gcc/config/aarch64/arm_neon.h

   1 /* ARM NEON intrinsics include file.
   2
   3    Copyright (C) 2011-2016 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 poly16x4_t
3510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3511 vreinterpret_p16_f16 (float16x4_t __a)
3512 {
3513   return (poly16x4_t) __a;
3514 }
3515
3516 __extension__ extern __inline poly16x4_t
3517 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3518 vreinterpret_p16_f64 (float64x1_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_s8 (int8x8_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_s16 (int16x4_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_s32 (int32x2_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_s64 (int64x1_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_f32 (float32x2_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_u8 (uint8x8_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_u16 (uint16x4_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_u32 (uint32x2_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_u64 (uint64x1_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_p8 (poly8x8_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_p64 (poly64x1_t __a)
3596 {
3597   return (poly16x4_t) __a;
3598 }
3599
3600 __extension__ extern __inline poly16x8_t
3601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3602 vreinterpretq_p16_f64 (float64x2_t __a)
3603 {
3604   return (poly16x8_t) __a;
3605 }
3606
3607 __extension__ extern __inline poly16x8_t
3608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3609 vreinterpretq_p16_s8 (int8x16_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_s16 (int16x8_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_s32 (int32x4_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_s64 (int64x2_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_f16 (float16x8_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_f32 (float32x4_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_u8 (uint8x16_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_u16 (uint16x8_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_u32 (uint32x4_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_u64 (uint64x2_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_p8 (poly8x16_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_p64 (poly64x2_t __a)
3687 {
3688   return (poly16x8_t) __a;
3689 }
3690
3691 __extension__ extern __inline poly64x1_t
3692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3693 vreinterpret_p64_f16 (float16x4_t __a)
3694 {
3695   return (poly64x1_t) __a;
3696 }
3697
3698 __extension__ extern __inline poly64x1_t
3699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3700 vreinterpret_p64_f64 (float64x1_t __a)
3701 {
3702   return (poly64x1_t) __a;
3703 }
3704
3705 __extension__ extern __inline poly64x1_t
3706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3707 vreinterpret_p64_s8 (int8x8_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_s16 (int16x4_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_s32 (int32x2_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_s64 (int64x1_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_f32 (float32x2_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_u8 (uint8x8_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_u16 (uint16x4_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_u32 (uint32x2_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_u64 (uint64x1_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_p8 (poly8x8_t __a)
3771 {
3772   return (poly64x1_t) __a;
3773 }
3774
3775 __extension__ extern __inline poly64x2_t
3776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3777 vreinterpretq_p64_f64 (float64x2_t __a)
3778 {
3779   return (poly64x2_t) __a;
3780 }
3781
3782 __extension__ extern __inline poly64x2_t
3783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3784 vreinterpretq_p64_s8 (int8x16_t __a)
3785 {
3786   return (poly64x2_t) __a;
3787 }
3788
3789 __extension__ extern __inline poly64x2_t
3790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3791 vreinterpretq_p64_s16 (int16x8_t __a)
3792 {
3793   return (poly64x2_t) __a;
3794 }
3795
3796 __extension__ extern __inline poly64x2_t
3797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3798 vreinterpretq_p64_s32 (int32x4_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_s64 (int64x2_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_f16 (float16x8_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_f32 (float32x4_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_u8 (uint8x16_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_u16 (uint16x8_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_u32 (uint32x4_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_u64 (uint64x2_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_p8 (poly8x16_t __a)
3855 {
3856   return (poly64x2_t) __a;
3857 }
3858
3859 __extension__ extern __inline float16x4_t
3860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3861 vreinterpret_f16_f64 (float64x1_t __a)
3862 {
3863   return (float16x4_t) __a;
3864 }
3865
3866 __extension__ extern __inline float16x4_t
3867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3868 vreinterpret_f16_s8 (int8x8_t __a)
3869 {
3870   return (float16x4_t) __a;
3871 }
3872
3873 __extension__ extern __inline float16x4_t
3874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3875 vreinterpret_f16_s16 (int16x4_t __a)
3876 {
3877   return (float16x4_t) __a;
3878 }
3879
3880 __extension__ extern __inline float16x4_t
3881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3882 vreinterpret_f16_s32 (int32x2_t __a)
3883 {
3884   return (float16x4_t) __a;
3885 }
3886
3887 __extension__ extern __inline float16x4_t
3888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3889 vreinterpret_f16_s64 (int64x1_t __a)
3890 {
3891   return (float16x4_t) __a;
3892 }
3893
3894 __extension__ extern __inline float16x4_t
3895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3896 vreinterpret_f16_f32 (float32x2_t __a)
3897 {
3898   return (float16x4_t) __a;
3899 }
3900
3901 __extension__ extern __inline float16x4_t
3902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3903 vreinterpret_f16_u8 (uint8x8_t __a)
3904 {
3905   return (float16x4_t) __a;
3906 }
3907
3908 __extension__ extern __inline float16x4_t
3909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3910 vreinterpret_f16_u16 (uint16x4_t __a)
3911 {
3912   return (float16x4_t) __a;
3913 }
3914
3915 __extension__ extern __inline float16x4_t
3916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3917 vreinterpret_f16_u32 (uint32x2_t __a)
3918 {
3919   return (float16x4_t) __a;
3920 }
3921
3922 __extension__ extern __inline float16x4_t
3923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3924 vreinterpret_f16_u64 (uint64x1_t __a)
3925 {
3926   return (float16x4_t) __a;
3927 }
3928
3929 __extension__ extern __inline float16x4_t
3930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3931 vreinterpret_f16_p8 (poly8x8_t __a)
3932 {
3933   return (float16x4_t) __a;
3934 }
3935
3936 __extension__ extern __inline float16x4_t
3937 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3938 vreinterpret_f16_p16 (poly16x4_t __a)
3939 {
3940   return (float16x4_t) __a;
3941 }
3942
3943 __extension__ extern __inline float16x4_t
3944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3945 vreinterpret_f16_p64 (poly64x1_t __a)
3946 {
3947   return (float16x4_t) __a;
3948 }
3949
3950 __extension__ extern __inline float16x8_t
3951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3952 vreinterpretq_f16_f64 (float64x2_t __a)
3953 {
3954   return (float16x8_t) __a;
3955 }
3956
3957 __extension__ extern __inline float16x8_t
3958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3959 vreinterpretq_f16_s8 (int8x16_t __a)
3960 {
3961   return (float16x8_t) __a;
3962 }
3963
3964 __extension__ extern __inline float16x8_t
3965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3966 vreinterpretq_f16_s16 (int16x8_t __a)
3967 {
3968   return (float16x8_t) __a;
3969 }
3970
3971 __extension__ extern __inline float16x8_t
3972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3973 vreinterpretq_f16_s32 (int32x4_t __a)
3974 {
3975   return (float16x8_t) __a;
3976 }
3977
3978 __extension__ extern __inline float16x8_t
3979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3980 vreinterpretq_f16_s64 (int64x2_t __a)
3981 {
3982   return (float16x8_t) __a;
3983 }
3984
3985 __extension__ extern __inline float16x8_t
3986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3987 vreinterpretq_f16_f32 (float32x4_t __a)
3988 {
3989   return (float16x8_t) __a;
3990 }
3991
3992 __extension__ extern __inline float16x8_t
3993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
3994 vreinterpretq_f16_u8 (uint8x16_t __a)
3995 {
3996   return (float16x8_t) __a;
3997 }
3998
3999 __extension__ extern __inline float16x8_t
4000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4001 vreinterpretq_f16_u16 (uint16x8_t __a)
4002 {
4003   return (float16x8_t) __a;
4004 }
4005
4006 __extension__ extern __inline float16x8_t
4007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4008 vreinterpretq_f16_u32 (uint32x4_t __a)
4009 {
4010   return (float16x8_t) __a;
4011 }
4012
4013 __extension__ extern __inline float16x8_t
4014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4015 vreinterpretq_f16_u64 (uint64x2_t __a)
4016 {
4017   return (float16x8_t) __a;
4018 }
4019
4020 __extension__ extern __inline float16x8_t
4021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4022 vreinterpretq_f16_p8 (poly8x16_t __a)
4023 {
4024   return (float16x8_t) __a;
4025 }
4026
4027 __extension__ extern __inline float16x8_t
4028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4029 vreinterpretq_f16_p16 (poly16x8_t __a)
4030 {
4031   return (float16x8_t) __a;
4032 }
4033
4034 __extension__ extern __inline float16x8_t
4035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4036 vreinterpretq_f16_p64 (poly64x2_t __a)
4037 {
4038   return (float16x8_t) __a;
4039 }
4040
4041 __extension__ extern __inline float32x2_t
4042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4043 vreinterpret_f32_f16 (float16x4_t __a)
4044 {
4045   return (float32x2_t) __a;
4046 }
4047
4048 __extension__ extern __inline float32x2_t
