t __b)
12760 {
12761   return (uint64x2_t) (__a == __b);
12762 }
12763
12764 __extension__ extern __inline uint8x16_t
12765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12766 vceqq_u8 (uint8x16_t __a, uint8x16_t __b)
12767 {
12768   return (__a == __b);
12769 }
12770
12771 __extension__ extern __inline uint16x8_t
12772 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12773 vceqq_u16 (uint16x8_t __a, uint16x8_t __b)
12774 {
12775   return (__a == __b);
12776 }
12777
12778 __extension__ extern __inline uint32x4_t
12779 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12780 vceqq_u32 (uint32x4_t __a, uint32x4_t __b)
12781 {
12782   return (__a == __b);
12783 }
12784
12785 __extension__ extern __inline uint64x2_t
12786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12787 vceqq_u64 (uint64x2_t __a, uint64x2_t __b)
12788 {
12789   return (__a == __b);
12790 }
12791
12792 /* vceq - scalar.  */
12793
12794 __extension__ extern __inline uint32_t
12795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12796 vceqs_f32 (float32_t __a, float32_t __b)
12797 {
12798   return __a == __b ? -1 : 0;
12799 }
12800
12801 __extension__ extern __inline uint64_t
12802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12803 vceqd_s64 (int64_t __a, int64_t __b)
12804 {
12805   return __a == __b ? -1ll : 0ll;
12806 }
12807
12808 __extension__ extern __inline uint64_t
12809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12810 vceqd_u64 (uint64_t __a, uint64_t __b)
12811 {
12812   return __a == __b ? -1ll : 0ll;
12813 }
12814
12815 __extension__ extern __inline uint64_t
12816 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12817 vceqd_f64 (float64_t __a, float64_t __b)
12818 {
12819   return __a == __b ? -1ll : 0ll;
12820 }
12821
12822 /* vceqz - vector.  */
12823
12824 __extension__ extern __inline uint32x2_t
12825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12826 vceqz_f32 (float32x2_t __a)
12827 {
12828   return (uint32x2_t) (__a == 0.0f);
12829 }
12830
12831 __extension__ extern __inline uint64x1_t
12832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12833 vceqz_f64 (float64x1_t __a)
12834 {
12835   return (uint64x1_t) (__a == (float64x1_t) {0.0});
12836 }
12837
12838 __extension__ extern __inline uint8x8_t
12839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12840 vceqz_p8 (poly8x8_t __a)
12841 {
12842   return (uint8x8_t) (__a == 0);
12843 }
12844
12845 __extension__ extern __inline uint8x8_t
12846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12847 vceqz_s8 (int8x8_t __a)
12848 {
12849   return (uint8x8_t) (__a == 0);
12850 }
12851
12852 __extension__ extern __inline uint16x4_t
12853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12854 vceqz_s16 (int16x4_t __a)
12855 {
12856   return (uint16x4_t) (__a == 0);
12857 }
12858
12859 __extension__ extern __inline uint32x2_t
12860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12861 vceqz_s32 (int32x2_t __a)
12862 {
12863   return (uint32x2_t) (__a == 0);
12864 }
12865
12866 __extension__ extern __inline uint64x1_t
12867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12868 vceqz_s64 (int64x1_t __a)
12869 {
12870   return (uint64x1_t) (__a == __AARCH64_INT64_C (0));
12871 }
12872
12873 __extension__ extern __inline uint8x8_t
12874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12875 vceqz_u8 (uint8x8_t __a)
12876 {
12877   return (__a == 0);
12878 }
12879
12880 __extension__ extern __inline uint16x4_t
12881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12882 vceqz_u16 (uint16x4_t __a)
12883 {
12884   return (__a == 0);
12885 }
12886
12887 __extension__ extern __inline uint32x2_t
12888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12889 vceqz_u32 (uint32x2_t __a)
12890 {
12891   return (__a == 0);
12892 }
12893
12894 __extension__ extern __inline uint64x1_t
12895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12896 vceqz_u64 (uint64x1_t __a)
12897 {
12898   return (__a == __AARCH64_UINT64_C (0));
12899 }
12900
12901 __extension__ extern __inline uint32x4_t
12902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12903 vceqzq_f32 (float32x4_t __a)
12904 {
12905   return (uint32x4_t) (__a == 0.0f);
12906 }
12907
12908 __extension__ extern __inline uint64x2_t
12909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12910 vceqzq_f64 (float64x2_t __a)
12911 {
12912   return (uint64x2_t) (__a == 0.0f);
12913 }
12914
12915 __extension__ extern __inline uint8x16_t
12916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12917 vceqzq_p8 (poly8x16_t __a)
12918 {
12919   return (uint8x16_t) (__a == 0);
12920 }
12921
12922 __extension__ extern __inline uint8x16_t
12923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12924 vceqzq_s8 (int8x16_t __a)
12925 {
12926   return (uint8x16_t) (__a == 0);
12927 }
12928
12929 __extension__ extern __inline uint16x8_t
12930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12931 vceqzq_s16 (int16x8_t __a)
12932 {
12933   return (uint16x8_t) (__a == 0);
12934 }
12935
12936 __extension__ extern __inline uint32x4_t
12937 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12938 vceqzq_s32 (int32x4_t __a)
12939 {
12940   return (uint32x4_t) (__a == 0);
12941 }
12942
12943 __extension__ extern __inline uint64x2_t
12944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12945 vceqzq_s64 (int64x2_t __a)
12946 {
12947   return (uint64x2_t) (__a == __AARCH64_INT64_C (0));
12948 }
12949
12950 __extension__ extern __inline uint8x16_t
12951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12952 vceqzq_u8 (uint8x16_t __a)
12953 {
12954   return (__a == 0);
12955 }
12956
12957 __extension__ extern __inline uint16x8_t
12958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12959 vceqzq_u16 (uint16x8_t __a)
12960 {
12961   return (__a == 0);
12962 }
12963
12964 __extension__ extern __inline uint32x4_t
12965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12966 vceqzq_u32 (uint32x4_t __a)
12967 {
12968   return (__a == 0);
12969 }
12970
12971 __extension__ extern __inline uint64x2_t
12972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12973 vceqzq_u64 (uint64x2_t __a)
12974 {
12975   return (__a == __AARCH64_UINT64_C (0));
12976 }
12977
12978 /* vceqz - scalar.  */
12979
12980 __extension__ extern __inline uint32_t
12981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12982 vceqzs_f32 (float32_t __a)
12983 {
12984   return __a == 0.0f ? -1 : 0;
12985 }
12986
12987 __extension__ extern __inline uint64_t
12988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12989 vceqzd_s64 (int64_t __a)
12990 {
12991   return __a == 0 ? -1ll : 0ll;
12992 }
12993
12994 __extension__ extern __inline uint64_t
12995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12996 vceqzd_u64 (uint64_t __a)
12997 {
12998   return __a == 0 ? -1ll : 0ll;
12999 }
13000
13001 __extension__ extern __inline uint64_t
13002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13003 vceqzd_f64 (float64_t __a)
13004 {
13005   return __a == 0.0 ? -1ll : 0ll;
13006 }
13007
13008 /* vcge - vector.  */
13009
13010 __extension__ extern __inline uint32x2_t
13011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13012 vcge_f32 (float32x2_t __a, float32x2_t __b)
13013 {
13014   return (uint32x2_t) (__a >= __b);
13015 }
13016
13017 __extension__ extern __inline uint64x1_t
13018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13019 vcge_f64 (float64x1_t __a, float64x1_t __b)
13020 {
13021   return (uint64x1_t) (__a >= __b);
13022 }
13023
13024 __extension__ extern __inline uint8x8_t
13025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13026 vcge_s8 (int8x8_t __a, int8x8_t __b)
13027 {
13028   return (uint8x8_t) (__a >= __b);
13029 }
13030
13031 __extension__ extern __inline uint16x4_t
13032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13033 vcge_s16 (int16x4_t __a, int16x4_t __b)
13034 {
13035   return (uint16x4_t) (__a >= __b);
13036 }
13037
13038 __extension__ extern __inline uint32x2_t
13039 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13040 vcge_s32 (int32x2_t __a, int32x2_t __b)
13041 {
13042   return (uint32x2_t) (__a >= __b);
13043 }
13044
13045 __extension__ extern __inline uint64x1_t
13046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13047 vcge_s64 (int64x1_t __a, int64x1_t __b)
13048 {
13049   return (uint64x1_t) (__a >= __b);
13050 }
13051
13052 __extension__ extern __inline uint8x8_t
13053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13054 vcge_u8 (uint8x8_t __a, uint8x8_t __b)
13055 {
13056   return (__a >= __b);
13057 }
13058
13059 __extension__ extern __inline uint16x4_t
13060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13061 vcge_u16 (uint16x4_t __a, uint16x4_t __b)
13062 {
13063   return (__a >= __b);
13064 }
13065
13066 __extension__ extern __inline uint32x2_t
13067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13068 vcge_u32 (uint32x2_t __a, uint32x2_t __b)
13069 {
13070   return (__a >= __b);
13071 }
13072
13073 __extension__ extern __inline uint64x1_t
13074 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13075 vcge_u64 (uint64x1_t __a, uint64x1_t __b)
13076 {
13077   return (__a >= __b);
13078 }
13079
13080 __extension__ extern __inline uint32x4_t
13081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13082 vcgeq_f32 (float32x4_t __a, float32x4_t __b)
13083 {
13084   return (uint32x4_t) (__a >= __b);
13085 }
13086
13087 __extension__ extern __inline uint64x2_t
13088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13089 vcgeq_f64 (float64x2_t __a, float64x2_t __b)
13090 {
13091   return (uint64x2_t) (__a >= __b);
13092 }
13093
13094 __extension__ extern __inline uint8x16_t
13095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13096 vcgeq_s8 (int8x16_t __a, int8x16_t __b)
13097 {
13098   return (uint8x16_t) (__a >= __b);
13099 }
13100
13101 __extension__ extern __inline uint16x8_t
13102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13103 vcgeq_s16 (int16x8_t __a, int16x8_t __b)
13104 {
13105   return (uint16x8_t) (__a >= __b);
13106 }
13107
13108 __extension__ extern __inline uint32x4_t
13109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13110 vcgeq_s32 (int32x4_t __a, int32x4_t __b)
13111 {
13112   return (uint32x4_t) (__a >= __b);
13113 }
13114
13115 __extension__ extern __inline uint64x2_t
13116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13117 vcgeq_s64 (int64x2_t __a, int64x2_t __b)
13118 {
13119   return (uint64x2_t) (__a >= __b);
13120 }
13121
13122 __extension__ extern __inline uint8x16_t
13123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13124 vcgeq_u8 (uint8x16_t __a, uint8x16_t __b)
13125 {
13126   return (__a >= __b);
13127 }
13128
13129 __extension__ extern __inline uint16x8_t
13130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13131 vcgeq_u16 (uint16x8_t __a, uint16x8_t __b)
13132 {
13133   return (__a >= __b);
13134 }
13135
13136 __extension__ extern __inline uint32x4_t
13137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13138 vcgeq_u32 (uint32x4_t __a, uint32x4_t __b)
13139 {
13140   return (__a >= __b);
13141 }
13142
13143 __extension__ extern __inline uint64x2_t
13144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13145 vcgeq_u64 (uint64x2_t __a, uint64x2_t __b)
13146 {
13147   return (__a >= __b);
13148 }
13149
13150 /* vcge - scalar.  */
13151
13152 __extension__ extern __inline uint32_t
13153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13154 vcges_f32 (float32_t __a, float32_t __b)
13155 {
13156   return __a >= __b ? -1 : 0;
13157 }
13158
13159 __extension__ extern __inline uint64_t
13160 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13161 vcged_s64 (int64_t __a, int64_t __b)
13162 {
13163   return __a >= __b ? -1ll : 0ll;
13164 }
13165
13166 __extension__ extern __inline uint64_t
13167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13168 vcged_u64 (uint64_t __a, uint64_t __b)
13169 {
13170   return __a >= __b ? -1ll : 0ll;
13171 }
13172
13173 __extension__ extern __inline uint64_t
13174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13175 vcged_f64 (float64_t __a, float64_t __b)
13176 {
13177   return __a >= __b ? -1ll : 0ll;
13178 }
13179
13180 /* vcgez - vector.  */
13181
13182 __extension__ extern __inline uint32x2_t
13183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13184 vcgez_f32 (float32x2_t __a)
13185 {
13186   return (uint32x2_t) (__a >= 0.0f);
13187 }
13188
13189 __extension__ extern __inline uint64x1_t
13190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13191 vcgez_f64 (float64x1_t __a)
13192 {
13193   return (uint64x1_t) (__a[0] >= (float64x1_t) {0.0});
13194 }
13195
13196 __extension__ extern __inline uint8x8_t
13197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13198 vcgez_s8 (int8x8_t __a)
13199 {
13200   return (uint8x8_t) (__a >= 0);
13201 }
13202
13203 __extension__ extern __inline uint16x4_t
13204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13205 vcgez_s16 (int16x4_t __a)
13206 {
13207   return (uint16x4_t) (__a >= 0);
13208 }
13209
13210 __extension__ extern __inline uint32x2_t
13211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13212 vcgez_s32 (int32x2_t __a)
13213 {
13214   return (uint32x2_t) (__a >= 0);
13215 }
13216
13217 __extension__ extern __inline uint64x1_t
13218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13219 vcgez_s64 (int64x1_t __a)
13220 {
13221   return (uint64x1_t) (__a >= __AARCH64_INT64_C (0));
13222 }
13223
13224 __extension__ extern __inline uint32x4_t
13225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13226 vcgezq_f32 (float32x4_t __a)
13227 {
13228   return (uint32x4_t) (__a >= 0.0f);
13229 }
13230
13231 __extension__ extern __inline uint64x2_t
13232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13233 vcgezq_f64 (float64x2_t __a)
13234 {
13235   return (uint64x2_t) (__a >= 0.0);
13236 }
13237
13238 __extension__ extern __inline uint8x16_t
13239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13240 vcgezq_s8 (int8x16_t __a)
13241 {
13242   return (uint8x16_t) (__a >= 0);
13243 }
13244
13245 __extension__ extern __inline uint16x8_t
13246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13247 vcgezq_s16 (int16x8_t __a)
13248 {
13249   return (uint16x8_t) (__a >= 0);
13250 }
13251
13252 __extension__ extern __inline uint32x4_t
13253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13254 vcgezq_s32 (int32x4_t __a)
13255 {
13256   return (uint32x4_t) (__a >= 0);
13257 }
13258
13259 __extension__ extern __inline uint64x2_t
13260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13261 vcgezq_s64 (int64x2_t __a)
13262 {
13263   return (uint64x2_t) (__a >= __AARCH64_INT64_C (0));
13264 }
13265
13266 /* vcgez - scalar.  */
13267
13268 __extension__ extern __inline uint32_t
13269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13270 vcgezs_f32 (float32_t __a)
13271 {
13272   return __a >= 0.0f ? -1 : 0;
13273 }
13274
13275 __extension__ extern __inline uint64_t
13276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13277 vcgezd_s64 (int64_t __a)
13278 {
13279   return __a >= 0 ? -1ll : 0ll;
13280 }
13281
13282 __extension__ extern __inline uint64_t
13283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13284 vcgezd_f64 (float64_t __a)
13285 {
13286   return __a >= 0.0 ? -1ll : 0ll;
13287 }
13288
13289 /* vcgt - vector.  */
13290
13291 __extension__ extern __inline uint32x2_t
13292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13293 vcgt_f32 (float32x2_t __a, float32x2_t __b)
13294 {
13295   return (uint32x2_t) (__a > __b);
13296 }
13297
13298 __extension__ extern __inline uint64x1_t
13299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13300 vcgt_f64 (float64x1_t __a, float64x1_t __b)
13301 {
13302   return (uint64x1_t) (__a > __b);
13303 }
13304
13305 __extension__ extern __inline uint8x8_t
13306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13307 vcgt_s8 (int8x8_t __a, int8x8_t __b)
13308 {
13309   return (uint8x8_t) (__a > __b);
13310 }
13311
13312 __extension__ extern __inline uint16x4_t
13313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13314 vcgt_s16 (int16x4_t __a, int16x4_t __b)
13315 {
13316   return (uint16x4_t) (__a > __b);
13317 }
13318
13319 __extension__ extern __inline uint32x2_t
13320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13321 vcgt_s32 (int32x2_t __a, int32x2_t __b)
13322 {
13323   return (uint32x2_t) (__a > __b);
13324 }
13325
13326 __extension__ extern __inline uint64x1_t
13327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13328 vcgt_s64 (int64x1_t __a, int64x1_t __b)
13329 {
13330   return (uint64x1_t) (__a > __b);
13331 }
13332
13333 __extension__ extern __inline uint8x8_t
13334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13335 vcgt_u8 (uint8x8_t __a, uint8x8_t __b)
13336 {
13337   return (__a > __b);
13338 }
13339
13340 __extension__ extern __inline uint16x4_t
13341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13342 vcgt_u16 (uint16x4_t __a, uint16x4_t __b)
13343 {
13344   return (__a > __b);
13345 }
13346
13347 __extension__ extern __inline uint32x2_t
13348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13349 vcgt_u32 (uint32x2_t __a, uint32x2_t __b)
13350 {
13351   return (__a > __b);
13352 }
13353
13354 __extension__ extern __inline uint64x1_t
13355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13356 vcgt_u64 (uint64x1_t __a, uint64x1_t __b)
13357 {
13358   return (__a > __b);
13359 }
13360
13361 __extension__ extern __inline uint32x4_t
13362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13363 vcgtq_f32 (float32x4_t __a, float32x4_t __b)
13364 {
13365   return (uint32x4_t) (__a > __b);
13366 }
13367
13368 __extension__ extern __inline uint64x2_t
13369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13370 vcgtq_f64 (float64x2_t __a, float64x2_t __b)
13371 {
13372   return (uint64x2_t) (__a > __b);
13373 }
13374
13375 __extension__ extern __inline uint8x16_t
13376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13377 vcgtq_s8 (int8x16_t __a, int8x16_t __b)
13378 {
13379   return (uint8x16_t) (__a > __b);
13380 }
13381
13382 __extension__ extern __inline uint16x8_t
13383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13384 vcgtq_s16 (int16x8_t __a, int16x8_t __b)
13385 {
13386   return (uint16x8_t) (__a > __b);
13387 }
13388
13389 __extension__ extern __inline uint32x4_t
13390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13391 vcgtq_s32 (int32x4_t __a, int32x4_t __b)
13392 {
13393   return (uint32x4_t) (__a > __b);
13394 }
13395
13396 __extension__ extern __inline uint64x2_t
13397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13398 vcgtq_s64 (int64x2_t __a, int64x2_t __b)
13399 {
13400   return (uint64x2_t) (__a > __b);
13401 }
13402
13403 __extension__ extern __inline uint8x16_t
13404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13405 vcgtq_u8 (uint8x16_t __a, uint8x16_t __b)
13406 {
13407   return (__a > __b);
13408 }
13409
13410 __extension__ extern __inline uint16x8_t
13411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13412 vcgtq_u16 (uint16x8_t __a, uint16x8_t __b)
13413 {
13414   return (__a > __b);
13415 }
13416
13417 __extension__ extern __inline uint32x4_t
13418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13419 vcgtq_u32 (uint32x4_t __a, uint32x4_t __b)
13420 {
13421   return (__a > __b);
13422 }
13423
13424 __extension__ extern __inline uint64x2_t
13425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13426 vcgtq_u64 (uint64x2_t __a, uint64x2_t __b)
13427 {
13428   return (__a > __b);
13429 }
13430
13431 /* vcgt - scalar.  */
13432
13433 __extension__ extern __inline uint32_t
13434 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13435 vcgts_f32 (float32_t __a, float32_t __b)
13436 {
13437   return __a > __b ? -1 : 0;
13438 }
13439
13440 __extension__ extern __inline uint64_t
13441 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13442 vcgtd_s64 (int64_t __a, int64_t __b)
13443 {
13444   return __a > __b ? -1ll : 0ll;
13445 }
13446
13447 __extension__ extern __inline uint64_t
13448 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13449 vcgtd_u64 (uint64_t __a, uint64_t __b)
13450 {
13451   return __a > __b ? -1ll : 0ll;
13452 }
13453
13454 __extension__ extern __inline uint64_t
13455 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13456 vcgtd_f64 (float64_t __a, float64_t __b)
13457 {
13458   return __a > __b ? -1ll : 0ll;
13459 }
13460
13461 /* vcgtz - vector.  */
13462
13463 __extension__ extern __inline uint32x2_t
13464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13465 vcgtz_f32 (float32x2_t __a)
13466 {
13467   return (uint32x2_t) (__a > 0.0f);
13468 }
13469
13470 __extension__ extern __inline uint64x1_t
13471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13472 vcgtz_f64 (float64x1_t __a)
13473 {
13474   return (uint64x1_t) (__a > (float64x1_t) {0.0});
13475 }
13476
13477 __extension__ extern __inline uint8x8_t
13478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13479 vcgtz_s8 (int8x8_t __a)
13480 {
13481   return (uint8x8_t) (__a > 0);
13482 }
13483
13484 __extension__ extern __inline uint16x4_t
13485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13486 vcgtz_s16 (int16x4_t __a)
13487 {
13488   return (uint16x4_t) (__a > 0);
13489 }
13490
13491 __extension__ extern __inline uint32x2_t
13492 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13493 vcgtz_s32 (int32x2_t __a)
13494 {
13495   return (uint32x2_t) (__a > 0);
13496 }
13497
13498 __extension__ extern __inline uint64x1_t
13499 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13500 vcgtz_s64 (int64x1_t __a)
13501 {
13502   return (uint64x1_t) (__a > __AARCH64_INT64_C (0));
13503 }
13504
13505 __extension__ extern __inline uint32x4_t
13506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13507 vcgtzq_f32 (float32x4_t __a)
13508 {
13509   return (uint32x4_t) (__a > 0.0f);
13510 }
13511
13512 __extension__ extern __inline uint64x2_t
13513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13514 vcgtzq_f64 (float64x2_t __a)
13515 {
13516     return (uint64x2_t) (__a > 0.0);
13517 }
13518
13519 __extension__ extern __inline uint8x16_t
13520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13521 vcgtzq_s8 (int8x16_t __a)
13522 {
13523   return (uint8x16_t) (__a > 0);
13524 }
13525
13526 __extension__ extern __inline uint16x8_t
13527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13528 vcgtzq_s16 (int16x8_t __a)
13529 {
13530   return (uint16x8_t) (__a > 0);
13531 }
13532
13533 __extension__ extern __inline uint32x4_t
13534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13535 vcgtzq_s32 (int32x4_t __a)
13536 {
13537   return (uint32x4_t) (__a > 0);
13538 }
13539
13540 __extension__ extern __inline uint64x2_t
13541 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13542 vcgtzq_s64 (int64x2_t __a)
13543 {
13544   return (uint64x2_t) (__a > __AARCH64_INT64_C (0));
13545 }
13546
13547 /* vcgtz - scalar.  */
13548
13549 __extension__ extern __inline uint32_t
13550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13551 vcgtzs_f32 (float32_t __a)
13552 {
13553   return __a > 0.0f ? -1 : 0;
13554 }
13555
13556 __extension__ extern __inline uint64_t
13557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13558 vcgtzd_s64 (int64_t __a)
13559 {
13560   return __a > 0 ? -1ll : 0ll;
13561 }
13562
13563 __extension__ extern __inline uint64_t
13564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13565 vcgtzd_f64 (float64_t __a)
13566 {
13567   return __a > 0.0 ? -1ll : 0ll;
13568 }
13569
13570 /* vcle - vector.  */
13571
13572 __extension__ extern __inline uint32x2_t
13573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13574 vcle_f32 (float32x2_t __a, float32x2_t __b)
13575 {
13576   return (uint32x2_t) (__a <= __b);
13577 }
13578
13579 __extension__ extern __inline uint64x1_t
13580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13581 vcle_f64 (float64x1_t __a, float64x1_t __b)
13582 {
13583   return (uint64x1_t) (__a <= __b);
13584 }
13585
13586 __extension__ extern __inline uint8x8_t
13587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13588 vcle_s8 (int8x8_t __a, int8x8_t __b)
13589 {
13590   return (uint8x8_t) (__a <= __b);
13591 }
13592
13593 __extension__ extern __inline uint16x4_t
13594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13595 vcle_s16 (int16x4_t __a, int16x4_t __b)
13596 {
13597   return (uint16x4_t) (__a <= __b);
13598 }
13599
13600 __extension__ extern __inline uint32x2_t
13601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13602 vcle_s32 (int32x2_t __a, int32x2_t __b)
13603 {
13604   return (uint32x2_t) (__a <= __b);
13605 }
13606
13607 __extension__ extern __inline uint64x1_t
13608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13609 vcle_s64 (int64x1_t __a, int64x1_t __b)
13610 {
13611   return (uint64x1_t) (__a <= __b);
13612 }
13613
13614 __extension__ extern __inline uint8x8_t
13615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13616 vcle_u8 (uint8x8_t __a, uint8x8_t __b)
13617 {
13618   return (__a <= __b);
13619 }
13620
13621 __extension__ extern __inline uint16x4_t
13622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13623 vcle_u16 (uint16x4_t __a, uint16x4_t __b)
13624 {
13625   return (__a <= __b);
13626 }
13627
13628 __extension__ extern __inline uint32x2_t
13629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13630 vcle_u32 (uint32x2_t __a, uint32x2_t __b)
13631 {
13632   return (__a <= __b);
13633 }
13634
13635 __extension__ extern __inline uint64x1_t
13636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13637 vcle_u64 (uint64x1_t __a, uint64x1_t __b)
13638 {
13639   return (__a <= __b);
13640 }
13641
13642 __extension__ extern __inline uint32x4_t
13643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13644 vcleq_f32 (float32x4_t __a, float32x4_t __b)
13645 {
13646   return (uint32x4_t) (__a <= __b);
13647 }
13648
13649 __extension__ extern __inline uint64x2_t
13650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13651 vcleq_f64 (float64x2_t __a, float64x2_t __b)
13652 {
13653   return (uint64x2_t) (__a <= __b);
13654 }
13655
13656 __extension__ extern __inline uint8x16_t
13657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13658 vcleq_s8 (int8x16_t __a, int8x16_t __b)
13659 {
13660   return (uint8x16_t) (__a <= __b);
13661 }
13662
13663 __extension__ extern __inline uint16x8_t
13664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13665 vcleq_s16 (int16x8_t __a, int16x8_t __b)
13666 {
13667   return (uint16x8_t) (__a <= __b);
13668 }
13669
13670 __extension__ extern __inline uint32x4_t
13671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13672 vcleq_s32 (int32x4_t __a, int32x4_t __b)
13673 {
13674   return (uint32x4_t) (__a <= __b);
13675 }
13676
13677 __extension__ extern __inline uint64x2_t
13678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13679 vcleq_s64 (int64x2_t __a, int64x2_t __b)
13680 {
13681   return (uint64x2_t) (__a <= __b);
13682 }
13683
13684 __extension__ extern __inline uint8x16_t
13685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13686 vcleq_u8 (uint8x16_t __a, uint8x16_t __b)
13687 {
13688   return (__a <= __b);
13689 }
13690
13691 __extension__ extern __inline uint16x8_t
13692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13693 vcleq_u16 (uint16x8_t __a, uint16x8_t __b)
13694 {
13695   return (__a <= __b);
13696 }
13697
13698 __extension__ extern __inline uint32x4_t
13699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13700 vcleq_u32 (uint32x4_t __a, uint32x4_t __b)
13701 {
13702   return (__a <= __b);
13703 }
13704
13705 __extension__ extern __inline uint64x2_t
13706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13707 vcleq_u64 (uint64x2_t __a, uint64x2_t __b)
13708 {
13709   return (__a <= __b);
13710 }
13711
13712 /* vcle - scalar.  */
13713
13714 __extension__ extern __inline uint32_t
13715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13716 vcles_f32 (float32_t __a, float32_t __b)
13717 {
13718   return __a <= __b ? -1 : 0;
13719 }
13720
13721 __extension__ extern __inline uint64_t
13722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13723 vcled_s64 (int64_t __a, int64_t __b)
13724 {
13725   return __a <= __b ? -1ll : 0ll;
13726 }
13727
13728 __extension__ extern __inline uint64_t
13729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13730 vcled_u64 (uint64_t __a, uint64_t __b)
13731 {
13732   return __a <= __b ? -1ll : 0ll;
13733 }
13734
13735 __extension__ extern __inline uint64_t
13736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13737 vcled_f64 (float64_t __a, float64_t __b)
13738 {
13739   return __a <= __b ? -1ll : 0ll;
13740 }
13741
13742 /* vclez - vector.  */
13743
13744 __extension__ extern __inline uint32x2_t
13745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13746 vclez_f32 (float32x2_t __a)
13747 {
13748   return (uint32x2_t) (__a <= 0.0f);
13749 }
13750
13751 __extension__ extern __inline uint64x1_t
13752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13753 vclez_f64 (float64x1_t __a)
13754 {
13755   return (uint64x1_t) (__a <= (float64x1_t) {0.0});
13756 }
13757
13758 __extension__ extern __inline uint8x8_t
13759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13760 vclez_s8 (int8x8_t __a)
13761 {
13762   return (uint8x8_t) (__a <= 0);
13763 }
13764
13765 __extension__ extern __inline uint16x4_t
13766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13767 vclez_s16 (int16x4_t __a)
13768 {
13769   return (uint16x4_t) (__a <= 0);
13770 }
13771
13772 __extension__ extern __inline uint32x2_t
13773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13774 vclez_s32 (int32x2_t __a)
13775 {
13776   return (uint32x2_t) (__a <= 0);
13777 }
13778
13779 __extension__ extern __inline uint64x1_t
13780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13781 vclez_s64 (int64x1_t __a)
13782 {
13783   return (uint64x1_t) (__a <= __AARCH64_INT64_C (0));
13784 }
13785
13786 __extension__ extern __inline uint32x4_t
13787 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13788 vclezq_f32 (float32x4_t __a)
13789 {
13790   return (uint32x4_t) (__a <= 0.0f);
13791 }
13792
13793 __extension__ extern __inline uint64x2_t
13794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13795 vclezq_f64 (float64x2_t __a)
13796 {
13797   return (uint64x2_t) (__a <= 0.0);
13798 }
13799
13800 __extension__ extern __inline uint8x16_t
13801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13802 vclezq_s8 (int8x16_t __a)
13803 {
13804   return (uint8x16_t) (__a <= 0);
13805 }
13806
13807 __extension__ extern __inline uint16x8_t
13808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13809 vclezq_s16 (int16x8_t __a)
13810 {
13811   return (uint16x8_t) (__a <= 0);
13812 }
13813
13814 __extension__ extern __inline uint32x4_t
13815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13816 vclezq_s32 (int32x4_t __a)
13817 {
13818   return (uint32x4_t) (__a <= 0);
13819 }
13820
13821 __extension__ extern __inline uint64x2_t
13822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13823 vclezq_s64 (int64x2_t __a)
13824 {
13825   return (uint64x2_t) (__a <= __AARCH64_INT64_C (0));
13826 }
13827
13828 /* vclez - scalar.  */
13829
13830 __extension__ extern __inline uint32_t
13831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13832 vclezs_f32 (float32_t __a)
13833 {
13834   return __a <= 0.0f ? -1 : 0;
13835 }
13836
13837 __extension__ extern __inline uint64_t
13838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13839 vclezd_s64 (int64_t __a)
13840 {
13841   return __a <= 0 ? -1ll : 0ll;
13842 }
13843
13844 __extension__ extern __inline uint64_t
13845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13846 vclezd_f64 (float64_t __a)
13847 {
13848   return __a <= 0.0 ? -1ll : 0ll;
13849 }
13850
13851 /* vclt - vector.  */
13852
13853 __extension__ extern __inline uint32x2_t
13854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13855 vclt_f32 (float32x2_t __a, float32x2_t __b)
13856 {
13857   return (uint32x2_t) (__a < __b);
13858 }
13859
13860 __extension__ extern __inline uint64x1_t
13861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13862 vclt_f64 (float64x1_t __a, float64x1_t __b)
13863 {
13864   return (uint64x1_t) (__a < __b);
13865 }
13866
13867 __extension__ extern __inline uint8x8_t
13868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13869 vclt_s8 (int8x8_t __a, int8x8_t __b)
13870 {
13871   return (uint8x8_t) (__a < __b);
13872 }
13873
13874 __extension__ extern __inline uint16x4_t
13875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13876 vclt_s16 (int16x4_t __a, int16x4_t __b)
13877 {
13878   return (uint16x4_t) (__a < __b);
13879 }
13880
13881 __extension__ extern __inline uint32x2_t
13882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13883 vclt_s32 (int32x2_t __a, int32x2_t __b)
13884 {
13885   return (uint32x2_t) (__a < __b);
13886 }
13887
13888 __extension__ extern __inline uint64x1_t
13889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13890 vclt_s64 (int64x1_t __a, int64x1_t __b)
13891 {
13892   return (uint64x1_t) (__a < __b);
13893 }
13894
13895 __extension__ extern __inline uint8x8_t
13896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13897 vclt_u8 (uint8x8_t __a, uint8x8_t __b)
13898 {
13899   return (__a < __b);
13900 }
13901
13902 __extension__ extern __inline uint16x4_t
13903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13904 vclt_u16 (uint16x4_t __a, uint16x4_t __b)
13905 {
13906   return (__a < __b);
13907 }
13908
13909 __extension__ extern __inline uint32x2_t
13910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13911 vclt_u32 (uint32x2_t __a, uint32x2_t __b)
13912 {
13913   return (__a < __b);
13914 }
13915
13916 __extension__ extern __inline uint64x1_t
13917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13918 vclt_u64 (uint64x1_t __a, uint64x1_t __b)
13919 {
13920   return (__a < __b);
13921 }
13922
13923 __extension__ extern __inline uint32x4_t
13924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13925 vcltq_f32 (float32x4_t __a, float32x4_t __b)
13926 {
13927   return (uint32x4_t) (__a < __b);
13928 }
13929
13930 __extension__ extern __inline uint64x2_t
13931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13932 vcltq_f64 (float64x2_t __a, float64x2_t __b)
13933 {
13934   return (uint64x2_t) (__a < __b);
13935 }
13936
13937 __extension__ extern __inline uint8x16_t
13938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13939 vcltq_s8 (int8x16_t __a, int8x16_t __b)
13940 {
13941   return (uint8x16_t) (__a < __b);
13942 }
13943
13944 __extension__ extern __inline uint16x8_t
13945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13946 vcltq_s16 (int16x8_t __a, int16x8_t __b)
13947 {
13948   return (uint16x8_t) (__a < __b);
13949 }
13950
13951 __extension__ extern __inline uint32x4_t
13952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13953 vcltq_s32 (int32x4_t __a, int32x4_t __b)
13954 {
13955   return (uint32x4_t) (__a < __b);
13956 }
13957
13958 __extension__ extern __inline uint64x2_t
13959 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13960 vcltq_s64 (int64x2_t __a, int64x2_t __b)
13961 {
13962   return (uint64x2_t) (__a < __b);
13963 }
13964
13965 __extension__ extern __inline uint8x16_t
13966 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13967 vcltq_u8 (uint8x16_t __a, uint8x16_t __b)
13968 {
13969   return (__a < __b);
13970 }
13971
13972 __extension__ extern __inline uint16x8_t
13973 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13974 vcltq_u16 (uint16x8_t __a, uint16x8_t __b)
13975 {
13976   return (__a < __b);
13977 }
13978
13979 __extension__ extern __inline uint32x4_t
13980 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13981 vcltq_u32 (uint32x4_t __a, uint32x4_t __b)
13982 {
13983   return (__a < __b);
13984 }
13985
13986 __extension__ extern __inline uint64x2_t
13987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13988 vcltq_u64 (uint64x2_t __a, uint64x2_t __b)
13989 {
13990   return (__a < __b);
13991 }
13992
13993 /* vclt - scalar.  */
13994
13995 __extension__ extern __inline uint32_t
13996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13997 vclts_f32 (float32_t __a, float32_t __b)
13998 {
13999   return __a < __b ? -1 : 0;
14000 }
14001
14002 __extension__ extern __inline uint64_t
14003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14004 vcltd_s64 (int64_t __a, int64_t __b)
14005 {
14006   return __a < __b ? -1ll : 0ll;
14007 }
14008
14009 __extension__ extern __inline uint64_t
14010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14011 vcltd_u64 (uint64_t __a, uint64_t __b)
14012 {
14013   return __a < __b ? -1ll : 0ll;
14014 }
14015
14016 __extension__ extern __inline uint64_t
14017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14018 vcltd_f64 (float64_t __a, float64_t __b)
14019 {
14020   return __a < __b ? -1ll : 0ll;
14021 }
14022
14023 /* vcltz - vector.  */
14024
14025 __extension__ extern __inline uint32x2_t
14026 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14027 vcltz_f32 (float32x2_t __a)
14028 {
14029   return (uint32x2_t) (__a < 0.0f);
14030 }
14031
14032 __extension__ extern __inline uint64x1_t
14033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14034 vcltz_f64 (float64x1_t __a)
14035 {
14036   return (uint64x1_t) (__a < (float64x1_t) {0.0});
14037 }
14038
14039 __extension__ extern __inline uint8x8_t
14040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14041 vcltz_s8 (int8x8_t __a)
14042 {
14043   return (uint8x8_t) (__a < 0);
14044 }
14045
14046 __extension__ extern __inline uint16x4_t
14047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14048 vcltz_s16 (int16x4_t __a)
14049 {
14050   return (uint16x4_t) (__a < 0);
14051 }
14052
14053 __extension__ extern __inline uint32x2_t
14054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14055 vcltz_s32 (int32x2_t __a)
14056 {
14057   return (uint32x2_t) (__a < 0);
14058 }
14059
14060 __extension__ extern __inline uint64x1_t
14061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14062 vcltz_s64 (int64x1_t __a)
14063 {
14064   return (uint64x1_t) (__a < __AARCH64_INT64_C (0));
14065 }
14066
14067 __extension__ extern __inline uint32x4_t
14068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14069 vcltzq_f32 (float32x4_t __a)
14070 {
14071   return (uint32x4_t) (__a < 0.0f);
14072 }
14073
14074 __extension__ extern __inline uint64x2_t
14075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14076 vcltzq_f64 (float64x2_t __a)
14077 {
14078   return (uint64x2_t) (__a < 0.0);
14079 }
14080
14081 __extension__ extern __inline uint8x16_t
14082 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14083 vcltzq_s8 (int8x16_t __a)
14084 {
14085   return (uint8x16_t) (__a < 0);
14086 }
14087
14088 __extension__ extern __inline uint16x8_t
14089 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14090 vcltzq_s16 (int16x8_t __a)
14091 {
14092   return (uint16x8_t) (__a < 0);
14093 }
14094
14095 __extension__ extern __inline uint32x4_t
14096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14097 vcltzq_s32 (int32x4_t __a)
14098 {
14099   return (uint32x4_t) (__a < 0);
14100 }
14101
14102 __extension__ extern __inline uint64x2_t
14103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14104 vcltzq_s64 (int64x2_t __a)
14105 {
14106   return (uint64x2_t) (__a < __AARCH64_INT64_C (0));
14107 }
14108
14109 /* vcltz - scalar.  */
14110
14111 __extension__ extern __inline uint32_t
14112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14113 vcltzs_f32 (float32_t __a)
14114 {
14115   return __a < 0.0f ? -1 : 0;
14116 }
14117
14118 __extension__ extern __inline uint64_t
14119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14120 vcltzd_s64 (int64_t __a)
14121 {
14122   return __a < 0 ? -1ll : 0ll;
14123 }
14124
14125 __extension__ extern __inline uint64_t
14126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14127 vcltzd_f64 (float64_t __a)
14128 {
14129   return __a < 0.0 ? -1ll : 0ll;
14130 }
14131
14132 /* vcls.  */
14133
14134 __extension__ extern __inline int8x8_t
14135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14136 vcls_s8 (int8x8_t __a)
14137 {
14138   return __builtin_aarch64_clrsbv8qi (__a);
14139 }
14140
14141 __extension__ extern __inline int16x4_t
14142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14143 vcls_s16 (int16x4_t __a)
14144 {
14145   return __builtin_aarch64_clrsbv4hi (__a);
14146 }
14147
14148 __extension__ extern __inline int32x2_t
14149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14150 vcls_s32 (int32x2_t __a)
14151 {
14152   return __builtin_aarch64_clrsbv2si (__a);
14153 }
14154
14155 __extension__ extern __inline int8x16_t
14156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14157 vclsq_s8 (int8x16_t __a)
14158 {
14159   return __builtin_aarch64_clrsbv16qi (__a);
14160 }
14161
14162 __extension__ extern __inline int16x8_t
14163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14164 vclsq_s16 (int16x8_t __a)
14165 {
14166   return __builtin_aarch64_clrsbv8hi (__a);
14167 }
14168
14169 __extension__ extern __inline int32x4_t
14170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14171 vclsq_s32 (int32x4_t __a)
14172 {
14173   return __builtin_aarch64_clrsbv4si (__a);
14174 }
14175
14176 /* vclz.  */
14177
14178 __extension__ extern __inline int8x8_t
14179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14180 vclz_s8 (int8x8_t __a)
14181 {
14182   return __builtin_aarch64_clzv8qi (__a);
14183 }
14184
14185 __extension__ extern __inline int16x4_t
14186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14187 vclz_s16 (int16x4_t __a)
14188 {
14189   return __builtin_aarch64_clzv4hi (__a);
14190 }
14191
14192 __extension__ extern __inline int32x2_t
14193 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14194 vclz_s32 (int32x2_t __a)
14195 {
14196   return __builtin_aarch64_clzv2si (__a);
14197 }
14198
14199 __extension__ extern __inline uint8x8_t
14200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14201 vclz_u8 (uint8x8_t __a)
14202 {
14203   return (uint8x8_t)__builtin_aarch64_clzv8qi ((int8x8_t)__a);
14204 }
14205
14206 __extension__ extern __inline uint16x4_t
14207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14208 vclz_u16 (uint16x4_t __a)
14209 {
14210   return (uint16x4_t)__builtin_aarch64_clzv4hi ((int16x4_t)__a);
14211 }
14212
14213 __extension__ extern __inline uint32x2_t
14214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14215 vclz_u32 (uint32x2_t __a)
14216 {
14217   return (uint32x2_t)__builtin_aarch64_clzv2si ((int32x2_t)__a);
14218 }
14219
14220 __extension__ extern __inline int8x16_t
14221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14222 vclzq_s8 (int8x16_t __a)
14223 {
14224   return __builtin_aarch64_clzv16qi (__a);
14225 }
14226
14227 __extension__ extern __inline int16x8_t
14228 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14229 vclzq_s16 (int16x8_t __a)
14230 {
14231   return __builtin_aarch64_clzv8hi (__a);
14232 }
14233
14234 __extension__ extern __inline int32x4_t
14235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14236 vclzq_s32 (int32x4_t __a)
14237 {
14238   return __builtin_aarch64_clzv4si (__a);
14239 }
14240
14241 __extension__ extern __inline uint8x16_t
14242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14243 vclzq_u8 (uint8x16_t __a)
14244 {
14245   return (uint8x16_t)__builtin_aarch64_clzv16qi ((int8x16_t)__a);
14246 }
14247
14248 __extension__ extern __inline uint16x8_t
14249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14250 vclzq_u16 (uint16x8_t __a)
14251 {
14252   return (uint16x8_t)__builtin_aarch64_clzv8hi ((int16x8_t)__a);
14253 }
14254
14255 __extension__ extern __inline uint32x4_t
14256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14257 vclzq_u32 (uint32x4_t __a)
14258 {
14259   return (uint32x4_t)__builtin_aarch64_clzv4si ((int32x4_t)__a);
14260 }
14261
14262 /* vcnt.  */
14263
14264 __extension__ extern __inline poly8x8_t
14265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14266 vcnt_p8 (poly8x8_t __a)
14267 {
14268   return (poly8x8_t) __builtin_aarch64_popcountv8qi ((int8x8_t) __a);
14269 }
14270
14271 __extension__ extern __inline int8x8_t
14272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14273 vcnt_s8 (int8x8_t __a)
14274 {
14275   return __builtin_aarch64_popcountv8qi (__a);
14276 }
14277
14278 __extension__ extern __inline uint8x8_t
14279 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14280 vcnt_u8 (uint8x8_t __a)
14281 {
14282   return (uint8x8_t) __builtin_aarch64_popcountv8qi ((int8x8_t) __a);
14283 }
14284
14285 __extension__ extern __inline poly8x16_t
14286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14287 vcntq_p8 (poly8x16_t __a)
14288 {
14289   return (poly8x16_t) __builtin_aarch64_popcountv16qi ((int8x16_t) __a);
14290 }
14291
14292 __extension__ extern __inline int8x16_t
14293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14294 vcntq_s8 (int8x16_t __a)
14295 {
14296   return __builtin_aarch64_popcountv16qi (__a);
14297 }
14298
14299 __extension__ extern __inline uint8x16_t
14300 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14301 vcntq_u8 (uint8x16_t __a)
14302 {
14303   return (uint8x16_t) __builtin_aarch64_popcountv16qi ((int8x16_t) __a);
14304 }
14305
14306 /* vcopy_lane.  */
14307
14308 __extension__ extern __inline float32x2_t
14309 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14310 vcopy_lane_f32 (float32x2_t __a, const int __lane1,
14311                 float32x2_t __b, const int __lane2)
14312 {
14313   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14314                                   __a, __lane1);
14315 }
14316
14317 __extension__ extern __inline float64x1_t
14318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14319 vcopy_lane_f64 (float64x1_t __a, const int __lane1,
14320                 float64x1_t __b, const int __lane2)
14321 {
14322   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14323                                   __a, __lane1);
14324 }
14325
14326 __extension__ extern __inline poly8x8_t
14327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14328 vcopy_lane_p8 (poly8x8_t __a, const int __lane1,
14329                poly8x8_t __b, const int __lane2)
14330 {
14331   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14332                                  __a, __lane1);
14333 }
14334
14335 __extension__ extern __inline poly16x4_t
14336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14337 vcopy_lane_p16 (poly16x4_t __a, const int __lane1,
14338                 poly16x4_t __b, const int __lane2)
14339 {
14340   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14341                                   __a, __lane1);
14342 }
14343
14344 __extension__ extern __inline poly64x1_t
14345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14346 vcopy_lane_p64 (poly64x1_t __a, const int __lane1,
14347                 poly64x1_t __b, const int __lane2)
14348 {
14349   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14350                                   __a, __lane1);
14351 }
14352
14353 __extension__ extern __inline int8x8_t
14354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14355 vcopy_lane_s8 (int8x8_t __a, const int __lane1,
14356                int8x8_t __b, const int __lane2)
14357 {
14358   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14359                                  __a, __lane1);
14360 }
14361
14362 __extension__ extern __inline int16x4_t
14363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14364 vcopy_lane_s16 (int16x4_t __a, const int __lane1,
14365                 int16x4_t __b, const int __lane2)
14366 {
14367   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14368                                   __a, __lane1);
14369 }
14370
14371 __extension__ extern __inline int32x2_t
14372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14373 vcopy_lane_s32 (int32x2_t __a, const int __lane1,
14374                 int32x2_t __b, const int __lane2)
14375 {
14376   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14377                                   __a, __lane1);
14378 }
14379
14380 __extension__ extern __inline int64x1_t
14381 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14382 vcopy_lane_s64 (int64x1_t __a, const int __lane1,
14383                 int64x1_t __b, const int __lane2)
14384 {
14385   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14386                                   __a, __lane1);
14387 }
14388
14389 __extension__ extern __inline uint8x8_t
14390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14391 vcopy_lane_u8 (uint8x8_t __a, const int __lane1,
14392                uint8x8_t __b, const int __lane2)
14393 {
14394   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14395                                  __a, __lane1);
14396 }
14397
14398 __extension__ extern __inline uint16x4_t
14399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14400 vcopy_lane_u16 (uint16x4_t __a, const int __lane1,
14401                 uint16x4_t __b, const int __lane2)
14402 {
14403   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14404                                   __a, __lane1);
14405 }
14406
14407 __extension__ extern __inline uint32x2_t
14408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14409 vcopy_lane_u32 (uint32x2_t __a, const int __lane1,
14410                 uint32x2_t __b, const int __lane2)
14411 {
14412   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14413                                   __a, __lane1);
14414 }
14415
14416 __extension__ extern __inline uint64x1_t
14417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14418 vcopy_lane_u64 (uint64x1_t __a, const int __lane1,
14419                 uint64x1_t __b, const int __lane2)
14420 {
14421   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14422                                   __a, __lane1);
14423 }
14424
14425 /* vcopy_laneq.  */
14426
14427 __extension__ extern __inline float32x2_t
14428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14429 vcopy_laneq_f32 (float32x2_t __a, const int __lane1,
14430                  float32x4_t __b, const int __lane2)
14431 {
14432   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14433                                   __a, __lane1);
14434 }
14435
14436 __extension__ extern __inline float64x1_t
14437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14438 vcopy_laneq_f64 (float64x1_t __a, const int __lane1,
14439                  float64x2_t __b, const int __lane2)
14440 {
14441   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14442                                   __a, __lane1);
14443 }
14444
14445 __extension__ extern __inline poly8x8_t
14446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14447 vcopy_laneq_p8 (poly8x8_t __a, const int __lane1,
14448                 poly8x16_t __b, const int __lane2)
14449 {
14450   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14451                                  __a, __lane1);
14452 }
14453
14454 __extension__ extern __inline poly16x4_t
14455 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14456 vcopy_laneq_p16 (poly16x4_t __a, const int __lane1,
14457                  poly16x8_t __b, const int __lane2)
14458 {
14459   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14460                                   __a, __lane1);
14461 }
14462
14463 __extension__ extern __inline poly64x1_t
14464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14465 vcopy_laneq_p64 (poly64x1_t __a, const int __lane1,
14466                  poly64x2_t __b, const int __lane2)
14467 {
14468   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14469                                   __a, __lane1);
14470 }
14471
14472 __extension__ extern __inline int8x8_t
14473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14474 vcopy_laneq_s8 (int8x8_t __a, const int __lane1,
14475                 int8x16_t __b, const int __lane2)
14476 {
14477   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14478                                  __a, __lane1);
14479 }
14480
14481 __extension__ extern __inline int16x4_t
14482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14483 vcopy_laneq_s16 (int16x4_t __a, const int __lane1,
14484                  int16x8_t __b, const int __lane2)
14485 {
14486   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14487                                   __a, __lane1);
14488 }
14489
14490 __extension__ extern __inline int32x2_t
14491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14492 vcopy_laneq_s32 (int32x2_t __a, const int __lane1,
14493                  int32x4_t __b, const int __lane2)
14494 {
14495   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14496                                   __a, __lane1);
14497 }
14498
14499 __extension__ extern __inline int64x1_t
14500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14501 vcopy_laneq_s64 (int64x1_t __a, const int __lane1,
14502                  int64x2_t __b, const int __lane2)
14503 {
14504   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14505                                   __a, __lane1);
14506 }
14507
14508 __extension__ extern __inline uint8x8_t
14509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14510 vcopy_laneq_u8 (uint8x8_t __a, const int __lane1,
14511                 uint8x16_t __b, const int __lane2)
14512 {
14513   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14514                                  __a, __lane1);
14515 }
14516
14517 __extension__ extern __inline uint16x4_t
14518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14519 vcopy_laneq_u16 (uint16x4_t __a, const int __lane1,
14520                  uint16x8_t __b, const int __lane2)
14521 {
14522   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14523                                   __a, __lane1);
14524 }
14525
14526 __extension__ extern __inline uint32x2_t
14527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14528 vcopy_laneq_u32 (uint32x2_t __a, const int __lane1,
14529                  uint32x4_t __b, const int __lane2)
14530 {
14531   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14532                                   __a, __lane1);
14533 }
14534
14535 __extension__ extern __inline uint64x1_t
14536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14537 vcopy_laneq_u64 (uint64x1_t __a, const int __lane1,
14538                  uint64x2_t __b, const int __lane2)
14539 {
14540   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14541                                   __a, __lane1);
14542 }
14543
14544 /* vcopyq_lane.  */
14545
14546 __extension__ extern __inline float32x4_t
14547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14548 vcopyq_lane_f32 (float32x4_t __a, const int __lane1,
14549                  float32x2_t __b, const int __lane2)
14550 {
14551   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14552                                    __a, __lane1);
14553 }
14554
14555 __extension__ extern __inline float64x2_t
14556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14557 vcopyq_lane_f64 (float64x2_t __a, const int __lane1,
14558                  float64x1_t __b, const int __lane2)
14559 {
14560   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14561                                    __a, __lane1);
14562 }
14563
14564 __extension__ extern __inline poly8x16_t
14565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14566 vcopyq_lane_p8 (poly8x16_t __a, const int __lane1,
14567                 poly8x8_t __b, const int __lane2)
14568 {
14569   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14570                                   __a, __lane1);
14571 }
14572
14573 __extension__ extern __inline poly16x8_t
14574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14575 vcopyq_lane_p16 (poly16x8_t __a, const int __lane1,
14576                  poly16x4_t __b, const int __lane2)
14577 {
14578   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14579                                    __a, __lane1);
14580 }
14581
14582 __extension__ extern __inline poly64x2_t
14583 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14584 vcopyq_lane_p64 (poly64x2_t __a, const int __lane1,
14585                  poly64x1_t __b, const int __lane2)
14586 {
14587   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14588                                    __a, __lane1);
14589 }
14590
14591 __extension__ extern __inline int8x16_t
14592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14593 vcopyq_lane_s8 (int8x16_t __a, const int __lane1,
14594                 int8x8_t __b, const int __lane2)
14595 {
14596   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14597                                   __a, __lane1);
14598 }
14599
14600 __extension__ extern __inline int16x8_t
14601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14602 vcopyq_lane_s16 (int16x8_t __a, const int __lane1,
14603                  int16x4_t __b, const int __lane2)
14604 {
14605   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14606                                    __a, __lane1);
14607 }
14608
14609 __extension__ extern __inline int32x4_t
14610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14611 vcopyq_lane_s32 (int32x4_t __a, const int __lane1,
14612                  int32x2_t __b, const int __lane2)
14613 {
14614   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14615                                    __a, __lane1);
14616 }
14617
14618 __extension__ extern __inline int64x2_t
14619 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14620 vcopyq_lane_s64 (int64x2_t __a, const int __lane1,
14621                  int64x1_t __b, const int __lane2)
14622 {
14623   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14624                                    __a, __lane1);
14625 }
14626
14627 __extension__ extern __inline uint8x16_t
14628 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14629 vcopyq_lane_u8 (uint8x16_t __a, const int __lane1,
14630                 uint8x8_t __b, const int __lane2)
14631 {
14632   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14633                                   __a, __lane1);
14634 }
14635
14636 __extension__ extern __inline uint16x8_t
14637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14638 vcopyq_lane_u16 (uint16x8_t __a, const int __lane1,
14639                  uint16x4_t __b, const int __lane2)
14640 {
14641   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14642                                    __a, __lane1);
14643 }
14644
14645 __extension__ extern __inline uint32x4_t
14646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14647 vcopyq_lane_u32 (uint32x4_t __a, const int __lane1,
14648                  uint32x2_t __b, const int __lane2)
14649 {
14650   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14651                                    __a, __lane1);
14652 }
14653
14654 __extension__ extern __inline uint64x2_t
14655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14656 vcopyq_lane_u64 (uint64x2_t __a, const int __lane1,
14657                  uint64x1_t __b, const int __lane2)
14658 {
14659   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14660                                    __a, __lane1);
14661 }
14662
14663 /* vcopyq_laneq.  */
14664
14665 __extension__ extern __inline float32x4_t
14666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14667 vcopyq_laneq_f32 (float32x4_t __a, const int __lane1,
14668                   float32x4_t __b, const int __lane2)
14669 {
14670   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14671                                    __a, __lane1);
14672 }
14673
14674 __extension__ extern __inline float64x2_t
14675 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14676 vcopyq_laneq_f64 (float64x2_t __a, const int __lane1,
14677                   float64x2_t __b, const int __lane2)
14678 {
14679   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14680                                    __a, __lane1);
14681 }
14682
14683 __extension__ extern __inline poly8x16_t
14684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14685 vcopyq_laneq_p8 (poly8x16_t __a, const int __lane1,
14686                  poly8x16_t __b, const int __lane2)
14687 {
14688   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14689                                   __a, __lane1);
14690 }
14691
14692 __extension__ extern __inline poly16x8_t
14693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14694 vcopyq_laneq_p16 (poly16x8_t __a, const int __lane1,
14695                   poly16x8_t __b, const int __lane2)
14696 {
14697   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14698                                    __a, __lane1);
14699 }
14700
14701 __extension__ extern __inline poly64x2_t
14702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14703 vcopyq_laneq_p64 (poly64x2_t __a, const int __lane1,
14704                   poly64x2_t __b, const int __lane2)
14705 {
14706   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14707                                    __a, __lane1);
14708 }
14709
14710 __extension__ extern __inline int8x16_t
14711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14712 vcopyq_laneq_s8 (int8x16_t __a, const int __lane1,
14713                  int8x16_t __b, const int __lane2)
14714 {
14715   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14716                                   __a, __lane1);
14717 }
14718
14719 __extension__ extern __inline int16x8_t
14720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14721 vcopyq_laneq_s16 (int16x8_t __a, const int __lane1,
14722                   int16x8_t __b, const int __lane2)
14723 {
14724   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14725                                    __a, __lane1);
14726 }
14727
14728 __extension__ extern __inline int32x4_t
14729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14730 vcopyq_laneq_s32 (int32x4_t __a, const int __lane1,
14731                   int32x4_t __b, const int __lane2)
14732 {
14733   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14734                                    __a, __lane1);
14735 }
14736
14737 __extension__ extern __inline int64x2_t
14738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14739 vcopyq_laneq_s64 (int64x2_t __a, const int __lane1,
14740                   int64x2_t __b, const int __lane2)
14741 {
14742   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14743                                    __a, __lane1);
14744 }
14745
14746 __extension__ extern __inline uint8x16_t
14747 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14748 vcopyq_laneq_u8 (uint8x16_t __a, const int __lane1,
14749                  uint8x16_t __b, const int __lane2)
14750 {
14751   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14752                                   __a, __lane1);
14753 }
14754
14755 __extension__ extern __inline uint16x8_t
14756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14757 vcopyq_laneq_u16 (uint16x8_t __a, const int __lane1,
14758                   uint16x8_t __b, const int __lane2)
14759 {
14760   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14761                                    __a, __lane1);
14762 }
14763
14764 __extension__ extern __inline uint32x4_t
14765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14766 vcopyq_laneq_u32 (uint32x4_t __a, const int __lane1,
14767                   uint32x4_t __b, const int __lane2)
14768 {
14769   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14770                                    __a, __lane1);
14771 }
14772
14773 __extension__ extern __inline uint64x2_t
14774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14775 vcopyq_laneq_u64 (uint64x2_t __a, const int __lane1,
14776                   uint64x2_t __b, const int __lane2)
14777 {
14778   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14779                                    __a, __lane1);
14780 }
14781
14782 /* vcvt (double -> float).  */
14783
14784 __extension__ extern __inline float16x4_t
14785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14786 vcvt_f16_f32 (float32x4_t __a)
14787 {
14788   return __builtin_aarch64_float_truncate_lo_v4hf (__a);
14789 }
14790
14791 __extension__ extern __inline float16x8_t
14792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14793 vcvt_high_f16_f32 (float16x4_t __a, float32x4_t __b)
14794 {
14795   return __builtin_aarch64_float_truncate_hi_v8hf (__a, __b);
14796 }
14797
14798 __extension__ extern __inline float32x2_t
14799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14800 vcvt_f32_f64 (float64x2_t __a)
14801 {
14802   return __builtin_aarch64_float_truncate_lo_v2sf (__a);
14803 }
14804
14805 __extension__ extern __inline float32x4_t
14806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14807 vcvt_high_f32_f64 (float32x2_t __a, float64x2_t __b)
14808 {
14809   return __builtin_aarch64_float_truncate_hi_v4sf (__a, __b);
14810 }
14811
14812 /* vcvt (float -> double).  */
14813
14814 __extension__ extern __inline float32x4_t
14815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14816 vcvt_f32_f16 (float16x4_t __a)
14817 {
14818   return __builtin_aarch64_float_extend_lo_v4sf (__a);
14819 }
14820
14821 __extension__ extern __inline float64x2_t
14822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14823 vcvt_f64_f32 (float32x2_t __a)
14824 {
14825
14826   return __builtin_aarch64_float_extend_lo_v2df (__a);
14827 }
14828
14829 __extension__ extern __inline float32x4_t
14830 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14831 vcvt_high_f32_f16 (float16x8_t __a)
14832 {
14833   return __builtin_aarch64_vec_unpacks_hi_v8hf (__a);
14834 }
14835
14836 __extension__ extern __inline float64x2_t
14837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14838 vcvt_high_f64_f32 (float32x4_t __a)
14839 {
14840   return __builtin_aarch64_vec_unpacks_hi_v4sf (__a);
14841 }
14842
14843 /* vcvt (<u>fixed-point -> float).  */
14844
14845 __extension__ extern __inline float64_t
14846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14847 vcvtd_n_f64_s64 (int64_t __a, const int __b)
14848 {
14849   return __builtin_aarch64_scvtfdi (__a, __b);
14850 }
14851
14852 __extension__ extern __inline float64_t
14853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14854 vcvtd_n_f64_u64 (uint64_t __a, const int __b)
14855 {
14856   return __builtin_aarch64_ucvtfdi_sus (__a, __b);
14857 }
14858
14859 __extension__ extern __inline float32_t
14860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14861 vcvts_n_f32_s32 (int32_t __a, const int __b)
14862 {
14863   return __builtin_aarch64_scvtfsi (__a, __b);
14864 }
14865
14866 __extension__ extern __inline float32_t
14867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14868 vcvts_n_f32_u32 (uint32_t __a, const int __b)
14869 {
14870   return __builtin_aarch64_ucvtfsi_sus (__a, __b);
14871 }
14872
14873 __extension__ extern __inline float32x2_t
14874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14875 vcvt_n_f32_s32 (int32x2_t __a, const int __b)
14876 {
14877   return __builtin_aarch64_scvtfv2si (__a, __b);
14878 }
14879
14880 __extension__ extern __inline float32x2_t
14881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14882 vcvt_n_f32_u32 (uint32x2_t __a, const int __b)
14883 {
14884   return __builtin_aarch64_ucvtfv2si_sus (__a, __b);
14885 }
14886
14887 __extension__ extern __inline float64x1_t
14888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14889 vcvt_n_f64_s64 (int64x1_t __a, const int __b)
14890 {
14891   return (float64x1_t)
14892     { __builtin_aarch64_scvtfdi (vget_lane_s64 (__a, 0), __b) };
14893 }
14894
14895 __extension__ extern __inline float64x1_t
14896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14897 vcvt_n_f64_u64 (uint64x1_t __a, const int __b)
14898 {
14899   return (float64x1_t)
14900     { __builtin_aarch64_ucvtfdi_sus (vget_lane_u64 (__a, 0), __b) };
14901 }
14902
14903 __extension__ extern __inline float32x4_t
14904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14905 vcvtq_n_f32_s32 (int32x4_t __a, const int __b)
14906 {
14907   return __builtin_aarch64_scvtfv4si (__a, __b);
14908 }
14909
14910 __extension__ extern __inline float32x4_t
14911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14912 vcvtq_n_f32_u32 (uint32x4_t __a, const int __b)
14913 {
14914   return __builtin_aarch64_ucvtfv4si_sus (__a, __b);
14915 }
14916
14917 __extension__ extern __inline float64x2_t
14918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14919 vcvtq_n_f64_s64 (int64x2_t __a, const int __b)
14920 {
14921   return __builtin_aarch64_scvtfv2di (__a, __b);
14922 }
14923
14924 __extension__ extern __inline float64x2_t
14925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14926 vcvtq_n_f64_u64 (uint64x2_t __a, const int __b)
14927 {
14928   return __builtin_aarch64_ucvtfv2di_sus (__a, __b);
14929 }
14930
14931 /* vcvt (float -> <u>fixed-point).  */
14932
14933 __extension__ extern __inline int64_t
14934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14935 vcvtd_n_s64_f64 (float64_t __a, const int __b)
14936 {
14937   return __builtin_aarch64_fcvtzsdf (__a, __b);
14938 }
14939
14940 __extension__ extern __inline uint64_t
14941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14942 vcvtd_n_u64_f64 (float64_t __a, const int __b)
14943 {
14944   return __builtin_aarch64_fcvtzudf_uss (__a, __b);
14945 }
14946
14947 __extension__ extern __inline int32_t
14948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14949 vcvts_n_s32_f32 (float32_t __a, const int __b)
14950 {
14951   return __builtin_aarch64_fcvtzssf (__a, __b);
14952 }
14953
14954 __extension__ extern __inline uint32_t
14955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14956 vcvts_n_u32_f32 (float32_t __a, const int __b)
14957 {
14958   return __builtin_aarch64_fcvtzusf_uss (__a, __b);
14959 }
14960
14961 __extension__ extern __inline int32x2_t
14962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14963 vcvt_n_s32_f32 (float32x2_t __a, const int __b)
14964 {
14965   return __builtin_aarch64_fcvtzsv2sf (__a, __b);
14966 }
14967
14968 __extension__ extern __inline uint32x2_t
14969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14970 vcvt_n_u32_f32 (float32x2_t __a, const int __b)
14971 {
14972   return __builtin_aarch64_fcvtzuv2sf_uss (__a, __b);
14973 }
14974
14975 __extension__ extern __inline int64x1_t
14976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14977 vcvt_n_s64_f64 (float64x1_t __a, const int __b)
14978 {
14979   return (int64x1_t)
14980     { __builtin_aarch64_fcvtzsdf (vget_lane_f64 (__a, 0), __b) };
14981 }
14982
14983 __extension__ extern __inline uint64x1_t
14984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14985 vcvt_n_u64_f64 (float64x1_t __a, const int __b)
14986 {
14987   return (uint64x1_t)
14988     { __builtin_aarch64_fcvtzudf_uss (vget_lane_f64 (__a, 0), __b) };
14989 }
14990
14991 __extension__ extern __inline int32x4_t
14992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14993 vcvtq_n_s32_f32 (float32x4_t __a, const int __b)
14994 {
14995   return __builtin_aarch64_fcvtzsv4sf (__a, __b);
14996 }
14997
14998 __extension__ extern __inline uint32x4_t
14999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15000 vcvtq_n_u32_f32 (float32x4_t __a, const int __b)
15001 {
15002   return __builtin_aarch64_fcvtzuv4sf_uss (__a, __b);
15003 }
15004
15005 __extension__ extern __inline int64x2_t
15006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15007 vcvtq_n_s64_f64 (float64x2_t __a, const int __b)
15008 {
15009   return __builtin_aarch64_fcvtzsv2df (__a, __b);
15010 }
15011
15012 __extension__ extern __inline uint64x2_t
15013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15014 vcvtq_n_u64_f64 (float64x2_t __a, const int __b)
15015 {
15016   return __builtin_aarch64_fcvtzuv2df_uss (__a, __b);
15017 }
15018
15019 /* vcvt  (<u>int -> float)  */
15020
15021 __extension__ extern __inline float64_t
15022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15023 vcvtd_f64_s64 (int64_t __a)
15024 {
15025   return (float64_t) __a;
15026 }
15027
15028 __extension__ extern __inline float64_t
15029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15030 vcvtd_f64_u64 (uint64_t __a)
15031 {
15032   return (float64_t) __a;
15033 }
15034
15035 __extension__ extern __inline float32_t
15036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15037 vcvts_f32_s32 (int32_t __a)
15038 {
15039   return (float32_t) __a;
15040 }
15041
15042 __extension__ extern __inline float32_t
15043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15044 vcvts_f32_u32 (uint32_t __a)
15045 {
15046   return (float32_t) __a;
15047 }
15048
15049 __extension__ extern __inline float32x2_t
15050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15051 vcvt_f32_s32 (int32x2_t __a)
15052 {
15053   return __builtin_aarch64_floatv2siv2sf (__a);
15054 }
15055
15056 __extension__ extern __inline float32x2_t
15057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15058 vcvt_f32_u32 (uint32x2_t __a)
15059 {
15060   return __builtin_aarch64_floatunsv2siv2sf ((int32x2_t) __a);
15061 }
15062
15063 __extension__ extern __inline float64x1_t
15064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15065 vcvt_f64_s64 (int64x1_t __a)
15066 {
15067   return (float64x1_t) { vget_lane_s64 (__a, 0) };
15068 }
15069
15070 __extension__ extern __inline float64x1_t
15071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15072 vcvt_f64_u64 (uint64x1_t __a)
15073 {
15074   return (float64x1_t) { vget_lane_u64 (__a, 0) };
15075 }
15076
15077 __extension__ extern __inline float32x4_t
15078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15079 vcvtq_f32_s32 (int32x4_t __a)
15080 {
15081   return __builtin_aarch64_floatv4siv4sf (__a);
15082 }
15083
15084 __extension__ extern __inline float32x4_t
15085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15086 vcvtq_f32_u32 (uint32x4_t __a)
15087 {
15088   return __builtin_aarch64_floatunsv4siv4sf ((int32x4_t) __a);
15089 }
15090
15091 __extension__ extern __inline float64x2_t
15092 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15093 vcvtq_f64_s64 (int64x2_t __a)
15094 {
15095   return __builtin_aarch64_floatv2div2df (__a);
15096 }
15097
15098 __extension__ extern __inline float64x2_t
15099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15100 vcvtq_f64_u64 (uint64x2_t __a)
15101 {
15102   return __builtin_aarch64_floatunsv2div2df ((int64x2_t) __a);
15103 }
15104
15105 /* vcvt (float -> <u>int)  */
15106
15107 __extension__ extern __inline int64_t
15108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15109 vcvtd_s64_f64 (float64_t __a)
15110 {
15111   return (int64_t) __a;
15112 }
15113
15114 __extension__ extern __inline uint64_t
15115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15116 vcvtd_u64_f64 (float64_t __a)
15117 {
15118   return (uint64_t) __a;
15119 }
15120
15121 __extension__ extern __inline int32_t
15122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15123 vcvts_s32_f32 (float32_t __a)
15124 {
15125   return (int32_t) __a;
15126 }
15127
15128 __extension__ extern __inline uint32_t
15129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15130 vcvts_u32_f32 (float32_t __a)
15131 {
15132   return (uint32_t) __a;
15133 }
15134
15135 __extension__ extern __inline int32x2_t
15136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15137 vcvt_s32_f32 (float32x2_t __a)
15138 {
15139   return __builtin_aarch64_lbtruncv2sfv2si (__a);
15140 }
15141
15142 __extension__ extern __inline uint32x2_t
15143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15144 vcvt_u32_f32 (float32x2_t __a)
15145 {
15146   return __builtin_aarch64_lbtruncuv2sfv2si_us (__a);
15147 }
15148
15149 __extension__ extern __inline int32x4_t
15150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15151 vcvtq_s32_f32 (float32x4_t __a)
15152 {
15153   return __builtin_aarch64_lbtruncv4sfv4si (__a);
15154 }
15155
15156 __extension__ extern __inline uint32x4_t
15157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15158 vcvtq_u32_f32 (float32x4_t __a)
15159 {
15160   return __builtin_aarch64_lbtruncuv4sfv4si_us (__a);
15161 }
15162
15163 __extension__ extern __inline int64x1_t
15164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15165 vcvt_s64_f64 (float64x1_t __a)
15166 {
15167   return (int64x1_t) {vcvtd_s64_f64 (__a[0])};
15168 }
15169
15170 __extension__ extern __inline uint64x1_t
15171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15172 vcvt_u64_f64 (float64x1_t __a)
15173 {
15174   return (uint64x1_t) {vcvtd_u64_f64 (__a[0])};
15175 }
15176
15177 __extension__ extern __inline int64x2_t
15178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15179 vcvtq_s64_f64 (float64x2_t __a)
15180 {
15181   return __builtin_aarch64_lbtruncv2dfv2di (__a);
15182 }
15183
15184 __extension__ extern __inline uint64x2_t
15185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15186 vcvtq_u64_f64 (float64x2_t __a)
15187 {
15188   return __builtin_aarch64_lbtruncuv2dfv2di_us (__a);
15189 }
15190
15191 /* vcvta  */
15192
15193 __extension__ extern __inline int64_t
15194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15195 vcvtad_s64_f64 (float64_t __a)
15196 {
15197   return __builtin_aarch64_lrounddfdi (__a);
15198 }
15199
15200 __extension__ extern __inline uint64_t
15201 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15202 vcvtad_u64_f64 (float64_t __a)
15203 {
15204   return __builtin_aarch64_lroundudfdi_us (__a);
15205 }
15206
15207 __extension__ extern __inline int32_t
15208 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15209 vcvtas_s32_f32 (float32_t __a)
15210 {
15211   return __builtin_aarch64_lroundsfsi (__a);
15212 }
15213
15214 __extension__ extern __inline uint32_t
15215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15216 vcvtas_u32_f32 (float32_t __a)
15217 {
15218   return __builtin_aarch64_lroundusfsi_us (__a);
15219 }
15220
15221 __extension__ extern __inline int32x2_t
15222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15223 vcvta_s32_f32 (float32x2_t __a)
15224 {
15225   return __builtin_aarch64_lroundv2sfv2si (__a);
15226 }
15227
15228 __extension__ extern __inline uint32x2_t
15229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15230 vcvta_u32_f32 (float32x2_t __a)
15231 {
15232   return __builtin_aarch64_lrounduv2sfv2si_us (__a);
15233 }
15234
15235 __extension__ extern __inline int32x4_t
15236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15237 vcvtaq_s32_f32 (float32x4_t __a)
15238 {
15239   return __builtin_aarch64_lroundv4sfv4si (__a);
15240 }
15241
15242 __extension__ extern __inline uint32x4_t
15243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15244 vcvtaq_u32_f32 (float32x4_t __a)
15245 {
15246   return __builtin_aarch64_lrounduv4sfv4si_us (__a);
15247 }
15248
15249 __extension__ extern __inline int64x1_t
15250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15251 vcvta_s64_f64 (float64x1_t __a)
15252 {
15253   return (int64x1_t) {vcvtad_s64_f64 (__a[0])};
15254 }
15255
15256 __extension__ extern __inline uint64x1_t
15257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15258 vcvta_u64_f64 (float64x1_t __a)
15259 {
15260   return (uint64x1_t) {vcvtad_u64_f64 (__a[0])};
15261 }
15262
15263 __extension__ extern __inline int64x2_t
15264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15265 vcvtaq_s64_f64 (float64x2_t __a)
15266 {
15267   return __builtin_aarch64_lroundv2dfv2di (__a);
15268 }
15269
15270 __extension__ extern __inline uint64x2_t
15271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15272 vcvtaq_u64_f64 (float64x2_t __a)
15273 {
15274   return __builtin_aarch64_lrounduv2dfv2di_us (__a);
15275 }
15276
15277 /* vcvtm  */
15278
15279 __extension__ extern __inline int64_t
15280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15281 vcvtmd_s64_f64 (float64_t __a)
15282 {
15283   return __builtin_llfloor (__a);
15284 }
15285
15286 __extension__ extern __inline uint64_t
15287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15288 vcvtmd_u64_f64 (float64_t __a)
15289 {
15290   return __builtin_aarch64_lfloorudfdi_us (__a);
15291 }
15292
15293 __extension__ extern __inline int32_t
15294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15295 vcvtms_s32_f32 (float32_t __a)
15296 {
15297   return __builtin_ifloorf (__a);
15298 }
15299
15300 __extension__ extern __inline uint32_t
15301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15302 vcvtms_u32_f32 (float32_t __a)
15303 {
15304   return __builtin_aarch64_lfloorusfsi_us (__a);
15305 }
15306
15307 __extension__ extern __inline int32x2_t
15308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15309 vcvtm_s32_f32 (float32x2_t __a)
15310 {
15311   return __builtin_aarch64_lfloorv2sfv2si (__a);
15312 }
15313
15314 __extension__ extern __inline uint32x2_t
15315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15316 vcvtm_u32_f32 (float32x2_t __a)
15317 {
15318   return __builtin_aarch64_lflooruv2sfv2si_us (__a);
15319 }
15320
15321 __extension__ extern __inline int32x4_t
15322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15323 vcvtmq_s32_f32 (float32x4_t __a)
15324 {
15325   return __builtin_aarch64_lfloorv4sfv4si (__a);
15326 }
15327
15328 __extension__ extern __inline uint32x4_t
15329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15330 vcvtmq_u32_f32 (float32x4_t __a)
15331 {
15332   return __builtin_aarch64_lflooruv4sfv4si_us (__a);
15333 }
15334
15335 __extension__ extern __inline int64x1_t
15336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15337 vcvtm_s64_f64 (float64x1_t __a)
15338 {
15339   return (int64x1_t) {vcvtmd_s64_f64 (__a[0])};
15340 }
15341
15342 __extension__ extern __inline uint64x1_t
15343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15344 vcvtm_u64_f64 (float64x1_t __a)
15345 {
15346   return (uint64x1_t) {vcvtmd_u64_f64 (__a[0])};
15347 }
15348
15349 __extension__ extern __inline int64x2_t
15350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15351 vcvtmq_s64_f64 (float64x2_t __a)
15352 {
15353   return __builtin_aarch64_lfloorv2dfv2di (__a);
15354 }
15355
15356 __extension__ extern __inline uint64x2_t
15357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15358 vcvtmq_u64_f64 (float64x2_t __a)
15359 {
15360   return __builtin_aarch64_lflooruv2dfv2di_us (__a);
15361 }
15362
15363 /* vcvtn  */
15364
15365 __extension__ extern __inline int64_t
15366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15367 vcvtnd_s64_f64 (float64_t __a)
15368 {
15369   return __builtin_aarch64_lfrintndfdi (__a);
15370 }
15371
15372 __extension__ extern __inline uint64_t
15373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15374 vcvtnd_u64_f64 (float64_t __a)
15375 {
15376   return __builtin_aarch64_lfrintnudfdi_us (__a);
15377 }
15378
15379 __extension__ extern __inline int32_t
15380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15381 vcvtns_s32_f32 (float32_t __a)
15382 {
15383   return __builtin_aarch64_lfrintnsfsi (__a);
15384 }
15385
15386 __extension__ extern __inline uint32_t
15387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15388 vcvtns_u32_f32 (float32_t __a)
15389 {
15390   return __builtin_aarch64_lfrintnusfsi_us (__a);
15391 }
15392
15393 __extension__ extern __inline int32x2_t
15394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15395 vcvtn_s32_f32 (float32x2_t __a)
15396 {
15397   return __builtin_aarch64_lfrintnv2sfv2si (__a);
15398 }
15399
15400 __extension__ extern __inline uint32x2_t
15401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15402 vcvtn_u32_f32 (float32x2_t __a)
15403 {
15404   return __builtin_aarch64_lfrintnuv2sfv2si_us (__a);
15405 }
15406
15407 __extension__ extern __inline int32x4_t
15408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15409 vcvtnq_s32_f32 (float32x4_t __a)
15410 {
15411   return __builtin_aarch64_lfrintnv4sfv4si (__a);
15412 }
15413
15414 __extension__ extern __inline uint32x4_t
15415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15416 vcvtnq_u32_f32 (float32x4_t __a)
15417 {
15418   return __builtin_aarch64_lfrintnuv4sfv4si_us (__a);
15419 }
15420
15421 __extension__ extern __inline int64x1_t
15422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15423 vcvtn_s64_f64 (float64x1_t __a)
15424 {
15425   return (int64x1_t) {vcvtnd_s64_f64 (__a[0])};
15426 }
15427
15428 __extension__ extern __inline uint64x1_t
15429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15430 vcvtn_u64_f64 (float64x1_t __a)
15431 {
15432   return (uint64x1_t) {vcvtnd_u64_f64 (__a[0])};
15433 }
15434
15435 __extension__ extern __inline int64x2_t
15436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15437 vcvtnq_s64_f64 (float64x2_t __a)
15438 {
15439   return __builtin_aarch64_lfrintnv2dfv2di (__a);
15440 }
15441
15442 __extension__ extern __inline uint64x2_t
15443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15444 vcvtnq_u64_f64 (float64x2_t __a)
15445 {
15446   return __builtin_aarch64_lfrintnuv2dfv2di_us (__a);
15447 }
15448
15449 /* vcvtp  */
15450
15451 __extension__ extern __inline int64_t
15452 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15453 vcvtpd_s64_f64 (float64_t __a)
15454 {
15455   return __builtin_llceil (__a);
15456 }
15457
15458 __extension__ extern __inline uint64_t
15459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15460 vcvtpd_u64_f64 (float64_t __a)
15461 {
15462   return __builtin_aarch64_lceiludfdi_us (__a);
15463 }
15464
15465 __extension__ extern __inline int32_t
15466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15467 vcvtps_s32_f32 (float32_t __a)
15468 {
15469   return __builtin_iceilf (__a);
15470 }
15471
15472 __extension__ extern __inline uint32_t
15473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15474 vcvtps_u32_f32 (float32_t __a)
15475 {
15476   return __builtin_aarch64_lceilusfsi_us (__a);
15477 }
15478
15479 __extension__ extern __inline int32x2_t
15480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15481 vcvtp_s32_f32 (float32x2_t __a)
15482 {
15483   return __builtin_aarch64_lceilv2sfv2si (__a);
15484 }
15485
15486 __extension__ extern __inline uint32x2_t
15487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15488 vcvtp_u32_f32 (float32x2_t __a)
15489 {
15490   return __builtin_aarch64_lceiluv2sfv2si_us (__a);
15491 }
15492
15493 __extension__ extern __inline int32x4_t
15494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15495 vcvtpq_s32_f32 (float32x4_t __a)
15496 {
15497   return __builtin_aarch64_lceilv4sfv4si (__a);
15498 }
15499
15500 __extension__ extern __inline uint32x4_t
15501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15502 vcvtpq_u32_f32 (float32x4_t __a)
15503 {
15504   return __builtin_aarch64_lceiluv4sfv4si_us (__a);
15505 }
15506
15507 __extension__ extern __inline int64x1_t
15508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15509 vcvtp_s64_f64 (float64x1_t __a)
15510 {
15511   return (int64x1_t) {vcvtpd_s64_f64 (__a[0])};
15512 }
15513
15514 __extension__ extern __inline uint64x1_t
15515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15516 vcvtp_u64_f64 (float64x1_t __a)
15517 {
15518   return (uint64x1_t) {vcvtpd_u64_f64 (__a[0])};
15519 }
15520
15521 __extension__ extern __inline int64x2_t
15522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15523 vcvtpq_s64_f64 (float64x2_t __a)
15524 {
15525   return __builtin_aarch64_lceilv2dfv2di (__a);
15526 }
15527
15528 __extension__ extern __inline uint64x2_t
15529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15530 vcvtpq_u64_f64 (float64x2_t __a)
15531 {
15532   return __builtin_aarch64_lceiluv2dfv2di_us (__a);
15533 }
15534
15535 /* vdup_n  */
15536
15537 __extension__ extern __inline float16x4_t
15538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15539 vdup_n_f16 (float16_t __a)
15540 {
15541   return (float16x4_t) {__a, __a, __a, __a};
15542 }
15543
15544 __extension__ extern __inline float32x2_t
15545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15546 vdup_n_f32 (float32_t __a)
15547 {
15548   return (float32x2_t) {__a, __a};
15549 }
15550
15551 __extension__ extern __inline float64x1_t
15552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15553 vdup_n_f64 (float64_t __a)
15554 {
15555   return (float64x1_t) {__a};
15556 }
15557
15558 __extension__ extern __inline poly8x8_t
15559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15560 vdup_n_p8 (poly8_t __a)
15561 {
15562   return (poly8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15563 }
15564
15565 __extension__ extern __inline poly16x4_t
15566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15567 vdup_n_p16 (poly16_t __a)
15568 {
15569   return (poly16x4_t) {__a, __a, __a, __a};
15570 }
15571
15572 __extension__ extern __inline poly64x1_t
15573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15574 vdup_n_p64 (poly64_t __a)
15575 {
15576   return (poly64x1_t) {__a};
15577 }
15578
15579 __extension__ extern __inline int8x8_t
15580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15581 vdup_n_s8 (int8_t __a)
15582 {
15583   return (int8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15584 }
15585
15586 __extension__ extern __inline int16x4_t
15587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15588 vdup_n_s16 (int16_t __a)
15589 {
15590   return (int16x4_t) {__a, __a, __a, __a};
15591 }
15592
15593 __extension__ extern __inline int32x2_t
15594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15595 vdup_n_s32 (int32_t __a)
15596 {
15597   return (int32x2_t) {__a, __a};
15598 }
15599
15600 __extension__ extern __inline int64x1_t
15601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15602 vdup_n_s64 (int64_t __a)
15603 {
15604   return (int64x1_t) {__a};
15605 }
15606
15607 __extension__ extern __inline uint8x8_t
15608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15609 vdup_n_u8 (uint8_t __a)
15610 {
15611   return (uint8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15612 }
15613
15614 __extension__ extern __inline uint16x4_t
15615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15616 vdup_n_u16 (uint16_t __a)
15617 {
15618   return (uint16x4_t) {__a, __a, __a, __a};
15619 }
15620
15621 __extension__ extern __inline uint32x2_t
15622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15623 vdup_n_u32 (uint32_t __a)
15624 {
15625   return (uint32x2_t) {__a, __a};
15626 }
15627
15628 __extension__ extern __inline uint64x1_t
15629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15630 vdup_n_u64 (uint64_t __a)
15631 {
15632   return (uint64x1_t) {__a};
15633 }
15634
15635 /* vdupq_n  */
15636
15637 __extension__ extern __inline float16x8_t
15638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15639 vdupq_n_f16 (float16_t __a)
15640 {
15641   return (float16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15642 }
15643
15644 __extension__ extern __inline float32x4_t
15645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15646 vdupq_n_f32 (float32_t __a)
15647 {
15648   return (float32x4_t) {__a, __a, __a, __a};
15649 }
15650
15651 __extension__ extern __inline float64x2_t
15652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15653 vdupq_n_f64 (float64_t __a)
15654 {
15655   return (float64x2_t) {__a, __a};
15656 }
15657
15658 __extension__ extern __inline poly8x16_t
15659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15660 vdupq_n_p8 (uint32_t __a)
15661 {
15662   return (poly8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15663                        __a, __a, __a, __a, __a, __a, __a, __a};
15664 }
15665
15666 __extension__ extern __inline poly16x8_t
15667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15668 vdupq_n_p16 (uint32_t __a)
15669 {
15670   return (poly16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15671 }
15672
15673 __extension__ extern __inline poly64x2_t
15674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15675 vdupq_n_p64 (uint64_t __a)
15676 {
15677   return (poly64x2_t) {__a, __a};
15678 }
15679
15680 __extension__ extern __inline int8x16_t
15681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15682 vdupq_n_s8 (int32_t __a)
15683 {
15684   return (int8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15685                       __a, __a, __a, __a, __a, __a, __a, __a};
15686 }
15687
15688 __extension__ extern __inline int16x8_t
15689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15690 vdupq_n_s16 (int32_t __a)
15691 {
15692   return (int16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15693 }
15694
15695 __extension__ extern __inline int32x4_t
15696 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15697 vdupq_n_s32 (int32_t __a)
15698 {
15699   return (int32x4_t) {__a, __a, __a, __a};
15700 }
15701
15702 __extension__ extern __inline int64x2_t
15703 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15704 vdupq_n_s64 (int64_t __a)
15705 {
15706   return (int64x2_t) {__a, __a};
15707 }
15708
15709 __extension__ extern __inline uint8x16_t
15710 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15711 vdupq_n_u8 (uint32_t __a)
15712 {
15713   return (uint8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15714                        __a, __a, __a, __a, __a, __a, __a, __a};
15715 }
15716
15717 __extension__ extern __inline uint16x8_t