4049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4050 vreinterpret_f32_f64 (float64x1_t __a)
4051 {
4052   return (float32x2_t) __a;
4053 }
4054
4055 __extension__ extern __inline float32x2_t
4056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4057 vreinterpret_f32_s8 (int8x8_t __a)
4058 {
4059   return (float32x2_t) __a;
4060 }
4061
4062 __extension__ extern __inline float32x2_t
4063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4064 vreinterpret_f32_s16 (int16x4_t __a)
4065 {
4066   return (float32x2_t) __a;
4067 }
4068
4069 __extension__ extern __inline float32x2_t
4070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4071 vreinterpret_f32_s32 (int32x2_t __a)
4072 {
4073   return (float32x2_t) __a;
4074 }
4075
4076 __extension__ extern __inline float32x2_t
4077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4078 vreinterpret_f32_s64 (int64x1_t __a)
4079 {
4080   return (float32x2_t) __a;
4081 }
4082
4083 __extension__ extern __inline float32x2_t
4084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4085 vreinterpret_f32_u8 (uint8x8_t __a)
4086 {
4087   return (float32x2_t) __a;
4088 }
4089
4090 __extension__ extern __inline float32x2_t
4091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4092 vreinterpret_f32_u16 (uint16x4_t __a)
4093 {
4094   return (float32x2_t) __a;
4095 }
4096
4097 __extension__ extern __inline float32x2_t
4098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4099 vreinterpret_f32_u32 (uint32x2_t __a)
4100 {
4101   return (float32x2_t) __a;
4102 }
4103
4104 __extension__ extern __inline float32x2_t
4105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4106 vreinterpret_f32_u64 (uint64x1_t __a)
4107 {
4108   return (float32x2_t) __a;
4109 }
4110
4111 __extension__ extern __inline float32x2_t
4112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4113 vreinterpret_f32_p8 (poly8x8_t __a)
4114 {
4115   return (float32x2_t) __a;
4116 }
4117
4118 __extension__ extern __inline float32x2_t
4119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4120 vreinterpret_f32_p16 (poly16x4_t __a)
4121 {
4122   return (float32x2_t) __a;
4123 }
4124
4125 __extension__ extern __inline float32x2_t
4126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4127 vreinterpret_f32_p64 (poly64x1_t __a)
4128 {
4129   return (float32x2_t) __a;
4130 }
4131
4132 __extension__ extern __inline float32x4_t
4133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4134 vreinterpretq_f32_f16 (float16x8_t __a)
4135 {
4136   return (float32x4_t) __a;
4137 }
4138
4139 __extension__ extern __inline float32x4_t
4140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4141 vreinterpretq_f32_f64 (float64x2_t __a)
4142 {
4143   return (float32x4_t) __a;
4144 }
4145
4146 __extension__ extern __inline float32x4_t
4147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4148 vreinterpretq_f32_s8 (int8x16_t __a)
4149 {
4150   return (float32x4_t) __a;
4151 }
4152
4153 __extension__ extern __inline float32x4_t
4154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4155 vreinterpretq_f32_s16 (int16x8_t __a)
4156 {
4157   return (float32x4_t) __a;
4158 }
4159
4160 __extension__ extern __inline float32x4_t
4161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4162 vreinterpretq_f32_s32 (int32x4_t __a)
4163 {
4164   return (float32x4_t) __a;
4165 }
4166
4167 __extension__ extern __inline float32x4_t
4168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4169 vreinterpretq_f32_s64 (int64x2_t __a)
4170 {
4171   return (float32x4_t) __a;
4172 }
4173
4174 __extension__ extern __inline float32x4_t
4175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4176 vreinterpretq_f32_u8 (uint8x16_t __a)
4177 {
4178   return (float32x4_t) __a;
4179 }
4180
4181 __extension__ extern __inline float32x4_t
4182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4183 vreinterpretq_f32_u16 (uint16x8_t __a)
4184 {
4185   return (float32x4_t) __a;
4186 }
4187
4188 __extension__ extern __inline float32x4_t
4189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4190 vreinterpretq_f32_u32 (uint32x4_t __a)
4191 {
4192   return (float32x4_t) __a;
4193 }
4194
4195 __extension__ extern __inline float32x4_t
4196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4197 vreinterpretq_f32_u64 (uint64x2_t __a)
4198 {
4199   return (float32x4_t) __a;
4200 }
4201
4202 __extension__ extern __inline float32x4_t
4203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4204 vreinterpretq_f32_p8 (poly8x16_t __a)
4205 {
4206   return (float32x4_t) __a;
4207 }
4208
4209 __extension__ extern __inline float32x4_t
4210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4211 vreinterpretq_f32_p16 (poly16x8_t __a)
4212 {
4213   return (float32x4_t) __a;
4214 }
4215
4216 __extension__ extern __inline float32x4_t
4217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4218 vreinterpretq_f32_p64 (poly64x2_t __a)
4219 {
4220   return (float32x4_t) __a;
4221 }
4222
4223 __extension__ extern __inline float64x1_t
4224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4225 vreinterpret_f64_f16 (float16x4_t __a)
4226 {
4227   return (float64x1_t) __a;
4228 }
4229
4230 __extension__ extern __inline float64x1_t
4231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4232 vreinterpret_f64_f32 (float32x2_t __a)
4233 {
4234   return (float64x1_t) __a;
4235 }
4236
4237 __extension__ extern __inline float64x1_t
4238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4239 vreinterpret_f64_p8 (poly8x8_t __a)
4240 {
4241   return (float64x1_t) __a;
4242 }
4243
4244 __extension__ extern __inline float64x1_t
4245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4246 vreinterpret_f64_p16 (poly16x4_t __a)
4247 {
4248   return (float64x1_t) __a;
4249 }
4250
4251 __extension__ extern __inline float64x1_t
4252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4253 vreinterpret_f64_p64 (poly64x1_t __a)
4254 {
4255   return (float64x1_t) __a;
4256 }
4257
4258 __extension__ extern __inline float64x1_t
4259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4260 vreinterpret_f64_s8 (int8x8_t __a)
4261 {
4262   return (float64x1_t) __a;
4263 }
4264
4265 __extension__ extern __inline float64x1_t
4266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4267 vreinterpret_f64_s16 (int16x4_t __a)
4268 {
4269   return (float64x1_t) __a;
4270 }
4271
4272 __extension__ extern __inline float64x1_t
4273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4274 vreinterpret_f64_s32 (int32x2_t __a)
4275 {
4276   return (float64x1_t) __a;
4277 }
4278
4279 __extension__ extern __inline float64x1_t
4280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4281 vreinterpret_f64_s64 (int64x1_t __a)
4282 {
4283   return (float64x1_t) __a;
4284 }
4285
4286 __extension__ extern __inline float64x1_t
4287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4288 vreinterpret_f64_u8 (uint8x8_t __a)
4289 {
4290   return (float64x1_t) __a;
4291 }
4292
4293 __extension__ extern __inline float64x1_t
4294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4295 vreinterpret_f64_u16 (uint16x4_t __a)
4296 {
4297   return (float64x1_t) __a;
4298 }
4299
4300 __extension__ extern __inline float64x1_t
4301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4302 vreinterpret_f64_u32 (uint32x2_t __a)
4303 {
4304   return (float64x1_t) __a;
4305 }
4306
4307 __extension__ extern __inline float64x1_t
4308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4309 vreinterpret_f64_u64 (uint64x1_t __a)
4310 {
4311   return (float64x1_t) __a;
4312 }
4313
4314 __extension__ extern __inline float64x2_t
4315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4316 vreinterpretq_f64_f16 (float16x8_t __a)
4317 {
4318   return (float64x2_t) __a;
4319 }
4320
4321 __extension__ extern __inline float64x2_t
4322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4323 vreinterpretq_f64_f32 (float32x4_t __a)
4324 {
4325   return (float64x2_t) __a;
4326 }
4327
4328 __extension__ extern __inline float64x2_t
4329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4330 vreinterpretq_f64_p8 (poly8x16_t __a)
4331 {
4332   return (float64x2_t) __a;
4333 }
4334
4335 __extension__ extern __inline float64x2_t
4336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4337 vreinterpretq_f64_p16 (poly16x8_t __a)
4338 {
4339   return (float64x2_t) __a;
4340 }
4341
4342 __extension__ extern __inline float64x2_t
4343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4344 vreinterpretq_f64_p64 (poly64x2_t __a)
4345 {
4346   return (float64x2_t) __a;
4347 }
4348
4349 __extension__ extern __inline float64x2_t
4350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4351 vreinterpretq_f64_s8 (int8x16_t __a)
4352 {
4353   return (float64x2_t) __a;
4354 }
4355
4356 __extension__ extern __inline float64x2_t
4357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4358 vreinterpretq_f64_s16 (int16x8_t __a)
4359 {
4360   return (float64x2_t) __a;
4361 }
4362
4363 __extension__ extern __inline float64x2_t
4364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4365 vreinterpretq_f64_s32 (int32x4_t __a)
4366 {
4367   return (float64x2_t) __a;
4368 }
4369
4370 __extension__ extern __inline float64x2_t
4371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4372 vreinterpretq_f64_s64 (int64x2_t __a)
4373 {
4374   return (float64x2_t) __a;
4375 }
4376
4377 __extension__ extern __inline float64x2_t
4378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4379 vreinterpretq_f64_u8 (uint8x16_t __a)
4380 {
4381   return (float64x2_t) __a;
4382 }
4383
4384 __extension__ extern __inline float64x2_t
4385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4386 vreinterpretq_f64_u16 (uint16x8_t __a)
4387 {
4388   return (float64x2_t) __a;
4389 }
4390
4391 __extension__ extern __inline float64x2_t
4392 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4393 vreinterpretq_f64_u32 (uint32x4_t __a)
4394 {
4395   return (float64x2_t) __a;
4396 }
4397
4398 __extension__ extern __inline float64x2_t
4399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4400 vreinterpretq_f64_u64 (uint64x2_t __a)
4401 {
4402   return (float64x2_t) __a;
4403 }
4404
4405 __extension__ extern __inline int64x1_t
4406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4407 vreinterpret_s64_f16 (float16x4_t __a)
4408 {
4409   return (int64x1_t) __a;
4410 }
4411
4412 __extension__ extern __inline int64x1_t
4413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4414 vreinterpret_s64_f64 (float64x1_t __a)