15718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15719 vdupq_n_u16 (uint32_t __a)
15720 {
15721   return (uint16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15722 }
15723
15724 __extension__ extern __inline uint32x4_t
15725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15726 vdupq_n_u32 (uint32_t __a)
15727 {
15728   return (uint32x4_t) {__a, __a, __a, __a};
15729 }
15730
15731 __extension__ extern __inline uint64x2_t
15732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15733 vdupq_n_u64 (uint64_t __a)
15734 {
15735   return (uint64x2_t) {__a, __a};
15736 }
15737
15738 /* vdup_lane  */
15739
15740 __extension__ extern __inline float16x4_t
15741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15742 vdup_lane_f16 (float16x4_t __a, const int __b)
15743 {
15744   return __aarch64_vdup_lane_f16 (__a, __b);
15745 }
15746
15747 __extension__ extern __inline float32x2_t
15748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15749 vdup_lane_f32 (float32x2_t __a, const int __b)
15750 {
15751   return __aarch64_vdup_lane_f32 (__a, __b);
15752 }
15753
15754 __extension__ extern __inline float64x1_t
15755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15756 vdup_lane_f64 (float64x1_t __a, const int __b)
15757 {
15758   return __aarch64_vdup_lane_f64 (__a, __b);
15759 }
15760
15761 __extension__ extern __inline poly8x8_t
15762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15763 vdup_lane_p8 (poly8x8_t __a, const int __b)
15764 {
15765   return __aarch64_vdup_lane_p8 (__a, __b);
15766 }
15767
15768 __extension__ extern __inline poly16x4_t
15769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15770 vdup_lane_p16 (poly16x4_t __a, const int __b)
15771 {
15772   return __aarch64_vdup_lane_p16 (__a, __b);
15773 }
15774
15775 __extension__ extern __inline poly64x1_t
15776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15777 vdup_lane_p64 (poly64x1_t __a, const int __b)
15778 {
15779   return __aarch64_vdup_lane_p64 (__a, __b);
15780 }
15781
15782 __extension__ extern __inline int8x8_t
15783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15784 vdup_lane_s8 (int8x8_t __a, const int __b)
15785 {
15786   return __aarch64_vdup_lane_s8 (__a, __b);
15787 }
15788
15789 __extension__ extern __inline int16x4_t
15790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15791 vdup_lane_s16 (int16x4_t __a, const int __b)
15792 {
15793   return __aarch64_vdup_lane_s16 (__a, __b);
15794 }
15795
15796 __extension__ extern __inline int32x2_t
15797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15798 vdup_lane_s32 (int32x2_t __a, const int __b)
15799 {
15800   return __aarch64_vdup_lane_s32 (__a, __b);
15801 }
15802
15803 __extension__ extern __inline int64x1_t
15804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15805 vdup_lane_s64 (int64x1_t __a, const int __b)
15806 {
15807   return __aarch64_vdup_lane_s64 (__a, __b);
15808 }
15809
15810 __extension__ extern __inline uint8x8_t
15811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15812 vdup_lane_u8 (uint8x8_t __a, const int __b)
15813 {
15814   return __aarch64_vdup_lane_u8 (__a, __b);
15815 }
15816
15817 __extension__ extern __inline uint16x4_t
15818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15819 vdup_lane_u16 (uint16x4_t __a, const int __b)
15820 {
15821   return __aarch64_vdup_lane_u16 (__a, __b);
15822 }
15823
15824 __extension__ extern __inline uint32x2_t
15825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15826 vdup_lane_u32 (uint32x2_t __a, const int __b)
15827 {
15828   return __aarch64_vdup_lane_u32 (__a, __b);
15829 }
15830
15831 __extension__ extern __inline uint64x1_t
15832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15833 vdup_lane_u64 (uint64x1_t __a, const int __b)
15834 {
15835   return __aarch64_vdup_lane_u64 (__a, __b);
15836 }
15837
15838 /* vdup_laneq  */
15839
15840 __extension__ extern __inline float16x4_t
15841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15842 vdup_laneq_f16 (float16x8_t __a, const int __b)
15843 {
15844   return __aarch64_vdup_laneq_f16 (__a, __b);
15845 }
15846
15847 __extension__ extern __inline float32x2_t
15848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15849 vdup_laneq_f32 (float32x4_t __a, const int __b)
15850 {
15851   return __aarch64_vdup_laneq_f32 (__a, __b);
15852 }
15853
15854 __extension__ extern __inline float64x1_t
15855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15856 vdup_laneq_f64 (float64x2_t __a, const int __b)
15857 {
15858   return __aarch64_vdup_laneq_f64 (__a, __b);
15859 }
15860
15861 __extension__ extern __inline poly8x8_t
15862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15863 vdup_laneq_p8 (poly8x16_t __a, const int __b)
15864 {
15865   return __aarch64_vdup_laneq_p8 (__a, __b);
15866 }
15867
15868 __extension__ extern __inline poly16x4_t
15869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15870 vdup_laneq_p16 (poly16x8_t __a, const int __b)
15871 {
15872   return __aarch64_vdup_laneq_p16 (__a, __b);
15873 }
15874
15875 __extension__ extern __inline poly64x1_t
15876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15877 vdup_laneq_p64 (poly64x2_t __a, const int __b)
15878 {
15879   return __aarch64_vdup_laneq_p64 (__a, __b);
15880 }
15881
15882 __extension__ extern __inline int8x8_t
15883 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15884 vdup_laneq_s8 (int8x16_t __a, const int __b)
15885 {
15886   return __aarch64_vdup_laneq_s8 (__a, __b);
15887 }
15888
15889 __extension__ extern __inline int16x4_t
15890 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15891 vdup_laneq_s16 (int16x8_t __a, const int __b)
15892 {
15893   return __aarch64_vdup_laneq_s16 (__a, __b);
15894 }
15895
15896 __extension__ extern __inline int32x2_t
15897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15898 vdup_laneq_s32 (int32x4_t __a, const int __b)
15899 {
15900   return __aarch64_vdup_laneq_s32 (__a, __b);
15901 }
15902
15903 __extension__ extern __inline int64x1_t
15904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15905 vdup_laneq_s64 (int64x2_t __a, const int __b)
15906 {
15907   return __aarch64_vdup_laneq_s64 (__a, __b);
15908 }
15909
15910 __extension__ extern __inline uint8x8_t
15911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15912 vdup_laneq_u8 (uint8x16_t __a, const int __b)
15913 {
15914   return __aarch64_vdup_laneq_u8 (__a, __b);
15915 }
15916
15917 __extension__ extern __inline uint16x4_t
15918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15919 vdup_laneq_u16 (uint16x8_t __a, const int __b)
15920 {
15921   return __aarch64_vdup_laneq_u16 (__a, __b);
15922 }
15923
15924 __extension__ extern __inline uint32x2_t
15925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15926 vdup_laneq_u32 (uint32x4_t __a, const int __b)
15927 {
15928   return __aarch64_vdup_laneq_u32 (__a, __b);
15929 }
15930
15931 __extension__ extern __inline uint64x1_t
15932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15933 vdup_laneq_u64 (uint64x2_t __a, const int __b)
15934 {
15935   return __aarch64_vdup_laneq_u64 (__a, __b);
15936 }
15937
15938 /* vdupq_lane  */
15939
15940 __extension__ extern __inline float16x8_t
15941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15942 vdupq_lane_f16 (float16x4_t __a, const int __b)
15943 {
15944   return __aarch64_vdupq_lane_f16 (__a, __b);
15945 }
15946
15947 __extension__ extern __inline float32x4_t
15948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15949 vdupq_lane_f32 (float32x2_t __a, const int __b)
15950 {
15951   return __aarch64_vdupq_lane_f32 (__a, __b);
15952 }
15953
15954 __extension__ extern __inline float64x2_t
15955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15956 vdupq_lane_f64 (float64x1_t __a, const int __b)
15957 {
15958   return __aarch64_vdupq_lane_f64 (__a, __b);
15959 }
15960
15961 __extension__ extern __inline poly8x16_t
15962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15963 vdupq_lane_p8 (poly8x8_t __a, const int __b)
15964 {
15965   return __aarch64_vdupq_lane_p8 (__a, __b);
15966 }
15967
15968 __extension__ extern __inline poly16x8_t
15969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15970 vdupq_lane_p16 (poly16x4_t __a, const int __b)
15971 {
15972   return __aarch64_vdupq_lane_p16 (__a, __b);
15973 }
15974
15975 __extension__ extern __inline poly64x2_t
15976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15977 vdupq_lane_p64 (poly64x1_t __a, const int __b)
15978 {
15979   return __aarch64_vdupq_lane_p64 (__a, __b);
15980 }
15981
15982 __extension__ extern __inline int8x16_t
15983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15984 vdupq_lane_s8 (int8x8_t __a, const int __b)
15985 {
15986   return __aarch64_vdupq_lane_s8 (__a, __b);
15987 }
15988
15989 __extension__ extern __inline int16x8_t
15990 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15991 vdupq_lane_s16 (int16x4_t __a, const int __b)
15992 {
15993   return __aarch64_vdupq_lane_s16 (__a, __b);
15994 }
15995
15996 __extension__ extern __inline int32x4_t
15997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15998 vdupq_lane_s32 (int32x2_t __a, const int __b)
15999 {
16000   return __aarch64_vdupq_lane_s32 (__a, __b);
16001 }
16002
16003 __extension__ extern __inline int64x2_t
16004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16005 vdupq_lane_s64 (int64x1_t __a, const int __b)
16006 {
16007   return __aarch64_vdupq_lane_s64 (__a, __b);
16008 }
16009
16010 __extension__ extern __inline uint8x16_t
16011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16012 vdupq_lane_u8 (uint8x8_t __a, const int __b)
16013 {
16014   return __aarch64_vdupq_lane_u8 (__a, __b);
16015 }
16016
16017 __extension__ extern __inline uint16x8_t
16018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16019 vdupq_lane_u16 (uint16x4_t __a, const int __b)
16020 {
16021   return __aarch64_vdupq_lane_u16 (__a, __b);
16022 }
16023
16024 __extension__ extern __inline uint32x4_t
16025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16026 vdupq_lane_u32 (uint32x2_t __a, const int __b)
16027 {
16028   return __aarch64_vdupq_lane_u32 (__a, __b);
16029 }
16030
16031 __extension__ extern __inline uint64x2_t
16032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16033 vdupq_lane_u64 (uint64x1_t __a, const int __b)
16034 {
16035   return __aarch64_vdupq_lane_u64 (__a, __b);
16036 }
16037
16038 /* vdupq_laneq  */
16039
16040 __extension__ extern __inline float16x8_t
16041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16042 vdupq_laneq_f16 (float16x8_t __a, const int __b)
16043 {
16044   return __aarch64_vdupq_laneq_f16 (__a, __b);
16045 }
16046
16047 __extension__ extern __inline float32x4_t
16048 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16049 vdupq_laneq_f32 (float32x4_t __a, const int __b)
16050 {
16051   return __aarch64_vdupq_laneq_f32 (__a, __b);
16052 }
16053
16054 __extension__ extern __inline float64x2_t
16055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16056 vdupq_laneq_f64 (float64x2_t __a, const int __b)
16057 {
16058   return __aarch64_vdupq_laneq_f64 (__a, __b);
16059 }
16060
16061 __extension__ extern __inline poly8x16_t
16062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16063 vdupq_laneq_p8 (poly8x16_t __a, const int __b)
16064 {
16065   return __aarch64_vdupq_laneq_p8 (__a, __b);
16066 }
16067
16068 __extension__ extern __inline poly16x8_t
16069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16070 vdupq_laneq_p16 (poly16x8_t __a, const int __b)
16071 {
16072   return __aarch64_vdupq_laneq_p16 (__a, __b);
16073 }
16074
16075 __extension__ extern __inline poly64x2_t
16076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16077 vdupq_laneq_p64 (poly64x2_t __a, const int __b)
16078 {
16079   return __aarch64_vdupq_laneq_p64 (__a, __b);
16080 }
16081
16082 __extension__ extern __inline int8x16_t
16083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16084 vdupq_laneq_s8 (int8x16_t __a, const int __b)
16085 {
16086   return __aarch64_vdupq_laneq_s8 (__a, __b);
16087 }
16088
16089 __extension__ extern __inline int16x8_t
16090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16091 vdupq_laneq_s16 (int16x8_t __a, const int __b)
16092 {
16093   return __aarch64_vdupq_laneq_s16 (__a, __b);
16094 }
16095
16096 __extension__ extern __inline int32x4_t
16097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16098 vdupq_laneq_s32 (int32x4_t __a, const int __b)
16099 {
16100   return __aarch64_vdupq_laneq_s32 (__a, __b);
16101 }
16102
16103 __extension__ extern __inline int64x2_t
16104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16105 vdupq_laneq_s64 (int64x2_t __a, const int __b)
16106 {
16107   return __aarch64_vdupq_laneq_s64 (__a, __b);
16108 }
16109
16110 __extension__ extern __inline uint8x16_t
16111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16112 vdupq_laneq_u8 (uint8x16_t __a, const int __b)
16113 {
16114   return __aarch64_vdupq_laneq_u8 (__a, __b);
16115 }
16116
16117 __extension__ extern __inline uint16x8_t
16118 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16119 vdupq_laneq_u16 (uint16x8_t __a, const int __b)
16120 {
16121   return __aarch64_vdupq_laneq_u16 (__a, __b);
16122 }
16123
16124 __extension__ extern __inline uint32x4_t
16125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16126 vdupq_laneq_u32 (uint32x4_t __a, const int __b)
16127 {
16128   return __aarch64_vdupq_laneq_u32 (__a, __b);
16129 }
16130
16131 __extension__ extern __inline uint64x2_t
16132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16133 vdupq_laneq_u64 (uint64x2_t __a, const int __b)
16134 {
16135   return __aarch64_vdupq_laneq_u64 (__a, __b);
16136 }
16137
16138 /* vdupb_lane  */
16139 __extension__ extern __inline poly8_t
16140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16141 vdupb_lane_p8 (poly8x8_t __a, const int __b)
16142 {
16143   return __aarch64_vget_lane_any (__a, __b);
16144 }
16145
16146 __extension__ extern __inline int8_t
16147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16148 vdupb_lane_s8 (int8x8_t __a, const int __b)
16149 {
16150   return __aarch64_vget_lane_any (__a, __b);
16151 }
16152
16153 __extension__ extern __inline uint8_t
16154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16155 vdupb_lane_u8 (uint8x8_t __a, const int __b)
16156 {
16157   return __aarch64_vget_lane_any (__a, __b);
16158 }
16159
16160 /* vduph_lane  */
16161
16162 __extension__ extern __inline float16_t
16163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16164 vduph_lane_f16 (float16x4_t __a, const int __b)
16165 {
16166   return __aarch64_vget_lane_any (__a, __b);
16167 }
16168
16169 __extension__ extern __inline poly16_t
16170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16171 vduph_lane_p16 (poly16x4_t __a, const int __b)
16172 {
16173   return __aarch64_vget_lane_any (__a, __b);
16174 }
16175
16176 __extension__ extern __inline int16_t
16177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16178 vduph_lane_s16 (int16x4_t __a, const int __b)
16179 {
16180   return __aarch64_vget_lane_any (__a, __b);
16181 }
16182
16183 __extension__ extern __inline uint16_t
16184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16185 vduph_lane_u16 (uint16x4_t __a, const int __b)
16186 {
16187   return __aarch64_vget_lane_any (__a, __b);
16188 }
16189
16190 /* vdups_lane  */
16191
16192 __extension__ extern __inline float32_t
16193 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16194 vdups_lane_f32 (float32x2_t __a, const int __b)
16195 {
16196   return __aarch64_vget_lane_any (__a, __b);
16197 }
16198
16199 __extension__ extern __inline int32_t
16200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16201 vdups_lane_s32 (int32x2_t __a, const int __b)
16202 {
16203   return __aarch64_vget_lane_any (__a, __b);
16204 }
16205
16206 __extension__ extern __inline uint32_t
16207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16208 vdups_lane_u32 (uint32x2_t __a, const int __b)
16209 {
16210   return __aarch64_vget_lane_any (__a, __b);
16211 }
16212
16213 /* vdupd_lane  */
16214 __extension__ extern __inline float64_t
16215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16216 vdupd_lane_f64 (float64x1_t __a, const int __b)
16217 {
16218   __AARCH64_LANE_CHECK (__a, __b);
16219   return __a[0];
16220 }
16221
16222 __extension__ extern __inline int64_t
16223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16224 vdupd_lane_s64 (int64x1_t __a, const int __b)
16225 {
16226   __AARCH64_LANE_CHECK (__a, __b);
16227   return __a[0];
16228 }
16229
16230 __extension__ extern __inline uint64_t
16231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16232 vdupd_lane_u64 (uint64x1_t __a, const int __b)
16233 {
16234   __AARCH64_LANE_CHECK (__a, __b);
16235   return __a[0];
16236 }
16237
16238 /* vdupb_laneq  */
16239 __extension__ extern __inline poly8_t
16240 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16241 vdupb_laneq_p8 (poly8x16_t __a, const int __b)
16242 {
16243   return __aarch64_vget_lane_any (__a, __b);
16244 }
16245
16246 __extension__ extern __inline int8_t
16247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16248 vdupb_laneq_s8 (int8x16_t __a, const int __b)
16249 {
16250   return __aarch64_vget_lane_any (__a, __b);
16251 }
16252
16253 __extension__ extern __inline uint8_t
16254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16255 vdupb_laneq_u8 (uint8x16_t __a, const int __b)
16256 {
16257   return __aarch64_vget_lane_any (__a, __b);
16258 }
16259
16260 /* vduph_laneq  */
16261
16262 __extension__ extern __inline float16_t
16263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16264 vduph_laneq_f16 (float16x8_t __a, const int __b)
16265 {
16266   return __aarch64_vget_lane_any (__a, __b);
16267 }
16268
16269 __extension__ extern __inline poly16_t
16270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16271 vduph_laneq_p16 (poly16x8_t __a, const int __b)
16272 {
16273   return __aarch64_vget_lane_any (__a, __b);
16274 }
16275
16276 __extension__ extern __inline int16_t
16277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16278 vduph_laneq_s16 (int16x8_t __a, const int __b)
16279 {
16280   return __aarch64_vget_lane_any (__a, __b);
16281 }
16282
16283 __extension__ extern __inline uint16_t
16284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16285 vduph_laneq_u16 (uint16x8_t __a, const int __b)
16286 {
16287   return __aarch64_vget_lane_any (__a, __b);
16288 }
16289
16290 /* vdups_laneq  */
16291
16292 __extension__ extern __inline float32_t
16293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16294 vdups_laneq_f32 (float32x4_t __a, const int __b)
16295 {
16296   return __aarch64_vget_lane_any (__a, __b);
16297 }
16298
16299 __extension__ extern __inline int32_t
16300 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16301 vdups_laneq_s32 (int32x4_t __a, const int __b)
16302 {
16303   return __aarch64_vget_lane_any (__a, __b);
16304 }
16305
16306 __extension__ extern __inline uint32_t
16307 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16308 vdups_laneq_u32 (uint32x4_t __a, const int __b)
16309 {
16310   return __aarch64_vget_lane_any (__a, __b);
16311 }
16312
16313 /* vdupd_laneq  */
16314 __extension__ extern __inline float64_t
16315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16316 vdupd_laneq_f64 (float64x2_t __a, const int __b)
16317 {
16318   return __aarch64_vget_lane_any (__a, __b);
16319 }
16320
16321 __extension__ extern __inline int64_t
16322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16323 vdupd_laneq_s64 (int64x2_t __a, const int __b)
16324 {
16325   return __aarch64_vget_lane_any (__a, __b);
16326 }
16327
16328 __extension__ extern __inline uint64_t
16329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16330 vdupd_laneq_u64 (uint64x2_t __a, const int __b)
16331 {
16332   return __aarch64_vget_lane_any (__a, __b);
16333 }
16334
16335 /* vext  */
16336
16337 __extension__ extern __inline float16x4_t
16338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16339 vext_f16 (float16x4_t __a, float16x4_t __b, __const int __c)
16340 {
16341   __AARCH64_LANE_CHECK (__a, __c);
16342 #ifdef __AARCH64EB__
16343   return __builtin_shuffle (__b, __a,
16344                             (uint16x4_t) {4 - __c, 5 - __c, 6 - __c, 7 - __c});
16345 #else
16346   return __builtin_shuffle (__a, __b,
16347                             (uint16x4_t) {__c, __c + 1, __c + 2, __c + 3});
16348 #endif
16349 }
16350
16351 __extension__ extern __inline float32x2_t
16352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16353 vext_f32 (float32x2_t __a, float32x2_t __b, __const int __c)
16354 {
16355   __AARCH64_LANE_CHECK (__a, __c);
16356 #ifdef __AARCH64EB__
16357   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16358 #else
16359   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16360 #endif
16361 }
16362
16363 __extension__ extern __inline float64x1_t
16364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16365 vext_f64 (float64x1_t __a, float64x1_t __b, __const int __c)
16366 {
16367   __AARCH64_LANE_CHECK (__a, __c);
16368   /* The only possible index to the assembler instruction returns element 0.  */
16369   return __a;
16370 }
16371 __extension__ extern __inline poly8x8_t
16372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16373 vext_p8 (poly8x8_t __a, poly8x8_t __b, __const int __c)
16374 {
16375   __AARCH64_LANE_CHECK (__a, __c);
16376 #ifdef __AARCH64EB__
16377   return __builtin_shuffle (__b, __a, (uint8x8_t)
16378       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16379 #else
16380   return __builtin_shuffle (__a, __b,
16381       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16382 #endif
16383 }
16384
16385 __extension__ extern __inline poly16x4_t
16386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16387 vext_p16 (poly16x4_t __a, poly16x4_t __b, __const int __c)
16388 {
16389   __AARCH64_LANE_CHECK (__a, __c);
16390 #ifdef __AARCH64EB__
16391   return __builtin_shuffle (__b, __a,
16392       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16393 #else
16394   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16395 #endif
16396 }
16397
16398 __extension__ extern __inline poly64x1_t
16399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16400 vext_p64 (poly64x1_t __a, poly64x1_t __b, __const int __c)
16401 {
16402   __AARCH64_LANE_CHECK (__a, __c);
16403   /* The only possible index to the assembler instruction returns element 0.  */
16404   return __a;
16405 }
16406
16407 __extension__ extern __inline int8x8_t
16408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16409 vext_s8 (int8x8_t __a, int8x8_t __b, __const int __c)
16410 {
16411   __AARCH64_LANE_CHECK (__a, __c);
16412 #ifdef __AARCH64EB__
16413   return __builtin_shuffle (__b, __a, (uint8x8_t)
16414       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16415 #else
16416   return __builtin_shuffle (__a, __b,
16417       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16418 #endif
16419 }
16420
16421 __extension__ extern __inline int16x4_t
16422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16423 vext_s16 (int16x4_t __a, int16x4_t __b, __const int __c)
16424 {
16425   __AARCH64_LANE_CHECK (__a, __c);
16426 #ifdef __AARCH64EB__
16427   return __builtin_shuffle (__b, __a,
16428       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16429 #else
16430   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16431 #endif
16432 }
16433
16434 __extension__ extern __inline int32x2_t
16435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16436 vext_s32 (int32x2_t __a, int32x2_t __b, __const int __c)
16437 {
16438   __AARCH64_LANE_CHECK (__a, __c);
16439 #ifdef __AARCH64EB__
16440   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16441 #else
16442   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16443 #endif
16444 }
16445
16446 __extension__ extern __inline int64x1_t
16447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16448 vext_s64 (int64x1_t __a, int64x1_t __b, __const int __c)
16449 {
16450   __AARCH64_LANE_CHECK (__a, __c);
16451   /* The only possible index to the assembler instruction returns element 0.  */
16452   return __a;
16453 }
16454
16455 __extension__ extern __inline uint8x8_t
16456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16457 vext_u8 (uint8x8_t __a, uint8x8_t __b, __const int __c)
16458 {
16459   __AARCH64_LANE_CHECK (__a, __c);
16460 #ifdef __AARCH64EB__
16461   return __builtin_shuffle (__b, __a, (uint8x8_t)
16462       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16463 #else
16464   return __builtin_shuffle (__a, __b,
16465       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16466 #endif
16467 }
16468
16469 __extension__ extern __inline uint16x4_t
16470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16471 vext_u16 (uint16x4_t __a, uint16x4_t __b, __const int __c)
16472 {
16473   __AARCH64_LANE_CHECK (__a, __c);
16474 #ifdef __AARCH64EB__
16475   return __builtin_shuffle (__b, __a,
16476       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16477 #else
16478   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16479 #endif
16480 }
16481
16482 __extension__ extern __inline uint32x2_t
16483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16484 vext_u32 (uint32x2_t __a, uint32x2_t __b, __const int __c)
16485 {
16486   __AARCH64_LANE_CHECK (__a, __c);
16487 #ifdef __AARCH64EB__
16488   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16489 #else
16490   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16491 #endif
16492 }
16493
16494 __extension__ extern __inline uint64x1_t
16495 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16496 vext_u64 (uint64x1_t __a, uint64x1_t __b, __const int __c)
16497 {
16498   __AARCH64_LANE_CHECK (__a, __c);
16499   /* The only possible index to the assembler instruction returns element 0.  */
16500   return __a;
16501 }
16502
16503 __extension__ extern __inline float16x8_t
16504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16505 vextq_f16 (float16x8_t __a, float16x8_t __b, __const int __c)
16506 {
16507   __AARCH64_LANE_CHECK (__a, __c);
16508 #ifdef __AARCH64EB__
16509   return __builtin_shuffle (__b, __a,
16510                             (uint16x8_t) {8 - __c, 9 - __c, 10 - __c, 11 - __c,
16511                                           12 - __c, 13 - __c, 14 - __c,
16512                                           15 - __c});
16513 #else
16514   return __builtin_shuffle (__a, __b,
16515                             (uint16x8_t) {__c, __c + 1, __c + 2, __c + 3,
16516                                           __c + 4, __c + 5, __c + 6, __c + 7});
16517 #endif
16518 }
16519
16520 __extension__ extern __inline float32x4_t
16521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16522 vextq_f32 (float32x4_t __a, float32x4_t __b, __const int __c)
16523 {
16524   __AARCH64_LANE_CHECK (__a, __c);
16525 #ifdef __AARCH64EB__
16526   return __builtin_shuffle (__b, __a,
16527       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16528 #else
16529   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16530 #endif
16531 }
16532
16533 __extension__ extern __inline float64x2_t
16534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16535 vextq_f64 (float64x2_t __a, float64x2_t __b, __const int __c)
16536 {
16537   __AARCH64_LANE_CHECK (__a, __c);
16538 #ifdef __AARCH64EB__
16539   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16540 #else
16541   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16542 #endif
16543 }
16544
16545 __extension__ extern __inline poly8x16_t
16546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16547 vextq_p8 (poly8x16_t __a, poly8x16_t __b, __const int __c)
16548 {
16549   __AARCH64_LANE_CHECK (__a, __c);
16550 #ifdef __AARCH64EB__
16551   return __builtin_shuffle (__b, __a, (uint8x16_t)
16552       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16553        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16554 #else
16555   return __builtin_shuffle (__a, __b, (uint8x16_t)
16556       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16557        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16558 #endif
16559 }
16560
16561 __extension__ extern __inline poly16x8_t
16562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16563 vextq_p16 (poly16x8_t __a, poly16x8_t __b, __const int __c)
16564 {
16565   __AARCH64_LANE_CHECK (__a, __c);
16566 #ifdef __AARCH64EB__
16567   return __builtin_shuffle (__b, __a, (uint16x8_t)
16568       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16569 #else
16570   return __builtin_shuffle (__a, __b,
16571       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16572 #endif
16573 }
16574
16575 __extension__ extern __inline poly64x2_t
16576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16577 vextq_p64 (poly64x2_t __a, poly64x2_t __b, __const int __c)
16578 {
16579   __AARCH64_LANE_CHECK (__a, __c);
16580 #ifdef __AARCH64EB__
16581   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16582 #else
16583   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16584 #endif
16585 }
16586
16587 __extension__ extern __inline int8x16_t
16588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16589 vextq_s8 (int8x16_t __a, int8x16_t __b, __const int __c)
16590 {
16591   __AARCH64_LANE_CHECK (__a, __c);
16592 #ifdef __AARCH64EB__
16593   return __builtin_shuffle (__b, __a, (uint8x16_t)
16594       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16595        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16596 #else
16597   return __builtin_shuffle (__a, __b, (uint8x16_t)
16598       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16599        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16600 #endif
16601 }
16602
16603 __extension__ extern __inline int16x8_t
16604 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16605 vextq_s16 (int16x8_t __a, int16x8_t __b, __const int __c)
16606 {
16607   __AARCH64_LANE_CHECK (__a, __c);
16608 #ifdef __AARCH64EB__
16609   return __builtin_shuffle (__b, __a, (uint16x8_t)
16610       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16611 #else
16612   return __builtin_shuffle (__a, __b,
16613       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16614 #endif
16615 }
16616
16617 __extension__ extern __inline int32x4_t
16618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16619 vextq_s32 (int32x4_t __a, int32x4_t __b, __const int __c)
16620 {
16621   __AARCH64_LANE_CHECK (__a, __c);
16622 #ifdef __AARCH64EB__
16623   return __builtin_shuffle (__b, __a,
16624       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16625 #else
16626   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16627 #endif
16628 }
16629
16630 __extension__ extern __inline int64x2_t
16631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16632 vextq_s64 (int64x2_t __a, int64x2_t __b, __const int __c)
16633 {
16634   __AARCH64_LANE_CHECK (__a, __c);
16635 #ifdef __AARCH64EB__
16636   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16637 #else
16638   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16639 #endif
16640 }
16641
16642 __extension__ extern __inline uint8x16_t
16643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16644 vextq_u8 (uint8x16_t __a, uint8x16_t __b, __const int __c)
16645 {
16646   __AARCH64_LANE_CHECK (__a, __c);
16647 #ifdef __AARCH64EB__
16648   return __builtin_shuffle (__b, __a, (uint8x16_t)
16649       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16650        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16651 #else
16652   return __builtin_shuffle (__a, __b, (uint8x16_t)
16653       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16654        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16655 #endif
16656 }
16657
16658 __extension__ extern __inline uint16x8_t
16659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16660 vextq_u16 (uint16x8_t __a, uint16x8_t __b, __const int __c)
16661 {
16662   __AARCH64_LANE_CHECK (__a, __c);
16663 #ifdef __AARCH64EB__
16664   return __builtin_shuffle (__b, __a, (uint16x8_t)
16665       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16666 #else
16667   return __builtin_shuffle (__a, __b,
16668       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16669 #endif
16670 }
16671
16672 __extension__ extern __inline uint32x4_t
16673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16674 vextq_u32 (uint32x4_t __a, uint32x4_t __b, __const int __c)
16675 {
16676   __AARCH64_LANE_CHECK (__a, __c);
16677 #ifdef __AARCH64EB__
16678   return __builtin_shuffle (__b, __a,
16679       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16680 #else
16681   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16682 #endif
16683 }
16684
16685 __extension__ extern __inline uint64x2_t
16686 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16687 vextq_u64 (uint64x2_t __a, uint64x2_t __b, __const int __c)
16688 {
16689   __AARCH64_LANE_CHECK (__a, __c);
16690 #ifdef __AARCH64EB__
16691   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16692 #else
16693   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16694 #endif
16695 }
16696
16697 /* vfma  */
16698
16699 __extension__ extern __inline float64x1_t
16700 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16701 vfma_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
16702 {
16703   return (float64x1_t) {__builtin_fma (__b[0], __c[0], __a[0])};
16704 }
16705
16706 __extension__ extern __inline float32x2_t
16707 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16708 vfma_f32 (float32x2_t __a, float32x2_t __b, float32x2_t __c)
16709 {
16710   return __builtin_aarch64_fmav2sf (__b, __c, __a);
16711 }
16712
16713 __extension__ extern __inline float32x4_t
16714 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16715 vfmaq_f32 (float32x4_t __a, float32x4_t __b, float32x4_t __c)
16716 {
16717   return __builtin_aarch64_fmav4sf (__b, __c, __a);
16718 }
16719
16720 __extension__ extern __inline float64x2_t
16721 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16722 vfmaq_f64 (float64x2_t __a, float64x2_t __b, float64x2_t __c)
16723 {
16724   return __builtin_aarch64_fmav2df (__b, __c, __a);
16725 }
16726
16727 __extension__ extern __inline float32x2_t
16728 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16729 vfma_n_f32 (float32x2_t __a, float32x2_t __b, float32_t __c)
16730 {
16731   return __builtin_aarch64_fmav2sf (__b, vdup_n_f32 (__c), __a);
16732 }
16733
16734 __extension__ extern __inline float64x1_t
16735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16736 vfma_n_f64 (float64x1_t __a, float64x1_t __b, float64_t __c)
16737 {
16738   return (float64x1_t) {__b[0] * __c + __a[0]};
16739 }
16740
16741 __extension__ extern __inline float32x4_t
16742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16743 vfmaq_n_f32 (float32x4_t __a, float32x4_t __b, float32_t __c)
16744 {
16745   return __builtin_aarch64_fmav4sf (__b, vdupq_n_f32 (__c), __a);
16746 }
16747
16748 __extension__ extern __inline float64x2_t
16749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16750 vfmaq_n_f64 (float64x2_t __a, float64x2_t __b, float64_t __c)
16751 {
16752   return __builtin_aarch64_fmav2df (__b, vdupq_n_f64 (__c), __a);
16753 }
16754
16755 /* vfma_lane  */
16756
16757 __extension__ extern __inline float32x2_t
16758 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16759 vfma_lane_f32 (float32x2_t __a, float32x2_t __b,
16760                float32x2_t __c, const int __lane)
16761 {
16762   return __builtin_aarch64_fmav2sf (__b,
16763                                     __aarch64_vdup_lane_f32 (__c, __lane),
16764                                     __a);
16765 }
16766
16767 __extension__ extern __inline float64x1_t
16768 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16769 vfma_lane_f64 (float64x1_t __a, float64x1_t __b,
16770                float64x1_t __c, const int __lane)
16771 {
16772   return (float64x1_t) {__builtin_fma (__b[0], __c[0], __a[0])};
16773 }
16774
16775 __extension__ extern __inline float64_t
16776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16777 vfmad_lane_f64 (float64_t __a, float64_t __b,
16778                 float64x1_t __c, const int __lane)
16779 {
16780   return __builtin_fma (__b, __c[0], __a);
16781 }
16782
16783 __extension__ extern __inline float32_t
16784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16785 vfmas_lane_f32 (float32_t __a, float32_t __b,
16786                 float32x2_t __c, const int __lane)
16787 {
16788   return __builtin_fmaf (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16789 }
16790
16791 /* vfma_laneq  */
16792
16793 __extension__ extern __inline float32x2_t
16794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16795 vfma_laneq_f32 (float32x2_t __a, float32x2_t __b,
16796                 float32x4_t __c, const int __lane)
16797 {
16798   return __builtin_aarch64_fmav2sf (__b,
16799                                     __aarch64_vdup_laneq_f32 (__c, __lane),
16800                                     __a);
16801 }
16802
16803 __extension__ extern __inline float64x1_t
16804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16805 vfma_laneq_f64 (float64x1_t __a, float64x1_t __b,
16806                 float64x2_t __c, const int __lane)
16807 {
16808   float64_t __c0 = __aarch64_vget_lane_any (__c, __lane);
16809   return (float64x1_t) {__builtin_fma (__b[0], __c0, __a[0])};
16810 }
16811
16812 __extension__ extern __inline float64_t
16813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16814 vfmad_laneq_f64 (float64_t __a, float64_t __b,
16815                  float64x2_t __c, const int __lane)
16816 {
16817   return __builtin_fma (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16818 }
16819
16820 __extension__ extern __inline float32_t
16821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16822 vfmas_laneq_f32 (float32_t __a, float32_t __b,
16823                  float32x4_t __c, const int __lane)
16824 {
16825   return __builtin_fmaf (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16826 }
16827
16828 /* vfmaq_lane  */
16829
16830 __extension__ extern __inline float32x4_t
16831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16832 vfmaq_lane_f32 (float32x4_t __a, float32x4_t __b,
16833                 float32x2_t __c, const int __lane)
16834 {
16835   return __builtin_aarch64_fmav4sf (__b,
16836                                     __aarch64_vdupq_lane_f32 (__c, __lane),
16837                                     __a);
16838 }
16839
16840 __extension__ extern __inline float64x2_t
16841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16842 vfmaq_lane_f64 (float64x2_t __a, float64x2_t __b,
16843                 float64x1_t __c, const int __lane)
16844 {
16845   return __builtin_aarch64_fmav2df (__b, vdupq_n_f64 (__c[0]), __a);
16846 }
16847
16848 /* vfmaq_laneq  */
16849
16850 __extension__ extern __inline float32x4_t
16851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16852 vfmaq_laneq_f32 (float32x4_t __a, float32x4_t __b,
16853                  float32x4_t __c, const int __lane)
16854 {
16855   return __builtin_aarch64_fmav4sf (__b,
16856                                     __aarch64_vdupq_laneq_f32 (__c, __lane),
16857                                     __a);
16858 }
16859
16860 __extension__ extern __inline float64x2_t
16861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16862 vfmaq_laneq_f64 (float64x2_t __a, float64x2_t __b,
16863                  float64x2_t __c, const int __lane)
16864 {
16865   return __builtin_aarch64_fmav2df (__b,
16866                                     __aarch64_vdupq_laneq_f64 (__c, __lane),
16867                                     __a);
16868 }
16869
16870 /* vfms  */
16871
16872 __extension__ extern __inline float64x1_t
16873 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16874 vfms_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
16875 {
16876   return (float64x1_t) {__builtin_fma (-__b[0], __c[0], __a[0])};
16877 }
16878
16879 __extension__ extern __inline float32x2_t
16880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16881 vfms_f32 (float32x2_t __a, float32x2_t __b, float32x2_t __c)
16882 {
16883   return __builtin_aarch64_fmav2sf (-__b, __c, __a);
16884 }
16885
16886 __extension__ extern __inline float32x4_t
16887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16888 vfmsq_f32 (float32x4_t __a, float32x4_t __b, float32x4_t __c)
16889 {
16890   return __builtin_aarch64_fmav4sf (-__b, __c, __a);
16891 }
16892
16893 __extension__ extern __inline float64x2_t
16894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16895 vfmsq_f64 (float64x2_t __a, float64x2_t __b, float64x2_t __c)
16896 {
16897   return __builtin_aarch64_fmav2df (-__b, __c, __a);
16898 }
16899
16900 __extension__ extern __inline float32x2_t
16901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16902 vfms_n_f32 (float32x2_t __a, float32x2_t __b, float32_t __c)
16903 {
16904   return __builtin_aarch64_fmav2sf (-__b, vdup_n_f32 (__c), __a);
16905 }
16906
16907 __extension__ extern __inline float64x1_t
16908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16909 vfms_n_f64 (float64x1_t __a, float64x1_t __b, float64_t __c)
16910 {
16911   return (float64x1_t) {-__b[0] * __c + __a[0]};
16912 }
16913
16914 __extension__ extern __inline float32x4_t
16915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16916 vfmsq_n_f32 (float32x4_t __a, float32x4_t __b, float32_t __c)
16917 {
16918   return __builtin_aarch64_fmav4sf (-__b, vdupq_n_f32 (__c), __a);
16919 }
16920
16921 __extension__ extern __inline float64x2_t
16922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16923 vfmsq_n_f64 (float64x2_t __a, float64x2_t __b, float64_t __c)
16924 {
16925   return __builtin_aarch64_fmav2df (-__b, vdupq_n_f64 (__c), __a);
16926 }
16927
16928 /* vfms_lane  */
16929
16930 __extension__ extern __inline float32x2_t
16931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16932 vfms_lane_f32 (float32x2_t __a, float32x2_t __b,
16933                float32x2_t __c, const int __lane)
16934 {
16935   return __builtin_aarch64_fmav2sf (-__b,
16936                                     __aarch64_vdup_lane_f32 (__c, __lane),
16937                                     __a);
16938 }
16939
16940 __extension__ extern __inline float64x1_t
16941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16942 vfms_lane_f64 (float64x1_t __a, float64x1_t __b,
16943                float64x1_t __c, const int __lane)
16944 {
16945   return (float64x1_t) {__builtin_fma (-__b[0], __c[0], __a[0])};
16946 }
16947
16948 __extension__ extern __inline float64_t
16949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16950 vfmsd_lane_f64 (float64_t __a, float64_t __b,
16951                 float64x1_t __c, const int __lane)
16952 {
16953   return __builtin_fma (-__b, __c[0], __a);
16954 }
16955
16956 __extension__ extern __inline float32_t
16957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16958 vfmss_lane_f32 (float32_t __a, float32_t __b,
16959                 float32x2_t __c, const int __lane)
16960 {
16961   return __builtin_fmaf (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16962 }
16963
16964 /* vfms_laneq  */
16965
16966 __extension__ extern __inline float32x2_t
16967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16968 vfms_laneq_f32 (float32x2_t __a, float32x2_t __b,
16969                 float32x4_t __c, const int __lane)
16970 {
16971   return __builtin_aarch64_fmav2sf (-__b,
16972                                     __aarch64_vdup_laneq_f32 (__c, __lane),
16973                                     __a);
16974 }
16975
16976 __extension__ extern __inline float64x1_t
16977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16978 vfms_laneq_f64 (float64x1_t __a, float64x1_t __b,
16979                 float64x2_t __c, const int __lane)
16980 {
16981   float64_t __c0 = __aarch64_vget_lane_any (__c, __lane);
16982   return (float64x1_t) {__builtin_fma (-__b[0], __c0, __a[0])};
16983 }
16984
16985 __extension__ extern __inline float64_t
16986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16987 vfmsd_laneq_f64 (float64_t __a, float64_t __b,
16988                  float64x2_t __c, const int __lane)
16989 {
16990   return __builtin_fma (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16991 }
16992
16993 __extension__ extern __inline float32_t
16994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16995 vfmss_laneq_f32 (float32_t __a, float32_t __b,
16996                  float32x4_t __c, const int __lane)
16997 {
16998   return __builtin_fmaf (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16999 }
17000
17001 /* vfmsq_lane  */
17002
17003 __extension__ extern __inline float32x4_t
17004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17005 vfmsq_lane_f32 (float32x4_t __a, float32x4_t __b,
17006                 float32x2_t __c, const int __lane)
17007 {
17008   return __builtin_aarch64_fmav4sf (-__b,
17009                                     __aarch64_vdupq_lane_f32 (__c, __lane),
17010                                     __a);
17011 }
17012
17013 __extension__ extern __inline float64x2_t
17014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17015 vfmsq_lane_f64 (float64x2_t __a, float64x2_t __b,
17016                 float64x1_t __c, const int __lane)
17017 {
17018   return __builtin_aarch64_fmav2df (-__b, vdupq_n_f64 (__c[0]), __a);
17019 }
17020
17021 /* vfmsq_laneq  */
17022
17023 __extension__ extern __inline float32x4_t
17024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17025 vfmsq_laneq_f32 (float32x4_t __a, float32x4_t __b,
17026                  float32x4_t __c, const int __lane)
17027 {
17028   return __builtin_aarch64_fmav4sf (-__b,
17029                                     __aarch64_vdupq_laneq_f32 (__c, __lane),
17030                                     __a);
17031 }
17032
17033 __extension__ extern __inline float64x2_t
17034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17035 vfmsq_laneq_f64 (float64x2_t __a, float64x2_t __b,
17036                  float64x2_t __c, const int __lane)
17037 {
17038   return __builtin_aarch64_fmav2df (-__b,
17039                                     __aarch64_vdupq_laneq_f64 (__c, __lane),
17040                                     __a);
17041 }
17042
17043 /* vld1 */
17044
17045 __extension__ extern __inline float16x4_t
17046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17047 vld1_f16 (const float16_t *__a)
17048 {
17049   return __builtin_aarch64_ld1v4hf (__a);
17050 }
17051
17052 __extension__ extern __inline float32x2_t
17053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17054 vld1_f32 (const float32_t *a)
17055 {
17056   return __builtin_aarch64_ld1v2sf ((const __builtin_aarch64_simd_sf *) a);
17057 }
17058
17059 __extension__ extern __inline float64x1_t
17060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17061 vld1_f64 (const float64_t *a)
17062 {
17063   return (float64x1_t) {*a};
17064 }
17065
17066 __extension__ extern __inline poly8x8_t
17067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17068 vld1_p8 (const poly8_t *a)
17069 {
17070   return (poly8x8_t)
17071     __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
17072 }
17073
17074 __extension__ extern __inline poly16x4_t
17075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17076 vld1_p16 (const poly16_t *a)
17077 {
17078   return (poly16x4_t)
17079     __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
17080 }
17081
17082 __extension__ extern __inline poly64x1_t
17083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17084 vld1_p64 (const poly64_t *a)
17085 {
17086   return (poly64x1_t) {*a};
17087 }
17088
17089 __extension__ extern __inline int8x8_t
17090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17091 vld1_s8 (const int8_t *a)
17092 {
17093   return __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
17094 }
17095
17096 __extension__ extern __inline int16x4_t
17097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17098 vld1_s16 (const int16_t *a)
17099 {
17100   return __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
17101 }
17102
17103 __extension__ extern __inline int32x2_t
17104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17105 vld1_s32 (const int32_t *a)
17106 {
17107   return __builtin_aarch64_ld1v2si ((const __builtin_aarch64_simd_si *) a);
17108 }
17109
17110 __extension__ extern __inline int64x1_t
17111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17112 vld1_s64 (const int64_t *a)
17113 {
17114   return (int64x1_t) {*a};
17115 }
17116
17117 __extension__ extern __inline uint8x8_t
17118 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17119 vld1_u8 (const uint8_t *a)
17120 {
17121   return (uint8x8_t)
17122     __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
17123 }
17124
17125 __extension__ extern __inline uint16x4_t
17126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17127 vld1_u16 (const uint16_t *a)
17128 {
17129   return (uint16x4_t)
17130     __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
17131 }
17132
17133 __extension__ extern __inline uint32x2_t
17134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17135 vld1_u32 (const uint32_t *a)
17136 {
17137   return (uint32x2_t)
17138     __builtin_aarch64_ld1v2si ((const __builtin_aarch64_simd_si *) a);
17139 }
17140
17141 __extension__ extern __inline uint64x1_t
17142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17143 vld1_u64 (const uint64_t *a)
17144 {
17145   return (uint64x1_t) {*a};
17146 }
17147
17148 /* vld1x3  */
17149
17150 __extension__ extern __inline uint8x8x3_t
17151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17152 vld1_u8_x3 (const uint8_t *__a)
17153 {
17154   uint8x8x3_t __i;
17155   __builtin_aarch64_simd_ci __o;
17156   __o = (__builtin_aarch64_simd_ci)__builtin_aarch64_ld1x3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17157   __i.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 0);
17158   __i.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 1);
17159   __i.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 2);
17160   return __i;
17161 }
17162
17163 __extension__ extern __inline int8x8x3_t
17164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17165 vld1_s8_x3 (const uint8_t *__a)
17166 {
17167   int8x8x3_t __i;
17168   __builtin_aarch64_simd_ci __o;
17169   __o = __builtin_aarch64_ld1x3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17170   __i.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 0);
17171   __i.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 1);
17172   __i.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 2);
17173   return __i;
17174 }
17175
17176 __extension__ extern __inline uint16x4x3_t
17177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17178 vld1_u16_x3 (const uint16_t *__a)
17179 {
17180   uint16x4x3_t __i;
17181   __builtin_aarch64_simd_ci __o;
17182   __o = __builtin_aarch64_ld1x3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17183   __i.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 0);
17184   __i.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 1);
17185   __i.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 2);
17186   return __i;
17187 }
17188
17189 __extension__ extern __inline int16x4x3_t
17190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17191 vld1_s16_x3 (const int16_t *__a)
17192 {
17193   int16x4x3_t __i;
17194   __builtin_aarch64_simd_ci __o;
17195   __o = __builtin_aarch64_ld1x3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17196   __i.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 0);
17197   __i.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 1);
17198   __i.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 2);
17199   return __i;
17200 }
17201
17202 __extension__ extern __inline uint32x2x3_t
17203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17204 vld1_u32_x3 (const uint32_t *__a)
17205 {
17206   uint32x2x3_t __i;
17207   __builtin_aarch64_simd_ci __o;
17208   __o = __builtin_aarch64_ld1x3v2si ((const __builtin_aarch64_simd_si *) __a);
17209   __i.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si  (__o, 0);
17210   __i.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si  (__o, 1);
17211   __i.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si  (__o, 2);
17212   return __i;
17213 }
17214
17215 __extension__ extern __inline int32x2x3_t
17216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17217 vld1_s32_x3 (const uint32_t *__a)
17218 {
17219   int32x2x3_t __i;
17220   __builtin_aarch64_simd_ci __o;
17221   __o = __builtin_aarch64_ld1x3v2si ((const __builtin_aarch64_simd_si *) __a);
17222   __i.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si  (__o, 0);
17223   __i.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si  (__o, 1);
17224   __i.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si  (__o, 2);
17225   return __i;
17226 }
17227
17228 __extension__ extern __inline uint64x1x3_t
17229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17230 vld1_u64_x3 (const uint64_t *__a)
17231 {
17232   uint64x1x3_t __i;
17233   __builtin_aarch64_simd_ci __o;
17234   __o = __builtin_aarch64_ld1x3di ((const __builtin_aarch64_simd_di *) __a);
17235   __i.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17236   __i.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17237   __i.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17238   return __i;
17239 }
17240
17241 __extension__ extern __inline int64x1x3_t
17242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17243 vld1_s64_x3 (const int64_t *__a)
17244 {
17245   int64x1x3_t __i;
17246   __builtin_aarch64_simd_ci __o;
17247   __o = __builtin_aarch64_ld1x3di ((const __builtin_aarch64_simd_di *) __a);
17248   __i.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17249   __i.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17250   __i.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17251
17252   return __i;
17253 }
17254
17255 __extension__ extern __inline float16x4x3_t
17256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17257 vld1_f16_x3 (const float16_t *__a)
17258 {
17259   float16x4x3_t __i;
17260   __builtin_aarch64_simd_ci __o;
17261   __o = __builtin_aarch64_ld1x3v4hf ((const __builtin_aarch64_simd_hf *) __a);
17262   __i.val[0] = (float16x4_t) __builtin_aarch64_get_dregciv4hf  (__o, 0);
17263   __i.val[1] = (float16x4_t) __builtin_aarch64_get_dregciv4hf  (__o, 1);
17264   __i.val[2] = (float16x4_t) __builtin_aarch64_get_dregciv4hf  (__o, 2);
17265   return __i;
17266 }
17267
17268 __extension__ extern __inline float32x2x3_t
17269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17270 vld1_f32_x3 (const float32_t *__a)
17271 {
17272   float32x2x3_t __i;
17273   __builtin_aarch64_simd_ci __o;
17274   __o = __builtin_aarch64_ld1x3v2sf ((const __builtin_aarch64_simd_sf *) __a);
17275   __i.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf  (__o, 0);
17276   __i.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf  (__o, 1);
17277   __i.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf  (__o, 2);
17278   return __i;
17279 }
17280
17281 __extension__ extern __inline float64x1x3_t
17282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17283 vld1_f64_x3 (const float64_t *__a)
17284 {
17285   float64x1x3_t __i;
17286   __builtin_aarch64_simd_ci __o;
17287   __o = __builtin_aarch64_ld1x3df ((const __builtin_aarch64_simd_df *) __a);
17288   __i.val[0] = (float64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17289   __i.val[1] = (float64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17290   __i.val[2] = (float64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17291   return __i;
17292 }
17293
17294 __extension__ extern __inline poly8x8x3_t
17295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17296 vld1_p8_x3 (const poly8_t *__a)
17297 {
17298   poly8x8x3_t __i;
17299   __builtin_aarch64_simd_ci __o;
17300   __o = __builtin_aarch64_ld1x3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17301   __i.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 0);
17302   __i.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 1);
17303   __i.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi  (__o, 2);
17304   return __i;
17305 }
17306
17307 __extension__ extern __inline poly16x4x3_t
17308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17309 vld1_p16_x3 (const poly16_t *__a)
17310 {
17311   poly16x4x3_t __i;
17312   __builtin_aarch64_simd_ci __o;
17313   __o = __builtin_aarch64_ld1x3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17314   __i.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 0);
17315   __i.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 1);
17316   __i.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi  (__o, 2);
17317   return __i;
17318 }
17319
17320 __extension__ extern __inline poly64x1x3_t
17321 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17322 vld1_p64_x3 (const poly64_t *__a)
17323 {
17324   poly64x1x3_t __i;
17325   __builtin_aarch64_simd_ci __o;
17326   __o = __builtin_aarch64_ld1x3di ((const __builtin_aarch64_simd_di *) __a);
17327   __i.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17328   __i.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17329   __i.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17330
17331 return __i;
17332 }
17333
17334 __extension__ extern __inline uint8x16x3_t
17335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17336 vld1q_u8_x3 (const uint8_t *__a)
17337 {
17338   uint8x16x3_t __i;
17339   __builtin_aarch64_simd_ci __o;
17340   __o = __builtin_aarch64_ld1x3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17341   __i.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 0);
17342   __i.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 1);
17343   __i.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 2);
17344   return __i;
17345 }
17346
17347 __extension__ extern __inline int8x16x3_t
17348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17349 vld1q_s8_x3 (const int8_t *__a)
17350 {
17351   int8x16x3_t __i;
17352   __builtin_aarch64_simd_ci __o;
17353   __o = __builtin_aarch64_ld1x3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17354   __i.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 0);
17355   __i.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 1);
17356   __i.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 2);
17357   return __i;
17358 }
17359
17360 __extension__ extern __inline uint16x8x3_t
17361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17362 vld1q_u16_x3 (const uint16_t *__a)
17363 {
17364   uint16x8x3_t __i;
17365   __builtin_aarch64_simd_ci __o;
17366   __o = __builtin_aarch64_ld1x3v8hi ((const __builtin_aarch64_simd_hi *) __a);
17367   __i.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 0);
17368   __i.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 1);
17369   __i.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 2);
17370   return __i;
17371 }
17372
17373 __extension__ extern __inline int16x8x3_t
17374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17375 vld1q_s16_x3 (const int16_t *__a)
17376 {
17377   int16x8x3_t __i;
17378   __builtin_aarch64_simd_ci __o;
17379   __o = __builtin_aarch64_ld1x3v8hi ((const __builtin_aarch64_simd_hi *) __a);
17380   __i.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 0);
17381   __i.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 1);
17382   __i.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 2);
17383   return __i;
17384 }
17385
17386 __extension__ extern __inline uint32x4x3_t
17387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17388 vld1q_u32_x3 (const uint32_t *__a)
17389 {
17390   uint32x4x3_t __i;
17391   __builtin_aarch64_simd_ci __o;
17392   __o = __builtin_aarch64_ld1x3v4si ((const __builtin_aarch64_simd_si *) __a);
17393   __i.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si  (__o, 0);
17394   __i.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si  (__o, 1);
17395   __i.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si  (__o, 2);
17396   return __i;
17397 }
17398
17399 __extension__ extern __inline int32x4x3_t
17400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17401 vld1q_s32_x3 (const int32_t *__a)
17402 {
17403   int32x4x3_t __i;
17404   __builtin_aarch64_simd_ci __o;
17405   __o = __builtin_aarch64_ld1x3v4si ((const __builtin_aarch64_simd_si *) __a);
17406   __i.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si  (__o, 0);
17407   __i.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si  (__o, 1);
17408   __i.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si  (__o, 2);
17409   return __i;
17410 }
17411
17412 __extension__ extern __inline uint64x2x3_t
17413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17414 vld1q_u64_x3 (const uint64_t *__a)
17415 {
17416   uint64x2x3_t __i;
17417   __builtin_aarch64_simd_ci __o;
17418   __o = __builtin_aarch64_ld1x3v2di ((const __builtin_aarch64_simd_di *) __a);
17419   __i.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 0);
17420   __i.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 1);
17421   __i.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 2);
17422   return __i;
17423 }
17424
17425 __extension__ extern __inline int64x2x3_t
17426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17427 vld1q_s64_x3 (const int64_t *__a)
17428 {
17429   int64x2x3_t __i;
17430   __builtin_aarch64_simd_ci __o;
17431   __o = __builtin_aarch64_ld1x3v2di ((const __builtin_aarch64_simd_di *) __a);
17432   __i.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 0);
17433   __i.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 1);
17434   __i.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 2);
17435   return __i;
17436 }
17437
17438 __extension__ extern __inline float16x8x3_t
17439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17440 vld1q_f16_x3 (const float16_t *__a)
17441 {
17442   float16x8x3_t __i;
17443   __builtin_aarch64_simd_ci __o;
17444   __o = __builtin_aarch64_ld1x3v8hf ((const __builtin_aarch64_simd_hf *) __a);
17445   __i.val[0] = (float16x8_t) __builtin_aarch64_get_qregciv8hf  (__o, 0);
17446   __i.val[1] = (float16x8_t) __builtin_aarch64_get_qregciv8hf  (__o, 1);
17447   __i.val[2] = (float16x8_t) __builtin_aarch64_get_qregciv8hf  (__o, 2);
17448   return __i;
17449 }
17450
17451 __extension__ extern __inline float32x4x3_t
17452 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17453 vld1q_f32_x3 (const float32_t *__a)
17454 {
17455   float32x4x3_t __i;
17456   __builtin_aarch64_simd_ci __o;
17457   __o = __builtin_aarch64_ld1x3v4sf ((const __builtin_aarch64_simd_sf *) __a);
17458   __i.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf  (__o, 0);
17459   __i.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf  (__o, 1);
17460   __i.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf  (__o, 2);
17461   return __i;
17462 }
17463
17464 __extension__ extern __inline float64x2x3_t
17465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17466 vld1q_f64_x3 (const float64_t *__a)
17467 {
17468   float64x2x3_t __i;
17469   __builtin_aarch64_simd_ci __o;
17470   __o = __builtin_aarch64_ld1x3v2df ((const __builtin_aarch64_simd_df *) __a);
17471   __i.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df  (__o, 0);
17472   __i.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df  (__o, 1);
17473   __i.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df  (__o, 2);
17474   return __i;
17475 }
17476
17477 __extension__ extern __inline poly8x16x3_t
17478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17479 vld1q_p8_x3 (const poly8_t *__a)
17480 {
17481   poly8x16x3_t __i;
17482   __builtin_aarch64_simd_ci __o;
17483   __o = __builtin_aarch64_ld1x3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17484   __i.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 0);
17485   __i.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 1);
17486   __i.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi  (__o, 2);
17487   return __i;
17488 }
17489
17490 __extension__ extern __inline poly16x8x3_t
17491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17492 vld1q_p16_x3 (const poly16_t *__a)
17493 {
17494   poly16x8x3_t __i;
17495   __builtin_aarch64_simd_ci __o;
17496   __o = __builtin_aarch64_ld1x3v8hi ((const __builtin_aarch64_simd_hi *) __a);
17497   __i.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 0);
17498   __i.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 1);
17499   __i.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi  (__o, 2);
17500   return __i;
17501 }
17502
17503 __extension__ extern __inline poly64x2x3_t
17504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17505 vld1q_p64_x3 (const poly64_t *__a)
17506 {
17507   poly64x2x3_t __i;
17508   __builtin_aarch64_simd_ci __o;
17509   __o = __builtin_aarch64_ld1x3v2di ((const __builtin_aarch64_simd_di *) __a);
17510   __i.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 0);
17511   __i.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 1);
17512   __i.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di  (__o, 2);
17513   return __i;
17514 }
17515
17516 /* vld1q */
17517
17518 __extension__ extern __inline float16x8_t
17519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17520 vld1q_f16 (const float16_t *__a)
17521 {
17522   return __builtin_aarch64_ld1v8hf (__a);
17523 }
17524
17525 __extension__ extern __inline float32x4_t
17526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17527 vld1q_f32 (const float32_t *a)
17528 {
17529   return __builtin_aarch64_ld1v4sf ((const __builtin_aarch64_simd_sf *) a);
17530 }
17531
17532 __extension__ extern __inline float64x2_t
17533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17534 vld1q_f64 (const float64_t *a)
17535 {
17536   return __builtin_aarch64_ld1v2df ((const __builtin_aarch64_simd_df *) a);
17537 }
17538
17539 __extension__ extern __inline poly8x16_t
17540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17541 vld1q_p8 (const poly8_t *a)
17542 {
17543   return (poly8x16_t)
17544     __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
17545 }
17546
17547 __extension__ extern __inline poly16x8_t
17548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17549 vld1q_p16 (const poly16_t *a)
17550 {
17551   return (poly16x8_t)
17552     __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
17553 }
17554
17555 __extension__ extern __inline poly64x2_t
17556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17557 vld1q_p64 (const poly64_t *a)
17558 {
17559   return (poly64x2_t)
17560     __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
17561 }
17562
17563 __extension__ extern __inline int8x16_t
17564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17565 vld1q_s8 (const int8_t *a)
17566 {
17567   return __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
17568 }
17569
17570 __extension__ extern __inline int16x8_t
17571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17572 vld1q_s16 (const int16_t *a)
17573 {
17574   return __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
17575 }
17576
17577 __extension__ extern __inline int32x4_t
17578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17579 vld1q_s32 (const int32_t *a)
17580 {
17581   return __builtin_aarch64_ld1v4si ((const __builtin_aarch64_simd_si *) a);
17582 }
17583
17584 __extension__ extern __inline int64x2_t
17585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17586 vld1q_s64 (const int64_t *a)
17587 {
17588   return __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
17589 }
17590
17591 __extension__ extern __inline uint8x16_t
17592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17593 vld1q_u8 (const uint8_t *a)
17594 {
17595   return (uint8x16_t)
17596     __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
17597 }
17598
17599 __extension__ extern __inline uint8x8x2_t
17600 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17601 vld1_u8_x2 (const uint8_t *__a)
17602 {
17603   uint8x8x2_t ret;
17604   __builtin_aarch64_simd_oi __o;
17605   __o = __builtin_aarch64_ld1x2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17606   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17607   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17608   return ret;
17609 }
17610
17611 __extension__ extern __inline int8x8x2_t
17612 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17613 vld1_s8_x2 (const int8_t *__a)
17614 {
17615   int8x8x2_t ret;
17616   __builtin_aarch64_simd_oi __o;
17617   __o = __builtin_aarch64_ld1x2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17618   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17619   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17620   return ret;
17621 }
17622
17623 __extension__ extern __inline uint16x4x2_t
17624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17625 vld1_u16_x2 (const uint16_t *__a)
17626 {
17627   uint16x4x2_t ret;
17628   __builtin_aarch64_simd_oi __o;
17629   __o = __builtin_aarch64_ld1x2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17630   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17631   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17632   return ret;
17633 }
17634
17635 __extension__ extern __inline int16x4x2_t
17636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17637 vld1_s16_x2 (const int16_t *__a)
17638 {
17639   int16x4x2_t ret;
17640   __builtin_aarch64_simd_oi __o;
17641   __o = __builtin_aarch64_ld1x2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17642   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17643   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17644   return ret;
17645 }
17646
17647 __extension__ extern __inline uint32x2x2_t
17648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17649 vld1_u32_x2 (const uint32_t *__a)
17650 {
17651   uint32x2x2_t ret;
17652   __builtin_aarch64_simd_oi __o;
17653   __o = __builtin_aarch64_ld1x2v2si ((const __builtin_aarch64_simd_si *) __a);
17654   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
17655   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
17656   return ret;
17657 }
17658
17659 __extension__ extern __inline int32x2x2_t
17660 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17661 vld1_s32_x2 (const int32_t *__a)
17662 {
17663   int32x2x2_t ret;
17664   __builtin_aarch64_simd_oi __o;
17665   __o = __builtin_aarch64_ld1x2v2si ((const __builtin_aarch64_simd_si *) __a);
17666   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
17667   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
17668   return ret;
17669 }
17670
17671 __extension__ extern __inline uint64x1x2_t
17672 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17673 vld1_u64_x2 (const uint64_t *__a)
17674 {
17675   uint64x1x2_t ret;
17676   __builtin_aarch64_simd_oi __o;
17677   __o = __builtin_aarch64_ld1x2di ((const __builtin_aarch64_simd_di *) __a);
17678   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17679   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17680   return ret;
17681 }
17682
17683 __extension__ extern __inline int64x1x2_t
17684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17685 vld1_s64_x2 (const int64_t *__a)
17686 {
17687   int64x1x2_t ret;
17688   __builtin_aarch64_simd_oi __o;
17689   __o = __builtin_aarch64_ld1x2di ((const __builtin_aarch64_simd_di *) __a);
17690   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17691   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17692   return ret;
17693 }
17694
17695 __extension__ extern __inline float16x4x2_t
17696 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17697 vld1_f16_x2 (const float16_t *__a)
17698 {
17699   float16x4x2_t ret;
17700   __builtin_aarch64_simd_oi __o;
17701   __o = __builtin_aarch64_ld1x2v4hf ((const __builtin_aarch64_simd_hf *) __a);
17702   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregoiv4hf (__o, 0);
17703   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregoiv4hf (__o, 1);
17704   return ret;
17705 }
17706
17707 __extension__ extern __inline float32x2x2_t
17708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17709 vld1_f32_x2 (const float32_t *__a)
17710 {
17711   float32x2x2_t ret;
17712   __builtin_aarch64_simd_oi __o;
17713   __o = __builtin_aarch64_ld1x2v2sf ((const __builtin_aarch64_simd_sf *) __a);
17714   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
17715   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
17716   return ret;
17717 }
17718
17719 __extension__ extern __inline float64x1x2_t
17720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17721 vld1_f64_x2 (const float64_t *__a)
17722 {
17723   float64x1x2_t ret;
17724   __builtin_aarch64_simd_oi __o;
17725   __o = __builtin_aarch64_ld1x2df ((const __builtin_aarch64_simd_df *) __a);
17726   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
17727   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
17728   return ret;
17729 }
17730
17731 __extension__ extern __inline poly8x8x2_t
17732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17733 vld1_p8_x2 (const poly8_t *__a)
17734 {
17735   poly8x8x2_t ret;
17736   __builtin_aarch64_simd_oi __o;
17737   __o = __builtin_aarch64_ld1x2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17738   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17739   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17740   return ret;
17741 }
17742
17743 __extension__ extern __inline poly16x4x2_t
17744 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17745 vld1_p16_x2 (const poly16_t *__a)
17746 {
17747   poly16x4x2_t ret;
17748   __builtin_aarch64_simd_oi __o;
17749   __o = __builtin_aarch64_ld1x2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17750   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17751   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17752   return ret;
17753 }
17754
17755 __extension__ extern __inline poly64x1x2_t
17756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17757 vld1_p64_x2 (const poly64_t *__a)
17758 {
17759   poly64x1x2_t ret;
17760   __builtin_aarch64_simd_oi __o;
17761   __o = __builtin_aarch64_ld1x2di ((const __builtin_aarch64_simd_di *) __a);
17762   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17763   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17764   return ret;
17765 }
17766
17767 __extension__ extern __inline uint8x16x2_t
17768 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17769 vld1q_u8_x2 (const uint8_t *__a)
17770 {
17771   uint8x16x2_t ret;
17772   __builtin_aarch64_simd_oi __o;
17773   __o = __builtin_aarch64_ld1x2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17774   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17775   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17776   return ret;
17777 }
17778
17779 __extension__ extern __inline int8x16x2_t
17780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17781 vld1q_s8_x2 (const int8_t *__a)
17782 {
17783   int8x16x2_t ret;
17784   __builtin_aarch64_simd_oi __o;
17785   __o = __builtin_aarch64_ld1x2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17786   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17787   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17788   return ret;
17789 }
17790
17791 __extension__ extern __inline uint16x8x2_t
17792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17793 vld1q_u16_x2 (const uint16_t *__a)
17794 {
17795   uint16x8x2_t ret;
17796   __builtin_aarch64_simd_oi __o;
17797   __o = __builtin_aarch64_ld1x2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17798   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17799   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17800   return ret;
17801 }
17802
17803 __extension__ extern __inline int16x8x2_t
17804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17805 vld1q_s16_x2 (const int16_t *__a)
17806 {
17807   int16x8x2_t ret;
17808   __builtin_aarch64_simd_oi __o;
17809   __o = __builtin_aarch64_ld1x2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17810   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17811   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17812   return ret;
17813 }
17814
17815 __extension__ extern __inline uint32x4x2_t
17816 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17817 vld1q_u32_x2 (const uint32_t *__a)
17818 {
17819   uint32x4x2_t ret;
17820   __builtin_aarch64_simd_oi __o;
17821   __o = __builtin_aarch64_ld1x2v4si ((const __builtin_aarch64_simd_si *) __a);
17822   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
17823   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
17824   return ret;
17825 }
17826
17827 __extension__ extern __inline int32x4x2_t
17828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17829 vld1q_s32_x2 (const int32_t *__a)
17830 {
17831   int32x4x2_t ret;
17832   __builtin_aarch64_simd_oi __o;
17833   __o = __builtin_aarch64_ld1x2v4si ((const __builtin_aarch64_simd_si *) __a);
17834   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
17835   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
17836   return ret;
17837 }
17838
17839 __extension__ extern __inline uint64x2x2_t
17840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17841 vld1q_u64_x2 (const uint64_t *__a)
17842 {
17843   uint64x2x2_t ret;
17844   __builtin_aarch64_simd_oi __o;
17845   __o = __builtin_aarch64_ld1x2v2di ((const __builtin_aarch64_simd_di *) __a);
17846   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17847   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17848   return ret;
17849 }
17850
17851 __extension__ extern __inline int64x2x2_t
17852 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17853 vld1q_s64_x2 (const int64_t *__a)
17854 {
17855   int64x2x2_t ret;
17856   __builtin_aarch64_simd_oi __o;
17857   __o = __builtin_aarch64_ld1x2v2di ((const __builtin_aarch64_simd_di *) __a);
17858   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17859   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17860   return ret;
17861 }
17862
17863 __extension__ extern __inline float16x8x2_t
17864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17865 vld1q_f16_x2 (const float16_t *__a)
17866 {
17867   float16x8x2_t ret;
17868   __builtin_aarch64_simd_oi __o;
17869   __o = __builtin_aarch64_ld1x2v8hf ((const __builtin_aarch64_simd_hf *) __a);
17870   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregoiv8hf (__o, 0);
17871   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregoiv8hf (__o, 1);
17872   return ret;
17873 }
17874
17875 __extension__ extern __inline float32x4x2_t
17876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17877 vld1q_f32_x2 (const float32_t *__a)
17878 {
17879   float32x4x2_t ret;
17880   __builtin_aarch64_simd_oi __o;
17881   __o = __builtin_aarch64_ld1x2v4sf ((const __builtin_aarch64_simd_sf *) __a);
17882   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
17883   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
17884   return ret;
17885 }
17886
17887 __extension__ extern __inline float64x2x2_t
17888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17889 vld1q_f64_x2 (const float64_t *__a)
17890 {
17891   float64x2x2_t ret;
17892   __builtin_aarch64_simd_oi __o;
17893   __o = __builtin_aarch64_ld1x2v2df ((const __builtin_aarch64_simd_df *) __a);
17894   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
17895   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
17896   return ret;
17897 }
17898
17899 __extension__ extern __inline poly8x16x2_t
17900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17901 vld1q_p8_x2 (const poly8_t *__a)
17902 {
17903   poly8x16x2_t ret;
17904   __builtin_aarch64_simd_oi __o;
17905   __o = __builtin_aarch64_ld1x2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17906   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17907   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17908   return ret;
17909 }
17910
17911 __extension__ extern __inline poly16x8x2_t
17912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17913 vld1q_p16_x2 (const poly16_t *__a)
17914 {
17915   poly16x8x2_t ret;
17916   __builtin_aarch64_simd_oi __o;
17917   __o = __builtin_aarch64_ld1x2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17918   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17919   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17920   return ret;
17921 }
17922
17923 __extension__ extern __inline poly64x2x2_t
17924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17925 vld1q_p64_x2 (const poly64_t *__a)
17926 {
17927   poly64x2x2_t ret;
17928   __builtin_aarch64_simd_oi __o;
17929   __o = __builtin_aarch64_ld1x2v2di ((const __builtin_aarch64_simd_di *) __a);
17930   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17931   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17932   return ret;
17933 }
17934
17935 __extension__ extern __inline uint16x8_t
17936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17937 vld1q_u16 (const uint16_t *a)
17938 {
17939   return (uint16x8_t)
17940     __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
17941 }
17942
17943 __extension__ extern __inline uint32x4_t
17944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17945 vld1q_u32 (const uint32_t *a)
17946 {
17947   return (uint32x4_t)
17948     __builtin_aarch64_ld1v4si ((const __builtin_aarch64_simd_si *) a);
17949 }
17950
17951 __extension__ extern __inline uint64x2_t
17952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17953 vld1q_u64 (const uint64_t *a)
17954 {
17955   return (uint64x2_t)
17956     __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
17957 }
17958
17959 /* vld1_dup  */
17960
17961 __extension__ extern __inline float16x4_t
17962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17963 vld1_dup_f16 (const float16_t* __a)
17964 {
17965   return vdup_n_f16 (*__a);
17966 }
17967
17968 __extension__ extern __inline float32x2_t
17969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17970 vld1_dup_f32 (const float32_t* __a)
17971 {
17972   return vdup_n_f32 (*__a);
17973 }
17974
17975 __extension__ extern __inline float64x1_t
17976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17977 vld1_dup_f64 (const float64_t* __a)
17978 {
17979   return vdup_n_f64 (*__a);
17980 }
17981
17982 __extension__ extern __inline poly8x8_t
17983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17984 vld1_dup_p8 (const poly8_t* __a)
17985 {
17986   return vdup_n_p8 (*__a);
17987 }
17988
17989 __extension__ extern __inline poly16x4_t
17990 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17991 vld1_dup_p16 (const poly16_t* __a)
17992 {
17993   return vdup_n_p16 (*__a);
17994 }
17995
17996 __extension__ extern __inline poly64x1_t
17997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17998 vld1_dup_p64 (const poly64_t* __a)
17999 {
18000   return vdup_n_p64 (*__a);
18001 }
18002
18003 __extension__ extern __inline int8x8_t
18004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18005 vld1_dup_s8 (const int8_t* __a)
18006 {
18007   return vdup_n_s8 (*__a);
18008 }
18009
18010 __extension__ extern __inline int16x4_t
18011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18012 vld1_dup_s16 (const int16_t* __a)
18013 {
18014   return vdup_n_s16 (*__a);
18015 }
18016
18017 __extension__ extern __inline int32x2_t
18018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18019 vld1_dup_s32 (const int32_t* __a)
18020 {
18021   return vdup_n_s32 (*__a);
18022 }
18023
18024 __extension__ extern __inline int64x1_t
18025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18026 vld1_dup_s64 (const int64_t* __a)
18027 {
18028   return vdup_n_s64 (*__a);
18029 }
18030
18031 __extension__ extern __inline uint8x8_t
18032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18033 vld1_dup_u8 (const uint8_t* __a)
18034 {
18035   return vdup_n_u8 (*__a);
18036 }
18037
18038 __extension__ extern __inline uint16x4_t
18039 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18040 vld1_dup_u16 (const uint16_t* __a)
18041 {
18042   return vdup_n_u16 (*__a);
18043 }
18044
18045 __extension__ extern __inline uint32x2_t
18046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18047 vld1_dup_u32 (const uint32_t* __a)
18048 {
18049   return vdup_n_u32 (*__a);
18050 }
18051
18052 __extension__ extern __inline uint64x1_t
18053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18054 vld1_dup_u64 (const uint64_t* __a)
18055 {
18056   return vdup_n_u64 (*__a);
18057 }
18058
18059 /* vld1q_dup  */
18060
18061 __extension__ extern __inline float16x8_t
18062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18063 vld1q_dup_f16 (const float16_t* __a)
18064 {
18065   return vdupq_n_f16 (*__a);
18066 }
18067
18068 __extension__ extern __inline float32x4_t
18069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18070 vld1q_dup_f32 (const float32_t* __a)
18071 {
18072   return vdupq_n_f32 (*__a);
18073 }
18074
18075 __extension__ extern __inline float64x2_t
18076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18077 vld1q_dup_f64 (const float64_t* __a)
18078 {
18079   return vdupq_n_f64 (*__a);
18080 }
18081
18082 __extension__ extern __inline poly8x16_t
18083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18084 vld1q_dup_p8 (const poly8_t* __a)
18085 {
18086   return vdupq_n_p8 (*__a);
18087 }
18088
18089 __extension__ extern __inline poly16x8_t
18090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18091 vld1q_dup_p16 (const poly16_t* __a)
18092 {
18093   return vdupq_n_p16 (*__a);
18094 }
18095
18096 __extension__ extern __inline poly64x2_t
18097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18098 vld1q_dup_p64 (const poly64_t* __a)
18099 {
18100   return vdupq_n_p64 (*__a);
18101 }
18102
18103  __extension__ extern __inline int8x16_t
18104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18105 vld1q_dup_s8 (const int8_t* __a)
18106 {
18107   return vdupq_n_s8 (*__a);
18108 }
18109
18110 __extension__ extern __inline int16x8_t
18111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18112 vld1q_dup_s16 (const int16_t* __a)
18113 {
18114   return vdupq_n_s16 (*__a);
18115 }
18116
18117 __extension__ extern __inline int32x4_t
18118 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18119 vld1q_dup_s32 (const int32_t* __a)
18120 {
18121   return vdupq_n_s32 (*__a);
18122 }
18123
18124 __extension__ extern __inline int64x2_t
18125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18126 vld1q_dup_s64 (const int64_t* __a)
18127 {
18128   return vdupq_n_s64 (*__a);
18129 }
18130
18131 __extension__ extern __inline uint8x16_t
18132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18133 vld1q_dup_u8 (const uint8_t* __a)
18134 {
18135   return vdupq_n_u8 (*__a);
18136 }
18137
18138 __extension__ extern __inline uint16x8_t
18139 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18140 vld1q_dup_u16 (const uint16_t* __a)
18141 {
18142   return vdupq_n_u16 (*__a);
18143 }
18144
18145 __extension__ extern __inline uint32x4_t
18146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18147 vld1q_dup_u32 (const uint32_t* __a)
18148 {
18149   return vdupq_n_u32 (*__a);
18150 }
18151
18152 __extension__ extern __inline uint64x2_t
18153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18154 vld1q_dup_u64 (const uint64_t* __a)
18155 {
18156   return vdupq_n_u64 (*__a);
18157 }
18158
18159 /* vld1_lane  */
18160
18161 __extension__ extern __inline float16x4_t
18162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18163 vld1_lane_f16 (const float16_t *__src, float16x4_t __vec, const int __lane)
18164 {
18165   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18166 }
18167
18168 __extension__ extern __inline float32x2_t
18169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18170 vld1_lane_f32 (const float32_t *__src, float32x2_t __vec, const int __lane)
18171 {
18172   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18173 }
18174
18175 __extension__ extern __inline float64x1_t
18176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18177 vld1_lane_f64 (const float64_t *__src, float64x1_t __vec, const int __lane)
18178 {
18179   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18180 }
18181
18182 __extension__ extern __inline poly8x8_t
18183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18184 vld1_lane_p8 (const poly8_t *__src, poly8x8_t __vec, const int __lane)
18185 {
18186   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18187 }
18188
18189 __extension__ extern __inline poly16x4_t
18190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18191 vld1_lane_p16 (const poly16_t *__src, poly16x4_t __vec, const int __lane)
18192 {
18193   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18194 }
18195
18196 __extension__ extern __inline poly64x1_t
18197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18198 vld1_lane_p64 (const poly64_t *__src, poly64x1_t __vec, const int __lane)
18199 {
18200   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18201 }
18202
18203 __extension__ extern __inline int8x8_t
18204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18205 vld1_lane_s8 (const int8_t *__src, int8x8_t __vec, const int __lane)
18206 {
18207   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18208 }
18209
18210 __extension__ extern __inline int16x4_t
18211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18212 vld1_lane_s16 (const int16_t *__src, int16x4_t __vec, const int __lane)
18213 {
18214   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18215 }
18216
18217 __extension__ extern __inline int32x2_t
18218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18219 vld1_lane_s32 (const int32_t *__src, int32x2_t __vec, const int __lane)
18220 {
18221   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18222 }
18223
18224 __extension__ extern __inline int64x1_t
18225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18226 vld1_lane_s64 (const int64_t *__src, int64x1_t __vec, const int __lane)
18227 {
18228   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18229 }
18230
18231 __extension__ extern __inline uint8x8_t
18232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18233 vld1_lane_u8 (const uint8_t *__src, uint8x8_t __vec, const int __lane)
18234 {
18235   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18236 }
18237
18238 __extension__ extern __inline uint16x4_t
18239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18240 vld1_lane_u16 (const uint16_t *__src, uint16x4_t __vec, const int __lane)
18241 {
18242   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18243 }
18244
18245 __extension__ extern __inline uint32x2_t
18246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18247 vld1_lane_u32 (const uint32_t *__src, uint32x2_t __vec, const int __lane)
18248 {
18249   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18250 }
18251
18252 __extension__ extern __inline uint64x1_t
18253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18254 vld1_lane_u64 (const uint64_t *__src, uint64x1_t __vec, const int __lane)
18255 {
18256   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18257 }
18258
18259 /* vld1q_lane  */
18260
18261 __extension__ extern __inline float16x8_t
18262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18263 vld1q_lane_f16 (const float16_t *__src, float16x8_t __vec, const int __lane)
18264 {
18265   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18266 }
18267
18268 __extension__ extern __inline float32x4_t
18269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18270 vld1q_lane_f32 (const float32_t *__src, float32x4_t __vec, const int __lane)
18271 {
18272   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18273 }
18274
18275 __extension__ extern __inline float64x2_t
18276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18277 vld1q_lane_f64 (const float64_t *__src, float64x2_t __vec, const int __lane)
18278 {
18279   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18280 }
18281
18282 __extension__ extern __inline poly8x16_t
18283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18284 vld1q_lane_p8 (const poly8_t *__src, poly8x16_t __vec, const int __lane)
18285 {
18286   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18287 }
18288
18289 __extension__ extern __inline poly16x8_t
18290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18291 vld1q_lane_p16 (const poly16_t *__src, poly16x8_t __vec, const int __lane)
18292 {
18293   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18294 }
18295
18296 __extension__ extern __inline poly64x2_t
18297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18298 vld1q_lane_p64 (const poly64_t *__src, poly64x2_t __vec, const int __lane)
18299 {
18300   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18301 }
18302
18303 __extension__ extern __inline int8x16_t
18304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18305 vld1q_lane_s8 (const int8_t *__src, int8x16_t __vec, const int __lane)
18306 {
18307   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18308 }
18309
18310 __extension__ extern __inline int16x8_t