4415 {
4416   return (int64x1_t) __a;
4417 }
4418
4419 __extension__ extern __inline int64x1_t
4420 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4421 vreinterpret_s64_s8 (int8x8_t __a)
4422 {
4423   return (int64x1_t) __a;
4424 }
4425
4426 __extension__ extern __inline int64x1_t
4427 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4428 vreinterpret_s64_s16 (int16x4_t __a)
4429 {
4430   return (int64x1_t) __a;
4431 }
4432
4433 __extension__ extern __inline int64x1_t
4434 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4435 vreinterpret_s64_s32 (int32x2_t __a)
4436 {
4437   return (int64x1_t) __a;
4438 }
4439
4440 __extension__ extern __inline int64x1_t
4441 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4442 vreinterpret_s64_f32 (float32x2_t __a)
4443 {
4444   return (int64x1_t) __a;
4445 }
4446
4447 __extension__ extern __inline int64x1_t
4448 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4449 vreinterpret_s64_u8 (uint8x8_t __a)
4450 {
4451   return (int64x1_t) __a;
4452 }
4453
4454 __extension__ extern __inline int64x1_t
4455 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4456 vreinterpret_s64_u16 (uint16x4_t __a)
4457 {
4458   return (int64x1_t) __a;
4459 }
4460
4461 __extension__ extern __inline int64x1_t
4462 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4463 vreinterpret_s64_u32 (uint32x2_t __a)
4464 {
4465   return (int64x1_t) __a;
4466 }
4467
4468 __extension__ extern __inline int64x1_t
4469 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4470 vreinterpret_s64_u64 (uint64x1_t __a)
4471 {
4472   return (int64x1_t) __a;
4473 }
4474
4475 __extension__ extern __inline int64x1_t
4476 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4477 vreinterpret_s64_p8 (poly8x8_t __a)
4478 {
4479   return (int64x1_t) __a;
4480 }
4481
4482 __extension__ extern __inline int64x1_t
4483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4484 vreinterpret_s64_p16 (poly16x4_t __a)
4485 {
4486   return (int64x1_t) __a;
4487 }
4488
4489 __extension__ extern __inline int64x1_t
4490 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4491 vreinterpret_s64_p64 (poly64x1_t __a)
4492 {
4493   return (int64x1_t) __a;
4494 }
4495
4496 __extension__ extern __inline int64x2_t
4497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4498 vreinterpretq_s64_f64 (float64x2_t __a)
4499 {
4500   return (int64x2_t) __a;
4501 }
4502
4503 __extension__ extern __inline int64x2_t
4504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4505 vreinterpretq_s64_s8 (int8x16_t __a)
4506 {
4507   return (int64x2_t) __a;
4508 }
4509
4510 __extension__ extern __inline int64x2_t
4511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4512 vreinterpretq_s64_s16 (int16x8_t __a)
4513 {
4514   return (int64x2_t) __a;
4515 }
4516
4517 __extension__ extern __inline int64x2_t
4518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4519 vreinterpretq_s64_s32 (int32x4_t __a)
4520 {
4521   return (int64x2_t) __a;
4522 }
4523
4524 __extension__ extern __inline int64x2_t
4525 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4526 vreinterpretq_s64_f16 (float16x8_t __a)
4527 {
4528   return (int64x2_t) __a;
4529 }
4530
4531 __extension__ extern __inline int64x2_t
4532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4533 vreinterpretq_s64_f32 (float32x4_t __a)
4534 {
4535   return (int64x2_t) __a;
4536 }
4537
4538 __extension__ extern __inline int64x2_t
4539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4540 vreinterpretq_s64_u8 (uint8x16_t __a)
4541 {
4542   return (int64x2_t) __a;
4543 }
4544
4545 __extension__ extern __inline int64x2_t
4546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4547 vreinterpretq_s64_u16 (uint16x8_t __a)
4548 {
4549   return (int64x2_t) __a;
4550 }
4551
4552 __extension__ extern __inline int64x2_t
4553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4554 vreinterpretq_s64_u32 (uint32x4_t __a)
4555 {
4556   return (int64x2_t) __a;
4557 }
4558
4559 __extension__ extern __inline int64x2_t
4560 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4561 vreinterpretq_s64_u64 (uint64x2_t __a)
4562 {
4563   return (int64x2_t) __a;
4564 }
4565
4566 __extension__ extern __inline int64x2_t
4567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4568 vreinterpretq_s64_p8 (poly8x16_t __a)
4569 {
4570   return (int64x2_t) __a;
4571 }
4572
4573 __extension__ extern __inline int64x2_t
4574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4575 vreinterpretq_s64_p16 (poly16x8_t __a)
4576 {
4577   return (int64x2_t) __a;
4578 }
4579
4580 __extension__ extern __inline int64x2_t
4581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4582 vreinterpretq_s64_p64 (poly64x2_t __a)
4583 {
4584   return (int64x2_t) __a;
4585 }
4586
4587 __extension__ extern __inline uint64x1_t
4588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4589 vreinterpret_u64_f16 (float16x4_t __a)
4590 {
4591   return (uint64x1_t) __a;
4592 }
4593
4594 __extension__ extern __inline uint64x1_t
4595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4596 vreinterpret_u64_f64 (float64x1_t __a)
4597 {
4598   return (uint64x1_t) __a;
4599 }
4600
4601 __extension__ extern __inline uint64x1_t
4602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4603 vreinterpret_u64_s8 (int8x8_t __a)
4604 {
4605   return (uint64x1_t) __a;
4606 }
4607
4608 __extension__ extern __inline uint64x1_t
4609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4610 vreinterpret_u64_s16 (int16x4_t __a)
4611 {
4612   return (uint64x1_t) __a;
4613 }
4614
4615 __extension__ extern __inline uint64x1_t
4616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4617 vreinterpret_u64_s32 (int32x2_t __a)
4618 {
4619   return (uint64x1_t) __a;
4620 }
4621
4622 __extension__ extern __inline uint64x1_t
4623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4624 vreinterpret_u64_s64 (int64x1_t __a)
4625 {
4626   return (uint64x1_t) __a;
4627 }
4628
4629 __extension__ extern __inline uint64x1_t
4630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4631 vreinterpret_u64_f32 (float32x2_t __a)
4632 {
4633   return (uint64x1_t) __a;
4634 }
4635
4636 __extension__ extern __inline uint64x1_t
4637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4638 vreinterpret_u64_u8 (uint8x8_t __a)
4639 {
4640   return (uint64x1_t) __a;
4641 }
4642
4643 __extension__ extern __inline uint64x1_t
4644 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4645 vreinterpret_u64_u16 (uint16x4_t __a)
4646 {
4647   return (uint64x1_t) __a;
4648 }
4649
4650 __extension__ extern __inline uint64x1_t
4651 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4652 vreinterpret_u64_u32 (uint32x2_t __a)
4653 {
4654   return (uint64x1_t) __a;
4655 }
4656
4657 __extension__ extern __inline uint64x1_t
4658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4659 vreinterpret_u64_p8 (poly8x8_t __a)
4660 {
4661   return (uint64x1_t) __a;
4662 }
4663
4664 __extension__ extern __inline uint64x1_t
4665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4666 vreinterpret_u64_p16 (poly16x4_t __a)
4667 {
4668   return (uint64x1_t) __a;
4669 }
4670
4671 __extension__ extern __inline uint64x1_t
4672 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4673 vreinterpret_u64_p64 (poly64x1_t __a)
4674 {
4675   return (uint64x1_t) __a;
4676 }
4677
4678 __extension__ extern __inline uint64x2_t
4679 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4680 vreinterpretq_u64_f64 (float64x2_t __a)
4681 {
4682   return (uint64x2_t) __a;
4683 }
4684
4685 __extension__ extern __inline uint64x2_t
4686 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4687 vreinterpretq_u64_s8 (int8x16_t __a)
4688 {
4689   return (uint64x2_t) __a;
4690 }
4691
4692 __extension__ extern __inline uint64x2_t
4693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4694 vreinterpretq_u64_s16 (int16x8_t __a)
4695 {
4696   return (uint64x2_t) __a;
4697 }
4698
4699 __extension__ extern __inline uint64x2_t
4700 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4701 vreinterpretq_u64_s32 (int32x4_t __a)
4702 {
4703   return (uint64x2_t) __a;
4704 }
4705
4706 __extension__ extern __inline uint64x2_t
4707 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4708 vreinterpretq_u64_s64 (int64x2_t __a)
4709 {
4710   return (uint64x2_t) __a;
4711 }
4712
4713 __extension__ extern __inline uint64x2_t
4714 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4715 vreinterpretq_u64_f16 (float16x8_t __a)
4716 {
4717   return (uint64x2_t) __a;
4718 }
4719
4720 __extension__ extern __inline uint64x2_t
4721 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4722 vreinterpretq_u64_f32 (float32x4_t __a)
4723 {
4724   return (uint64x2_t) __a;
4725 }
4726
4727 __extension__ extern __inline uint64x2_t
4728 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4729 vreinterpretq_u64_u8 (uint8x16_t __a)
4730 {
4731   return (uint64x2_t) __a;
4732 }
4733
4734 __extension__ extern __inline uint64x2_t
4735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4736 vreinterpretq_u64_u16 (uint16x8_t __a)
4737 {
4738   return (uint64x2_t) __a;
4739 }
4740
4741 __extension__ extern __inline uint64x2_t
4742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4743 vreinterpretq_u64_u32 (uint32x4_t __a)
4744 {
4745   return (uint64x2_t) __a;
4746 }
4747
4748 __extension__ extern __inline uint64x2_t
4749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4750 vreinterpretq_u64_p8 (poly8x16_t __a)
4751 {
4752   return (uint64x2_t) __a;
4753 }
4754
4755 __extension__ extern __inline uint64x2_t
4756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4757 vreinterpretq_u64_p16 (poly16x8_t __a)
4758 {
4759   return (uint64x2_t) __a;
4760 }
4761
4762 __extension__ extern __inline uint64x2_t
4763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4764 vreinterpretq_u64_p64 (poly64x2_t __a)
4765 {
4766   return (uint64x2_t) __a;
4767 }
4768
4769 __extension__ extern __inline int8x8_t
4770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4771 vreinterpret_s8_f16 (float16x4_t __a)
4772 {
4773   return (int8x8_t) __a;
4774 }
4775
4776 __extension__ extern __inline int8x8_t
4777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4778 vreinterpret_s8_f64 (float64x1_t __a)
4779 {
4780   return (int8x8_t) __a;
4781 }
4782
4783 __extension__ extern __inline int8x8_t
4784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4785 vreinterpret_s8_s16 (int16x4_t __a)
4786 {
4787   return (int8x8_t) __a;
4788 }
4789
4790 __extension__ extern __inline int8x8_t
4791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4792 vreinterpret_s8_s32 (int32x2_t __a)
4793 {
4794   return (int8x8_t) __a;
4795 }
4796
4797 __extension__ extern __inline int8x8_t
4798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4799 vreinterpret_s8_s64 (int64x1_t __a)
4800 {
4801   return (int8x8_t) __a;
4802 }
4803
4804 __extension__ extern __inline int8x8_t
4805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4806 vreinterpret_s8_f32 (float32x2_t __a)
4807 {
4808   return (int8x8_t) __a;
4809 }
4810
4811 __extension__ extern __inline int8x8_t
4812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4813 vreinterpret_s8_u8 (uint8x8_t __a)
4814 {
4815   return (int8x8_t) __a;
4816 }
4817
4818 __extension__ extern __inline int8x8_t
4819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4820 vreinterpret_s8_u16 (uint16x4_t __a)
4821 {
4822   return (int8x8_t) __a;
4823 }
4824
4825 __extension__ extern __inline int8x8_t
4826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4827 vreinterpret_s8_u32 (uint32x2_t __a)
4828 {
4829   return (int8x8_t) __a;
4830 }
4831
4832 __extension__ extern __inline int8x8_t
4833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4834 vreinterpret_s8_u64 (uint64x1_t __a)
4835 {
4836   return (int8x8_t) __a;
4837 }
4838
4839 __extension__ extern __inline int8x8_t
4840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4841 vreinterpret_s8_p8 (poly8x8_t __a)
4842 {
4843   return (int8x8_t) __a;
4844 }
4845
4846 __extension__ extern __inline int8x8_t
4847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4848 vreinterpret_s8_p16 (poly16x4_t __a)
4849 {
4850   return (int8x8_t) __a;
4851 }
4852
4853 __extension__ extern __inline int8x8_t
4854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4855 vreinterpret_s8_p64 (poly64x1_t __a)
4856 {
4857   return (int8x8_t) __a;
4858 }
4859
4860 __extension__ extern __inline int8x16_t
4861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4862 vreinterpretq_s8_f64 (float64x2_t __a)
4863 {
4864   return (int8x16_t) __a;
4865 }
4866
4867 __extension__ extern __inline int8x16_t
4868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4869 vreinterpretq_s8_s16 (int16x8_t __a)
4870 {
4871   return (int8x16_t) __a;
4872 }
4873
4874 __extension__ extern __inline int8x16_t
4875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4876 vreinterpretq_s8_s32 (int32x4_t __a)
4877 {
4878   return (int8x16_t) __a;
4879 }
4880
4881 __extension__ extern __inline int8x16_t
4882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4883 vreinterpretq_s8_s64 (int64x2_t __a)
4884 {
4885   return (int8x16_t) __a;
4886 }
4887
4888 __extension__ extern __inline int8x16_t
4889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4890 vreinterpretq_s8_f16 (float16x8_t __a)
4891 {
4892   return (int8x16_t) __a;
4893 }
4894
4895 __extension__ extern __inline int8x16_t
4896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4897 vreinterpretq_s8_f32 (float32x4_t __a)
4898 {
4899   return (int8x16_t) __a;
4900 }
4901
4902 __extension__ extern __inline int8x16_t
4903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4904 vreinterpretq_s8_u8 (uint8x16_t __a)
4905 {
4906   return (int8x16_t) __a;
4907 }
4908
4909 __extension__ extern __inline int8x16_t
4910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4911 vreinterpretq_s8_u16 (uint16x8_t __a)
4912 {
4913   return (int8x16_t) __a;
4914 }
4915
4916 __extension__ extern __inline int8x16_t
4917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4918 vreinterpretq_s8_u32 (uint32x4_t __a)
4919 {
4920   return (int8x16_t) __a;
4921 }
4922
4923 __extension__ extern __inline int8x16_t
4924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4925 vreinterpretq_s8_u64 (uint64x2_t __a)
4926 {
4927   return (int8x16_t) __a;
4928 }
4929
4930 __extension__ extern __inline int8x16_t
4931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4932 vreinterpretq_s8_p8 (poly8x16_t __a)
4933 {
4934   return (int8x16_t) __a;
4935 }
4936
4937 __extension__ extern __inline int8x16_t
4938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4939 vreinterpretq_s8_p16 (poly16x8_t __a)
4940 {
4941   return (int8x16_t) __a;
4942 }
4943
4944 __extension__ extern __inline int8x16_t
4945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4946 vreinterpretq_s8_p64 (poly64x2_t __a)
4947 {
4948   return (int8x16_t) __a;
4949 }
4950
4951 __extension__ extern __inline int16x4_t
4952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4953 vreinterpret_s16_f16 (float16x4_t __a)
4954 {
4955   return (int16x4_t) __a;
4956 }
4957
4958 __extension__ extern __inline int16x4_t
4959 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4960 vreinterpret_s16_f64 (float64x1_t __a)
4961 {
4962   return (int16x4_t) __a;
4963 }
4964
4965 __extension__ extern __inline int16x4_t
4966 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4967 vreinterpret_s16_s8 (int8x8_t __a)
4968 {
4969   return (int16x4_t) __a;
4970 }
4971
4972 __extension__ extern __inline int16x4_t
4973 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4974 vreinterpret_s16_s32 (int32x2_t __a)
4975 {
4976   return (int16x4_t) __a;
4977 }
4978
4979 __extension__ extern __inline int16x4_t
4980 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4981 vreinterpret_s16_s64 (int64x1_t __a)
4982 {
4983   return (int16x4_t) __a;
4984 }
4985
4986 __extension__ extern __inline int16x4_t
4987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4988 vreinterpret_s16_f32 (float32x2_t __a)
4989 {
4990   return (int16x4_t) __a;
4991 }
4992
4993 __extension__ extern __inline int16x4_t
4994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
4995 vreinterpret_s16_u8 (uint8x8_t __a)
4996 {
4997   return (int16x4_t) __a;
4998 }
4999
5000 __extension__ extern __inline int16x4_t
5001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5002 vreinterpret_s16_u16 (uint16x4_t __a)
5003 {
5004   return (int16x4_t) __a;
5005 }
5006
5007 __extension__ extern __inline int16x4_t
5008 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5009 vreinterpret_s16_u32 (uint32x2_t __a)
5010 {
5011   return (int16x4_t) __a;
5012 }
5013
5014 __extension__ extern __inline int16x4_t
5015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5016 vreinterpret_s16_u64 (uint64x1_t __a)
5017 {
5018   return (int16x4_t) __a;
5019 }
5020
5021 __extension__ extern __inline int16x4_t
5022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5023 vreinterpret_s16_p8 (poly8x8_t __a)
5024 {
5025   return (int16x4_t) __a;
5026 }
5027
5028 __extension__ extern __inline int16x4_t
5029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5030 vreinterpret_s16_p16 (poly16x4_t __a)
5031 {
5032   return (int16x4_t) __a;
5033 }
5034
5035 __extension__ extern __inline int16x4_t
5036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5037 vreinterpret_s16_p64 (poly64x1_t __a)
5038 {
5039   return (int16x4_t) __a;
5040 }
5041
5042 __extension__ extern __inline int16x8_t
5043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5044 vreinterpretq_s16_f64 (float64x2_t __a)
5045 {
5046   return (int16x8_t) __a;
5047 }
5048
5049 __extension__ extern __inline int16x8_t
5050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5051 vreinterpretq_s16_s8 (int8x16_t __a)
5052 {
5053   return (int16x8_t) __a;
5054 }
5055
5056 __extension__ extern __inline int16x8_t
5057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5058 vreinterpretq_s16_s32 (int32x4_t __a)
5059 {
5060   return (int16x8_t) __a;
5061 }
5062
5063 __extension__ extern __inline int16x8_t
5064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5065 vreinterpretq_s16_s64 (int64x2_t __a)
5066 {
5067   return (int16x8_t) __a;
5068 }
5069
5070 __extension__ extern __inline int16x8_t
5071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5072 vreinterpretq_s16_f16 (float16x8_t __a)
5073 {
5074   return (int16x8_t) __a;
5075 }
5076
5077 __extension__ extern __inline int16x8_t
5078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5079 vreinterpretq_s16_f32 (float32x4_t __a)
5080 {
5081   return (int16x8_t) __a;
5082 }
5083
5084 __extension__ extern __inline int16x8_t
5085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5086 vreinterpretq_s16_u8 (uint8x16_t __a)
5087 {
5088   return (int16x8_t) __a;
5089 }
5090
5091 __extension__ extern __inline int16x8_t
5092 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5093 vreinterpretq_s16_u16 (uint16x8_t __a)
5094 {
5095   return (int16x8_t) __a;
5096 }
5097
5098 __extension__ extern __inline int16x8_t
5099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5100 vreinterpretq_s16_u32 (uint32x4_t __a)
5101 {
5102   return (int16x8_t) __a;
5103 }
5104
5105 __extension__ extern __inline int16x8_t
5106 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5107 vreinterpretq_s16_u64 (uint64x2_t __a)
5108 {
5109   return (int16x8_t) __a;
5110 }
5111
5112 __extension__ extern __inline int16x8_t
5113 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5114 vreinterpretq_s16_p8 (poly8x16_t __a)
5115 {
5116   return (int16x8_t) __a;
5117 }
5118
5119 __extension__ extern __inline int16x8_t
5120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5121 vreinterpretq_s16_p16 (poly16x8_t __a)
5122 {
5123   return (int16x8_t) __a;
5124 }
5125
5126 __extension__ extern __inline int16x8_t
5127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5128 vreinterpretq_s16_p64 (poly64x2_t __a)
5129 {
5130   return (int16x8_t) __a;
5131 }
5132
5133 __extension__ extern __inline int32x2_t
5134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5135 vreinterpret_s32_f16 (float16x4_t __a)
5136 {
5137   return (int32x2_t) __a;
5138 }
5139
5140 __extension__ extern __inline int32x2_t
5141 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5142 vreinterpret_s32_f64 (float64x1_t __a)
5143 {
5144   return (int32x2_t) __a;
5145 }
5146
5147 __extension__ extern __inline int32x2_t
5148 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5149 vreinterpret_s32_s8 (int8x8_t __a)
5150 {
5151   return (int32x2_t) __a;
5152 }
5153
5154 __extension__ extern __inline int32x2_t
5155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5156 vreinterpret_s32_s16 (int16x4_t __a)
5157 {
5158   return (int32x2_t) __a;
5159 }
5160