18311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18312 vld1q_lane_s16 (const int16_t *__src, int16x8_t __vec, const int __lane)
18313 {
18314   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18315 }
18316
18317 __extension__ extern __inline int32x4_t
18318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18319 vld1q_lane_s32 (const int32_t *__src, int32x4_t __vec, const int __lane)
18320 {
18321   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18322 }
18323
18324 __extension__ extern __inline int64x2_t
18325 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18326 vld1q_lane_s64 (const int64_t *__src, int64x2_t __vec, const int __lane)
18327 {
18328   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18329 }
18330
18331 __extension__ extern __inline uint8x16_t
18332 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18333 vld1q_lane_u8 (const uint8_t *__src, uint8x16_t __vec, const int __lane)
18334 {
18335   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18336 }
18337
18338 __extension__ extern __inline uint16x8_t
18339 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18340 vld1q_lane_u16 (const uint16_t *__src, uint16x8_t __vec, const int __lane)
18341 {
18342   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18343 }
18344
18345 __extension__ extern __inline uint32x4_t
18346 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18347 vld1q_lane_u32 (const uint32_t *__src, uint32x4_t __vec, const int __lane)
18348 {
18349   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18350 }
18351
18352 __extension__ extern __inline uint64x2_t
18353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18354 vld1q_lane_u64 (const uint64_t *__src, uint64x2_t __vec, const int __lane)
18355 {
18356   return __aarch64_vset_lane_any (*__src, __vec, __lane);
18357 }
18358
18359 /* vldn */
18360
18361 __extension__ extern __inline int64x1x2_t
18362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18363 vld2_s64 (const int64_t * __a)
18364 {
18365   int64x1x2_t ret;
18366   __builtin_aarch64_simd_oi __o;
18367   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
18368   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
18369   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
18370   return ret;
18371 }
18372
18373 __extension__ extern __inline uint64x1x2_t
18374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18375 vld2_u64 (const uint64_t * __a)
18376 {
18377   uint64x1x2_t ret;
18378   __builtin_aarch64_simd_oi __o;
18379   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
18380   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
18381   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
18382   return ret;
18383 }
18384
18385 __extension__ extern __inline float64x1x2_t
18386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18387 vld2_f64 (const float64_t * __a)
18388 {
18389   float64x1x2_t ret;
18390   __builtin_aarch64_simd_oi __o;
18391   __o = __builtin_aarch64_ld2df ((const __builtin_aarch64_simd_df *) __a);
18392   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
18393   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
18394   return ret;
18395 }
18396
18397 __extension__ extern __inline int8x8x2_t
18398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18399 vld2_s8 (const int8_t * __a)
18400 {
18401   int8x8x2_t ret;
18402   __builtin_aarch64_simd_oi __o;
18403   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
18404   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18405   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18406   return ret;
18407 }
18408
18409 __extension__ extern __inline poly8x8x2_t
18410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18411 vld2_p8 (const poly8_t * __a)
18412 {
18413   poly8x8x2_t ret;
18414   __builtin_aarch64_simd_oi __o;
18415   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
18416   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18417   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18418   return ret;
18419 }
18420
18421 __extension__ extern __inline poly64x1x2_t
18422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18423 vld2_p64 (const poly64_t * __a)
18424 {
18425   poly64x1x2_t ret;
18426   __builtin_aarch64_simd_oi __o;
18427   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
18428   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 0);
18429   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 1);
18430   return ret;
18431 }
18432
18433 __extension__ extern __inline int16x4x2_t
18434 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18435 vld2_s16 (const int16_t * __a)
18436 {
18437   int16x4x2_t ret;
18438   __builtin_aarch64_simd_oi __o;
18439   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
18440   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18441   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18442   return ret;
18443 }
18444
18445 __extension__ extern __inline poly16x4x2_t
18446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18447 vld2_p16 (const poly16_t * __a)
18448 {
18449   poly16x4x2_t ret;
18450   __builtin_aarch64_simd_oi __o;
18451   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
18452   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18453   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18454   return ret;
18455 }
18456
18457 __extension__ extern __inline int32x2x2_t
18458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18459 vld2_s32 (const int32_t * __a)
18460 {
18461   int32x2x2_t ret;
18462   __builtin_aarch64_simd_oi __o;
18463   __o = __builtin_aarch64_ld2v2si ((const __builtin_aarch64_simd_si *) __a);
18464   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
18465   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
18466   return ret;
18467 }
18468
18469 __extension__ extern __inline uint8x8x2_t
18470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18471 vld2_u8 (const uint8_t * __a)
18472 {
18473   uint8x8x2_t ret;
18474   __builtin_aarch64_simd_oi __o;
18475   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
18476   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18477   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18478   return ret;
18479 }
18480
18481 __extension__ extern __inline uint16x4x2_t
18482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18483 vld2_u16 (const uint16_t * __a)
18484 {
18485   uint16x4x2_t ret;
18486   __builtin_aarch64_simd_oi __o;
18487   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
18488   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18489   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18490   return ret;
18491 }
18492
18493 __extension__ extern __inline uint32x2x2_t
18494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18495 vld2_u32 (const uint32_t * __a)
18496 {
18497   uint32x2x2_t ret;
18498   __builtin_aarch64_simd_oi __o;
18499   __o = __builtin_aarch64_ld2v2si ((const __builtin_aarch64_simd_si *) __a);
18500   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
18501   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
18502   return ret;
18503 }
18504
18505 __extension__ extern __inline float16x4x2_t
18506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18507 vld2_f16 (const float16_t * __a)
18508 {
18509   float16x4x2_t ret;
18510   __builtin_aarch64_simd_oi __o;
18511   __o = __builtin_aarch64_ld2v4hf (__a);
18512   ret.val[0] = __builtin_aarch64_get_dregoiv4hf (__o, 0);
18513   ret.val[1] = __builtin_aarch64_get_dregoiv4hf (__o, 1);
18514   return ret;
18515 }
18516
18517 __extension__ extern __inline float32x2x2_t
18518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18519 vld2_f32 (const float32_t * __a)
18520 {
18521   float32x2x2_t ret;
18522   __builtin_aarch64_simd_oi __o;
18523   __o = __builtin_aarch64_ld2v2sf ((const __builtin_aarch64_simd_sf *) __a);
18524   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
18525   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
18526   return ret;
18527 }
18528
18529 __extension__ extern __inline int8x16x2_t
18530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18531 vld2q_s8 (const int8_t * __a)
18532 {
18533   int8x16x2_t ret;
18534   __builtin_aarch64_simd_oi __o;
18535   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
18536   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18537   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18538   return ret;
18539 }
18540
18541 __extension__ extern __inline poly8x16x2_t
18542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18543 vld2q_p8 (const poly8_t * __a)
18544 {
18545   poly8x16x2_t ret;
18546   __builtin_aarch64_simd_oi __o;
18547   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
18548   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18549   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18550   return ret;
18551 }
18552
18553 __extension__ extern __inline int16x8x2_t
18554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18555 vld2q_s16 (const int16_t * __a)
18556 {
18557   int16x8x2_t ret;
18558   __builtin_aarch64_simd_oi __o;
18559   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
18560   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18561   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18562   return ret;
18563 }
18564
18565 __extension__ extern __inline poly16x8x2_t
18566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18567 vld2q_p16 (const poly16_t * __a)
18568 {
18569   poly16x8x2_t ret;
18570   __builtin_aarch64_simd_oi __o;
18571   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
18572   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18573   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18574   return ret;
18575 }
18576
18577 __extension__ extern __inline poly64x2x2_t
18578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18579 vld2q_p64 (const poly64_t * __a)
18580 {
18581   poly64x2x2_t ret;
18582   __builtin_aarch64_simd_oi __o;
18583   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
18584   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 0);
18585   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 1);
18586   return ret;
18587 }
18588
18589 __extension__ extern __inline int32x4x2_t
18590 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18591 vld2q_s32 (const int32_t * __a)
18592 {
18593   int32x4x2_t ret;
18594   __builtin_aarch64_simd_oi __o;
18595   __o = __builtin_aarch64_ld2v4si ((const __builtin_aarch64_simd_si *) __a);
18596   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
18597   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
18598   return ret;
18599 }
18600
18601 __extension__ extern __inline int64x2x2_t
18602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18603 vld2q_s64 (const int64_t * __a)
18604 {
18605   int64x2x2_t ret;
18606   __builtin_aarch64_simd_oi __o;
18607   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
18608   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
18609   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
18610   return ret;
18611 }
18612
18613 __extension__ extern __inline uint8x16x2_t
18614 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18615 vld2q_u8 (const uint8_t * __a)
18616 {
18617   uint8x16x2_t ret;
18618   __builtin_aarch64_simd_oi __o;
18619   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
18620   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18621   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18622   return ret;
18623 }
18624
18625 __extension__ extern __inline uint16x8x2_t
18626 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18627 vld2q_u16 (const uint16_t * __a)
18628 {
18629   uint16x8x2_t ret;
18630   __builtin_aarch64_simd_oi __o;
18631   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
18632   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18633   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18634   return ret;
18635 }
18636
18637 __extension__ extern __inline uint32x4x2_t
18638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18639 vld2q_u32 (const uint32_t * __a)
18640 {
18641   uint32x4x2_t ret;
18642   __builtin_aarch64_simd_oi __o;
18643   __o = __builtin_aarch64_ld2v4si ((const __builtin_aarch64_simd_si *) __a);
18644   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
18645   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
18646   return ret;
18647 }
18648
18649 __extension__ extern __inline uint64x2x2_t
18650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18651 vld2q_u64 (const uint64_t * __a)
18652 {
18653   uint64x2x2_t ret;
18654   __builtin_aarch64_simd_oi __o;
18655   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
18656   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
18657   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
18658   return ret;
18659 }
18660
18661 __extension__ extern __inline float16x8x2_t
18662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18663 vld2q_f16 (const float16_t * __a)
18664 {
18665   float16x8x2_t ret;
18666   __builtin_aarch64_simd_oi __o;
18667   __o = __builtin_aarch64_ld2v8hf (__a);
18668   ret.val[0] = __builtin_aarch64_get_qregoiv8hf (__o, 0);
18669   ret.val[1] = __builtin_aarch64_get_qregoiv8hf (__o, 1);
18670   return ret;
18671 }
18672
18673 __extension__ extern __inline float32x4x2_t
18674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18675 vld2q_f32 (const float32_t * __a)
18676 {
18677   float32x4x2_t ret;
18678   __builtin_aarch64_simd_oi __o;
18679   __o = __builtin_aarch64_ld2v4sf ((const __builtin_aarch64_simd_sf *) __a);
18680   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
18681   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
18682   return ret;
18683 }
18684
18685 __extension__ extern __inline float64x2x2_t
18686 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18687 vld2q_f64 (const float64_t * __a)
18688 {
18689   float64x2x2_t ret;
18690   __builtin_aarch64_simd_oi __o;
18691   __o = __builtin_aarch64_ld2v2df ((const __builtin_aarch64_simd_df *) __a);
18692   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
18693   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
18694   return ret;
18695 }
18696
18697 __extension__ extern __inline int64x1x3_t
18698 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18699 vld3_s64 (const int64_t * __a)
18700 {
18701   int64x1x3_t ret;
18702   __builtin_aarch64_simd_ci __o;
18703   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
18704   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
18705   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
18706   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
18707   return ret;
18708 }
18709
18710 __extension__ extern __inline uint64x1x3_t
18711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18712 vld3_u64 (const uint64_t * __a)
18713 {
18714   uint64x1x3_t ret;
18715   __builtin_aarch64_simd_ci __o;
18716   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
18717   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
18718   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
18719   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
18720   return ret;
18721 }
18722
18723 __extension__ extern __inline float64x1x3_t
18724 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18725 vld3_f64 (const float64_t * __a)
18726 {
18727   float64x1x3_t ret;
18728   __builtin_aarch64_simd_ci __o;
18729   __o = __builtin_aarch64_ld3df ((const __builtin_aarch64_simd_df *) __a);
18730   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 0)};
18731   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 1)};
18732   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 2)};
18733   return ret;
18734 }
18735
18736 __extension__ extern __inline int8x8x3_t
18737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18738 vld3_s8 (const int8_t * __a)
18739 {
18740   int8x8x3_t ret;
18741   __builtin_aarch64_simd_ci __o;
18742   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
18743   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18744   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18745   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18746   return ret;
18747 }
18748
18749 __extension__ extern __inline poly8x8x3_t
18750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18751 vld3_p8 (const poly8_t * __a)
18752 {
18753   poly8x8x3_t ret;
18754   __builtin_aarch64_simd_ci __o;
18755   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
18756   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18757   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18758   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18759   return ret;
18760 }
18761
18762 __extension__ extern __inline int16x4x3_t
18763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18764 vld3_s16 (const int16_t * __a)
18765 {
18766   int16x4x3_t ret;
18767   __builtin_aarch64_simd_ci __o;
18768   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
18769   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18770   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18771   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18772   return ret;
18773 }
18774
18775 __extension__ extern __inline poly16x4x3_t
18776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18777 vld3_p16 (const poly16_t * __a)
18778 {
18779   poly16x4x3_t ret;
18780   __builtin_aarch64_simd_ci __o;
18781   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
18782   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18783   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18784   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18785   return ret;
18786 }
18787
18788 __extension__ extern __inline int32x2x3_t
18789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18790 vld3_s32 (const int32_t * __a)
18791 {
18792   int32x2x3_t ret;
18793   __builtin_aarch64_simd_ci __o;
18794   __o = __builtin_aarch64_ld3v2si ((const __builtin_aarch64_simd_si *) __a);
18795   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
18796   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
18797   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
18798   return ret;
18799 }
18800
18801 __extension__ extern __inline uint8x8x3_t
18802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18803 vld3_u8 (const uint8_t * __a)
18804 {
18805   uint8x8x3_t ret;
18806   __builtin_aarch64_simd_ci __o;
18807   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
18808   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18809   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18810   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18811   return ret;
18812 }
18813
18814 __extension__ extern __inline uint16x4x3_t
18815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18816 vld3_u16 (const uint16_t * __a)
18817 {
18818   uint16x4x3_t ret;
18819   __builtin_aarch64_simd_ci __o;
18820   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
18821   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18822   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18823   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18824   return ret;
18825 }
18826
18827 __extension__ extern __inline uint32x2x3_t
18828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18829 vld3_u32 (const uint32_t * __a)
18830 {
18831   uint32x2x3_t ret;
18832   __builtin_aarch64_simd_ci __o;
18833   __o = __builtin_aarch64_ld3v2si ((const __builtin_aarch64_simd_si *) __a);
18834   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
18835   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
18836   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
18837   return ret;
18838 }
18839
18840 __extension__ extern __inline float16x4x3_t
18841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18842 vld3_f16 (const float16_t * __a)
18843 {
18844   float16x4x3_t ret;
18845   __builtin_aarch64_simd_ci __o;
18846   __o = __builtin_aarch64_ld3v4hf (__a);
18847   ret.val[0] = __builtin_aarch64_get_dregciv4hf (__o, 0);
18848   ret.val[1] = __builtin_aarch64_get_dregciv4hf (__o, 1);
18849   ret.val[2] = __builtin_aarch64_get_dregciv4hf (__o, 2);
18850   return ret;
18851 }
18852
18853 __extension__ extern __inline float32x2x3_t
18854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18855 vld3_f32 (const float32_t * __a)
18856 {
18857   float32x2x3_t ret;
18858   __builtin_aarch64_simd_ci __o;
18859   __o = __builtin_aarch64_ld3v2sf ((const __builtin_aarch64_simd_sf *) __a);
18860   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 0);
18861   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 1);
18862   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 2);
18863   return ret;
18864 }
18865
18866 __extension__ extern __inline poly64x1x3_t
18867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18868 vld3_p64 (const poly64_t * __a)
18869 {
18870   poly64x1x3_t ret;
18871   __builtin_aarch64_simd_ci __o;
18872   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
18873   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 0);
18874   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 1);
18875   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 2);
18876   return ret;
18877 }
18878
18879 __extension__ extern __inline int8x16x3_t
18880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18881 vld3q_s8 (const int8_t * __a)
18882 {
18883   int8x16x3_t ret;
18884   __builtin_aarch64_simd_ci __o;
18885   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
18886   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
18887   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
18888   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
18889   return ret;
18890 }
18891
18892 __extension__ extern __inline poly8x16x3_t
18893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18894 vld3q_p8 (const poly8_t * __a)
18895 {
18896   poly8x16x3_t ret;
18897   __builtin_aarch64_simd_ci __o;
18898   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
18899   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
18900   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
18901   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
18902   return ret;
18903 }
18904
18905 __extension__ extern __inline int16x8x3_t
18906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18907 vld3q_s16 (const int16_t * __a)
18908 {
18909   int16x8x3_t ret;
18910   __builtin_aarch64_simd_ci __o;
18911   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18912   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18913   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18914   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18915   return ret;
18916 }
18917
18918 __extension__ extern __inline poly16x8x3_t
18919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18920 vld3q_p16 (const poly16_t * __a)
18921 {
18922   poly16x8x3_t ret;
18923   __builtin_aarch64_simd_ci __o;
18924   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18925   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18926   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18927   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18928   return ret;
18929 }
18930
18931 __extension__ extern __inline int32x4x3_t
18932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18933 vld3q_s32 (const int32_t * __a)
18934 {
18935   int32x4x3_t ret;
18936   __builtin_aarch64_simd_ci __o;
18937   __o = __builtin_aarch64_ld3v4si ((const __builtin_aarch64_simd_si *) __a);
18938   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
18939   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
18940   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
18941   return ret;
18942 }
18943
18944 __extension__ extern __inline int64x2x3_t
18945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18946 vld3q_s64 (const int64_t * __a)
18947 {
18948   int64x2x3_t ret;
18949   __builtin_aarch64_simd_ci __o;
18950   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
18951   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
18952   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
18953   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
18954   return ret;
18955 }
18956
18957 __extension__ extern __inline uint8x16x3_t
18958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18959 vld3q_u8 (const uint8_t * __a)
18960 {
18961   uint8x16x3_t ret;
18962   __builtin_aarch64_simd_ci __o;
18963   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
18964   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
18965   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
18966   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
18967   return ret;
18968 }
18969
18970 __extension__ extern __inline uint16x8x3_t
18971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18972 vld3q_u16 (const uint16_t * __a)
18973 {
18974   uint16x8x3_t ret;
18975   __builtin_aarch64_simd_ci __o;
18976   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18977   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18978   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18979   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18980   return ret;
18981 }
18982
18983 __extension__ extern __inline uint32x4x3_t
18984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18985 vld3q_u32 (const uint32_t * __a)
18986 {
18987   uint32x4x3_t ret;
18988   __builtin_aarch64_simd_ci __o;
18989   __o = __builtin_aarch64_ld3v4si ((const __builtin_aarch64_simd_si *) __a);
18990   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
18991   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
18992   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
18993   return ret;
18994 }
18995
18996 __extension__ extern __inline uint64x2x3_t
18997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18998 vld3q_u64 (const uint64_t * __a)
18999 {
19000   uint64x2x3_t ret;
19001   __builtin_aarch64_simd_ci __o;
19002   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
19003   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
19004   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
19005   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
19006   return ret;
19007 }
19008
19009 __extension__ extern __inline float16x8x3_t
19010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19011 vld3q_f16 (const float16_t * __a)
19012 {
19013   float16x8x3_t ret;
19014   __builtin_aarch64_simd_ci __o;
19015   __o = __builtin_aarch64_ld3v8hf (__a);
19016   ret.val[0] = __builtin_aarch64_get_qregciv8hf (__o, 0);
19017   ret.val[1] = __builtin_aarch64_get_qregciv8hf (__o, 1);
19018   ret.val[2] = __builtin_aarch64_get_qregciv8hf (__o, 2);
19019   return ret;
19020 }
19021
19022 __extension__ extern __inline float32x4x3_t
19023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19024 vld3q_f32 (const float32_t * __a)
19025 {
19026   float32x4x3_t ret;
19027   __builtin_aarch64_simd_ci __o;
19028   __o = __builtin_aarch64_ld3v4sf ((const __builtin_aarch64_simd_sf *) __a);
19029   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 0);
19030   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 1);
19031   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 2);
19032   return ret;
19033 }
19034
19035 __extension__ extern __inline float64x2x3_t
19036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19037 vld3q_f64 (const float64_t * __a)
19038 {
19039   float64x2x3_t ret;
19040   __builtin_aarch64_simd_ci __o;
19041   __o = __builtin_aarch64_ld3v2df ((const __builtin_aarch64_simd_df *) __a);
19042   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 0);
19043   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 1);
19044   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 2);
19045   return ret;
19046 }
19047
19048 __extension__ extern __inline poly64x2x3_t
19049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19050 vld3q_p64 (const poly64_t * __a)
19051 {
19052   poly64x2x3_t ret;
19053   __builtin_aarch64_simd_ci __o;
19054   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
19055   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 0);
19056   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 1);
19057   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 2);
19058   return ret;
19059 }
19060
19061 __extension__ extern __inline int64x1x4_t
19062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19063 vld4_s64 (const int64_t * __a)
19064 {
19065   int64x1x4_t ret;
19066   __builtin_aarch64_simd_xi __o;
19067   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
19068   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
19069   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
19070   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
19071   ret.val[3] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
19072   return ret;
19073 }
19074
19075 __extension__ extern __inline uint64x1x4_t
19076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19077 vld4_u64 (const uint64_t * __a)
19078 {
19079   uint64x1x4_t ret;
19080   __builtin_aarch64_simd_xi __o;
19081   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
19082   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
19083   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
19084   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
19085   ret.val[3] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
19086   return ret;
19087 }
19088
19089 __extension__ extern __inline float64x1x4_t
19090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19091 vld4_f64 (const float64_t * __a)
19092 {
19093   float64x1x4_t ret;
19094   __builtin_aarch64_simd_xi __o;
19095   __o = __builtin_aarch64_ld4df ((const __builtin_aarch64_simd_df *) __a);
19096   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 0)};
19097   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 1)};
19098   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 2)};
19099   ret.val[3] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 3)};
19100   return ret;
19101 }
19102
19103 __extension__ extern __inline int8x8x4_t
19104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19105 vld4_s8 (const int8_t * __a)
19106 {
19107   int8x8x4_t ret;
19108   __builtin_aarch64_simd_xi __o;
19109   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
19110   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19111   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19112   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19113   ret.val[3] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19114   return ret;
19115 }
19116
19117 __extension__ extern __inline poly8x8x4_t
19118 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19119 vld4_p8 (const poly8_t * __a)
19120 {
19121   poly8x8x4_t ret;
19122   __builtin_aarch64_simd_xi __o;
19123   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
19124   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19125   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19126   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19127   ret.val[3] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19128   return ret;
19129 }
19130
19131 __extension__ extern __inline int16x4x4_t
19132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19133 vld4_s16 (const int16_t * __a)
19134 {
19135   int16x4x4_t ret;
19136   __builtin_aarch64_simd_xi __o;
19137   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
19138   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19139   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19140   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19141   ret.val[3] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19142   return ret;
19143 }
19144
19145 __extension__ extern __inline poly16x4x4_t
19146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19147 vld4_p16 (const poly16_t * __a)
19148 {
19149   poly16x4x4_t ret;
19150   __builtin_aarch64_simd_xi __o;
19151   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
19152   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19153   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19154   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19155   ret.val[3] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19156   return ret;
19157 }
19158
19159 __extension__ extern __inline int32x2x4_t
19160 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19161 vld4_s32 (const int32_t * __a)
19162 {
19163   int32x2x4_t ret;
19164   __builtin_aarch64_simd_xi __o;
19165   __o = __builtin_aarch64_ld4v2si ((const __builtin_aarch64_simd_si *) __a);
19166   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
19167   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
19168   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
19169   ret.val[3] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
19170   return ret;
19171 }
19172
19173 __extension__ extern __inline uint8x8x4_t
19174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19175 vld4_u8 (const uint8_t * __a)
19176 {
19177   uint8x8x4_t ret;
19178   __builtin_aarch64_simd_xi __o;
19179   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
19180   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19181   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19182   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19183   ret.val[3] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19184   return ret;
19185 }
19186
19187 __extension__ extern __inline uint16x4x4_t
19188 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19189 vld4_u16 (const uint16_t * __a)
19190 {
19191   uint16x4x4_t ret;
19192   __builtin_aarch64_simd_xi __o;
19193   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
19194   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19195   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19196   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19197   ret.val[3] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19198   return ret;
19199 }
19200
19201 __extension__ extern __inline uint32x2x4_t
19202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19203 vld4_u32 (const uint32_t * __a)
19204 {
19205   uint32x2x4_t ret;
19206   __builtin_aarch64_simd_xi __o;
19207   __o = __builtin_aarch64_ld4v2si ((const __builtin_aarch64_simd_si *) __a);
19208   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
19209   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
19210   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
19211   ret.val[3] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
19212   return ret;
19213 }
19214
19215 __extension__ extern __inline float16x4x4_t
19216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19217 vld4_f16 (const float16_t * __a)
19218 {
19219   float16x4x4_t ret;
19220   __builtin_aarch64_simd_xi __o;
19221   __o = __builtin_aarch64_ld4v4hf (__a);
19222   ret.val[0] = __builtin_aarch64_get_dregxiv4hf (__o, 0);
19223   ret.val[1] = __builtin_aarch64_get_dregxiv4hf (__o, 1);
19224   ret.val[2] = __builtin_aarch64_get_dregxiv4hf (__o, 2);
19225   ret.val[3] = __builtin_aarch64_get_dregxiv4hf (__o, 3);
19226   return ret;
19227 }
19228
19229 __extension__ extern __inline float32x2x4_t
19230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19231 vld4_f32 (const float32_t * __a)
19232 {
19233   float32x2x4_t ret;
19234   __builtin_aarch64_simd_xi __o;
19235   __o = __builtin_aarch64_ld4v2sf ((const __builtin_aarch64_simd_sf *) __a);
19236   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 0);
19237   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 1);
19238   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 2);
19239   ret.val[3] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 3);
19240   return ret;
19241 }
19242
19243 __extension__ extern __inline poly64x1x4_t
19244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19245 vld4_p64 (const poly64_t * __a)
19246 {
19247   poly64x1x4_t  ret;
19248   __builtin_aarch64_simd_xi __o;
19249   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
19250   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 0);
19251   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 1);
19252   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 2);
19253   ret.val[3] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 3);
19254   return ret;
19255 }
19256
19257 __extension__ extern __inline int8x16x4_t
19258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19259 vld4q_s8 (const int8_t * __a)
19260 {
19261   int8x16x4_t ret;
19262   __builtin_aarch64_simd_xi __o;
19263   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
19264   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19265   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19266   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19267   ret.val[3] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19268   return ret;
19269 }
19270
19271 __extension__ extern __inline poly8x16x4_t
19272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19273 vld4q_p8 (const poly8_t * __a)
19274 {
19275   poly8x16x4_t ret;
19276   __builtin_aarch64_simd_xi __o;
19277   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
19278   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19279   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19280   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19281   ret.val[3] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19282   return ret;
19283 }
19284
19285 __extension__ extern __inline int16x8x4_t
19286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19287 vld4q_s16 (const int16_t * __a)
19288 {
19289   int16x8x4_t ret;
19290   __builtin_aarch64_simd_xi __o;
19291   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
19292   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19293   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19294   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19295   ret.val[3] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19296   return ret;
19297 }
19298
19299 __extension__ extern __inline poly16x8x4_t
19300 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19301 vld4q_p16 (const poly16_t * __a)
19302 {
19303   poly16x8x4_t ret;
19304   __builtin_aarch64_simd_xi __o;
19305   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
19306   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19307   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19308   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19309   ret.val[3] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19310   return ret;
19311 }
19312
19313 __extension__ extern __inline int32x4x4_t
19314 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19315 vld4q_s32 (const int32_t * __a)
19316 {
19317   int32x4x4_t ret;
19318   __builtin_aarch64_simd_xi __o;
19319   __o = __builtin_aarch64_ld4v4si ((const __builtin_aarch64_simd_si *) __a);
19320   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
19321   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
19322   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
19323   ret.val[3] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
19324   return ret;
19325 }
19326
19327 __extension__ extern __inline int64x2x4_t
19328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19329 vld4q_s64 (const int64_t * __a)
19330 {
19331   int64x2x4_t ret;
19332   __builtin_aarch64_simd_xi __o;
19333   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
19334   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
19335   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
19336   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
19337   ret.val[3] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
19338   return ret;
19339 }
19340
19341 __extension__ extern __inline uint8x16x4_t
19342 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19343 vld4q_u8 (const uint8_t * __a)
19344 {
19345   uint8x16x4_t ret;
19346   __builtin_aarch64_simd_xi __o;
19347   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
19348   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19349   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19350   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19351   ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19352   return ret;
19353 }
19354
19355 __extension__ extern __inline uint16x8x4_t
19356 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19357 vld4q_u16 (const uint16_t * __a)
19358 {
19359   uint16x8x4_t ret;
19360   __builtin_aarch64_simd_xi __o;
19361   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
19362   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19363   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19364   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19365   ret.val[3] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19366   return ret;
19367 }
19368
19369 __extension__ extern __inline uint32x4x4_t
19370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19371 vld4q_u32 (const uint32_t * __a)
19372 {
19373   uint32x4x4_t ret;
19374   __builtin_aarch64_simd_xi __o;
19375   __o = __builtin_aarch64_ld4v4si ((const __builtin_aarch64_simd_si *) __a);
19376   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
19377   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
19378   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
19379   ret.val[3] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
19380   return ret;
19381 }
19382
19383 __extension__ extern __inline uint64x2x4_t
19384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19385 vld4q_u64 (const uint64_t * __a)
19386 {
19387   uint64x2x4_t ret;
19388   __builtin_aarch64_simd_xi __o;
19389   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
19390   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
19391   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
19392   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
19393   ret.val[3] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
19394   return ret;
19395 }
19396
19397 __extension__ extern __inline float16x8x4_t
19398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19399 vld4q_f16 (const float16_t * __a)
19400 {
19401   float16x8x4_t ret;
19402   __builtin_aarch64_simd_xi __o;
19403   __o = __builtin_aarch64_ld4v8hf (__a);
19404   ret.val[0] = __builtin_aarch64_get_qregxiv8hf (__o, 0);
19405   ret.val[1] = __builtin_aarch64_get_qregxiv8hf (__o, 1);
19406   ret.val[2] = __builtin_aarch64_get_qregxiv8hf (__o, 2);
19407   ret.val[3] = __builtin_aarch64_get_qregxiv8hf (__o, 3);
19408   return ret;
19409 }
19410
19411 __extension__ extern __inline float32x4x4_t
19412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19413 vld4q_f32 (const float32_t * __a)
19414 {
19415   float32x4x4_t ret;
19416   __builtin_aarch64_simd_xi __o;
19417   __o = __builtin_aarch64_ld4v4sf ((const __builtin_aarch64_simd_sf *) __a);
19418   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 0);
19419   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 1);
19420   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 2);
19421   ret.val[3] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 3);
19422   return ret;
19423 }
19424
19425 __extension__ extern __inline float64x2x4_t
19426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19427 vld4q_f64 (const float64_t * __a)
19428 {
19429   float64x2x4_t ret;
19430   __builtin_aarch64_simd_xi __o;
19431   __o = __builtin_aarch64_ld4v2df ((const __builtin_aarch64_simd_df *) __a);
19432   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 0);
19433   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 1);
19434   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 2);
19435   ret.val[3] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 3);
19436   return ret;
19437 }
19438
19439 __extension__ extern __inline poly64x2x4_t
19440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19441 vld4q_p64 (const poly64_t * __a)
19442 {
19443   poly64x2x4_t  ret;
19444   __builtin_aarch64_simd_xi __o;
19445   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
19446   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 0);
19447   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 1);
19448   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 2);
19449   ret.val[3] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 3);
19450   return ret;
19451 }
19452
19453 /* vldn_dup */
19454
19455 __extension__ extern __inline int8x8x2_t
19456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19457 vld2_dup_s8 (const int8_t * __a)
19458 {
19459   int8x8x2_t ret;
19460   __builtin_aarch64_simd_oi __o;
19461   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19462   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
19463   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
19464   return ret;
19465 }
19466
19467 __extension__ extern __inline int16x4x2_t
19468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19469 vld2_dup_s16 (const int16_t * __a)
19470 {
19471   int16x4x2_t ret;
19472   __builtin_aarch64_simd_oi __o;
19473   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19474   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
19475   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
19476   return ret;
19477 }
19478
19479 __extension__ extern __inline int32x2x2_t
19480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19481 vld2_dup_s32 (const int32_t * __a)
19482 {
19483   int32x2x2_t ret;
19484   __builtin_aarch64_simd_oi __o;
19485   __o = __builtin_aarch64_ld2rv2si ((const __builtin_aarch64_simd_si *) __a);
19486   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
19487   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
19488   return ret;
19489 }
19490
19491 __extension__ extern __inline float16x4x2_t
19492 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19493 vld2_dup_f16 (const float16_t * __a)
19494 {
19495   float16x4x2_t ret;
19496   __builtin_aarch64_simd_oi __o;
19497   __o = __builtin_aarch64_ld2rv4hf ((const __builtin_aarch64_simd_hf *) __a);
19498   ret.val[0] = __builtin_aarch64_get_dregoiv4hf (__o, 0);
19499   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregoiv4hf (__o, 1);
19500   return ret;
19501 }
19502
19503 __extension__ extern __inline float32x2x2_t
19504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19505 vld2_dup_f32 (const float32_t * __a)
19506 {
19507   float32x2x2_t ret;
19508   __builtin_aarch64_simd_oi __o;
19509   __o = __builtin_aarch64_ld2rv2sf ((const __builtin_aarch64_simd_sf *) __a);
19510   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
19511   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
19512   return ret;
19513 }
19514
19515 __extension__ extern __inline float64x1x2_t
19516 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19517 vld2_dup_f64 (const float64_t * __a)
19518 {
19519   float64x1x2_t ret;
19520   __builtin_aarch64_simd_oi __o;
19521   __o = __builtin_aarch64_ld2rdf ((const __builtin_aarch64_simd_df *) __a);
19522   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
19523   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
19524   return ret;
19525 }
19526
19527 __extension__ extern __inline uint8x8x2_t
19528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19529 vld2_dup_u8 (const uint8_t * __a)
19530 {
19531   uint8x8x2_t ret;
19532   __builtin_aarch64_simd_oi __o;
19533   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19534   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
19535   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
19536   return ret;
19537 }
19538
19539 __extension__ extern __inline uint16x4x2_t
19540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19541 vld2_dup_u16 (const uint16_t * __a)
19542 {
19543   uint16x4x2_t ret;
19544   __builtin_aarch64_simd_oi __o;
19545   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19546   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
19547   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
19548   return ret;
19549 }
19550
19551 __extension__ extern __inline uint32x2x2_t
19552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19553 vld2_dup_u32 (const uint32_t * __a)
19554 {
19555   uint32x2x2_t ret;
19556   __builtin_aarch64_simd_oi __o;
19557   __o = __builtin_aarch64_ld2rv2si ((const __builtin_aarch64_simd_si *) __a);
19558   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
19559   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
19560   return ret;
19561 }
19562
19563 __extension__ extern __inline poly8x8x2_t
19564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19565 vld2_dup_p8 (const poly8_t * __a)
19566 {
19567   poly8x8x2_t ret;
19568   __builtin_aarch64_simd_oi __o;
19569   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19570   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
19571   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
19572   return ret;
19573 }
19574
19575 __extension__ extern __inline poly16x4x2_t
19576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19577 vld2_dup_p16 (const poly16_t * __a)
19578 {
19579   poly16x4x2_t ret;
19580   __builtin_aarch64_simd_oi __o;
19581   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19582   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
19583   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
19584   return ret;
19585 }
19586
19587 __extension__ extern __inline poly64x1x2_t
19588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19589 vld2_dup_p64 (const poly64_t * __a)
19590 {
19591   poly64x1x2_t ret;
19592   __builtin_aarch64_simd_oi __o;
19593   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
19594   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 0);
19595   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 1);
19596   return ret;
19597 }
19598
19599
19600 __extension__ extern __inline int64x1x2_t
19601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19602 vld2_dup_s64 (const int64_t * __a)
19603 {
19604   int64x1x2_t ret;
19605   __builtin_aarch64_simd_oi __o;
19606   __o = __builtin_aarch64_ld2rdi ((const __builtin_aarch64_simd_di *) __a);
19607   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
19608   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
19609   return ret;
19610 }
19611
19612 __extension__ extern __inline uint64x1x2_t
19613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19614 vld2_dup_u64 (const uint64_t * __a)
19615 {
19616   uint64x1x2_t ret;
19617   __builtin_aarch64_simd_oi __o;
19618   __o = __builtin_aarch64_ld2rdi ((const __builtin_aarch64_simd_di *) __a);
19619   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
19620   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
19621   return ret;
19622 }
19623
19624 __extension__ extern __inline int8x16x2_t
19625 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19626 vld2q_dup_s8 (const int8_t * __a)
19627 {
19628   int8x16x2_t ret;
19629   __builtin_aarch64_simd_oi __o;
19630   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19631   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
19632   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
19633   return ret;
19634 }
19635
19636 __extension__ extern __inline poly8x16x2_t
19637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19638 vld2q_dup_p8 (const poly8_t * __a)
19639 {
19640   poly8x16x2_t ret;
19641   __builtin_aarch64_simd_oi __o;
19642   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19643   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
19644   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
19645   return ret;
19646 }
19647
19648 __extension__ extern __inline int16x8x2_t
19649 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19650 vld2q_dup_s16 (const int16_t * __a)
19651 {
19652   int16x8x2_t ret;
19653   __builtin_aarch64_simd_oi __o;
19654   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19655   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
19656   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
19657   return ret;
19658 }
19659
19660 __extension__ extern __inline poly16x8x2_t
19661 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19662 vld2q_dup_p16 (const poly16_t * __a)
19663 {
19664   poly16x8x2_t ret;
19665   __builtin_aarch64_simd_oi __o;
19666   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19667   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
19668   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
19669   return ret;
19670 }
19671
19672 __extension__ extern __inline int32x4x2_t
19673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19674 vld2q_dup_s32 (const int32_t * __a)
19675 {
19676   int32x4x2_t ret;
19677   __builtin_aarch64_simd_oi __o;
19678   __o = __builtin_aarch64_ld2rv4si ((const __builtin_aarch64_simd_si *) __a);
19679   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
19680   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
19681   return ret;
19682 }
19683
19684 __extension__ extern __inline int64x2x2_t
19685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19686 vld2q_dup_s64 (const int64_t * __a)
19687 {
19688   int64x2x2_t ret;
19689   __builtin_aarch64_simd_oi __o;
19690   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
19691   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
19692   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
19693   return ret;
19694 }
19695
19696 __extension__ extern __inline uint8x16x2_t
19697 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19698 vld2q_dup_u8 (const uint8_t * __a)
19699 {
19700   uint8x16x2_t ret;
19701   __builtin_aarch64_simd_oi __o;
19702   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19703   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
19704   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
19705   return ret;
19706 }
19707
19708 __extension__ extern __inline uint16x8x2_t
19709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19710 vld2q_dup_u16 (const uint16_t * __a)
19711 {
19712   uint16x8x2_t ret;
19713   __builtin_aarch64_simd_oi __o;
19714   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19715   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
19716   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
19717   return ret;
19718 }
19719
19720 __extension__ extern __inline uint32x4x2_t
19721 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19722 vld2q_dup_u32 (const uint32_t * __a)
19723 {
19724   uint32x4x2_t ret;
19725   __builtin_aarch64_simd_oi __o;
19726   __o = __builtin_aarch64_ld2rv4si ((const __builtin_aarch64_simd_si *) __a);
19727   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
19728   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
19729   return ret;
19730 }
19731
19732 __extension__ extern __inline uint64x2x2_t
19733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19734 vld2q_dup_u64 (const uint64_t * __a)
19735 {
19736   uint64x2x2_t ret;
19737   __builtin_aarch64_simd_oi __o;
19738   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
19739   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
19740   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
19741   return ret;
19742 }
19743
19744 __extension__ extern __inline float16x8x2_t
19745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19746 vld2q_dup_f16 (const float16_t * __a)
19747 {
19748   float16x8x2_t ret;
19749   __builtin_aarch64_simd_oi __o;
19750   __o = __builtin_aarch64_ld2rv8hf ((const __builtin_aarch64_simd_hf *) __a);
19751   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregoiv8hf (__o, 0);
19752   ret.val[1] = __builtin_aarch64_get_qregoiv8hf (__o, 1);
19753   return ret;
19754 }
19755
19756 __extension__ extern __inline float32x4x2_t
19757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19758 vld2q_dup_f32 (const float32_t * __a)
19759 {
19760   float32x4x2_t ret;
19761   __builtin_aarch64_simd_oi __o;
19762   __o = __builtin_aarch64_ld2rv4sf ((const __builtin_aarch64_simd_sf *) __a);
19763   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
19764   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
19765   return ret;
19766 }
19767
19768 __extension__ extern __inline float64x2x2_t
19769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19770 vld2q_dup_f64 (const float64_t * __a)
19771 {
19772   float64x2x2_t ret;
19773   __builtin_aarch64_simd_oi __o;
19774   __o = __builtin_aarch64_ld2rv2df ((const __builtin_aarch64_simd_df *) __a);
19775   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
19776   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
19777   return ret;
19778 }
19779
19780 __extension__ extern __inline poly64x2x2_t
19781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19782 vld2q_dup_p64 (const poly64_t * __a)
19783 {
19784   poly64x2x2_t ret;
19785   __builtin_aarch64_simd_oi __o;
19786   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
19787   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 0);
19788   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 1);
19789   return ret;
19790 }
19791
19792 __extension__ extern __inline int64x1x3_t
19793 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19794 vld3_dup_s64 (const int64_t * __a)
19795 {
19796   int64x1x3_t ret;
19797   __builtin_aarch64_simd_ci __o;
19798   __o = __builtin_aarch64_ld3rdi ((const __builtin_aarch64_simd_di *) __a);
19799   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
19800   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
19801   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
19802   return ret;
19803 }
19804
19805 __extension__ extern __inline uint64x1x3_t
19806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19807 vld3_dup_u64 (const uint64_t * __a)
19808 {
19809   uint64x1x3_t ret;
19810   __builtin_aarch64_simd_ci __o;
19811   __o = __builtin_aarch64_ld3rdi ((const __builtin_aarch64_simd_di *) __a);
19812   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
19813   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
19814   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
19815   return ret;
19816 }
19817
19818 __extension__ extern __inline float64x1x3_t
19819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19820 vld3_dup_f64 (const float64_t * __a)
19821 {
19822   float64x1x3_t ret;
19823   __builtin_aarch64_simd_ci __o;
19824   __o = __builtin_aarch64_ld3rdf ((const __builtin_aarch64_simd_df *) __a);
19825   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 0)};
19826   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 1)};
19827   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 2)};
19828   return ret;
19829 }
19830
19831 __extension__ extern __inline int8x8x3_t
19832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19833 vld3_dup_s8 (const int8_t * __a)
19834 {
19835   int8x8x3_t ret;
19836   __builtin_aarch64_simd_ci __o;
19837   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19838   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
19839   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
19840   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
19841   return ret;
19842 }
19843
19844 __extension__ extern __inline poly8x8x3_t
19845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19846 vld3_dup_p8 (const poly8_t * __a)
19847 {
19848   poly8x8x3_t ret;
19849   __builtin_aarch64_simd_ci __o;
19850   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19851   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
19852   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
19853   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
19854   return ret;
19855 }
19856
19857 __extension__ extern __inline int16x4x3_t
19858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19859 vld3_dup_s16 (const int16_t * __a)
19860 {
19861   int16x4x3_t ret;
19862   __builtin_aarch64_simd_ci __o;
19863   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19864   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
19865   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
19866   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
19867   return ret;
19868 }
19869
19870 __extension__ extern __inline poly16x4x3_t
19871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19872 vld3_dup_p16 (const poly16_t * __a)
19873 {
19874   poly16x4x3_t ret;
19875   __builtin_aarch64_simd_ci __o;
19876   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19877   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
19878   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
19879   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
19880   return ret;
19881 }
19882
19883 __extension__ extern __inline int32x2x3_t
19884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19885 vld3_dup_s32 (const int32_t * __a)
19886 {
19887   int32x2x3_t ret;
19888   __builtin_aarch64_simd_ci __o;
19889   __o = __builtin_aarch64_ld3rv2si ((const __builtin_aarch64_simd_si *) __a);
19890   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
19891   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
19892   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
19893   return ret;
19894 }
19895
19896 __extension__ extern __inline uint8x8x3_t
19897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19898 vld3_dup_u8 (const uint8_t * __a)
19899 {
19900   uint8x8x3_t ret;
19901   __builtin_aarch64_simd_ci __o;
19902   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19903   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
19904   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
19905   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
19906   return ret;
19907 }
19908
19909 __extension__ extern __inline uint16x4x3_t
19910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19911 vld3_dup_u16 (const uint16_t * __a)
19912 {
19913   uint16x4x3_t ret;
19914   __builtin_aarch64_simd_ci __o;
19915   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19916   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
19917   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
19918   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
19919   return ret;
19920 }
19921
19922 __extension__ extern __inline uint32x2x3_t
19923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19924 vld3_dup_u32 (const uint32_t * __a)
19925 {
19926   uint32x2x3_t ret;
19927   __builtin_aarch64_simd_ci __o;
19928   __o = __builtin_aarch64_ld3rv2si ((const __builtin_aarch64_simd_si *) __a);
19929   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
19930   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
19931   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
19932   return ret;
19933 }
19934
19935 __extension__ extern __inline float16x4x3_t
19936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19937 vld3_dup_f16 (const float16_t * __a)
19938 {
19939   float16x4x3_t ret;
19940   __builtin_aarch64_simd_ci __o;
19941   __o = __builtin_aarch64_ld3rv4hf ((const __builtin_aarch64_simd_hf *) __a);
19942   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 0);
19943   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 1);
19944   ret.val[2] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 2);
19945   return ret;
19946 }
19947
19948 __extension__ extern __inline float32x2x3_t
19949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19950 vld3_dup_f32 (const float32_t * __a)
19951 {
19952   float32x2x3_t ret;
19953   __builtin_aarch64_simd_ci __o;
19954   __o = __builtin_aarch64_ld3rv2sf ((const __builtin_aarch64_simd_sf *) __a);
19955   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 0);
19956   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 1);
19957   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 2);
19958   return ret;
19959 }
19960
19961 __extension__ extern __inline poly64x1x3_t
19962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19963 vld3_dup_p64 (const poly64_t * __a)
19964 {
19965   poly64x1x3_t ret;
19966   __builtin_aarch64_simd_ci __o;
19967   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19968   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 0);
19969   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 1);
19970   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 2);
19971   return ret;
19972 }
19973
19974 __extension__ extern __inline int8x16x3_t
19975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19976 vld3q_dup_s8 (const int8_t * __a)
19977 {
19978   int8x16x3_t ret;
19979   __builtin_aarch64_simd_ci __o;
19980   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19981   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19982   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19983   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19984   return ret;
19985 }
19986
19987 __extension__ extern __inline poly8x16x3_t
19988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19989 vld3q_dup_p8 (const poly8_t * __a)
19990 {
19991   poly8x16x3_t ret;
19992   __builtin_aarch64_simd_ci __o;
19993   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19994   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19995   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19996   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19997   return ret;
19998 }
19999
20000 __extension__ extern __inline int16x8x3_t
20001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20002 vld3q_dup_s16 (const int16_t * __a)
20003 {
20004   int16x8x3_t ret;
20005   __builtin_aarch64_simd_ci __o;
20006   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
20007   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
20008   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
20009   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
20010   return ret;
20011 }
20012
20013 __extension__ extern __inline poly16x8x3_t
20014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20015 vld3q_dup_p16 (const poly16_t * __a)
20016 {
20017   poly16x8x3_t ret;
20018   __builtin_aarch64_simd_ci __o;
20019   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
20020   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
20021   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
20022   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
20023   return ret;
20024 }
20025
20026 __extension__ extern __inline int32x4x3_t
20027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20028 vld3q_dup_s32 (const int32_t * __a)
20029 {
20030   int32x4x3_t ret;
20031   __builtin_aarch64_simd_ci __o;
20032   __o = __builtin_aarch64_ld3rv4si ((const __builtin_aarch64_simd_si *) __a);
20033   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
20034   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
20035   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
20036   return ret;
20037 }
20038
20039 __extension__ extern __inline int64x2x3_t
20040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20041 vld3q_dup_s64 (const int64_t * __a)
20042 {
20043   int64x2x3_t ret;
20044   __builtin_aarch64_simd_ci __o;
20045   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
20046   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
20047   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
20048   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
20049   return ret;
20050 }
20051
20052 __extension__ extern __inline uint8x16x3_t
20053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20054 vld3q_dup_u8 (const uint8_t * __a)
20055 {
20056   uint8x16x3_t ret;
20057   __builtin_aarch64_simd_ci __o;
20058   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
20059   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
20060   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
20061   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
20062   return ret;
20063 }
20064
20065 __extension__ extern __inline uint16x8x3_t
20066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20067 vld3q_dup_u16 (const uint16_t * __a)
20068 {
20069   uint16x8x3_t ret;
20070   __builtin_aarch64_simd_ci __o;
20071   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
20072   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
20073   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
20074   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
20075   return ret;
20076 }
20077
20078 __extension__ extern __inline uint32x4x3_t
20079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20080 vld3q_dup_u32 (const uint32_t * __a)
20081 {
20082   uint32x4x3_t ret;
20083   __builtin_aarch64_simd_ci __o;
20084   __o = __builtin_aarch64_ld3rv4si ((const __builtin_aarch64_simd_si *) __a);
20085   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
20086   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
20087   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
20088   return ret;
20089 }
20090
20091 __extension__ extern __inline uint64x2x3_t
20092 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20093 vld3q_dup_u64 (const uint64_t * __a)
20094 {
20095   uint64x2x3_t ret;
20096   __builtin_aarch64_simd_ci __o;
20097   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
20098   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
20099   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
20100   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
20101   return ret;
20102 }
20103
20104 __extension__ extern __inline float16x8x3_t
20105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20106 vld3q_dup_f16 (const float16_t * __a)
20107 {
20108   float16x8x3_t ret;
20109   __builtin_aarch64_simd_ci __o;
20110   __o = __builtin_aarch64_ld3rv8hf ((const __builtin_aarch64_simd_hf *) __a);
20111   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 0);
20112   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 1);
20113   ret.val[2] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 2);
20114   return ret;
20115 }
20116
20117 __extension__ extern __inline float32x4x3_t
20118 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20119 vld3q_dup_f32 (const float32_t * __a)
20120 {
20121   float32x4x3_t ret;
20122   __builtin_aarch64_simd_ci __o;
20123   __o = __builtin_aarch64_ld3rv4sf ((const __builtin_aarch64_simd_sf *) __a);
20124   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 0);
20125   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 1);
20126   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 2);
20127   return ret;
20128 }
20129
20130 __extension__ extern __inline float64x2x3_t
20131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20132 vld3q_dup_f64 (const float64_t * __a)
20133 {
20134   float64x2x3_t ret;
20135   __builtin_aarch64_simd_ci __o;
20136   __o = __builtin_aarch64_ld3rv2df ((const __builtin_aarch64_simd_df *) __a);
20137   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 0);
20138   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 1);
20139   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 2);
20140   return ret;
20141 }
20142
20143 __extension__ extern __inline poly64x2x3_t
20144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20145 vld3q_dup_p64 (const poly64_t * __a)
20146 {
20147   poly64x2x3_t ret;
20148   __builtin_aarch64_simd_ci __o;
20149   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
20150   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 0);
20151   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 1);
20152   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 2);
20153   return ret;
20154 }
20155
20156 __extension__ extern __inline int64x1x4_t
20157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20158 vld4_dup_s64 (const int64_t * __a)
20159 {
20160   int64x1x4_t ret;
20161   __builtin_aarch64_simd_xi __o;
20162   __o = __builtin_aarch64_ld4rdi ((const __builtin_aarch64_simd_di *) __a);
20163   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
20164   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
20165   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
20166   ret.val[3] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
20167   return ret;
20168 }
20169
20170 __extension__ extern __inline uint64x1x4_t
20171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20172 vld4_dup_u64 (const uint64_t * __a)
20173 {
20174   uint64x1x4_t ret;
20175   __builtin_aarch64_simd_xi __o;
20176   __o = __builtin_aarch64_ld4rdi ((const __builtin_aarch64_simd_di *) __a);
20177   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
20178   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
20179   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
20180   ret.val[3] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
20181   return ret;
20182 }
20183
20184 __extension__ extern __inline float64x1x4_t
20185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20186 vld4_dup_f64 (const float64_t * __a)
20187 {
20188   float64x1x4_t ret;
20189   __builtin_aarch64_simd_xi __o;
20190   __o = __builtin_aarch64_ld4rdf ((const __builtin_aarch64_simd_df *) __a);
20191   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 0)};
20192   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 1)};
20193   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 2)};
20194   ret.val[3] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 3)};
20195   return ret;
20196 }
20197
20198 __extension__ extern __inline int8x8x4_t
20199 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20200 vld4_dup_s8 (const int8_t * __a)
20201 {
20202   int8x8x4_t ret;
20203   __builtin_aarch64_simd_xi __o;
20204   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
20205   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
20206   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
20207   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
20208   ret.val[3] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
20209   return ret;
20210 }
20211
20212 __extension__ extern __inline poly8x8x4_t
20213 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20214 vld4_dup_p8 (const poly8_t * __a)
20215 {
20216   poly8x8x4_t ret;
20217   __builtin_aarch64_simd_xi __o;
20218   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
20219   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
20220   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
20221   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
20222   ret.val[3] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
20223   return ret;
20224 }
20225
20226 __extension__ extern __inline int16x4x4_t
20227 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20228 vld4_dup_s16 (const int16_t * __a)
20229 {
20230   int16x4x4_t ret;
20231   __builtin_aarch64_simd_xi __o;
20232   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
20233   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
20234   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
20235   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
20236   ret.val[3] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
20237   return ret;
20238 }
20239
20240 __extension__ extern __inline poly16x4x4_t
20241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20242 vld4_dup_p16 (const poly16_t * __a)
20243 {
20244   poly16x4x4_t ret;
20245   __builtin_aarch64_simd_xi __o;
20246   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
20247   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
20248   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
20249   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
20250   ret.val[3] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
20251   return ret;
20252 }
20253
20254 __extension__ extern __inline int32x2x4_t
20255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20256 vld4_dup_s32 (const int32_t * __a)
20257 {
20258   int32x2x4_t ret;
20259   __builtin_aarch64_simd_xi __o;
20260   __o = __builtin_aarch64_ld4rv2si ((const __builtin_aarch64_simd_si *) __a);
20261   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
20262   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
20263   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
20264   ret.val[3] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
20265   return ret;
20266 }
20267
20268 __extension__ extern __inline uint8x8x4_t
20269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20270 vld4_dup_u8 (const uint8_t * __a)
20271 {
20272   uint8x8x4_t ret;
20273   __builtin_aarch64_simd_xi __o;
20274   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
20275   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
20276   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
20277   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
20278   ret.val[3] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
20279   return ret;
20280 }
20281
20282 __extension__ extern __inline uint16x4x4_t
20283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20284 vld4_dup_u16 (const uint16_t * __a)
20285 {
20286   uint16x4x4_t ret;
20287   __builtin_aarch64_simd_xi __o;
20288   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
20289   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
20290   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
20291   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
20292   ret.val[3] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
20293   return ret;
20294 }
20295
20296 __extension__ extern __inline uint32x2x4_t
20297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20298 vld4_dup_u32 (const uint32_t * __a)
20299 {
20300   uint32x2x4_t ret;
20301   __builtin_aarch64_simd_xi __o;
20302   __o = __builtin_aarch64_ld4rv2si ((const __builtin_aarch64_simd_si *) __a);
20303   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
20304   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
20305   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
20306   ret.val[3] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
20307   return ret;
20308 }
20309
20310 __extension__ extern __inline float16x4x4_t
20311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20312 vld4_dup_f16 (const float16_t * __a)
20313 {
20314   float16x4x4_t ret;
20315   __builtin_aarch64_simd_xi __o;
20316   __o = __builtin_aarch64_ld4rv4hf ((const __builtin_aarch64_simd_hf *) __a);
20317   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 0);
20318   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 1);
20319   ret.val[2] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 2);
20320   ret.val[3] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 3);
20321   return ret;
20322 }
20323
20324 __extension__ extern __inline float32x2x4_t
20325 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20326 vld4_dup_f32 (const float32_t * __a)
20327 {
20328   float32x2x4_t ret;
20329   __builtin_aarch64_simd_xi __o;
20330   __o = __builtin_aarch64_ld4rv2sf ((const __builtin_aarch64_simd_sf *) __a);
20331   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 0);
20332   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 1);
20333   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 2);
20334   ret.val[3] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 3);
20335   return ret;
20336 }
20337
20338 __extension__ extern __inline poly64x1x4_t
20339 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20340 vld4_dup_p64 (const poly64_t * __a)
20341 {
20342   poly64x1x4_t ret;
20343   __builtin_aarch64_simd_xi __o;
20344   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
20345   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 0);
20346   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 1);
20347   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 2);
20348   ret.val[3] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 3);
20349   return ret;
20350 }
20351
20352 __extension__ extern __inline int8x16x4_t
20353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20354 vld4q_dup_s8 (const int8_t * __a)
20355 {
20356   int8x16x4_t ret;
20357   __builtin_aarch64_simd_xi __o;
20358   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
20359   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
20360   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
20361   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
20362   ret.val[3] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
20363   return ret;
20364 }
20365
20366 __extension__ extern __inline poly8x16x4_t
20367 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20368 vld4q_dup_p8 (const poly8_t * __a)
20369 {
20370   poly8x16x4_t ret;
20371   __builtin_aarch64_simd_xi __o;
20372   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
20373   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
20374   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
20375   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
20376   ret.val[3] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
20377   return ret;
20378 }
20379
20380 __extension__ extern __inline int16x8x4_t
20381 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20382 vld4q_dup_s16 (const int16_t * __a)
20383 {
20384   int16x8x4_t ret;
20385   __builtin_aarch64_simd_xi __o;
20386   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
20387   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
20388   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
20389   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
20390   ret.val[3] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
20391   return ret;
20392 }
20393
20394 __extension__ extern __inline poly16x8x4_t
20395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20396 vld4q_dup_p16 (const poly16_t * __a)
20397 {
20398   poly16x8x4_t ret;
20399   __builtin_aarch64_simd_xi __o;
20400   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
20401   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
20402   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
20403   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
20404   ret.val[3] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
20405   return ret;
20406 }
20407
20408 __extension__ extern __inline int32x4x4_t
20409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20410 vld4q_dup_s32 (const int32_t * __a)
20411 {
20412   int32x4x4_t ret;
20413   __builtin_aarch64_simd_xi __o;
20414   __o = __builtin_aarch64_ld4rv4si ((const __builtin_aarch64_simd_si *) __a);
20415   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
20416   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
20417   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
20418   ret.val[3] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
20419   return ret;
20420 }
20421
20422 __extension__ extern __inline int64x2x4_t
20423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20424 vld4q_dup_s64 (const int64_t * __a)
20425 {
20426   int64x2x4_t ret;
20427   __builtin_aarch64_simd_xi __o;
20428   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
20429   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
20430   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
20431   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
20432   ret.val[3] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
20433   return ret;
20434 }
20435
20436 __extension__ extern __inline uint8x16x4_t
20437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20438 vld4q_dup_u8 (const uint8_t * __a)
20439 {
20440   uint8x16x4_t ret;
20441   __builtin_aarch64_simd_xi __o;
20442   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
20443   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
20444   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
20445   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
20446   ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
20447   return ret;
20448 }
20449
20450 __extension__ extern __inline uint16x8x4_t
20451 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20452 vld4q_dup_u16 (const uint16_t * __a)
20453 {
20454   uint16x8x4_t ret;
20455   __builtin_aarch64_simd_xi __o;
20456   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
20457   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
20458   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
20459   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
20460   ret.val[3] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
20461   return ret;
20462 }
20463
20464 __extension__ extern __inline uint32x4x4_t
20465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20466 vld4q_dup_u32 (const uint32_t * __a)
20467 {
20468   uint32x4x4_t ret;
20469   __builtin_aarch64_simd_xi __o;
20470   __o = __builtin_aarch64_ld4rv4si ((const __builtin_aarch64_simd_si *) __a);
20471   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
20472   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
20473   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
20474   ret.val[3] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
20475   return ret;
20476 }
20477
20478 __extension__ extern __inline uint64x2x4_t
20479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20480 vld4q_dup_u64 (const uint64_t * __a)
20481 {
20482   uint64x2x4_t ret;
20483   __builtin_aarch64_simd_xi __o;
20484   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
20485   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
20486   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
20487   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
20488   ret.val[3] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
20489   return ret;
20490 }
20491
20492 __extension__ extern __inline float16x8x4_t
20493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20494 vld4q_dup_f16 (const float16_t * __a)
20495 {
20496   float16x8x4_t ret;
20497   __builtin_aarch64_simd_xi __o;
20498   __o = __builtin_aarch64_ld4rv8hf ((const __builtin_aarch64_simd_hf *) __a);
20499   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 0);
20500   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 1);
20501   ret.val[2] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 2);
20502   ret.val[3] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 3);
20503   return ret;
20504 }
20505
20506 __extension__ extern __inline float32x4x4_t
20507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20508 vld4q_dup_f32 (const float32_t * __a)
20509 {
20510   float32x4x4_t ret;
20511   __builtin_aarch64_simd_xi __o;
20512   __o = __builtin_aarch64_ld4rv4sf ((const __builtin_aarch64_simd_sf *) __a);
20513   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 0);
20514   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 1);
20515   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 2);
20516   ret.val[3] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 3);
20517   return ret;
20518 }
20519
20520 __extension__ extern __inline float64x2x4_t
20521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20522 vld4q_dup_f64 (const float64_t * __a)
20523 {
20524   float64x2x4_t ret;
20525   __builtin_aarch64_simd_xi __o;
20526   __o = __builtin_aarch64_ld4rv2df ((const __builtin_aarch64_simd_df *) __a);
20527   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 0);
20528   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 1);
20529   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 2);
20530   ret.val[3] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 3);
20531   return ret;
20532 }
20533
20534 __extension__ extern __inline poly64x2x4_t
20535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20536 vld4q_dup_p64 (const poly64_t * __a)
20537 {
20538   poly64x2x4_t ret;
20539   __builtin_aarch64_simd_xi __o;
20540   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
20541   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 0);
20542   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 1);
20543   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 2);
20544   ret.val[3] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 3);
20545   return ret;
20546 }
20547
20548 /* vld2_lane */
20549
20550 #define __LD2_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
20551                          qmode, ptrmode, funcsuffix, signedtype)           \
20552 __extension__ extern __inline intype \
20553 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
20554 vld2_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
20555 {                                                                          \
20556   __builtin_aarch64_simd_oi __o;                                           \
20557   largetype __temp;                                                        \
20558   __temp.val[0] =                                                          \
20559     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
20560   __temp.val[1] =                                                          \
20561     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
20562   __o = __builtin_aarch64_set_qregoi##qmode (__o,                          \
20563                                             (signedtype) __temp.val[0],    \
20564                                             0);                            \
20565   __o = __builtin_aarch64_set_qregoi##qmode (__o,                          \
20566                                             (signedtype) __temp.val[1],    \
20567                                             1);                            \
20568   __o = __builtin_aarch64_ld2_lane##mode (                                 \
20569           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
20570   __b.val[0] = (vectype) __builtin_aarch64_get_dregoidi (__o, 0);          \
20571   __b.val[1] = (vectype) __builtin_aarch64_get_dregoidi (__o, 1);          \
20572   return __b;                                                              \
20573 }
20574
20575 __LD2_LANE_FUNC (float16x4x2_t, float16x4_t, float16x8x2_t, float16_t, v4hf,
20576                  v8hf, hf, f16, float16x8_t)
20577 __LD2_LANE_FUNC (float32x2x2_t, float32x2_t, float32x4x2_t, float32_t, v2sf, v4sf,
20578                  sf, f32, float32x4_t)
20579 __LD2_LANE_FUNC (float64x1x2_t, float64x1_t, float64x2x2_t, float64_t, df, v2df,
20580                  df, f64, float64x2_t)
20581 __LD2_LANE_FUNC (poly8x8x2_t, poly8x8_t, poly8x16x2_t, poly8_t, v8qi, v16qi, qi, p8,
20582                  int8x16_t)
20583 __LD2_LANE_FUNC (poly16x4x2_t, poly16x4_t, poly16x8x2_t, poly16_t, v4hi, v8hi, hi,
20584                  p16, int16x8_t)
20585 __LD2_LANE_FUNC (poly64x1x2_t, poly64x1_t, poly64x2x2_t, poly64_t, di,
20586                  v2di_ssps, di, p64, poly64x2_t)
20587 __LD2_LANE_FUNC (int8x8x2_t, int8x8_t, int8x16x2_t, int8_t, v8qi, v16qi, qi, s8,
20588                  int8x16_t)
20589 __LD2_LANE_FUNC (int16x4x2_t, int16x4_t, int16x8x2_t, int16_t, v4hi, v8hi, hi, s16,
20590                  int16x8_t)
20591 __LD2_LANE_FUNC (int32x2x2_t, int32x2_t, int32x4x2_t, int32_t, v2si, v4si, si, s32,
20592                  int32x4_t)
20593 __LD2_LANE_FUNC (int64x1x2_t, int64x1_t, int64x2x2_t, int64_t, di, v2di, di, s64,
20594                  int64x2_t)
20595 __LD2_LANE_FUNC (uint8x8x2_t, uint8x8_t, uint8x16x2_t, uint8_t, v8qi, v16qi, qi, u8,
20596                  int8x16_t)
20597 __LD2_LANE_FUNC (uint16x4x2_t, uint16x4_t, uint16x8x2_t, uint16_t, v4hi, v8hi, hi,
20598                  u16, int16x8_t)
20599 __LD2_LANE_FUNC (uint32x2x2_t, uint32x2_t, uint32x4x2_t, uint32_t, v2si, v4si, si,
20600                  u32, int32x4_t)
20601 __LD2_LANE_FUNC (uint64x1x2_t, uint64x1_t, uint64x2x2_t, uint64_t, di, v2di, di,
20602                  u64, int64x2_t)
20603
20604 #undef __LD2_LANE_FUNC
20605
20606 /* vld2q_lane */
20607
20608 #define __LD2_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
20609 __extension__ extern __inline intype \
20610 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
20611 vld2q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
20612 {                                                                          \
20613   __builtin_aarch64_simd_oi __o;                                           \
20614   intype ret;                                                              \
20615   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) __b.val[0], 0); \
20616   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) __b.val[1], 1); \
20617   __o = __builtin_aarch64_ld2_lane##mode (                                 \
20618         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
20619   ret.val[0] = (vtype) __builtin_aarch64_get_qregoiv4si (__o, 0);          \
20620   ret.val[1] = (vtype) __builtin_aarch64_get_qregoiv4si (__o, 1);          \
20621   return ret;                                                              \
20622 }
20623
20624 __LD2_LANE_FUNC (float16x8x2_t, float16x8_t, float16_t, v8hf, hf, f16)
20625 __LD2_LANE_FUNC (float32x4x2_t, float32x4_t, float32_t, v4sf, sf, f32)
20626 __LD2_LANE_FUNC (float64x2x2_t, float64x2_t, float64_t, v2df, df, f64)
20627 __LD2_LANE_FUNC (poly8x16x2_t, poly8x16_t, poly8_t, v16qi, qi, p8)