5161 __extension__ extern __inline int32x2_t
5162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5163 vreinterpret_s32_s64 (int64x1_t __a)
5164 {
5165   return (int32x2_t) __a;
5166 }
5167
5168 __extension__ extern __inline int32x2_t
5169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5170 vreinterpret_s32_f32 (float32x2_t __a)
5171 {
5172   return (int32x2_t) __a;
5173 }
5174
5175 __extension__ extern __inline int32x2_t
5176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5177 vreinterpret_s32_u8 (uint8x8_t __a)
5178 {
5179   return (int32x2_t) __a;
5180 }
5181
5182 __extension__ extern __inline int32x2_t
5183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5184 vreinterpret_s32_u16 (uint16x4_t __a)
5185 {
5186   return (int32x2_t) __a;
5187 }
5188
5189 __extension__ extern __inline int32x2_t
5190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5191 vreinterpret_s32_u32 (uint32x2_t __a)
5192 {
5193   return (int32x2_t) __a;
5194 }
5195
5196 __extension__ extern __inline int32x2_t
5197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5198 vreinterpret_s32_u64 (uint64x1_t __a)
5199 {
5200   return (int32x2_t) __a;
5201 }
5202
5203 __extension__ extern __inline int32x2_t
5204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5205 vreinterpret_s32_p8 (poly8x8_t __a)
5206 {
5207   return (int32x2_t) __a;
5208 }
5209
5210 __extension__ extern __inline int32x2_t
5211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5212 vreinterpret_s32_p16 (poly16x4_t __a)
5213 {
5214   return (int32x2_t) __a;
5215 }
5216
5217 __extension__ extern __inline int32x2_t
5218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5219 vreinterpret_s32_p64 (poly64x1_t __a)
5220 {
5221   return (int32x2_t) __a;
5222 }
5223
5224 __extension__ extern __inline int32x4_t
5225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5226 vreinterpretq_s32_f64 (float64x2_t __a)
5227 {
5228   return (int32x4_t) __a;
5229 }
5230
5231 __extension__ extern __inline int32x4_t
5232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5233 vreinterpretq_s32_s8 (int8x16_t __a)
5234 {
5235   return (int32x4_t) __a;
5236 }
5237
5238 __extension__ extern __inline int32x4_t
5239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5240 vreinterpretq_s32_s16 (int16x8_t __a)
5241 {
5242   return (int32x4_t) __a;
5243 }
5244
5245 __extension__ extern __inline int32x4_t
5246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5247 vreinterpretq_s32_s64 (int64x2_t __a)
5248 {
5249   return (int32x4_t) __a;
5250 }
5251
5252 __extension__ extern __inline int32x4_t
5253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5254 vreinterpretq_s32_f16 (float16x8_t __a)
5255 {
5256   return (int32x4_t) __a;
5257 }
5258
5259 __extension__ extern __inline int32x4_t
5260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5261 vreinterpretq_s32_f32 (float32x4_t __a)
5262 {
5263   return (int32x4_t) __a;
5264 }
5265
5266 __extension__ extern __inline int32x4_t
5267 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5268 vreinterpretq_s32_u8 (uint8x16_t __a)
5269 {
5270   return (int32x4_t) __a;
5271 }
5272
5273 __extension__ extern __inline int32x4_t
5274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5275 vreinterpretq_s32_u16 (uint16x8_t __a)
5276 {
5277   return (int32x4_t) __a;
5278 }
5279
5280 __extension__ extern __inline int32x4_t
5281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5282 vreinterpretq_s32_u32 (uint32x4_t __a)
5283 {
5284   return (int32x4_t) __a;
5285 }
5286
5287 __extension__ extern __inline int32x4_t
5288 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5289 vreinterpretq_s32_u64 (uint64x2_t __a)
5290 {
5291   return (int32x4_t) __a;
5292 }
5293
5294 __extension__ extern __inline int32x4_t
5295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5296 vreinterpretq_s32_p8 (poly8x16_t __a)
5297 {
5298   return (int32x4_t) __a;
5299 }
5300
5301 __extension__ extern __inline int32x4_t
5302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5303 vreinterpretq_s32_p16 (poly16x8_t __a)
5304 {
5305   return (int32x4_t) __a;
5306 }
5307
5308 __extension__ extern __inline int32x4_t
5309 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5310 vreinterpretq_s32_p64 (poly64x2_t __a)
5311 {
5312   return (int32x4_t) __a;
5313 }
5314
5315 __extension__ extern __inline uint8x8_t
5316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5317 vreinterpret_u8_f16 (float16x4_t __a)
5318 {
5319   return (uint8x8_t) __a;
5320 }
5321
5322 __extension__ extern __inline uint8x8_t
5323 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5324 vreinterpret_u8_f64 (float64x1_t __a)
5325 {
5326   return (uint8x8_t) __a;
5327 }
5328
5329 __extension__ extern __inline uint8x8_t
5330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5331 vreinterpret_u8_s8 (int8x8_t __a)
5332 {
5333   return (uint8x8_t) __a;
5334 }
5335
5336 __extension__ extern __inline uint8x8_t
5337 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5338 vreinterpret_u8_s16 (int16x4_t __a)
5339 {
5340   return (uint8x8_t) __a;
5341 }
5342
5343 __extension__ extern __inline uint8x8_t
5344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5345 vreinterpret_u8_s32 (int32x2_t __a)
5346 {
5347   return (uint8x8_t) __a;
5348 }
5349
5350 __extension__ extern __inline uint8x8_t
5351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5352 vreinterpret_u8_s64 (int64x1_t __a)
5353 {
5354   return (uint8x8_t) __a;
5355 }
5356
5357 __extension__ extern __inline uint8x8_t
5358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5359 vreinterpret_u8_f32 (float32x2_t __a)
5360 {
5361   return (uint8x8_t) __a;
5362 }
5363
5364 __extension__ extern __inline uint8x8_t
5365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5366 vreinterpret_u8_u16 (uint16x4_t __a)
5367 {
5368   return (uint8x8_t) __a;
5369 }
5370
5371 __extension__ extern __inline uint8x8_t
5372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5373 vreinterpret_u8_u32 (uint32x2_t __a)
5374 {
5375   return (uint8x8_t) __a;
5376 }
5377
5378 __extension__ extern __inline uint8x8_t
5379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5380 vreinterpret_u8_u64 (uint64x1_t __a)
5381 {
5382   return (uint8x8_t) __a;
5383 }
5384
5385 __extension__ extern __inline uint8x8_t
5386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5387 vreinterpret_u8_p8 (poly8x8_t __a)
5388 {
5389   return (uint8x8_t) __a;
5390 }
5391
5392 __extension__ extern __inline uint8x8_t
5393 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5394 vreinterpret_u8_p16 (poly16x4_t __a)
5395 {
5396   return (uint8x8_t) __a;
5397 }
5398
5399 __extension__ extern __inline uint8x8_t
5400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5401 vreinterpret_u8_p64 (poly64x1_t __a)
5402 {
5403   return (uint8x8_t) __a;
5404 }
5405
5406 __extension__ extern __inline uint8x16_t
5407 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5408 vreinterpretq_u8_f64 (float64x2_t __a)
5409 {
5410   return (uint8x16_t) __a;
5411 }
5412
5413 __extension__ extern __inline uint8x16_t
5414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5415 vreinterpretq_u8_s8 (int8x16_t __a)
5416 {
5417   return (uint8x16_t) __a;
5418 }
5419
5420 __extension__ extern __inline uint8x16_t
5421 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5422 vreinterpretq_u8_s16 (int16x8_t __a)
5423 {
5424   return (uint8x16_t) __a;
5425 }
5426
5427 __extension__ extern __inline uint8x16_t
5428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5429 vreinterpretq_u8_s32 (int32x4_t __a)
5430 {
5431   return (uint8x16_t) __a;
5432 }
5433
5434 __extension__ extern __inline uint8x16_t
5435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5436 vreinterpretq_u8_s64 (int64x2_t __a)
5437 {
5438   return (uint8x16_t) __a;
5439 }
5440
5441 __extension__ extern __inline uint8x16_t
5442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5443 vreinterpretq_u8_f16 (float16x8_t __a)
5444 {
5445   return (uint8x16_t) __a;
5446 }
5447
5448 __extension__ extern __inline uint8x16_t
5449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5450 vreinterpretq_u8_f32 (float32x4_t __a)
5451 {
5452   return (uint8x16_t) __a;
5453 }
5454
5455 __extension__ extern __inline uint8x16_t
5456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5457 vreinterpretq_u8_u16 (uint16x8_t __a)
5458 {
5459   return (uint8x16_t) __a;
5460 }
5461
5462 __extension__ extern __inline uint8x16_t
5463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5464 vreinterpretq_u8_u32 (uint32x4_t __a)
5465 {
5466   return (uint8x16_t) __a;
5467 }
5468
5469 __extension__ extern __inline uint8x16_t
5470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5471 vreinterpretq_u8_u64 (uint64x2_t __a)
5472 {
5473   return (uint8x16_t) __a;
5474 }
5475
5476 __extension__ extern __inline uint8x16_t
5477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5478 vreinterpretq_u8_p8 (poly8x16_t __a)
5479 {
5480   return (uint8x16_t) __a;
5481 }
5482
5483 __extension__ extern __inline uint8x16_t
5484 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5485 vreinterpretq_u8_p16 (poly16x8_t __a)
5486 {
5487   return (uint8x16_t) __a;
5488 }
5489
5490 __extension__ extern __inline uint8x16_t
5491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5492 vreinterpretq_u8_p64 (poly64x2_t __a)
5493 {
5494   return (uint8x16_t) __a;
5495 }
5496
5497 __extension__ extern __inline uint16x4_t
5498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5499 vreinterpret_u16_f16 (float16x4_t __a)
5500 {
5501   return (uint16x4_t) __a;
5502 }
5503
5504 __extension__ extern __inline uint16x4_t
5505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5506 vreinterpret_u16_f64 (float64x1_t __a)
5507 {
5508   return (uint16x4_t) __a;
5509 }
5510
5511 __extension__ extern __inline uint16x4_t
5512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5513 vreinterpret_u16_s8 (int8x8_t __a)
5514 {
5515   return (uint16x4_t) __a;
5516 }
5517
5518 __extension__ extern __inline uint16x4_t
5519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5520 vreinterpret_u16_s16 (int16x4_t __a)
5521 {
5522   return (uint16x4_t) __a;
5523 }
5524
5525 __extension__ extern __inline uint16x4_t
5526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5527 vreinterpret_u16_s32 (int32x2_t __a)
5528 {
5529   return (uint16x4_t) __a;
5530 }
5531
5532 __extension__ extern __inline uint16x4_t
5533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5534 vreinterpret_u16_s64 (int64x1_t __a)
5535 {
5536   return (uint16x4_t) __a;
5537 }
5538
5539 __extension__ extern __inline uint16x4_t
5540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5541 vreinterpret_u16_f32 (float32x2_t __a)
5542 {
5543   return (uint16x4_t) __a;
5544 }
5545
5546 __extension__ extern __inline uint16x4_t
5547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5548 vreinterpret_u16_u8 (uint8x8_t __a)
5549 {
5550   return (uint16x4_t) __a;
5551 }
5552
5553 __extension__ extern __inline uint16x4_t
5554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5555 vreinterpret_u16_u32 (uint32x2_t __a)
5556 {
5557   return (uint16x4_t) __a;
5558 }
5559
5560 __extension__ extern __inline uint16x4_t
5561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5562 vreinterpret_u16_u64 (uint64x1_t __a)
5563 {
5564   return (uint16x4_t) __a;
5565 }
5566
5567 __extension__ extern __inline uint16x4_t
5568 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5569 vreinterpret_u16_p8 (poly8x8_t __a)
5570 {
5571   return (uint16x4_t) __a;
5572 }
5573
5574 __extension__ extern __inline uint16x4_t
5575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5576 vreinterpret_u16_p16 (poly16x4_t __a)
5577 {
5578   return (uint16x4_t) __a;
5579 }
5580
5581 __extension__ extern __inline uint16x4_t
5582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5583 vreinterpret_u16_p64 (poly64x1_t __a)
5584 {
5585   return (uint16x4_t) __a;
5586 }
5587
5588 __extension__ extern __inline uint16x8_t
5589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5590 vreinterpretq_u16_f64 (float64x2_t __a)
5591 {
5592   return (uint16x8_t) __a;
5593 }
5594
5595 __extension__ extern __inline uint16x8_t
5596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5597 vreinterpretq_u16_s8 (int8x16_t __a)
5598 {
5599   return (uint16x8_t) __a;
5600 }
5601
5602 __extension__ extern __inline uint16x8_t
5603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5604 vreinterpretq_u16_s16 (int16x8_t __a)
5605 {
5606   return (uint16x8_t) __a;
5607 }
5608
5609 __extension__ extern __inline uint16x8_t
5610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5611 vreinterpretq_u16_s32 (int32x4_t __a)
5612 {
5613   return (uint16x8_t) __a;
5614 }
5615
5616 __extension__ extern __inline uint16x8_t
5617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5618 vreinterpretq_u16_s64 (int64x2_t __a)
5619 {
5620   return (uint16x8_t) __a;
5621 }
5622
5623 __extension__ extern __inline uint16x8_t
5624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5625 vreinterpretq_u16_f16 (float16x8_t __a)
5626 {
5627   return (uint16x8_t) __a;
5628 }
5629
5630 __extension__ extern __inline uint16x8_t
5631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5632 vreinterpretq_u16_f32 (float32x4_t __a)
5633 {
5634   return (uint16x8_t) __a;
5635 }
5636
5637 __extension__ extern __inline uint16x8_t
5638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5639 vreinterpretq_u16_u8 (uint8x16_t __a)
5640 {
5641   return (uint16x8_t) __a;
5642 }
5643
5644 __extension__ extern __inline uint16x8_t
5645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5646 vreinterpretq_u16_u32 (uint32x4_t __a)
5647 {
5648   return (uint16x8_t) __a;
5649 }
5650
5651 __extension__ extern __inline uint16x8_t
5652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5653 vreinterpretq_u16_u64 (uint64x2_t __a)
5654 {
5655   return (uint16x8_t) __a;
5656 }
5657
5658 __extension__ extern __inline uint16x8_t
5659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5660 vreinterpretq_u16_p8 (poly8x16_t __a)
5661 {
5662   return (uint16x8_t) __a;
5663 }
5664
5665 __extension__ extern __inline uint16x8_t
5666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5667 vreinterpretq_u16_p16 (poly16x8_t __a)
5668 {
5669   return (uint16x8_t) __a;
5670 }
5671
5672 __extension__ extern __inline uint16x8_t
5673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5674 vreinterpretq_u16_p64 (poly64x2_t __a)
5675 {
5676   return (uint16x8_t) __a;
5677 }
5678
5679 __extension__ extern __inline uint32x2_t
5680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5681 vreinterpret_u32_f16 (float16x4_t __a)
5682 {
5683   return (uint32x2_t) __a;
5684 }
5685
5686 __extension__ extern __inline uint32x2_t
5687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5688 vreinterpret_u32_f64 (float64x1_t __a)
5689 {
5690   return (uint32x2_t) __a;
5691 }
5692
5693 __extension__ extern __inline uint32x2_t
5694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5695 vreinterpret_u32_s8 (int8x8_t __a)
5696 {
5697   return (uint32x2_t) __a;
5698 }
5699
5700 __extension__ extern __inline uint32x2_t
5701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5702 vreinterpret_u32_s16 (int16x4_t __a)
5703 {
5704   return (uint32x2_t) __a;
5705 }
5706
5707 __extension__ extern __inline uint32x2_t
5708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5709 vreinterpret_u32_s32 (int32x2_t __a)
5710 {
5711   return (uint32x2_t) __a;
5712 }
5713
5714 __extension__ extern __inline uint32x2_t
5715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5716 vreinterpret_u32_s64 (int64x1_t __a)
5717 {
5718   return (uint32x2_t) __a;
5719 }
5720
5721 __extension__ extern __inline uint32x2_t
5722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5723 vreinterpret_u32_f32 (float32x2_t __a)
5724 {
5725   return (uint32x2_t) __a;
5726 }
5727
5728 __extension__ extern __inline uint32x2_t
5729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5730 vreinterpret_u32_u8 (uint8x8_t __a)
5731 {
5732   return (uint32x2_t) __a;
5733 }
5734
5735 __extension__ extern __inline uint32x2_t
5736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5737 vreinterpret_u32_u16 (uint16x4_t __a)
5738 {
5739   return (uint32x2_t) __a;
5740 }
5741
5742 __extension__ extern __inline uint32x2_t
5743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5744 vreinterpret_u32_u64 (uint64x1_t __a)
5745 {
5746   return (uint32x2_t) __a;
5747 }
5748
5749 __extension__ extern __inline uint32x2_t
5750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5751 vreinterpret_u32_p8 (poly8x8_t __a)
5752 {
5753   return (uint32x2_t) __a;
5754 }
5755
5756 __extension__ extern __inline uint32x2_t
5757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5758 vreinterpret_u32_p16 (poly16x4_t __a)
5759 {
5760   return (uint32x2_t) __a;
5761 }
5762
5763 __extension__ extern __inline uint32x2_t
5764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5765 vreinterpret_u32_p64 (poly64x1_t __a)
5766 {
5767   return (uint32x2_t) __a;
5768 }
5769
5770 __extension__ extern __inline uint32x4_t
5771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5772 vreinterpretq_u32_f64 (float64x2_t __a)
5773 {
5774   return (uint32x4_t) __a;
5775 }
5776
5777 __extension__ extern __inline uint32x4_t
5778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5779 vreinterpretq_u32_s8 (int8x16_t __a)
5780 {
5781   return (uint32x4_t) __a;
5782 }
5783
5784 __extension__ extern __inline uint32x4_t
5785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5786 vreinterpretq_u32_s16 (int16x8_t __a)
5787 {
5788   return (uint32x4_t) __a;
5789 }
5790
5791 __extension__ extern __inline uint32x4_t
5792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5793 vreinterpretq_u32_s32 (int32x4_t __a)
5794 {
5795   return (uint32x4_t) __a;
5796 }
5797
5798 __extension__ extern __inline uint32x4_t
5799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5800 vreinterpretq_u32_s64 (int64x2_t __a)
5801 {
5802   return (uint32x4_t) __a;
5803 }
5804
5805 __extension__ extern __inline uint32x4_t
5806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5807 vreinterpretq_u32_f16 (float16x8_t __a)
5808 {
5809   return (uint32x4_t) __a;
5810 }
5811
5812 __extension__ extern __inline uint32x4_t
5813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5814 vreinterpretq_u32_f32 (float32x4_t __a)
5815 {
5816   return (uint32x4_t) __a;
5817 }
5818
5819 __extension__ extern __inline uint32x4_t
5820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5821 vreinterpretq_u32_u8 (uint8x16_t __a)
5822 {
5823   return (uint32x4_t) __a;
5824 }
5825
5826 __extension__ extern __inline uint32x4_t
5827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5828 vreinterpretq_u32_u16 (uint16x8_t __a)
5829 {
5830   return (uint32x4_t) __a;
5831 }
5832
5833 __extension__ extern __inline uint32x4_t
5834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5835 vreinterpretq_u32_u64 (uint64x2_t __a)
5836 {
5837   return (uint32x4_t) __a;
5838 }
5839
5840 __extension__ extern __inline uint32x4_t
5841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5842 vreinterpretq_u32_p8 (poly8x16_t __a)
5843 {
5844   return (uint32x4_t) __a;
5845 }
5846
5847 __extension__ extern __inline uint32x4_t
5848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5849 vreinterpretq_u32_p16 (poly16x8_t __a)
5850 {
5851   return (uint32x4_t) __a;
5852 }
5853
5854 __extension__ extern __inline uint32x4_t
5855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5856 vreinterpretq_u32_p64 (poly64x2_t __a)
5857 {
5858   return (uint32x4_t) __a;
5859 }
5860
5861 /* vset_lane  */
5862
5863 __extension__ extern __inline float16x4_t
5864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5865 vset_lane_f16 (float16_t __elem, float16x4_t __vec, const int __index)
5866 {
5867   return __aarch64_vset_lane_any (__elem, __vec, __index);
5868 }
5869
5870 __extension__ extern __inline float32x2_t
5871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5872 vset_lane_f32 (float32_t __elem, float32x2_t __vec, const int __index)
5873 {
5874   return __aarch64_vset_lane_any (__elem, __vec, __index);
5875 }
5876
5877 __extension__ extern __inline float64x1_t
5878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5879 vset_lane_f64 (float64_t __elem, float64x1_t __vec, const int __index)
5880 {
5881   return __aarch64_vset_lane_any (__elem, __vec, __index);
5882 }
5883
5884 __extension__ extern __inline poly8x8_t
5885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5886 vset_lane_p8 (poly8_t __elem, poly8x8_t __vec, const int __index)
5887 {
5888   return __aarch64_vset_lane_any (__elem, __vec, __index);
5889 }
5890
5891 __extension__ extern __inline poly16x4_t