20628 __LD2_LANE_FUNC (poly16x8x2_t, poly16x8_t, poly16_t, v8hi, hi, p16)
20629 __LD2_LANE_FUNC (poly64x2x2_t, poly64x2_t, poly64_t, v2di, di, p64)
20630 __LD2_LANE_FUNC (int8x16x2_t, int8x16_t, int8_t, v16qi, qi, s8)
20631 __LD2_LANE_FUNC (int16x8x2_t, int16x8_t, int16_t, v8hi, hi, s16)
20632 __LD2_LANE_FUNC (int32x4x2_t, int32x4_t, int32_t, v4si, si, s32)
20633 __LD2_LANE_FUNC (int64x2x2_t, int64x2_t, int64_t, v2di, di, s64)
20634 __LD2_LANE_FUNC (uint8x16x2_t, uint8x16_t, uint8_t, v16qi, qi, u8)
20635 __LD2_LANE_FUNC (uint16x8x2_t, uint16x8_t, uint16_t, v8hi, hi, u16)
20636 __LD2_LANE_FUNC (uint32x4x2_t, uint32x4_t, uint32_t, v4si, si, u32)
20637 __LD2_LANE_FUNC (uint64x2x2_t, uint64x2_t, uint64_t, v2di, di, u64)
20638
20639 #undef __LD2_LANE_FUNC
20640
20641 /* vld3_lane */
20642
20643 #define __LD3_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
20644                          qmode, ptrmode, funcsuffix, signedtype)           \
20645 __extension__ extern __inline intype \
20646 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
20647 vld3_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
20648 {                                                                          \
20649   __builtin_aarch64_simd_ci __o;                                           \
20650   largetype __temp;                                                        \
20651   __temp.val[0] =                                                          \
20652     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
20653   __temp.val[1] =                                                          \
20654     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
20655   __temp.val[2] =                                                          \
20656     vcombine_##funcsuffix (__b.val[2], vcreate_##funcsuffix (0));          \
20657   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
20658                                             (signedtype) __temp.val[0],    \
20659                                             0);                            \
20660   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
20661                                             (signedtype) __temp.val[1],    \
20662                                             1);                            \
20663   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
20664                                             (signedtype) __temp.val[2],    \
20665                                             2);                            \
20666   __o = __builtin_aarch64_ld3_lane##mode (                                 \
20667           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
20668   __b.val[0] = (vectype) __builtin_aarch64_get_dregcidi (__o, 0);          \
20669   __b.val[1] = (vectype) __builtin_aarch64_get_dregcidi (__o, 1);          \
20670   __b.val[2] = (vectype) __builtin_aarch64_get_dregcidi (__o, 2);          \
20671   return __b;                                                              \
20672 }
20673
20674 __LD3_LANE_FUNC (float16x4x3_t, float16x4_t, float16x8x3_t, float16_t, v4hf,
20675                  v8hf, hf, f16, float16x8_t)
20676 __LD3_LANE_FUNC (float32x2x3_t, float32x2_t, float32x4x3_t, float32_t, v2sf, v4sf,
20677                  sf, f32, float32x4_t)
20678 __LD3_LANE_FUNC (float64x1x3_t, float64x1_t, float64x2x3_t, float64_t, df, v2df,
20679                  df, f64, float64x2_t)
20680 __LD3_LANE_FUNC (poly8x8x3_t, poly8x8_t, poly8x16x3_t, poly8_t, v8qi, v16qi, qi, p8,
20681                  int8x16_t)
20682 __LD3_LANE_FUNC (poly16x4x3_t, poly16x4_t, poly16x8x3_t, poly16_t, v4hi, v8hi, hi,
20683                  p16, int16x8_t)
20684 __LD3_LANE_FUNC (poly64x1x3_t, poly64x1_t, poly64x2x3_t, poly64_t, di,
20685                  v2di_ssps, di, p64, poly64x2_t)
20686 __LD3_LANE_FUNC (int8x8x3_t, int8x8_t, int8x16x3_t, int8_t, v8qi, v16qi, qi, s8,
20687                  int8x16_t)
20688 __LD3_LANE_FUNC (int16x4x3_t, int16x4_t, int16x8x3_t, int16_t, v4hi, v8hi, hi, s16,
20689                  int16x8_t)
20690 __LD3_LANE_FUNC (int32x2x3_t, int32x2_t, int32x4x3_t, int32_t, v2si, v4si, si, s32,
20691                  int32x4_t)
20692 __LD3_LANE_FUNC (int64x1x3_t, int64x1_t, int64x2x3_t, int64_t, di, v2di, di, s64,
20693                  int64x2_t)
20694 __LD3_LANE_FUNC (uint8x8x3_t, uint8x8_t, uint8x16x3_t, uint8_t, v8qi, v16qi, qi, u8,
20695                  int8x16_t)
20696 __LD3_LANE_FUNC (uint16x4x3_t, uint16x4_t, uint16x8x3_t, uint16_t, v4hi, v8hi, hi,
20697                  u16, int16x8_t)
20698 __LD3_LANE_FUNC (uint32x2x3_t, uint32x2_t, uint32x4x3_t, uint32_t, v2si, v4si, si,
20699                  u32, int32x4_t)
20700 __LD3_LANE_FUNC (uint64x1x3_t, uint64x1_t, uint64x2x3_t, uint64_t, di, v2di, di,
20701                  u64, int64x2_t)
20702
20703 #undef __LD3_LANE_FUNC
20704
20705 /* vld3q_lane */
20706
20707 #define __LD3_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
20708 __extension__ extern __inline intype \
20709 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
20710 vld3q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
20711 {                                                                          \
20712   __builtin_aarch64_simd_ci __o;                                           \
20713   intype ret;                                                              \
20714   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[0], 0); \
20715   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[1], 1); \
20716   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[2], 2); \
20717   __o = __builtin_aarch64_ld3_lane##mode (                                 \
20718         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
20719   ret.val[0] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 0);          \
20720   ret.val[1] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 1);          \
20721   ret.val[2] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 2);          \
20722   return ret;                                                              \
20723 }
20724
20725 __LD3_LANE_FUNC (float16x8x3_t, float16x8_t, float16_t, v8hf, hf, f16)
20726 __LD3_LANE_FUNC (float32x4x3_t, float32x4_t, float32_t, v4sf, sf, f32)
20727 __LD3_LANE_FUNC (float64x2x3_t, float64x2_t, float64_t, v2df, df, f64)
20728 __LD3_LANE_FUNC (poly8x16x3_t, poly8x16_t, poly8_t, v16qi, qi, p8)
20729 __LD3_LANE_FUNC (poly16x8x3_t, poly16x8_t, poly16_t, v8hi, hi, p16)
20730 __LD3_LANE_FUNC (poly64x2x3_t, poly64x2_t, poly64_t, v2di, di, p64)
20731 __LD3_LANE_FUNC (int8x16x3_t, int8x16_t, int8_t, v16qi, qi, s8)
20732 __LD3_LANE_FUNC (int16x8x3_t, int16x8_t, int16_t, v8hi, hi, s16)
20733 __LD3_LANE_FUNC (int32x4x3_t, int32x4_t, int32_t, v4si, si, s32)
20734 __LD3_LANE_FUNC (int64x2x3_t, int64x2_t, int64_t, v2di, di, s64)
20735 __LD3_LANE_FUNC (uint8x16x3_t, uint8x16_t, uint8_t, v16qi, qi, u8)
20736 __LD3_LANE_FUNC (uint16x8x3_t, uint16x8_t, uint16_t, v8hi, hi, u16)
20737 __LD3_LANE_FUNC (uint32x4x3_t, uint32x4_t, uint32_t, v4si, si, u32)
20738 __LD3_LANE_FUNC (uint64x2x3_t, uint64x2_t, uint64_t, v2di, di, u64)
20739
20740 #undef __LD3_LANE_FUNC
20741
20742 /* vld4_lane */
20743
20744 #define __LD4_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
20745                          qmode, ptrmode, funcsuffix, signedtype)           \
20746 __extension__ extern __inline intype \
20747 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
20748 vld4_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
20749 {                                                                          \
20750   __builtin_aarch64_simd_xi __o;                                           \
20751   largetype __temp;                                                        \
20752   __temp.val[0] =                                                          \
20753     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
20754   __temp.val[1] =                                                          \
20755     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
20756   __temp.val[2] =                                                          \
20757     vcombine_##funcsuffix (__b.val[2], vcreate_##funcsuffix (0));          \
20758   __temp.val[3] =                                                          \
20759     vcombine_##funcsuffix (__b.val[3], vcreate_##funcsuffix (0));          \
20760   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
20761                                             (signedtype) __temp.val[0],    \
20762                                             0);                            \
20763   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
20764                                             (signedtype) __temp.val[1],    \
20765                                             1);                            \
20766   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
20767                                             (signedtype) __temp.val[2],    \
20768                                             2);                            \
20769   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
20770                                             (signedtype) __temp.val[3],    \
20771                                             3);                            \
20772   __o = __builtin_aarch64_ld4_lane##mode (                                 \
20773           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
20774   __b.val[0] = (vectype) __builtin_aarch64_get_dregxidi (__o, 0);          \
20775   __b.val[1] = (vectype) __builtin_aarch64_get_dregxidi (__o, 1);          \
20776   __b.val[2] = (vectype) __builtin_aarch64_get_dregxidi (__o, 2);          \
20777   __b.val[3] = (vectype) __builtin_aarch64_get_dregxidi (__o, 3);          \
20778   return __b;                                                              \
20779 }
20780
20781 /* vld4q_lane */
20782
20783 __LD4_LANE_FUNC (float16x4x4_t, float16x4_t, float16x8x4_t, float16_t, v4hf,
20784                  v8hf, hf, f16, float16x8_t)
20785 __LD4_LANE_FUNC (float32x2x4_t, float32x2_t, float32x4x4_t, float32_t, v2sf, v4sf,
20786                  sf, f32, float32x4_t)
20787 __LD4_LANE_FUNC (float64x1x4_t, float64x1_t, float64x2x4_t, float64_t, df, v2df,
20788                  df, f64, float64x2_t)
20789 __LD4_LANE_FUNC (poly8x8x4_t, poly8x8_t, poly8x16x4_t, poly8_t, v8qi, v16qi, qi, p8,
20790                  int8x16_t)
20791 __LD4_LANE_FUNC (poly16x4x4_t, poly16x4_t, poly16x8x4_t, poly16_t, v4hi, v8hi, hi,
20792                  p16, int16x8_t)
20793 __LD4_LANE_FUNC (poly64x1x4_t, poly64x1_t, poly64x2x4_t, poly64_t, di,
20794                  v2di_ssps, di, p64, poly64x2_t)
20795 __LD4_LANE_FUNC (int8x8x4_t, int8x8_t, int8x16x4_t, int8_t, v8qi, v16qi, qi, s8,
20796                  int8x16_t)
20797 __LD4_LANE_FUNC (int16x4x4_t, int16x4_t, int16x8x4_t, int16_t, v4hi, v8hi, hi, s16,
20798                  int16x8_t)
20799 __LD4_LANE_FUNC (int32x2x4_t, int32x2_t, int32x4x4_t, int32_t, v2si, v4si, si, s32,
20800                  int32x4_t)
20801 __LD4_LANE_FUNC (int64x1x4_t, int64x1_t, int64x2x4_t, int64_t, di, v2di, di, s64,
20802                  int64x2_t)
20803 __LD4_LANE_FUNC (uint8x8x4_t, uint8x8_t, uint8x16x4_t, uint8_t, v8qi, v16qi, qi, u8,
20804                  int8x16_t)
20805 __LD4_LANE_FUNC (uint16x4x4_t, uint16x4_t, uint16x8x4_t, uint16_t, v4hi, v8hi, hi,
20806                  u16, int16x8_t)
20807 __LD4_LANE_FUNC (uint32x2x4_t, uint32x2_t, uint32x4x4_t, uint32_t, v2si, v4si, si,
20808                  u32, int32x4_t)
20809 __LD4_LANE_FUNC (uint64x1x4_t, uint64x1_t, uint64x2x4_t, uint64_t, di, v2di, di,
20810                  u64, int64x2_t)
20811
20812 #undef __LD4_LANE_FUNC
20813
20814 /* vld4q_lane */
20815
20816 #define __LD4_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
20817 __extension__ extern __inline intype \
20818 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
20819 vld4q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
20820 {                                                                          \
20821   __builtin_aarch64_simd_xi __o;                                           \
20822   intype ret;                                                              \
20823   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[0], 0); \
20824   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[1], 1); \
20825   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[2], 2); \
20826   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[3], 3); \
20827   __o = __builtin_aarch64_ld4_lane##mode (                                 \
20828         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
20829   ret.val[0] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 0);          \
20830   ret.val[1] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 1);          \
20831   ret.val[2] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 2);          \
20832   ret.val[3] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 3);          \
20833   return ret;                                                              \
20834 }
20835
20836 __LD4_LANE_FUNC (float16x8x4_t, float16x8_t, float16_t, v8hf, hf, f16)
20837 __LD4_LANE_FUNC (float32x4x4_t, float32x4_t, float32_t, v4sf, sf, f32)
20838 __LD4_LANE_FUNC (float64x2x4_t, float64x2_t, float64_t, v2df, df, f64)
20839 __LD4_LANE_FUNC (poly8x16x4_t, poly8x16_t, poly8_t, v16qi, qi, p8)
20840 __LD4_LANE_FUNC (poly16x8x4_t, poly16x8_t, poly16_t, v8hi, hi, p16)
20841 __LD4_LANE_FUNC (poly64x2x4_t, poly64x2_t, poly64_t, v2di, di, p64)
20842 __LD4_LANE_FUNC (int8x16x4_t, int8x16_t, int8_t, v16qi, qi, s8)
20843 __LD4_LANE_FUNC (int16x8x4_t, int16x8_t, int16_t, v8hi, hi, s16)
20844 __LD4_LANE_FUNC (int32x4x4_t, int32x4_t, int32_t, v4si, si, s32)
20845 __LD4_LANE_FUNC (int64x2x4_t, int64x2_t, int64_t, v2di, di, s64)
20846 __LD4_LANE_FUNC (uint8x16x4_t, uint8x16_t, uint8_t, v16qi, qi, u8)
20847 __LD4_LANE_FUNC (uint16x8x4_t, uint16x8_t, uint16_t, v8hi, hi, u16)
20848 __LD4_LANE_FUNC (uint32x4x4_t, uint32x4_t, uint32_t, v4si, si, u32)
20849 __LD4_LANE_FUNC (uint64x2x4_t, uint64x2_t, uint64_t, v2di, di, u64)
20850
20851 #undef __LD4_LANE_FUNC
20852
20853 /* vmax */
20854
20855 __extension__ extern __inline float32x2_t
20856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20857 vmax_f32 (float32x2_t __a, float32x2_t __b)
20858 {
20859   return __builtin_aarch64_smax_nanv2sf (__a, __b);
20860 }
20861
20862 __extension__ extern __inline float64x1_t
20863 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20864 vmax_f64 (float64x1_t __a, float64x1_t __b)
20865 {
20866     return (float64x1_t)
20867       { __builtin_aarch64_smax_nandf (vget_lane_f64 (__a, 0),
20868                                       vget_lane_f64 (__b, 0)) };
20869 }
20870
20871 __extension__ extern __inline int8x8_t
20872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20873 vmax_s8 (int8x8_t __a, int8x8_t __b)
20874 {
20875   return __builtin_aarch64_smaxv8qi (__a, __b);
20876 }
20877
20878 __extension__ extern __inline int16x4_t
20879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20880 vmax_s16 (int16x4_t __a, int16x4_t __b)
20881 {
20882   return __builtin_aarch64_smaxv4hi (__a, __b);
20883 }
20884
20885 __extension__ extern __inline int32x2_t
20886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20887 vmax_s32 (int32x2_t __a, int32x2_t __b)
20888 {
20889   return __builtin_aarch64_smaxv2si (__a, __b);
20890 }
20891
20892 __extension__ extern __inline uint8x8_t
20893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20894 vmax_u8 (uint8x8_t __a, uint8x8_t __b)
20895 {
20896   return (uint8x8_t) __builtin_aarch64_umaxv8qi ((int8x8_t) __a,
20897                                                  (int8x8_t) __b);
20898 }
20899
20900 __extension__ extern __inline uint16x4_t
20901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20902 vmax_u16 (uint16x4_t __a, uint16x4_t __b)
20903 {
20904   return (uint16x4_t) __builtin_aarch64_umaxv4hi ((int16x4_t) __a,
20905                                                   (int16x4_t) __b);
20906 }
20907
20908 __extension__ extern __inline uint32x2_t
20909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20910 vmax_u32 (uint32x2_t __a, uint32x2_t __b)
20911 {
20912   return (uint32x2_t) __builtin_aarch64_umaxv2si ((int32x2_t) __a,
20913                                                   (int32x2_t) __b);
20914 }
20915
20916 __extension__ extern __inline float32x4_t
20917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20918 vmaxq_f32 (float32x4_t __a, float32x4_t __b)
20919 {
20920   return __builtin_aarch64_smax_nanv4sf (__a, __b);
20921 }
20922
20923 __extension__ extern __inline float64x2_t
20924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20925 vmaxq_f64 (float64x2_t __a, float64x2_t __b)
20926 {
20927   return __builtin_aarch64_smax_nanv2df (__a, __b);
20928 }
20929
20930 __extension__ extern __inline int8x16_t
20931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20932 vmaxq_s8 (int8x16_t __a, int8x16_t __b)
20933 {
20934   return __builtin_aarch64_smaxv16qi (__a, __b);
20935 }
20936
20937 __extension__ extern __inline int16x8_t
20938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20939 vmaxq_s16 (int16x8_t __a, int16x8_t __b)
20940 {
20941   return __builtin_aarch64_smaxv8hi (__a, __b);
20942 }
20943
20944 __extension__ extern __inline int32x4_t
20945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20946 vmaxq_s32 (int32x4_t __a, int32x4_t __b)
20947 {
20948   return __builtin_aarch64_smaxv4si (__a, __b);
20949 }
20950
20951 __extension__ extern __inline uint8x16_t
20952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20953 vmaxq_u8 (uint8x16_t __a, uint8x16_t __b)
20954 {
20955   return (uint8x16_t) __builtin_aarch64_umaxv16qi ((int8x16_t) __a,
20956                                                    (int8x16_t) __b);
20957 }
20958
20959 __extension__ extern __inline uint16x8_t
20960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20961 vmaxq_u16 (uint16x8_t __a, uint16x8_t __b)
20962 {
20963   return (uint16x8_t) __builtin_aarch64_umaxv8hi ((int16x8_t) __a,
20964                                                   (int16x8_t) __b);
20965 }
20966
20967 __extension__ extern __inline uint32x4_t
20968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20969 vmaxq_u32 (uint32x4_t __a, uint32x4_t __b)
20970 {
20971   return (uint32x4_t) __builtin_aarch64_umaxv4si ((int32x4_t) __a,
20972                                                   (int32x4_t) __b);
20973 }
20974 /* vmulx */
20975
20976 __extension__ extern __inline float32x2_t
20977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20978 vmulx_f32 (float32x2_t __a, float32x2_t __b)
20979 {
20980   return __builtin_aarch64_fmulxv2sf (__a, __b);
20981 }
20982
20983 __extension__ extern __inline float32x4_t
20984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20985 vmulxq_f32 (float32x4_t __a, float32x4_t __b)
20986 {
20987   return __builtin_aarch64_fmulxv4sf (__a, __b);
20988 }
20989
20990 __extension__ extern __inline float64x1_t
20991 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20992 vmulx_f64 (float64x1_t __a, float64x1_t __b)
20993 {
20994   return (float64x1_t) {__builtin_aarch64_fmulxdf (__a[0], __b[0])};
20995 }
20996
20997 __extension__ extern __inline float64x2_t
20998 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20999 vmulxq_f64 (float64x2_t __a, float64x2_t __b)
21000 {
21001   return __builtin_aarch64_fmulxv2df (__a, __b);
21002 }
21003
21004 __extension__ extern __inline float32_t
21005 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21006 vmulxs_f32 (float32_t __a, float32_t __b)
21007 {
21008   return __builtin_aarch64_fmulxsf (__a, __b);
21009 }
21010
21011 __extension__ extern __inline float64_t
21012 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21013 vmulxd_f64 (float64_t __a, float64_t __b)
21014 {
21015   return __builtin_aarch64_fmulxdf (__a, __b);
21016 }
21017
21018 __extension__ extern __inline float32x2_t
21019 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21020 vmulx_lane_f32 (float32x2_t __a, float32x2_t __v, const int __lane)
21021 {
21022   return vmulx_f32 (__a, __aarch64_vdup_lane_f32 (__v, __lane));
21023 }
21024
21025 __extension__ extern __inline float64x1_t
21026 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21027 vmulx_lane_f64 (float64x1_t __a, float64x1_t __v, const int __lane)
21028 {
21029   return vmulx_f64 (__a, __aarch64_vdup_lane_f64 (__v, __lane));
21030 }
21031
21032 __extension__ extern __inline float32x4_t
21033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21034 vmulxq_lane_f32 (float32x4_t __a, float32x2_t __v, const int __lane)
21035 {
21036   return vmulxq_f32 (__a, __aarch64_vdupq_lane_f32 (__v, __lane));
21037 }
21038
21039 __extension__ extern __inline float64x2_t
21040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21041 vmulxq_lane_f64 (float64x2_t __a, float64x1_t __v, const int __lane)
21042 {
21043   return vmulxq_f64 (__a, __aarch64_vdupq_lane_f64 (__v, __lane));
21044 }
21045
21046 __extension__ extern __inline float32x2_t
21047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21048 vmulx_laneq_f32 (float32x2_t __a, float32x4_t __v, const int __lane)
21049 {
21050   return vmulx_f32 (__a, __aarch64_vdup_laneq_f32 (__v, __lane));
21051 }
21052
21053 __extension__ extern __inline float64x1_t
21054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21055 vmulx_laneq_f64 (float64x1_t __a, float64x2_t __v, const int __lane)
21056 {
21057   return vmulx_f64 (__a, __aarch64_vdup_laneq_f64 (__v, __lane));
21058 }
21059
21060 __extension__ extern __inline float32x4_t
21061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21062 vmulxq_laneq_f32 (float32x4_t __a, float32x4_t __v, const int __lane)
21063 {
21064   return vmulxq_f32 (__a, __aarch64_vdupq_laneq_f32 (__v, __lane));
21065 }
21066
21067 __extension__ extern __inline float64x2_t
21068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21069 vmulxq_laneq_f64 (float64x2_t __a, float64x2_t __v, const int __lane)
21070 {
21071   return vmulxq_f64 (__a, __aarch64_vdupq_laneq_f64 (__v, __lane));
21072 }
21073
21074 __extension__ extern __inline float32_t
21075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21076 vmulxs_lane_f32 (float32_t __a, float32x2_t __v, const int __lane)
21077 {
21078   return vmulxs_f32 (__a, __aarch64_vget_lane_any (__v, __lane));
21079 }
21080
21081 __extension__ extern __inline float32_t
21082 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21083 vmulxs_laneq_f32 (float32_t __a, float32x4_t __v, const int __lane)
21084 {
21085   return vmulxs_f32 (__a, __aarch64_vget_lane_any (__v, __lane));
21086 }
21087
21088 __extension__ extern __inline float64_t
21089 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21090 vmulxd_lane_f64 (float64_t __a, float64x1_t __v, const int __lane)
21091 {
21092   return vmulxd_f64 (__a, __aarch64_vget_lane_any (__v, __lane));
21093 }
21094
21095 __extension__ extern __inline float64_t
21096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21097 vmulxd_laneq_f64 (float64_t __a, float64x2_t __v, const int __lane)
21098 {
21099   return vmulxd_f64 (__a, __aarch64_vget_lane_any (__v, __lane));
21100 }
21101
21102 /* vpmax  */
21103
21104 __extension__ extern __inline int8x8_t
21105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21106 vpmax_s8 (int8x8_t a, int8x8_t b)
21107 {
21108   return __builtin_aarch64_smaxpv8qi (a, b);
21109 }
21110
21111 __extension__ extern __inline int16x4_t
21112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21113 vpmax_s16 (int16x4_t a, int16x4_t b)
21114 {
21115   return __builtin_aarch64_smaxpv4hi (a, b);
21116 }
21117
21118 __extension__ extern __inline int32x2_t
21119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21120 vpmax_s32 (int32x2_t a, int32x2_t b)
21121 {
21122   return __builtin_aarch64_smaxpv2si (a, b);
21123 }
21124
21125 __extension__ extern __inline uint8x8_t
21126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21127 vpmax_u8 (uint8x8_t a, uint8x8_t b)
21128 {
21129   return (uint8x8_t) __builtin_aarch64_umaxpv8qi ((int8x8_t) a,
21130                                                   (int8x8_t) b);
21131 }
21132
21133 __extension__ extern __inline uint16x4_t
21134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21135 vpmax_u16 (uint16x4_t a, uint16x4_t b)
21136 {
21137   return (uint16x4_t) __builtin_aarch64_umaxpv4hi ((int16x4_t) a,
21138                                                    (int16x4_t) b);
21139 }
21140
21141 __extension__ extern __inline uint32x2_t
21142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21143 vpmax_u32 (uint32x2_t a, uint32x2_t b)
21144 {
21145   return (uint32x2_t) __builtin_aarch64_umaxpv2si ((int32x2_t) a,
21146                                                    (int32x2_t) b);
21147 }
21148
21149 __extension__ extern __inline int8x16_t
21150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21151 vpmaxq_s8 (int8x16_t a, int8x16_t b)
21152 {
21153   return __builtin_aarch64_smaxpv16qi (a, b);
21154 }
21155
21156 __extension__ extern __inline int16x8_t
21157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21158 vpmaxq_s16 (int16x8_t a, int16x8_t b)
21159 {
21160   return __builtin_aarch64_smaxpv8hi (a, b);
21161 }
21162
21163 __extension__ extern __inline int32x4_t
21164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21165 vpmaxq_s32 (int32x4_t a, int32x4_t b)
21166 {
21167   return __builtin_aarch64_smaxpv4si (a, b);
21168 }
21169
21170 __extension__ extern __inline uint8x16_t
21171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21172 vpmaxq_u8 (uint8x16_t a, uint8x16_t b)
21173 {
21174   return (uint8x16_t) __builtin_aarch64_umaxpv16qi ((int8x16_t) a,
21175                                                     (int8x16_t) b);
21176 }
21177
21178 __extension__ extern __inline uint16x8_t
21179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21180 vpmaxq_u16 (uint16x8_t a, uint16x8_t b)
21181 {
21182   return (uint16x8_t) __builtin_aarch64_umaxpv8hi ((int16x8_t) a,
21183                                                    (int16x8_t) b);
21184 }
21185
21186 __extension__ extern __inline uint32x4_t
21187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21188 vpmaxq_u32 (uint32x4_t a, uint32x4_t b)
21189 {
21190   return (uint32x4_t) __builtin_aarch64_umaxpv4si ((int32x4_t) a,
21191                                                    (int32x4_t) b);
21192 }
21193
21194 __extension__ extern __inline float32x2_t
21195 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21196 vpmax_f32 (float32x2_t a, float32x2_t b)
21197 {
21198   return __builtin_aarch64_smax_nanpv2sf (a, b);
21199 }
21200
21201 __extension__ extern __inline float32x4_t
21202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21203 vpmaxq_f32 (float32x4_t a, float32x4_t b)
21204 {
21205   return __builtin_aarch64_smax_nanpv4sf (a, b);
21206 }
21207
21208 __extension__ extern __inline float64x2_t
21209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21210 vpmaxq_f64 (float64x2_t a, float64x2_t b)
21211 {
21212   return __builtin_aarch64_smax_nanpv2df (a, b);
21213 }
21214
21215 __extension__ extern __inline float64_t
21216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21217 vpmaxqd_f64 (float64x2_t a)
21218 {
21219   return __builtin_aarch64_reduc_smax_nan_scal_v2df (a);
21220 }
21221
21222 __extension__ extern __inline float32_t
21223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21224 vpmaxs_f32 (float32x2_t a)
21225 {
21226   return __builtin_aarch64_reduc_smax_nan_scal_v2sf (a);
21227 }
21228
21229 /* vpmaxnm  */
21230
21231 __extension__ extern __inline float32x2_t
21232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21233 vpmaxnm_f32 (float32x2_t a, float32x2_t b)
21234 {
21235   return __builtin_aarch64_smaxpv2sf (a, b);
21236 }
21237
21238 __extension__ extern __inline float32x4_t
21239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21240 vpmaxnmq_f32 (float32x4_t a, float32x4_t b)
21241 {
21242   return __builtin_aarch64_smaxpv4sf (a, b);
21243 }
21244
21245 __extension__ extern __inline float64x2_t
21246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21247 vpmaxnmq_f64 (float64x2_t a, float64x2_t b)
21248 {
21249   return __builtin_aarch64_smaxpv2df (a, b);
21250 }
21251
21252 __extension__ extern __inline float64_t
21253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21254 vpmaxnmqd_f64 (float64x2_t a)
21255 {
21256   return __builtin_aarch64_reduc_smax_scal_v2df (a);
21257 }
21258
21259 __extension__ extern __inline float32_t
21260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21261 vpmaxnms_f32 (float32x2_t a)
21262 {
21263   return __builtin_aarch64_reduc_smax_scal_v2sf (a);
21264 }
21265
21266 /* vpmin  */
21267
21268 __extension__ extern __inline int8x8_t
21269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21270 vpmin_s8 (int8x8_t a, int8x8_t b)
21271 {
21272   return __builtin_aarch64_sminpv8qi (a, b);
21273 }
21274
21275 __extension__ extern __inline int16x4_t
21276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21277 vpmin_s16 (int16x4_t a, int16x4_t b)
21278 {
21279   return __builtin_aarch64_sminpv4hi (a, b);
21280 }
21281
21282 __extension__ extern __inline int32x2_t
21283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21284 vpmin_s32 (int32x2_t a, int32x2_t b)
21285 {
21286   return __builtin_aarch64_sminpv2si (a, b);
21287 }
21288
21289 __extension__ extern __inline uint8x8_t
21290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21291 vpmin_u8 (uint8x8_t a, uint8x8_t b)
21292 {
21293   return (uint8x8_t) __builtin_aarch64_uminpv8qi ((int8x8_t) a,
21294                                                   (int8x8_t) b);
21295 }
21296
21297 __extension__ extern __inline uint16x4_t
21298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21299 vpmin_u16 (uint16x4_t a, uint16x4_t b)
21300 {
21301   return (uint16x4_t) __builtin_aarch64_uminpv4hi ((int16x4_t) a,
21302                                                    (int16x4_t) b);
21303 }
21304
21305 __extension__ extern __inline uint32x2_t
21306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21307 vpmin_u32 (uint32x2_t a, uint32x2_t b)
21308 {
21309   return (uint32x2_t) __builtin_aarch64_uminpv2si ((int32x2_t) a,
21310                                                    (int32x2_t) b);
21311 }
21312
21313 __extension__ extern __inline int8x16_t
21314 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21315 vpminq_s8 (int8x16_t a, int8x16_t b)
21316 {
21317   return __builtin_aarch64_sminpv16qi (a, b);
21318 }
21319
21320 __extension__ extern __inline int16x8_t
21321 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21322 vpminq_s16 (int16x8_t a, int16x8_t b)
21323 {
21324   return __builtin_aarch64_sminpv8hi (a, b);
21325 }
21326
21327 __extension__ extern __inline int32x4_t
21328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21329 vpminq_s32 (int32x4_t a, int32x4_t b)
21330 {
21331   return __builtin_aarch64_sminpv4si (a, b);
21332 }
21333
21334 __extension__ extern __inline uint8x16_t
21335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21336 vpminq_u8 (uint8x16_t a, uint8x16_t b)
21337 {
21338   return (uint8x16_t) __builtin_aarch64_uminpv16qi ((int8x16_t) a,
21339                                                     (int8x16_t) b);
21340 }
21341
21342 __extension__ extern __inline uint16x8_t
21343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21344 vpminq_u16 (uint16x8_t a, uint16x8_t b)
21345 {
21346   return (uint16x8_t) __builtin_aarch64_uminpv8hi ((int16x8_t) a,
21347                                                    (int16x8_t) b);
21348 }
21349
21350 __extension__ extern __inline uint32x4_t
21351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21352 vpminq_u32 (uint32x4_t a, uint32x4_t b)
21353 {
21354   return (uint32x4_t) __builtin_aarch64_uminpv4si ((int32x4_t) a,
21355                                                    (int32x4_t) b);
21356 }
21357
21358 __extension__ extern __inline float32x2_t
21359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21360 vpmin_f32 (float32x2_t a, float32x2_t b)
21361 {
21362   return __builtin_aarch64_smin_nanpv2sf (a, b);
21363 }
21364
21365 __extension__ extern __inline float32x4_t
21366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21367 vpminq_f32 (float32x4_t a, float32x4_t b)
21368 {
21369   return __builtin_aarch64_smin_nanpv4sf (a, b);
21370 }
21371
21372 __extension__ extern __inline float64x2_t
21373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21374 vpminq_f64 (float64x2_t a, float64x2_t b)
21375 {
21376   return __builtin_aarch64_smin_nanpv2df (a, b);
21377 }
21378
21379 __extension__ extern __inline float64_t
21380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21381 vpminqd_f64 (float64x2_t a)
21382 {
21383   return __builtin_aarch64_reduc_smin_nan_scal_v2df (a);
21384 }
21385
21386 __extension__ extern __inline float32_t
21387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21388 vpmins_f32 (float32x2_t a)
21389 {
21390   return __builtin_aarch64_reduc_smin_nan_scal_v2sf (a);
21391 }
21392
21393 /* vpminnm  */
21394
21395 __extension__ extern __inline float32x2_t
21396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21397 vpminnm_f32 (float32x2_t a, float32x2_t b)
21398 {
21399   return __builtin_aarch64_sminpv2sf (a, b);
21400 }
21401
21402 __extension__ extern __inline float32x4_t
21403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21404 vpminnmq_f32 (float32x4_t a, float32x4_t b)
21405 {
21406   return __builtin_aarch64_sminpv4sf (a, b);
21407 }
21408
21409 __extension__ extern __inline float64x2_t
21410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21411 vpminnmq_f64 (float64x2_t a, float64x2_t b)
21412 {
21413   return __builtin_aarch64_sminpv2df (a, b);
21414 }
21415
21416 __extension__ extern __inline float64_t
21417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21418 vpminnmqd_f64 (float64x2_t a)
21419 {
21420   return __builtin_aarch64_reduc_smin_scal_v2df (a);
21421 }
21422
21423 __extension__ extern __inline float32_t
21424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21425 vpminnms_f32 (float32x2_t a)
21426 {
21427   return __builtin_aarch64_reduc_smin_scal_v2sf (a);
21428 }
21429
21430 /* vmaxnm  */
21431
21432 __extension__ extern __inline float32x2_t
21433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21434 vmaxnm_f32 (float32x2_t __a, float32x2_t __b)
21435 {
21436   return __builtin_aarch64_fmaxv2sf (__a, __b);
21437 }
21438
21439 __extension__ extern __inline float64x1_t
21440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21441 vmaxnm_f64 (float64x1_t __a, float64x1_t __b)
21442 {
21443   return (float64x1_t)
21444     { __builtin_aarch64_fmaxdf (vget_lane_f64 (__a, 0),
21445                                 vget_lane_f64 (__b, 0)) };
21446 }
21447
21448 __extension__ extern __inline float32x4_t
21449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21450 vmaxnmq_f32 (float32x4_t __a, float32x4_t __b)
21451 {
21452   return __builtin_aarch64_fmaxv4sf (__a, __b);
21453 }
21454
21455 __extension__ extern __inline float64x2_t
21456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21457 vmaxnmq_f64 (float64x2_t __a, float64x2_t __b)
21458 {
21459   return __builtin_aarch64_fmaxv2df (__a, __b);
21460 }
21461
21462 /* vmaxv  */
21463
21464 __extension__ extern __inline float32_t
21465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21466 vmaxv_f32 (float32x2_t __a)
21467 {
21468   return __builtin_aarch64_reduc_smax_nan_scal_v2sf (__a);
21469 }
21470
21471 __extension__ extern __inline int8_t
21472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21473 vmaxv_s8 (int8x8_t __a)
21474 {
21475   return __builtin_aarch64_reduc_smax_scal_v8qi (__a);
21476 }
21477
21478 __extension__ extern __inline int16_t
21479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21480 vmaxv_s16 (int16x4_t __a)
21481 {
21482   return __builtin_aarch64_reduc_smax_scal_v4hi (__a);
21483 }
21484
21485 __extension__ extern __inline int32_t
21486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21487 vmaxv_s32 (int32x2_t __a)
21488 {
21489   return __builtin_aarch64_reduc_smax_scal_v2si (__a);
21490 }
21491
21492 __extension__ extern __inline uint8_t
21493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21494 vmaxv_u8 (uint8x8_t __a)
21495 {
21496   return __builtin_aarch64_reduc_umax_scal_v8qi_uu (__a);
21497 }
21498
21499 __extension__ extern __inline uint16_t
21500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21501 vmaxv_u16 (uint16x4_t __a)
21502 {
21503   return __builtin_aarch64_reduc_umax_scal_v4hi_uu (__a);
21504 }
21505
21506 __extension__ extern __inline uint32_t
21507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21508 vmaxv_u32 (uint32x2_t __a)
21509 {
21510   return __builtin_aarch64_reduc_umax_scal_v2si_uu (__a);
21511 }
21512
21513 __extension__ extern __inline float32_t
21514 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21515 vmaxvq_f32 (float32x4_t __a)
21516 {
21517   return __builtin_aarch64_reduc_smax_nan_scal_v4sf (__a);
21518 }
21519
21520 __extension__ extern __inline float64_t
21521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21522 vmaxvq_f64 (float64x2_t __a)
21523 {
21524   return __builtin_aarch64_reduc_smax_nan_scal_v2df (__a);
21525 }
21526
21527 __extension__ extern __inline int8_t
21528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21529 vmaxvq_s8 (int8x16_t __a)
21530 {
21531   return __builtin_aarch64_reduc_smax_scal_v16qi (__a);
21532 }
21533
21534 __extension__ extern __inline int16_t
21535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21536 vmaxvq_s16 (int16x8_t __a)
21537 {
21538   return __builtin_aarch64_reduc_smax_scal_v8hi (__a);
21539 }
21540
21541 __extension__ extern __inline int32_t
21542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21543 vmaxvq_s32 (int32x4_t __a)
21544 {
21545   return __builtin_aarch64_reduc_smax_scal_v4si (__a);
21546 }
21547
21548 __extension__ extern __inline uint8_t
21549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21550 vmaxvq_u8 (uint8x16_t __a)
21551 {
21552   return __builtin_aarch64_reduc_umax_scal_v16qi_uu (__a);
21553 }
21554
21555 __extension__ extern __inline uint16_t
21556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21557 vmaxvq_u16 (uint16x8_t __a)
21558 {
21559   return __builtin_aarch64_reduc_umax_scal_v8hi_uu (__a);
21560 }
21561
21562 __extension__ extern __inline uint32_t
21563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21564 vmaxvq_u32 (uint32x4_t __a)
21565 {
21566   return __builtin_aarch64_reduc_umax_scal_v4si_uu (__a);
21567 }
21568
21569 /* vmaxnmv  */
21570
21571 __extension__ extern __inline float32_t
21572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21573 vmaxnmv_f32 (float32x2_t __a)
21574 {
21575   return __builtin_aarch64_reduc_smax_scal_v2sf (__a);
21576 }
21577
21578 __extension__ extern __inline float32_t
21579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21580 vmaxnmvq_f32 (float32x4_t __a)
21581 {
21582   return __builtin_aarch64_reduc_smax_scal_v4sf (__a);
21583 }
21584
21585 __extension__ extern __inline float64_t
21586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21587 vmaxnmvq_f64 (float64x2_t __a)
21588 {
21589   return __builtin_aarch64_reduc_smax_scal_v2df (__a);
21590 }
21591
21592 /* vmin  */
21593
21594 __extension__ extern __inline float32x2_t
21595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21596 vmin_f32 (float32x2_t __a, float32x2_t __b)
21597 {
21598   return __builtin_aarch64_smin_nanv2sf (__a, __b);
21599 }
21600
21601 __extension__ extern __inline float64x1_t
21602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21603 vmin_f64 (float64x1_t __a, float64x1_t __b)
21604 {
21605     return (float64x1_t)
21606           { __builtin_aarch64_smin_nandf (vget_lane_f64 (__a, 0),
21607                                           vget_lane_f64 (__b, 0)) };
21608 }
21609
21610 __extension__ extern __inline int8x8_t
21611 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21612 vmin_s8 (int8x8_t __a, int8x8_t __b)
21613 {
21614   return __builtin_aarch64_sminv8qi (__a, __b);
21615 }
21616
21617 __extension__ extern __inline int16x4_t
21618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21619 vmin_s16 (int16x4_t __a, int16x4_t __b)
21620 {
21621   return __builtin_aarch64_sminv4hi (__a, __b);
21622 }
21623
21624 __extension__ extern __inline int32x2_t
21625 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21626 vmin_s32 (int32x2_t __a, int32x2_t __b)
21627 {
21628   return __builtin_aarch64_sminv2si (__a, __b);
21629 }
21630
21631 __extension__ extern __inline uint8x8_t
21632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21633 vmin_u8 (uint8x8_t __a, uint8x8_t __b)
21634 {
21635   return (uint8x8_t) __builtin_aarch64_uminv8qi ((int8x8_t) __a,
21636                                                  (int8x8_t) __b);
21637 }
21638
21639 __extension__ extern __inline uint16x4_t
21640 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21641 vmin_u16 (uint16x4_t __a, uint16x4_t __b)
21642 {
21643   return (uint16x4_t) __builtin_aarch64_uminv4hi ((int16x4_t) __a,
21644                                                   (int16x4_t) __b);
21645 }
21646
21647 __extension__ extern __inline uint32x2_t
21648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21649 vmin_u32 (uint32x2_t __a, uint32x2_t __b)
21650 {
21651   return (uint32x2_t) __builtin_aarch64_uminv2si ((int32x2_t) __a,
21652                                                   (int32x2_t) __b);
21653 }
21654
21655 __extension__ extern __inline float32x4_t
21656 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21657 vminq_f32 (float32x4_t __a, float32x4_t __b)
21658 {
21659   return __builtin_aarch64_smin_nanv4sf (__a, __b);
21660 }
21661
21662 __extension__ extern __inline float64x2_t
21663 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21664 vminq_f64 (float64x2_t __a, float64x2_t __b)
21665 {
21666   return __builtin_aarch64_smin_nanv2df (__a, __b);
21667 }
21668
21669 __extension__ extern __inline int8x16_t
21670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21671 vminq_s8 (int8x16_t __a, int8x16_t __b)
21672 {
21673   return __builtin_aarch64_sminv16qi (__a, __b);
21674 }
21675
21676 __extension__ extern __inline int16x8_t
21677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21678 vminq_s16 (int16x8_t __a, int16x8_t __b)
21679 {
21680   return __builtin_aarch64_sminv8hi (__a, __b);
21681 }
21682
21683 __extension__ extern __inline int32x4_t
21684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21685 vminq_s32 (int32x4_t __a, int32x4_t __b)
21686 {
21687   return __builtin_aarch64_sminv4si (__a, __b);
21688 }
21689
21690 __extension__ extern __inline uint8x16_t
21691 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21692 vminq_u8 (uint8x16_t __a, uint8x16_t __b)
21693 {
21694   return (uint8x16_t) __builtin_aarch64_uminv16qi ((int8x16_t) __a,
21695                                                    (int8x16_t) __b);
21696 }
21697
21698 __extension__ extern __inline uint16x8_t
21699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21700 vminq_u16 (uint16x8_t __a, uint16x8_t __b)
21701 {
21702   return (uint16x8_t) __builtin_aarch64_uminv8hi ((int16x8_t) __a,
21703                                                   (int16x8_t) __b);
21704 }
21705
21706 __extension__ extern __inline uint32x4_t
21707 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21708 vminq_u32 (uint32x4_t __a, uint32x4_t __b)
21709 {
21710   return (uint32x4_t) __builtin_aarch64_uminv4si ((int32x4_t) __a,
21711                                                   (int32x4_t) __b);
21712 }
21713
21714 /* vminnm  */
21715
21716 __extension__ extern __inline float32x2_t
21717 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21718 vminnm_f32 (float32x2_t __a, float32x2_t __b)
21719 {
21720   return __builtin_aarch64_fminv2sf (__a, __b);
21721 }
21722
21723 __extension__ extern __inline float64x1_t
21724 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21725 vminnm_f64 (float64x1_t __a, float64x1_t __b)
21726 {
21727   return (float64x1_t)
21728     { __builtin_aarch64_fmindf (vget_lane_f64 (__a, 0),
21729                                 vget_lane_f64 (__b, 0)) };
21730 }
21731
21732 __extension__ extern __inline float32x4_t
21733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21734 vminnmq_f32 (float32x4_t __a, float32x4_t __b)
21735 {
21736   return __builtin_aarch64_fminv4sf (__a, __b);
21737 }
21738
21739 __extension__ extern __inline float64x2_t
21740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21741 vminnmq_f64 (float64x2_t __a, float64x2_t __b)
21742 {
21743   return __builtin_aarch64_fminv2df (__a, __b);
21744 }
21745
21746 /* vminv  */
21747
21748 __extension__ extern __inline float32_t
21749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21750 vminv_f32 (float32x2_t __a)
21751 {
21752   return __builtin_aarch64_reduc_smin_nan_scal_v2sf (__a);
21753 }
21754
21755 __extension__ extern __inline int8_t
21756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21757 vminv_s8 (int8x8_t __a)
21758 {
21759   return __builtin_aarch64_reduc_smin_scal_v8qi (__a);
21760 }
21761
21762 __extension__ extern __inline int16_t
21763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21764 vminv_s16 (int16x4_t __a)
21765 {
21766   return __builtin_aarch64_reduc_smin_scal_v4hi (__a);
21767 }
21768
21769 __extension__ extern __inline int32_t
21770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21771 vminv_s32 (int32x2_t __a)
21772 {
21773   return __builtin_aarch64_reduc_smin_scal_v2si (__a);
21774 }
21775
21776 __extension__ extern __inline uint8_t
21777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21778 vminv_u8 (uint8x8_t __a)
21779 {
21780   return __builtin_aarch64_reduc_umin_scal_v8qi_uu (__a);
21781 }
21782
21783 __extension__ extern __inline uint16_t
21784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21785 vminv_u16 (uint16x4_t __a)
21786 {
21787   return __builtin_aarch64_reduc_umin_scal_v4hi_uu (__a);
21788 }
21789
21790 __extension__ extern __inline uint32_t
21791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21792 vminv_u32 (uint32x2_t __a)
21793 {
21794   return __builtin_aarch64_reduc_umin_scal_v2si_uu (__a);
21795 }
21796
21797 __extension__ extern __inline float32_t
21798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21799 vminvq_f32 (float32x4_t __a)
21800 {
21801   return __builtin_aarch64_reduc_smin_nan_scal_v4sf (__a);
21802 }
21803
21804 __extension__ extern __inline float64_t
21805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21806 vminvq_f64 (float64x2_t __a)
21807 {
21808   return __builtin_aarch64_reduc_smin_nan_scal_v2df (__a);
21809 }
21810
21811 __extension__ extern __inline int8_t
21812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21813 vminvq_s8 (int8x16_t __a)
21814 {
21815   return __builtin_aarch64_reduc_smin_scal_v16qi (__a);
21816 }
21817
21818 __extension__ extern __inline int16_t
21819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21820 vminvq_s16 (int16x8_t __a)
21821 {
21822   return __builtin_aarch64_reduc_smin_scal_v8hi (__a);
21823 }
21824
21825 __extension__ extern __inline int32_t
21826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21827 vminvq_s32 (int32x4_t __a)
21828 {
21829   return __builtin_aarch64_reduc_smin_scal_v4si (__a);
21830 }
21831
21832 __extension__ extern __inline uint8_t
21833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21834 vminvq_u8 (uint8x16_t __a)
21835 {
21836   return __builtin_aarch64_reduc_umin_scal_v16qi_uu (__a);
21837 }
21838
21839 __extension__ extern __inline uint16_t
21840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21841 vminvq_u16 (uint16x8_t __a)
21842 {
21843   return __builtin_aarch64_reduc_umin_scal_v8hi_uu (__a);
21844 }
21845
21846 __extension__ extern __inline uint32_t
21847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21848 vminvq_u32 (uint32x4_t __a)
21849 {
21850   return __builtin_aarch64_reduc_umin_scal_v4si_uu (__a);
21851 }
21852
21853 /* vminnmv  */
21854
21855 __extension__ extern __inline float32_t
21856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21857 vminnmv_f32 (float32x2_t __a)
21858 {
21859   return __builtin_aarch64_reduc_smin_scal_v2sf (__a);
21860 }
21861
21862 __extension__ extern __inline float32_t
21863 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21864 vminnmvq_f32 (float32x4_t __a)
21865 {
21866   return __builtin_aarch64_reduc_smin_scal_v4sf (__a);
21867 }
21868
21869 __extension__ extern __inline float64_t
21870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21871 vminnmvq_f64 (float64x2_t __a)
21872 {
21873   return __builtin_aarch64_reduc_smin_scal_v2df (__a);
21874 }
21875
21876 /* vmla */
21877
21878 __extension__ extern __inline float32x2_t
21879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21880 vmla_f32 (float32x2_t a, float32x2_t b, float32x2_t c)
21881 {
21882   return a + b * c;
21883 }
21884
21885 __extension__ extern __inline float64x1_t
21886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21887 vmla_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
21888 {
21889   return __a + __b * __c;
21890 }
21891
21892 __extension__ extern __inline float32x4_t
21893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21894 vmlaq_f32 (float32x4_t a, float32x4_t b, float32x4_t c)
21895 {
21896   return a + b * c;
21897 }
21898
21899 __extension__ extern __inline float64x2_t
21900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21901 vmlaq_f64 (float64x2_t a, float64x2_t b, float64x2_t c)
21902 {
21903   return a + b * c;
21904 }
21905
21906 /* vmla_lane  */
21907
21908 __extension__ extern __inline float32x2_t
21909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21910 vmla_lane_f32 (float32x2_t __a, float32x2_t __b,
21911                float32x2_t __c, const int __lane)
21912 {
21913   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21914 }
21915
21916 __extension__ extern __inline int16x4_t
21917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21918 vmla_lane_s16 (int16x4_t __a, int16x4_t __b,
21919                 int16x4_t __c, const int __lane)
21920 {
21921   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21922 }
21923
21924 __extension__ extern __inline int32x2_t
21925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21926 vmla_lane_s32 (int32x2_t __a, int32x2_t __b,
21927                 int32x2_t __c, const int __lane)
21928 {
21929   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21930 }
21931
21932 __extension__ extern __inline uint16x4_t
21933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21934 vmla_lane_u16 (uint16x4_t __a, uint16x4_t __b,
21935                 uint16x4_t __c, const int __lane)
21936 {
21937   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21938 }
21939
21940 __extension__ extern __inline uint32x2_t
21941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21942 vmla_lane_u32 (uint32x2_t __a, uint32x2_t __b,
21943                uint32x2_t __c, const int __lane)
21944 {
21945   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21946 }
21947
21948 /* vmla_laneq  */
21949
21950 __extension__ extern __inline float32x2_t
21951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21952 vmla_laneq_f32 (float32x2_t __a, float32x2_t __b,
21953                 float32x4_t __c, const int __lane)
21954 {
21955   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21956 }
21957
21958 __extension__ extern __inline int16x4_t
21959 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21960 vmla_laneq_s16 (int16x4_t __a, int16x4_t __b,
21961                 int16x8_t __c, const int __lane)
21962 {
21963   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21964 }
21965
21966 __extension__ extern __inline int32x2_t
21967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21968 vmla_laneq_s32 (int32x2_t __a, int32x2_t __b,
21969                 int32x4_t __c, const int __lane)
21970 {
21971   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21972 }
21973
21974 __extension__ extern __inline uint16x4_t
21975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21976 vmla_laneq_u16 (uint16x4_t __a, uint16x4_t __b,
21977                 uint16x8_t __c, const int __lane)
21978 {
21979   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21980 }
21981
21982 __extension__ extern __inline uint32x2_t
21983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21984 vmla_laneq_u32 (uint32x2_t __a, uint32x2_t __b,
21985                 uint32x4_t __c, const int __lane)
21986 {
21987   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21988 }
21989
21990 /* vmlaq_lane  */
21991
21992 __extension__ extern __inline float32x4_t
21993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21994 vmlaq_lane_f32 (float32x4_t __a, float32x4_t __b,
21995                 float32x2_t __c, const int __lane)
21996 {
21997   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21998 }
21999
22000 __extension__ extern __inline int16x8_t
22001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22002 vmlaq_lane_s16 (int16x8_t __a, int16x8_t __b,
22003                 int16x4_t __c, const int __lane)
22004 {
22005   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22006 }
22007
22008 __extension__ extern __inline int32x4_t
22009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22010 vmlaq_lane_s32 (int32x4_t __a, int32x4_t __b,
22011                 int32x2_t __c, const int __lane)
22012 {
22013   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22014 }
22015
22016 __extension__ extern __inline uint16x8_t
22017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22018 vmlaq_lane_u16 (uint16x8_t __a, uint16x8_t __b,
22019                 uint16x4_t __c, const int __lane)
22020 {
22021   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22022 }
22023
22024 __extension__ extern __inline uint32x4_t
22025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22026 vmlaq_lane_u32 (uint32x4_t __a, uint32x4_t __b,
22027                 uint32x2_t __c, const int __lane)
22028 {
22029   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22030 }
22031
22032   /* vmlaq_laneq  */
22033
22034 __extension__ extern __inline float32x4_t
22035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22036 vmlaq_laneq_f32 (float32x4_t __a, float32x4_t __b,
22037                  float32x4_t __c, const int __lane)
22038 {
22039   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22040 }
22041
22042 __extension__ extern __inline int16x8_t
22043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22044 vmlaq_laneq_s16 (int16x8_t __a, int16x8_t __b,
22045                 int16x8_t __c, const int __lane)
22046 {
22047   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22048 }
22049
22050 __extension__ extern __inline int32x4_t
22051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22052 vmlaq_laneq_s32 (int32x4_t __a, int32x4_t __b,
22053                 int32x4_t __c, const int __lane)
22054 {
22055   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22056 }
22057
22058 __extension__ extern __inline uint16x8_t
22059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22060 vmlaq_laneq_u16 (uint16x8_t __a, uint16x8_t __b,
22061                 uint16x8_t __c, const int __lane)
22062 {
22063   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22064 }
22065
22066 __extension__ extern __inline uint32x4_t
22067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22068 vmlaq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
22069                 uint32x4_t __c, const int __lane)
22070 {
22071   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
22072 }
22073
22074 /* vmls  */
22075
22076 __extension__ extern __inline float32x2_t
22077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22078 vmls_f32 (float32x2_t a, float32x2_t b, float32x2_t c)
22079 {
22080   return a - b * c;
22081 }
22082
22083 __extension__ extern __inline float64x1_t
22084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22085 vmls_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
22086 {
22087   return __a - __b * __c;
22088 }
22089
22090 __extension__ extern __inline float32x4_t
22091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22092 vmlsq_f32 (float32x4_t a, float32x4_t b, float32x4_t c)
22093 {
22094   return a - b * c;
22095 }
22096
22097 __extension__ extern __inline float64x2_t
22098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22099 vmlsq_f64 (float64x2_t a, float64x2_t b, float64x2_t c)
22100 {
22101   return a - b * c;
22102 }
22103
22104 /* vmls_lane  */
22105
22106 __extension__ extern __inline float32x2_t
22107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22108 vmls_lane_f32 (float32x2_t __a, float32x2_t __b,
22109                float32x2_t __c, const int __lane)
22110 {
22111   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22112 }
22113
22114 __extension__ extern __inline int16x4_t
22115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22116 vmls_lane_s16 (int16x4_t __a, int16x4_t __b,
22117                 int16x4_t __c, const int __lane)
22118 {
22119   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22120 }
22121
22122 __extension__ extern __inline int32x2_t
22123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22124 vmls_lane_s32 (int32x2_t __a, int32x2_t __b,
22125                 int32x2_t __c, const int __lane)
22126 {
22127   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22128 }
22129
22130 __extension__ extern __inline uint16x4_t
22131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22132 vmls_lane_u16 (uint16x4_t __a, uint16x4_t __b,
22133                 uint16x4_t __c, const int __lane)
22134 {
22135   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22136 }
22137
22138 __extension__ extern __inline uint32x2_t
22139 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22140 vmls_lane_u32 (uint32x2_t __a, uint32x2_t __b,
22141                uint32x2_t __c, const int __lane)
22142 {
22143   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22144 }
22145
22146 /* vmls_laneq  */
22147
22148 __extension__ extern __inline float32x2_t
22149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22150 vmls_laneq_f32 (float32x2_t __a, float32x2_t __b,
22151                float32x4_t __c, const int __lane)
22152 {
22153   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22154 }
22155
22156 __extension__ extern __inline int16x4_t
22157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22158 vmls_laneq_s16 (int16x4_t __a, int16x4_t __b,
22159                 int16x8_t __c, const int __lane)
22160 {
22161   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22162 }
22163
22164 __extension__ extern __inline int32x2_t
22165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22166 vmls_laneq_s32 (int32x2_t __a, int32x2_t __b,
22167                 int32x4_t __c, const int __lane)
22168 {
22169   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22170 }
22171
22172 __extension__ extern __inline uint16x4_t
22173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22174 vmls_laneq_u16 (uint16x4_t __a, uint16x4_t __b,
22175                 uint16x8_t __c, const int __lane)
22176 {
22177   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22178 }
22179
22180 __extension__ extern __inline uint32x2_t
22181 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22182 vmls_laneq_u32 (uint32x2_t __a, uint32x2_t __b,
22183                 uint32x4_t __c, const int __lane)
22184 {
22185   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22186 }
22187
22188 /* vmlsq_lane  */
22189
22190 __extension__ extern __inline float32x4_t
22191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22192 vmlsq_lane_f32 (float32x4_t __a, float32x4_t __b,
22193                 float32x2_t __c, const int __lane)
22194 {
22195   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22196 }
22197
22198 __extension__ extern __inline int16x8_t
22199 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22200 vmlsq_lane_s16 (int16x8_t __a, int16x8_t __b,
22201                 int16x4_t __c, const int __lane)
22202 {
22203   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22204 }
22205
22206 __extension__ extern __inline int32x4_t
22207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22208 vmlsq_lane_s32 (int32x4_t __a, int32x4_t __b,
22209                 int32x2_t __c, const int __lane)
22210 {
22211   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22212 }
22213
22214 __extension__ extern __inline uint16x8_t
22215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22216 vmlsq_lane_u16 (uint16x8_t __a, uint16x8_t __b,
22217                 uint16x4_t __c, const int __lane)
22218 {
22219   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22220 }
22221
22222 __extension__ extern __inline uint32x4_t
22223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22224 vmlsq_lane_u32 (uint32x4_t __a, uint32x4_t __b,
22225                 uint32x2_t __c, const int __lane)
22226 {
22227   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22228 }
22229
22230   /* vmlsq_laneq  */
22231
22232 __extension__ extern __inline float32x4_t
22233 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22234 vmlsq_laneq_f32 (float32x4_t __a, float32x4_t __b,
22235                 float32x4_t __c, const int __lane)
22236 {
22237   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22238 }
22239
22240 __extension__ extern __inline int16x8_t
22241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22242 vmlsq_laneq_s16 (int16x8_t __a, int16x8_t __b,
22243                 int16x8_t __c, const int __lane)
22244 {
22245   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22246 }
22247
22248 __extension__ extern __inline int32x4_t
22249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22250 vmlsq_laneq_s32 (int32x4_t __a, int32x4_t __b,
22251                 int32x4_t __c, const int __lane)
22252 {
22253   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22254 }
22255 __extension__ extern __inline uint16x8_t
22256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22257 vmlsq_laneq_u16 (uint16x8_t __a, uint16x8_t __b,
22258                 uint16x8_t __c, const int __lane)
22259 {
22260   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22261 }
22262
22263 __extension__ extern __inline uint32x4_t
22264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22265 vmlsq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
22266                 uint32x4_t __c, const int __lane)
22267 {
22268   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
22269 }
22270
22271 /* vmov_n_  */
22272
22273 __extension__ extern __inline float16x4_t
22274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22275 vmov_n_f16 (float16_t __a)
22276 {
22277   return vdup_n_f16 (__a);
22278 }
22279
22280 __extension__ extern __inline float32x2_t
22281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22282 vmov_n_f32 (float32_t __a)
22283 {
22284   return vdup_n_f32 (__a);
22285 }
22286
22287 __extension__ extern __inline float64x1_t
22288 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22289 vmov_n_f64 (float64_t __a)
22290 {
22291   return (float64x1_t) {__a};
22292 }
22293
22294 __extension__ extern __inline poly8x8_t
22295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22296 vmov_n_p8 (poly8_t __a)
22297 {
22298   return vdup_n_p8 (__a);
22299 }
22300
22301 __extension__ extern __inline poly16x4_t
22302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22303 vmov_n_p16 (poly16_t __a)
22304 {
22305   return vdup_n_p16 (__a);
22306 }
22307
22308 __extension__ extern __inline poly64x1_t
22309 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22310 vmov_n_p64 (poly64_t __a)
22311 {
22312   return vdup_n_p64 (__a);
22313 }
22314
22315 __extension__ extern __inline int8x8_t
22316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22317 vmov_n_s8 (int8_t __a)
22318 {
22319   return vdup_n_s8 (__a);
22320 }
22321
22322 __extension__ extern __inline int16x4_t
22323 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22324 vmov_n_s16 (int16_t __a)
22325 {
22326   return vdup_n_s16 (__a);
22327 }
22328
22329 __extension__ extern __inline int32x2_t
22330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22331 vmov_n_s32 (int32_t __a)
22332 {
22333   return vdup_n_s32 (__a);
22334 }
22335
22336 __extension__ extern __inline int64x1_t
22337 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22338 vmov_n_s64 (int64_t __a)
22339 {
22340   return (int64x1_t) {__a};
22341 }
22342
22343 __extension__ extern __inline uint8x8_t
22344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22345 vmov_n_u8 (uint8_t __a)
22346 {
22347   return vdup_n_u8 (__a);
22348 }
22349
22350 __extension__ extern __inline uint16x4_t
22351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22352 vmov_n_u16 (uint16_t __a)
22353 {
22354     return vdup_n_u16 (__a);
22355 }
22356
22357 __extension__ extern __inline uint32x2_t
22358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22359 vmov_n_u32 (uint32_t __a)
22360 {
22361    return vdup_n_u32 (__a);
22362 }
22363
22364 __extension__ extern __inline uint64x1_t
22365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22366 vmov_n_u64 (uint64_t __a)
22367 {
22368   return (uint64x1_t) {__a};
22369 }
22370
22371 __extension__ extern __inline float16x8_t
22372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22373 vmovq_n_f16 (float16_t __a)
22374 {
22375   return vdupq_n_f16 (__a);
22376 }
22377
22378 __extension__ extern __inline float32x4_t
22379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22380 vmovq_n_f32 (float32_t __a)
22381 {
22382   return vdupq_n_f32 (__a);
22383 }
22384
22385 __extension__ extern __inline float64x2_t
22386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22387 vmovq_n_f64 (float64_t __a)
22388 {
22389   return vdupq_n_f64 (__a);
22390 }
22391
22392 __extension__ extern __inline poly8x16_t
22393 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22394 vmovq_n_p8 (poly8_t __a)
22395 {
22396   return vdupq_n_p8 (__a);
22397 }
22398
22399 __extension__ extern __inline poly16x8_t
22400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22401 vmovq_n_p16 (poly16_t __a)
22402 {
22403   return vdupq_n_p16 (__a);
22404 }
22405
22406 __extension__ extern __inline poly64x2_t
22407 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22408 vmovq_n_p64 (poly64_t __a)
22409 {
22410   return vdupq_n_p64 (__a);
22411 }
22412
22413 __extension__ extern __inline int8x16_t
22414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22415 vmovq_n_s8 (int8_t __a)
22416 {
22417   return vdupq_n_s8 (__a);
22418 }
22419
22420 __extension__ extern __inline int16x8_t
22421 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22422 vmovq_n_s16 (int16_t __a)
22423 {
22424   return vdupq_n_s16 (__a);
22425 }
22426
22427 __extension__ extern __inline int32x4_t
22428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22429 vmovq_n_s32 (int32_t __a)
22430 {
22431   return vdupq_n_s32 (__a);
22432 }
22433
22434 __extension__ extern __inline int64x2_t
22435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22436 vmovq_n_s64 (int64_t __a)
22437 {
22438   return vdupq_n_s64 (__a);
22439 }
22440
22441 __extension__ extern __inline uint8x16_t
22442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22443 vmovq_n_u8 (uint8_t __a)
22444 {
22445   return vdupq_n_u8 (__a);
22446 }
22447
22448 __extension__ extern __inline uint16x8_t
22449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22450 vmovq_n_u16 (uint16_t __a)
22451 {
22452   return vdupq_n_u16 (__a);
22453 }
22454
22455 __extension__ extern __inline uint32x4_t
22456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22457 vmovq_n_u32 (uint32_t __a)
22458 {
22459   return vdupq_n_u32 (__a);
22460 }
22461
22462 __extension__ extern __inline uint64x2_t
22463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22464 vmovq_n_u64 (uint64_t __a)
22465 {
22466   return vdupq_n_u64 (__a);
22467 }
22468
22469 /* vmul_lane  */
22470
22471 __extension__ extern __inline float32x2_t
22472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22473 vmul_lane_f32 (float32x2_t __a, float32x2_t __b, const int __lane)
22474 {
22475   return __a * __aarch64_vget_lane_any (__b, __lane);
22476 }
22477
22478 __extension__ extern __inline float64x1_t
22479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22480 vmul_lane_f64 (float64x1_t __a, float64x1_t __b, const int __lane)
22481 {
22482   return __a * __b;
22483 }
22484
22485 __extension__ extern __inline int16x4_t
22486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22487 vmul_lane_s16 (int16x4_t __a, int16x4_t __b, const int __lane)
22488 {
22489   return __a * __aarch64_vget_lane_any (__b, __lane);
22490 }
22491
22492 __extension__ extern __inline int32x2_t
22493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22494 vmul_lane_s32 (int32x2_t __a, int32x2_t __b, const int __lane)
22495 {
22496   return __a * __aarch64_vget_lane_any (__b, __lane);
22497 }
22498
22499 __extension__ extern __inline uint16x4_t
22500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22501 vmul_lane_u16 (uint16x4_t __a, uint16x4_t __b, const int __lane)
22502 {
22503   return __a * __aarch64_vget_lane_any (__b, __lane);
22504 }
22505
22506 __extension__ extern __inline uint32x2_t
22507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22508 vmul_lane_u32 (uint32x2_t __a, uint32x2_t __b, const int __lane)
22509 {
22510   return __a * __aarch64_vget_lane_any (__b, __lane);
22511 }
22512
22513 /* vmuld_lane  */
22514
22515 __extension__ extern __inline float64_t
22516 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22517 vmuld_lane_f64 (float64_t __a, float64x1_t __b, const int __lane)
22518 {
22519   return __a * __aarch64_vget_lane_any (__b, __lane);
22520 }
22521
22522 __extension__ extern __inline float64_t
22523 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22524 vmuld_laneq_f64 (float64_t __a, float64x2_t __b, const int __lane)
22525 {
22526   return __a * __aarch64_vget_lane_any (__b, __lane);
22527 }
22528
22529 /* vmuls_lane  */
22530
22531 __extension__ extern __inline float32_t
22532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22533 vmuls_lane_f32 (float32_t __a, float32x2_t __b, const int __lane)
22534 {
22535   return __a * __aarch64_vget_lane_any (__b, __lane);
22536 }
22537
22538 __extension__ extern __inline float32_t
22539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22540 vmuls_laneq_f32 (float32_t __a, float32x4_t __b, const int __lane)
22541 {
22542   return __a * __aarch64_vget_lane_any (__b, __lane);
22543 }
22544
22545 /* vmul_laneq  */
22546
22547 __extension__ extern __inline float32x2_t
22548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22549 vmul_laneq_f32 (float32x2_t __a, float32x4_t __b, const int __lane)
22550 {
22551   return __a * __aarch64_vget_lane_any (__b, __lane);
22552 }
22553
22554 __extension__ extern __inline float64x1_t
22555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22556 vmul_laneq_f64 (float64x1_t __a, float64x2_t __b, const int __lane)
22557 {
22558   return __a * __aarch64_vget_lane_any (__b, __lane);
22559 }
22560
22561 __extension__ extern __inline int16x4_t
22562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22563 vmul_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __lane)
22564 {
22565   return __a * __aarch64_vget_lane_any (__b, __lane);
22566 }
22567
22568 __extension__ extern __inline int32x2_t
22569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22570 vmul_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __lane)
22571 {
22572   return __a * __aarch64_vget_lane_any (__b, __lane);
22573 }
22574
22575 __extension__ extern __inline uint16x4_t
22576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22577 vmul_laneq_u16 (uint16x4_t __a, uint16x8_t __b, const int __lane)
22578 {
22579   return __a * __aarch64_vget_lane_any (__b, __lane);
22580 }
22581
22582 __extension__ extern __inline uint32x2_t
22583 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22584 vmul_laneq_u32 (uint32x2_t __a, uint32x4_t __b, const int __lane)
22585 {
22586   return __a * __aarch64_vget_lane_any (__b, __lane);
22587 }
22588
22589 /* vmul_n  */
22590
22591 __extension__ extern __inline float64x1_t
22592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22593 vmul_n_f64  (float64x1_t __a, float64_t __b)
22594 {
22595   return (float64x1_t) { vget_lane_f64 (__a, 0) * __b };
22596 }
22597
22598 /* vmulq_lane  */
22599
22600 __extension__ extern __inline float32x4_t
22601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22602 vmulq_lane_f32 (float32x4_t __a, float32x2_t __b, const int __lane)
22603 {
22604   return __a * __aarch64_vget_lane_any (__b, __lane);
22605 }
22606
22607 __extension__ extern __inline float64x2_t
22608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22609 vmulq_lane_f64 (float64x2_t __a, float64x1_t __b, const int __lane)
22610 {
22611   __AARCH64_LANE_CHECK (__a, __lane);
22612   return __a * __b[0];
22613 }
22614
22615 __extension__ extern __inline int16x8_t
22616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22617 vmulq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __lane)
22618 {
22619   return __a * __aarch64_vget_lane_any (__b, __lane);
22620 }
22621
22622 __extension__ extern __inline int32x4_t
22623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22624 vmulq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __lane)
22625 {
22626   return __a * __aarch64_vget_lane_any (__b, __lane);
22627 }
22628
22629 __extension__ extern __inline uint16x8_t
22630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22631 vmulq_lane_u16 (uint16x8_t __a, uint16x4_t __b, const int __lane)
22632 {
22633   return __a * __aarch64_vget_lane_any (__b, __lane);
22634 }
22635
22636 __extension__ extern __inline uint32x4_t
22637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22638 vmulq_lane_u32 (uint32x4_t __a, uint32x2_t __b, const int __lane)
22639 {
22640   return __a * __aarch64_vget_lane_any (__b, __lane);
22641 }
22642
22643 /* vmulq_laneq  */
22644
22645 __extension__ extern __inline float32x4_t
22646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22647 vmulq_laneq_f32 (float32x4_t __a, float32x4_t __b, const int __lane)
22648 {
22649   return __a * __aarch64_vget_lane_any (__b, __lane);
22650 }
22651
22652 __extension__ extern __inline float64x2_t
22653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22654 vmulq_laneq_f64 (float64x2_t __a, float64x2_t __b, const int __lane)
22655 {
22656   return __a * __aarch64_vget_lane_any (__b, __lane);
22657 }
22658
22659 __extension__ extern __inline int16x8_t
22660 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22661 vmulq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __lane)
22662 {
22663   return __a * __aarch64_vget_lane_any (__b, __lane);
22664 }
22665
22666 __extension__ extern __inline int32x4_t
22667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22668 vmulq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __lane)
22669 {
22670   return __a * __aarch64_vget_lane_any (__b, __lane);
22671 }
22672
22673 __extension__ extern __inline uint16x8_t
22674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22675 vmulq_laneq_u16 (uint16x8_t __a, uint16x8_t __b, const int __lane)
22676 {
22677   return __a * __aarch64_vget_lane_any (__b, __lane);
22678 }
22679
22680 __extension__ extern __inline uint32x4_t
22681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22682 vmulq_laneq_u32 (uint32x4_t __a, uint32x4_t __b, const int __lane)
22683 {
22684   return __a * __aarch64_vget_lane_any (__b, __lane);
22685 }
22686
22687 /* vmul_n.  */
22688
22689 __extension__ extern __inline float32x2_t
22690 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22691 vmul_n_f32 (float32x2_t __a, float32_t __b)
22692 {
22693   return __a * __b;
22694 }
22695
22696 __extension__ extern __inline float32x4_t
22697 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22698 vmulq_n_f32 (float32x4_t __a, float32_t __b)
22699 {
22700   return __a * __b;
22701 }
22702
22703 __extension__ extern __inline float64x2_t
22704 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22705 vmulq_n_f64 (float64x2_t __a, float64_t __b)
22706 {
22707   return __a * __b;
22708 }
22709
22710 __extension__ extern __inline int16x4_t
22711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22712 vmul_n_s16 (int16x4_t __a, int16_t __b)
22713 {
22714   return __a * __b;
22715 }
22716
22717 __extension__ extern __inline int16x8_t
22718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22719 vmulq_n_s16 (int16x8_t __a, int16_t __b)
22720 {
22721   return __a * __b;
22722 }
22723
22724 __extension__ extern __inline int32x2_t
22725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22726 vmul_n_s32 (int32x2_t __a, int32_t __b)
22727 {
22728   return __a * __b;
22729 }
22730
22731 __extension__ extern __inline int32x4_t
22732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22733 vmulq_n_s32 (int32x4_t __a, int32_t __b)
22734 {
22735   return __a * __b;
22736 }
22737
22738 __extension__ extern __inline uint16x4_t
22739 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22740 vmul_n_u16 (uint16x4_t __a, uint16_t __b)
22741 {
22742   return __a * __b;
22743 }
22744
22745 __extension__ extern __inline uint16x8_t
22746 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22747 vmulq_n_u16 (uint16x8_t __a, uint16_t __b)
22748 {
22749   return __a * __b;
22750 }
22751
22752 __extension__ extern __inline uint32x2_t
22753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22754 vmul_n_u32 (uint32x2_t __a, uint32_t __b)
22755 {
22756   return __a * __b;
22757 }
22758
22759 __extension__ extern __inline uint32x4_t
22760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22761 vmulq_n_u32 (uint32x4_t __a, uint32_t __b)
22762 {
22763   return __a * __b;
22764 }
22765
22766 /* vmvn  */
22767
22768 __extension__ extern __inline poly8x8_t
22769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22770 vmvn_p8 (poly8x8_t __a)
22771 {
22772   return (poly8x8_t) ~((int8x8_t) __a);
22773 }
22774
22775 __extension__ extern __inline int8x8_t
22776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22777 vmvn_s8 (int8x8_t __a)
22778 {
22779   return ~__a;
22780 }
22781
22782 __extension__ extern __inline int16x4_t
22783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22784 vmvn_s16 (int16x4_t __a)
22785 {
22786   return ~__a;
22787 }
22788
22789 __extension__ extern __inline int32x2_t
22790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22791 vmvn_s32 (int32x2_t __a)
22792 {
22793   return ~__a;
22794 }
22795
22796 __extension__ extern __inline uint8x8_t
22797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22798 vmvn_u8 (uint8x8_t __a)
22799 {
22800   return ~__a;
22801 }
22802
22803 __extension__ extern __inline uint16x4_t
22804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22805 vmvn_u16 (uint16x4_t __a)
22806 {
22807   return ~__a;
22808 }
22809
22810 __extension__ extern __inline uint32x2_t
22811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22812 vmvn_u32 (uint32x2_t __a)
22813 {
22814   return ~__a;
22815 }
22816
22817 __extension__ extern __inline poly8x16_t
22818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22819 vmvnq_p8 (poly8x16_t __a)
22820 {
22821   return (poly8x16_t) ~((int8x16_t) __a);
22822 }
22823
22824 __extension__ extern __inline int8x16_t
22825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22826 vmvnq_s8 (int8x16_t __a)
22827 {
22828   return ~__a;
22829 }
22830
22831 __extension__ extern __inline int16x8_t
22832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22833 vmvnq_s16 (int16x8_t __a)
22834 {
22835   return ~__a;
22836 }
22837
22838 __extension__ extern __inline int32x4_t
22839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22840 vmvnq_s32 (int32x4_t __a)
22841 {
22842   return ~__a;
22843 }
22844
22845 __extension__ extern __inline uint8x16_t
22846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22847 vmvnq_u8 (uint8x16_t __a)
22848 {
22849   return ~__a;
22850 }
22851
22852 __extension__ extern __inline uint16x8_t
22853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22854 vmvnq_u16 (uint16x8_t __a)
22855 {
22856   return ~__a;
22857 }
22858
22859 __extension__ extern __inline uint32x4_t
22860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22861 vmvnq_u32 (uint32x4_t __a)
22862 {
22863   return ~__a;
22864 }
22865
22866 /* vneg  */
22867
22868 __extension__ extern __inline float32x2_t
22869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22870 vneg_f32 (float32x2_t __a)
22871 {
22872   return -__a;
22873 }
22874
22875 __extension__ extern __inline float64x1_t
22876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22877 vneg_f64 (float64x1_t __a)
22878 {
22879   return -__a;
22880 }
22881
22882 __extension__ extern __inline int8x8_t
22883 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22884 vneg_s8 (int8x8_t __a)
22885 {
22886   return -__a;
22887 }
22888
22889 __extension__ extern __inline int16x4_t
22890 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22891 vneg_s16 (int16x4_t __a)
22892 {
22893   return -__a;
22894 }
22895
22896 __extension__ extern __inline int32x2_t
22897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22898 vneg_s32 (int32x2_t __a)
22899 {
22900   return -__a;
22901 }
22902
22903 __extension__ extern __inline int64x1_t
22904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22905 vneg_s64 (int64x1_t __a)
22906 {
22907   return -__a;
22908 }
22909
22910 __extension__ extern __inline float32x4_t
22911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22912 vnegq_f32 (float32x4_t __a)
22913 {
22914   return -__a;
22915 }
22916
22917 __extension__ extern __inline float64x2_t
22918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22919 vnegq_f64 (float64x2_t __a)
22920 {
22921   return -__a;
22922 }
22923
22924 __extension__ extern __inline int8x16_t
22925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22926 vnegq_s8 (int8x16_t __a)
22927 {
22928   return -__a;
22929 }
22930
22931 __extension__ extern __inline int16x8_t
22932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22933 vnegq_s16 (int16x8_t __a)
22934 {
22935   return -__a;
22936 }
22937
22938 __extension__ extern __inline int32x4_t
22939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22940 vnegq_s32 (int32x4_t __a)
22941 {
22942   return -__a;
22943 }
22944
22945 __extension__ extern __inline int64x2_t
22946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22947 vnegq_s64 (int64x2_t __a)
22948 {
22949   return -__a;
22950 }
22951
22952 /* vpadd  */
22953
22954 __extension__ extern __inline float32x2_t
22955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22956 vpadd_f32 (float32x2_t __a, float32x2_t __b)
22957 {
22958   return __builtin_aarch64_faddpv2sf (__a, __b);
22959 }
22960
22961 __extension__ extern __inline float32x4_t
22962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22963 vpaddq_f32 (float32x4_t __a, float32x4_t __b)
22964 {
22965   return __builtin_aarch64_faddpv4sf (__a, __b);
22966 }
22967
22968 __extension__ extern __inline float64x2_t
22969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22970 vpaddq_f64 (float64x2_t __a, float64x2_t __b)
22971 {
22972   return __builtin_aarch64_faddpv2df (__a, __b);
22973 }
22974
22975 __extension__ extern __inline int8x8_t
22976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22977 vpadd_s8 (int8x8_t __a, int8x8_t __b)
22978 {
22979   return __builtin_aarch64_addpv8qi (__a, __b);
22980 }
22981
22982 __extension__ extern __inline int16x4_t
22983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22984 vpadd_s16 (int16x4_t __a, int16x4_t __b)
22985 {
22986   return __builtin_aarch64_addpv4hi (__a, __b);
22987 }
22988
22989 __extension__ extern __inline int32x2_t
22990 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22991 vpadd_s32 (int32x2_t __a, int32x2_t __b)
22992 {
22993   return __builtin_aarch64_addpv2si (__a, __b);
22994 }
22995
22996 __extension__ extern __inline uint8x8_t
22997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22998 vpadd_u8 (uint8x8_t __a, uint8x8_t __b)
22999 {
23000   return (uint8x8_t) __builtin_aarch64_addpv8qi ((int8x8_t) __a,
23001                                                  (int8x8_t) __b);
23002 }
23003
23004 __extension__ extern __inline uint16x4_t
23005 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23006 vpadd_u16 (uint16x4_t __a, uint16x4_t __b)
23007 {
23008   return (uint16x4_t) __builtin_aarch64_addpv4hi ((int16x4_t) __a,
23009                                                   (int16x4_t) __b);
23010 }
23011
23012 __extension__ extern __inline uint32x2_t
23013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23014 vpadd_u32 (uint32x2_t __a, uint32x2_t __b)
23015 {
23016   return (uint32x2_t) __builtin_aarch64_addpv2si ((int32x2_t) __a,
23017                                                   (int32x2_t) __b);
23018 }
23019
23020 __extension__ extern __inline float32_t
23021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23022 vpadds_f32 (float32x2_t __a)
23023 {
23024   return __builtin_aarch64_reduc_plus_scal_v2sf (__a);
23025 }
23026
23027 __extension__ extern __inline float64_t
23028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23029 vpaddd_f64 (float64x2_t __a)
23030 {
23031   return __builtin_aarch64_reduc_plus_scal_v2df (__a);
23032 }
23033
23034 __extension__ extern __inline int64_t
23035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23036 vpaddd_s64 (int64x2_t __a)
23037 {
23038   return __builtin_aarch64_addpdi (__a);
23039 }
23040
23041 __extension__ extern __inline uint64_t
23042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23043 vpaddd_u64 (uint64x2_t __a)
23044 {
23045   return __builtin_aarch64_addpdi ((int64x2_t) __a);
23046 }
23047
23048 /* vqabs */
23049
23050 __extension__ extern __inline int64x2_t
23051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23052 vqabsq_s64 (int64x2_t __a)
23053 {
23054   return (int64x2_t) __builtin_aarch64_sqabsv2di (__a);
23055 }
23056
23057 __extension__ extern __inline int8_t
23058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23059 vqabsb_s8 (int8_t __a)
23060 {
23061   return (int8_t) __builtin_aarch64_sqabsqi (__a);
23062 }
23063
23064 __extension__ extern __inline int16_t
23065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23066 vqabsh_s16 (int16_t __a)
23067 {
23068   return (int16_t) __builtin_aarch64_sqabshi (__a);
23069 }
23070
23071 __extension__ extern __inline int32_t
23072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23073 vqabss_s32 (int32_t __a)
23074 {
23075   return (int32_t) __builtin_aarch64_sqabssi (__a);
23076 }
23077
23078 __extension__ extern __inline int64_t
23079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23080 vqabsd_s64 (int64_t __a)
23081 {
23082   return __builtin_aarch64_sqabsdi (__a);
23083 }
23084
23085 /* vqadd */
23086
23087 __extension__ extern __inline int8_t
23088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23089 vqaddb_s8 (int8_t __a, int8_t __b)
23090 {
23091   return (int8_t) __builtin_aarch64_sqaddqi (__a, __b);
23092 }
23093
23094 __extension__ extern __inline int16_t
23095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23096 vqaddh_s16 (int16_t __a, int16_t __b)
23097 {
23098   return (int16_t) __builtin_aarch64_sqaddhi (__a, __b);
23099 }
23100
23101 __extension__ extern __inline int32_t
23102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23103 vqadds_s32 (int32_t __a, int32_t __b)
23104 {
23105   return (int32_t) __builtin_aarch64_sqaddsi (__a, __b);
23106 }
23107
23108 __extension__ extern __inline int64_t
23109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23110 vqaddd_s64 (int64_t __a, int64_t __b)
23111 {
23112   return __builtin_aarch64_sqadddi (__a, __b);
23113 }
23114
23115 __extension__ extern __inline uint8_t
23116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23117 vqaddb_u8 (uint8_t __a, uint8_t __b)
23118 {
23119   return (uint8_t) __builtin_aarch64_uqaddqi_uuu (__a, __b);
23120 }
23121
23122 __extension__ extern __inline uint16_t
23123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23124 vqaddh_u16 (uint16_t __a, uint16_t __b)
23125 {
23126   return (uint16_t) __builtin_aarch64_uqaddhi_uuu (__a, __b);
23127 }
23128
23129 __extension__ extern __inline uint32_t
23130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23131 vqadds_u32 (uint32_t __a, uint32_t __b)
23132 {
23133   return (uint32_t) __builtin_aarch64_uqaddsi_uuu (__a, __b);
23134 }
23135
23136 __extension__ extern __inline uint64_t
23137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23138 vqaddd_u64 (uint64_t __a, uint64_t __b)
23139 {
23140   return __builtin_aarch64_uqadddi_uuu (__a, __b);
23141 }
23142
23143 /* vqdmlal */
23144
23145 __extension__ extern __inline int32x4_t
23146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23147 vqdmlal_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c)
23148 {
23149   return __builtin_aarch64_sqdmlalv4hi (__a, __b, __c);
23150 }
23151
23152 __extension__ extern __inline int32x4_t
23153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23154 vqdmlal_high_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c)
23155 {
23156   return __builtin_aarch64_sqdmlal2v8hi (__a, __b, __c);
23157 }
23158
23159 __extension__ extern __inline int32x4_t
23160 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23161 vqdmlal_high_lane_s16 (int32x4_t __a, int16x8_t __b, int16x4_t __c,
23162                        int const __d)
23163 {
23164   return __builtin_aarch64_sqdmlal2_lanev8hi (__a, __b, __c, __d);
23165 }
23166
23167 __extension__ extern __inline int32x4_t
23168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23169 vqdmlal_high_laneq_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c,
23170                         int const __d)
23171 {
23172   return __builtin_aarch64_sqdmlal2_laneqv8hi (__a, __b, __c, __d);
23173 }
23174
23175 __extension__ extern __inline int32x4_t
23176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23177 vqdmlal_high_n_s16 (int32x4_t __a, int16x8_t __b, int16_t __c)
23178 {
23179   return __builtin_aarch64_sqdmlal2_nv8hi (__a, __b, __c);
23180 }
23181
23182 __extension__ extern __inline int32x4_t
23183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23184 vqdmlal_lane_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c, int const __d)
23185 {
23186   return __builtin_aarch64_sqdmlal_lanev4hi (__a, __b, __c, __d);
23187 }
23188
23189 __extension__ extern __inline int32x4_t
23190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23191 vqdmlal_laneq_s16 (int32x4_t __a, int16x4_t __b, int16x8_t __c, int const __d)
23192 {
23193   return __builtin_aarch64_sqdmlal_laneqv4hi (__a, __b, __c, __d);
23194 }
23195
23196 __extension__ extern __inline int32x4_t
23197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23198 vqdmlal_n_s16 (int32x4_t __a, int16x4_t __b, int16_t __c)
23199 {
23200   return __builtin_aarch64_sqdmlal_nv4hi (__a, __b, __c);
23201 }
23202
23203 __extension__ extern __inline int64x2_t
23204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23205 vqdmlal_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c)
23206 {
23207   return __builtin_aarch64_sqdmlalv2si (__a, __b, __c);
23208 }
23209
23210 __extension__ extern __inline int64x2_t
23211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23212 vqdmlal_high_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c)
23213 {
23214   return __builtin_aarch64_sqdmlal2v4si (__a, __b, __c);
23215 }
23216
23217 __extension__ extern __inline int64x2_t
23218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23219 vqdmlal_high_lane_s32 (int64x2_t __a, int32x4_t __b, int32x2_t __c,
23220                        int const __d)
23221 {
23222   return __builtin_aarch64_sqdmlal2_lanev4si (__a, __b, __c, __d);
23223 }
23224
23225 __extension__ extern __inline int64x2_t
23226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23227 vqdmlal_high_laneq_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c,
23228                         int const __d)
23229 {
23230   return __builtin_aarch64_sqdmlal2_laneqv4si (__a, __b, __c, __d);
23231 }
23232
23233 __extension__ extern __inline int64x2_t
23234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23235 vqdmlal_high_n_s32 (int64x2_t __a, int32x4_t __b, int32_t __c)
23236 {
23237   return __builtin_aarch64_sqdmlal2_nv4si (__a, __b, __c);
23238 }
23239
23240 __extension__ extern __inline int64x2_t
23241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23242 vqdmlal_lane_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c, int const __d)
23243 {
23244   return __builtin_aarch64_sqdmlal_lanev2si (__a, __b, __c, __d);
23245 }
23246
23247 __extension__ extern __inline int64x2_t
23248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23249 vqdmlal_laneq_s32 (int64x2_t __a, int32x2_t __b, int32x4_t __c, int const __d)