5892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5893 vset_lane_p16 (poly16_t __elem, poly16x4_t __vec, const int __index)
5894 {
5895   return __aarch64_vset_lane_any (__elem, __vec, __index);
5896 }
5897
5898 __extension__ extern __inline poly64x1_t
5899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5900 vset_lane_p64 (poly64_t __elem, poly64x1_t __vec, const int __index)
5901 {
5902   return __aarch64_vset_lane_any (__elem, __vec, __index);
5903 }
5904
5905 __extension__ extern __inline int8x8_t
5906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5907 vset_lane_s8 (int8_t __elem, int8x8_t __vec, const int __index)
5908 {
5909   return __aarch64_vset_lane_any (__elem, __vec, __index);
5910 }
5911
5912 __extension__ extern __inline int16x4_t
5913 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5914 vset_lane_s16 (int16_t __elem, int16x4_t __vec, const int __index)
5915 {
5916   return __aarch64_vset_lane_any (__elem, __vec, __index);
5917 }
5918
5919 __extension__ extern __inline int32x2_t
5920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5921 vset_lane_s32 (int32_t __elem, int32x2_t __vec, const int __index)
5922 {
5923   return __aarch64_vset_lane_any (__elem, __vec, __index);
5924 }
5925
5926 __extension__ extern __inline int64x1_t
5927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5928 vset_lane_s64 (int64_t __elem, int64x1_t __vec, const int __index)
5929 {
5930   return __aarch64_vset_lane_any (__elem, __vec, __index);
5931 }
5932
5933 __extension__ extern __inline uint8x8_t
5934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5935 vset_lane_u8 (uint8_t __elem, uint8x8_t __vec, const int __index)
5936 {
5937   return __aarch64_vset_lane_any (__elem, __vec, __index);
5938 }
5939
5940 __extension__ extern __inline uint16x4_t
5941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5942 vset_lane_u16 (uint16_t __elem, uint16x4_t __vec, const int __index)
5943 {
5944   return __aarch64_vset_lane_any (__elem, __vec, __index);
5945 }
5946
5947 __extension__ extern __inline uint32x2_t
5948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5949 vset_lane_u32 (uint32_t __elem, uint32x2_t __vec, const int __index)
5950 {
5951   return __aarch64_vset_lane_any (__elem, __vec, __index);
5952 }
5953
5954 __extension__ extern __inline uint64x1_t
5955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5956 vset_lane_u64 (uint64_t __elem, uint64x1_t __vec, const int __index)
5957 {
5958   return __aarch64_vset_lane_any (__elem, __vec, __index);
5959 }
5960
5961 /* vsetq_lane  */
5962
5963 __extension__ extern __inline float16x8_t
5964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5965 vsetq_lane_f16 (float16_t __elem, float16x8_t __vec, const int __index)
5966 {
5967   return __aarch64_vset_lane_any (__elem, __vec, __index);
5968 }
5969
5970 __extension__ extern __inline float32x4_t
5971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5972 vsetq_lane_f32 (float32_t __elem, float32x4_t __vec, const int __index)
5973 {
5974   return __aarch64_vset_lane_any (__elem, __vec, __index);
5975 }
5976
5977 __extension__ extern __inline float64x2_t
5978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5979 vsetq_lane_f64 (float64_t __elem, float64x2_t __vec, const int __index)
5980 {
5981   return __aarch64_vset_lane_any (__elem, __vec, __index);
5982 }
5983
5984 __extension__ extern __inline poly8x16_t
5985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5986 vsetq_lane_p8 (poly8_t __elem, poly8x16_t __vec, const int __index)
5987 {
5988   return __aarch64_vset_lane_any (__elem, __vec, __index);
5989 }
5990
5991 __extension__ extern __inline poly16x8_t
5992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
5993 vsetq_lane_p16 (poly16_t __elem, poly16x8_t __vec, const int __index)
5994 {
5995   return __aarch64_vset_lane_any (__elem, __vec, __index);
5996 }
5997
5998 __extension__ extern __inline poly64x2_t
5999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6000 vsetq_lane_p64 (poly64_t __elem, poly64x2_t __vec, const int __index)
6001 {
6002   return __aarch64_vset_lane_any (__elem, __vec, __index);
6003 }
6004
6005 __extension__ extern __inline int8x16_t
6006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6007 vsetq_lane_s8 (int8_t __elem, int8x16_t __vec, const int __index)
6008 {
6009   return __aarch64_vset_lane_any (__elem, __vec, __index);
6010 }
6011
6012 __extension__ extern __inline int16x8_t
6013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6014 vsetq_lane_s16 (int16_t __elem, int16x8_t __vec, const int __index)
6015 {
6016   return __aarch64_vset_lane_any (__elem, __vec, __index);
6017 }
6018
6019 __extension__ extern __inline int32x4_t
6020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6021 vsetq_lane_s32 (int32_t __elem, int32x4_t __vec, const int __index)
6022 {
6023   return __aarch64_vset_lane_any (__elem, __vec, __index);
6024 }
6025
6026 __extension__ extern __inline int64x2_t
6027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6028 vsetq_lane_s64 (int64_t __elem, int64x2_t __vec, const int __index)
6029 {
6030   return __aarch64_vset_lane_any (__elem, __vec, __index);
6031 }
6032
6033 __extension__ extern __inline uint8x16_t
6034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6035 vsetq_lane_u8 (uint8_t __elem, uint8x16_t __vec, const int __index)
6036 {
6037   return __aarch64_vset_lane_any (__elem, __vec, __index);
6038 }
6039
6040 __extension__ extern __inline uint16x8_t
6041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6042 vsetq_lane_u16 (uint16_t __elem, uint16x8_t __vec, const int __index)
6043 {
6044   return __aarch64_vset_lane_any (__elem, __vec, __index);
6045 }
6046
6047 __extension__ extern __inline uint32x4_t
6048 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6049 vsetq_lane_u32 (uint32_t __elem, uint32x4_t __vec, const int __index)
6050 {
6051   return __aarch64_vset_lane_any (__elem, __vec, __index);
6052 }
6053
6054 __extension__ extern __inline uint64x2_t
6055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6056 vsetq_lane_u64 (uint64_t __elem, uint64x2_t __vec, const int __index)
6057 {
6058   return __aarch64_vset_lane_any (__elem, __vec, __index);
6059 }
6060
6061 #define __GET_LOW(__TYPE) \
6062   uint64x2_t tmp = vreinterpretq_u64_##__TYPE (__a);  \
6063   uint64x1_t lo = vcreate_u64 (vgetq_lane_u64 (tmp, 0));  \
6064   return vreinterpret_##__TYPE##_u64 (lo);
6065
6066 __extension__ extern __inline float16x4_t
6067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6068 vget_low_f16 (float16x8_t __a)
6069 {
6070   __GET_LOW (f16);
6071 }
6072
6073 __extension__ extern __inline float32x2_t
6074 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6075 vget_low_f32 (float32x4_t __a)
6076 {
6077   __GET_LOW (f32);
6078 }
6079
6080 __extension__ extern __inline float64x1_t
6081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6082 vget_low_f64 (float64x2_t __a)
6083 {
6084   return (float64x1_t) {vgetq_lane_f64 (__a, 0)};
6085 }
6086
6087 __extension__ extern __inline poly8x8_t
6088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6089 vget_low_p8 (poly8x16_t __a)
6090 {
6091   __GET_LOW (p8);
6092 }
6093
6094 __extension__ extern __inline poly16x4_t
6095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6096 vget_low_p16 (poly16x8_t __a)
6097 {
6098   __GET_LOW (p16);
6099 }
6100
6101 __extension__ extern __inline poly64x1_t
6102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6103 vget_low_p64 (poly64x2_t __a)
6104 {
6105   __GET_LOW (p64);
6106 }
6107
6108 __extension__ extern __inline int8x8_t
6109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6110 vget_low_s8 (int8x16_t __a)
6111 {
6112   __GET_LOW (s8);
6113 }
6114
6115 __extension__ extern __inline int16x4_t
6116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6117 vget_low_s16 (int16x8_t __a)
6118 {
6119   __GET_LOW (s16);
6120 }
6121
6122 __extension__ extern __inline int32x2_t
6123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6124 vget_low_s32 (int32x4_t __a)
6125 {
6126   __GET_LOW (s32);
6127 }
6128
6129 __extension__ extern __inline int64x1_t
6130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6131 vget_low_s64 (int64x2_t __a)
6132 {
6133   __GET_LOW (s64);
6134 }
6135
6136 __extension__ extern __inline uint8x8_t
6137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6138 vget_low_u8 (uint8x16_t __a)
6139 {
6140   __GET_LOW (u8);
6141 }
6142
6143 __extension__ extern __inline uint16x4_t
6144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6145 vget_low_u16 (uint16x8_t __a)
6146 {
6147   __GET_LOW (u16);
6148 }
6149
6150 __extension__ extern __inline uint32x2_t
6151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6152 vget_low_u32 (uint32x4_t __a)
6153 {
6154   __GET_LOW (u32);
6155 }
6156
6157 __extension__ extern __inline uint64x1_t
6158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6159 vget_low_u64 (uint64x2_t __a)
6160 {
6161   return vcreate_u64 (vgetq_lane_u64 (__a, 0));
6162 }
6163
6164 #undef __GET_LOW
6165
6166 #define __GET_HIGH(__TYPE)                                      \
6167   uint64x2_t tmp = vreinterpretq_u64_##__TYPE (__a);            \
6168   uint64x1_t hi = vcreate_u64 (vgetq_lane_u64 (tmp, 1));        \
6169   return vreinterpret_##__TYPE##_u64 (hi);
6170
6171 __extension__ extern __inline float16x4_t
6172 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6173 vget_high_f16 (float16x8_t __a)
6174 {
6175   __GET_HIGH (f16);
6176 }
6177
6178 __extension__ extern __inline float32x2_t
6179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6180 vget_high_f32 (float32x4_t __a)
6181 {
6182   __GET_HIGH (f32);
6183 }
6184
6185 __extension__ extern __inline float64x1_t
6186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6187 vget_high_f64 (float64x2_t __a)
6188 {
6189   __GET_HIGH (f64);
6190 }
6191
6192 __extension__ extern __inline poly8x8_t
6193 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6194 vget_high_p8 (poly8x16_t __a)
6195 {
6196   __GET_HIGH (p8);
6197 }
6198
6199 __extension__ extern __inline poly16x4_t
6200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6201 vget_high_p16 (poly16x8_t __a)
6202 {
6203   __GET_HIGH (p16);
6204 }
6205
6206 __extension__ extern __inline poly64x1_t
6207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6208 vget_high_p64 (poly64x2_t __a)
6209 {
6210   __GET_HIGH (p64);
6211 }
6212
6213 __extension__ extern __inline int8x8_t