23250 {
23251   return __builtin_aarch64_sqdmlal_laneqv2si (__a, __b, __c, __d);
23252 }
23253
23254 __extension__ extern __inline int64x2_t
23255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23256 vqdmlal_n_s32 (int64x2_t __a, int32x2_t __b, int32_t __c)
23257 {
23258   return __builtin_aarch64_sqdmlal_nv2si (__a, __b, __c);
23259 }
23260
23261 __extension__ extern __inline int32_t
23262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23263 vqdmlalh_s16 (int32_t __a, int16_t __b, int16_t __c)
23264 {
23265   return __builtin_aarch64_sqdmlalhi (__a, __b, __c);
23266 }
23267
23268 __extension__ extern __inline int32_t
23269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23270 vqdmlalh_lane_s16 (int32_t __a, int16_t __b, int16x4_t __c, const int __d)
23271 {
23272   return __builtin_aarch64_sqdmlal_lanehi (__a, __b, __c, __d);
23273 }
23274
23275 __extension__ extern __inline int32_t
23276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23277 vqdmlalh_laneq_s16 (int32_t __a, int16_t __b, int16x8_t __c, const int __d)
23278 {
23279   return __builtin_aarch64_sqdmlal_laneqhi (__a, __b, __c, __d);
23280 }
23281
23282 __extension__ extern __inline int64_t
23283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23284 vqdmlals_s32 (int64_t __a, int32_t __b, int32_t __c)
23285 {
23286   return __builtin_aarch64_sqdmlalsi (__a, __b, __c);
23287 }
23288
23289 __extension__ extern __inline int64_t
23290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23291 vqdmlals_lane_s32 (int64_t __a, int32_t __b, int32x2_t __c, const int __d)
23292 {
23293   return __builtin_aarch64_sqdmlal_lanesi (__a, __b, __c, __d);
23294 }
23295
23296 __extension__ extern __inline int64_t
23297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23298 vqdmlals_laneq_s32 (int64_t __a, int32_t __b, int32x4_t __c, const int __d)
23299 {
23300   return __builtin_aarch64_sqdmlal_laneqsi (__a, __b, __c, __d);
23301 }
23302
23303 /* vqdmlsl */
23304
23305 __extension__ extern __inline int32x4_t
23306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23307 vqdmlsl_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c)
23308 {
23309   return __builtin_aarch64_sqdmlslv4hi (__a, __b, __c);
23310 }
23311
23312 __extension__ extern __inline int32x4_t
23313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23314 vqdmlsl_high_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c)
23315 {
23316   return __builtin_aarch64_sqdmlsl2v8hi (__a, __b, __c);
23317 }
23318
23319 __extension__ extern __inline int32x4_t
23320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23321 vqdmlsl_high_lane_s16 (int32x4_t __a, int16x8_t __b, int16x4_t __c,
23322                        int const __d)
23323 {
23324   return __builtin_aarch64_sqdmlsl2_lanev8hi (__a, __b, __c, __d);
23325 }
23326
23327 __extension__ extern __inline int32x4_t
23328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23329 vqdmlsl_high_laneq_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c,
23330                         int const __d)
23331 {
23332   return __builtin_aarch64_sqdmlsl2_laneqv8hi (__a, __b, __c, __d);
23333 }
23334
23335 __extension__ extern __inline int32x4_t
23336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23337 vqdmlsl_high_n_s16 (int32x4_t __a, int16x8_t __b, int16_t __c)
23338 {
23339   return __builtin_aarch64_sqdmlsl2_nv8hi (__a, __b, __c);
23340 }
23341
23342 __extension__ extern __inline int32x4_t
23343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23344 vqdmlsl_lane_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c, int const __d)
23345 {
23346   return __builtin_aarch64_sqdmlsl_lanev4hi (__a, __b, __c, __d);
23347 }
23348
23349 __extension__ extern __inline int32x4_t
23350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23351 vqdmlsl_laneq_s16 (int32x4_t __a, int16x4_t __b, int16x8_t __c, int const __d)
23352 {
23353   return __builtin_aarch64_sqdmlsl_laneqv4hi (__a, __b, __c, __d);
23354 }
23355
23356 __extension__ extern __inline int32x4_t
23357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23358 vqdmlsl_n_s16 (int32x4_t __a, int16x4_t __b, int16_t __c)
23359 {
23360   return __builtin_aarch64_sqdmlsl_nv4hi (__a, __b, __c);
23361 }
23362
23363 __extension__ extern __inline int64x2_t
23364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23365 vqdmlsl_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c)
23366 {
23367   return __builtin_aarch64_sqdmlslv2si (__a, __b, __c);
23368 }
23369
23370 __extension__ extern __inline int64x2_t
23371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23372 vqdmlsl_high_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c)
23373 {
23374   return __builtin_aarch64_sqdmlsl2v4si (__a, __b, __c);
23375 }
23376
23377 __extension__ extern __inline int64x2_t
23378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23379 vqdmlsl_high_lane_s32 (int64x2_t __a, int32x4_t __b, int32x2_t __c,
23380                        int const __d)
23381 {
23382   return __builtin_aarch64_sqdmlsl2_lanev4si (__a, __b, __c, __d);
23383 }
23384
23385 __extension__ extern __inline int64x2_t
23386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23387 vqdmlsl_high_laneq_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c,
23388                         int const __d)
23389 {
23390   return __builtin_aarch64_sqdmlsl2_laneqv4si (__a, __b, __c, __d);
23391 }
23392
23393 __extension__ extern __inline int64x2_t
23394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23395 vqdmlsl_high_n_s32 (int64x2_t __a, int32x4_t __b, int32_t __c)
23396 {
23397   return __builtin_aarch64_sqdmlsl2_nv4si (__a, __b, __c);
23398 }
23399
23400 __extension__ extern __inline int64x2_t
23401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23402 vqdmlsl_lane_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c, int const __d)
23403 {
23404   return __builtin_aarch64_sqdmlsl_lanev2si (__a, __b, __c, __d);
23405 }
23406
23407 __extension__ extern __inline int64x2_t
23408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23409 vqdmlsl_laneq_s32 (int64x2_t __a, int32x2_t __b, int32x4_t __c, int const __d)
23410 {
23411   return __builtin_aarch64_sqdmlsl_laneqv2si (__a, __b, __c, __d);
23412 }
23413
23414 __extension__ extern __inline int64x2_t
23415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23416 vqdmlsl_n_s32 (int64x2_t __a, int32x2_t __b, int32_t __c)
23417 {
23418   return __builtin_aarch64_sqdmlsl_nv2si (__a, __b, __c);
23419 }
23420
23421 __extension__ extern __inline int32_t
23422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23423 vqdmlslh_s16 (int32_t __a, int16_t __b, int16_t __c)
23424 {
23425   return __builtin_aarch64_sqdmlslhi (__a, __b, __c);
23426 }
23427
23428 __extension__ extern __inline int32_t
23429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23430 vqdmlslh_lane_s16 (int32_t __a, int16_t __b, int16x4_t __c, const int __d)
23431 {
23432   return __builtin_aarch64_sqdmlsl_lanehi (__a, __b, __c, __d);
23433 }
23434
23435 __extension__ extern __inline int32_t
23436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23437 vqdmlslh_laneq_s16 (int32_t __a, int16_t __b, int16x8_t __c, const int __d)
23438 {
23439   return __builtin_aarch64_sqdmlsl_laneqhi (__a, __b, __c, __d);
23440 }
23441
23442 __extension__ extern __inline int64_t
23443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23444 vqdmlsls_s32 (int64_t __a, int32_t __b, int32_t __c)
23445 {
23446   return __builtin_aarch64_sqdmlslsi (__a, __b, __c);
23447 }
23448
23449 __extension__ extern __inline int64_t
23450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23451 vqdmlsls_lane_s32 (int64_t __a, int32_t __b, int32x2_t __c, const int __d)
23452 {
23453   return __builtin_aarch64_sqdmlsl_lanesi (__a, __b, __c, __d);
23454 }
23455
23456 __extension__ extern __inline int64_t
23457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23458 vqdmlsls_laneq_s32 (int64_t __a, int32_t __b, int32x4_t __c, const int __d)
23459 {
23460   return __builtin_aarch64_sqdmlsl_laneqsi (__a, __b, __c, __d);
23461 }
23462
23463 /* vqdmulh */
23464
23465 __extension__ extern __inline int16x4_t
23466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23467 vqdmulh_lane_s16 (int16x4_t __a, int16x4_t __b, const int __c)
23468 {
23469   return __builtin_aarch64_sqdmulh_lanev4hi (__a, __b, __c);
23470 }
23471
23472 __extension__ extern __inline int32x2_t
23473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23474 vqdmulh_lane_s32 (int32x2_t __a, int32x2_t __b, const int __c)
23475 {
23476   return __builtin_aarch64_sqdmulh_lanev2si (__a, __b, __c);
23477 }
23478
23479 __extension__ extern __inline int16x8_t
23480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23481 vqdmulhq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __c)
23482 {
23483   return __builtin_aarch64_sqdmulh_lanev8hi (__a, __b, __c);
23484 }
23485
23486 __extension__ extern __inline int32x4_t
23487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23488 vqdmulhq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __c)
23489 {
23490   return __builtin_aarch64_sqdmulh_lanev4si (__a, __b, __c);
23491 }
23492
23493 __extension__ extern __inline int16_t
23494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23495 vqdmulhh_s16 (int16_t __a, int16_t __b)
23496 {
23497   return (int16_t) __builtin_aarch64_sqdmulhhi (__a, __b);
23498 }
23499
23500 __extension__ extern __inline int16_t
23501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23502 vqdmulhh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
23503 {
23504   return __builtin_aarch64_sqdmulh_lanehi (__a, __b, __c);
23505 }
23506
23507 __extension__ extern __inline int16_t
23508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23509 vqdmulhh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
23510 {
23511   return __builtin_aarch64_sqdmulh_laneqhi (__a, __b, __c);
23512 }
23513
23514 __extension__ extern __inline int32_t
23515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23516 vqdmulhs_s32 (int32_t __a, int32_t __b)
23517 {
23518   return (int32_t) __builtin_aarch64_sqdmulhsi (__a, __b);
23519 }
23520
23521 __extension__ extern __inline int32_t
23522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23523 vqdmulhs_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
23524 {
23525   return __builtin_aarch64_sqdmulh_lanesi (__a, __b, __c);
23526 }
23527
23528 __extension__ extern __inline int32_t
23529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23530 vqdmulhs_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
23531 {
23532   return __builtin_aarch64_sqdmulh_laneqsi (__a, __b, __c);
23533 }
23534
23535 /* vqdmull */
23536
23537 __extension__ extern __inline int32x4_t
23538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23539 vqdmull_s16 (int16x4_t __a, int16x4_t __b)
23540 {
23541   return __builtin_aarch64_sqdmullv4hi (__a, __b);
23542 }
23543
23544 __extension__ extern __inline int32x4_t
23545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23546 vqdmull_high_s16 (int16x8_t __a, int16x8_t __b)
23547 {
23548   return __builtin_aarch64_sqdmull2v8hi (__a, __b);
23549 }
23550
23551 __extension__ extern __inline int32x4_t
23552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23553 vqdmull_high_lane_s16 (int16x8_t __a, int16x4_t __b, int const __c)
23554 {
23555   return __builtin_aarch64_sqdmull2_lanev8hi (__a, __b,__c);
23556 }
23557
23558 __extension__ extern __inline int32x4_t
23559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23560 vqdmull_high_laneq_s16 (int16x8_t __a, int16x8_t __b, int const __c)
23561 {
23562   return __builtin_aarch64_sqdmull2_laneqv8hi (__a, __b,__c);
23563 }
23564
23565 __extension__ extern __inline int32x4_t
23566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23567 vqdmull_high_n_s16 (int16x8_t __a, int16_t __b)
23568 {
23569   return __builtin_aarch64_sqdmull2_nv8hi (__a, __b);
23570 }
23571
23572 __extension__ extern __inline int32x4_t
23573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23574 vqdmull_lane_s16 (int16x4_t __a, int16x4_t __b, int const __c)
23575 {
23576   return __builtin_aarch64_sqdmull_lanev4hi (__a, __b, __c);
23577 }
23578
23579 __extension__ extern __inline int32x4_t
23580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23581 vqdmull_laneq_s16 (int16x4_t __a, int16x8_t __b, int const __c)
23582 {
23583   return __builtin_aarch64_sqdmull_laneqv4hi (__a, __b, __c);
23584 }
23585
23586 __extension__ extern __inline int32x4_t
23587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23588 vqdmull_n_s16 (int16x4_t __a, int16_t __b)
23589 {
23590   return __builtin_aarch64_sqdmull_nv4hi (__a, __b);
23591 }
23592
23593 __extension__ extern __inline int64x2_t
23594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23595 vqdmull_s32 (int32x2_t __a, int32x2_t __b)
23596 {
23597   return __builtin_aarch64_sqdmullv2si (__a, __b);
23598 }
23599
23600 __extension__ extern __inline int64x2_t
23601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23602 vqdmull_high_s32 (int32x4_t __a, int32x4_t __b)
23603 {
23604   return __builtin_aarch64_sqdmull2v4si (__a, __b);
23605 }
23606
23607 __extension__ extern __inline int64x2_t
23608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23609 vqdmull_high_lane_s32 (int32x4_t __a, int32x2_t __b, int const __c)
23610 {
23611   return __builtin_aarch64_sqdmull2_lanev4si (__a, __b, __c);
23612 }
23613
23614 __extension__ extern __inline int64x2_t
23615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23616 vqdmull_high_laneq_s32 (int32x4_t __a, int32x4_t __b, int const __c)
23617 {
23618   return __builtin_aarch64_sqdmull2_laneqv4si (__a, __b, __c);
23619 }
23620
23621 __extension__ extern __inline int64x2_t
23622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23623 vqdmull_high_n_s32 (int32x4_t __a, int32_t __b)
23624 {
23625   return __builtin_aarch64_sqdmull2_nv4si (__a, __b);
23626 }
23627
23628 __extension__ extern __inline int64x2_t
23629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23630 vqdmull_lane_s32 (int32x2_t __a, int32x2_t __b, int const __c)
23631 {
23632   return __builtin_aarch64_sqdmull_lanev2si (__a, __b, __c);
23633 }
23634
23635 __extension__ extern __inline int64x2_t
23636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23637 vqdmull_laneq_s32 (int32x2_t __a, int32x4_t __b, int const __c)
23638 {
23639   return __builtin_aarch64_sqdmull_laneqv2si (__a, __b, __c);
23640 }
23641
23642 __extension__ extern __inline int64x2_t
23643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23644 vqdmull_n_s32 (int32x2_t __a, int32_t __b)
23645 {
23646   return __builtin_aarch64_sqdmull_nv2si (__a, __b);
23647 }
23648
23649 __extension__ extern __inline int32_t
23650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23651 vqdmullh_s16 (int16_t __a, int16_t __b)
23652 {
23653   return (int32_t) __builtin_aarch64_sqdmullhi (__a, __b);
23654 }
23655
23656 __extension__ extern __inline int32_t
23657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23658 vqdmullh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
23659 {
23660   return __builtin_aarch64_sqdmull_lanehi (__a, __b, __c);
23661 }
23662
23663 __extension__ extern __inline int32_t
23664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23665 vqdmullh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
23666 {
23667   return __builtin_aarch64_sqdmull_laneqhi (__a, __b, __c);
23668 }
23669
23670 __extension__ extern __inline int64_t
23671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23672 vqdmulls_s32 (int32_t __a, int32_t __b)
23673 {
23674   return __builtin_aarch64_sqdmullsi (__a, __b);
23675 }
23676
23677 __extension__ extern __inline int64_t
23678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23679 vqdmulls_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
23680 {
23681   return __builtin_aarch64_sqdmull_lanesi (__a, __b, __c);
23682 }
23683
23684 __extension__ extern __inline int64_t
23685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23686 vqdmulls_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
23687 {
23688   return __builtin_aarch64_sqdmull_laneqsi (__a, __b, __c);
23689 }
23690
23691 /* vqmovn */
23692
23693 __extension__ extern __inline int8x8_t
23694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23695 vqmovn_s16 (int16x8_t __a)
23696 {
23697   return (int8x8_t) __builtin_aarch64_sqmovnv8hi (__a);
23698 }
23699
23700 __extension__ extern __inline int16x4_t
23701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23702 vqmovn_s32 (int32x4_t __a)
23703 {
23704   return (int16x4_t) __builtin_aarch64_sqmovnv4si (__a);
23705 }
23706
23707 __extension__ extern __inline int32x2_t
23708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23709 vqmovn_s64 (int64x2_t __a)
23710 {
23711   return (int32x2_t) __builtin_aarch64_sqmovnv2di (__a);
23712 }
23713
23714 __extension__ extern __inline uint8x8_t
23715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23716 vqmovn_u16 (uint16x8_t __a)
23717 {
23718   return (uint8x8_t) __builtin_aarch64_uqmovnv8hi ((int16x8_t) __a);
23719 }
23720
23721 __extension__ extern __inline uint16x4_t
23722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23723 vqmovn_u32 (uint32x4_t __a)
23724 {
23725   return (uint16x4_t) __builtin_aarch64_uqmovnv4si ((int32x4_t) __a);
23726 }
23727
23728 __extension__ extern __inline uint32x2_t
23729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23730 vqmovn_u64 (uint64x2_t __a)
23731 {
23732   return (uint32x2_t) __builtin_aarch64_uqmovnv2di ((int64x2_t) __a);
23733 }
23734
23735 __extension__ extern __inline int8_t
23736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23737 vqmovnh_s16 (int16_t __a)
23738 {
23739   return (int8_t) __builtin_aarch64_sqmovnhi (__a);
23740 }
23741
23742 __extension__ extern __inline int16_t
23743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23744 vqmovns_s32 (int32_t __a)
23745 {
23746   return (int16_t) __builtin_aarch64_sqmovnsi (__a);
23747 }
23748
23749 __extension__ extern __inline int32_t
23750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23751 vqmovnd_s64 (int64_t __a)
23752 {
23753   return (int32_t) __builtin_aarch64_sqmovndi (__a);
23754 }
23755
23756 __extension__ extern __inline uint8_t
23757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23758 vqmovnh_u16 (uint16_t __a)
23759 {
23760   return (uint8_t) __builtin_aarch64_uqmovnhi (__a);
23761 }
23762
23763 __extension__ extern __inline uint16_t
23764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23765 vqmovns_u32 (uint32_t __a)
23766 {
23767   return (uint16_t) __builtin_aarch64_uqmovnsi (__a);
23768 }
23769
23770 __extension__ extern __inline uint32_t
23771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23772 vqmovnd_u64 (uint64_t __a)
23773 {
23774   return (uint32_t) __builtin_aarch64_uqmovndi (__a);
23775 }
23776
23777 /* vqmovun */
23778
23779 __extension__ extern __inline uint8x8_t
23780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23781 vqmovun_s16 (int16x8_t __a)
23782 {
23783   return (uint8x8_t) __builtin_aarch64_sqmovunv8hi (__a);
23784 }
23785
23786 __extension__ extern __inline uint16x4_t
23787 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23788 vqmovun_s32 (int32x4_t __a)
23789 {
23790   return (uint16x4_t) __builtin_aarch64_sqmovunv4si (__a);
23791 }
23792
23793 __extension__ extern __inline uint32x2_t
23794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23795 vqmovun_s64 (int64x2_t __a)
23796 {
23797   return (uint32x2_t) __builtin_aarch64_sqmovunv2di (__a);
23798 }
23799
23800 __extension__ extern __inline int8_t
23801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23802 vqmovunh_s16 (int16_t __a)
23803 {
23804   return (int8_t) __builtin_aarch64_sqmovunhi (__a);
23805 }
23806
23807 __extension__ extern __inline int16_t
23808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23809 vqmovuns_s32 (int32_t __a)
23810 {
23811   return (int16_t) __builtin_aarch64_sqmovunsi (__a);
23812 }
23813
23814 __extension__ extern __inline int32_t
23815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23816 vqmovund_s64 (int64_t __a)
23817 {
23818   return (int32_t) __builtin_aarch64_sqmovundi (__a);
23819 }
23820
23821 /* vqneg */
23822
23823 __extension__ extern __inline int64x2_t
23824 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23825 vqnegq_s64 (int64x2_t __a)
23826 {
23827   return (int64x2_t) __builtin_aarch64_sqnegv2di (__a);
23828 }
23829
23830 __extension__ extern __inline int8_t
23831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23832 vqnegb_s8 (int8_t __a)
23833 {
23834   return (int8_t) __builtin_aarch64_sqnegqi (__a);
23835 }
23836
23837 __extension__ extern __inline int16_t
23838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23839 vqnegh_s16 (int16_t __a)
23840 {
23841   return (int16_t) __builtin_aarch64_sqneghi (__a);
23842 }
23843
23844 __extension__ extern __inline int32_t
23845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23846 vqnegs_s32 (int32_t __a)
23847 {
23848   return (int32_t) __builtin_aarch64_sqnegsi (__a);
23849 }
23850
23851 __extension__ extern __inline int64_t
23852 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23853 vqnegd_s64 (int64_t __a)
23854 {
23855   return __builtin_aarch64_sqnegdi (__a);
23856 }
23857
23858 /* vqrdmulh */
23859
23860 __extension__ extern __inline int16x4_t
23861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23862 vqrdmulh_lane_s16 (int16x4_t __a, int16x4_t __b, const int __c)
23863 {
23864   return  __builtin_aarch64_sqrdmulh_lanev4hi (__a, __b, __c);
23865 }
23866
23867 __extension__ extern __inline int32x2_t
23868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23869 vqrdmulh_lane_s32 (int32x2_t __a, int32x2_t __b, const int __c)
23870 {
23871   return __builtin_aarch64_sqrdmulh_lanev2si (__a, __b, __c);
23872 }
23873
23874 __extension__ extern __inline int16x8_t
23875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23876 vqrdmulhq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __c)
23877 {
23878   return __builtin_aarch64_sqrdmulh_lanev8hi (__a, __b, __c);
23879 }
23880
23881 __extension__ extern __inline int32x4_t
23882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23883 vqrdmulhq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __c)
23884 {
23885   return __builtin_aarch64_sqrdmulh_lanev4si (__a, __b, __c);
23886 }
23887
23888 __extension__ extern __inline int16_t
23889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23890 vqrdmulhh_s16 (int16_t __a, int16_t __b)
23891 {
23892   return (int16_t) __builtin_aarch64_sqrdmulhhi (__a, __b);
23893 }
23894
23895 __extension__ extern __inline int16_t
23896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23897 vqrdmulhh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
23898 {
23899   return __builtin_aarch64_sqrdmulh_lanehi (__a, __b, __c);
23900 }
23901
23902 __extension__ extern __inline int16_t
23903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23904 vqrdmulhh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
23905 {
23906   return __builtin_aarch64_sqrdmulh_laneqhi (__a, __b, __c);
23907 }
23908
23909 __extension__ extern __inline int32_t
23910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23911 vqrdmulhs_s32 (int32_t __a, int32_t __b)
23912 {
23913   return (int32_t) __builtin_aarch64_sqrdmulhsi (__a, __b);
23914 }
23915
23916 __extension__ extern __inline int32_t
23917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23918 vqrdmulhs_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
23919 {
23920   return __builtin_aarch64_sqrdmulh_lanesi (__a, __b, __c);
23921 }
23922
23923 __extension__ extern __inline int32_t
23924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23925 vqrdmulhs_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
23926 {
23927   return __builtin_aarch64_sqrdmulh_laneqsi (__a, __b, __c);
23928 }
23929
23930 /* vqrshl */
23931
23932 __extension__ extern __inline int8x8_t
23933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23934 vqrshl_s8 (int8x8_t __a, int8x8_t __b)
23935 {
23936   return __builtin_aarch64_sqrshlv8qi (__a, __b);
23937 }
23938
23939 __extension__ extern __inline int16x4_t
23940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23941 vqrshl_s16 (int16x4_t __a, int16x4_t __b)
23942 {
23943   return __builtin_aarch64_sqrshlv4hi (__a, __b);
23944 }
23945
23946 __extension__ extern __inline int32x2_t
23947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23948 vqrshl_s32 (int32x2_t __a, int32x2_t __b)
23949 {
23950   return __builtin_aarch64_sqrshlv2si (__a, __b);
23951 }
23952
23953 __extension__ extern __inline int64x1_t
23954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23955 vqrshl_s64 (int64x1_t __a, int64x1_t __b)
23956 {
23957   return (int64x1_t) {__builtin_aarch64_sqrshldi (__a[0], __b[0])};
23958 }
23959
23960 __extension__ extern __inline uint8x8_t
23961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23962 vqrshl_u8 (uint8x8_t __a, int8x8_t __b)
23963 {
23964   return __builtin_aarch64_uqrshlv8qi_uus ( __a, __b);
23965 }
23966
23967 __extension__ extern __inline uint16x4_t
23968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23969 vqrshl_u16 (uint16x4_t __a, int16x4_t __b)
23970 {
23971   return __builtin_aarch64_uqrshlv4hi_uus ( __a, __b);
23972 }
23973
23974 __extension__ extern __inline uint32x2_t
23975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23976 vqrshl_u32 (uint32x2_t __a, int32x2_t __b)
23977 {
23978   return __builtin_aarch64_uqrshlv2si_uus ( __a, __b);
23979 }
23980
23981 __extension__ extern __inline uint64x1_t
23982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23983 vqrshl_u64 (uint64x1_t __a, int64x1_t __b)
23984 {
23985   return (uint64x1_t) {__builtin_aarch64_uqrshldi_uus (__a[0], __b[0])};
23986 }
23987
23988 __extension__ extern __inline int8x16_t
23989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23990 vqrshlq_s8 (int8x16_t __a, int8x16_t __b)
23991 {
23992   return __builtin_aarch64_sqrshlv16qi (__a, __b);
23993 }
23994
23995 __extension__ extern __inline int16x8_t
23996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23997 vqrshlq_s16 (int16x8_t __a, int16x8_t __b)
23998 {
23999   return __builtin_aarch64_sqrshlv8hi (__a, __b);
24000 }
24001
24002 __extension__ extern __inline int32x4_t
24003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24004 vqrshlq_s32 (int32x4_t __a, int32x4_t __b)
24005 {
24006   return __builtin_aarch64_sqrshlv4si (__a, __b);
24007 }
24008
24009 __extension__ extern __inline int64x2_t
24010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24011 vqrshlq_s64 (int64x2_t __a, int64x2_t __b)
24012 {
24013   return __builtin_aarch64_sqrshlv2di (__a, __b);
24014 }
24015
24016 __extension__ extern __inline uint8x16_t
24017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24018 vqrshlq_u8 (uint8x16_t __a, int8x16_t __b)
24019 {
24020   return __builtin_aarch64_uqrshlv16qi_uus ( __a, __b);
24021 }
24022
24023 __extension__ extern __inline uint16x8_t
24024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24025 vqrshlq_u16 (uint16x8_t __a, int16x8_t __b)
24026 {
24027   return __builtin_aarch64_uqrshlv8hi_uus ( __a, __b);
24028 }
24029
24030 __extension__ extern __inline uint32x4_t
24031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24032 vqrshlq_u32 (uint32x4_t __a, int32x4_t __b)
24033 {
24034   return __builtin_aarch64_uqrshlv4si_uus ( __a, __b);
24035 }
24036
24037 __extension__ extern __inline uint64x2_t
24038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24039 vqrshlq_u64 (uint64x2_t __a, int64x2_t __b)
24040 {
24041   return __builtin_aarch64_uqrshlv2di_uus ( __a, __b);
24042 }
24043
24044 __extension__ extern __inline int8_t
24045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24046 vqrshlb_s8 (int8_t __a, int8_t __b)
24047 {
24048   return __builtin_aarch64_sqrshlqi (__a, __b);
24049 }
24050
24051 __extension__ extern __inline int16_t
24052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24053 vqrshlh_s16 (int16_t __a, int16_t __b)
24054 {
24055   return __builtin_aarch64_sqrshlhi (__a, __b);
24056 }
24057
24058 __extension__ extern __inline int32_t
24059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24060 vqrshls_s32 (int32_t __a, int32_t __b)
24061 {
24062   return __builtin_aarch64_sqrshlsi (__a, __b);
24063 }
24064
24065 __extension__ extern __inline int64_t
24066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24067 vqrshld_s64 (int64_t __a, int64_t __b)
24068 {
24069   return __builtin_aarch64_sqrshldi (__a, __b);
24070 }
24071
24072 __extension__ extern __inline uint8_t
24073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24074 vqrshlb_u8 (uint8_t __a, uint8_t __b)
24075 {
24076   return __builtin_aarch64_uqrshlqi_uus (__a, __b);
24077 }
24078
24079 __extension__ extern __inline uint16_t
24080 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24081 vqrshlh_u16 (uint16_t __a, uint16_t __b)
24082 {
24083   return __builtin_aarch64_uqrshlhi_uus (__a, __b);
24084 }
24085
24086 __extension__ extern __inline uint32_t
24087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24088 vqrshls_u32 (uint32_t __a, uint32_t __b)
24089 {
24090   return __builtin_aarch64_uqrshlsi_uus (__a, __b);
24091 }
24092
24093 __extension__ extern __inline uint64_t
24094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24095 vqrshld_u64 (uint64_t __a, uint64_t __b)
24096 {
24097   return __builtin_aarch64_uqrshldi_uus (__a, __b);
24098 }
24099
24100 /* vqrshrn */
24101
24102 __extension__ extern __inline int8x8_t
24103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24104 vqrshrn_n_s16 (int16x8_t __a, const int __b)
24105 {
24106   return (int8x8_t) __builtin_aarch64_sqrshrn_nv8hi (__a, __b);
24107 }
24108
24109 __extension__ extern __inline int16x4_t
24110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24111 vqrshrn_n_s32 (int32x4_t __a, const int __b)
24112 {
24113   return (int16x4_t) __builtin_aarch64_sqrshrn_nv4si (__a, __b);
24114 }
24115
24116 __extension__ extern __inline int32x2_t
24117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24118 vqrshrn_n_s64 (int64x2_t __a, const int __b)
24119 {
24120   return (int32x2_t) __builtin_aarch64_sqrshrn_nv2di (__a, __b);
24121 }
24122
24123 __extension__ extern __inline uint8x8_t
24124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24125 vqrshrn_n_u16 (uint16x8_t __a, const int __b)
24126 {
24127   return __builtin_aarch64_uqrshrn_nv8hi_uus ( __a, __b);
24128 }
24129
24130 __extension__ extern __inline uint16x4_t
24131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24132 vqrshrn_n_u32 (uint32x4_t __a, const int __b)
24133 {
24134   return __builtin_aarch64_uqrshrn_nv4si_uus ( __a, __b);
24135 }
24136
24137 __extension__ extern __inline uint32x2_t
24138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24139 vqrshrn_n_u64 (uint64x2_t __a, const int __b)
24140 {
24141   return __builtin_aarch64_uqrshrn_nv2di_uus ( __a, __b);
24142 }
24143
24144 __extension__ extern __inline int8_t
24145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24146 vqrshrnh_n_s16 (int16_t __a, const int __b)
24147 {
24148   return (int8_t) __builtin_aarch64_sqrshrn_nhi (__a, __b);
24149 }
24150
24151 __extension__ extern __inline int16_t
24152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24153 vqrshrns_n_s32 (int32_t __a, const int __b)
24154 {
24155   return (int16_t) __builtin_aarch64_sqrshrn_nsi (__a, __b);
24156 }
24157
24158 __extension__ extern __inline int32_t
24159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24160 vqrshrnd_n_s64 (int64_t __a, const int __b)
24161 {
24162   return (int32_t) __builtin_aarch64_sqrshrn_ndi (__a, __b);
24163 }
24164
24165 __extension__ extern __inline uint8_t
24166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24167 vqrshrnh_n_u16 (uint16_t __a, const int __b)
24168 {
24169   return __builtin_aarch64_uqrshrn_nhi_uus (__a, __b);
24170 }
24171
24172 __extension__ extern __inline uint16_t
24173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24174 vqrshrns_n_u32 (uint32_t __a, const int __b)
24175 {
24176   return __builtin_aarch64_uqrshrn_nsi_uus (__a, __b);
24177 }
24178
24179 __extension__ extern __inline uint32_t
24180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24181 vqrshrnd_n_u64 (uint64_t __a, const int __b)
24182 {
24183   return __builtin_aarch64_uqrshrn_ndi_uus (__a, __b);
24184 }
24185
24186 /* vqrshrun */
24187
24188 __extension__ extern __inline uint8x8_t
24189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24190 vqrshrun_n_s16 (int16x8_t __a, const int __b)
24191 {
24192   return (uint8x8_t) __builtin_aarch64_sqrshrun_nv8hi (__a, __b);
24193 }
24194
24195 __extension__ extern __inline uint16x4_t
24196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24197 vqrshrun_n_s32 (int32x4_t __a, const int __b)
24198 {
24199   return (uint16x4_t) __builtin_aarch64_sqrshrun_nv4si (__a, __b);
24200 }
24201
24202 __extension__ extern __inline uint32x2_t
24203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24204 vqrshrun_n_s64 (int64x2_t __a, const int __b)
24205 {
24206   return (uint32x2_t) __builtin_aarch64_sqrshrun_nv2di (__a, __b);
24207 }
24208
24209 __extension__ extern __inline int8_t
24210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24211 vqrshrunh_n_s16 (int16_t __a, const int __b)
24212 {
24213   return (int8_t) __builtin_aarch64_sqrshrun_nhi (__a, __b);
24214 }
24215
24216 __extension__ extern __inline int16_t
24217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24218 vqrshruns_n_s32 (int32_t __a, const int __b)
24219 {
24220   return (int16_t) __builtin_aarch64_sqrshrun_nsi (__a, __b);
24221 }
24222
24223 __extension__ extern __inline int32_t
24224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24225 vqrshrund_n_s64 (int64_t __a, const int __b)
24226 {
24227   return (int32_t) __builtin_aarch64_sqrshrun_ndi (__a, __b);
24228 }
24229
24230 /* vqshl */
24231
24232 __extension__ extern __inline int8x8_t
24233 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24234 vqshl_s8 (int8x8_t __a, int8x8_t __b)
24235 {
24236   return __builtin_aarch64_sqshlv8qi (__a, __b);
24237 }
24238
24239 __extension__ extern __inline int16x4_t
24240 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24241 vqshl_s16 (int16x4_t __a, int16x4_t __b)
24242 {
24243   return __builtin_aarch64_sqshlv4hi (__a, __b);
24244 }
24245
24246 __extension__ extern __inline int32x2_t
24247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24248 vqshl_s32 (int32x2_t __a, int32x2_t __b)
24249 {
24250   return __builtin_aarch64_sqshlv2si (__a, __b);
24251 }
24252
24253 __extension__ extern __inline int64x1_t
24254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24255 vqshl_s64 (int64x1_t __a, int64x1_t __b)
24256 {
24257   return (int64x1_t) {__builtin_aarch64_sqshldi (__a[0], __b[0])};
24258 }
24259
24260 __extension__ extern __inline uint8x8_t
24261 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24262 vqshl_u8 (uint8x8_t __a, int8x8_t __b)
24263 {
24264   return __builtin_aarch64_uqshlv8qi_uus ( __a, __b);
24265 }
24266
24267 __extension__ extern __inline uint16x4_t
24268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24269 vqshl_u16 (uint16x4_t __a, int16x4_t __b)
24270 {
24271   return __builtin_aarch64_uqshlv4hi_uus ( __a, __b);
24272 }
24273
24274 __extension__ extern __inline uint32x2_t
24275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24276 vqshl_u32 (uint32x2_t __a, int32x2_t __b)
24277 {
24278   return __builtin_aarch64_uqshlv2si_uus ( __a, __b);
24279 }
24280
24281 __extension__ extern __inline uint64x1_t
24282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24283 vqshl_u64 (uint64x1_t __a, int64x1_t __b)
24284 {
24285   return (uint64x1_t) {__builtin_aarch64_uqshldi_uus (__a[0], __b[0])};
24286 }
24287
24288 __extension__ extern __inline int8x16_t
24289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24290 vqshlq_s8 (int8x16_t __a, int8x16_t __b)
24291 {
24292   return __builtin_aarch64_sqshlv16qi (__a, __b);
24293 }
24294
24295 __extension__ extern __inline int16x8_t
24296 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24297 vqshlq_s16 (int16x8_t __a, int16x8_t __b)
24298 {
24299   return __builtin_aarch64_sqshlv8hi (__a, __b);
24300 }
24301
24302 __extension__ extern __inline int32x4_t
24303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24304 vqshlq_s32 (int32x4_t __a, int32x4_t __b)
24305 {
24306   return __builtin_aarch64_sqshlv4si (__a, __b);
24307 }
24308
24309 __extension__ extern __inline int64x2_t
24310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24311 vqshlq_s64 (int64x2_t __a, int64x2_t __b)
24312 {
24313   return __builtin_aarch64_sqshlv2di (__a, __b);
24314 }
24315
24316 __extension__ extern __inline uint8x16_t
24317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24318 vqshlq_u8 (uint8x16_t __a, int8x16_t __b)
24319 {
24320   return __builtin_aarch64_uqshlv16qi_uus ( __a, __b);
24321 }
24322
24323 __extension__ extern __inline uint16x8_t
24324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24325 vqshlq_u16 (uint16x8_t __a, int16x8_t __b)
24326 {
24327   return __builtin_aarch64_uqshlv8hi_uus ( __a, __b);
24328 }
24329
24330 __extension__ extern __inline uint32x4_t
24331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24332 vqshlq_u32 (uint32x4_t __a, int32x4_t __b)
24333 {
24334   return __builtin_aarch64_uqshlv4si_uus ( __a, __b);
24335 }
24336
24337 __extension__ extern __inline uint64x2_t
24338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24339 vqshlq_u64 (uint64x2_t __a, int64x2_t __b)
24340 {
24341   return __builtin_aarch64_uqshlv2di_uus ( __a, __b);
24342 }
24343
24344 __extension__ extern __inline int8_t
24345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24346 vqshlb_s8 (int8_t __a, int8_t __b)
24347 {
24348   return __builtin_aarch64_sqshlqi (__a, __b);
24349 }
24350
24351 __extension__ extern __inline int16_t
24352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24353 vqshlh_s16 (int16_t __a, int16_t __b)
24354 {
24355   return __builtin_aarch64_sqshlhi (__a, __b);
24356 }
24357
24358 __extension__ extern __inline int32_t
24359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24360 vqshls_s32 (int32_t __a, int32_t __b)
24361 {
24362   return __builtin_aarch64_sqshlsi (__a, __b);
24363 }
24364
24365 __extension__ extern __inline int64_t
24366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24367 vqshld_s64 (int64_t __a, int64_t __b)
24368 {
24369   return __builtin_aarch64_sqshldi (__a, __b);
24370 }
24371
24372 __extension__ extern __inline uint8_t
24373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24374 vqshlb_u8 (uint8_t __a, uint8_t __b)
24375 {
24376   return __builtin_aarch64_uqshlqi_uus (__a, __b);
24377 }
24378
24379 __extension__ extern __inline uint16_t
24380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24381 vqshlh_u16 (uint16_t __a, uint16_t __b)
24382 {
24383   return __builtin_aarch64_uqshlhi_uus (__a, __b);
24384 }
24385
24386 __extension__ extern __inline uint32_t
24387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24388 vqshls_u32 (uint32_t __a, uint32_t __b)
24389 {
24390   return __builtin_aarch64_uqshlsi_uus (__a, __b);
24391 }
24392
24393 __extension__ extern __inline uint64_t
24394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24395 vqshld_u64 (uint64_t __a, uint64_t __b)
24396 {
24397   return __builtin_aarch64_uqshldi_uus (__a, __b);
24398 }
24399
24400 __extension__ extern __inline int8x8_t
24401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24402 vqshl_n_s8 (int8x8_t __a, const int __b)
24403 {
24404   return (int8x8_t) __builtin_aarch64_sqshl_nv8qi (__a, __b);
24405 }
24406
24407 __extension__ extern __inline int16x4_t
24408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24409 vqshl_n_s16 (int16x4_t __a, const int __b)
24410 {
24411   return (int16x4_t) __builtin_aarch64_sqshl_nv4hi (__a, __b);
24412 }
24413
24414 __extension__ extern __inline int32x2_t
24415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24416 vqshl_n_s32 (int32x2_t __a, const int __b)
24417 {
24418   return (int32x2_t) __builtin_aarch64_sqshl_nv2si (__a, __b);
24419 }
24420
24421 __extension__ extern __inline int64x1_t
24422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24423 vqshl_n_s64 (int64x1_t __a, const int __b)
24424 {
24425   return (int64x1_t) {__builtin_aarch64_sqshl_ndi (__a[0], __b)};
24426 }
24427
24428 __extension__ extern __inline uint8x8_t
24429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24430 vqshl_n_u8 (uint8x8_t __a, const int __b)
24431 {
24432   return __builtin_aarch64_uqshl_nv8qi_uus (__a, __b);
24433 }
24434
24435 __extension__ extern __inline uint16x4_t
24436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24437 vqshl_n_u16 (uint16x4_t __a, const int __b)
24438 {
24439   return __builtin_aarch64_uqshl_nv4hi_uus (__a, __b);
24440 }
24441
24442 __extension__ extern __inline uint32x2_t
24443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24444 vqshl_n_u32 (uint32x2_t __a, const int __b)
24445 {
24446   return __builtin_aarch64_uqshl_nv2si_uus (__a, __b);
24447 }
24448
24449 __extension__ extern __inline uint64x1_t
24450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24451 vqshl_n_u64 (uint64x1_t __a, const int __b)
24452 {
24453   return (uint64x1_t) {__builtin_aarch64_uqshl_ndi_uus (__a[0], __b)};
24454 }
24455
24456 __extension__ extern __inline int8x16_t
24457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24458 vqshlq_n_s8 (int8x16_t __a, const int __b)
24459 {
24460   return (int8x16_t) __builtin_aarch64_sqshl_nv16qi (__a, __b);
24461 }
24462
24463 __extension__ extern __inline int16x8_t
24464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24465 vqshlq_n_s16 (int16x8_t __a, const int __b)
24466 {
24467   return (int16x8_t) __builtin_aarch64_sqshl_nv8hi (__a, __b);
24468 }
24469
24470 __extension__ extern __inline int32x4_t
24471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24472 vqshlq_n_s32 (int32x4_t __a, const int __b)
24473 {
24474   return (int32x4_t) __builtin_aarch64_sqshl_nv4si (__a, __b);
24475 }
24476
24477 __extension__ extern __inline int64x2_t
24478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24479 vqshlq_n_s64 (int64x2_t __a, const int __b)
24480 {
24481   return (int64x2_t) __builtin_aarch64_sqshl_nv2di (__a, __b);
24482 }
24483
24484 __extension__ extern __inline uint8x16_t
24485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24486 vqshlq_n_u8 (uint8x16_t __a, const int __b)
24487 {
24488   return __builtin_aarch64_uqshl_nv16qi_uus (__a, __b);
24489 }
24490
24491 __extension__ extern __inline uint16x8_t
24492 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24493 vqshlq_n_u16 (uint16x8_t __a, const int __b)
24494 {
24495   return __builtin_aarch64_uqshl_nv8hi_uus (__a, __b);
24496 }
24497
24498 __extension__ extern __inline uint32x4_t
24499 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24500 vqshlq_n_u32 (uint32x4_t __a, const int __b)
24501 {
24502   return __builtin_aarch64_uqshl_nv4si_uus (__a, __b);
24503 }
24504
24505 __extension__ extern __inline uint64x2_t
24506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24507 vqshlq_n_u64 (uint64x2_t __a, const int __b)
24508 {
24509   return __builtin_aarch64_uqshl_nv2di_uus (__a, __b);
24510 }
24511
24512 __extension__ extern __inline int8_t
24513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24514 vqshlb_n_s8 (int8_t __a, const int __b)
24515 {
24516   return (int8_t) __builtin_aarch64_sqshl_nqi (__a, __b);
24517 }
24518
24519 __extension__ extern __inline int16_t
24520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24521 vqshlh_n_s16 (int16_t __a, const int __b)
24522 {
24523   return (int16_t) __builtin_aarch64_sqshl_nhi (__a, __b);
24524 }
24525
24526 __extension__ extern __inline int32_t
24527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24528 vqshls_n_s32 (int32_t __a, const int __b)
24529 {
24530   return (int32_t) __builtin_aarch64_sqshl_nsi (__a, __b);
24531 }
24532
24533 __extension__ extern __inline int64_t
24534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24535 vqshld_n_s64 (int64_t __a, const int __b)
24536 {
24537   return __builtin_aarch64_sqshl_ndi (__a, __b);
24538 }
24539
24540 __extension__ extern __inline uint8_t
24541 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24542 vqshlb_n_u8 (uint8_t __a, const int __b)
24543 {
24544   return __builtin_aarch64_uqshl_nqi_uus (__a, __b);
24545 }
24546
24547 __extension__ extern __inline uint16_t
24548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24549 vqshlh_n_u16 (uint16_t __a, const int __b)
24550 {
24551   return __builtin_aarch64_uqshl_nhi_uus (__a, __b);
24552 }
24553
24554 __extension__ extern __inline uint32_t
24555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24556 vqshls_n_u32 (uint32_t __a, const int __b)
24557 {
24558   return __builtin_aarch64_uqshl_nsi_uus (__a, __b);
24559 }
24560
24561 __extension__ extern __inline uint64_t
24562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24563 vqshld_n_u64 (uint64_t __a, const int __b)
24564 {
24565   return __builtin_aarch64_uqshl_ndi_uus (__a, __b);
24566 }
24567
24568 /* vqshlu */
24569
24570 __extension__ extern __inline uint8x8_t
24571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24572 vqshlu_n_s8 (int8x8_t __a, const int __b)
24573 {
24574   return __builtin_aarch64_sqshlu_nv8qi_uss (__a, __b);
24575 }
24576
24577 __extension__ extern __inline uint16x4_t
24578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24579 vqshlu_n_s16 (int16x4_t __a, const int __b)
24580 {
24581   return __builtin_aarch64_sqshlu_nv4hi_uss (__a, __b);
24582 }
24583
24584 __extension__ extern __inline uint32x2_t
24585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24586 vqshlu_n_s32 (int32x2_t __a, const int __b)
24587 {
24588   return __builtin_aarch64_sqshlu_nv2si_uss (__a, __b);
24589 }
24590
24591 __extension__ extern __inline uint64x1_t
24592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24593 vqshlu_n_s64 (int64x1_t __a, const int __b)
24594 {
24595   return (uint64x1_t) {__builtin_aarch64_sqshlu_ndi_uss (__a[0], __b)};
24596 }
24597
24598 __extension__ extern __inline uint8x16_t
24599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24600 vqshluq_n_s8 (int8x16_t __a, const int __b)
24601 {
24602   return __builtin_aarch64_sqshlu_nv16qi_uss (__a, __b);
24603 }
24604
24605 __extension__ extern __inline uint16x8_t
24606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24607 vqshluq_n_s16 (int16x8_t __a, const int __b)
24608 {
24609   return __builtin_aarch64_sqshlu_nv8hi_uss (__a, __b);
24610 }
24611
24612 __extension__ extern __inline uint32x4_t
24613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24614 vqshluq_n_s32 (int32x4_t __a, const int __b)
24615 {
24616   return __builtin_aarch64_sqshlu_nv4si_uss (__a, __b);
24617 }
24618
24619 __extension__ extern __inline uint64x2_t
24620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24621 vqshluq_n_s64 (int64x2_t __a, const int __b)
24622 {
24623   return __builtin_aarch64_sqshlu_nv2di_uss (__a, __b);
24624 }
24625
24626 __extension__ extern __inline int8_t
24627 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24628 vqshlub_n_s8 (int8_t __a, const int __b)
24629 {
24630   return (int8_t) __builtin_aarch64_sqshlu_nqi_uss (__a, __b);
24631 }
24632
24633 __extension__ extern __inline int16_t
24634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24635 vqshluh_n_s16 (int16_t __a, const int __b)
24636 {
24637   return (int16_t) __builtin_aarch64_sqshlu_nhi_uss (__a, __b);
24638 }
24639
24640 __extension__ extern __inline int32_t
24641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24642 vqshlus_n_s32 (int32_t __a, const int __b)
24643 {
24644   return (int32_t) __builtin_aarch64_sqshlu_nsi_uss (__a, __b);
24645 }
24646
24647 __extension__ extern __inline uint64_t
24648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24649 vqshlud_n_s64 (int64_t __a, const int __b)
24650 {
24651   return __builtin_aarch64_sqshlu_ndi_uss (__a, __b);
24652 }
24653
24654 /* vqshrn */
24655
24656 __extension__ extern __inline int8x8_t
24657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24658 vqshrn_n_s16 (int16x8_t __a, const int __b)
24659 {
24660   return (int8x8_t) __builtin_aarch64_sqshrn_nv8hi (__a, __b);
24661 }
24662
24663 __extension__ extern __inline int16x4_t
24664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24665 vqshrn_n_s32 (int32x4_t __a, const int __b)
24666 {
24667   return (int16x4_t) __builtin_aarch64_sqshrn_nv4si (__a, __b);
24668 }
24669
24670 __extension__ extern __inline int32x2_t
24671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24672 vqshrn_n_s64 (int64x2_t __a, const int __b)
24673 {
24674   return (int32x2_t) __builtin_aarch64_sqshrn_nv2di (__a, __b);
24675 }
24676
24677 __extension__ extern __inline uint8x8_t
24678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24679 vqshrn_n_u16 (uint16x8_t __a, const int __b)
24680 {
24681   return __builtin_aarch64_uqshrn_nv8hi_uus ( __a, __b);
24682 }
24683
24684 __extension__ extern __inline uint16x4_t
24685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24686 vqshrn_n_u32 (uint32x4_t __a, const int __b)
24687 {
24688   return __builtin_aarch64_uqshrn_nv4si_uus ( __a, __b);
24689 }
24690
24691 __extension__ extern __inline uint32x2_t
24692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24693 vqshrn_n_u64 (uint64x2_t __a, const int __b)
24694 {
24695   return __builtin_aarch64_uqshrn_nv2di_uus ( __a, __b);
24696 }
24697
24698 __extension__ extern __inline int8_t
24699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24700 vqshrnh_n_s16 (int16_t __a, const int __b)
24701 {
24702   return (int8_t) __builtin_aarch64_sqshrn_nhi (__a, __b);
24703 }
24704
24705 __extension__ extern __inline int16_t
24706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24707 vqshrns_n_s32 (int32_t __a, const int __b)
24708 {
24709   return (int16_t) __builtin_aarch64_sqshrn_nsi (__a, __b);
24710 }
24711
24712 __extension__ extern __inline int32_t
24713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24714 vqshrnd_n_s64 (int64_t __a, const int __b)
24715 {
24716   return (int32_t) __builtin_aarch64_sqshrn_ndi (__a, __b);
24717 }
24718
24719 __extension__ extern __inline uint8_t
24720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24721 vqshrnh_n_u16 (uint16_t __a, const int __b)
24722 {
24723   return __builtin_aarch64_uqshrn_nhi_uus (__a, __b);
24724 }
24725
24726 __extension__ extern __inline uint16_t
24727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24728 vqshrns_n_u32 (uint32_t __a, const int __b)
24729 {
24730   return __builtin_aarch64_uqshrn_nsi_uus (__a, __b);
24731 }
24732
24733 __extension__ extern __inline uint32_t
24734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24735 vqshrnd_n_u64 (uint64_t __a, const int __b)
24736 {
24737   return __builtin_aarch64_uqshrn_ndi_uus (__a, __b);
24738 }
24739
24740 /* vqshrun */
24741
24742 __extension__ extern __inline uint8x8_t
24743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24744 vqshrun_n_s16 (int16x8_t __a, const int __b)
24745 {
24746   return (uint8x8_t) __builtin_aarch64_sqshrun_nv8hi (__a, __b);
24747 }
24748
24749 __extension__ extern __inline uint16x4_t
24750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24751 vqshrun_n_s32 (int32x4_t __a, const int __b)
24752 {
24753   return (uint16x4_t) __builtin_aarch64_sqshrun_nv4si (__a, __b);
24754 }
24755
24756 __extension__ extern __inline uint32x2_t
24757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24758 vqshrun_n_s64 (int64x2_t __a, const int __b)
24759 {
24760   return (uint32x2_t) __builtin_aarch64_sqshrun_nv2di (__a, __b);
24761 }
24762
24763 __extension__ extern __inline int8_t
24764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24765 vqshrunh_n_s16 (int16_t __a, const int __b)
24766 {
24767   return (int8_t) __builtin_aarch64_sqshrun_nhi (__a, __b);
24768 }
24769
24770 __extension__ extern __inline int16_t
24771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24772 vqshruns_n_s32 (int32_t __a, const int __b)
24773 {
24774   return (int16_t) __builtin_aarch64_sqshrun_nsi (__a, __b);
24775 }
24776
24777 __extension__ extern __inline int32_t
24778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24779 vqshrund_n_s64 (int64_t __a, const int __b)
24780 {
24781   return (int32_t) __builtin_aarch64_sqshrun_ndi (__a, __b);
24782 }
24783
24784 /* vqsub */
24785
24786 __extension__ extern __inline int8_t
24787 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24788 vqsubb_s8 (int8_t __a, int8_t __b)
24789 {
24790   return (int8_t) __builtin_aarch64_sqsubqi (__a, __b);
24791 }
24792
24793 __extension__ extern __inline int16_t
24794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24795 vqsubh_s16 (int16_t __a, int16_t __b)
24796 {
24797   return (int16_t) __builtin_aarch64_sqsubhi (__a, __b);
24798 }
24799
24800 __extension__ extern __inline int32_t
24801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24802 vqsubs_s32 (int32_t __a, int32_t __b)
24803 {
24804   return (int32_t) __builtin_aarch64_sqsubsi (__a, __b);
24805 }
24806
24807 __extension__ extern __inline int64_t
24808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24809 vqsubd_s64 (int64_t __a, int64_t __b)
24810 {
24811   return __builtin_aarch64_sqsubdi (__a, __b);
24812 }
24813
24814 __extension__ extern __inline uint8_t
24815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24816 vqsubb_u8 (uint8_t __a, uint8_t __b)
24817 {
24818   return (uint8_t) __builtin_aarch64_uqsubqi_uuu (__a, __b);
24819 }
24820
24821 __extension__ extern __inline uint16_t
24822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24823 vqsubh_u16 (uint16_t __a, uint16_t __b)
24824 {
24825   return (uint16_t) __builtin_aarch64_uqsubhi_uuu (__a, __b);
24826 }
24827
24828 __extension__ extern __inline uint32_t
24829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24830 vqsubs_u32 (uint32_t __a, uint32_t __b)
24831 {
24832   return (uint32_t) __builtin_aarch64_uqsubsi_uuu (__a, __b);
24833 }
24834
24835 __extension__ extern __inline uint64_t
24836 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24837 vqsubd_u64 (uint64_t __a, uint64_t __b)
24838 {
24839   return __builtin_aarch64_uqsubdi_uuu (__a, __b);
24840 }
24841
24842 /* vqtbl2 */
24843
24844 __extension__ extern __inline int8x8_t
24845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24846 vqtbl2_s8 (int8x16x2_t tab, uint8x8_t idx)
24847 {
24848   __builtin_aarch64_simd_oi __o;
24849   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
24850   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
24851   return __builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
24852 }
24853
24854 __extension__ extern __inline uint8x8_t
24855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24856 vqtbl2_u8 (uint8x16x2_t tab, uint8x8_t idx)
24857 {
24858   __builtin_aarch64_simd_oi __o;
24859   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24860   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24861   return (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
24862 }
24863
24864 __extension__ extern __inline poly8x8_t
24865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24866 vqtbl2_p8 (poly8x16x2_t tab, uint8x8_t idx)
24867 {
24868   __builtin_aarch64_simd_oi __o;
24869   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24870   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24871   return (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
24872 }
24873
24874 __extension__ extern __inline int8x16_t
24875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24876 vqtbl2q_s8 (int8x16x2_t tab, uint8x16_t idx)
24877 {
24878   __builtin_aarch64_simd_oi __o;
24879   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24880   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24881   return __builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
24882 }
24883
24884 __extension__ extern __inline uint8x16_t
24885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24886 vqtbl2q_u8 (uint8x16x2_t tab, uint8x16_t idx)
24887 {
24888   __builtin_aarch64_simd_oi __o;
24889   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24890   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24891   return (uint8x16_t)__builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
24892 }
24893
24894 __extension__ extern __inline poly8x16_t
24895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24896 vqtbl2q_p8 (poly8x16x2_t tab, uint8x16_t idx)
24897 {
24898   __builtin_aarch64_simd_oi __o;
24899   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24900   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24901   return (poly8x16_t)__builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
24902 }
24903
24904 /* vqtbl3 */
24905
24906 __extension__ extern __inline int8x8_t
24907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24908 vqtbl3_s8 (int8x16x3_t tab, uint8x8_t idx)
24909 {
24910   __builtin_aarch64_simd_ci __o;
24911   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24912   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24913   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24914   return __builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
24915 }
24916
24917 __extension__ extern __inline uint8x8_t
24918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24919 vqtbl3_u8 (uint8x16x3_t tab, uint8x8_t idx)
24920 {
24921   __builtin_aarch64_simd_ci __o;
24922   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24923   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24924   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24925   return (uint8x8_t)__builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
24926 }
24927
24928 __extension__ extern __inline poly8x8_t
24929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24930 vqtbl3_p8 (poly8x16x3_t tab, uint8x8_t idx)
24931 {
24932   __builtin_aarch64_simd_ci __o;
24933   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24934   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24935   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24936   return (poly8x8_t)__builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
24937 }
24938
24939 __extension__ extern __inline int8x16_t
24940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24941 vqtbl3q_s8 (int8x16x3_t tab, uint8x16_t idx)
24942 {
24943   __builtin_aarch64_simd_ci __o;
24944   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24945   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24946   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24947   return __builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24948 }
24949
24950 __extension__ extern __inline uint8x16_t
24951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24952 vqtbl3q_u8 (uint8x16x3_t tab, uint8x16_t idx)
24953 {
24954   __builtin_aarch64_simd_ci __o;
24955   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24956   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24957   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24958   return (uint8x16_t)__builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24959 }
24960
24961 __extension__ extern __inline poly8x16_t
24962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24963 vqtbl3q_p8 (poly8x16x3_t tab, uint8x16_t idx)
24964 {
24965   __builtin_aarch64_simd_ci __o;
24966   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24967   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24968   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24969   return (poly8x16_t)__builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24970 }
24971
24972 /* vqtbl4 */
24973
24974 __extension__ extern __inline int8x8_t
24975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24976 vqtbl4_s8 (int8x16x4_t tab, uint8x8_t idx)
24977 {
24978   __builtin_aarch64_simd_xi __o;
24979   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24980   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24981   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24982   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24983   return __builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24984 }
24985
24986 __extension__ extern __inline uint8x8_t
24987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24988 vqtbl4_u8 (uint8x16x4_t tab, uint8x8_t idx)
24989 {
24990   __builtin_aarch64_simd_xi __o;
24991   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24992   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24993   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24994   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24995   return (uint8x8_t)__builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24996 }
24997
24998 __extension__ extern __inline poly8x8_t
24999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25000 vqtbl4_p8 (poly8x16x4_t tab, uint8x8_t idx)
25001 {
25002   __builtin_aarch64_simd_xi __o;
25003   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25004   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25005   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25006   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25007   return (poly8x8_t)__builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
25008 }
25009
25010 __extension__ extern __inline int8x16_t
25011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25012 vqtbl4q_s8 (int8x16x4_t tab, uint8x16_t idx)
25013 {
25014   __builtin_aarch64_simd_xi __o;
25015   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25016   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25017   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25018   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25019   return __builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
25020 }
25021
25022 __extension__ extern __inline uint8x16_t
25023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25024 vqtbl4q_u8 (uint8x16x4_t tab, uint8x16_t idx)
25025 {
25026   __builtin_aarch64_simd_xi __o;
25027   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25028   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25029   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25030   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25031   return (uint8x16_t)__builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
25032 }
25033
25034 __extension__ extern __inline poly8x16_t
25035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25036 vqtbl4q_p8 (poly8x16x4_t tab, uint8x16_t idx)
25037 {
25038   __builtin_aarch64_simd_xi __o;
25039   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25040   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25041   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25042   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25043   return (poly8x16_t)__builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
25044 }
25045
25046
25047 /* vqtbx2 */
25048 __extension__ extern __inline int8x8_t
25049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25050 vqtbx2_s8 (int8x8_t r, int8x16x2_t tab, uint8x8_t idx)
25051 {
25052   __builtin_aarch64_simd_oi __o;
25053   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
25054   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
25055   return __builtin_aarch64_tbx4v8qi (r, __o, (int8x8_t)idx);
25056 }
25057
25058 __extension__ extern __inline uint8x8_t
25059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25060 vqtbx2_u8 (uint8x8_t r, uint8x16x2_t tab, uint8x8_t idx)
25061 {
25062   __builtin_aarch64_simd_oi __o;
25063   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
25064   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
25065   return (uint8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)r, __o,
25066                                                 (int8x8_t)idx);
25067 }
25068
25069 __extension__ extern __inline poly8x8_t
25070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25071 vqtbx2_p8 (poly8x8_t r, poly8x16x2_t tab, uint8x8_t idx)
25072 {
25073   __builtin_aarch64_simd_oi __o;
25074   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
25075   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
25076   return (poly8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)r, __o,
25077                                                 (int8x8_t)idx);
25078 }
25079
25080 __extension__ extern __inline int8x16_t
25081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25082 vqtbx2q_s8 (int8x16_t r, int8x16x2_t tab, uint8x16_t idx)
25083 {
25084   __builtin_aarch64_simd_oi __o;
25085   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
25086   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
25087   return __builtin_aarch64_tbx4v16qi (r, __o, (int8x16_t)idx);
25088 }
25089
25090 __extension__ extern __inline uint8x16_t
25091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25092 vqtbx2q_u8 (uint8x16_t r, uint8x16x2_t tab, uint8x16_t idx)
25093 {
25094   __builtin_aarch64_simd_oi __o;
25095   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
25096   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
25097   return (uint8x16_t)__builtin_aarch64_tbx4v16qi ((int8x16_t)r, __o,
25098                                                   (int8x16_t)idx);
25099 }
25100
25101 __extension__ extern __inline poly8x16_t
25102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25103 vqtbx2q_p8 (poly8x16_t r, poly8x16x2_t tab, uint8x16_t idx)
25104 {
25105   __builtin_aarch64_simd_oi __o;
25106   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
25107   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
25108   return (poly8x16_t)__builtin_aarch64_tbx4v16qi ((int8x16_t)r, __o,
25109                                                   (int8x16_t)idx);
25110 }
25111
25112 /* vqtbx3 */
25113 __extension__ extern __inline int8x8_t
25114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25115 vqtbx3_s8 (int8x8_t r, int8x16x3_t tab, uint8x8_t idx)
25116 {
25117   __builtin_aarch64_simd_ci __o;
25118   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[0], 0);
25119   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[1], 1);
25120   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[2], 2);
25121   return __builtin_aarch64_qtbx3v8qi (r, __o, (int8x8_t)idx);
25122 }
25123
25124 __extension__ extern __inline uint8x8_t
25125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25126 vqtbx3_u8 (uint8x8_t r, uint8x16x3_t tab, uint8x8_t idx)
25127 {
25128   __builtin_aarch64_simd_ci __o;
25129   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
25130   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
25131   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
25132   return (uint8x8_t)__builtin_aarch64_qtbx3v8qi ((int8x8_t)r, __o,
25133                                                  (int8x8_t)idx);
25134 }
25135
25136 __extension__ extern __inline poly8x8_t
25137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25138 vqtbx3_p8 (poly8x8_t r, poly8x16x3_t tab, uint8x8_t idx)
25139 {
25140   __builtin_aarch64_simd_ci __o;
25141   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
25142   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
25143   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
25144   return (poly8x8_t)__builtin_aarch64_qtbx3v8qi ((int8x8_t)r, __o,
25145                                                  (int8x8_t)idx);
25146 }
25147
25148 __extension__ extern __inline int8x16_t
25149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25150 vqtbx3q_s8 (int8x16_t r, int8x16x3_t tab, uint8x16_t idx)
25151 {
25152   __builtin_aarch64_simd_ci __o;
25153   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[0], 0);
25154   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[1], 1);
25155   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[2], 2);
25156   return __builtin_aarch64_qtbx3v16qi (r, __o, (int8x16_t)idx);
25157 }
25158
25159 __extension__ extern __inline uint8x16_t
25160 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25161 vqtbx3q_u8 (uint8x16_t r, uint8x16x3_t tab, uint8x16_t idx)
25162 {
25163   __builtin_aarch64_simd_ci __o;
25164   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
25165   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
25166   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
25167   return (uint8x16_t)__builtin_aarch64_qtbx3v16qi ((int8x16_t)r, __o,
25168                                                    (int8x16_t)idx);
25169 }
25170
25171 __extension__ extern __inline poly8x16_t
25172 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25173 vqtbx3q_p8 (poly8x16_t r, poly8x16x3_t tab, uint8x16_t idx)
25174 {
25175   __builtin_aarch64_simd_ci __o;
25176   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
25177   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
25178   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
25179   return (poly8x16_t)__builtin_aarch64_qtbx3v16qi ((int8x16_t)r, __o,
25180                                                    (int8x16_t)idx);
25181 }
25182
25183 /* vqtbx4 */
25184
25185 __extension__ extern __inline int8x8_t
25186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25187 vqtbx4_s8 (int8x8_t r, int8x16x4_t tab, uint8x8_t idx)
25188 {
25189   __builtin_aarch64_simd_xi __o;
25190   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[0], 0);
25191   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[1], 1);
25192   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[2], 2);
25193   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[3], 3);
25194   return __builtin_aarch64_qtbx4v8qi (r, __o, (int8x8_t)idx);
25195 }
25196
25197 __extension__ extern __inline uint8x8_t
25198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25199 vqtbx4_u8 (uint8x8_t r, uint8x16x4_t tab, uint8x8_t idx)
25200 {
25201   __builtin_aarch64_simd_xi __o;
25202   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25203   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25204   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25205   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25206   return (uint8x8_t)__builtin_aarch64_qtbx4v8qi ((int8x8_t)r, __o,
25207                                                  (int8x8_t)idx);
25208 }
25209
25210 __extension__ extern __inline poly8x8_t
25211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25212 vqtbx4_p8 (poly8x8_t r, poly8x16x4_t tab, uint8x8_t idx)
25213 {
25214   __builtin_aarch64_simd_xi __o;
25215   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25216   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25217   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25218   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25219   return (poly8x8_t)__builtin_aarch64_qtbx4v8qi ((int8x8_t)r, __o,
25220                                                  (int8x8_t)idx);
25221 }
25222
25223 __extension__ extern __inline int8x16_t
25224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25225 vqtbx4q_s8 (int8x16_t r, int8x16x4_t tab, uint8x16_t idx)
25226 {
25227   __builtin_aarch64_simd_xi __o;
25228   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[0], 0);
25229   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[1], 1);
25230   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[2], 2);
25231   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[3], 3);
25232   return __builtin_aarch64_qtbx4v16qi (r, __o, (int8x16_t)idx);
25233 }
25234
25235 __extension__ extern __inline uint8x16_t
25236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25237 vqtbx4q_u8 (uint8x16_t r, uint8x16x4_t tab, uint8x16_t idx)
25238 {
25239   __builtin_aarch64_simd_xi __o;
25240   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25241   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25242   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25243   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25244   return (uint8x16_t)__builtin_aarch64_qtbx4v16qi ((int8x16_t)r, __o,
25245                                                    (int8x16_t)idx);
25246 }
25247
25248 __extension__ extern __inline poly8x16_t
25249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25250 vqtbx4q_p8 (poly8x16_t r, poly8x16x4_t tab, uint8x16_t idx)
25251 {
25252   __builtin_aarch64_simd_xi __o;
25253   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
25254   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
25255   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
25256   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
25257   return (poly8x16_t)__builtin_aarch64_qtbx4v16qi ((int8x16_t)r, __o,
25258                                                    (int8x16_t)idx);
25259 }
25260
25261 /* vrbit  */
25262
25263 __extension__ extern __inline poly8x8_t
25264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25265 vrbit_p8 (poly8x8_t __a)
25266 {
25267   return (poly8x8_t) __builtin_aarch64_rbitv8qi ((int8x8_t) __a);
25268 }
25269
25270 __extension__ extern __inline int8x8_t
25271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25272 vrbit_s8 (int8x8_t __a)
25273 {
25274   return __builtin_aarch64_rbitv8qi (__a);
25275 }
25276
25277 __extension__ extern __inline uint8x8_t
25278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25279 vrbit_u8 (uint8x8_t __a)
25280 {
25281   return (uint8x8_t) __builtin_aarch64_rbitv8qi ((int8x8_t) __a);
25282 }
25283
25284 __extension__ extern __inline poly8x16_t
25285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25286 vrbitq_p8 (poly8x16_t __a)
25287 {
25288   return (poly8x16_t) __builtin_aarch64_rbitv16qi ((int8x16_t)__a);
25289 }
25290
25291 __extension__ extern __inline int8x16_t
25292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25293 vrbitq_s8 (int8x16_t __a)
25294 {
25295   return __builtin_aarch64_rbitv16qi (__a);
25296 }
25297
25298 __extension__ extern __inline uint8x16_t
25299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25300 vrbitq_u8 (uint8x16_t __a)
25301 {
25302   return (uint8x16_t) __builtin_aarch64_rbitv16qi ((int8x16_t) __a);
25303 }
25304
25305 /* vrecpe  */
25306
25307 __extension__ extern __inline uint32x2_t
25308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25309 vrecpe_u32 (uint32x2_t __a)
25310 {
25311   return (uint32x2_t) __builtin_aarch64_urecpev2si ((int32x2_t) __a);
25312 }
25313
25314 __extension__ extern __inline uint32x4_t
25315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25316 vrecpeq_u32 (uint32x4_t __a)
25317 {
25318   return (uint32x4_t) __builtin_aarch64_urecpev4si ((int32x4_t) __a);
25319 }
25320
25321 __extension__ extern __inline float32_t
25322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25323 vrecpes_f32 (float32_t __a)
25324 {
25325   return __builtin_aarch64_frecpesf (__a);
25326 }
25327
25328 __extension__ extern __inline float64_t
25329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25330 vrecped_f64 (float64_t __a)
25331 {
25332   return __builtin_aarch64_frecpedf (__a);
25333 }
25334
25335 __extension__ extern __inline float32x2_t
25336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25337 vrecpe_f32 (float32x2_t __a)
25338 {
25339   return __builtin_aarch64_frecpev2sf (__a);
25340 }
25341
25342 __extension__ extern __inline float64x1_t
25343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25344 vrecpe_f64 (float64x1_t __a)
25345 {
25346   return (float64x1_t) { vrecped_f64 (vget_lane_f64 (__a, 0)) };
25347 }
25348
25349 __extension__ extern __inline float32x4_t
25350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25351 vrecpeq_f32 (float32x4_t __a)
25352 {
25353   return __builtin_aarch64_frecpev4sf (__a);
25354 }
25355
25356 __extension__ extern __inline float64x2_t
25357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25358 vrecpeq_f64 (float64x2_t __a)
25359 {
25360   return __builtin_aarch64_frecpev2df (__a);
25361 }
25362
25363 /* vrecps  */
25364
25365 __extension__ extern __inline float32_t
25366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25367 vrecpss_f32 (float32_t __a, float32_t __b)
25368 {
25369   return __builtin_aarch64_frecpssf (__a, __b);
25370 }
25371
25372 __extension__ extern __inline float64_t
25373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25374 vrecpsd_f64 (float64_t __a, float64_t __b)
25375 {
25376   return __builtin_aarch64_frecpsdf (__a, __b);
25377 }
25378
25379 __extension__ extern __inline float32x2_t
25380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25381 vrecps_f32 (float32x2_t __a, float32x2_t __b)
25382 {
25383   return __builtin_aarch64_frecpsv2sf (__a, __b);
25384 }
25385
25386 __extension__ extern __inline float64x1_t
25387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25388 vrecps_f64 (float64x1_t __a, float64x1_t __b)
25389 {
25390   return (float64x1_t) { vrecpsd_f64 (vget_lane_f64 (__a, 0),
25391                                       vget_lane_f64 (__b, 0)) };
25392 }
25393
25394 __extension__ extern __inline float32x4_t
25395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25396 vrecpsq_f32 (float32x4_t __a, float32x4_t __b)
25397 {
25398   return __builtin_aarch64_frecpsv4sf (__a, __b);
25399 }
25400
25401 __extension__ extern __inline float64x2_t
25402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25403 vrecpsq_f64 (float64x2_t __a, float64x2_t __b)
25404 {
25405   return __builtin_aarch64_frecpsv2df (__a, __b);
25406 }
25407
25408 /* vrecpx  */
25409
25410 __extension__ extern __inline float32_t
25411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25412 vrecpxs_f32 (float32_t __a)
25413 {
25414   return __builtin_aarch64_frecpxsf (__a);
25415 }
25416
25417 __extension__ extern __inline float64_t
25418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25419 vrecpxd_f64 (float64_t __a)
25420 {
25421   return __builtin_aarch64_frecpxdf (__a);
25422 }
25423
25424
25425 /* vrev  */
25426
25427 __extension__ extern __inline poly8x8_t
25428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25429 vrev16_p8 (poly8x8_t a)
25430 {
25431   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
25432 }
25433
25434 __extension__ extern __inline int8x8_t
25435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25436 vrev16_s8 (int8x8_t a)
25437 {
25438   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
25439 }
25440
25441 __extension__ extern __inline uint8x8_t
25442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25443 vrev16_u8 (uint8x8_t a)
25444 {
25445   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
25446 }
25447
25448 __extension__ extern __inline poly8x16_t
25449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25450 vrev16q_p8 (poly8x16_t a)
25451 {
25452   return __builtin_shuffle (a,
25453       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
25454 }
25455
25456 __extension__ extern __inline int8x16_t
25457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25458 vrev16q_s8 (int8x16_t a)
25459 {
25460   return __builtin_shuffle (a,
25461       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
25462 }
25463
25464 __extension__ extern __inline uint8x16_t
25465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25466 vrev16q_u8 (uint8x16_t a)
25467 {
25468   return __builtin_shuffle (a,
25469       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
25470 }
25471
25472 __extension__ extern __inline poly8x8_t
25473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25474 vrev32_p8 (poly8x8_t a)
25475 {
25476   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25477 }
25478
25479 __extension__ extern __inline poly16x4_t
25480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25481 vrev32_p16 (poly16x4_t a)
25482 {
25483   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
25484 }
25485
25486 __extension__ extern __inline int8x8_t
25487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25488 vrev32_s8 (int8x8_t a)
25489 {
25490   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25491 }
25492
25493 __extension__ extern __inline int16x4_t
25494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25495 vrev32_s16 (int16x4_t a)
25496 {
25497   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
25498 }
25499
25500 __extension__ extern __inline uint8x8_t
25501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25502 vrev32_u8 (uint8x8_t a)
25503 {
25504   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25505 }
25506
25507 __extension__ extern __inline uint16x4_t
25508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25509 vrev32_u16 (uint16x4_t a)
25510 {
25511   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
25512 }
25513
25514 __extension__ extern __inline poly8x16_t
25515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25516 vrev32q_p8 (poly8x16_t a)
25517 {
25518   return __builtin_shuffle (a,
25519       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
25520 }
25521
25522 __extension__ extern __inline poly16x8_t
25523 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25524 vrev32q_p16 (poly16x8_t a)
25525 {
25526   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
25527 }
25528
25529 __extension__ extern __inline int8x16_t
25530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25531 vrev32q_s8 (int8x16_t a)
25532 {
25533   return __builtin_shuffle (a,
25534       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
25535 }
25536
25537 __extension__ extern __inline int16x8_t
25538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25539 vrev32q_s16 (int16x8_t a)
25540 {
25541   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
25542 }
25543
25544 __extension__ extern __inline uint8x16_t
25545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25546 vrev32q_u8 (uint8x16_t a)
25547 {
25548   return __builtin_shuffle (a,
25549       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
25550 }
25551
25552 __extension__ extern __inline uint16x8_t
25553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25554 vrev32q_u16 (uint16x8_t a)
25555 {
25556   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
25557 }
25558
25559 __extension__ extern __inline float16x4_t
25560 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25561 vrev64_f16 (float16x4_t __a)
25562 {
25563   return __builtin_shuffle (__a, (uint16x4_t) { 3, 2, 1, 0 });
25564 }
25565
25566 __extension__ extern __inline float32x2_t
25567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25568 vrev64_f32 (float32x2_t a)
25569 {
25570   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
25571 }
25572
25573 __extension__ extern __inline poly8x8_t
25574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25575 vrev64_p8 (poly8x8_t a)
25576 {
25577   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
25578 }
25579
25580 __extension__ extern __inline poly16x4_t
25581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25582 vrev64_p16 (poly16x4_t a)
25583 {
25584   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
25585 }
25586
25587 __extension__ extern __inline int8x8_t
25588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25589 vrev64_s8 (int8x8_t a)
25590 {
25591   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
25592 }
25593
25594 __extension__ extern __inline int16x4_t
25595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25596 vrev64_s16 (int16x4_t a)
25597 {
25598   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
25599 }
25600
25601 __extension__ extern __inline int32x2_t
25602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25603 vrev64_s32 (int32x2_t a)
25604 {
25605   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
25606 }
25607
25608 __extension__ extern __inline uint8x8_t
25609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25610 vrev64_u8 (uint8x8_t a)
25611 {
25612   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
25613 }
25614
25615 __extension__ extern __inline uint16x4_t
25616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25617 vrev64_u16 (uint16x4_t a)
25618 {
25619   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
25620 }
25621
25622 __extension__ extern __inline uint32x2_t
25623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25624 vrev64_u32 (uint32x2_t a)
25625 {
25626   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
25627 }
25628
25629 __extension__ extern __inline float16x8_t
25630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25631 vrev64q_f16 (float16x8_t __a)
25632 {
25633   return __builtin_shuffle (__a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25634 }
25635
25636 __extension__ extern __inline float32x4_t
25637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25638 vrev64q_f32 (float32x4_t a)
25639 {
25640   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
25641 }
25642
25643 __extension__ extern __inline poly8x16_t
25644 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25645 vrev64q_p8 (poly8x16_t a)
25646 {
25647   return __builtin_shuffle (a,
25648       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
25649 }
25650
25651 __extension__ extern __inline poly16x8_t
25652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25653 vrev64q_p16 (poly16x8_t a)
25654 {
25655   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25656 }
25657
25658 __extension__ extern __inline int8x16_t
25659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25660 vrev64q_s8 (int8x16_t a)
25661 {
25662   return __builtin_shuffle (a,
25663       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
25664 }
25665
25666 __extension__ extern __inline int16x8_t
25667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25668 vrev64q_s16 (int16x8_t a)
25669 {
25670   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25671 }
25672
25673 __extension__ extern __inline int32x4_t
25674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25675 vrev64q_s32 (int32x4_t a)
25676 {
25677   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
25678 }
25679
25680 __extension__ extern __inline uint8x16_t
25681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25682 vrev64q_u8 (uint8x16_t a)
25683 {
25684   return __builtin_shuffle (a,
25685       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
25686 }
25687
25688 __extension__ extern __inline uint16x8_t
25689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25690 vrev64q_u16 (uint16x8_t a)
25691 {
25692   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
25693 }
25694
25695 __extension__ extern __inline uint32x4_t
25696 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25697 vrev64q_u32 (uint32x4_t a)
25698 {
25699   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
25700 }
25701
25702 /* vrnd  */
25703
25704 __extension__ extern __inline float32x2_t
25705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25706 vrnd_f32 (float32x2_t __a)
25707 {
25708   return __builtin_aarch64_btruncv2sf (__a);
25709 }
25710
25711 __extension__ extern __inline float64x1_t
25712 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25713 vrnd_f64 (float64x1_t __a)
25714 {
25715   return vset_lane_f64 (__builtin_trunc (vget_lane_f64 (__a, 0)), __a, 0);
25716 }
25717
25718 __extension__ extern __inline float32x4_t
25719 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25720 vrndq_f32 (float32x4_t __a)
25721 {
25722   return __builtin_aarch64_btruncv4sf (__a);
25723 }
25724
25725 __extension__ extern __inline float64x2_t
25726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25727 vrndq_f64 (float64x2_t __a)
25728 {
25729   return __builtin_aarch64_btruncv2df (__a);
25730 }
25731
25732 /* vrnda  */
25733
25734 __extension__ extern __inline float32x2_t
25735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25736 vrnda_f32 (float32x2_t __a)
25737 {
25738   return __builtin_aarch64_roundv2sf (__a);
25739 }
25740
25741 __extension__ extern __inline float64x1_t
25742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25743 vrnda_f64 (float64x1_t __a)
25744 {
25745   return vset_lane_f64 (__builtin_round (vget_lane_f64 (__a, 0)), __a, 0);
25746 }
25747
25748 __extension__ extern __inline float32x4_t
25749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25750 vrndaq_f32 (float32x4_t __a)
25751 {
25752   return __builtin_aarch64_roundv4sf (__a);
25753 }
25754
25755 __extension__ extern __inline float64x2_t
25756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25757 vrndaq_f64 (float64x2_t __a)
25758 {
25759   return __builtin_aarch64_roundv2df (__a);
25760 }
25761
25762 /* vrndi  */
25763
25764 __extension__ extern __inline float32x2_t
25765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25766 vrndi_f32 (float32x2_t __a)
25767 {
25768   return __builtin_aarch64_nearbyintv2sf (__a);
25769 }
25770
25771 __extension__ extern __inline float64x1_t
25772 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25773 vrndi_f64 (float64x1_t __a)
25774 {
25775   return vset_lane_f64 (__builtin_nearbyint (vget_lane_f64 (__a, 0)), __a, 0);
25776 }
25777
25778 __extension__ extern __inline float32x4_t
25779 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25780 vrndiq_f32 (float32x4_t __a)
25781 {
25782   return __builtin_aarch64_nearbyintv4sf (__a);
25783 }
25784
25785 __extension__ extern __inline float64x2_t
25786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25787 vrndiq_f64 (float64x2_t __a)
25788 {
25789   return __builtin_aarch64_nearbyintv2df (__a);
25790 }
25791
25792 /* vrndm  */
25793
25794 __extension__ extern __inline float32x2_t
25795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25796 vrndm_f32 (float32x2_t __a)
25797 {
25798   return __builtin_aarch64_floorv2sf (__a);
25799 }
25800
25801 __extension__ extern __inline float64x1_t
25802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25803 vrndm_f64 (float64x1_t __a)
25804 {
25805   return vset_lane_f64 (__builtin_floor (vget_lane_f64 (__a, 0)), __a, 0);
25806 }
25807
25808 __extension__ extern __inline float32x4_t
25809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25810 vrndmq_f32 (float32x4_t __a)
25811 {
25812   return __builtin_aarch64_floorv4sf (__a);
25813 }
25814
25815 __extension__ extern __inline float64x2_t
25816 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25817 vrndmq_f64 (float64x2_t __a)
25818 {
25819   return __builtin_aarch64_floorv2df (__a);
25820 }
25821
25822 /* vrndn  */
25823
25824 __extension__ extern __inline float32x2_t
25825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25826 vrndn_f32 (float32x2_t __a)
25827 {
25828   return __builtin_aarch64_frintnv2sf (__a);
25829 }
25830
25831 __extension__ extern __inline float64x1_t
25832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25833 vrndn_f64 (float64x1_t __a)
25834 {
25835   return (float64x1_t) {__builtin_aarch64_frintndf (__a[0])};
25836 }
25837
25838 __extension__ extern __inline float32x4_t
25839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25840 vrndnq_f32 (float32x4_t __a)
25841 {
25842   return __builtin_aarch64_frintnv4sf (__a);
25843 }
25844
25845 __extension__ extern __inline float64x2_t
25846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25847 vrndnq_f64 (float64x2_t __a)
25848 {
25849   return __builtin_aarch64_frintnv2df (__a);
25850 }
25851
25852 /* vrndp  */
25853
25854 __extension__ extern __inline float32x2_t
25855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25856 vrndp_f32 (float32x2_t __a)
25857 {
25858   return __builtin_aarch64_ceilv2sf (__a);
25859 }
25860
25861 __extension__ extern __inline float64x1_t
25862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25863 vrndp_f64 (float64x1_t __a)
25864 {
25865   return vset_lane_f64 (__builtin_ceil (vget_lane_f64 (__a, 0)), __a, 0);
25866 }
25867
25868 __extension__ extern __inline float32x4_t
25869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25870 vrndpq_f32 (float32x4_t __a)
25871 {
25872   return __builtin_aarch64_ceilv4sf (__a);
25873 }
25874
25875 __extension__ extern __inline float64x2_t
25876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25877 vrndpq_f64 (float64x2_t __a)
25878 {
25879   return __builtin_aarch64_ceilv2df (__a);
25880 }
25881
25882 /* vrndx  */
25883
25884 __extension__ extern __inline float32x2_t
25885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25886 vrndx_f32 (float32x2_t __a)
25887 {
25888   return __builtin_aarch64_rintv2sf (__a);
25889 }
25890
25891 __extension__ extern __inline float64x1_t
25892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25893 vrndx_f64 (float64x1_t __a)
25894 {
25895   return vset_lane_f64 (__builtin_rint (vget_lane_f64 (__a, 0)), __a, 0);
25896 }
25897
25898 __extension__ extern __inline float32x4_t
25899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25900 vrndxq_f32 (float32x4_t __a)
25901 {
25902   return __builtin_aarch64_rintv4sf (__a);
25903 }
25904
25905 __extension__ extern __inline float64x2_t
25906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25907 vrndxq_f64 (float64x2_t __a)
25908 {
25909   return __builtin_aarch64_rintv2df (__a);
25910 }
25911
25912 /* vrshl */
25913
25914 __extension__ extern __inline int8x8_t
25915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25916 vrshl_s8 (int8x8_t __a, int8x8_t __b)
25917 {
25918   return (int8x8_t) __builtin_aarch64_srshlv8qi (__a, __b);
25919 }
25920
25921 __extension__ extern __inline int16x4_t
25922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25923 vrshl_s16 (int16x4_t __a, int16x4_t __b)
25924 {
25925   return (int16x4_t) __builtin_aarch64_srshlv4hi (__a, __b);
25926 }
25927
25928 __extension__ extern __inline int32x2_t
25929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25930 vrshl_s32 (int32x2_t __a, int32x2_t __b)
25931 {
25932   return (int32x2_t) __builtin_aarch64_srshlv2si (__a, __b);
25933 }
25934
25935 __extension__ extern __inline int64x1_t
25936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25937 vrshl_s64 (int64x1_t __a, int64x1_t __b)
25938 {
25939   return (int64x1_t) {__builtin_aarch64_srshldi (__a[0], __b[0])};
25940 }
25941
25942 __extension__ extern __inline uint8x8_t
25943 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25944 vrshl_u8 (uint8x8_t __a, int8x8_t __b)
25945 {
25946   return __builtin_aarch64_urshlv8qi_uus (__a, __b);
25947 }
25948
25949 __extension__ extern __inline uint16x4_t
25950 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25951 vrshl_u16 (uint16x4_t __a, int16x4_t __b)
25952 {
25953   return __builtin_aarch64_urshlv4hi_uus (__a, __b);
25954 }
25955
25956 __extension__ extern __inline uint32x2_t
25957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25958 vrshl_u32 (uint32x2_t __a, int32x2_t __b)
25959 {
25960   return __builtin_aarch64_urshlv2si_uus (__a, __b);
25961 }
25962
25963 __extension__ extern __inline uint64x1_t
25964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25965 vrshl_u64 (uint64x1_t __a, int64x1_t __b)
25966 {
25967   return (uint64x1_t) {__builtin_aarch64_urshldi_uus (__a[0], __b[0])};
25968 }
25969
25970 __extension__ extern __inline int8x16_t
25971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25972 vrshlq_s8 (int8x16_t __a, int8x16_t __b)
25973 {
25974   return (int8x16_t) __builtin_aarch64_srshlv16qi (__a, __b);
25975 }
25976
25977 __extension__ extern __inline int16x8_t
25978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25979 vrshlq_s16 (int16x8_t __a, int16x8_t __b)
25980 {
25981   return (int16x8_t) __builtin_aarch64_srshlv8hi (__a, __b);
25982 }
25983
25984 __extension__ extern __inline int32x4_t
25985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25986 vrshlq_s32 (int32x4_t __a, int32x4_t __b)
25987 {
25988   return (int32x4_t) __builtin_aarch64_srshlv4si (__a, __b);
25989 }
25990
25991 __extension__ extern __inline int64x2_t
25992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25993 vrshlq_s64 (int64x2_t __a, int64x2_t __b)
25994 {
25995   return (int64x2_t) __builtin_aarch64_srshlv2di (__a, __b);
25996 }
25997
25998 __extension__ extern __inline uint8x16_t
25999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26000 vrshlq_u8 (uint8x16_t __a, int8x16_t __b)
26001 {
26002   return __builtin_aarch64_urshlv16qi_uus (__a, __b);
26003 }
26004
26005 __extension__ extern __inline uint16x8_t
26006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26007 vrshlq_u16 (uint16x8_t __a, int16x8_t __b)
26008 {
26009   return __builtin_aarch64_urshlv8hi_uus (__a, __b);
26010 }
26011
26012 __extension__ extern __inline uint32x4_t
26013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26014 vrshlq_u32 (uint32x4_t __a, int32x4_t __b)
26015 {
26016   return __builtin_aarch64_urshlv4si_uus (__a, __b);
26017 }
26018
26019 __extension__ extern __inline uint64x2_t
26020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26021 vrshlq_u64 (uint64x2_t __a, int64x2_t __b)
26022 {
26023   return __builtin_aarch64_urshlv2di_uus (__a, __b);
26024 }
26025
26026 __extension__ extern __inline int64_t
26027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26028 vrshld_s64 (int64_t __a, int64_t __b)
26029 {
26030   return __builtin_aarch64_srshldi (__a, __b);
26031 }
26032
26033 __extension__ extern __inline uint64_t
26034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26035 vrshld_u64 (uint64_t __a, int64_t __b)
26036 {
26037   return __builtin_aarch64_urshldi_uus (__a, __b);
26038 }
26039
26040 /* vrshr */
26041
26042 __extension__ extern __inline int8x8_t
26043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26044 vrshr_n_s8 (int8x8_t __a, const int __b)
26045 {
26046   return (int8x8_t) __builtin_aarch64_srshr_nv8qi (__a, __b);
26047 }
26048
26049 __extension__ extern __inline int16x4_t
26050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26051 vrshr_n_s16 (int16x4_t __a, const int __b)
26052 {
26053   return (int16x4_t) __builtin_aarch64_srshr_nv4hi (__a, __b);
26054 }
26055
26056 __extension__ extern __inline int32x2_t
26057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26058 vrshr_n_s32 (int32x2_t __a, const int __b)
26059 {
26060   return (int32x2_t) __builtin_aarch64_srshr_nv2si (__a, __b);
26061 }
26062
26063 __extension__ extern __inline int64x1_t
26064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26065 vrshr_n_s64 (int64x1_t __a, const int __b)
26066 {
26067   return (int64x1_t) {__builtin_aarch64_srshr_ndi (__a[0], __b)};
26068 }
26069
26070 __extension__ extern __inline uint8x8_t
26071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26072 vrshr_n_u8 (uint8x8_t __a, const int __b)
26073 {
26074   return __builtin_aarch64_urshr_nv8qi_uus (__a, __b);
26075 }
26076
26077 __extension__ extern __inline uint16x4_t
26078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26079 vrshr_n_u16 (uint16x4_t __a, const int __b)
26080 {
26081   return __builtin_aarch64_urshr_nv4hi_uus (__a, __b);
26082 }
26083
26084 __extension__ extern __inline uint32x2_t
26085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26086 vrshr_n_u32 (uint32x2_t __a, const int __b)
26087 {
26088   return __builtin_aarch64_urshr_nv2si_uus (__a, __b);
26089 }
26090
26091 __extension__ extern __inline uint64x1_t
26092 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26093 vrshr_n_u64 (uint64x1_t __a, const int __b)
26094 {
26095   return (uint64x1_t) {__builtin_aarch64_urshr_ndi_uus (__a[0], __b)};
26096 }
26097
26098 __extension__ extern __inline int8x16_t
26099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26100 vrshrq_n_s8 (int8x16_t __a, const int __b)
26101 {
26102   return (int8x16_t) __builtin_aarch64_srshr_nv16qi (__a, __b);
26103 }
26104
26105 __extension__ extern __inline int16x8_t
26106 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26107 vrshrq_n_s16 (int16x8_t __a, const int __b)
26108 {
26109   return (int16x8_t) __builtin_aarch64_srshr_nv8hi (__a, __b);
26110 }
26111
26112 __extension__ extern __inline int32x4_t
26113 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26114 vrshrq_n_s32 (int32x4_t __a, const int __b)
26115 {
26116   return (int32x4_t) __builtin_aarch64_srshr_nv4si (__a, __b);
26117 }
26118
26119 __extension__ extern __inline int64x2_t
26120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26121 vrshrq_n_s64 (int64x2_t __a, const int __b)
26122 {
26123   return (int64x2_t) __builtin_aarch64_srshr_nv2di (__a, __b);
26124 }
26125
26126 __extension__ extern __inline uint8x16_t
26127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26128 vrshrq_n_u8 (uint8x16_t __a, const int __b)
26129 {
26130   return __builtin_aarch64_urshr_nv16qi_uus (__a, __b);
26131 }
26132
26133 __extension__ extern __inline uint16x8_t
26134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26135 vrshrq_n_u16 (uint16x8_t __a, const int __b)
26136 {
26137   return __builtin_aarch64_urshr_nv8hi_uus (__a, __b);
26138 }
26139
26140 __extension__ extern __inline uint32x4_t
26141 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26142 vrshrq_n_u32 (uint32x4_t __a, const int __b)
26143 {
26144   return __builtin_aarch64_urshr_nv4si_uus (__a, __b);
26145 }
26146
26147 __extension__ extern __inline uint64x2_t
26148 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26149 vrshrq_n_u64 (uint64x2_t __a, const int __b)
26150 {
26151   return __builtin_aarch64_urshr_nv2di_uus (__a, __b);
26152 }
26153
26154 __extension__ extern __inline int64_t
26155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26156 vrshrd_n_s64 (int64_t __a, const int __b)
26157 {
26158   return __builtin_aarch64_srshr_ndi (__a, __b);
26159 }
26160
26161 __extension__ extern __inline uint64_t
26162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26163 vrshrd_n_u64 (uint64_t __a, const int __b)
26164 {
26165   return __builtin_aarch64_urshr_ndi_uus (__a, __b);
26166 }
26167
26168 /* vrsqrte.  */
26169
26170 __extension__ extern __inline float32_t
26171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26172 vrsqrtes_f32 (float32_t __a)
26173 {
26174   return __builtin_aarch64_rsqrtesf (__a);
26175 }
26176
26177 __extension__ extern __inline float64_t
26178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26179 vrsqrted_f64 (float64_t __a)
26180 {
26181   return __builtin_aarch64_rsqrtedf (__a);
26182 }
26183
26184 __extension__ extern __inline float32x2_t
26185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26186 vrsqrte_f32 (float32x2_t __a)
26187 {
26188   return __builtin_aarch64_rsqrtev2sf (__a);
26189 }
26190
26191 __extension__ extern __inline float64x1_t
26192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26193 vrsqrte_f64 (float64x1_t __a)
26194 {
26195   return (float64x1_t) {vrsqrted_f64 (vget_lane_f64 (__a, 0))};
26196 }
26197
26198 __extension__ extern __inline float32x4_t
26199 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26200 vrsqrteq_f32 (float32x4_t __a)
26201 {
26202   return __builtin_aarch64_rsqrtev4sf (__a);
26203 }
26204
26205 __extension__ extern __inline float64x2_t
26206 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26207 vrsqrteq_f64 (float64x2_t __a)
26208 {
26209   return __builtin_aarch64_rsqrtev2df (__a);
26210 }
26211
26212 /* vrsqrts.  */
26213
26214 __extension__ extern __inline float32_t
26215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26216 vrsqrtss_f32 (float32_t __a, float32_t __b)
26217 {
26218   return __builtin_aarch64_rsqrtssf (__a, __b);
26219 }
26220
26221 __extension__ extern __inline float64_t
26222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26223 vrsqrtsd_f64 (float64_t __a, float64_t __b)
26224 {
26225   return __builtin_aarch64_rsqrtsdf (__a, __b);
26226 }
26227
26228 __extension__ extern __inline float32x2_t
26229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26230 vrsqrts_f32 (float32x2_t __a, float32x2_t __b)
26231 {
26232   return __builtin_aarch64_rsqrtsv2sf (__a, __b);
26233 }
26234
26235 __extension__ extern __inline float64x1_t
26236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26237 vrsqrts_f64 (float64x1_t __a, float64x1_t __b)
26238 {
26239   return (float64x1_t) {vrsqrtsd_f64 (vget_lane_f64 (__a, 0),
26240                                       vget_lane_f64 (__b, 0))};
26241 }
26242
26243 __extension__ extern __inline float32x4_t
26244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26245 vrsqrtsq_f32 (float32x4_t __a, float32x4_t __b)
26246 {
26247   return __builtin_aarch64_rsqrtsv4sf (__a, __b);
26248 }
26249
26250 __extension__ extern __inline float64x2_t
26251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26252 vrsqrtsq_f64 (float64x2_t __a, float64x2_t __b)
26253 {
26254   return __builtin_aarch64_rsqrtsv2df (__a, __b);
26255 }
26256
26257 /* vrsra */
26258
26259 __extension__ extern __inline int8x8_t
26260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26261 vrsra_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26262 {
26263   return (int8x8_t) __builtin_aarch64_srsra_nv8qi (__a, __b, __c);
26264 }
26265
26266 __extension__ extern __inline int16x4_t
26267 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26268 vrsra_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26269 {
26270   return (int16x4_t) __builtin_aarch64_srsra_nv4hi (__a, __b, __c);
26271 }
26272
26273 __extension__ extern __inline int32x2_t
26274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26275 vrsra_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26276 {
26277   return (int32x2_t) __builtin_aarch64_srsra_nv2si (__a, __b, __c);
26278 }
26279
26280 __extension__ extern __inline int64x1_t
26281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26282 vrsra_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26283 {
26284   return (int64x1_t) {__builtin_aarch64_srsra_ndi (__a[0], __b[0], __c)};
26285 }
26286
26287 __extension__ extern __inline uint8x8_t
26288 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26289 vrsra_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26290 {
26291   return __builtin_aarch64_ursra_nv8qi_uuus (__a, __b, __c);
26292 }
26293
26294 __extension__ extern __inline uint16x4_t
26295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26296 vrsra_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26297 {
26298   return __builtin_aarch64_ursra_nv4hi_uuus (__a, __b, __c);
26299 }
26300
26301 __extension__ extern __inline uint32x2_t
26302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26303 vrsra_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26304 {
26305   return __builtin_aarch64_ursra_nv2si_uuus (__a, __b, __c);
26306 }
26307
26308 __extension__ extern __inline uint64x1_t
26309 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26310 vrsra_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26311 {
26312   return (uint64x1_t) {__builtin_aarch64_ursra_ndi_uuus (__a[0], __b[0], __c)};
26313 }
26314
26315 __extension__ extern __inline int8x16_t
26316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26317 vrsraq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26318 {
26319   return (int8x16_t) __builtin_aarch64_srsra_nv16qi (__a, __b, __c);
26320 }
26321
26322 __extension__ extern __inline int16x8_t
26323 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26324 vrsraq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26325 {
26326   return (int16x8_t) __builtin_aarch64_srsra_nv8hi (__a, __b, __c);
26327 }
26328
26329 __extension__ extern __inline int32x4_t
26330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26331 vrsraq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26332 {
26333   return (int32x4_t) __builtin_aarch64_srsra_nv4si (__a, __b, __c);
26334 }
26335
26336 __extension__ extern __inline int64x2_t
26337 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26338 vrsraq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26339 {
26340   return (int64x2_t) __builtin_aarch64_srsra_nv2di (__a, __b, __c);
26341 }
26342
26343 __extension__ extern __inline uint8x16_t
26344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26345 vrsraq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26346 {
26347   return __builtin_aarch64_ursra_nv16qi_uuus (__a, __b, __c);
26348 }
26349
26350 __extension__ extern __inline uint16x8_t
26351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26352 vrsraq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26353 {
26354   return __builtin_aarch64_ursra_nv8hi_uuus (__a, __b, __c);
26355 }
26356
26357 __extension__ extern __inline uint32x4_t
26358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26359 vrsraq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26360 {
26361   return __builtin_aarch64_ursra_nv4si_uuus (__a, __b, __c);
26362 }
26363
26364 __extension__ extern __inline uint64x2_t
26365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26366 vrsraq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26367 {
26368   return __builtin_aarch64_ursra_nv2di_uuus (__a, __b, __c);
26369 }
26370
26371 __extension__ extern __inline int64_t
26372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26373 vrsrad_n_s64 (int64_t __a, int64_t __b, const int __c)
26374 {
26375   return __builtin_aarch64_srsra_ndi (__a, __b, __c);
26376 }
26377
26378 __extension__ extern __inline uint64_t
26379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26380 vrsrad_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26381 {
26382   return __builtin_aarch64_ursra_ndi_uuus (__a, __b, __c);
26383 }
26384
26385 #pragma GCC push_options
26386 #pragma GCC target ("+nothing+crypto")
26387
26388 /* vsha1  */
26389
26390 __extension__ extern __inline uint32x4_t
26391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26392 vsha1cq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
26393 {
26394   return __builtin_aarch64_crypto_sha1cv4si_uuuu (hash_abcd, hash_e, wk);
26395 }
26396
26397 __extension__ extern __inline uint32x4_t
26398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26399 vsha1mq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
26400 {
26401   return __builtin_aarch64_crypto_sha1mv4si_uuuu (hash_abcd, hash_e, wk);
26402 }
26403
26404 __extension__ extern __inline uint32x4_t
26405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26406 vsha1pq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
26407 {
26408   return __builtin_aarch64_crypto_sha1pv4si_uuuu (hash_abcd, hash_e, wk);
26409 }
26410
26411 __extension__ extern __inline uint32_t
26412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26413 vsha1h_u32 (uint32_t hash_e)
26414 {
26415   return __builtin_aarch64_crypto_sha1hsi_uu (hash_e);
26416 }
26417
26418 __extension__ extern __inline uint32x4_t
26419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26420 vsha1su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7, uint32x4_t w8_11)
26421 {
26422   return __builtin_aarch64_crypto_sha1su0v4si_uuuu (w0_3, w4_7, w8_11);
26423 }
26424
26425 __extension__ extern __inline uint32x4_t
26426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26427 vsha1su1q_u32 (uint32x4_t tw0_3, uint32x4_t w12_15)
26428 {
26429   return __builtin_aarch64_crypto_sha1su1v4si_uuu (tw0_3, w12_15);
26430 }
26431
26432 __extension__ extern __inline uint32x4_t
26433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26434 vsha256hq_u32 (uint32x4_t hash_abcd, uint32x4_t hash_efgh, uint32x4_t wk)
26435 {
26436   return __builtin_aarch64_crypto_sha256hv4si_uuuu (hash_abcd, hash_efgh, wk);
26437 }
26438
26439 __extension__ extern __inline uint32x4_t
26440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26441 vsha256h2q_u32 (uint32x4_t hash_efgh, uint32x4_t hash_abcd, uint32x4_t wk)
26442 {
26443   return __builtin_aarch64_crypto_sha256h2v4si_uuuu (hash_efgh, hash_abcd, wk);
26444 }
26445
26446 __extension__ extern __inline uint32x4_t
26447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26448 vsha256su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7)
26449 {
26450   return __builtin_aarch64_crypto_sha256su0v4si_uuu (w0_3, w4_7);
26451 }
26452
26453 __extension__ extern __inline uint32x4_t
26454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26455 vsha256su1q_u32 (uint32x4_t tw0_3, uint32x4_t w8_11, uint32x4_t w12_15)
26456 {
26457   return __builtin_aarch64_crypto_sha256su1v4si_uuuu (tw0_3, w8_11, w12_15);
26458 }
26459
26460 __extension__ extern __inline poly128_t
26461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26462 vmull_p64 (poly64_t a, poly64_t b)
26463 {
26464   return
26465     __builtin_aarch64_crypto_pmulldi_ppp (a, b);
26466 }
26467
26468 __extension__ extern __inline poly128_t
26469 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26470 vmull_high_p64 (poly64x2_t a, poly64x2_t b)
26471 {
26472   return __builtin_aarch64_crypto_pmullv2di_ppp (a, b);
26473 }
26474
26475 #pragma GCC pop_options
26476
26477 /* vshl */
26478
26479 __extension__ extern __inline int8x8_t
26480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26481 vshl_n_s8 (int8x8_t __a, const int __b)
26482 {
26483   return (int8x8_t) __builtin_aarch64_ashlv8qi (__a, __b);
26484 }
26485
26486 __extension__ extern __inline int16x4_t
26487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26488 vshl_n_s16 (int16x4_t __a, const int __b)
26489 {
26490   return (int16x4_t) __builtin_aarch64_ashlv4hi (__a, __b);
26491 }
26492
26493 __extension__ extern __inline int32x2_t
26494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26495 vshl_n_s32 (int32x2_t __a, const int __b)
26496 {
26497   return (int32x2_t) __builtin_aarch64_ashlv2si (__a, __b);
26498 }
26499
26500 __extension__ extern __inline int64x1_t
26501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26502 vshl_n_s64 (int64x1_t __a, const int __b)
26503 {
26504   return (int64x1_t) {__builtin_aarch64_ashldi (__a[0], __b)};
26505 }
26506
26507 __extension__ extern __inline uint8x8_t
26508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26509 vshl_n_u8 (uint8x8_t __a, const int __b)
26510 {
26511   return (uint8x8_t) __builtin_aarch64_ashlv8qi ((int8x8_t) __a, __b);
26512 }
26513
26514 __extension__ extern __inline uint16x4_t
26515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26516 vshl_n_u16 (uint16x4_t __a, const int __b)
26517 {
26518   return (uint16x4_t) __builtin_aarch64_ashlv4hi ((int16x4_t) __a, __b);
26519 }
26520
26521 __extension__ extern __inline uint32x2_t
26522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26523 vshl_n_u32 (uint32x2_t __a, const int __b)
26524 {
26525   return (uint32x2_t) __builtin_aarch64_ashlv2si ((int32x2_t) __a, __b);
26526 }
26527
26528 __extension__ extern __inline uint64x1_t
26529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26530 vshl_n_u64 (uint64x1_t __a, const int __b)
26531 {
26532   return (uint64x1_t) {__builtin_aarch64_ashldi ((int64_t) __a[0], __b)};
26533 }
26534
26535 __extension__ extern __inline int8x16_t
26536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26537 vshlq_n_s8 (int8x16_t __a, const int __b)
26538 {
26539   return (int8x16_t) __builtin_aarch64_ashlv16qi (__a, __b);
26540 }
26541
26542 __extension__ extern __inline int16x8_t
26543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26544 vshlq_n_s16 (int16x8_t __a, const int __b)
26545 {
26546   return (int16x8_t) __builtin_aarch64_ashlv8hi (__a, __b);
26547 }
26548
26549 __extension__ extern __inline int32x4_t
26550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26551 vshlq_n_s32 (int32x4_t __a, const int __b)
26552 {
26553   return (int32x4_t) __builtin_aarch64_ashlv4si (__a, __b);
26554 }
26555
26556 __extension__ extern __inline int64x2_t
26557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26558 vshlq_n_s64 (int64x2_t __a, const int __b)
26559 {
26560   return (int64x2_t) __builtin_aarch64_ashlv2di (__a, __b);
26561 }
26562
26563 __extension__ extern __inline uint8x16_t
26564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26565 vshlq_n_u8 (uint8x16_t __a, const int __b)
26566 {
26567   return (uint8x16_t) __builtin_aarch64_ashlv16qi ((int8x16_t) __a, __b);
26568 }
26569
26570 __extension__ extern __inline uint16x8_t
26571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26572 vshlq_n_u16 (uint16x8_t __a, const int __b)
26573 {
26574   return (uint16x8_t) __builtin_aarch64_ashlv8hi ((int16x8_t) __a, __b);
26575 }
26576
26577 __extension__ extern __inline uint32x4_t
26578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26579 vshlq_n_u32 (uint32x4_t __a, const int __b)
26580 {
26581   return (uint32x4_t) __builtin_aarch64_ashlv4si ((int32x4_t) __a, __b);
26582 }
26583
26584 __extension__ extern __inline uint64x2_t
26585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26586 vshlq_n_u64 (uint64x2_t __a, const int __b)
26587 {
26588   return (uint64x2_t) __builtin_aarch64_ashlv2di ((int64x2_t) __a, __b);
26589 }
26590
26591 __extension__ extern __inline int64_t
26592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26593 vshld_n_s64 (int64_t __a, const int __b)
26594 {
26595   return __builtin_aarch64_ashldi (__a, __b);
26596 }
26597
26598 __extension__ extern __inline uint64_t
26599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26600 vshld_n_u64 (uint64_t __a, const int __b)
26601 {
26602   return (uint64_t) __builtin_aarch64_ashldi (__a, __b);
26603 }
26604
26605 __extension__ extern __inline int8x8_t
26606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26607 vshl_s8 (int8x8_t __a, int8x8_t __b)
26608 {
26609   return __builtin_aarch64_sshlv8qi (__a, __b);
26610 }
26611
26612 __extension__ extern __inline int16x4_t
26613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26614 vshl_s16 (int16x4_t __a, int16x4_t __b)
26615 {
26616   return __builtin_aarch64_sshlv4hi (__a, __b);
26617 }
26618
26619 __extension__ extern __inline int32x2_t
26620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26621 vshl_s32 (int32x2_t __a, int32x2_t __b)
26622 {
26623   return __builtin_aarch64_sshlv2si (__a, __b);
26624 }
26625
26626 __extension__ extern __inline int64x1_t
26627 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26628 vshl_s64 (int64x1_t __a, int64x1_t __b)
26629 {
26630   return (int64x1_t) {__builtin_aarch64_sshldi (__a[0], __b[0])};
26631 }
26632
26633 __extension__ extern __inline uint8x8_t
26634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26635 vshl_u8 (uint8x8_t __a, int8x8_t __b)
26636 {
26637   return __builtin_aarch64_ushlv8qi_uus (__a, __b);
26638 }
26639
26640 __extension__ extern __inline uint16x4_t
26641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26642 vshl_u16 (uint16x4_t __a, int16x4_t __b)
26643 {
26644   return __builtin_aarch64_ushlv4hi_uus (__a, __b);
26645 }
26646
26647 __extension__ extern __inline uint32x2_t
26648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26649 vshl_u32 (uint32x2_t __a, int32x2_t __b)
26650 {
26651   return __builtin_aarch64_ushlv2si_uus (__a, __b);
26652 }
26653
26654 __extension__ extern __inline uint64x1_t
26655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26656 vshl_u64 (uint64x1_t __a, int64x1_t __b)
26657 {
26658   return (uint64x1_t) {__builtin_aarch64_ushldi_uus (__a[0], __b[0])};
26659 }
26660
26661 __extension__ extern __inline int8x16_t
26662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26663 vshlq_s8 (int8x16_t __a, int8x16_t __b)
26664 {
26665   return __builtin_aarch64_sshlv16qi (__a, __b);
26666 }
26667
26668 __extension__ extern __inline int16x8_t
26669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26670 vshlq_s16 (int16x8_t __a, int16x8_t __b)
26671 {
26672   return __builtin_aarch64_sshlv8hi (__a, __b);
26673 }
26674
26675 __extension__ extern __inline int32x4_t
26676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26677 vshlq_s32 (int32x4_t __a, int32x4_t __b)
26678 {
26679   return __builtin_aarch64_sshlv4si (__a, __b);
26680 }
26681
26682 __extension__ extern __inline int64x2_t
26683 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26684 vshlq_s64 (int64x2_t __a, int64x2_t __b)
26685 {
26686   return __builtin_aarch64_sshlv2di (__a, __b);
26687 }
26688
26689 __extension__ extern __inline uint8x16_t
26690 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26691 vshlq_u8 (uint8x16_t __a, int8x16_t __b)
26692 {
26693   return __builtin_aarch64_ushlv16qi_uus (__a, __b);
26694 }
26695
26696 __extension__ extern __inline uint16x8_t
26697 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26698 vshlq_u16 (uint16x8_t __a, int16x8_t __b)
26699 {
26700   return __builtin_aarch64_ushlv8hi_uus (__a, __b);
26701 }
26702
26703 __extension__ extern __inline uint32x4_t
26704 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26705 vshlq_u32 (uint32x4_t __a, int32x4_t __b)
26706 {
26707   return __builtin_aarch64_ushlv4si_uus (__a, __b);
26708 }
26709
26710 __extension__ extern __inline uint64x2_t
26711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26712 vshlq_u64 (uint64x2_t __a, int64x2_t __b)
26713 {
26714   return __builtin_aarch64_ushlv2di_uus (__a, __b);
26715 }
26716
26717 __extension__ extern __inline int64_t
26718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26719 vshld_s64 (int64_t __a, int64_t __b)
26720 {
26721   return __builtin_aarch64_sshldi (__a, __b);
26722 }
26723
26724 __extension__ extern __inline uint64_t
26725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26726 vshld_u64 (uint64_t __a, uint64_t __b)
26727 {
26728   return __builtin_aarch64_ushldi_uus (__a, __b);
26729 }
26730
26731 __extension__ extern __inline int16x8_t
26732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26733 vshll_high_n_s8 (int8x16_t __a, const int __b)
26734 {
26735   return __builtin_aarch64_sshll2_nv16qi (__a, __b);
26736 }
26737
26738 __extension__ extern __inline int32x4_t
26739 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26740 vshll_high_n_s16 (int16x8_t __a, const int __b)
26741 {
26742   return __builtin_aarch64_sshll2_nv8hi (__a, __b);
26743 }
26744
26745 __extension__ extern __inline int64x2_t
26746 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26747 vshll_high_n_s32 (int32x4_t __a, const int __b)
26748 {
26749   return __builtin_aarch64_sshll2_nv4si (__a, __b);
26750 }
26751
26752 __extension__ extern __inline uint16x8_t
26753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26754 vshll_high_n_u8 (uint8x16_t __a, const int __b)
26755 {
26756   return (uint16x8_t) __builtin_aarch64_ushll2_nv16qi ((int8x16_t) __a, __b);
26757 }
26758
26759 __extension__ extern __inline uint32x4_t
26760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26761 vshll_high_n_u16 (uint16x8_t __a, const int __b)
26762 {
26763   return (uint32x4_t) __builtin_aarch64_ushll2_nv8hi ((int16x8_t) __a, __b);
26764 }
26765
26766 __extension__ extern __inline uint64x2_t
26767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26768 vshll_high_n_u32 (uint32x4_t __a, const int __b)
26769 {
26770   return (uint64x2_t) __builtin_aarch64_ushll2_nv4si ((int32x4_t) __a, __b);
26771 }
26772
26773 __extension__ extern __inline int16x8_t
26774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26775 vshll_n_s8 (int8x8_t __a, const int __b)
26776 {
26777   return __builtin_aarch64_sshll_nv8qi (__a, __b);
26778 }
26779
26780 __extension__ extern __inline int32x4_t
26781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26782 vshll_n_s16 (int16x4_t __a, const int __b)
26783 {
26784   return __builtin_aarch64_sshll_nv4hi (__a, __b);
26785 }
26786
26787 __extension__ extern __inline int64x2_t
26788 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26789 vshll_n_s32 (int32x2_t __a, const int __b)
26790 {
26791   return __builtin_aarch64_sshll_nv2si (__a, __b);
26792 }
26793
26794 __extension__ extern __inline uint16x8_t
26795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26796 vshll_n_u8 (uint8x8_t __a, const int __b)
26797 {
26798   return __builtin_aarch64_ushll_nv8qi_uus (__a, __b);
26799 }
26800
26801 __extension__ extern __inline uint32x4_t
26802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26803 vshll_n_u16 (uint16x4_t __a, const int __b)
26804 {
26805   return __builtin_aarch64_ushll_nv4hi_uus (__a, __b);
26806 }
26807
26808 __extension__ extern __inline uint64x2_t
26809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26810 vshll_n_u32 (uint32x2_t __a, const int __b)
26811 {
26812   return __builtin_aarch64_ushll_nv2si_uus (__a, __b);
26813 }
26814
26815 /* vshr */
26816
26817 __extension__ extern __inline int8x8_t
26818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26819 vshr_n_s8 (int8x8_t __a, const int __b)
26820 {
26821   return (int8x8_t) __builtin_aarch64_ashrv8qi (__a, __b);
26822 }
26823
26824 __extension__ extern __inline int16x4_t
26825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26826 vshr_n_s16 (int16x4_t __a, const int __b)
26827 {
26828   return (int16x4_t) __builtin_aarch64_ashrv4hi (__a, __b);
26829 }
26830
26831 __extension__ extern __inline int32x2_t
26832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26833 vshr_n_s32 (int32x2_t __a, const int __b)
26834 {
26835   return (int32x2_t) __builtin_aarch64_ashrv2si (__a, __b);
26836 }
26837
26838 __extension__ extern __inline int64x1_t
26839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26840 vshr_n_s64 (int64x1_t __a, const int __b)
26841 {
26842   return (int64x1_t) {__builtin_aarch64_ashr_simddi (__a[0], __b)};
26843 }
26844
26845 __extension__ extern __inline uint8x8_t
26846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26847 vshr_n_u8 (uint8x8_t __a, const int __b)
26848 {
26849   return (uint8x8_t) __builtin_aarch64_lshrv8qi ((int8x8_t) __a, __b);
26850 }
26851
26852 __extension__ extern __inline uint16x4_t
26853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26854 vshr_n_u16 (uint16x4_t __a, const int __b)
26855 {
26856   return (uint16x4_t) __builtin_aarch64_lshrv4hi ((int16x4_t) __a, __b);
26857 }
26858
26859 __extension__ extern __inline uint32x2_t
26860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26861 vshr_n_u32 (uint32x2_t __a, const int __b)
26862 {
26863   return (uint32x2_t) __builtin_aarch64_lshrv2si ((int32x2_t) __a, __b);
26864 }
26865
26866 __extension__ extern __inline uint64x1_t
26867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26868 vshr_n_u64 (uint64x1_t __a, const int __b)
26869 {
26870   return (uint64x1_t) {__builtin_aarch64_lshr_simddi_uus ( __a[0], __b)};
26871 }
26872
26873 __extension__ extern __inline int8x16_t
26874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26875 vshrq_n_s8 (int8x16_t __a, const int __b)
26876 {
26877   return (int8x16_t) __builtin_aarch64_ashrv16qi (__a, __b);
26878 }
26879
26880 __extension__ extern __inline int16x8_t
26881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26882 vshrq_n_s16 (int16x8_t __a, const int __b)
26883 {
26884   return (int16x8_t) __builtin_aarch64_ashrv8hi (__a, __b);
26885 }
26886
26887 __extension__ extern __inline int32x4_t
26888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26889 vshrq_n_s32 (int32x4_t __a, const int __b)
26890 {
26891   return (int32x4_t) __builtin_aarch64_ashrv4si (__a, __b);
26892 }
26893
26894 __extension__ extern __inline int64x2_t
26895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26896 vshrq_n_s64 (int64x2_t __a, const int __b)
26897 {
26898   return (int64x2_t) __builtin_aarch64_ashrv2di (__a, __b);
26899 }
26900
26901 __extension__ extern __inline uint8x16_t
26902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26903 vshrq_n_u8 (uint8x16_t __a, const int __b)
26904 {
26905   return (uint8x16_t) __builtin_aarch64_lshrv16qi ((int8x16_t) __a, __b);
26906 }
26907
26908 __extension__ extern __inline uint16x8_t
26909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26910 vshrq_n_u16 (uint16x8_t __a, const int __b)
26911 {
26912   return (uint16x8_t) __builtin_aarch64_lshrv8hi ((int16x8_t) __a, __b);
26913 }
26914
26915 __extension__ extern __inline uint32x4_t
26916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26917 vshrq_n_u32 (uint32x4_t __a, const int __b)
26918 {
26919   return (uint32x4_t) __builtin_aarch64_lshrv4si ((int32x4_t) __a, __b);
26920 }
26921
26922 __extension__ extern __inline uint64x2_t
26923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26924 vshrq_n_u64 (uint64x2_t __a, const int __b)
26925 {
26926   return (uint64x2_t) __builtin_aarch64_lshrv2di ((int64x2_t) __a, __b);
26927 }
26928
26929 __extension__ extern __inline int64_t
26930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26931 vshrd_n_s64 (int64_t __a, const int __b)
26932 {
26933   return __builtin_aarch64_ashr_simddi (__a, __b);
26934 }
26935
26936 __extension__ extern __inline uint64_t
26937 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26938 vshrd_n_u64 (uint64_t __a, const int __b)
26939 {
26940   return __builtin_aarch64_lshr_simddi_uus (__a, __b);
26941 }
26942
26943 /* vsli */
26944
26945 __extension__ extern __inline int8x8_t
26946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26947 vsli_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26948 {
26949   return (int8x8_t) __builtin_aarch64_ssli_nv8qi (__a, __b, __c);
26950 }
26951
26952 __extension__ extern __inline int16x4_t
26953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26954 vsli_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26955 {
26956   return (int16x4_t) __builtin_aarch64_ssli_nv4hi (__a, __b, __c);
26957 }
26958
26959 __extension__ extern __inline int32x2_t
26960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26961 vsli_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26962 {
26963   return (int32x2_t) __builtin_aarch64_ssli_nv2si (__a, __b, __c);
26964 }
26965
26966 __extension__ extern __inline int64x1_t
26967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26968 vsli_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26969 {
26970   return (int64x1_t) {__builtin_aarch64_ssli_ndi (__a[0], __b[0], __c)};
26971 }
26972
26973 __extension__ extern __inline uint8x8_t
26974 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26975 vsli_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26976 {
26977   return __builtin_aarch64_usli_nv8qi_uuus (__a, __b, __c);
26978 }
26979
26980 __extension__ extern __inline uint16x4_t
26981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26982 vsli_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26983 {
26984   return __builtin_aarch64_usli_nv4hi_uuus (__a, __b, __c);
26985 }
26986
26987 __extension__ extern __inline uint32x2_t
26988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26989 vsli_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26990 {
26991   return __builtin_aarch64_usli_nv2si_uuus (__a, __b, __c);
26992 }
26993
26994 __extension__ extern __inline uint64x1_t
26995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26996 vsli_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26997 {
26998   return (uint64x1_t) {__builtin_aarch64_usli_ndi_uuus (__a[0], __b[0], __c)};
26999 }
27000
27001 __extension__ extern __inline poly64x1_t
27002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27003 vsli_n_p64 (poly64x1_t __a, poly64x1_t __b, const int __c)
27004 {
27005   return (poly64x1_t) {__builtin_aarch64_ssli_ndi_ppps (__a[0], __b[0], __c)};
27006 }
27007
27008 __extension__ extern __inline int8x16_t
27009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27010 vsliq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
27011 {
27012   return (int8x16_t) __builtin_aarch64_ssli_nv16qi (__a, __b, __c);
27013 }
27014
27015 __extension__ extern __inline int16x8_t
27016 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27017 vsliq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
27018 {
27019   return (int16x8_t) __builtin_aarch64_ssli_nv8hi (__a, __b, __c);
27020 }
27021
27022 __extension__ extern __inline int32x4_t
27023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27024 vsliq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
27025 {
27026   return (int32x4_t) __builtin_aarch64_ssli_nv4si (__a, __b, __c);
27027 }
27028
27029 __extension__ extern __inline int64x2_t
27030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27031 vsliq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
27032 {
27033   return (int64x2_t) __builtin_aarch64_ssli_nv2di (__a, __b, __c);
27034 }
27035
27036 __extension__ extern __inline uint8x16_t
27037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27038 vsliq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
27039 {
27040   return __builtin_aarch64_usli_nv16qi_uuus (__a, __b, __c);
27041 }
27042
27043 __extension__ extern __inline uint16x8_t
27044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27045 vsliq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
27046 {
27047   return __builtin_aarch64_usli_nv8hi_uuus (__a, __b, __c);
27048 }
27049
27050 __extension__ extern __inline uint32x4_t
27051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27052 vsliq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
27053 {
27054   return __builtin_aarch64_usli_nv4si_uuus (__a, __b, __c);
27055 }
27056
27057 __extension__ extern __inline uint64x2_t
27058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27059 vsliq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
27060 {
27061   return __builtin_aarch64_usli_nv2di_uuus (__a, __b, __c);
27062 }
27063
27064 __extension__ extern __inline poly64x2_t
27065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27066 vsliq_n_p64 (poly64x2_t __a, poly64x2_t __b, const int __c)
27067 {
27068   return __builtin_aarch64_ssli_nv2di_ppps (__a, __b, __c);
27069 }
27070
27071 __extension__ extern __inline int64_t
27072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27073 vslid_n_s64 (int64_t __a, int64_t __b, const int __c)
27074 {
27075   return __builtin_aarch64_ssli_ndi (__a, __b, __c);
27076 }
27077
27078 __extension__ extern __inline uint64_t
27079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27080 vslid_n_u64 (uint64_t __a, uint64_t __b, const int __c)
27081 {
27082   return __builtin_aarch64_usli_ndi_uuus (__a, __b, __c);
27083 }
27084
27085 /* vsqadd */
27086
27087 __extension__ extern __inline uint8x8_t
27088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27089 vsqadd_u8 (uint8x8_t __a, int8x8_t __b)
27090 {
27091   return __builtin_aarch64_usqaddv8qi_uus (__a, __b);
27092 }
27093
27094 __extension__ extern __inline uint16x4_t
27095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27096 vsqadd_u16 (uint16x4_t __a, int16x4_t __b)
27097 {
27098   return __builtin_aarch64_usqaddv4hi_uus (__a, __b);
27099 }
27100
27101 __extension__ extern __inline uint32x2_t
27102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27103 vsqadd_u32 (uint32x2_t __a, int32x2_t __b)
27104 {
27105   return __builtin_aarch64_usqaddv2si_uus (__a, __b);
27106 }
27107
27108 __extension__ extern __inline uint64x1_t
27109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27110 vsqadd_u64 (uint64x1_t __a, int64x1_t __b)
27111 {
27112   return (uint64x1_t) {__builtin_aarch64_usqadddi_uus (__a[0], __b[0])};
27113 }
27114
27115 __extension__ extern __inline uint8x16_t
27116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27117 vsqaddq_u8 (uint8x16_t __a, int8x16_t __b)
27118 {
27119   return __builtin_aarch64_usqaddv16qi_uus (__a, __b);
27120 }
27121
27122 __extension__ extern __inline uint16x8_t
27123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27124 vsqaddq_u16 (uint16x8_t __a, int16x8_t __b)
27125 {
27126   return __builtin_aarch64_usqaddv8hi_uus (__a, __b);
27127 }
27128
27129 __extension__ extern __inline uint32x4_t
27130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27131 vsqaddq_u32 (uint32x4_t __a, int32x4_t __b)
27132 {
27133   return __builtin_aarch64_usqaddv4si_uus (__a, __b);
27134 }
27135
27136 __extension__ extern __inline uint64x2_t
27137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27138 vsqaddq_u64 (uint64x2_t __a, int64x2_t __b)
27139 {
27140   return __builtin_aarch64_usqaddv2di_uus (__a, __b);
27141 }
27142
27143 __extension__ extern __inline uint8_t
27144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27145 vsqaddb_u8 (uint8_t __a, int8_t __b)
27146 {
27147   return __builtin_aarch64_usqaddqi_uus (__a, __b);
27148 }
27149
27150 __extension__ extern __inline uint16_t
27151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27152 vsqaddh_u16 (uint16_t __a, int16_t __b)
27153 {
27154   return __builtin_aarch64_usqaddhi_uus (__a, __b);
27155 }
27156
27157 __extension__ extern __inline uint32_t
27158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27159 vsqadds_u32 (uint32_t __a, int32_t __b)
27160 {
27161   return __builtin_aarch64_usqaddsi_uus (__a, __b);
27162 }
27163
27164 __extension__ extern __inline uint64_t
27165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27166 vsqaddd_u64 (uint64_t __a, int64_t __b)
27167 {
27168   return __builtin_aarch64_usqadddi_uus (__a, __b);
27169 }
27170
27171 /* vsqrt */
27172 __extension__ extern __inline float32x2_t
27173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27174 vsqrt_f32 (float32x2_t a)
27175 {
27176   return __builtin_aarch64_sqrtv2sf (a);
27177 }
27178
27179 __extension__ extern __inline float32x4_t
27180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27181 vsqrtq_f32 (float32x4_t a)
27182 {
27183   return __builtin_aarch64_sqrtv4sf (a);
27184 }
27185
27186 __extension__ extern __inline float64x1_t
27187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27188 vsqrt_f64 (float64x1_t a)
27189 {
27190   return (float64x1_t) { __builtin_aarch64_sqrtdf (a[0]) };
27191 }
27192
27193 __extension__ extern __inline float64x2_t
27194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27195 vsqrtq_f64 (float64x2_t a)
27196 {
27197   return __builtin_aarch64_sqrtv2df (a);
27198 }
27199
27200 /* vsra */
27201
27202 __extension__ extern __inline int8x8_t
27203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27204 vsra_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
27205 {
27206   return (int8x8_t) __builtin_aarch64_ssra_nv8qi (__a, __b, __c);
27207 }
27208
27209 __extension__ extern __inline int16x4_t
27210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27211 vsra_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
27212 {
27213   return (int16x4_t) __builtin_aarch64_ssra_nv4hi (__a, __b, __c);
27214 }
27215
27216 __extension__ extern __inline int32x2_t
27217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27218 vsra_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
27219 {
27220   return (int32x2_t) __builtin_aarch64_ssra_nv2si (__a, __b, __c);
27221 }
27222
27223 __extension__ extern __inline int64x1_t
27224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27225 vsra_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
27226 {
27227   return (int64x1_t) {__builtin_aarch64_ssra_ndi (__a[0], __b[0], __c)};
27228 }
27229
27230 __extension__ extern __inline uint8x8_t
27231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27232 vsra_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
27233 {
27234   return __builtin_aarch64_usra_nv8qi_uuus (__a, __b, __c);
27235 }
27236
27237 __extension__ extern __inline uint16x4_t
27238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27239 vsra_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
27240 {
27241   return __builtin_aarch64_usra_nv4hi_uuus (__a, __b, __c);
27242 }
27243
27244 __extension__ extern __inline uint32x2_t
27245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27246 vsra_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
27247 {
27248   return __builtin_aarch64_usra_nv2si_uuus (__a, __b, __c);
27249 }
27250
27251 __extension__ extern __inline uint64x1_t
27252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27253 vsra_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
27254 {
27255   return (uint64x1_t) {__builtin_aarch64_usra_ndi_uuus (__a[0], __b[0], __c)};
27256 }
27257
27258 __extension__ extern __inline int8x16_t
27259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27260 vsraq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
27261 {
27262   return (int8x16_t) __builtin_aarch64_ssra_nv16qi (__a, __b, __c);
27263 }
27264
27265 __extension__ extern __inline int16x8_t
27266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27267 vsraq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
27268 {
27269   return (int16x8_t) __builtin_aarch64_ssra_nv8hi (__a, __b, __c);
27270 }
27271
27272 __extension__ extern __inline int32x4_t
27273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27274 vsraq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
27275 {
27276   return (int32x4_t) __builtin_aarch64_ssra_nv4si (__a, __b, __c);
27277 }
27278
27279 __extension__ extern __inline int64x2_t
27280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27281 vsraq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
27282 {
27283   return (int64x2_t) __builtin_aarch64_ssra_nv2di (__a, __b, __c);
27284 }
27285
27286 __extension__ extern __inline uint8x16_t
27287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27288 vsraq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
27289 {
27290   return __builtin_aarch64_usra_nv16qi_uuus (__a, __b, __c);
27291 }
27292
27293 __extension__ extern __inline uint16x8_t
27294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27295 vsraq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
27296 {
27297   return __builtin_aarch64_usra_nv8hi_uuus (__a, __b, __c);
27298 }
27299
27300 __extension__ extern __inline uint32x4_t
27301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27302 vsraq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
27303 {
27304   return __builtin_aarch64_usra_nv4si_uuus (__a, __b, __c);
27305 }
27306
27307 __extension__ extern __inline uint64x2_t
27308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27309 vsraq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
27310 {
27311   return __builtin_aarch64_usra_nv2di_uuus (__a, __b, __c);
27312 }
27313
27314 __extension__ extern __inline int64_t
27315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27316 vsrad_n_s64 (int64_t __a, int64_t __b, const int __c)
27317 {
27318   return __builtin_aarch64_ssra_ndi (__a, __b, __c);
27319 }
27320
27321 __extension__ extern __inline uint64_t
27322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27323 vsrad_n_u64 (uint64_t __a, uint64_t __b, const int __c)
27324 {
27325   return __builtin_aarch64_usra_ndi_uuus (__a, __b, __c);
27326 }
27327
27328 /* vsri */
27329
27330 __extension__ extern __inline int8x8_t
27331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27332 vsri_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
27333 {
27334   return (int8x8_t) __builtin_aarch64_ssri_nv8qi (__a, __b, __c);
27335 }
27336
27337 __extension__ extern __inline int16x4_t
27338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27339 vsri_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
27340 {
27341   return (int16x4_t) __builtin_aarch64_ssri_nv4hi (__a, __b, __c);
27342 }
27343
27344 __extension__ extern __inline int32x2_t
27345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27346 vsri_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
27347 {
27348   return (int32x2_t) __builtin_aarch64_ssri_nv2si (__a, __b, __c);
27349 }
27350
27351 __extension__ extern __inline int64x1_t
27352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27353 vsri_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
27354 {
27355   return (int64x1_t) {__builtin_aarch64_ssri_ndi (__a[0], __b[0], __c)};
27356 }
27357
27358 __extension__ extern __inline uint8x8_t
27359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27360 vsri_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
27361 {
27362   return __builtin_aarch64_usri_nv8qi_uuus (__a, __b, __c);
27363 }
27364
27365 __extension__ extern __inline uint16x4_t
27366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27367 vsri_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
27368 {
27369   return __builtin_aarch64_usri_nv4hi_uuus (__a, __b, __c);
27370 }
27371
27372 __extension__ extern __inline uint32x2_t
27373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27374 vsri_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
27375 {
27376   return __builtin_aarch64_usri_nv2si_uuus (__a, __b, __c);
27377 }
27378
27379 __extension__ extern __inline uint64x1_t
27380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27381 vsri_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
27382 {
27383   return (uint64x1_t) {__builtin_aarch64_usri_ndi_uuus (__a[0], __b[0], __c)};
27384 }
27385
27386 __extension__ extern __inline int8x16_t
27387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27388 vsriq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
27389 {
27390   return (int8x16_t) __builtin_aarch64_ssri_nv16qi (__a, __b, __c);
27391 }
27392
27393 __extension__ extern __inline int16x8_t
27394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27395 vsriq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
27396 {
27397   return (int16x8_t) __builtin_aarch64_ssri_nv8hi (__a, __b, __c);
27398 }
27399
27400 __extension__ extern __inline int32x4_t
27401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27402 vsriq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
27403 {
27404   return (int32x4_t) __builtin_aarch64_ssri_nv4si (__a, __b, __c);
27405 }
27406
27407 __extension__ extern __inline int64x2_t
27408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27409 vsriq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
27410 {
27411   return (int64x2_t) __builtin_aarch64_ssri_nv2di (__a, __b, __c);
27412 }
27413
27414 __extension__ extern __inline uint8x16_t
27415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27416 vsriq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
27417 {
27418   return __builtin_aarch64_usri_nv16qi_uuus (__a, __b, __c);
27419 }
27420
27421 __extension__ extern __inline uint16x8_t
27422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27423 vsriq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
27424 {
27425   return __builtin_aarch64_usri_nv8hi_uuus (__a, __b, __c);
27426 }
27427
27428 __extension__ extern __inline uint32x4_t
27429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27430 vsriq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
27431 {
27432   return __builtin_aarch64_usri_nv4si_uuus (__a, __b, __c);
27433 }
27434
27435 __extension__ extern __inline uint64x2_t
27436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27437 vsriq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
27438 {
27439   return __builtin_aarch64_usri_nv2di_uuus (__a, __b, __c);
27440 }
27441
27442 __extension__ extern __inline int64_t
27443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27444 vsrid_n_s64 (int64_t __a, int64_t __b, const int __c)
27445 {
27446   return __builtin_aarch64_ssri_ndi (__a, __b, __c);
27447 }
27448
27449 __extension__ extern __inline uint64_t
27450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27451 vsrid_n_u64 (uint64_t __a, uint64_t __b, const int __c)
27452 {
27453   return __builtin_aarch64_usri_ndi_uuus (__a, __b, __c);
27454 }
27455
27456 /* vst1 */
27457
27458 __extension__ extern __inline void
27459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27460 vst1_f16 (float16_t *__a, float16x4_t __b)
27461 {
27462   __builtin_aarch64_st1v4hf (__a, __b);
27463 }
27464
27465 __extension__ extern __inline void
27466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27467 vst1_f32 (float32_t *a, float32x2_t b)
27468 {
27469   __builtin_aarch64_st1v2sf ((__builtin_aarch64_simd_sf *) a, b);
27470 }
27471
27472 __extension__ extern __inline void
27473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27474 vst1_f64 (float64_t *a, float64x1_t b)
27475 {
27476   *a = b[0];
27477 }
27478
27479 __extension__ extern __inline void
27480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27481 vst1_p8 (poly8_t *a, poly8x8_t b)
27482 {
27483   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a,
27484                              (int8x8_t) b);
27485 }
27486
27487 __extension__ extern __inline void
27488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27489 vst1_p16 (poly16_t *a, poly16x4_t b)
27490 {
27491   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a,
27492                              (int16x4_t) b);
27493 }
27494
27495 __extension__ extern __inline void
27496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27497 vst1_p64 (poly64_t *a, poly64x1_t b)
27498 {
27499   *a = b[0];
27500 }
27501
27502 __extension__ extern __inline void
27503 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27504 vst1_s8 (int8_t *a, int8x8_t b)
27505 {
27506   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a, b);
27507 }
27508
27509 __extension__ extern __inline void
27510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27511 vst1_s16 (int16_t *a, int16x4_t b)
27512 {
27513   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a, b);
27514 }
27515
27516 __extension__ extern __inline void
27517 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27518 vst1_s32 (int32_t *a, int32x2_t b)
27519 {
27520   __builtin_aarch64_st1v2si ((__builtin_aarch64_simd_si *) a, b);
27521 }
27522
27523 __extension__ extern __inline void
27524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27525 vst1_s64 (int64_t *a, int64x1_t b)
27526 {
27527   *a = b[0];
27528 }
27529
27530 __extension__ extern __inline void
27531 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27532 vst1_u8 (uint8_t *a, uint8x8_t b)
27533 {
27534   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a,
27535                              (int8x8_t) b);
27536 }
27537
27538 __extension__ extern __inline void
27539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27540 vst1_u16 (uint16_t *a, uint16x4_t b)
27541 {
27542   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a,
27543                              (int16x4_t) b);
27544 }
27545
27546 __extension__ extern __inline void
27547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27548 vst1_u32 (uint32_t *a, uint32x2_t b)
27549 {
27550   __builtin_aarch64_st1v2si ((__builtin_aarch64_simd_si *) a,
27551                              (int32x2_t) b);
27552 }
27553
27554 __extension__ extern __inline void
27555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27556 vst1_u64 (uint64_t *a, uint64x1_t b)
27557 {
27558   *a = b[0];
27559 }
27560
27561 /* vst1q */
27562
27563 __extension__ extern __inline void
27564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27565 vst1q_f16 (float16_t *__a, float16x8_t __b)
27566 {
27567   __builtin_aarch64_st1v8hf (__a, __b);
27568 }
27569
27570 __extension__ extern __inline void
27571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27572 vst1q_f32 (float32_t *a, float32x4_t b)
27573 {
27574   __builtin_aarch64_st1v4sf ((__builtin_aarch64_simd_sf *) a, b);
27575 }
27576
27577 __extension__ extern __inline void
27578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27579 vst1q_f64 (float64_t *a, float64x2_t b)
27580 {
27581   __builtin_aarch64_st1v2df ((__builtin_aarch64_simd_df *) a, b);
27582 }
27583
27584 __extension__ extern __inline void
27585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27586 vst1q_p8 (poly8_t *a, poly8x16_t b)
27587 {
27588   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a,
27589                               (int8x16_t) b);
27590 }
27591
27592 __extension__ extern __inline void
27593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27594 vst1q_p16 (poly16_t *a, poly16x8_t b)
27595 {
27596   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a,
27597                              (int16x8_t) b);
27598 }
27599
27600 __extension__ extern __inline void
27601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27602 vst1q_p64 (poly64_t *a, poly64x2_t b)
27603 {
27604   __builtin_aarch64_st1v2di_sp ((__builtin_aarch64_simd_di *) a,
27605                                 (poly64x2_t) b);
27606 }
27607
27608 __extension__ extern __inline void
27609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27610 vst1q_s8 (int8_t *a, int8x16_t b)
27611 {
27612   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a, b);
27613 }
27614
27615 __extension__ extern __inline void
27616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27617 vst1q_s16 (int16_t *a, int16x8_t b)
27618 {
27619   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a, b);
27620 }
27621
27622 __extension__ extern __inline void
27623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27624 vst1q_s32 (int32_t *a, int32x4_t b)
27625 {
27626   __builtin_aarch64_st1v4si ((__builtin_aarch64_simd_si *) a, b);
27627 }
27628
27629 __extension__ extern __inline void
27630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27631 vst1q_s64 (int64_t *a, int64x2_t b)
27632 {
27633   __builtin_aarch64_st1v2di ((__builtin_aarch64_simd_di *) a, b);
27634 }
27635
27636 __extension__ extern __inline void
27637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27638 vst1q_u8 (uint8_t *a, uint8x16_t b)
27639 {
27640   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a,
27641                               (int8x16_t) b);
27642 }
27643
27644 __extension__ extern __inline void
27645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27646 vst1q_u16 (uint16_t *a, uint16x8_t b)
27647 {
27648   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a,
27649                              (int16x8_t) b);
27650 }
27651
27652 __extension__ extern __inline void
27653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27654 vst1q_u32 (uint32_t *a, uint32x4_t b)
27655 {
27656   __builtin_aarch64_st1v4si ((__builtin_aarch64_simd_si *) a,
27657                              (int32x4_t) b);
27658 }
27659
27660 __extension__ extern __inline void
27661 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27662 vst1q_u64 (uint64_t *a, uint64x2_t b)
27663 {
27664   __builtin_aarch64_st1v2di ((__builtin_aarch64_simd_di *) a,
27665                              (int64x2_t) b);
27666 }
27667
27668 /* vst1_lane */
27669
27670 __extension__ extern __inline void
27671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27672 vst1_lane_f16 (float16_t *__a, float16x4_t __b, const int __lane)
27673 {
27674   *__a = __aarch64_vget_lane_any (__b, __lane);
27675 }
27676
27677 __extension__ extern __inline void
27678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27679 vst1_lane_f32 (float32_t *__a, float32x2_t __b, const int __lane)
27680 {
27681   *__a = __aarch64_vget_lane_any (__b, __lane);
27682 }
27683
27684 __extension__ extern __inline void
27685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27686 vst1_lane_f64 (float64_t *__a, float64x1_t __b, const int __lane)
27687 {
27688   *__a = __aarch64_vget_lane_any (__b, __lane);
27689 }
27690
27691 __extension__ extern __inline void
27692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27693 vst1_lane_p8 (poly8_t *__a, poly8x8_t __b, const int __lane)
27694 {
27695   *__a = __aarch64_vget_lane_any (__b, __lane);
27696 }
27697
27698 __extension__ extern __inline void
27699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27700 vst1_lane_p16 (poly16_t *__a, poly16x4_t __b, const int __lane)
27701 {
27702   *__a = __aarch64_vget_lane_any (__b, __lane);
27703 }
27704
27705 __extension__ extern __inline void
27706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27707 vst1_lane_p64 (poly64_t *__a, poly64x1_t __b, const int __lane)
27708 {
27709   *__a = __aarch64_vget_lane_any (__b, __lane);
27710 }
27711
27712 __extension__ extern __inline void
27713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27714 vst1_lane_s8 (int8_t *__a, int8x8_t __b, const int __lane)
27715 {
27716   *__a = __aarch64_vget_lane_any (__b, __lane);
27717 }
27718
27719 __extension__ extern __inline void
27720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27721 vst1_lane_s16 (int16_t *__a, int16x4_t __b, const int __lane)
27722 {
27723   *__a = __aarch64_vget_lane_any (__b, __lane);
27724 }
27725
27726 __extension__ extern __inline void
27727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27728 vst1_lane_s32 (int32_t *__a, int32x2_t __b, const int __lane)
27729 {
27730   *__a = __aarch64_vget_lane_any (__b, __lane);
27731 }
27732
27733 __extension__ extern __inline void
27734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27735 vst1_lane_s64 (int64_t *__a, int64x1_t __b, const int __lane)
27736 {
27737   *__a = __aarch64_vget_lane_any (__b, __lane);
27738 }
27739
27740 __extension__ extern __inline void
27741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27742 vst1_lane_u8 (uint8_t *__a, uint8x8_t __b, const int __lane)
27743 {
27744   *__a = __aarch64_vget_lane_any (__b, __lane);
27745 }
27746
27747 __extension__ extern __inline void
27748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27749 vst1_lane_u16 (uint16_t *__a, uint16x4_t __b, const int __lane)
27750 {
27751   *__a = __aarch64_vget_lane_any (__b, __lane);
27752 }
27753
27754 __extension__ extern __inline void
27755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27756 vst1_lane_u32 (uint32_t *__a, uint32x2_t __b, const int __lane)
27757 {
27758   *__a = __aarch64_vget_lane_any (__b, __lane);
27759 }
27760
27761 __extension__ extern __inline void
27762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27763 vst1_lane_u64 (uint64_t *__a, uint64x1_t __b, const int __lane)
27764 {
27765   *__a = __aarch64_vget_lane_any (__b, __lane);
27766 }
27767
27768 /* vst1q_lane */
27769
27770 __extension__ extern __inline void
27771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27772 vst1q_lane_f16 (float16_t *__a, float16x8_t __b, const int __lane)
27773 {
27774   *__a = __aarch64_vget_lane_any (__b, __lane);
27775 }
27776
27777 __extension__ extern __inline void
27778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27779 vst1q_lane_f32 (float32_t *__a, float32x4_t __b, const int __lane)
27780 {
27781   *__a = __aarch64_vget_lane_any (__b, __lane);
27782 }
27783
27784 __extension__ extern __inline void
27785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27786 vst1q_lane_f64 (float64_t *__a, float64x2_t __b, const int __lane)
27787 {
27788   *__a = __aarch64_vget_lane_any (__b, __lane);
27789 }
27790
27791 __extension__ extern __inline void
27792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27793 vst1q_lane_p8 (poly8_t *__a, poly8x16_t __b, const int __lane)
27794 {
27795   *__a = __aarch64_vget_lane_any (__b, __lane);
27796 }
27797
27798 __extension__ extern __inline void
27799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27800 vst1q_lane_p16 (poly16_t *__a, poly16x8_t __b, const int __lane)
27801 {
27802   *__a = __aarch64_vget_lane_any (__b, __lane);
27803 }
27804
27805 __extension__ extern __inline void
27806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27807 vst1q_lane_p64 (poly64_t *__a, poly64x2_t __b, const int __lane)
27808 {
27809   *__a = __aarch64_vget_lane_any (__b, __lane);
27810 }
27811
27812 __extension__ extern __inline void
27813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27814 vst1q_lane_s8 (int8_t *__a, int8x16_t __b, const int __lane)
27815 {
27816   *__a = __aarch64_vget_lane_any (__b, __lane);
27817 }
27818
27819 __extension__ extern __inline void
27820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27821 vst1q_lane_s16 (int16_t *__a, int16x8_t __b, const int __lane)
27822 {
27823   *__a = __aarch64_vget_lane_any (__b, __lane);
27824 }
27825
27826 __extension__ extern __inline void
27827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27828 vst1q_lane_s32 (int32_t *__a, int32x4_t __b, const int __lane)
27829 {
27830   *__a = __aarch64_vget_lane_any (__b, __lane);
27831 }
27832
27833 __extension__ extern __inline void
27834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27835 vst1q_lane_s64 (int64_t *__a, int64x2_t __b, const int __lane)
27836 {
27837   *__a = __aarch64_vget_lane_any (__b, __lane);
27838 }
27839
27840 __extension__ extern __inline void
27841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27842 vst1q_lane_u8 (uint8_t *__a, uint8x16_t __b, const int __lane)
27843 {
27844   *__a = __aarch64_vget_lane_any (__b, __lane);
27845 }
27846
27847 __extension__ extern __inline void
27848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27849 vst1q_lane_u16 (uint16_t *__a, uint16x8_t __b, const int __lane)
27850 {
27851   *__a = __aarch64_vget_lane_any (__b, __lane);
27852 }
27853
27854 __extension__ extern __inline void
27855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27856 vst1q_lane_u32 (uint32_t *__a, uint32x4_t __b, const int __lane)
27857 {
27858   *__a = __aarch64_vget_lane_any (__b, __lane);
27859 }
27860
27861 __extension__ extern __inline void
27862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27863 vst1q_lane_u64 (uint64_t *__a, uint64x2_t __b, const int __lane)
27864 {
27865   *__a = __aarch64_vget_lane_any (__b, __lane);
27866 }
27867
27868 /* vst1x2 */
27869
27870 __extension__ extern __inline void
27871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27872 vst1_s64_x2 (int64_t * __a, int64x1x2_t val)
27873 {
27874   __builtin_aarch64_simd_oi __o;
27875   int64x2x2_t temp;
27876   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
27877   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
27878   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
27879   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
27880   __builtin_aarch64_st1x2di ((__builtin_aarch64_simd_di *) __a, __o);
27881 }
27882
27883 __extension__ extern __inline void
27884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27885 vst1_u64_x2 (uint64_t * __a, uint64x1x2_t val)
27886 {
27887   __builtin_aarch64_simd_oi __o;
27888   uint64x2x2_t temp;
27889   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
27890   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
27891   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
27892   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
27893   __builtin_aarch64_st1x2di ((__builtin_aarch64_simd_di *) __a, __o);
27894 }
27895
27896 __extension__ extern __inline void
27897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27898 vst1_f64_x2 (float64_t * __a, float64x1x2_t val)
27899 {
27900   __builtin_aarch64_simd_oi __o;
27901   float64x2x2_t temp;
27902   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
27903   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
27904   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[0], 0);
27905   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[1], 1);
27906   __builtin_aarch64_st1x2df ((__builtin_aarch64_simd_df *) __a, __o);
27907 }
27908
27909 __extension__ extern __inline void
27910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27911 vst1_s8_x2 (int8_t * __a, int8x8x2_t val)
27912 {
27913   __builtin_aarch64_simd_oi __o;
27914   int8x16x2_t temp;
27915   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
27916   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
27917   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27918   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27919   __builtin_aarch64_st1x2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27920 }
27921
27922 __extension__ extern __inline void
27923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27924 vst1_p8_x2 (poly8_t * __a, poly8x8x2_t val)
27925 {
27926   __builtin_aarch64_simd_oi __o;
27927   poly8x16x2_t temp;
27928   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
27929   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
27930   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27931   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27932   __builtin_aarch64_st1x2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27933 }
27934
27935 __extension__ extern __inline void
27936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27937 vst1_s16_x2 (int16_t * __a, int16x4x2_t val)
27938 {
27939   __builtin_aarch64_simd_oi __o;
27940   int16x8x2_t temp;
27941   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
27942   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
27943   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27944   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27945   __builtin_aarch64_st1x2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27946 }
27947
27948 __extension__ extern __inline void
27949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27950 vst1_p16_x2 (poly16_t * __a, poly16x4x2_t val)
27951 {
27952   __builtin_aarch64_simd_oi __o;
27953   poly16x8x2_t temp;
27954   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
27955   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
27956   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27957   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27958   __builtin_aarch64_st1x2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27959 }
27960
27961 __extension__ extern __inline void
27962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27963 vst1_s32_x2 (int32_t * __a, int32x2x2_t val)
27964 {
27965   __builtin_aarch64_simd_oi __o;
27966   int32x4x2_t temp;
27967   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27968   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27969   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
27970   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
27971   __builtin_aarch64_st1x2v2si ((__builtin_aarch64_simd_si *) __a, __o);
27972 }
27973
27974 __extension__ extern __inline void
27975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27976 vst1_u8_x2 (uint8_t * __a, uint8x8x2_t val)
27977 {
27978   __builtin_aarch64_simd_oi __o;
27979   uint8x16x2_t temp;
27980   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27981   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27982   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27983   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27984   __builtin_aarch64_st1x2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27985 }
27986
27987 __extension__ extern __inline void
27988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27989 vst1_u16_x2 (uint16_t * __a, uint16x4x2_t val)
27990 {
27991   __builtin_aarch64_simd_oi __o;
27992   uint16x8x2_t temp;
27993   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27994   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27995   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27996   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27997   __builtin_aarch64_st1x2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27998 }
27999
28000 __extension__ extern __inline void
28001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28002 vst1_u32_x2 (uint32_t * __a, uint32x2x2_t val)
28003 {
28004   __builtin_aarch64_simd_oi __o;
28005   uint32x4x2_t temp;
28006   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
28007   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
28008   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
28009   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
28010   __builtin_aarch64_st1x2v2si ((__builtin_aarch64_simd_si *) __a, __o);
28011 }
28012
28013 __extension__ extern __inline void
28014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28015 vst1_f16_x2 (float16_t * __a, float16x4x2_t val)
28016 {
28017   __builtin_aarch64_simd_oi __o;
28018   float16x8x2_t temp;
28019   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
28020   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
28021   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[0], 0);
28022   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[1], 1);
28023   __builtin_aarch64_st1x2v4hf (__a, __o);
28024 }
28025
28026 __extension__ extern __inline void
28027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28028 vst1_f32_x2 (float32_t * __a, float32x2x2_t val)
28029 {
28030   __builtin_aarch64_simd_oi __o;
28031   float32x4x2_t temp;
28032   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
28033   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
28034   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[0], 0);
28035   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[1], 1);
28036   __builtin_aarch64_st1x2v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
28037 }
28038
28039 __extension__ extern __inline void
28040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28041 vst1_p64_x2 (poly64_t * __a, poly64x1x2_t val)
28042 {
28043   __builtin_aarch64_simd_oi __o;
28044   poly64x2x2_t temp;
28045   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
28046   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
28047   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28048                                                (poly64x2_t) temp.val[0], 0);
28049   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28050                                                (poly64x2_t) temp.val[1], 1);
28051   __builtin_aarch64_st1x2di ((__builtin_aarch64_simd_di *) __a, __o);
28052 }
28053
28054 __extension__ extern __inline void
28055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28056 vst1q_s8_x2 (int8_t * __a, int8x16x2_t val)
28057 {
28058   __builtin_aarch64_simd_oi __o;
28059   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
28060   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
28061   __builtin_aarch64_st1x2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28062 }
28063
28064 __extension__ extern __inline void
28065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28066 vst1q_p8_x2 (poly8_t * __a, poly8x16x2_t val)
28067 {
28068   __builtin_aarch64_simd_oi __o;
28069   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
28070   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
28071   __builtin_aarch64_st1x2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28072 }
28073
28074 __extension__ extern __inline void
28075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28076 vst1q_s16_x2 (int16_t * __a, int16x8x2_t val)
28077 {
28078   __builtin_aarch64_simd_oi __o;
28079   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
28080   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
28081   __builtin_aarch64_st1x2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28082 }
28083
28084 __extension__ extern __inline void
28085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28086 vst1q_p16_x2 (poly16_t * __a, poly16x8x2_t val)
28087 {
28088   __builtin_aarch64_simd_oi __o;
28089   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
28090   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
28091   __builtin_aarch64_st1x2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28092 }
28093
28094 __extension__ extern __inline void
28095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28096 vst1q_s32_x2 (int32_t * __a, int32x4x2_t val)
28097 {
28098   __builtin_aarch64_simd_oi __o;
28099   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
28100   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
28101   __builtin_aarch64_st1x2v4si ((__builtin_aarch64_simd_si *) __a, __o);
28102 }
28103
28104 __extension__ extern __inline void
28105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28106 vst1q_s64_x2 (int64_t * __a, int64x2x2_t val)
28107 {
28108   __builtin_aarch64_simd_oi __o;
28109   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
28110   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
28111   __builtin_aarch64_st1x2v2di ((__builtin_aarch64_simd_di *) __a, __o);
28112 }
28113
28114 __extension__ extern __inline void
28115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28116 vst1q_u8_x2 (uint8_t * __a, uint8x16x2_t val)
28117 {
28118   __builtin_aarch64_simd_oi __o;
28119   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
28120   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
28121   __builtin_aarch64_st1x2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28122 }
28123
28124 __extension__ extern __inline void
28125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28126 vst1q_u16_x2 (uint16_t * __a, uint16x8x2_t val)
28127 {
28128   __builtin_aarch64_simd_oi __o;
28129   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
28130   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
28131   __builtin_aarch64_st1x2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28132 }
28133
28134 __extension__ extern __inline void
28135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28136 vst1q_u32_x2 (uint32_t * __a, uint32x4x2_t val)
28137 {
28138   __builtin_aarch64_simd_oi __o;
28139   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
28140   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
28141   __builtin_aarch64_st1x2v4si ((__builtin_aarch64_simd_si *) __a, __o);
28142 }
28143
28144 __extension__ extern __inline void
28145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28146 vst1q_u64_x2 (uint64_t * __a, uint64x2x2_t val)
28147 {
28148   __builtin_aarch64_simd_oi __o;
28149   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
28150   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
28151   __builtin_aarch64_st1x2v2di ((__builtin_aarch64_simd_di *) __a, __o);
28152 }
28153
28154 __extension__ extern __inline void
28155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28156 vst1q_f16_x2 (float16_t * __a, float16x8x2_t val)
28157 {
28158   __builtin_aarch64_simd_oi __o;
28159   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[0], 0);
28160   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[1], 1);
28161   __builtin_aarch64_st1x2v8hf (__a, __o);
28162 }
28163
28164 __extension__ extern __inline void
28165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28166 vst1q_f32_x2 (float32_t * __a, float32x4x2_t val)
28167 {
28168   __builtin_aarch64_simd_oi __o;
28169   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[0], 0);
28170   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[1], 1);
28171   __builtin_aarch64_st1x2v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
28172 }
28173
28174 __extension__ extern __inline void
28175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28176 vst1q_f64_x2 (float64_t * __a, float64x2x2_t val)
28177 {
28178   __builtin_aarch64_simd_oi __o;
28179   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[0], 0);
28180   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[1], 1);
28181   __builtin_aarch64_st1x2v2df ((__builtin_aarch64_simd_df *) __a, __o);
28182 }
28183
28184 __extension__ extern __inline void
28185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28186 vst1q_p64_x2 (poly64_t * __a, poly64x2x2_t val)
28187 {
28188   __builtin_aarch64_simd_oi __o;
28189   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28190                                                (poly64x2_t) val.val[0], 0);
28191   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28192                                                (poly64x2_t) val.val[1], 1);
28193   __builtin_aarch64_st1x2v2di ((__builtin_aarch64_simd_di *) __a, __o);
28194 }
28195
28196 /* vst1x3 */
28197
28198 __extension__ extern __inline void
28199 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28200 vst1_s64_x3 (int64_t * __a, int64x1x3_t val)
28201 {
28202   __builtin_aarch64_simd_ci __o;
28203   int64x2x3_t temp;
28204   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
28205   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
28206   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
28207   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
28208   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
28209   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
28210   __builtin_aarch64_st1x3di ((__builtin_aarch64_simd_di *) __a, __o);
28211 }
28212
28213 __extension__ extern __inline void
28214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28215 vst1_u64_x3 (uint64_t * __a, uint64x1x3_t val)
28216 {
28217   __builtin_aarch64_simd_ci __o;
28218   uint64x2x3_t temp;
28219   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
28220   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
28221   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
28222   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
28223   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
28224   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
28225   __builtin_aarch64_st1x3di ((__builtin_aarch64_simd_di *) __a, __o);
28226 }
28227
28228 __extension__ extern __inline void
28229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28230 vst1_f64_x3 (float64_t * __a, float64x1x3_t val)
28231 {
28232   __builtin_aarch64_simd_ci __o;
28233   float64x2x3_t temp;
28234   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
28235   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
28236   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
28237   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[0], 0);
28238   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[1], 1);
28239   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[2], 2);
28240   __builtin_aarch64_st1x3df ((__builtin_aarch64_simd_df *) __a, __o);
28241 }
28242
28243 __extension__ extern __inline void
28244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28245 vst1_s8_x3 (int8_t * __a, int8x8x3_t val)
28246 {
28247   __builtin_aarch64_simd_ci __o;
28248   int8x16x3_t temp;
28249   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
28250   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
28251   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
28252   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
28253   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
28254   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
28255   __builtin_aarch64_st1x3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28256 }
28257
28258 __extension__ extern __inline void
28259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28260 vst1_p8_x3 (poly8_t * __a, poly8x8x3_t val)
28261 {
28262   __builtin_aarch64_simd_ci __o;
28263   poly8x16x3_t temp;
28264   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
28265   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
28266   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
28267   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
28268   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
28269   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
28270   __builtin_aarch64_st1x3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28271 }
28272
28273 __extension__ extern __inline void
28274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28275 vst1_s16_x3 (int16_t * __a, int16x4x3_t val)
28276 {
28277   __builtin_aarch64_simd_ci __o;
28278   int16x8x3_t temp;
28279   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
28280   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
28281   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
28282   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
28283   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
28284   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
28285   __builtin_aarch64_st1x3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28286 }
28287
28288 __extension__ extern __inline void
28289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28290 vst1_p16_x3 (poly16_t * __a, poly16x4x3_t val)
28291 {
28292   __builtin_aarch64_simd_ci __o;
28293   poly16x8x3_t temp;
28294   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
28295   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
28296   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
28297   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
28298   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
28299   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
28300   __builtin_aarch64_st1x3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28301 }
28302
28303 __extension__ extern __inline void
28304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28305 vst1_s32_x3 (int32_t * __a, int32x2x3_t val)
28306 {
28307   __builtin_aarch64_simd_ci __o;
28308   int32x4x3_t temp;
28309   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
28310   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
28311   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
28312   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
28313   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
28314   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
28315   __builtin_aarch64_st1x3v2si ((__builtin_aarch64_simd_si *) __a, __o);
28316 }
28317
28318 __extension__ extern __inline void
28319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28320 vst1_u8_x3 (uint8_t * __a, uint8x8x3_t val)
28321 {
28322   __builtin_aarch64_simd_ci __o;
28323   uint8x16x3_t temp;
28324   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
28325   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
28326   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
28327   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
28328   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
28329   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
28330   __builtin_aarch64_st1x3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28331 }
28332
28333 __extension__ extern __inline void
28334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28335 vst1_u16_x3 (uint16_t * __a, uint16x4x3_t val)
28336 {
28337   __builtin_aarch64_simd_ci __o;
28338   uint16x8x3_t temp;
28339   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
28340   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
28341   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
28342   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
28343   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
28344   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
28345   __builtin_aarch64_st1x3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28346 }
28347
28348 __extension__ extern __inline void
28349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28350 vst1_u32_x3 (uint32_t * __a, uint32x2x3_t val)
28351 {
28352   __builtin_aarch64_simd_ci __o;
28353   uint32x4x3_t temp;
28354   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
28355   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
28356   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
28357   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
28358   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
28359   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
28360   __builtin_aarch64_st1x3v2si ((__builtin_aarch64_simd_si *) __a, __o);
28361 }
28362
28363 __extension__ extern __inline void
28364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28365 vst1_f16_x3 (float16_t * __a, float16x4x3_t val)
28366 {
28367   __builtin_aarch64_simd_ci __o;
28368   float16x8x3_t temp;
28369   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
28370   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
28371   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
28372   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[0], 0);
28373   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[1], 1);
28374   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[2], 2);
28375   __builtin_aarch64_st1x3v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
28376 }
28377
28378 __extension__ extern __inline void
28379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28380 vst1_f32_x3 (float32_t * __a, float32x2x3_t val)
28381 {
28382   __builtin_aarch64_simd_ci __o;
28383   float32x4x3_t temp;
28384   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
28385   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
28386   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
28387   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[0], 0);
28388   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[1], 1);
28389   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[2], 2);
28390   __builtin_aarch64_st1x3v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
28391 }
28392
28393 __extension__ extern __inline void
28394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28395 vst1_p64_x3 (poly64_t * __a, poly64x1x3_t val)
28396 {
28397   __builtin_aarch64_simd_ci __o;
28398   poly64x2x3_t temp;
28399   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
28400   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
28401   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
28402   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
28403                                                (poly64x2_t) temp.val[0], 0);
28404   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
28405                                                (poly64x2_t) temp.val[1], 1);
28406   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
28407                                                (poly64x2_t) temp.val[2], 2);
28408   __builtin_aarch64_st1x3di ((__builtin_aarch64_simd_di *) __a, __o);
28409 }
28410
28411 __extension__ extern __inline void
28412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28413 vst1q_s8_x3 (int8_t * __a, int8x16x3_t val)
28414 {
28415   __builtin_aarch64_simd_ci __o;
28416   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
28417   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
28418   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
28419   __builtin_aarch64_st1x3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28420 }
28421
28422 __extension__ extern __inline void
28423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28424 vst1q_p8_x3 (poly8_t * __a, poly8x16x3_t val)
28425 {
28426   __builtin_aarch64_simd_ci __o;
28427   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
28428   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
28429   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
28430   __builtin_aarch64_st1x3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28431 }
28432
28433 __extension__ extern __inline void
28434 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28435 vst1q_s16_x3 (int16_t * __a, int16x8x3_t val)
28436 {
28437   __builtin_aarch64_simd_ci __o;
28438   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
28439   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
28440   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
28441   __builtin_aarch64_st1x3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28442 }
28443
28444 __extension__ extern __inline void
28445 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28446 vst1q_p16_x3 (poly16_t * __a, poly16x8x3_t val)
28447 {
28448   __builtin_aarch64_simd_ci __o;
28449   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
28450   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
28451   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
28452   __builtin_aarch64_st1x3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28453 }
28454
28455 __extension__ extern __inline void
28456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28457 vst1q_s32_x3 (int32_t * __a, int32x4x3_t val)
28458 {
28459   __builtin_aarch64_simd_ci __o;
28460   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
28461   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
28462   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
28463   __builtin_aarch64_st1x3v4si ((__builtin_aarch64_simd_si *) __a, __o);
28464 }
28465
28466 __extension__ extern __inline void
28467 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28468 vst1q_s64_x3 (int64_t * __a, int64x2x3_t val)
28469 {
28470   __builtin_aarch64_simd_ci __o;
28471   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
28472   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
28473   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
28474   __builtin_aarch64_st1x3v2di ((__builtin_aarch64_simd_di *) __a, __o);
28475 }
28476
28477 __extension__ extern __inline void
28478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28479 vst1q_u8_x3 (uint8_t * __a, uint8x16x3_t val)
28480 {
28481   __builtin_aarch64_simd_ci __o;
28482   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
28483   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
28484   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
28485   __builtin_aarch64_st1x3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28486 }
28487
28488 __extension__ extern __inline void
28489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28490 vst1q_u16_x3 (uint16_t * __a, uint16x8x3_t val)
28491 {
28492   __builtin_aarch64_simd_ci __o;
28493   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
28494   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
28495   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
28496   __builtin_aarch64_st1x3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28497 }
28498
28499 __extension__ extern __inline void
28500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28501 vst1q_u32_x3 (uint32_t * __a, uint32x4x3_t val)
28502 {
28503   __builtin_aarch64_simd_ci __o;
28504   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
28505   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
28506   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
28507   __builtin_aarch64_st1x3v4si ((__builtin_aarch64_simd_si *) __a, __o);
28508 }
28509
28510 __extension__ extern __inline void
28511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28512 vst1q_u64_x3 (uint64_t * __a, uint64x2x3_t val)
28513 {
28514   __builtin_aarch64_simd_ci __o;
28515   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
28516   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
28517   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
28518   __builtin_aarch64_st1x3v2di ((__builtin_aarch64_simd_di *) __a, __o);
28519 }
28520
28521 __extension__ extern __inline void
28522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28523 vst1q_f16_x3 (float16_t * __a, float16x8x3_t val)
28524 {
28525   __builtin_aarch64_simd_ci __o;
28526   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[0], 0);
28527   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[1], 1);
28528   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[2], 2);
28529   __builtin_aarch64_st1x3v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
28530 }
28531
28532 __extension__ extern __inline void
28533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28534 vst1q_f32_x3 (float32_t * __a, float32x4x3_t val)
28535 {
28536   __builtin_aarch64_simd_ci __o;
28537   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[0], 0);
28538   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[1], 1);
28539   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[2], 2);
28540   __builtin_aarch64_st1x3v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
28541 }
28542
28543 __extension__ extern __inline void
28544 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28545 vst1q_f64_x3 (float64_t * __a, float64x2x3_t val)
28546 {
28547   __builtin_aarch64_simd_ci __o;
28548   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[0], 0);
28549   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[1], 1);
28550   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[2], 2);
28551   __builtin_aarch64_st1x3v2df ((__builtin_aarch64_simd_df *) __a, __o);
28552 }
28553
28554 __extension__ extern __inline void
28555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28556 vst1q_p64_x3 (poly64_t * __a, poly64x2x3_t val)
28557 {
28558   __builtin_aarch64_simd_ci __o;
28559   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
28560                                                (poly64x2_t) val.val[0], 0);
28561   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
28562                                                (poly64x2_t) val.val[1], 1);
28563   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
28564                                                (poly64x2_t) val.val[2], 2);
28565   __builtin_aarch64_st1x3v2di ((__builtin_aarch64_simd_di *) __a, __o);
28566 }
28567
28568 /* vstn */
28569
28570 __extension__ extern __inline void
28571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28572 vst2_s64 (int64_t * __a, int64x1x2_t val)
28573 {
28574   __builtin_aarch64_simd_oi __o;
28575   int64x2x2_t temp;
28576   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
28577   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
28578   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
28579   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
28580   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
28581 }
28582
28583 __extension__ extern __inline void
28584 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28585 vst2_u64 (uint64_t * __a, uint64x1x2_t val)
28586 {
28587   __builtin_aarch64_simd_oi __o;
28588   uint64x2x2_t temp;
28589   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
28590   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
28591   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
28592   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
28593   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
28594 }
28595
28596 __extension__ extern __inline void
28597 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28598 vst2_f64 (float64_t * __a, float64x1x2_t val)
28599 {
28600   __builtin_aarch64_simd_oi __o;
28601   float64x2x2_t temp;
28602   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
28603   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
28604   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[0], 0);
28605   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[1], 1);
28606   __builtin_aarch64_st2df ((__builtin_aarch64_simd_df *) __a, __o);
28607 }
28608
28609 __extension__ extern __inline void
28610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28611 vst2_s8 (int8_t * __a, int8x8x2_t val)
28612 {
28613   __builtin_aarch64_simd_oi __o;
28614   int8x16x2_t temp;
28615   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
28616   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
28617   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
28618   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
28619   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28620 }
28621
28622 __extension__ extern __inline void
28623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28624 vst2_p8 (poly8_t * __a, poly8x8x2_t val)
28625 {
28626   __builtin_aarch64_simd_oi __o;
28627   poly8x16x2_t temp;
28628   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
28629   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
28630   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
28631   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
28632   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28633 }
28634
28635 __extension__ extern __inline void
28636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28637 vst2_s16 (int16_t * __a, int16x4x2_t val)
28638 {
28639   __builtin_aarch64_simd_oi __o;
28640   int16x8x2_t temp;
28641   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
28642   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
28643   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
28644   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
28645   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28646 }
28647
28648 __extension__ extern __inline void
28649 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28650 vst2_p16 (poly16_t * __a, poly16x4x2_t val)
28651 {
28652   __builtin_aarch64_simd_oi __o;
28653   poly16x8x2_t temp;
28654   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
28655   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
28656   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
28657   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
28658   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28659 }
28660
28661 __extension__ extern __inline void
28662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28663 vst2_s32 (int32_t * __a, int32x2x2_t val)
28664 {
28665   __builtin_aarch64_simd_oi __o;
28666   int32x4x2_t temp;
28667   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
28668   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
28669   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
28670   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
28671   __builtin_aarch64_st2v2si ((__builtin_aarch64_simd_si *) __a, __o);
28672 }
28673
28674 __extension__ extern __inline void
28675 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28676 vst2_u8 (uint8_t * __a, uint8x8x2_t val)
28677 {
28678   __builtin_aarch64_simd_oi __o;
28679   uint8x16x2_t temp;
28680   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
28681   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
28682   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
28683   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
28684   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28685 }
28686
28687 __extension__ extern __inline void
28688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28689 vst2_u16 (uint16_t * __a, uint16x4x2_t val)
28690 {
28691   __builtin_aarch64_simd_oi __o;
28692   uint16x8x2_t temp;
28693   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
28694   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
28695   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
28696   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
28697   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28698 }
28699
28700 __extension__ extern __inline void
28701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28702 vst2_u32 (uint32_t * __a, uint32x2x2_t val)
28703 {
28704   __builtin_aarch64_simd_oi __o;
28705   uint32x4x2_t temp;
28706   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
28707   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
28708   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
28709   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
28710   __builtin_aarch64_st2v2si ((__builtin_aarch64_simd_si *) __a, __o);
28711 }
28712
28713 __extension__ extern __inline void
28714 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28715 vst2_f16 (float16_t * __a, float16x4x2_t val)
28716 {
28717   __builtin_aarch64_simd_oi __o;
28718   float16x8x2_t temp;
28719   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
28720   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
28721   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[0], 0);
28722   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[1], 1);
28723   __builtin_aarch64_st2v4hf (__a, __o);
28724 }
28725
28726 __extension__ extern __inline void
28727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28728 vst2_f32 (float32_t * __a, float32x2x2_t val)
28729 {
28730   __builtin_aarch64_simd_oi __o;
28731   float32x4x2_t temp;
28732   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
28733   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
28734   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[0], 0);
28735   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[1], 1);
28736   __builtin_aarch64_st2v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
28737 }
28738
28739 __extension__ extern __inline void
28740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28741 vst2_p64 (poly64_t * __a, poly64x1x2_t val)
28742 {
28743   __builtin_aarch64_simd_oi __o;
28744   poly64x2x2_t temp;
28745   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
28746   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
28747   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28748                                                (poly64x2_t) temp.val[0], 0);
28749   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28750                                                (poly64x2_t) temp.val[1], 1);
28751   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
28752 }
28753
28754 __extension__ extern __inline void
28755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28756 vst2q_s8 (int8_t * __a, int8x16x2_t val)
28757 {
28758   __builtin_aarch64_simd_oi __o;
28759   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
28760   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
28761   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28762 }
28763
28764 __extension__ extern __inline void
28765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28766 vst2q_p8 (poly8_t * __a, poly8x16x2_t val)
28767 {
28768   __builtin_aarch64_simd_oi __o;
28769   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
28770   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
28771   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28772 }
28773
28774 __extension__ extern __inline void
28775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28776 vst2q_s16 (int16_t * __a, int16x8x2_t val)
28777 {
28778   __builtin_aarch64_simd_oi __o;
28779   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
28780   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
28781   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28782 }
28783
28784 __extension__ extern __inline void
28785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28786 vst2q_p16 (poly16_t * __a, poly16x8x2_t val)
28787 {
28788   __builtin_aarch64_simd_oi __o;
28789   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
28790   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
28791   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28792 }
28793
28794 __extension__ extern __inline void
28795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28796 vst2q_s32 (int32_t * __a, int32x4x2_t val)
28797 {
28798   __builtin_aarch64_simd_oi __o;
28799   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
28800   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
28801   __builtin_aarch64_st2v4si ((__builtin_aarch64_simd_si *) __a, __o);
28802 }
28803
28804 __extension__ extern __inline void
28805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28806 vst2q_s64 (int64_t * __a, int64x2x2_t val)
28807 {
28808   __builtin_aarch64_simd_oi __o;
28809   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
28810   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
28811   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
28812 }
28813
28814 __extension__ extern __inline void
28815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28816 vst2q_u8 (uint8_t * __a, uint8x16x2_t val)
28817 {
28818   __builtin_aarch64_simd_oi __o;
28819   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
28820   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
28821   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
28822 }
28823
28824 __extension__ extern __inline void
28825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28826 vst2q_u16 (uint16_t * __a, uint16x8x2_t val)
28827 {
28828   __builtin_aarch64_simd_oi __o;
28829   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
28830   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
28831   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28832 }
28833
28834 __extension__ extern __inline void
28835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28836 vst2q_u32 (uint32_t * __a, uint32x4x2_t val)
28837 {
28838   __builtin_aarch64_simd_oi __o;
28839   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
28840   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
28841   __builtin_aarch64_st2v4si ((__builtin_aarch64_simd_si *) __a, __o);
28842 }
28843
28844 __extension__ extern __inline void
28845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28846 vst2q_u64 (uint64_t * __a, uint64x2x2_t val)
28847 {
28848   __builtin_aarch64_simd_oi __o;
28849   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
28850   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
28851   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
28852 }
28853
28854 __extension__ extern __inline void
28855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28856 vst2q_f16 (float16_t * __a, float16x8x2_t val)
28857 {
28858   __builtin_aarch64_simd_oi __o;
28859   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[0], 0);
28860   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[1], 1);
28861   __builtin_aarch64_st2v8hf (__a, __o);
28862 }
28863
28864 __extension__ extern __inline void
28865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28866 vst2q_f32 (float32_t * __a, float32x4x2_t val)
28867 {
28868   __builtin_aarch64_simd_oi __o;
28869   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[0], 0);
28870   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[1], 1);
28871   __builtin_aarch64_st2v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
28872 }
28873
28874 __extension__ extern __inline void
28875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28876 vst2q_f64 (float64_t * __a, float64x2x2_t val)
28877 {
28878   __builtin_aarch64_simd_oi __o;
28879   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[0], 0);
28880   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[1], 1);
28881   __builtin_aarch64_st2v2df ((__builtin_aarch64_simd_df *) __a, __o);
28882 }
28883
28884 __extension__ extern __inline void
28885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28886 vst2q_p64 (poly64_t * __a, poly64x2x2_t val)
28887 {
28888   __builtin_aarch64_simd_oi __o;
28889   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28890                                                (poly64x2_t) val.val[0], 0);
28891   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
28892                                                (poly64x2_t) val.val[1], 1);
28893   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
28894 }
28895
28896 __extension__ extern __inline void
28897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28898 vst3_s64 (int64_t * __a, int64x1x3_t val)
28899 {
28900   __builtin_aarch64_simd_ci __o;
28901   int64x2x3_t temp;
28902   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
28903   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
28904   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
28905   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
28906   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
28907   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
28908   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
28909 }
28910
28911 __extension__ extern __inline void
28912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28913 vst3_u64 (uint64_t * __a, uint64x1x3_t val)
28914 {
28915   __builtin_aarch64_simd_ci __o;
28916   uint64x2x3_t temp;
28917   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
28918   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
28919   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
28920   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
28921   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
28922   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
28923   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
28924 }
28925
28926 __extension__ extern __inline void
28927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28928 vst3_f64 (float64_t * __a, float64x1x3_t val)
28929 {
28930   __builtin_aarch64_simd_ci __o;
28931   float64x2x3_t temp;
28932   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
28933   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
28934   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
28935   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[0], 0);
28936   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[1], 1);
28937   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[2], 2);
28938   __builtin_aarch64_st3df ((__builtin_aarch64_simd_df *) __a, __o);
28939 }
28940
28941 __extension__ extern __inline void
28942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28943 vst3_s8 (int8_t * __a, int8x8x3_t val)
28944 {
28945   __builtin_aarch64_simd_ci __o;
28946   int8x16x3_t temp;
28947   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
28948   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
28949   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
28950   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
28951   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
28952   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
28953   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28954 }
28955
28956 __extension__ extern __inline void
28957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28958 vst3_p8 (poly8_t * __a, poly8x8x3_t val)
28959 {
28960   __builtin_aarch64_simd_ci __o;
28961   poly8x16x3_t temp;
28962   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
28963   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
28964   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
28965   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
28966   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
28967   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
28968   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
28969 }
28970
28971 __extension__ extern __inline void
28972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28973 vst3_s16 (int16_t * __a, int16x4x3_t val)
28974 {
28975   __builtin_aarch64_simd_ci __o;
28976   int16x8x3_t temp;
28977   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
28978   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
28979   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
28980   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
28981   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
28982   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
28983   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28984 }
28985
28986 __extension__ extern __inline void
28987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28988 vst3_p16 (poly16_t * __a, poly16x4x3_t val)
28989 {
28990   __builtin_aarch64_simd_ci __o;
28991   poly16x8x3_t temp;
28992   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
28993   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
28994   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
28995   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
28996   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
28997   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
28998   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
28999 }
29000
29001 __extension__ extern __inline void
29002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29003 vst3_s32 (int32_t * __a, int32x2x3_t val)
29004 {
29005   __builtin_aarch64_simd_ci __o;
29006   int32x4x3_t temp;
29007   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
29008   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
29009   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
29010   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
29011   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
29012   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
29013   __builtin_aarch64_st3v2si ((__builtin_aarch64_simd_si *) __a, __o);
29014 }
29015
29016 __extension__ extern __inline void
29017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29018 vst3_u8 (uint8_t * __a, uint8x8x3_t val)
29019 {
29020   __builtin_aarch64_simd_ci __o;
29021   uint8x16x3_t temp;
29022   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
29023   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
29024   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
29025   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
29026   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
29027   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
29028   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
29029 }
29030
29031 __extension__ extern __inline void
29032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29033 vst3_u16 (uint16_t * __a, uint16x4x3_t val)
29034 {
29035   __builtin_aarch64_simd_ci __o;
29036   uint16x8x3_t temp;
29037   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
29038   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
29039   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
29040   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
29041   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
29042   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
29043   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
29044 }
29045
29046 __extension__ extern __inline void
29047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29048 vst3_u32 (uint32_t * __a, uint32x2x3_t val)
29049 {
29050   __builtin_aarch64_simd_ci __o;
29051   uint32x4x3_t temp;
29052   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
29053   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
29054   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
29055   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
29056   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
29057   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
29058   __builtin_aarch64_st3v2si ((__builtin_aarch64_simd_si *) __a, __o);
29059 }
29060
29061 __extension__ extern __inline void
29062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29063 vst3_f16 (float16_t * __a, float16x4x3_t val)
29064 {
29065   __builtin_aarch64_simd_ci __o;
29066   float16x8x3_t temp;
29067   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
29068   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
29069   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
29070   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[0], 0);
29071   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[1], 1);
29072   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[2], 2);
29073   __builtin_aarch64_st3v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
29074 }
29075
29076 __extension__ extern __inline void
29077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29078 vst3_f32 (float32_t * __a, float32x2x3_t val)
29079 {
29080   __builtin_aarch64_simd_ci __o;
29081   float32x4x3_t temp;
29082   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
29083   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
29084   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
29085   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[0], 0);
29086   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[1], 1);
29087   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[2], 2);
29088   __builtin_aarch64_st3v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
29089 }
29090
29091 __extension__ extern __inline void
29092 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29093 vst3_p64 (poly64_t * __a, poly64x1x3_t val)
29094 {
29095   __builtin_aarch64_simd_ci __o;
29096   poly64x2x3_t temp;
29097   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
29098   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
29099   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
29100   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
29101                                                (poly64x2_t) temp.val[0], 0);
29102   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
29103                                                (poly64x2_t) temp.val[1], 1);
29104   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
29105                                                (poly64x2_t) temp.val[2], 2);
29106   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
29107 }
29108
29109 __extension__ extern __inline void
29110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29111 vst3q_s8 (int8_t * __a, int8x16x3_t val)
29112 {
29113   __builtin_aarch64_simd_ci __o;
29114   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
29115   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
29116   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
29117   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
29118 }
29119
29120 __extension__ extern __inline void
29121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29122 vst3q_p8 (poly8_t * __a, poly8x16x3_t val)
29123 {
29124   __builtin_aarch64_simd_ci __o;
29125   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
29126   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
29127   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
29128   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
29129 }
29130
29131 __extension__ extern __inline void
29132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29133 vst3q_s16 (int16_t * __a, int16x8x3_t val)
29134 {
29135   __builtin_aarch64_simd_ci __o;
29136   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
29137   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
29138   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
29139   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
29140 }
29141
29142 __extension__ extern __inline void
29143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29144 vst3q_p16 (poly16_t * __a, poly16x8x3_t val)
29145 {
29146   __builtin_aarch64_simd_ci __o;
29147   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
29148   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
29149   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
29150   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
29151 }
29152
29153 __extension__ extern __inline void
29154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29155 vst3q_s32 (int32_t * __a, int32x4x3_t val)
29156 {
29157   __builtin_aarch64_simd_ci __o;
29158   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
29159   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
29160   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
29161   __builtin_aarch64_st3v4si ((__builtin_aarch64_simd_si *) __a, __o);
29162 }
29163
29164 __extension__ extern __inline void
29165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29166 vst3q_s64 (int64_t * __a, int64x2x3_t val)
29167 {
29168   __builtin_aarch64_simd_ci __o;
29169   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
29170   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
29171   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
29172   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
29173 }
29174
29175 __extension__ extern __inline void
29176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29177 vst3q_u8 (uint8_t * __a, uint8x16x3_t val)
29178 {
29179   __builtin_aarch64_simd_ci __o;
29180   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
29181   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
29182   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
29183   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
29184 }
29185
29186 __extension__ extern __inline void
29187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29188 vst3q_u16 (uint16_t * __a, uint16x8x3_t val)
29189 {
29190   __builtin_aarch64_simd_ci __o;
29191   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
29192   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
29193   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
29194   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
29195 }
29196
29197 __extension__ extern __inline void
29198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29199 vst3q_u32 (uint32_t * __a, uint32x4x3_t val)
29200 {
29201   __builtin_aarch64_simd_ci __o;
29202   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
29203   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
29204   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
29205   __builtin_aarch64_st3v4si ((__builtin_aarch64_simd_si *) __a, __o);
29206 }
29207
29208 __extension__ extern __inline void
29209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29210 vst3q_u64 (uint64_t * __a, uint64x2x3_t val)
29211 {
29212   __builtin_aarch64_simd_ci __o;
29213   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
29214   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
29215   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
29216   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
29217 }
29218
29219 __extension__ extern __inline void
29220 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29221 vst3q_f16 (float16_t * __a, float16x8x3_t val)
29222 {
29223   __builtin_aarch64_simd_ci __o;
29224   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[0], 0);
29225   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[1], 1);
29226   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[2], 2);
29227   __builtin_aarch64_st3v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
29228 }
29229
29230 __extension__ extern __inline void
29231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29232 vst3q_f32 (float32_t * __a, float32x4x3_t val)
29233 {
29234   __builtin_aarch64_simd_ci __o;
29235   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[0], 0);
29236   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[1], 1);
29237   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[2], 2);
29238   __builtin_aarch64_st3v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
29239 }
29240
29241 __extension__ extern __inline void
29242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29243 vst3q_f64 (float64_t * __a, float64x2x3_t val)
29244 {
29245   __builtin_aarch64_simd_ci __o;
29246   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[0], 0);
29247   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[1], 1);
29248   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[2], 2);
29249   __builtin_aarch64_st3v2df ((__builtin_aarch64_simd_df *) __a, __o);
29250 }
29251
29252 __extension__ extern __inline void
29253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29254 vst3q_p64 (poly64_t * __a, poly64x2x3_t val)
29255 {
29256   __builtin_aarch64_simd_ci __o;
29257   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
29258                                                (poly64x2_t) val.val[0], 0);
29259   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
29260                                                (poly64x2_t) val.val[1], 1);
29261   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
29262                                                (poly64x2_t) val.val[2], 2);
29263   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
29264 }
29265
29266 __extension__ extern __inline void
29267 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29268 vst4_s64 (int64_t * __a, int64x1x4_t val)
29269 {
29270   __builtin_aarch64_simd_xi __o;
29271   int64x2x4_t temp;
29272   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
29273   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
29274   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
29275   temp.val[3] = vcombine_s64 (val.val[3], vcreate_s64 (__AARCH64_INT64_C (0)));
29276   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[0], 0);
29277   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[1], 1);
29278   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[2], 2);
29279   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[3], 3);
29280   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
29281 }
29282
29283 __extension__ extern __inline void
29284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29285 vst4_u64 (uint64_t * __a, uint64x1x4_t val)
29286 {
29287   __builtin_aarch64_simd_xi __o;
29288   uint64x2x4_t temp;
29289   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
29290   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
29291   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
29292   temp.val[3] = vcombine_u64 (val.val[3], vcreate_u64 (__AARCH64_UINT64_C (0)));
29293   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[0], 0);
29294   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[1], 1);
29295   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[2], 2);
29296   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[3], 3);
29297   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
29298 }
29299
29300 __extension__ extern __inline void
29301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29302 vst4_f64 (float64_t * __a, float64x1x4_t val)
29303 {
29304   __builtin_aarch64_simd_xi __o;
29305   float64x2x4_t temp;
29306   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
29307   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
29308   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
29309   temp.val[3] = vcombine_f64 (val.val[3], vcreate_f64 (__AARCH64_UINT64_C (0)));
29310   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[0], 0);
29311   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[1], 1);
29312   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[2], 2);
29313   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[3], 3);
29314   __builtin_aarch64_st4df ((__builtin_aarch64_simd_df *) __a, __o);
29315 }
29316
29317 __extension__ extern __inline void
29318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29319 vst4_s8 (int8_t * __a, int8x8x4_t val)
29320 {
29321   __builtin_aarch64_simd_xi __o;
29322   int8x16x4_t temp;
29323   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
29324   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
29325   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
29326   temp.val[3] = vcombine_s8 (val.val[3], vcreate_s8 (__AARCH64_INT64_C (0)));
29327   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
29328   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
29329   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
29330   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
29331   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
29332 }
29333
29334 __extension__ extern __inline void
29335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29336 vst4_p8 (poly8_t * __a, poly8x8x4_t val)
29337 {
29338   __builtin_aarch64_simd_xi __o;
29339   poly8x16x4_t temp;
29340   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
29341   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
29342   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
29343   temp.val[3] = vcombine_p8 (val.val[3], vcreate_p8 (__AARCH64_UINT64_C (0)));
29344   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
29345   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
29346   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
29347   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
29348   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
29349 }
29350
29351 __extension__ extern __inline void
29352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29353 vst4_s16 (int16_t * __a, int16x4x4_t val)
29354 {
29355   __builtin_aarch64_simd_xi __o;
29356   int16x8x4_t temp;
29357   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
29358   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
29359   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
29360   temp.val[3] = vcombine_s16 (val.val[3], vcreate_s16 (__AARCH64_INT64_C (0)));
29361   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
29362   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
29363   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
29364   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
29365   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
29366 }
29367
29368 __extension__ extern __inline void
29369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29370 vst4_p16 (poly16_t * __a, poly16x4x4_t val)
29371 {
29372   __builtin_aarch64_simd_xi __o;
29373   poly16x8x4_t temp;
29374   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
29375   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
29376   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
29377   temp.val[3] = vcombine_p16 (val.val[3], vcreate_p16 (__AARCH64_UINT64_C (0)));
29378   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
29379   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
29380   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
29381   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
29382   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
29383 }
29384
29385 __extension__ extern __inline void
29386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29387 vst4_s32 (int32_t * __a, int32x2x4_t val)
29388 {
29389   __builtin_aarch64_simd_xi __o;
29390   int32x4x4_t temp;
29391   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
29392   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
29393   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
29394   temp.val[3] = vcombine_s32 (val.val[3], vcreate_s32 (__AARCH64_INT64_C (0)));
29395   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[0], 0);
29396   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[1], 1);
29397   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[2], 2);
29398   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[3], 3);
29399   __builtin_aarch64_st4v2si ((__builtin_aarch64_simd_si *) __a, __o);
29400 }
29401
29402 __extension__ extern __inline void
29403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29404 vst4_u8 (uint8_t * __a, uint8x8x4_t val)
29405 {
29406   __builtin_aarch64_simd_xi __o;
29407   uint8x16x4_t temp;
29408   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
29409   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
29410   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
29411   temp.val[3] = vcombine_u8 (val.val[3], vcreate_u8 (__AARCH64_UINT64_C (0)));
29412   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
29413   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
29414   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
29415   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
29416   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
29417 }
29418
29419 __extension__ extern __inline void
29420 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29421 vst4_u16 (uint16_t * __a, uint16x4x4_t val)
29422 {
29423   __builtin_aarch64_simd_xi __o;
29424   uint16x8x4_t temp;
29425   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
29426   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
29427   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
29428   temp.val[3] = vcombine_u16 (val.val[3], vcreate_u16 (__AARCH64_UINT64_C (0)));
29429   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
29430   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
29431   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
29432   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
29433   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
29434 }
29435
29436 __extension__ extern __inline void
29437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29438 vst4_u32 (uint32_t * __a, uint32x2x4_t val)
29439 {
29440   __builtin_aarch64_simd_xi __o;
29441   uint32x4x4_t temp;
29442   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
29443   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
29444   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
29445   temp.val[3] = vcombine_u32 (val.val[3], vcreate_u32 (__AARCH64_UINT64_C (0)));
29446   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[0], 0);
29447   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[1], 1);
29448   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[2], 2);
29449   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[3], 3);
29450   __builtin_aarch64_st4v2si ((__builtin_aarch64_simd_si *) __a, __o);
29451 }
29452
29453 __extension__ extern __inline void
29454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29455 vst4_f16 (float16_t * __a, float16x4x4_t val)
29456 {
29457   __builtin_aarch64_simd_xi __o;
29458   float16x8x4_t temp;
29459   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
29460   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
29461   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
29462   temp.val[3] = vcombine_f16 (val.val[3], vcreate_f16 (__AARCH64_UINT64_C (0)));
29463   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[0], 0);
29464   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[1], 1);
29465   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[2], 2);
29466   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[3], 3);
29467   __builtin_aarch64_st4v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
29468 }
29469
29470 __extension__ extern __inline void
29471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29472 vst4_f32 (float32_t * __a, float32x2x4_t val)
29473 {
29474   __builtin_aarch64_simd_xi __o;
29475   float32x4x4_t temp;
29476   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
29477   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
29478   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
29479   temp.val[3] = vcombine_f32 (val.val[3], vcreate_f32 (__AARCH64_UINT64_C (0)));
29480   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[0], 0);
29481   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[1], 1);
29482   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[2], 2);
29483   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[3], 3);
29484   __builtin_aarch64_st4v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
29485 }
29486
29487 __extension__ extern __inline void
29488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29489 vst4_p64 (poly64_t * __a, poly64x1x4_t val)
29490 {
29491   __builtin_aarch64_simd_xi __o;
29492   poly64x2x4_t temp;
29493   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
29494   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
29495   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
29496   temp.val[3] = vcombine_p64 (val.val[3], vcreate_p64 (__AARCH64_UINT64_C (0)));
29497   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29498                                                (poly64x2_t) temp.val[0], 0);
29499   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29500                                                (poly64x2_t) temp.val[1], 1);
29501   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29502                                                (poly64x2_t) temp.val[2], 2);
29503   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29504                                                (poly64x2_t) temp.val[3], 3);
29505   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
29506 }
29507
29508 __extension__ extern __inline void
29509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29510 vst4q_s8 (int8_t * __a, int8x16x4_t val)
29511 {
29512   __builtin_aarch64_simd_xi __o;
29513   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
29514   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
29515   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
29516   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
29517   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
29518 }
29519
29520 __extension__ extern __inline void
29521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29522 vst4q_p8 (poly8_t * __a, poly8x16x4_t val)
29523 {
29524   __builtin_aarch64_simd_xi __o;
29525   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
29526   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
29527   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
29528   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
29529   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
29530 }
29531
29532 __extension__ extern __inline void
29533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29534 vst4q_s16 (int16_t * __a, int16x8x4_t val)
29535 {
29536   __builtin_aarch64_simd_xi __o;
29537   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
29538   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
29539   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
29540   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
29541   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
29542 }
29543
29544 __extension__ extern __inline void
29545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29546 vst4q_p16 (poly16_t * __a, poly16x8x4_t val)
29547 {
29548   __builtin_aarch64_simd_xi __o;
29549   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
29550   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
29551   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
29552   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
29553   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
29554 }
29555
29556 __extension__ extern __inline void
29557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29558 vst4q_s32 (int32_t * __a, int32x4x4_t val)
29559 {
29560   __builtin_aarch64_simd_xi __o;
29561   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[0], 0);
29562   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[1], 1);
29563   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[2], 2);
29564   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[3], 3);
29565   __builtin_aarch64_st4v4si ((__builtin_aarch64_simd_si *) __a, __o);
29566 }
29567
29568 __extension__ extern __inline void
29569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29570 vst4q_s64 (int64_t * __a, int64x2x4_t val)
29571 {
29572   __builtin_aarch64_simd_xi __o;
29573   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[0], 0);
29574   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[1], 1);
29575   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[2], 2);
29576   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[3], 3);
29577   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
29578 }
29579
29580 __extension__ extern __inline void
29581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29582 vst4q_u8 (uint8_t * __a, uint8x16x4_t val)
29583 {
29584   __builtin_aarch64_simd_xi __o;
29585   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
29586   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
29587   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
29588   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
29589   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
29590 }
29591
29592 __extension__ extern __inline void
29593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29594 vst4q_u16 (uint16_t * __a, uint16x8x4_t val)
29595 {
29596   __builtin_aarch64_simd_xi __o;
29597   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
29598   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
29599   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
29600   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
29601   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
29602 }
29603
29604 __extension__ extern __inline void
29605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29606 vst4q_u32 (uint32_t * __a, uint32x4x4_t val)
29607 {
29608   __builtin_aarch64_simd_xi __o;
29609   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[0], 0);
29610   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[1], 1);
29611   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[2], 2);
29612   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[3], 3);
29613   __builtin_aarch64_st4v4si ((__builtin_aarch64_simd_si *) __a, __o);
29614 }
29615
29616 __extension__ extern __inline void
29617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29618 vst4q_u64 (uint64_t * __a, uint64x2x4_t val)
29619 {
29620   __builtin_aarch64_simd_xi __o;
29621   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[0], 0);
29622   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[1], 1);
29623   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[2], 2);
29624   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[3], 3);
29625   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
29626 }
29627
29628 __extension__ extern __inline void
29629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29630 vst4q_f16 (float16_t * __a, float16x8x4_t val)
29631 {
29632   __builtin_aarch64_simd_xi __o;
29633   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[0], 0);
29634   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[1], 1);
29635   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[2], 2);
29636   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[3], 3);
29637   __builtin_aarch64_st4v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
29638 }
29639
29640 __extension__ extern __inline void
29641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29642 vst4q_f32 (float32_t * __a, float32x4x4_t val)
29643 {
29644   __builtin_aarch64_simd_xi __o;
29645   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[0], 0);
29646   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[1], 1);
29647   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[2], 2);
29648   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[3], 3);
29649   __builtin_aarch64_st4v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
29650 }
29651
29652 __extension__ extern __inline void
29653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29654 vst4q_f64 (float64_t * __a, float64x2x4_t val)
29655 {
29656   __builtin_aarch64_simd_xi __o;
29657   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[0], 0);
29658   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[1], 1);
29659   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[2], 2);
29660   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[3], 3);
29661   __builtin_aarch64_st4v2df ((__builtin_aarch64_simd_df *) __a, __o);
29662 }
29663
29664 __extension__ extern __inline void
29665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29666 vst4q_p64 (poly64_t * __a, poly64x2x4_t val)
29667 {
29668   __builtin_aarch64_simd_xi __o;
29669   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29670                                                (poly64x2_t) val.val[0], 0);
29671   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29672                                                (poly64x2_t) val.val[1], 1);
29673   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29674                                                (poly64x2_t) val.val[2], 2);
29675   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
29676                                                (poly64x2_t) val.val[3], 3);
29677   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
29678 }
29679
29680 /* vsub */
29681
29682 __extension__ extern __inline int64_t
29683 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29684 vsubd_s64 (int64_t __a, int64_t __b)
29685 {
29686   return __a - __b;
29687 }
29688
29689 __extension__ extern __inline uint64_t
29690 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29691 vsubd_u64 (uint64_t __a, uint64_t __b)
29692 {
29693   return __a - __b;
29694 }
29695
29696 /* vtbx1  */
29697
29698 __extension__ extern __inline int8x8_t
29699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29700 vtbx1_s8 (int8x8_t __r, int8x8_t __tab, int8x8_t __idx)
29701 {
29702   uint8x8_t __mask = vclt_u8 (vreinterpret_u8_s8 (__idx),
29703                               vmov_n_u8 (8));
29704   int8x8_t __tbl = vtbl1_s8 (__tab, __idx);
29705
29706   return vbsl_s8 (__mask, __tbl, __r);
29707 }
29708
29709 __extension__ extern __inline uint8x8_t
29710 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29711 vtbx1_u8 (uint8x8_t __r, uint8x8_t __tab, uint8x8_t __idx)
29712 {
29713   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (8));
29714   uint8x8_t __tbl = vtbl1_u8 (__tab, __idx);
29715
29716   return vbsl_u8 (__mask, __tbl, __r);
29717 }
29718
29719 __extension__ extern __inline poly8x8_t
29720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29721 vtbx1_p8 (poly8x8_t __r, poly8x8_t __tab, uint8x8_t __idx)
29722 {
29723   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (8));
29724   poly8x8_t __tbl = vtbl1_p8 (__tab, __idx);
29725
29726   return vbsl_p8 (__mask, __tbl, __r);
29727 }
29728
29729 /* vtbx3  */
29730
29731 __extension__ extern __inline int8x8_t
29732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29733 vtbx3_s8 (int8x8_t __r, int8x8x3_t __tab, int8x8_t __idx)
29734 {
29735   uint8x8_t __mask = vclt_u8 (vreinterpret_u8_s8 (__idx),
29736                               vmov_n_u8 (24));
29737   int8x8_t __tbl = vtbl3_s8 (__tab, __idx);
29738
29739   return vbsl_s8 (__mask, __tbl, __r);
29740 }
29741
29742 __extension__ extern __inline uint8x8_t
29743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29744 vtbx3_u8 (uint8x8_t __r, uint8x8x3_t __tab, uint8x8_t __idx)
29745 {
29746   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (24));
29747   uint8x8_t __tbl = vtbl3_u8 (__tab, __idx);
29748
29749   return vbsl_u8 (__mask, __tbl, __r);
29750 }
29751
29752 __extension__ extern __inline poly8x8_t
29753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29754 vtbx3_p8 (poly8x8_t __r, poly8x8x3_t __tab, uint8x8_t __idx)
29755 {
29756   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (24));
29757   poly8x8_t __tbl = vtbl3_p8 (__tab, __idx);
29758
29759   return vbsl_p8 (__mask, __tbl, __r);
29760 }
29761
29762 /* vtbx4  */
29763
29764 __extension__ extern __inline int8x8_t
29765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29766 vtbx4_s8 (int8x8_t __r, int8x8x4_t __tab, int8x8_t __idx)
29767 {
29768   int8x8_t result;
29769   int8x16x2_t temp;
29770   __builtin_aarch64_simd_oi __o;
29771   temp.val[0] = vcombine_s8 (__tab.val[0], __tab.val[1]);
29772   temp.val[1] = vcombine_s8 (__tab.val[2], __tab.val[3]);
29773   __o = __builtin_aarch64_set_qregoiv16qi (__o,
29774                                            (int8x16_t) temp.val[0], 0);
29775   __o = __builtin_aarch64_set_qregoiv16qi (__o,
29776                                            (int8x16_t) temp.val[1], 1);
29777   result = __builtin_aarch64_tbx4v8qi (__r, __o, __idx);
29778   return result;
29779 }
29780
29781 __extension__ extern __inline uint8x8_t
29782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29783 vtbx4_u8 (uint8x8_t __r, uint8x8x4_t __tab, uint8x8_t __idx)
29784 {
29785   uint8x8_t result;
29786   uint8x16x2_t temp;
29787   __builtin_aarch64_simd_oi __o;
29788   temp.val[0] = vcombine_u8 (__tab.val[0], __tab.val[1]);
29789   temp.val[1] = vcombine_u8 (__tab.val[2], __tab.val[3]);
29790   __o = __builtin_aarch64_set_qregoiv16qi (__o,
29791                                            (int8x16_t) temp.val[0], 0);
29792   __o = __builtin_aarch64_set_qregoiv16qi (__o,
29793                                            (int8x16_t) temp.val[1], 1);
29794   result = (uint8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)__r, __o,
29795                                                   (int8x8_t)__idx);
29796   return result;
29797 }
29798
29799 __extension__ extern __inline poly8x8_t
29800 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29801 vtbx4_p8 (poly8x8_t __r, poly8x8x4_t __tab, uint8x8_t __idx)
29802 {
29803   poly8x8_t result;
29804   poly8x16x2_t temp;
29805   __builtin_aarch64_simd_oi __o;
29806   temp.val[0] = vcombine_p8 (__tab.val[0], __tab.val[1]);
29807   temp.val[1] = vcombine_p8 (__tab.val[2], __tab.val[3]);
29808   __o = __builtin_aarch64_set_qregoiv16qi (__o,
29809                                            (int8x16_t) temp.val[0], 0);
29810   __o = __builtin_aarch64_set_qregoiv16qi (__o,
29811                                            (int8x16_t) temp.val[1], 1);
29812   result = (poly8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)__r, __o,
29813                                                   (int8x8_t)__idx);
29814   return result;
29815 }
29816
29817 /* vtrn */
29818
29819 __extension__ extern __inline float16x4_t
29820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29821 vtrn1_f16 (float16x4_t __a, float16x4_t __b)
29822 {
29823 #ifdef __AARCH64EB__
29824   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
29825 #else
29826   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
29827 #endif
29828 }
29829
29830 __extension__ extern __inline float32x2_t
29831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29832 vtrn1_f32 (float32x2_t __a, float32x2_t __b)
29833 {
29834 #ifdef __AARCH64EB__
29835   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29836 #else
29837   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29838 #endif
29839 }
29840
29841 __extension__ extern __inline poly8x8_t
29842 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29843 vtrn1_p8 (poly8x8_t __a, poly8x8_t __b)
29844 {
29845 #ifdef __AARCH64EB__
29846   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
29847 #else
29848   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
29849 #endif
29850 }
29851
29852 __extension__ extern __inline poly16x4_t
29853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29854 vtrn1_p16 (poly16x4_t __a, poly16x4_t __b)
29855 {
29856 #ifdef __AARCH64EB__
29857   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
29858 #else
29859   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
29860 #endif
29861 }
29862
29863 __extension__ extern __inline int8x8_t
29864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29865 vtrn1_s8 (int8x8_t __a, int8x8_t __b)
29866 {
29867 #ifdef __AARCH64EB__
29868   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
29869 #else
29870   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
29871 #endif
29872 }
29873
29874 __extension__ extern __inline int16x4_t
29875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29876 vtrn1_s16 (int16x4_t __a, int16x4_t __b)
29877 {
29878 #ifdef __AARCH64EB__
29879   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
29880 #else
29881   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
29882 #endif
29883 }
29884
29885 __extension__ extern __inline int32x2_t
29886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29887 vtrn1_s32 (int32x2_t __a, int32x2_t __b)
29888 {
29889 #ifdef __AARCH64EB__
29890   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29891 #else
29892   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29893 #endif
29894 }
29895
29896 __extension__ extern __inline uint8x8_t
29897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29898 vtrn1_u8 (uint8x8_t __a, uint8x8_t __b)
29899 {
29900 #ifdef __AARCH64EB__
29901   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
29902 #else
29903   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
29904 #endif
29905 }
29906
29907 __extension__ extern __inline uint16x4_t
29908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29909 vtrn1_u16 (uint16x4_t __a, uint16x4_t __b)
29910 {
29911 #ifdef __AARCH64EB__
29912   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
29913 #else
29914   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
29915 #endif
29916 }
29917
29918 __extension__ extern __inline uint32x2_t
29919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29920 vtrn1_u32 (uint32x2_t __a, uint32x2_t __b)
29921 {
29922 #ifdef __AARCH64EB__
29923   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29924 #else
29925   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29926 #endif
29927 }
29928
29929 __extension__ extern __inline float16x8_t
29930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29931 vtrn1q_f16 (float16x8_t __a, float16x8_t __b)
29932 {
29933 #ifdef __AARCH64EB__
29934   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
29935 #else
29936   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
29937 #endif
29938 }
29939
29940 __extension__ extern __inline float32x4_t
29941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29942 vtrn1q_f32 (float32x4_t __a, float32x4_t __b)
29943 {
29944 #ifdef __AARCH64EB__
29945   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
29946 #else
29947   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
29948 #endif
29949 }
29950
29951 __extension__ extern __inline float64x2_t
29952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29953 vtrn1q_f64 (float64x2_t __a, float64x2_t __b)
29954 {
29955 #ifdef __AARCH64EB__
29956   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29957 #else
29958   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29959 #endif
29960 }
29961
29962 __extension__ extern __inline poly8x16_t
29963 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29964 vtrn1q_p8 (poly8x16_t __a, poly8x16_t __b)
29965 {
29966 #ifdef __AARCH64EB__
29967   return __builtin_shuffle (__a, __b,
29968       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
29969 #else
29970   return __builtin_shuffle (__a, __b,
29971       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
29972 #endif
29973 }
29974
29975 __extension__ extern __inline poly16x8_t
29976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29977 vtrn1q_p16 (poly16x8_t __a, poly16x8_t __b)
29978 {
29979 #ifdef __AARCH64EB__
29980   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
29981 #else
29982   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
29983 #endif
29984 }
29985
29986 __extension__ extern __inline int8x16_t
29987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29988 vtrn1q_s8 (int8x16_t __a, int8x16_t __b)
29989 {
29990 #ifdef __AARCH64EB__
29991   return __builtin_shuffle (__a, __b,
29992       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
29993 #else
29994   return __builtin_shuffle (__a, __b,
29995       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
29996 #endif
29997 }
29998
29999 __extension__ extern __inline int16x8_t
30000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30001 vtrn1q_s16 (int16x8_t __a, int16x8_t __b)
30002 {
30003 #ifdef __AARCH64EB__
30004   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
30005 #else
30006   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
30007 #endif
30008 }
30009
30010 __extension__ extern __inline int32x4_t
30011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30012 vtrn1q_s32 (int32x4_t __a, int32x4_t __b)
30013 {
30014 #ifdef __AARCH64EB__
30015   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
30016 #else
30017   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
30018 #endif
30019 }
30020
30021 __extension__ extern __inline int64x2_t
30022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30023 vtrn1q_s64 (int64x2_t __a, int64x2_t __b)
30024 {
30025 #ifdef __AARCH64EB__
30026   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
30027 #else
30028   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
30029 #endif
30030 }
30031
30032 __extension__ extern __inline uint8x16_t
30033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30034 vtrn1q_u8 (uint8x16_t __a, uint8x16_t __b)
30035 {
30036 #ifdef __AARCH64EB__
30037   return __builtin_shuffle (__a, __b,
30038       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
30039 #else
30040   return __builtin_shuffle (__a, __b,
30041       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
30042 #endif
30043 }
30044
30045 __extension__ extern __inline uint16x8_t
30046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30047 vtrn1q_u16 (uint16x8_t __a, uint16x8_t __b)
30048 {
30049 #ifdef __AARCH64EB__
30050   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
30051 #else
30052   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
30053 #endif
30054 }
30055
30056 __extension__ extern __inline uint32x4_t
30057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30058 vtrn1q_u32 (uint32x4_t __a, uint32x4_t __b)
30059 {
30060 #ifdef __AARCH64EB__
30061   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
30062 #else
30063   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
30064 #endif
30065 }
30066
30067 __extension__ extern __inline uint64x2_t
30068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30069 vtrn1q_u64 (uint64x2_t __a, uint64x2_t __b)
30070 {
30071 #ifdef __AARCH64EB__
30072   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
30073 #else
30074   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
30075 #endif
30076 }
30077
30078 __extension__ extern __inline float16x4_t
30079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30080 vtrn2_f16 (float16x4_t __a, float16x4_t __b)
30081 {
30082 #ifdef __AARCH64EB__
30083   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
30084 #else
30085   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
30086 #endif
30087 }
30088
30089 __extension__ extern __inline float32x2_t
30090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30091 vtrn2_f32 (float32x2_t __a, float32x2_t __b)
30092 {
30093 #ifdef __AARCH64EB__
30094   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
30095 #else
30096   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
30097 #endif
30098 }
30099
30100 __extension__ extern __inline poly8x8_t
30101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30102 vtrn2_p8 (poly8x8_t __a, poly8x8_t __b)
30103 {
30104 #ifdef __AARCH64EB__
30105   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30106 #else
30107   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30108 #endif
30109 }
30110
30111 __extension__ extern __inline poly16x4_t
30112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30113 vtrn2_p16 (poly16x4_t __a, poly16x4_t __b)
30114 {
30115 #ifdef __AARCH64EB__
30116   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
30117 #else
30118   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
30119 #endif
30120 }
30121
30122 __extension__ extern __inline int8x8_t
30123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30124 vtrn2_s8 (int8x8_t __a, int8x8_t __b)
30125 {
30126 #ifdef __AARCH64EB__
30127   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30128 #else
30129   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30130 #endif
30131 }
30132
30133 __extension__ extern __inline int16x4_t
30134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30135 vtrn2_s16 (int16x4_t __a, int16x4_t __b)
30136 {
30137 #ifdef __AARCH64EB__
30138   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
30139 #else
30140   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
30141 #endif
30142 }
30143
30144 __extension__ extern __inline int32x2_t
30145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30146 vtrn2_s32 (int32x2_t __a, int32x2_t __b)
30147 {
30148 #ifdef __AARCH64EB__
30149   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
30150 #else
30151   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
30152 #endif
30153 }
30154
30155 __extension__ extern __inline uint8x8_t
30156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30157 vtrn2_u8 (uint8x8_t __a, uint8x8_t __b)
30158 {
30159 #ifdef __AARCH64EB__
30160   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30161 #else
30162   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30163 #endif
30164 }
30165
30166 __extension__ extern __inline uint16x4_t
30167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30168 vtrn2_u16 (uint16x4_t __a, uint16x4_t __b)
30169 {
30170 #ifdef __AARCH64EB__
30171   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
30172 #else
30173   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
30174 #endif
30175 }
30176
30177 __extension__ extern __inline uint32x2_t
30178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30179 vtrn2_u32 (uint32x2_t __a, uint32x2_t __b)
30180 {
30181 #ifdef __AARCH64EB__
30182   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
30183 #else
30184   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
30185 #endif
30186 }
30187
30188 __extension__ extern __inline float16x8_t
30189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30190 vtrn2q_f16 (float16x8_t __a, float16x8_t __b)
30191 {
30192 #ifdef __AARCH64EB__
30193   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30194 #else
30195   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30196 #endif
30197 }
30198
30199 __extension__ extern __inline float32x4_t
30200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30201 vtrn2q_f32 (float32x4_t __a, float32x4_t __b)
30202 {
30203 #ifdef __AARCH64EB__
30204   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
30205 #else
30206   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
30207 #endif
30208 }
30209
30210 __extension__ extern __inline float64x2_t
30211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30212 vtrn2q_f64 (float64x2_t __a, float64x2_t __b)
30213 {
30214 #ifdef __AARCH64EB__
30215   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30216 #else
30217   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30218 #endif
30219 }
30220
30221 __extension__ extern __inline poly8x16_t
30222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30223 vtrn2q_p8 (poly8x16_t __a, poly8x16_t __b)
30224 {
30225 #ifdef __AARCH64EB__
30226   return __builtin_shuffle (__a, __b,
30227       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
30228 #else
30229   return __builtin_shuffle (__a, __b,
30230       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
30231 #endif
30232 }
30233
30234 __extension__ extern __inline poly16x8_t
30235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30236 vtrn2q_p16 (poly16x8_t __a, poly16x8_t __b)
30237 {
30238 #ifdef __AARCH64EB__
30239   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30240 #else
30241   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30242 #endif
30243 }
30244
30245 __extension__ extern __inline int8x16_t
30246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30247 vtrn2q_s8 (int8x16_t __a, int8x16_t __b)
30248 {
30249 #ifdef __AARCH64EB__
30250   return __builtin_shuffle (__a, __b,
30251       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
30252 #else
30253   return __builtin_shuffle (__a, __b,
30254       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
30255 #endif
30256 }
30257
30258 __extension__ extern __inline int16x8_t
30259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30260 vtrn2q_s16 (int16x8_t __a, int16x8_t __b)
30261 {
30262 #ifdef __AARCH64EB__
30263   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30264 #else
30265   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30266 #endif
30267 }
30268
30269 __extension__ extern __inline int32x4_t
30270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30271 vtrn2q_s32 (int32x4_t __a, int32x4_t __b)
30272 {
30273 #ifdef __AARCH64EB__
30274   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
30275 #else
30276   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
30277 #endif
30278 }
30279
30280 __extension__ extern __inline int64x2_t
30281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30282 vtrn2q_s64 (int64x2_t __a, int64x2_t __b)
30283 {
30284 #ifdef __AARCH64EB__
30285   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30286 #else
30287   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30288 #endif
30289 }
30290
30291 __extension__ extern __inline uint8x16_t
30292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30293 vtrn2q_u8 (uint8x16_t __a, uint8x16_t __b)
30294 {
30295 #ifdef __AARCH64EB__
30296   return __builtin_shuffle (__a, __b,
30297       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
30298 #else
30299   return __builtin_shuffle (__a, __b,
30300       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
30301 #endif
30302 }
30303
30304 __extension__ extern __inline uint16x8_t
30305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30306 vtrn2q_u16 (uint16x8_t __a, uint16x8_t __b)
30307 {
30308 #ifdef __AARCH64EB__
30309   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
30310 #else
30311   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
30312 #endif
30313 }
30314
30315 __extension__ extern __inline uint32x4_t
30316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30317 vtrn2q_u32 (uint32x4_t __a, uint32x4_t __b)
30318 {
30319 #ifdef __AARCH64EB__
30320   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
30321 #else
30322   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
30323 #endif
30324 }
30325
30326 __extension__ extern __inline uint64x2_t
30327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30328 vtrn2q_u64 (uint64x2_t __a, uint64x2_t __b)
30329 {
30330 #ifdef __AARCH64EB__
30331   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30332 #else
30333   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30334 #endif
30335 }
30336
30337 __extension__ extern __inline float16x4x2_t
30338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30339 vtrn_f16 (float16x4_t __a, float16x4_t __b)
30340 {
30341   return (float16x4x2_t) {vtrn1_f16 (__a, __b), vtrn2_f16 (__a, __b)};
30342 }
30343
30344 __extension__ extern __inline float32x2x2_t
30345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30346 vtrn_f32 (float32x2_t a, float32x2_t b)
30347 {
30348   return (float32x2x2_t) {vtrn1_f32 (a, b), vtrn2_f32 (a, b)};
30349 }
30350
30351 __extension__ extern __inline poly8x8x2_t
30352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30353 vtrn_p8 (poly8x8_t a, poly8x8_t b)
30354 {
30355   return (poly8x8x2_t) {vtrn1_p8 (a, b), vtrn2_p8 (a, b)};
30356 }
30357
30358 __extension__ extern __inline poly16x4x2_t
30359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30360 vtrn_p16 (poly16x4_t a, poly16x4_t b)
30361 {
30362   return (poly16x4x2_t) {vtrn1_p16 (a, b), vtrn2_p16 (a, b)};
30363 }
30364
30365 __extension__ extern __inline int8x8x2_t
30366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30367 vtrn_s8 (int8x8_t a, int8x8_t b)
30368 {
30369   return (int8x8x2_t) {vtrn1_s8 (a, b), vtrn2_s8 (a, b)};
30370 }
30371
30372 __extension__ extern __inline int16x4x2_t
30373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30374 vtrn_s16 (int16x4_t a, int16x4_t b)
30375 {
30376   return (int16x4x2_t) {vtrn1_s16 (a, b), vtrn2_s16 (a, b)};
30377 }
30378
30379 __extension__ extern __inline int32x2x2_t
30380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30381 vtrn_s32 (int32x2_t a, int32x2_t b)
30382 {
30383   return (int32x2x2_t) {vtrn1_s32 (a, b), vtrn2_s32 (a, b)};
30384 }
30385
30386 __extension__ extern __inline uint8x8x2_t
30387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30388 vtrn_u8 (uint8x8_t a, uint8x8_t b)
30389 {
30390   return (uint8x8x2_t) {vtrn1_u8 (a, b), vtrn2_u8 (a, b)};
30391 }
30392
30393 __extension__ extern __inline uint16x4x2_t
30394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30395 vtrn_u16 (uint16x4_t a, uint16x4_t b)
30396 {
30397   return (uint16x4x2_t) {vtrn1_u16 (a, b), vtrn2_u16 (a, b)};
30398 }
30399
30400 __extension__ extern __inline uint32x2x2_t
30401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30402 vtrn_u32 (uint32x2_t a, uint32x2_t b)
30403 {
30404   return (uint32x2x2_t) {vtrn1_u32 (a, b), vtrn2_u32 (a, b)};
30405 }
30406
30407 __extension__ extern __inline float16x8x2_t
30408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30409 vtrnq_f16 (float16x8_t __a, float16x8_t __b)
30410 {
30411   return (float16x8x2_t) {vtrn1q_f16 (__a, __b), vtrn2q_f16 (__a, __b)};
30412 }
30413
30414 __extension__ extern __inline float32x4x2_t
30415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30416 vtrnq_f32 (float32x4_t a, float32x4_t b)
30417 {
30418   return (float32x4x2_t) {vtrn1q_f32 (a, b), vtrn2q_f32 (a, b)};
30419 }
30420
30421 __extension__ extern __inline poly8x16x2_t
30422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30423 vtrnq_p8 (poly8x16_t a, poly8x16_t b)
30424 {
30425   return (poly8x16x2_t) {vtrn1q_p8 (a, b), vtrn2q_p8 (a, b)};
30426 }
30427
30428 __extension__ extern __inline poly16x8x2_t
30429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30430 vtrnq_p16 (poly16x8_t a, poly16x8_t b)
30431 {
30432   return (poly16x8x2_t) {vtrn1q_p16 (a, b), vtrn2q_p16 (a, b)};
30433 }
30434
30435 __extension__ extern __inline int8x16x2_t
30436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30437 vtrnq_s8 (int8x16_t a, int8x16_t b)
30438 {
30439   return (int8x16x2_t) {vtrn1q_s8 (a, b), vtrn2q_s8 (a, b)};
30440 }
30441
30442 __extension__ extern __inline int16x8x2_t
30443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30444 vtrnq_s16 (int16x8_t a, int16x8_t b)
30445 {
30446   return (int16x8x2_t) {vtrn1q_s16 (a, b), vtrn2q_s16 (a, b)};
30447 }
30448
30449 __extension__ extern __inline int32x4x2_t
30450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30451 vtrnq_s32 (int32x4_t a, int32x4_t b)
30452 {
30453   return (int32x4x2_t) {vtrn1q_s32 (a, b), vtrn2q_s32 (a, b)};
30454 }
30455
30456 __extension__ extern __inline uint8x16x2_t
30457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30458 vtrnq_u8 (uint8x16_t a, uint8x16_t b)
30459 {
30460   return (uint8x16x2_t) {vtrn1q_u8 (a, b), vtrn2q_u8 (a, b)};
30461 }
30462
30463 __extension__ extern __inline uint16x8x2_t
30464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30465 vtrnq_u16 (uint16x8_t a, uint16x8_t b)
30466 {
30467   return (uint16x8x2_t) {vtrn1q_u16 (a, b), vtrn2q_u16 (a, b)};
30468 }
30469
30470 __extension__ extern __inline uint32x4x2_t
30471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30472 vtrnq_u32 (uint32x4_t a, uint32x4_t b)
30473 {
30474   return (uint32x4x2_t) {vtrn1q_u32 (a, b), vtrn2q_u32 (a, b)};
30475 }
30476
30477 /* vtst */
30478
30479 __extension__ extern __inline uint8x8_t
30480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30481 vtst_s8 (int8x8_t __a, int8x8_t __b)
30482 {
30483   return (uint8x8_t) ((__a & __b) != 0);
30484 }
30485
30486 __extension__ extern __inline uint16x4_t
30487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30488 vtst_s16 (int16x4_t __a, int16x4_t __b)
30489 {
30490   return (uint16x4_t) ((__a & __b) != 0);
30491 }
30492
30493 __extension__ extern __inline uint32x2_t
30494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30495 vtst_s32 (int32x2_t __a, int32x2_t __b)
30496 {
30497   return (uint32x2_t) ((__a & __b) != 0);
30498 }
30499
30500 __extension__ extern __inline uint64x1_t
30501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30502 vtst_s64 (int64x1_t __a, int64x1_t __b)
30503 {
30504   return (uint64x1_t) ((__a & __b) != __AARCH64_INT64_C (0));
30505 }
30506
30507 __extension__ extern __inline uint8x8_t
30508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30509 vtst_u8 (uint8x8_t __a, uint8x8_t __b)
30510 {
30511   return ((__a & __b) != 0);
30512 }
30513
30514 __extension__ extern __inline uint16x4_t
30515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30516 vtst_u16 (uint16x4_t __a, uint16x4_t __b)
30517 {
30518   return ((__a & __b) != 0);
30519 }
30520
30521 __extension__ extern __inline uint32x2_t
30522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30523 vtst_u32 (uint32x2_t __a, uint32x2_t __b)
30524 {
30525   return ((__a & __b) != 0);
30526 }
30527
30528 __extension__ extern __inline uint64x1_t
30529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30530 vtst_u64 (uint64x1_t __a, uint64x1_t __b)
30531 {
30532   return ((__a & __b) != __AARCH64_UINT64_C (0));
30533 }
30534
30535 __extension__ extern __inline uint8x16_t
30536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30537 vtstq_s8 (int8x16_t __a, int8x16_t __b)
30538 {
30539   return (uint8x16_t) ((__a & __b) != 0);
30540 }
30541
30542 __extension__ extern __inline uint16x8_t
30543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30544 vtstq_s16 (int16x8_t __a, int16x8_t __b)
30545 {
30546   return (uint16x8_t) ((__a & __b) != 0);
30547 }
30548
30549 __extension__ extern __inline uint32x4_t
30550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30551 vtstq_s32 (int32x4_t __a, int32x4_t __b)
30552 {
30553   return (uint32x4_t) ((__a & __b) != 0);
30554 }
30555
30556 __extension__ extern __inline uint64x2_t
30557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30558 vtstq_s64 (int64x2_t __a, int64x2_t __b)
30559 {
30560   return (uint64x2_t) ((__a & __b) != __AARCH64_INT64_C (0));
30561 }
30562
30563 __extension__ extern __inline uint8x16_t
30564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30565 vtstq_u8 (uint8x16_t __a, uint8x16_t __b)
30566 {
30567   return ((__a & __b) != 0);
30568 }
30569
30570 __extension__ extern __inline uint16x8_t
30571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30572 vtstq_u16 (uint16x8_t __a, uint16x8_t __b)
30573 {
30574   return ((__a & __b) != 0);
30575 }
30576
30577 __extension__ extern __inline uint32x4_t
30578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30579 vtstq_u32 (uint32x4_t __a, uint32x4_t __b)
30580 {
30581   return ((__a & __b) != 0);
30582 }
30583
30584 __extension__ extern __inline uint64x2_t
30585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30586 vtstq_u64 (uint64x2_t __a, uint64x2_t __b)
30587 {
30588   return ((__a & __b) != __AARCH64_UINT64_C (0));
30589 }
30590
30591 __extension__ extern __inline uint64_t
30592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30593 vtstd_s64 (int64_t __a, int64_t __b)
30594 {
30595   return (__a & __b) ? -1ll : 0ll;
30596 }
30597
30598 __extension__ extern __inline uint64_t
30599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30600 vtstd_u64 (uint64_t __a, uint64_t __b)
30601 {
30602   return (__a & __b) ? -1ll : 0ll;
30603 }
30604
30605 /* vuqadd */
30606
30607 __extension__ extern __inline int8x8_t
30608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30609 vuqadd_s8 (int8x8_t __a, uint8x8_t __b)
30610 {
30611   return __builtin_aarch64_suqaddv8qi_ssu (__a,  __b);
30612 }
30613
30614 __extension__ extern __inline int16x4_t
30615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30616 vuqadd_s16 (int16x4_t __a, uint16x4_t __b)
30617 {
30618   return __builtin_aarch64_suqaddv4hi_ssu (__a,  __b);
30619 }
30620
30621 __extension__ extern __inline int32x2_t
30622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30623 vuqadd_s32 (int32x2_t __a, uint32x2_t __b)
30624 {
30625   return __builtin_aarch64_suqaddv2si_ssu (__a,  __b);
30626 }
30627
30628 __extension__ extern __inline int64x1_t
30629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30630 vuqadd_s64 (int64x1_t __a, uint64x1_t __b)
30631 {
30632   return (int64x1_t) {__builtin_aarch64_suqadddi_ssu (__a[0], __b[0])};
30633 }
30634
30635 __extension__ extern __inline int8x16_t
30636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30637 vuqaddq_s8 (int8x16_t __a, uint8x16_t __b)
30638 {
30639   return __builtin_aarch64_suqaddv16qi_ssu (__a,  __b);
30640 }
30641
30642 __extension__ extern __inline int16x8_t
30643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30644 vuqaddq_s16 (int16x8_t __a, uint16x8_t __b)
30645 {
30646   return __builtin_aarch64_suqaddv8hi_ssu (__a,  __b);
30647 }
30648
30649 __extension__ extern __inline int32x4_t
30650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30651 vuqaddq_s32 (int32x4_t __a, uint32x4_t __b)
30652 {
30653   return __builtin_aarch64_suqaddv4si_ssu (__a,  __b);
30654 }
30655
30656 __extension__ extern __inline int64x2_t
30657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30658 vuqaddq_s64 (int64x2_t __a, uint64x2_t __b)
30659 {
30660   return __builtin_aarch64_suqaddv2di_ssu (__a,  __b);
30661 }
30662
30663 __extension__ extern __inline int8_t
30664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30665 vuqaddb_s8 (int8_t __a, uint8_t __b)
30666 {
30667   return __builtin_aarch64_suqaddqi_ssu (__a,  __b);
30668 }
30669
30670 __extension__ extern __inline int16_t
30671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30672 vuqaddh_s16 (int16_t __a, uint16_t __b)
30673 {
30674   return __builtin_aarch64_suqaddhi_ssu (__a,  __b);
30675 }
30676
30677 __extension__ extern __inline int32_t
30678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30679 vuqadds_s32 (int32_t __a, uint32_t __b)
30680 {
30681   return __builtin_aarch64_suqaddsi_ssu (__a,  __b);
30682 }
30683
30684 __extension__ extern __inline int64_t
30685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30686 vuqaddd_s64 (int64_t __a, uint64_t __b)
30687 {
30688   return __builtin_aarch64_suqadddi_ssu (__a,  __b);
30689 }
30690
30691 #define __DEFINTERLEAVE(op, rettype, intype, funcsuffix, Q)             \
30692   __extension__ extern __inline rettype                                 \
30693   __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
30694   v ## op ## Q ## _ ## funcsuffix (intype a, intype b)                  \
30695   {                                                                     \
30696     return (rettype) {v ## op ## 1 ## Q ## _ ## funcsuffix (a, b),      \
30697                       v ## op ## 2 ## Q ## _ ## funcsuffix (a, b)};     \
30698   }
30699
30700 #define __INTERLEAVE_LIST(op)                                   \
30701   __DEFINTERLEAVE (op, float16x4x2_t, float16x4_t, f16,)        \
30702   __DEFINTERLEAVE (op, float32x2x2_t, float32x2_t, f32,)        \
30703   __DEFINTERLEAVE (op, poly8x8x2_t, poly8x8_t, p8,)             \
30704   __DEFINTERLEAVE (op, poly16x4x2_t, poly16x4_t, p16,)          \
30705   __DEFINTERLEAVE (op, int8x8x2_t, int8x8_t, s8,)               \
30706   __DEFINTERLEAVE (op, int16x4x2_t, int16x4_t, s16,)            \
30707   __DEFINTERLEAVE (op, int32x2x2_t, int32x2_t, s32,)            \
30708   __DEFINTERLEAVE (op, uint8x8x2_t, uint8x8_t, u8,)             \
30709   __DEFINTERLEAVE (op, uint16x4x2_t, uint16x4_t, u16,)          \
30710   __DEFINTERLEAVE (op, uint32x2x2_t, uint32x2_t, u32,)          \
30711   __DEFINTERLEAVE (op, float16x8x2_t, float16x8_t, f16, q)      \
30712   __DEFINTERLEAVE (op, float32x4x2_t, float32x4_t, f32, q)      \
30713   __DEFINTERLEAVE (op, poly8x16x2_t, poly8x16_t, p8, q)         \
30714   __DEFINTERLEAVE (op, poly16x8x2_t, poly16x8_t, p16, q)        \
30715   __DEFINTERLEAVE (op, int8x16x2_t, int8x16_t, s8, q)           \
30716   __DEFINTERLEAVE (op, int16x8x2_t, int16x8_t, s16, q)          \
30717   __DEFINTERLEAVE (op, int32x4x2_t, int32x4_t, s32, q)          \
30718   __DEFINTERLEAVE (op, uint8x16x2_t, uint8x16_t, u8, q)         \
30719   __DEFINTERLEAVE (op, uint16x8x2_t, uint16x8_t, u16, q)        \
30720   __DEFINTERLEAVE (op, uint32x4x2_t, uint32x4_t, u32, q)
30721
30722 /* vuzp */
30723
30724 __extension__ extern __inline float16x4_t
30725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30726 vuzp1_f16 (float16x4_t __a, float16x4_t __b)
30727 {
30728 #ifdef __AARCH64EB__
30729   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
30730 #else
30731   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
30732 #endif
30733 }
30734
30735 __extension__ extern __inline float32x2_t
30736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30737 vuzp1_f32 (float32x2_t __a, float32x2_t __b)
30738 {
30739 #ifdef __AARCH64EB__
30740   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
30741 #else
30742   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
30743 #endif
30744 }
30745
30746 __extension__ extern __inline poly8x8_t
30747 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30748 vuzp1_p8 (poly8x8_t __a, poly8x8_t __b)
30749 {
30750 #ifdef __AARCH64EB__
30751   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30752 #else
30753   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30754 #endif
30755 }
30756
30757 __extension__ extern __inline poly16x4_t
30758 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30759 vuzp1_p16 (poly16x4_t __a, poly16x4_t __b)
30760 {
30761 #ifdef __AARCH64EB__
30762   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
30763 #else
30764   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
30765 #endif
30766 }
30767
30768 __extension__ extern __inline int8x8_t
30769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30770 vuzp1_s8 (int8x8_t __a, int8x8_t __b)
30771 {
30772 #ifdef __AARCH64EB__
30773   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30774 #else
30775   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30776 #endif
30777 }
30778
30779 __extension__ extern __inline int16x4_t
30780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30781 vuzp1_s16 (int16x4_t __a, int16x4_t __b)
30782 {
30783 #ifdef __AARCH64EB__
30784   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
30785 #else
30786   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
30787 #endif
30788 }
30789
30790 __extension__ extern __inline int32x2_t
30791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30792 vuzp1_s32 (int32x2_t __a, int32x2_t __b)
30793 {
30794 #ifdef __AARCH64EB__
30795   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
30796 #else
30797   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
30798 #endif
30799 }
30800
30801 __extension__ extern __inline uint8x8_t
30802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30803 vuzp1_u8 (uint8x8_t __a, uint8x8_t __b)
30804 {
30805 #ifdef __AARCH64EB__
30806   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30807 #else
30808   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30809 #endif
30810 }
30811
30812 __extension__ extern __inline uint16x4_t
30813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30814 vuzp1_u16 (uint16x4_t __a, uint16x4_t __b)
30815 {
30816 #ifdef __AARCH64EB__
30817   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
30818 #else
30819   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
30820 #endif
30821 }
30822
30823 __extension__ extern __inline uint32x2_t
30824 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30825 vuzp1_u32 (uint32x2_t __a, uint32x2_t __b)
30826 {
30827 #ifdef __AARCH64EB__
30828   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
30829 #else
30830   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
30831 #endif
30832 }
30833
30834 __extension__ extern __inline float16x8_t
30835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30836 vuzp1q_f16 (float16x8_t __a, float16x8_t __b)
30837 {
30838 #ifdef __AARCH64EB__
30839   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30840 #else
30841   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30842 #endif
30843 }
30844
30845 __extension__ extern __inline float32x4_t
30846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30847 vuzp1q_f32 (float32x4_t __a, float32x4_t __b)
30848 {
30849 #ifdef __AARCH64EB__
30850   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
30851 #else
30852   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
30853 #endif
30854 }
30855
30856 __extension__ extern __inline float64x2_t
30857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30858 vuzp1q_f64 (float64x2_t __a, float64x2_t __b)
30859 {
30860 #ifdef __AARCH64EB__
30861   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
30862 #else
30863   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
30864 #endif
30865 }
30866
30867 __extension__ extern __inline poly8x16_t
30868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30869 vuzp1q_p8 (poly8x16_t __a, poly8x16_t __b)
30870 {
30871 #ifdef __AARCH64EB__
30872   return __builtin_shuffle (__a, __b, (uint8x16_t)
30873       {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
30874 #else
30875   return __builtin_shuffle (__a, __b, (uint8x16_t)
30876       {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
30877 #endif
30878 }
30879
30880 __extension__ extern __inline poly16x8_t
30881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30882 vuzp1q_p16 (poly16x8_t __a, poly16x8_t __b)
30883 {
30884 #ifdef __AARCH64EB__
30885   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30886 #else
30887   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30888 #endif
30889 }
30890
30891 __extension__ extern __inline int8x16_t
30892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30893 vuzp1q_s8 (int8x16_t __a, int8x16_t __b)
30894 {
30895 #ifdef __AARCH64EB__
30896   return __builtin_shuffle (__a, __b,
30897       (uint8x16_t) {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
30898 #else
30899   return __builtin_shuffle (__a, __b,
30900       (uint8x16_t) {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
30901 #endif
30902 }
30903
30904 __extension__ extern __inline int16x8_t
30905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30906 vuzp1q_s16 (int16x8_t __a, int16x8_t __b)
30907 {
30908 #ifdef __AARCH64EB__
30909   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30910 #else
30911   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30912 #endif
30913 }
30914
30915 __extension__ extern __inline int32x4_t
30916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30917 vuzp1q_s32 (int32x4_t __a, int32x4_t __b)
30918 {
30919 #ifdef __AARCH64EB__
30920   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
30921 #else
30922   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
30923 #endif
30924 }
30925
30926 __extension__ extern __inline int64x2_t
30927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30928 vuzp1q_s64 (int64x2_t __a, int64x2_t __b)
30929 {
30930 #ifdef __AARCH64EB__
30931   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
30932 #else
30933   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
30934 #endif
30935 }
30936
30937 __extension__ extern __inline uint8x16_t
30938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30939 vuzp1q_u8 (uint8x16_t __a, uint8x16_t __b)
30940 {
30941 #ifdef __AARCH64EB__
30942   return __builtin_shuffle (__a, __b,
30943       (uint8x16_t) {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
30944 #else
30945   return __builtin_shuffle (__a, __b,
30946       (uint8x16_t) {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
30947 #endif
30948 }
30949
30950 __extension__ extern __inline uint16x8_t
30951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30952 vuzp1q_u16 (uint16x8_t __a, uint16x8_t __b)
30953 {
30954 #ifdef __AARCH64EB__
30955   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
30956 #else
30957   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
30958 #endif
30959 }
30960
30961 __extension__ extern __inline uint32x4_t
30962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30963 vuzp1q_u32 (uint32x4_t __a, uint32x4_t __b)
30964 {
30965 #ifdef __AARCH64EB__
30966   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
30967 #else
30968   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
30969 #endif
30970 }
30971
30972 __extension__ extern __inline uint64x2_t
30973 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30974 vuzp1q_u64 (uint64x2_t __a, uint64x2_t __b)
30975 {
30976 #ifdef __AARCH64EB__
30977   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
30978 #else
30979   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
30980 #endif
30981 }
30982
30983 __extension__ extern __inline float16x4_t
30984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30985 vuzp2_f16 (float16x4_t __a, float16x4_t __b)
30986 {
30987 #ifdef __AARCH64EB__
30988   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
30989 #else
30990   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
30991 #endif
30992 }
30993
30994 __extension__ extern __inline float32x2_t
30995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30996 vuzp2_f32 (float32x2_t __a, float32x2_t __b)
30997 {
30998 #ifdef __AARCH64EB__
30999   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
31000 #else
31001   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
31002 #endif
31003 }
31004
31005 __extension__ extern __inline poly8x8_t
31006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31007 vuzp2_p8 (poly8x8_t __a, poly8x8_t __b)
31008 {
31009 #ifdef __AARCH64EB__
31010   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31011 #else
31012   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31013 #endif
31014 }
31015
31016 __extension__ extern __inline poly16x4_t
31017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31018 vuzp2_p16 (poly16x4_t __a, poly16x4_t __b)
31019 {
31020 #ifdef __AARCH64EB__
31021   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
31022 #else
31023   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
31024 #endif
31025 }
31026
31027 __extension__ extern __inline int8x8_t
31028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31029 vuzp2_s8 (int8x8_t __a, int8x8_t __b)
31030 {
31031 #ifdef __AARCH64EB__
31032   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31033 #else
31034   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31035 #endif
31036 }
31037
31038 __extension__ extern __inline int16x4_t
31039 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31040 vuzp2_s16 (int16x4_t __a, int16x4_t __b)
31041 {
31042 #ifdef __AARCH64EB__
31043   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
31044 #else
31045   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
31046 #endif
31047 }
31048
31049 __extension__ extern __inline int32x2_t
31050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31051 vuzp2_s32 (int32x2_t __a, int32x2_t __b)
31052 {
31053 #ifdef __AARCH64EB__
31054   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
31055 #else
31056   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
31057 #endif
31058 }
31059
31060 __extension__ extern __inline uint8x8_t
31061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31062 vuzp2_u8 (uint8x8_t __a, uint8x8_t __b)
31063 {
31064 #ifdef __AARCH64EB__
31065   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31066 #else
31067   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31068 #endif
31069 }
31070
31071 __extension__ extern __inline uint16x4_t
31072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31073 vuzp2_u16 (uint16x4_t __a, uint16x4_t __b)
31074 {
31075 #ifdef __AARCH64EB__
31076   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
31077 #else
31078   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
31079 #endif
31080 }
31081
31082 __extension__ extern __inline uint32x2_t
31083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31084 vuzp2_u32 (uint32x2_t __a, uint32x2_t __b)
31085 {
31086 #ifdef __AARCH64EB__
31087   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
31088 #else
31089   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
31090 #endif
31091 }
31092
31093 __extension__ extern __inline float16x8_t
31094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31095 vuzp2q_f16 (float16x8_t __a, float16x8_t __b)
31096 {
31097 #ifdef __AARCH64EB__
31098   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31099 #else
31100   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31101 #endif
31102 }
31103
31104 __extension__ extern __inline float32x4_t
31105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31106 vuzp2q_f32 (float32x4_t __a, float32x4_t __b)
31107 {
31108 #ifdef __AARCH64EB__
31109   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
31110 #else
31111   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
31112 #endif
31113 }
31114
31115 __extension__ extern __inline float64x2_t
31116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31117 vuzp2q_f64 (float64x2_t __a, float64x2_t __b)
31118 {
31119 #ifdef __AARCH64EB__
31120   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
31121 #else
31122   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
31123 #endif
31124 }
31125
31126 __extension__ extern __inline poly8x16_t
31127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31128 vuzp2q_p8 (poly8x16_t __a, poly8x16_t __b)
31129 {
31130 #ifdef __AARCH64EB__
31131   return __builtin_shuffle (__a, __b,
31132       (uint8x16_t) {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
31133 #else
31134   return __builtin_shuffle (__a, __b,
31135       (uint8x16_t) {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
31136 #endif
31137 }
31138
31139 __extension__ extern __inline poly16x8_t
31140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31141 vuzp2q_p16 (poly16x8_t __a, poly16x8_t __b)
31142 {
31143 #ifdef __AARCH64EB__
31144   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31145 #else
31146   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31147 #endif
31148 }
31149
31150 __extension__ extern __inline int8x16_t
31151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31152 vuzp2q_s8 (int8x16_t __a, int8x16_t __b)
31153 {
31154 #ifdef __AARCH64EB__
31155   return __builtin_shuffle (__a, __b,
31156       (uint8x16_t) {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
31157 #else
31158   return __builtin_shuffle (__a, __b,
31159       (uint8x16_t) {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
31160 #endif
31161 }
31162
31163 __extension__ extern __inline int16x8_t
31164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31165 vuzp2q_s16 (int16x8_t __a, int16x8_t __b)
31166 {
31167 #ifdef __AARCH64EB__
31168   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31169 #else
31170   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31171 #endif
31172 }
31173
31174 __extension__ extern __inline int32x4_t
31175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31176 vuzp2q_s32 (int32x4_t __a, int32x4_t __b)
31177 {
31178 #ifdef __AARCH64EB__
31179   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
31180 #else
31181   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
31182 #endif
31183 }
31184
31185 __extension__ extern __inline int64x2_t
31186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31187 vuzp2q_s64 (int64x2_t __a, int64x2_t __b)
31188 {
31189 #ifdef __AARCH64EB__
31190   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
31191 #else
31192   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
31193 #endif
31194 }
31195
31196 __extension__ extern __inline uint8x16_t
31197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31198 vuzp2q_u8 (uint8x16_t __a, uint8x16_t __b)
31199 {
31200 #ifdef __AARCH64EB__
31201   return __builtin_shuffle (__a, __b, (uint8x16_t)
31202       {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
31203 #else
31204   return __builtin_shuffle (__a, __b, (uint8x16_t)
31205       {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
31206 #endif
31207 }
31208
31209 __extension__ extern __inline uint16x8_t
31210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31211 vuzp2q_u16 (uint16x8_t __a, uint16x8_t __b)
31212 {
31213 #ifdef __AARCH64EB__
31214   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
31215 #else
31216   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
31217 #endif
31218 }
31219
31220 __extension__ extern __inline uint32x4_t
31221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31222 vuzp2q_u32 (uint32x4_t __a, uint32x4_t __b)
31223 {
31224 #ifdef __AARCH64EB__
31225   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
31226 #else
31227   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
31228 #endif
31229 }
31230
31231 __extension__ extern __inline uint64x2_t
31232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31233 vuzp2q_u64 (uint64x2_t __a, uint64x2_t __b)
31234 {
31235 #ifdef __AARCH64EB__
31236   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
31237 #else
31238   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
31239 #endif
31240 }
31241
31242 __INTERLEAVE_LIST (uzp)
31243
31244 /* vzip */
31245
31246 __extension__ extern __inline float16x4_t
31247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31248 vzip1_f16 (float16x4_t __a, float16x4_t __b)
31249 {
31250 #ifdef __AARCH64EB__
31251   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
31252 #else
31253   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
31254 #endif
31255 }
31256
31257 __extension__ extern __inline float32x2_t
31258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31259 vzip1_f32 (float32x2_t __a, float32x2_t __b)
31260 {
31261 #ifdef __AARCH64EB__
31262   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
31263 #else
31264   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
31265 #endif
31266 }
31267
31268 __extension__ extern __inline poly8x8_t
31269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31270 vzip1_p8 (poly8x8_t __a, poly8x8_t __b)
31271 {
31272 #ifdef __AARCH64EB__
31273   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
31274 #else
31275   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31276 #endif
31277 }
31278
31279 __extension__ extern __inline poly16x4_t
31280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31281 vzip1_p16 (poly16x4_t __a, poly16x4_t __b)
31282 {
31283 #ifdef __AARCH64EB__
31284   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
31285 #else
31286   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
31287 #endif
31288 }
31289
31290 __extension__ extern __inline int8x8_t
31291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31292 vzip1_s8 (int8x8_t __a, int8x8_t __b)
31293 {
31294 #ifdef __AARCH64EB__
31295   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
31296 #else
31297   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31298 #endif
31299 }
31300
31301 __extension__ extern __inline int16x4_t
31302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31303 vzip1_s16 (int16x4_t __a, int16x4_t __b)
31304 {
31305 #ifdef __AARCH64EB__
31306   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
31307 #else
31308   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
31309 #endif
31310 }
31311
31312 __extension__ extern __inline int32x2_t
31313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31314 vzip1_s32 (int32x2_t __a, int32x2_t __b)
31315 {
31316 #ifdef __AARCH64EB__
31317   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
31318 #else
31319   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
31320 #endif
31321 }
31322
31323 __extension__ extern __inline uint8x8_t
31324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31325 vzip1_u8 (uint8x8_t __a, uint8x8_t __b)
31326 {
31327 #ifdef __AARCH64EB__
31328   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
31329 #else
31330   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31331 #endif
31332 }
31333
31334 __extension__ extern __inline uint16x4_t
31335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31336 vzip1_u16 (uint16x4_t __a, uint16x4_t __b)
31337 {
31338 #ifdef __AARCH64EB__
31339   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
31340 #else
31341   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
31342 #endif
31343 }
31344
31345 __extension__ extern __inline uint32x2_t
31346 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31347 vzip1_u32 (uint32x2_t __a, uint32x2_t __b)
31348 {
31349 #ifdef __AARCH64EB__
31350   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
31351 #else
31352   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
31353 #endif
31354 }
31355
31356 __extension__ extern __inline float16x8_t
31357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31358 vzip1q_f16 (float16x8_t __a, float16x8_t __b)
31359 {
31360 #ifdef __AARCH64EB__
31361   return __builtin_shuffle (__a, __b,
31362                             (uint16x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
31363 #else
31364   return __builtin_shuffle (__a, __b,
31365                             (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31366 #endif
31367 }
31368
31369 __extension__ extern __inline float32x4_t
31370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31371 vzip1q_f32 (float32x4_t __a, float32x4_t __b)
31372 {
31373 #ifdef __AARCH64EB__
31374   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
31375 #else
31376   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
31377 #endif
31378 }
31379
31380 __extension__ extern __inline float64x2_t
31381 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31382 vzip1q_f64 (float64x2_t __a, float64x2_t __b)
31383 {
31384 #ifdef __AARCH64EB__
31385   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
31386 #else
31387   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
31388 #endif
31389 }
31390
31391 __extension__ extern __inline poly8x16_t
31392 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31393 vzip1q_p8 (poly8x16_t __a, poly8x16_t __b)
31394 {
31395 #ifdef __AARCH64EB__
31396   return __builtin_shuffle (__a, __b, (uint8x16_t)
31397       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
31398 #else
31399   return __builtin_shuffle (__a, __b, (uint8x16_t)
31400       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
31401 #endif
31402 }
31403
31404 __extension__ extern __inline poly16x8_t
31405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31406 vzip1q_p16 (poly16x8_t __a, poly16x8_t __b)
31407 {
31408 #ifdef __AARCH64EB__
31409   return __builtin_shuffle (__a, __b, (uint16x8_t)
31410       {12, 4, 13, 5, 14, 6, 15, 7});
31411 #else
31412   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31413 #endif
31414 }
31415
31416 __extension__ extern __inline int8x16_t
31417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31418 vzip1q_s8 (int8x16_t __a, int8x16_t __b)
31419 {
31420 #ifdef __AARCH64EB__
31421   return __builtin_shuffle (__a, __b, (uint8x16_t)
31422       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
31423 #else
31424   return __builtin_shuffle (__a, __b, (uint8x16_t)
31425       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
31426 #endif
31427 }
31428
31429 __extension__ extern __inline int16x8_t
31430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31431 vzip1q_s16 (int16x8_t __a, int16x8_t __b)
31432 {
31433 #ifdef __AARCH64EB__
31434   return __builtin_shuffle (__a, __b, (uint16x8_t)
31435       {12, 4, 13, 5, 14, 6, 15, 7});
31436 #else
31437   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31438 #endif
31439 }
31440
31441 __extension__ extern __inline int32x4_t
31442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31443 vzip1q_s32 (int32x4_t __a, int32x4_t __b)
31444 {
31445 #ifdef __AARCH64EB__
31446   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
31447 #else
31448   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
31449 #endif
31450 }
31451
31452 __extension__ extern __inline int64x2_t
31453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31454 vzip1q_s64 (int64x2_t __a, int64x2_t __b)
31455 {
31456 #ifdef __AARCH64EB__
31457   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
31458 #else
31459   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
31460 #endif
31461 }
31462
31463 __extension__ extern __inline uint8x16_t
31464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31465 vzip1q_u8 (uint8x16_t __a, uint8x16_t __b)
31466 {
31467 #ifdef __AARCH64EB__
31468   return __builtin_shuffle (__a, __b, (uint8x16_t)
31469       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
31470 #else
31471   return __builtin_shuffle (__a, __b, (uint8x16_t)
31472       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
31473 #endif
31474 }
31475
31476 __extension__ extern __inline uint16x8_t
31477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31478 vzip1q_u16 (uint16x8_t __a, uint16x8_t __b)
31479 {
31480 #ifdef __AARCH64EB__
31481   return __builtin_shuffle (__a, __b, (uint16x8_t)
31482       {12, 4, 13, 5, 14, 6, 15, 7});
31483 #else
31484   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
31485 #endif
31486 }
31487
31488 __extension__ extern __inline uint32x4_t
31489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31490 vzip1q_u32 (uint32x4_t __a, uint32x4_t __b)
31491 {
31492 #ifdef __AARCH64EB__
31493   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
31494 #else
31495   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
31496 #endif
31497 }
31498
31499 __extension__ extern __inline uint64x2_t
31500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31501 vzip1q_u64 (uint64x2_t __a, uint64x2_t __b)
31502 {
31503 #ifdef __AARCH64EB__
31504   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
31505 #else
31506   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
31507 #endif
31508 }
31509
31510 __extension__ extern __inline float16x4_t
31511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31512 vzip2_f16 (float16x4_t __a, float16x4_t __b)
31513 {
31514 #ifdef __AARCH64EB__
31515   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
31516 #else
31517   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
31518 #endif
31519 }
31520
31521 __extension__ extern __inline float32x2_t
31522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31523 vzip2_f32 (float32x2_t __a, float32x2_t __b)
31524 {
31525 #ifdef __AARCH64EB__
31526   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
31527 #else
31528   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
31529 #endif
31530 }
31531
31532 __extension__ extern __inline poly8x8_t
31533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31534 vzip2_p8 (poly8x8_t __a, poly8x8_t __b)
31535 {
31536 #ifdef __AARCH64EB__
31537   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31538 #else
31539   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
31540 #endif
31541 }
31542
31543 __extension__ extern __inline poly16x4_t
31544 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31545 vzip2_p16 (poly16x4_t __a, poly16x4_t __b)
31546 {
31547 #ifdef __AARCH64EB__
31548   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
31549 #else
31550   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
31551 #endif
31552 }
31553
31554 __extension__ extern __inline int8x8_t
31555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31556 vzip2_s8 (int8x8_t __a, int8x8_t __b)
31557 {
31558 #ifdef __AARCH64EB__
31559   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31560 #else
31561   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
31562 #endif
31563 }
31564
31565 __extension__ extern __inline int16x4_t
31566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31567 vzip2_s16 (int16x4_t __a, int16x4_t __b)
31568 {
31569 #ifdef __AARCH64EB__
31570   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
31571 #else
31572   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
31573 #endif
31574 }
31575
31576 __extension__ extern __inline int32x2_t
31577 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31578 vzip2_s32 (int32x2_t __a, int32x2_t __b)
31579 {
31580 #ifdef __AARCH64EB__
31581   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
31582 #else
31583   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
31584 #endif
31585 }
31586
31587 __extension__ extern __inline uint8x8_t
31588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31589 vzip2_u8 (uint8x8_t __a, uint8x8_t __b)
31590 {
31591 #ifdef __AARCH64EB__
31592   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31593 #else
31594   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
31595 #endif
31596 }
31597
31598 __extension__ extern __inline uint16x4_t
31599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31600 vzip2_u16 (uint16x4_t __a, uint16x4_t __b)
31601 {
31602 #ifdef __AARCH64EB__
31603   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
31604 #else
31605   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
31606 #endif
31607 }
31608
31609 __extension__ extern __inline uint32x2_t
31610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31611 vzip2_u32 (uint32x2_t __a, uint32x2_t __b)
31612 {
31613 #ifdef __AARCH64EB__
31614   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
31615 #else
31616   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
31617 #endif
31618 }
31619
31620 __extension__ extern __inline float16x8_t
31621 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31622 vzip2q_f16 (float16x8_t __a, float16x8_t __b)
31623 {
31624 #ifdef __AARCH64EB__
31625   return __builtin_shuffle (__a, __b,
31626                             (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31627 #else
31628   return __builtin_shuffle (__a, __b,
31629                             (uint16x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
31630 #endif
31631 }
31632
31633 __extension__ extern __inline float32x4_t
31634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31635 vzip2q_f32 (float32x4_t __a, float32x4_t __b)
31636 {
31637 #ifdef __AARCH64EB__
31638   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
31639 #else
31640   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
31641 #endif
31642 }
31643
31644 __extension__ extern __inline float64x2_t
31645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31646 vzip2q_f64 (float64x2_t __a, float64x2_t __b)
31647 {
31648 #ifdef __AARCH64EB__
31649   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
31650 #else
31651   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
31652 #endif
31653 }
31654
31655 __extension__ extern __inline poly8x16_t
31656 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31657 vzip2q_p8 (poly8x16_t __a, poly8x16_t __b)
31658 {
31659 #ifdef __AARCH64EB__
31660   return __builtin_shuffle (__a, __b, (uint8x16_t)
31661       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
31662 #else
31663   return __builtin_shuffle (__a, __b, (uint8x16_t)
31664       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
31665 #endif
31666 }
31667
31668 __extension__ extern __inline poly16x8_t
31669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31670 vzip2q_p16 (poly16x8_t __a, poly16x8_t __b)
31671 {
31672 #ifdef __AARCH64EB__
31673   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31674 #else
31675   return __builtin_shuffle (__a, __b, (uint16x8_t)
31676       {4, 12, 5, 13, 6, 14, 7, 15});
31677 #endif
31678 }
31679
31680 __extension__ extern __inline int8x16_t
31681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31682 vzip2q_s8 (int8x16_t __a, int8x16_t __b)
31683 {
31684 #ifdef __AARCH64EB__
31685   return __builtin_shuffle (__a, __b, (uint8x16_t)
31686       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
31687 #else
31688   return __builtin_shuffle (__a, __b, (uint8x16_t)
31689       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
31690 #endif
31691 }
31692
31693 __extension__ extern __inline int16x8_t
31694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31695 vzip2q_s16 (int16x8_t __a, int16x8_t __b)
31696 {
31697 #ifdef __AARCH64EB__
31698   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31699 #else
31700   return __builtin_shuffle (__a, __b, (uint16x8_t)
31701       {4, 12, 5, 13, 6, 14, 7, 15});
31702 #endif
31703 }
31704
31705 __extension__ extern __inline int32x4_t
31706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31707 vzip2q_s32 (int32x4_t __a, int32x4_t __b)
31708 {
31709 #ifdef __AARCH64EB__
31710   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
31711 #else
31712   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
31713 #endif
31714 }
31715
31716 __extension__ extern __inline int64x2_t
31717 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31718 vzip2q_s64 (int64x2_t __a, int64x2_t __b)
31719 {
31720 #ifdef __AARCH64EB__
31721   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
31722 #else
31723   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
31724 #endif
31725 }
31726
31727 __extension__ extern __inline uint8x16_t
31728 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31729 vzip2q_u8 (uint8x16_t __a, uint8x16_t __b)
31730 {
31731 #ifdef __AARCH64EB__
31732   return __builtin_shuffle (__a, __b, (uint8x16_t)
31733       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
31734 #else
31735   return __builtin_shuffle (__a, __b, (uint8x16_t)
31736       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
31737 #endif
31738 }
31739
31740 __extension__ extern __inline uint16x8_t
31741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31742 vzip2q_u16 (uint16x8_t __a, uint16x8_t __b)
31743 {
31744 #ifdef __AARCH64EB__
31745   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
31746 #else
31747   return __builtin_shuffle (__a, __b, (uint16x8_t)
31748       {4, 12, 5, 13, 6, 14, 7, 15});
31749 #endif
31750 }
31751
31752 __extension__ extern __inline uint32x4_t
31753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31754 vzip2q_u32 (uint32x4_t __a, uint32x4_t __b)
31755 {
31756 #ifdef __AARCH64EB__
31757   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
31758 #else
31759   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
31760 #endif
31761 }
31762
31763 __extension__ extern __inline uint64x2_t
31764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31765 vzip2q_u64 (uint64x2_t __a, uint64x2_t __b)
31766 {
31767 #ifdef __AARCH64EB__
31768   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
31769 #else
31770   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
31771 #endif
31772 }
31773
31774 __INTERLEAVE_LIST (zip)
31775
31776 #undef __INTERLEAVE_LIST
31777 #undef __DEFINTERLEAVE
31778
31779 /* End of optimal implementations in approved order.  */
31780
31781 #pragma GCC pop_options
31782
31783 /* ARMv8.2-A FP16 intrinsics.  */
31784
31785 #include "arm_fp16.h"
31786
31787 #pragma GCC push_options
31788 #pragma GCC target ("arch=armv8.2-a+fp16")
31789
31790 /* ARMv8.2-A FP16 one operand vector intrinsics.  */
31791
31792 __extension__ extern __inline float16x4_t
31793 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31794 vabs_f16 (float16x4_t __a)
31795 {
31796   return __builtin_aarch64_absv4hf (__a);
31797 }
31798
31799 __extension__ extern __inline float16x8_t
31800 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31801 vabsq_f16 (float16x8_t __a)
31802 {
31803   return __builtin_aarch64_absv8hf (__a);
31804 }
31805
31806 __extension__ extern __inline uint16x4_t
31807 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31808 vceqz_f16 (float16x4_t __a)
31809 {
31810   return __builtin_aarch64_cmeqv4hf_uss (__a, vdup_n_f16 (0.0f));
31811 }
31812
31813 __extension__ extern __inline uint16x8_t
31814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31815 vceqzq_f16 (float16x8_t __a)
31816 {
31817   return __builtin_aarch64_cmeqv8hf_uss (__a, vdupq_n_f16 (0.0f));
31818 }
31819
31820 __extension__ extern __inline uint16x4_t
31821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31822 vcgez_f16 (float16x4_t __a)
31823 {
31824   return __builtin_aarch64_cmgev4hf_uss (__a, vdup_n_f16 (0.0f));
31825 }
31826
31827 __extension__ extern __inline uint16x8_t
31828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31829 vcgezq_f16 (float16x8_t __a)
31830 {
31831   return __builtin_aarch64_cmgev8hf_uss (__a, vdupq_n_f16 (0.0f));
31832 }
31833
31834 __extension__ extern __inline uint16x4_t
31835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31836 vcgtz_f16 (float16x4_t __a)
31837 {
31838   return __builtin_aarch64_cmgtv4hf_uss (__a, vdup_n_f16 (0.0f));
31839 }
31840
31841 __extension__ extern __inline uint16x8_t
31842 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31843 vcgtzq_f16 (float16x8_t __a)
31844 {
31845   return __builtin_aarch64_cmgtv8hf_uss (__a, vdupq_n_f16 (0.0f));
31846 }
31847
31848 __extension__ extern __inline uint16x4_t
31849 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31850 vclez_f16 (float16x4_t __a)
31851 {
31852   return __builtin_aarch64_cmlev4hf_uss (__a, vdup_n_f16 (0.0f));
31853 }
31854
31855 __extension__ extern __inline uint16x8_t
31856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31857 vclezq_f16 (float16x8_t __a)
31858 {
31859   return __builtin_aarch64_cmlev8hf_uss (__a, vdupq_n_f16 (0.0f));
31860 }
31861
31862 __extension__ extern __inline uint16x4_t
31863 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31864 vcltz_f16 (float16x4_t __a)
31865 {
31866   return __builtin_aarch64_cmltv4hf_uss (__a, vdup_n_f16 (0.0f));
31867 }
31868
31869 __extension__ extern __inline uint16x8_t
31870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31871 vcltzq_f16 (float16x8_t __a)
31872 {
31873   return __builtin_aarch64_cmltv8hf_uss (__a, vdupq_n_f16 (0.0f));
31874 }
31875
31876 __extension__ extern __inline float16x4_t
31877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31878 vcvt_f16_s16 (int16x4_t __a)
31879 {
31880   return __builtin_aarch64_floatv4hiv4hf (__a);
31881 }
31882
31883 __extension__ extern __inline float16x8_t
31884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31885 vcvtq_f16_s16 (int16x8_t __a)
31886 {
31887   return __builtin_aarch64_floatv8hiv8hf (__a);
31888 }
31889
31890 __extension__ extern __inline float16x4_t
31891 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31892 vcvt_f16_u16 (uint16x4_t __a)
31893 {
31894   return __builtin_aarch64_floatunsv4hiv4hf ((int16x4_t) __a);
31895 }
31896
31897 __extension__ extern __inline float16x8_t
31898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31899 vcvtq_f16_u16 (uint16x8_t __a)
31900 {
31901   return __builtin_aarch64_floatunsv8hiv8hf ((int16x8_t) __a);
31902 }
31903
31904 __extension__ extern __inline int16x4_t
31905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31906 vcvt_s16_f16 (float16x4_t __a)
31907 {
31908   return __builtin_aarch64_lbtruncv4hfv4hi (__a);
31909 }
31910
31911 __extension__ extern __inline int16x8_t
31912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31913 vcvtq_s16_f16 (float16x8_t __a)
31914 {
31915   return __builtin_aarch64_lbtruncv8hfv8hi (__a);
31916 }
31917
31918 __extension__ extern __inline uint16x4_t
31919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31920 vcvt_u16_f16 (float16x4_t __a)
31921 {
31922   return __builtin_aarch64_lbtruncuv4hfv4hi_us (__a);
31923 }
31924
31925 __extension__ extern __inline uint16x8_t
31926 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31927 vcvtq_u16_f16 (float16x8_t __a)
31928 {
31929   return __builtin_aarch64_lbtruncuv8hfv8hi_us (__a);
31930 }
31931
31932 __extension__ extern __inline int16x4_t
31933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31934 vcvta_s16_f16 (float16x4_t __a)
31935 {
31936   return __builtin_aarch64_lroundv4hfv4hi (__a);
31937 }
31938
31939 __extension__ extern __inline int16x8_t
31940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31941 vcvtaq_s16_f16 (float16x8_t __a)
31942 {
31943   return __builtin_aarch64_lroundv8hfv8hi (__a);
31944 }
31945
31946 __extension__ extern __inline uint16x4_t
31947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31948 vcvta_u16_f16 (float16x4_t __a)
31949 {
31950   return __builtin_aarch64_lrounduv4hfv4hi_us (__a);
31951 }
31952
31953 __extension__ extern __inline uint16x8_t
31954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31955 vcvtaq_u16_f16 (float16x8_t __a)
31956 {
31957   return __builtin_aarch64_lrounduv8hfv8hi_us (__a);
31958 }
31959
31960 __extension__ extern __inline int16x4_t
31961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31962 vcvtm_s16_f16 (float16x4_t __a)
31963 {
31964   return __builtin_aarch64_lfloorv4hfv4hi (__a);
31965 }
31966
31967 __extension__ extern __inline int16x8_t
31968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31969 vcvtmq_s16_f16 (float16x8_t __a)
31970 {
31971   return __builtin_aarch64_lfloorv8hfv8hi (__a);
31972 }
31973
31974 __extension__ extern __inline uint16x4_t
31975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31976 vcvtm_u16_f16 (float16x4_t __a)
31977 {
31978   return __builtin_aarch64_lflooruv4hfv4hi_us (__a);
31979 }
31980
31981 __extension__ extern __inline uint16x8_t
31982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31983 vcvtmq_u16_f16 (float16x8_t __a)
31984 {
31985   return __builtin_aarch64_lflooruv8hfv8hi_us (__a);
31986 }
31987
31988 __extension__ extern __inline int16x4_t
31989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31990 vcvtn_s16_f16 (float16x4_t __a)
31991 {
31992   return __builtin_aarch64_lfrintnv4hfv4hi (__a);
31993 }
31994
31995 __extension__ extern __inline int16x8_t
31996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31997 vcvtnq_s16_f16 (float16x8_t __a)
31998 {
31999   return __builtin_aarch64_lfrintnv8hfv8hi (__a);
32000 }
32001
32002 __extension__ extern __inline uint16x4_t
32003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32004 vcvtn_u16_f16 (float16x4_t __a)
32005 {
32006   return __builtin_aarch64_lfrintnuv4hfv4hi_us (__a);
32007 }
32008
32009 __extension__ extern __inline uint16x8_t
32010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32011 vcvtnq_u16_f16 (float16x8_t __a)
32012 {
32013   return __builtin_aarch64_lfrintnuv8hfv8hi_us (__a);
32014 }
32015
32016 __extension__ extern __inline int16x4_t
32017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32018 vcvtp_s16_f16 (float16x4_t __a)
32019 {
32020   return __builtin_aarch64_lceilv4hfv4hi (__a);
32021 }
32022
32023 __extension__ extern __inline int16x8_t
32024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32025 vcvtpq_s16_f16 (float16x8_t __a)
32026 {
32027   return __builtin_aarch64_lceilv8hfv8hi (__a);
32028 }
32029
32030 __extension__ extern __inline uint16x4_t
32031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32032 vcvtp_u16_f16 (float16x4_t __a)
32033 {
32034   return __builtin_aarch64_lceiluv4hfv4hi_us (__a);
32035 }
32036
32037 __extension__ extern __inline uint16x8_t
32038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32039 vcvtpq_u16_f16 (float16x8_t __a)
32040 {
32041   return __builtin_aarch64_lceiluv8hfv8hi_us (__a);
32042 }
32043
32044 __extension__ extern __inline float16x4_t
32045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32046 vneg_f16 (float16x4_t __a)
32047 {
32048   return -__a;
32049 }
32050
32051 __extension__ extern __inline float16x8_t
32052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32053 vnegq_f16 (float16x8_t __a)
32054 {
32055   return -__a;
32056 }
32057
32058 __extension__ extern __inline float16x4_t
32059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32060 vrecpe_f16 (float16x4_t __a)
32061 {
32062   return __builtin_aarch64_frecpev4hf (__a);
32063 }
32064
32065 __extension__ extern __inline float16x8_t
32066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32067 vrecpeq_f16 (float16x8_t __a)
32068 {
32069   return __builtin_aarch64_frecpev8hf (__a);
32070 }
32071
32072 __extension__ extern __inline float16x4_t
32073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32074 vrnd_f16 (float16x4_t __a)
32075 {
32076   return __builtin_aarch64_btruncv4hf (__a);
32077 }
32078
32079 __extension__ extern __inline float16x8_t
32080 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32081 vrndq_f16 (float16x8_t __a)
32082 {
32083   return __builtin_aarch64_btruncv8hf (__a);
32084 }
32085
32086 __extension__ extern __inline float16x4_t
32087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32088 vrnda_f16 (float16x4_t __a)
32089 {
32090   return __builtin_aarch64_roundv4hf (__a);
32091 }
32092
32093 __extension__ extern __inline float16x8_t
32094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32095 vrndaq_f16 (float16x8_t __a)
32096 {
32097   return __builtin_aarch64_roundv8hf (__a);
32098 }
32099
32100 __extension__ extern __inline float16x4_t
32101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32102 vrndi_f16 (float16x4_t __a)
32103 {
32104   return __builtin_aarch64_nearbyintv4hf (__a);
32105 }
32106
32107 __extension__ extern __inline float16x8_t
32108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32109 vrndiq_f16 (float16x8_t __a)
32110 {
32111   return __builtin_aarch64_nearbyintv8hf (__a);
32112 }
32113
32114 __extension__ extern __inline float16x4_t
32115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32116 vrndm_f16 (float16x4_t __a)
32117 {
32118   return __builtin_aarch64_floorv4hf (__a);
32119 }
32120
32121 __extension__ extern __inline float16x8_t
32122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32123 vrndmq_f16 (float16x8_t __a)
32124 {
32125   return __builtin_aarch64_floorv8hf (__a);
32126 }
32127
32128 __extension__ extern __inline float16x4_t
32129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32130 vrndn_f16 (float16x4_t __a)
32131 {
32132   return __builtin_aarch64_frintnv4hf (__a);
32133 }
32134
32135 __extension__ extern __inline float16x8_t
32136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32137 vrndnq_f16 (float16x8_t __a)
32138 {
32139   return __builtin_aarch64_frintnv8hf (__a);
32140 }
32141
32142 __extension__ extern __inline float16x4_t
32143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32144 vrndp_f16 (float16x4_t __a)
32145 {
32146   return __builtin_aarch64_ceilv4hf (__a);
32147 }
32148
32149 __extension__ extern __inline float16x8_t
32150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32151 vrndpq_f16 (float16x8_t __a)
32152 {
32153   return __builtin_aarch64_ceilv8hf (__a);
32154 }
32155
32156 __extension__ extern __inline float16x4_t
32157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32158 vrndx_f16 (float16x4_t __a)
32159 {
32160   return __builtin_aarch64_rintv4hf (__a);
32161 }
32162
32163 __extension__ extern __inline float16x8_t
32164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32165 vrndxq_f16 (float16x8_t __a)
32166 {
32167   return __builtin_aarch64_rintv8hf (__a);
32168 }
32169
32170 __extension__ extern __inline float16x4_t
32171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32172 vrsqrte_f16 (float16x4_t a)
32173 {
32174   return __builtin_aarch64_rsqrtev4hf (a);
32175 }
32176
32177 __extension__ extern __inline float16x8_t
32178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32179 vrsqrteq_f16 (float16x8_t a)
32180 {
32181   return __builtin_aarch64_rsqrtev8hf (a);
32182 }
32183
32184 __extension__ extern __inline float16x4_t
32185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32186 vsqrt_f16 (float16x4_t a)
32187 {
32188   return __builtin_aarch64_sqrtv4hf (a);
32189 }
32190
32191 __extension__ extern __inline float16x8_t
32192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32193 vsqrtq_f16 (float16x8_t a)
32194 {
32195   return __builtin_aarch64_sqrtv8hf (a);
32196 }
32197
32198 /* ARMv8.2-A FP16 two operands vector intrinsics.  */
32199
32200 __extension__ extern __inline float16x4_t
32201 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32202 vadd_f16 (float16x4_t __a, float16x4_t __b)
32203 {
32204   return __a + __b;
32205 }
32206
32207 __extension__ extern __inline float16x8_t
32208 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32209 vaddq_f16 (float16x8_t __a, float16x8_t __b)
32210 {
32211   return __a + __b;
32212 }
32213
32214 __extension__ extern __inline float16x4_t
32215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32216 vabd_f16 (float16x4_t a, float16x4_t b)
32217 {
32218   return __builtin_aarch64_fabdv4hf (a, b);
32219 }
32220
32221 __extension__ extern __inline float16x8_t
32222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32223 vabdq_f16 (float16x8_t a, float16x8_t b)
32224 {
32225   return __builtin_aarch64_fabdv8hf (a, b);
32226 }
32227
32228 __extension__ extern __inline uint16x4_t
32229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32230 vcage_f16 (float16x4_t __a, float16x4_t __b)
32231 {
32232   return __builtin_aarch64_facgev4hf_uss (__a, __b);
32233 }
32234
32235 __extension__ extern __inline uint16x8_t
32236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32237 vcageq_f16 (float16x8_t __a, float16x8_t __b)
32238 {
32239   return __builtin_aarch64_facgev8hf_uss (__a, __b);
32240 }
32241
32242 __extension__ extern __inline uint16x4_t
32243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32244 vcagt_f16 (float16x4_t __a, float16x4_t __b)
32245 {
32246   return __builtin_aarch64_facgtv4hf_uss (__a, __b);
32247 }
32248
32249 __extension__ extern __inline uint16x8_t
32250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32251 vcagtq_f16 (float16x8_t __a, float16x8_t __b)
32252 {
32253   return __builtin_aarch64_facgtv8hf_uss (__a, __b);
32254 }
32255
32256 __extension__ extern __inline uint16x4_t
32257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32258 vcale_f16 (float16x4_t __a, float16x4_t __b)
32259 {
32260   return __builtin_aarch64_faclev4hf_uss (__a, __b);
32261 }
32262
32263 __extension__ extern __inline uint16x8_t
32264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32265 vcaleq_f16 (float16x8_t __a, float16x8_t __b)
32266 {
32267   return __builtin_aarch64_faclev8hf_uss (__a, __b);
32268 }
32269
32270 __extension__ extern __inline uint16x4_t
32271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32272 vcalt_f16 (float16x4_t __a, float16x4_t __b)
32273 {
32274   return __builtin_aarch64_facltv4hf_uss (__a, __b);
32275 }
32276
32277 __extension__ extern __inline uint16x8_t
32278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32279 vcaltq_f16 (float16x8_t __a, float16x8_t __b)
32280 {
32281   return __builtin_aarch64_facltv8hf_uss (__a, __b);
32282 }
32283
32284 __extension__ extern __inline uint16x4_t
32285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32286 vceq_f16 (float16x4_t __a, float16x4_t __b)
32287 {
32288   return __builtin_aarch64_cmeqv4hf_uss (__a, __b);
32289 }
32290
32291 __extension__ extern __inline uint16x8_t
32292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32293 vceqq_f16 (float16x8_t __a, float16x8_t __b)
32294 {
32295   return __builtin_aarch64_cmeqv8hf_uss (__a, __b);
32296 }
32297
32298 __extension__ extern __inline uint16x4_t
32299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32300 vcge_f16 (float16x4_t __a, float16x4_t __b)
32301 {
32302   return __builtin_aarch64_cmgev4hf_uss (__a, __b);
32303 }
32304
32305 __extension__ extern __inline uint16x8_t
32306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32307 vcgeq_f16 (float16x8_t __a, float16x8_t __b)
32308 {
32309   return __builtin_aarch64_cmgev8hf_uss (__a, __b);
32310 }
32311
32312 __extension__ extern __inline uint16x4_t
32313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32314 vcgt_f16 (float16x4_t __a, float16x4_t __b)
32315 {
32316   return __builtin_aarch64_cmgtv4hf_uss (__a, __b);
32317 }
32318
32319 __extension__ extern __inline uint16x8_t
32320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32321 vcgtq_f16 (float16x8_t __a, float16x8_t __b)
32322 {
32323   return __builtin_aarch64_cmgtv8hf_uss (__a, __b);
32324 }
32325
32326 __extension__ extern __inline uint16x4_t
32327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32328 vcle_f16 (float16x4_t __a, float16x4_t __b)
32329 {
32330   return __builtin_aarch64_cmlev4hf_uss (__a, __b);
32331 }
32332
32333 __extension__ extern __inline uint16x8_t
32334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32335 vcleq_f16 (float16x8_t __a, float16x8_t __b)
32336 {
32337   return __builtin_aarch64_cmlev8hf_uss (__a, __b);
32338 }
32339
32340 __extension__ extern __inline uint16x4_t
32341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32342 vclt_f16 (float16x4_t __a, float16x4_t __b)
32343 {
32344   return __builtin_aarch64_cmltv4hf_uss (__a, __b);
32345 }
32346
32347 __extension__ extern __inline uint16x8_t
32348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32349 vcltq_f16 (float16x8_t __a, float16x8_t __b)
32350 {
32351   return __builtin_aarch64_cmltv8hf_uss (__a, __b);
32352 }
32353
32354 __extension__ extern __inline float16x4_t
32355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32356 vcvt_n_f16_s16 (int16x4_t __a, const int __b)
32357 {
32358   return __builtin_aarch64_scvtfv4hi (__a, __b);
32359 }
32360
32361 __extension__ extern __inline float16x8_t
32362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32363 vcvtq_n_f16_s16 (int16x8_t __a, const int __b)
32364 {
32365   return __builtin_aarch64_scvtfv8hi (__a, __b);
32366 }
32367
32368 __extension__ extern __inline float16x4_t
32369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32370 vcvt_n_f16_u16 (uint16x4_t __a, const int __b)
32371 {
32372   return __builtin_aarch64_ucvtfv4hi_sus (__a, __b);
32373 }
32374
32375 __extension__ extern __inline float16x8_t
32376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32377 vcvtq_n_f16_u16 (uint16x8_t __a, const int __b)
32378 {
32379   return __builtin_aarch64_ucvtfv8hi_sus (__a, __b);
32380 }
32381
32382 __extension__ extern __inline int16x4_t
32383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32384 vcvt_n_s16_f16 (float16x4_t __a, const int __b)
32385 {
32386   return __builtin_aarch64_fcvtzsv4hf (__a, __b);
32387 }
32388
32389 __extension__ extern __inline int16x8_t
32390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32391 vcvtq_n_s16_f16 (float16x8_t __a, const int __b)
32392 {
32393   return __builtin_aarch64_fcvtzsv8hf (__a, __b);
32394 }
32395
32396 __extension__ extern __inline uint16x4_t
32397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32398 vcvt_n_u16_f16 (float16x4_t __a, const int __b)
32399 {
32400   return __builtin_aarch64_fcvtzuv4hf_uss (__a, __b);
32401 }
32402
32403 __extension__ extern __inline uint16x8_t
32404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32405 vcvtq_n_u16_f16 (float16x8_t __a, const int __b)
32406 {
32407   return __builtin_aarch64_fcvtzuv8hf_uss (__a, __b);
32408 }
32409
32410 __extension__ extern __inline float16x4_t
32411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32412 vdiv_f16 (float16x4_t __a, float16x4_t __b)
32413 {
32414   return __a / __b;
32415 }
32416
32417 __extension__ extern __inline float16x8_t
32418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32419 vdivq_f16 (float16x8_t __a, float16x8_t __b)
32420 {
32421   return __a / __b;
32422 }
32423
32424 __extension__ extern __inline float16x4_t
32425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32426 vmax_f16 (float16x4_t __a, float16x4_t __b)
32427 {
32428   return __builtin_aarch64_smax_nanv4hf (__a, __b);
32429 }
32430
32431 __extension__ extern __inline float16x8_t
32432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32433 vmaxq_f16 (float16x8_t __a, float16x8_t __b)
32434 {
32435   return __builtin_aarch64_smax_nanv8hf (__a, __b);
32436 }
32437
32438 __extension__ extern __inline float16x4_t
32439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32440 vmaxnm_f16 (float16x4_t __a, float16x4_t __b)
32441 {
32442   return __builtin_aarch64_fmaxv4hf (__a, __b);
32443 }
32444
32445 __extension__ extern __inline float16x8_t
32446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32447 vmaxnmq_f16 (float16x8_t __a, float16x8_t __b)
32448 {
32449   return __builtin_aarch64_fmaxv8hf (__a, __b);
32450 }
32451
32452 __extension__ extern __inline float16x4_t
32453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32454 vmin_f16 (float16x4_t __a, float16x4_t __b)
32455 {
32456   return __builtin_aarch64_smin_nanv4hf (__a, __b);
32457 }
32458
32459 __extension__ extern __inline float16x8_t
32460 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32461 vminq_f16 (float16x8_t __a, float16x8_t __b)
32462 {
32463   return __builtin_aarch64_smin_nanv8hf (__a, __b);
32464 }
32465
32466 __extension__ extern __inline float16x4_t
32467 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32468 vminnm_f16 (float16x4_t __a, float16x4_t __b)
32469 {
32470   return __builtin_aarch64_fminv4hf (__a, __b);
32471 }
32472
32473 __extension__ extern __inline float16x8_t
32474 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32475 vminnmq_f16 (float16x8_t __a, float16x8_t __b)
32476 {
32477   return __builtin_aarch64_fminv8hf (__a, __b);
32478 }
32479
32480 __extension__ extern __inline float16x4_t
32481 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32482 vmul_f16 (float16x4_t __a, float16x4_t __b)
32483 {
32484   return __a * __b;
32485 }
32486
32487 __extension__ extern __inline float16x8_t
32488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32489 vmulq_f16 (float16x8_t __a, float16x8_t __b)
32490 {
32491   return __a * __b;
32492 }
32493
32494 __extension__ extern __inline float16x4_t
32495 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32496 vmulx_f16 (float16x4_t __a, float16x4_t __b)
32497 {
32498   return __builtin_aarch64_fmulxv4hf (__a, __b);
32499 }
32500
32501 __extension__ extern __inline float16x8_t
32502 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32503 vmulxq_f16 (float16x8_t __a, float16x8_t __b)
32504 {
32505   return __builtin_aarch64_fmulxv8hf (__a, __b);
32506 }
32507
32508 __extension__ extern __inline float16x4_t
32509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32510 vpadd_f16 (float16x4_t a, float16x4_t b)
32511 {
32512   return __builtin_aarch64_faddpv4hf (a, b);
32513 }
32514
32515 __extension__ extern __inline float16x8_t
32516 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32517 vpaddq_f16 (float16x8_t a, float16x8_t b)
32518 {
32519   return __builtin_aarch64_faddpv8hf (a, b);
32520 }
32521
32522 __extension__ extern __inline float16x4_t
32523 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32524 vpmax_f16 (float16x4_t a, float16x4_t b)
32525 {
32526   return __builtin_aarch64_smax_nanpv4hf (a, b);
32527 }
32528
32529 __extension__ extern __inline float16x8_t
32530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32531 vpmaxq_f16 (float16x8_t a, float16x8_t b)
32532 {
32533   return __builtin_aarch64_smax_nanpv8hf (a, b);
32534 }
32535
32536 __extension__ extern __inline float16x4_t
32537 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32538 vpmaxnm_f16 (float16x4_t a, float16x4_t b)
32539 {
32540   return __builtin_aarch64_smaxpv4hf (a, b);
32541 }
32542
32543 __extension__ extern __inline float16x8_t
32544 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32545 vpmaxnmq_f16 (float16x8_t a, float16x8_t b)
32546 {
32547   return __builtin_aarch64_smaxpv8hf (a, b);
32548 }
32549
32550 __extension__ extern __inline float16x4_t
32551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32552 vpmin_f16 (float16x4_t a, float16x4_t b)
32553 {
32554   return __builtin_aarch64_smin_nanpv4hf (a, b);
32555 }
32556
32557 __extension__ extern __inline float16x8_t
32558 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32559 vpminq_f16 (float16x8_t a, float16x8_t b)
32560 {
32561   return __builtin_aarch64_smin_nanpv8hf (a, b);
32562 }
32563
32564 __extension__ extern __inline float16x4_t
32565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32566 vpminnm_f16 (float16x4_t a, float16x4_t b)
32567 {
32568   return __builtin_aarch64_sminpv4hf (a, b);
32569 }
32570
32571 __extension__ extern __inline float16x8_t
32572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32573 vpminnmq_f16 (float16x8_t a, float16x8_t b)
32574 {
32575   return __builtin_aarch64_sminpv8hf (a, b);
32576 }
32577
32578 __extension__ extern __inline float16x4_t
32579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32580 vrecps_f16 (float16x4_t __a, float16x4_t __b)
32581 {
32582   return __builtin_aarch64_frecpsv4hf (__a, __b);
32583 }
32584
32585 __extension__ extern __inline float16x8_t
32586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32587 vrecpsq_f16 (float16x8_t __a, float16x8_t __b)
32588 {
32589   return __builtin_aarch64_frecpsv8hf (__a, __b);
32590 }
32591
32592 __extension__ extern __inline float16x4_t
32593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32594 vrsqrts_f16 (float16x4_t a, float16x4_t b)
32595 {
32596   return __builtin_aarch64_rsqrtsv4hf (a, b);
32597 }
32598
32599 __extension__ extern __inline float16x8_t
32600 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32601 vrsqrtsq_f16 (float16x8_t a, float16x8_t b)
32602 {
32603   return __builtin_aarch64_rsqrtsv8hf (a, b);
32604 }
32605
32606 __extension__ extern __inline float16x4_t
32607 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32608 vsub_f16 (float16x4_t __a, float16x4_t __b)
32609 {
32610   return __a - __b;
32611 }
32612
32613 __extension__ extern __inline float16x8_t
32614 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32615 vsubq_f16 (float16x8_t __a, float16x8_t __b)
32616 {
32617   return __a - __b;
32618 }
32619
32620 /* ARMv8.2-A FP16 three operands vector intrinsics.  */
32621
32622 __extension__ extern __inline float16x4_t
32623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32624 vfma_f16 (float16x4_t __a, float16x4_t __b, float16x4_t __c)
32625 {
32626   return __builtin_aarch64_fmav4hf (__b, __c, __a);
32627 }
32628
32629 __extension__ extern __inline float16x8_t
32630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32631 vfmaq_f16 (float16x8_t __a, float16x8_t __b, float16x8_t __c)
32632 {
32633   return __builtin_aarch64_fmav8hf (__b, __c, __a);
32634 }
32635
32636 __extension__ extern __inline float16x4_t
32637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32638 vfms_f16 (float16x4_t __a, float16x4_t __b, float16x4_t __c)
32639 {
32640   return __builtin_aarch64_fnmav4hf (__b, __c, __a);
32641 }
32642
32643 __extension__ extern __inline float16x8_t
32644 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32645 vfmsq_f16 (float16x8_t __a, float16x8_t __b, float16x8_t __c)
32646 {
32647   return __builtin_aarch64_fnmav8hf (__b, __c, __a);
32648 }
32649
32650 /* ARMv8.2-A FP16 lane vector intrinsics.  */
32651
32652 __extension__ extern __inline float16_t
32653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32654 vfmah_lane_f16 (float16_t __a, float16_t __b,
32655                 float16x4_t __c, const int __lane)
32656 {
32657   return vfmah_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
32658 }
32659
32660 __extension__ extern __inline float16_t
32661 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32662 vfmah_laneq_f16 (float16_t __a, float16_t __b,
32663                  float16x8_t __c, const int __lane)
32664 {
32665   return vfmah_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
32666 }
32667
32668 __extension__ extern __inline float16x4_t
32669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32670 vfma_lane_f16 (float16x4_t __a, float16x4_t __b,
32671                float16x4_t __c, const int __lane)
32672 {
32673   return vfma_f16 (__a, __b, __aarch64_vdup_lane_f16 (__c, __lane));
32674 }
32675
32676 __extension__ extern __inline float16x8_t
32677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32678 vfmaq_lane_f16 (float16x8_t __a, float16x8_t __b,
32679                 float16x4_t __c, const int __lane)
32680 {
32681   return vfmaq_f16 (__a, __b, __aarch64_vdupq_lane_f16 (__c, __lane));
32682 }
32683
32684 __extension__ extern __inline float16x4_t
32685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32686 vfma_laneq_f16 (float16x4_t __a, float16x4_t __b,
32687                 float16x8_t __c, const int __lane)
32688 {
32689   return vfma_f16 (__a, __b, __aarch64_vdup_laneq_f16 (__c, __lane));
32690 }
32691
32692 __extension__ extern __inline float16x8_t
32693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32694 vfmaq_laneq_f16 (float16x8_t __a, float16x8_t __b,
32695                  float16x8_t __c, const int __lane)
32696 {
32697   return vfmaq_f16 (__a, __b, __aarch64_vdupq_laneq_f16 (__c, __lane));
32698 }
32699
32700 __extension__ extern __inline float16x4_t
32701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32702 vfma_n_f16 (float16x4_t __a, float16x4_t __b, float16_t __c)
32703 {
32704   return vfma_f16 (__a, __b, vdup_n_f16 (__c));
32705 }
32706
32707 __extension__ extern __inline float16x8_t
32708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32709 vfmaq_n_f16 (float16x8_t __a, float16x8_t __b, float16_t __c)
32710 {
32711   return vfmaq_f16 (__a, __b, vdupq_n_f16 (__c));
32712 }
32713
32714 __extension__ extern __inline float16_t
32715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32716 vfmsh_lane_f16 (float16_t __a, float16_t __b,
32717                 float16x4_t __c, const int __lane)
32718 {
32719   return vfmsh_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
32720 }
32721
32722 __extension__ extern __inline float16_t
32723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32724 vfmsh_laneq_f16 (float16_t __a, float16_t __b,
32725                  float16x8_t __c, const int __lane)
32726 {
32727   return vfmsh_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
32728 }
32729
32730 __extension__ extern __inline float16x4_t
32731 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32732 vfms_lane_f16 (float16x4_t __a, float16x4_t __b,
32733                float16x4_t __c, const int __lane)
32734 {
32735   return vfms_f16 (__a, __b, __aarch64_vdup_lane_f16 (__c, __lane));
32736 }
32737
32738 __extension__ extern __inline float16x8_t
32739 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32740 vfmsq_lane_f16 (float16x8_t __a, float16x8_t __b,
32741                 float16x4_t __c, const int __lane)
32742 {
32743   return vfmsq_f16 (__a, __b, __aarch64_vdupq_lane_f16 (__c, __lane));
32744 }
32745
32746 __extension__ extern __inline float16x4_t
32747 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32748 vfms_laneq_f16 (float16x4_t __a, float16x4_t __b,
32749                 float16x8_t __c, const int __lane)
32750 {
32751   return vfms_f16 (__a, __b, __aarch64_vdup_laneq_f16 (__c, __lane));
32752 }
32753
32754 __extension__ extern __inline float16x8_t
32755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32756 vfmsq_laneq_f16 (float16x8_t __a, float16x8_t __b,
32757                  float16x8_t __c, const int __lane)
32758 {
32759   return vfmsq_f16 (__a, __b, __aarch64_vdupq_laneq_f16 (__c, __lane));
32760 }
32761
32762 __extension__ extern __inline float16x4_t
32763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32764 vfms_n_f16 (float16x4_t __a, float16x4_t __b, float16_t __c)
32765 {
32766   return vfms_f16 (__a, __b, vdup_n_f16 (__c));
32767 }
32768
32769 __extension__ extern __inline float16x8_t
32770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32771 vfmsq_n_f16 (float16x8_t __a, float16x8_t __b, float16_t __c)
32772 {
32773   return vfmsq_f16 (__a, __b, vdupq_n_f16 (__c));
32774 }
32775
32776 __extension__ extern __inline float16_t
32777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32778 vmulh_lane_f16 (float16_t __a, float16x4_t __b, const int __lane)
32779 {
32780   return __a * __aarch64_vget_lane_any (__b, __lane);
32781 }
32782
32783 __extension__ extern __inline float16x4_t
32784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32785 vmul_lane_f16 (float16x4_t __a, float16x4_t __b, const int __lane)
32786 {
32787   return vmul_f16 (__a, vdup_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
32788 }
32789
32790 __extension__ extern __inline float16x8_t
32791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32792 vmulq_lane_f16 (float16x8_t __a, float16x4_t __b, const int __lane)
32793 {
32794   return vmulq_f16 (__a, vdupq_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
32795 }
32796
32797 __extension__ extern __inline float16_t
32798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32799 vmulh_laneq_f16 (float16_t __a, float16x8_t __b, const int __lane)
32800 {
32801   return __a * __aarch64_vget_lane_any (__b, __lane);
32802 }
32803
32804 __extension__ extern __inline float16x4_t
32805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32806 vmul_laneq_f16 (float16x4_t __a, float16x8_t __b, const int __lane)
32807 {
32808   return vmul_f16 (__a, vdup_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
32809 }
32810
32811 __extension__ extern __inline float16x8_t
32812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32813 vmulq_laneq_f16 (float16x8_t __a, float16x8_t __b, const int __lane)
32814 {
32815   return vmulq_f16 (__a, vdupq_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
32816 }
32817
32818 __extension__ extern __inline float16x4_t
32819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32820 vmul_n_f16 (float16x4_t __a, float16_t __b)
32821 {
32822   return vmul_lane_f16 (__a, vdup_n_f16 (__b), 0);
32823 }
32824
32825 __extension__ extern __inline float16x8_t
32826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32827 vmulq_n_f16 (float16x8_t __a, float16_t __b)
32828 {
32829   return vmulq_laneq_f16 (__a, vdupq_n_f16 (__b), 0);
32830 }
32831
32832 __extension__ extern __inline float16_t
32833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32834 vmulxh_lane_f16 (float16_t __a, float16x4_t __b, const int __lane)
32835 {
32836   return vmulxh_f16 (__a, __aarch64_vget_lane_any (__b, __lane));
32837 }
32838
32839 __extension__ extern __inline float16x4_t
32840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32841 vmulx_lane_f16 (float16x4_t __a, float16x4_t __b, const int __lane)
32842 {
32843   return vmulx_f16 (__a, __aarch64_vdup_lane_f16 (__b, __lane));
32844 }
32845
32846 __extension__ extern __inline float16x8_t
32847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32848 vmulxq_lane_f16 (float16x8_t __a, float16x4_t __b, const int __lane)
32849 {
32850   return vmulxq_f16 (__a, __aarch64_vdupq_lane_f16 (__b, __lane));
32851 }
32852
32853 __extension__ extern __inline float16_t
32854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32855 vmulxh_laneq_f16 (float16_t __a, float16x8_t __b, const int __lane)
32856 {
32857   return vmulxh_f16 (__a, __aarch64_vget_lane_any (__b, __lane));
32858 }
32859
32860 __extension__ extern __inline float16x4_t
32861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32862 vmulx_laneq_f16 (float16x4_t __a, float16x8_t __b, const int __lane)
32863 {
32864   return vmulx_f16 (__a, __aarch64_vdup_laneq_f16 (__b, __lane));
32865 }
32866
32867 __extension__ extern __inline float16x8_t
32868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32869 vmulxq_laneq_f16 (float16x8_t __a, float16x8_t __b, const int __lane)
32870 {
32871   return vmulxq_f16 (__a, __aarch64_vdupq_laneq_f16 (__b, __lane));
32872 }
32873
32874 __extension__ extern __inline float16x4_t
32875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32876 vmulx_n_f16 (float16x4_t __a, float16_t __b)
32877 {
32878   return vmulx_f16 (__a, vdup_n_f16 (__b));
32879 }
32880
32881 __extension__ extern __inline float16x8_t
32882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32883 vmulxq_n_f16 (float16x8_t __a, float16_t __b)
32884 {
32885   return vmulxq_f16 (__a, vdupq_n_f16 (__b));
32886 }
32887
32888 /* ARMv8.2-A FP16 reduction vector intrinsics.  */
32889
32890 __extension__ extern __inline float16_t
32891 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32892 vmaxv_f16 (float16x4_t __a)
32893 {
32894   return __builtin_aarch64_reduc_smax_nan_scal_v4hf (__a);
32895 }
32896
32897 __extension__ extern __inline float16_t
32898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32899 vmaxvq_f16 (float16x8_t __a)
32900 {
32901   return __builtin_aarch64_reduc_smax_nan_scal_v8hf (__a);
32902 }
32903
32904 __extension__ extern __inline float16_t
32905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32906 vminv_f16 (float16x4_t __a)
32907 {
32908   return __builtin_aarch64_reduc_smin_nan_scal_v4hf (__a);
32909 }
32910
32911 __extension__ extern __inline float16_t
32912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32913 vminvq_f16 (float16x8_t __a)
32914 {
32915   return __builtin_aarch64_reduc_smin_nan_scal_v8hf (__a);
32916 }
32917
32918 __extension__ extern __inline float16_t
32919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32920 vmaxnmv_f16 (float16x4_t __a)
32921 {
32922   return __builtin_aarch64_reduc_smax_scal_v4hf (__a);
32923 }
32924
32925 __extension__ extern __inline float16_t
32926 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32927 vmaxnmvq_f16 (float16x8_t __a)
32928 {
32929   return __builtin_aarch64_reduc_smax_scal_v8hf (__a);
32930 }
32931
32932 __extension__ extern __inline float16_t
32933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32934 vminnmv_f16 (float16x4_t __a)
32935 {
32936   return __builtin_aarch64_reduc_smin_scal_v4hf (__a);
32937 }
32938
32939 __extension__ extern __inline float16_t
32940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32941 vminnmvq_f16 (float16x8_t __a)
32942 {
32943   return __builtin_aarch64_reduc_smin_scal_v8hf (__a);
32944 }
32945
32946 #pragma GCC pop_options
32947
32948 /* AdvSIMD Dot Product intrinsics.  */
32949
32950 #pragma GCC push_options
32951 #pragma GCC target ("arch=armv8.2-a+dotprod")
32952
32953 __extension__ extern __inline uint32x2_t
32954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32955 vdot_u32 (uint32x2_t __r, uint8x8_t __a, uint8x8_t __b)
32956 {
32957   return __builtin_aarch64_udotv8qi_uuuu (__r, __a, __b);
32958 }
32959
32960 __extension__ extern __inline uint32x4_t
32961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32962 vdotq_u32 (uint32x4_t __r, uint8x16_t __a, uint8x16_t __b)
32963 {
32964   return __builtin_aarch64_udotv16qi_uuuu (__r, __a, __b);
32965 }
32966
32967 __extension__ extern __inline int32x2_t
32968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32969 vdot_s32 (int32x2_t __r, int8x8_t __a, int8x8_t __b)
32970 {
32971   return __builtin_aarch64_sdotv8qi (__r, __a, __b);
32972 }
32973
32974 __extension__ extern __inline int32x4_t
32975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32976 vdotq_s32 (int32x4_t __r, int8x16_t __a, int8x16_t __b)
32977 {
32978   return __builtin_aarch64_sdotv16qi (__r, __a, __b);
32979 }
32980
32981 __extension__ extern __inline uint32x2_t
32982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32983 vdot_lane_u32 (uint32x2_t __r, uint8x8_t __a, uint8x8_t __b, const int __index)
32984 {
32985   return __builtin_aarch64_udot_lanev8qi_uuuus (__r, __a, __b, __index);
32986 }
32987
32988 __extension__ extern __inline uint32x2_t
32989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32990 vdot_laneq_u32 (uint32x2_t __r, uint8x8_t __a, uint8x16_t __b,
32991                 const int __index)
32992 {
32993   return __builtin_aarch64_udot_laneqv8qi_uuuus (__r, __a, __b, __index);
32994 }
32995
32996 __extension__ extern __inline uint32x4_t
32997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
32998 vdotq_lane_u32 (uint32x4_t __r, uint8x16_t __a, uint8x8_t __b,
32999                 const int __index)
33000 {
33001   return __builtin_aarch64_udot_lanev16qi_uuuus (__r, __a, __b, __index);
33002 }
33003
33004 __extension__ extern __inline uint32x4_t
33005 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33006 vdotq_laneq_u32 (uint32x4_t __r, uint8x16_t __a, uint8x16_t __b,
33007                  const int __index)
33008 {
33009   return __builtin_aarch64_udot_laneqv16qi_uuuus (__r, __a, __b, __index);
33010 }
33011
33012 __extension__ extern __inline int32x2_t
33013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33014 vdot_lane_s32 (int32x2_t __r, int8x8_t __a, int8x8_t __b, const int __index)
33015 {
33016   return __builtin_aarch64_sdot_lanev8qi (__r, __a, __b, __index);
33017 }
33018
33019 __extension__ extern __inline int32x2_t
33020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33021 vdot_laneq_s32 (int32x2_t __r, int8x8_t __a, int8x16_t __b, const int __index)
33022 {
33023   return __builtin_aarch64_sdot_laneqv8qi (__r, __a, __b, __index);
33024 }
33025
33026 __extension__ extern __inline int32x4_t
33027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33028 vdotq_lane_s32 (int32x4_t __r, int8x16_t __a, int8x8_t __b, const int __index)
33029 {
33030   return __builtin_aarch64_sdot_lanev16qi (__r, __a, __b, __index);
33031 }
33032
33033 __extension__ extern __inline int32x4_t
33034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33035 vdotq_laneq_s32 (int32x4_t __r, int8x16_t __a, int8x16_t __b, const int __index)
33036 {
33037   return __builtin_aarch64_sdot_laneqv16qi (__r, __a, __b, __index);
33038 }
33039 #pragma GCC pop_options
33040
33041 #pragma GCC push_options
33042 #pragma GCC target(("arch=armv8.2-a+sm4"))
33043
33044 __extension__ extern __inline uint32x4_t
33045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33046 vsm3ss1q_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
33047 {
33048   return __builtin_aarch64_sm3ss1qv4si_uuuu (__a, __b, __c);
33049 }
33050
33051 __extension__ extern __inline uint32x4_t
33052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33053 vsm3tt1aq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c, const int __imm2)
33054 {
33055   return __builtin_aarch64_sm3tt1aqv4si_uuuus (__a, __b, __c, __imm2);
33056 }
33057
33058 __extension__ extern __inline uint32x4_t
33059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33060 vsm3tt1bq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c, const int __imm2)
33061 {
33062   return __builtin_aarch64_sm3tt1bqv4si_uuuus (__a, __b, __c, __imm2);
33063 }
33064
33065 __extension__ extern __inline uint32x4_t
33066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33067 vsm3tt2aq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c, const int __imm2)
33068 {
33069   return __builtin_aarch64_sm3tt2aqv4si_uuuus (__a, __b, __c, __imm2);
33070 }
33071
33072 __extension__ extern __inline uint32x4_t
33073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33074 vsm3tt2bq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c, const int __imm2)
33075 {
33076   return __builtin_aarch64_sm3tt2bqv4si_uuuus (__a, __b, __c, __imm2);
33077 }
33078
33079 __extension__ extern __inline uint32x4_t
33080 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33081 vsm3partw1q_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
33082 {
33083   return __builtin_aarch64_sm3partw1qv4si_uuuu (__a, __b, __c);
33084 }
33085 __extension__ extern __inline uint32x4_t
33086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33087 vsm3partw2q_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
33088 {
33089   return __builtin_aarch64_sm3partw2qv4si_uuuu (__a, __b, __c);
33090 }
33091
33092 __extension__ extern __inline uint32x4_t
33093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33094 vsm4eq_u32 (uint32x4_t __a, uint32x4_t __b)
33095 {
33096   return __builtin_aarch64_sm4eqv4si_uuu (__a, __b);
33097 }
33098
33099 __extension__ extern __inline uint32x4_t
33100 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33101 vsm4ekeyq_u32 (uint32x4_t __a, uint32x4_t __b)
33102 {
33103   return __builtin_aarch64_sm4ekeyqv4si_uuu (__a, __b);
33104 }
33105
33106 #pragma GCC pop_options
33107
33108 #pragma GCC push_options
33109 #pragma GCC target(("arch=armv8.2-a+crypto"))
33110
33111 __extension__ extern __inline uint64x2_t
33112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33113 vsha512hq_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
33114 {
33115   return __builtin_aarch64_crypto_sha512hqv2di_uuuu (__a, __b, __c);
33116 }
33117
33118 __extension__ extern __inline uint64x2_t
33119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33120 vsha512h2q_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
33121 {
33122   return __builtin_aarch64_crypto_sha512h2qv2di_uuuu (__a, __b, __c);
33123 }
33124
33125 __extension__ extern __inline uint64x2_t
33126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33127 vsha512su0q_u64 (uint64x2_t __a, uint64x2_t __b)
33128 {
33129   return __builtin_aarch64_crypto_sha512su0qv2di_uuu (__a, __b);
33130 }
33131
33132 __extension__ extern __inline uint64x2_t
33133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33134 vsha512su1q_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
33135 {
33136   return __builtin_aarch64_crypto_sha512su1qv2di_uuuu (__a, __b, __c);
33137 }
33138
33139 __extension__ extern __inline uint8x16_t
33140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33141 veor3q_u8 (uint8x16_t __a, uint8x16_t __b, uint8x16_t __c)
33142 {
33143   return __builtin_aarch64_eor3qv16qi_uuuu (__a, __b, __c);
33144 }
33145
33146 __extension__ extern __inline uint16x8_t
33147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33148 veor3q_u16 (uint16x8_t __a, uint16x8_t __b, uint16x8_t __c)
33149 {
33150   return __builtin_aarch64_eor3qv8hi_uuuu (__a, __b, __c);
33151 }
33152
33153 __extension__ extern __inline uint32x4_t
33154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33155 veor3q_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
33156 {
33157   return __builtin_aarch64_eor3qv4si_uuuu (__a, __b, __c);
33158 }
33159
33160 __extension__ extern __inline uint64x2_t
33161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33162 veor3q_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
33163 {
33164   return __builtin_aarch64_eor3qv2di_uuuu (__a, __b, __c);
33165 }
33166
33167
33168 __extension__ extern __inline int8x16_t
33169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33170 veor3q_s8 (int8x16_t __a, int8x16_t __b, int8x16_t __c)
33171 {
33172   return __builtin_aarch64_eor3qv16qi (__a, __b, __c);
33173 }
33174
33175 __extension__ extern __inline int16x8_t
33176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33177 veor3q_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
33178 {
33179   return __builtin_aarch64_eor3qv8hi (__a, __b, __c);
33180 }
33181
33182 __extension__ extern __inline int32x4_t
33183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33184 veor3q_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
33185 {
33186   return __builtin_aarch64_eor3qv4si (__a, __b, __c);
33187 }
33188
33189 __extension__ extern __inline int64x2_t
33190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33191 veor3q_s64 (int64x2_t __a, int64x2_t __b, int64x2_t __c)
33192 {
33193   return __builtin_aarch64_eor3qv2di (__a, __b, __c);
33194 }
33195
33196 __extension__ extern __inline uint64x2_t
33197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33198 vrax1q_u64 (uint64x2_t __a, uint64x2_t __b)
33199 {
33200   return __builtin_aarch64_rax1qv2di_uuu (__a, __b);
33201 }
33202
33203 __extension__ extern __inline uint64x2_t
33204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33205 vxarq_u64 (uint64x2_t __a, uint64x2_t __b, const int imm6)
33206 {
33207   return __builtin_aarch64_xarqv2di_uuus (__a, __b,imm6);
33208 }
33209
33210 __extension__ extern __inline uint8x16_t
33211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33212 vbcaxq_u8 (uint8x16_t __a, uint8x16_t __b, uint8x16_t __c)
33213 {
33214   return __builtin_aarch64_bcaxqv16qi_uuuu (__a, __b, __c);
33215 }
33216
33217 __extension__ extern __inline uint16x8_t
33218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33219 vbcaxq_u16 (uint16x8_t __a, uint16x8_t __b, uint16x8_t __c)
33220 {
33221   return __builtin_aarch64_bcaxqv8hi_uuuu (__a, __b, __c);
33222 }
33223
33224 __extension__ extern __inline uint32x4_t
33225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33226 vbcaxq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
33227 {
33228   return __builtin_aarch64_bcaxqv4si_uuuu (__a, __b, __c);
33229 }
33230
33231 __extension__ extern __inline uint64x2_t
33232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33233 vbcaxq_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
33234 {
33235   return __builtin_aarch64_bcaxqv2di_uuuu (__a, __b, __c);
33236 }
33237
33238 __extension__ extern __inline int8x16_t
33239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33240 vbcaxq_s8 (int8x16_t __a, int8x16_t __b, int8x16_t __c)
33241 {
33242   return __builtin_aarch64_bcaxqv16qi (__a, __b, __c);
33243 }
33244
33245 __extension__ extern __inline int16x8_t
33246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33247 vbcaxq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
33248 {
33249   return __builtin_aarch64_bcaxqv8hi (__a, __b, __c);
33250 }
33251
33252 __extension__ extern __inline int32x4_t
33253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33254 vbcaxq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
33255 {
33256   return __builtin_aarch64_bcaxqv4si (__a, __b, __c);
33257 }
33258
33259 __extension__ extern __inline int64x2_t
33260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33261 vbcaxq_s64 (int64x2_t __a, int64x2_t __b, int64x2_t __c)
33262 {
33263   return __builtin_aarch64_bcaxqv2di (__a, __b, __c);
33264 }
33265
33266
33267 #pragma GCC pop_options
33268
33269 #pragma GCC push_options
33270 #pragma GCC target(("arch=armv8.2-a+fp16fml"))
33271
33272 __extension__ extern __inline float32x2_t
33273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33274 vfmlal_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b)
33275 {
33276   return __builtin_aarch64_fmlal_lowv2sf (__r, __a, __b);
33277 }
33278
33279 __extension__ extern __inline float32x2_t
33280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33281 vfmlsl_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b)
33282 {
33283   return __builtin_aarch64_fmlsl_lowv2sf (__r, __a, __b);
33284 }
33285
33286 __extension__ extern __inline float32x4_t
33287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33288 vfmlalq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b)
33289 {
33290   return __builtin_aarch64_fmlalq_lowv4sf (__r, __a, __b);
33291 }
33292
33293 __extension__ extern __inline float32x4_t
33294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33295 vfmlslq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b)
33296 {
33297   return __builtin_aarch64_fmlslq_lowv4sf (__r, __a, __b);
33298 }
33299
33300 __extension__ extern __inline float32x2_t
33301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33302 vfmlal_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b)
33303 {
33304   return __builtin_aarch64_fmlal_highv2sf (__r, __a, __b);
33305 }
33306
33307 __extension__ extern __inline float32x2_t
33308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33309 vfmlsl_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b)
33310 {
33311   return __builtin_aarch64_fmlsl_highv2sf (__r, __a, __b);
33312 }
33313
33314 __extension__ extern __inline float32x4_t
33315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33316 vfmlalq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b)
33317 {
33318   return __builtin_aarch64_fmlalq_highv4sf (__r, __a, __b);
33319 }
33320
33321 __extension__ extern __inline float32x4_t
33322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33323 vfmlslq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b)
33324 {
33325   return __builtin_aarch64_fmlslq_highv4sf (__r, __a, __b);
33326 }
33327
33328 __extension__ extern __inline float32x2_t
33329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33330 vfmlal_lane_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b,
33331                      const int __lane)
33332 {
33333   return __builtin_aarch64_fmlal_lane_lowv2sf (__r, __a, __b, __lane);
33334 }
33335
33336 __extension__ extern __inline float32x2_t
33337 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33338 vfmlsl_lane_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b,
33339                      const int __lane)
33340 {
33341   return __builtin_aarch64_fmlsl_lane_lowv2sf (__r, __a, __b, __lane);
33342 }
33343
33344 __extension__ extern __inline float32x2_t
33345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33346 vfmlal_laneq_low_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b,
33347                       const int __lane)
33348 {
33349   return __builtin_aarch64_fmlal_laneq_lowv2sf (__r, __a, __b, __lane);
33350 }
33351
33352 __extension__ extern __inline float32x2_t
33353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33354 vfmlsl_laneq_low_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b,
33355                       const int __lane)
33356 {
33357   return __builtin_aarch64_fmlsl_laneq_lowv2sf (__r, __a, __b, __lane);
33358 }
33359
33360 __extension__ extern __inline float32x4_t
33361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33362 vfmlalq_lane_low_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b,
33363                       const int __lane)
33364 {
33365   return __builtin_aarch64_fmlalq_lane_lowv4sf (__r, __a, __b, __lane);
33366 }
33367
33368 __extension__ extern __inline float32x4_t
33369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33370 vfmlslq_lane_low_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b,
33371                       const int __lane)
33372 {
33373   return __builtin_aarch64_fmlslq_lane_lowv4sf (__r, __a, __b, __lane);
33374 }
33375
33376 __extension__ extern __inline float32x4_t
33377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33378 vfmlalq_laneq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b,
33379                        const int __lane)
33380 {
33381   return __builtin_aarch64_fmlalq_laneq_lowv4sf (__r, __a, __b, __lane);
33382 }
33383
33384 __extension__ extern __inline float32x4_t
33385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33386 vfmlslq_laneq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b,
33387                       const int __lane)
33388 {
33389   return __builtin_aarch64_fmlslq_laneq_lowv4sf (__r, __a, __b, __lane);
33390 }
33391
33392 __extension__ extern __inline float32x2_t
33393 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33394 vfmlal_lane_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b,
33395                      const int __lane)
33396 {
33397   return __builtin_aarch64_fmlal_lane_highv2sf (__r, __a, __b, __lane);
33398 }
33399
33400 __extension__ extern __inline float32x2_t
33401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33402 vfmlsl_lane_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b,
33403                      const int __lane)
33404 {
33405   return __builtin_aarch64_fmlsl_lane_highv2sf (__r, __a, __b, __lane);
33406 }
33407
33408 __extension__ extern __inline float32x2_t
33409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33410 vfmlal_laneq_high_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b,
33411                       const int __lane)
33412 {
33413   return __builtin_aarch64_fmlal_laneq_highv2sf (__r, __a, __b, __lane);
33414 }
33415
33416 __extension__ extern __inline float32x2_t
33417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33418 vfmlsl_laneq_high_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b,
33419                       const int __lane)
33420 {
33421   return __builtin_aarch64_fmlsl_laneq_highv2sf (__r, __a, __b, __lane);
33422 }
33423
33424 __extension__ extern __inline float32x4_t
33425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33426 vfmlalq_lane_high_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b,
33427                       const int __lane)
33428 {
33429   return __builtin_aarch64_fmlalq_lane_highv4sf (__r, __a, __b, __lane);
33430 }
33431
33432 __extension__ extern __inline float32x4_t
33433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33434 vfmlslq_lane_high_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b,
33435                       const int __lane)
33436 {
33437   return __builtin_aarch64_fmlslq_lane_highv4sf (__r, __a, __b, __lane);
33438 }
33439
33440 __extension__ extern __inline float32x4_t
33441 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33442 vfmlalq_laneq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b,
33443                        const int __lane)
33444 {
33445   return __builtin_aarch64_fmlalq_laneq_highv4sf (__r, __a, __b, __lane);
33446 }
33447
33448 __extension__ extern __inline float32x4_t
33449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
33450 vfmlslq_laneq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b,
33451                       const int __lane)
33452 {
33453   return __builtin_aarch64_fmlslq_laneq_highv4sf (__r, __a, __b, __lane);
33454 }
33455
33456 #pragma GCC pop_options
33457
33458 #undef __aarch64_vget_lane_any
33459
33460 #undef __aarch64_vdup_lane_any
33461 #undef __aarch64_vdup_lane_f16
33462 #undef __aarch64_vdup_lane_f32
33463 #undef __aarch64_vdup_lane_f64
33464 #undef __aarch64_vdup_lane_p8
33465 #undef __aarch64_vdup_lane_p16
33466 #undef __aarch64_vdup_lane_s8
33467 #undef __aarch64_vdup_lane_s16
33468 #undef __aarch64_vdup_lane_s32
33469 #undef __aarch64_vdup_lane_s64
33470 #undef __aarch64_vdup_lane_u8
33471 #undef __aarch64_vdup_lane_u16
33472 #undef __aarch64_vdup_lane_u32
33473 #undef __aarch64_vdup_lane_u64
33474 #undef __aarch64_vdup_laneq_f16
33475 #undef __aarch64_vdup_laneq_f32
33476 #undef __aarch64_vdup_laneq_f64
33477 #undef __aarch64_vdup_laneq_p8
33478 #undef __aarch64_vdup_laneq_p16
33479 #undef __aarch64_vdup_laneq_s8
33480 #undef __aarch64_vdup_laneq_s16
33481 #undef __aarch64_vdup_laneq_s32
33482 #undef __aarch64_vdup_laneq_s64
33483 #undef __aarch64_vdup_laneq_u8
33484 #undef __aarch64_vdup_laneq_u16
33485 #undef __aarch64_vdup_laneq_u32
33486 #undef __aarch64_vdup_laneq_u64
33487 #undef __aarch64_vdupq_lane_f16
33488 #undef __aarch64_vdupq_lane_f32
33489 #undef __aarch64_vdupq_lane_f64
33490 #undef __aarch64_vdupq_lane_p8
33491 #undef __aarch64_vdupq_lane_p16
33492 #undef __aarch64_vdupq_lane_s8
33493 #undef __aarch64_vdupq_lane_s16
33494 #undef __aarch64_vdupq_lane_s32
33495 #undef __aarch64_vdupq_lane_s64
33496 #undef __aarch64_vdupq_lane_u8
33497 #undef __aarch64_vdupq_lane_u16
33498 #undef __aarch64_vdupq_lane_u32
33499 #undef __aarch64_vdupq_lane_u64
33500 #undef __aarch64_vdupq_laneq_f16
33501 #undef __aarch64_vdupq_laneq_f32
33502 #undef __aarch64_vdupq_laneq_f64
33503 #undef __aarch64_vdupq_laneq_p8
33504 #undef __aarch64_vdupq_laneq_p16
33505 #undef __aarch64_vdupq_laneq_s8
33506 #undef __aarch64_vdupq_laneq_s16
33507 #undef __aarch64_vdupq_laneq_s32
33508 #undef __aarch64_vdupq_laneq_s64
33509 #undef __aarch64_vdupq_laneq_u8
33510 #undef __aarch64_vdupq_laneq_u16
33511 #undef __aarch64_vdupq_laneq_u32
33512 #undef __aarch64_vdupq_laneq_u64
33513
33514 #endif