6214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6215 vget_high_s8 (int8x16_t __a)
6216 {
6217   __GET_HIGH (s8);
6218 }
6219
6220 __extension__ extern __inline int16x4_t
6221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6222 vget_high_s16 (int16x8_t __a)
6223 {
6224   __GET_HIGH (s16);
6225 }
6226
6227 __extension__ extern __inline int32x2_t
6228 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6229 vget_high_s32 (int32x4_t __a)
6230 {
6231   __GET_HIGH (s32);
6232 }
6233
6234 __extension__ extern __inline int64x1_t
6235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6236 vget_high_s64 (int64x2_t __a)
6237 {
6238   __GET_HIGH (s64);
6239 }
6240
6241 __extension__ extern __inline uint8x8_t
6242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6243 vget_high_u8 (uint8x16_t __a)
6244 {
6245   __GET_HIGH (u8);
6246 }
6247
6248 __extension__ extern __inline uint16x4_t
6249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6250 vget_high_u16 (uint16x8_t __a)
6251 {
6252   __GET_HIGH (u16);
6253 }
6254
6255 __extension__ extern __inline uint32x2_t
6256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6257 vget_high_u32 (uint32x4_t __a)
6258 {
6259   __GET_HIGH (u32);
6260 }
6261
6262 #undef __GET_HIGH
6263
6264 __extension__ extern __inline uint64x1_t
6265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6266 vget_high_u64 (uint64x2_t __a)
6267 {
6268   return vcreate_u64 (vgetq_lane_u64 (__a, 1));
6269 }
6270
6271 __extension__ extern __inline int8x16_t
6272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6273 vcombine_s8 (int8x8_t __a, int8x8_t __b)
6274 {
6275   return (int8x16_t) __builtin_aarch64_combinev8qi (__a, __b);
6276 }
6277
6278 __extension__ extern __inline int16x8_t
6279 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6280 vcombine_s16 (int16x4_t __a, int16x4_t __b)
6281 {
6282   return (int16x8_t) __builtin_aarch64_combinev4hi (__a, __b);
6283 }
6284
6285 __extension__ extern __inline int32x4_t
6286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6287 vcombine_s32 (int32x2_t __a, int32x2_t __b)
6288 {
6289   return (int32x4_t) __builtin_aarch64_combinev2si (__a, __b);
6290 }
6291
6292 __extension__ extern __inline int64x2_t
6293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6294 vcombine_s64 (int64x1_t __a, int64x1_t __b)
6295 {
6296   return __builtin_aarch64_combinedi (__a[0], __b[0]);
6297 }
6298
6299 __extension__ extern __inline float16x8_t
6300 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6301 vcombine_f16 (float16x4_t __a, float16x4_t __b)
6302 {
6303   return __builtin_aarch64_combinev4hf (__a, __b);
6304 }
6305
6306 __extension__ extern __inline float32x4_t
6307 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6308 vcombine_f32 (float32x2_t __a, float32x2_t __b)
6309 {
6310   return (float32x4_t) __builtin_aarch64_combinev2sf (__a, __b);
6311 }
6312
6313 __extension__ extern __inline uint8x16_t
6314 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6315 vcombine_u8 (uint8x8_t __a, uint8x8_t __b)
6316 {
6317   return (uint8x16_t) __builtin_aarch64_combinev8qi ((int8x8_t) __a,
6318                                                      (int8x8_t) __b);
6319 }
6320
6321 __extension__ extern __inline uint16x8_t
6322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6323 vcombine_u16 (uint16x4_t __a, uint16x4_t __b)
6324 {
6325   return (uint16x8_t) __builtin_aarch64_combinev4hi ((int16x4_t) __a,
6326                                                      (int16x4_t) __b);
6327 }
6328
6329 __extension__ extern __inline uint32x4_t
6330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6331 vcombine_u32 (uint32x2_t __a, uint32x2_t __b)
6332 {
6333   return (uint32x4_t) __builtin_aarch64_combinev2si ((int32x2_t) __a,
6334                                                      (int32x2_t) __b);
6335 }
6336
6337 __extension__ extern __inline uint64x2_t
6338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6339 vcombine_u64 (uint64x1_t __a, uint64x1_t __b)
6340 {
6341   return (uint64x2_t) __builtin_aarch64_combinedi (__a[0], __b[0]);
6342 }
6343
6344 __extension__ extern __inline float64x2_t
6345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6346 vcombine_f64 (float64x1_t __a, float64x1_t __b)
6347 {
6348   return __builtin_aarch64_combinedf (__a[0], __b[0]);
6349 }
6350
6351 __extension__ extern __inline poly8x16_t
6352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6353 vcombine_p8 (poly8x8_t __a, poly8x8_t __b)
6354 {
6355   return (poly8x16_t) __builtin_aarch64_combinev8qi ((int8x8_t) __a,
6356                                                      (int8x8_t) __b);
6357 }
6358
6359 __extension__ extern __inline poly16x8_t
6360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6361 vcombine_p16 (poly16x4_t __a, poly16x4_t __b)
6362 {
6363   return (poly16x8_t) __builtin_aarch64_combinev4hi ((int16x4_t) __a,
6364                                                      (int16x4_t) __b);
6365 }
6366
6367 __extension__ extern __inline poly64x2_t
6368 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6369 vcombine_p64 (poly64x1_t __a, poly64x1_t __b)
6370 {
6371   return (poly64x2_t) __builtin_aarch64_combinedi_ppp (__a[0], __b[0]);
6372 }
6373
6374 /* Start of temporary inline asm implementations.  */
6375
6376 __extension__ extern __inline int8x8_t
6377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6378 vaba_s8 (int8x8_t a, int8x8_t b, int8x8_t c)
6379 {
6380   int8x8_t result;
6381   __asm__ ("saba %0.8b,%2.8b,%3.8b"
6382            : "=w"(result)
6383            : "0"(a), "w"(b), "w"(c)
6384            : /* No clobbers */);
6385   return result;
6386 }
6387
6388 __extension__ extern __inline int16x4_t
6389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6390 vaba_s16 (int16x4_t a, int16x4_t b, int16x4_t c)
6391 {
6392   int16x4_t result;
6393   __asm__ ("saba %0.4h,%2.4h,%3.4h"
6394            : "=w"(result)
6395            : "0"(a), "w"(b), "w"(c)
6396            : /* No clobbers */);
6397   return result;
6398 }
6399
6400 __extension__ extern __inline int32x2_t
6401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6402 vaba_s32 (int32x2_t a, int32x2_t b, int32x2_t c)
6403 {
6404   int32x2_t result;
6405   __asm__ ("saba %0.2s,%2.2s,%3.2s"
6406            : "=w"(result)
6407            : "0"(a), "w"(b), "w"(c)
6408            : /* No clobbers */);
6409   return result;
6410 }
6411
6412 __extension__ extern __inline uint8x8_t
6413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6414 vaba_u8 (uint8x8_t a, uint8x8_t b, uint8x8_t c)
6415 {
6416   uint8x8_t result;
6417   __asm__ ("uaba %0.8b,%2.8b,%3.8b"
6418            : "=w"(result)
6419            : "0"(a), "w"(b), "w"(c)
6420            : /* No clobbers */);
6421   return result;
6422 }
6423
6424 __extension__ extern __inline uint16x4_t
6425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6426 vaba_u16 (uint16x4_t a, uint16x4_t b, uint16x4_t c)
6427 {
6428   uint16x4_t result;
6429   __asm__ ("uaba %0.4h,%2.4h,%3.4h"
6430            : "=w"(result)
6431            : "0"(a), "w"(b), "w"(c)
6432            : /* No clobbers */);
6433   return result;
6434 }
6435
6436 __extension__ extern __inline uint32x2_t
6437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6438 vaba_u32 (uint32x2_t a, uint32x2_t b, uint32x2_t c)
6439 {
6440   uint32x2_t result;
6441   __asm__ ("uaba %0.2s,%2.2s,%3.2s"
6442            : "=w"(result)
6443            : "0"(a), "w"(b), "w"(c)
6444            : /* No clobbers */);
6445   return result;
6446 }
6447
6448 __extension__ extern __inline int16x8_t
6449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6450 vabal_high_s8 (int16x8_t a, int8x16_t b, int8x16_t c)
6451 {
6452   int16x8_t result;
6453   __asm__ ("sabal2 %0.8h,%2.16b,%3.16b"
6454            : "=w"(result)
6455            : "0"(a), "w"(b), "w"(c)
6456            : /* No clobbers */);
6457   return result;
6458 }
6459
6460 __extension__ extern __inline int32x4_t
6461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6462 vabal_high_s16 (int32x4_t a, int16x8_t b, int16x8_t c)
6463 {
6464   int32x4_t result;
6465   __asm__ ("sabal2 %0.4s,%2.8h,%3.8h"
6466            : "=w"(result)
6467            : "0"(a), "w"(b), "w"(c)
6468            : /* No clobbers */);
6469   return result;
6470 }
6471
6472 __extension__ extern __inline int64x2_t
6473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6474 vabal_high_s32 (int64x2_t a, int32x4_t b, int32x4_t c)
6475 {
6476   int64x2_t result;
6477   __asm__ ("sabal2 %0.2d,%2.4s,%3.4s"
6478            : "=w"(result)
6479            : "0"(a), "w"(b), "w"(c)
6480            : /* No clobbers */);
6481   return result;
6482 }
6483
6484 __extension__ extern __inline uint16x8_t
6485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6486 vabal_high_u8 (uint16x8_t a, uint8x16_t b, uint8x16_t c)
6487 {
6488   uint16x8_t result;
6489   __asm__ ("uabal2 %0.8h,%2.16b,%3.16b"
6490            : "=w"(result)
6491            : "0"(a), "w"(b), "w"(c)
6492            : /* No clobbers */);
6493   return result;
6494 }
6495
6496 __extension__ extern __inline uint32x4_t
6497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6498 vabal_high_u16 (uint32x4_t a, uint16x8_t b, uint16x8_t c)
6499 {
6500   uint32x4_t result;
6501   __asm__ ("uabal2 %0.4s,%2.8h,%3.8h"
6502            : "=w"(result)
6503            : "0"(a), "w"(b), "w"(c)
6504            : /* No clobbers */);
6505   return result;
6506 }
6507
6508 __extension__ extern __inline uint64x2_t
6509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6510 vabal_high_u32 (uint64x2_t a, uint32x4_t b, uint32x4_t c)
6511 {
6512   uint64x2_t result;
6513   __asm__ ("uabal2 %0.2d,%2.4s,%3.4s"
6514            : "=w"(result)
6515            : "0"(a), "w"(b), "w"(c)
6516            : /* No clobbers */);
6517   return result;
6518 }
6519
6520 __extension__ extern __inline int16x8_t
6521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6522 vabal_s8 (int16x8_t a, int8x8_t b, int8x8_t c)
6523 {
6524   int16x8_t result;
6525   __asm__ ("sabal %0.8h,%2.8b,%3.8b"
6526            : "=w"(result)
6527            : "0"(a), "w"(b), "w"(c)
6528            : /* No clobbers */);
6529   return result;
6530 }
6531
6532 __extension__ extern __inline int32x4_t
6533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6534 vabal_s16 (int32x4_t a, int16x4_t b, int16x4_t c)
6535 {
6536   int32x4_t result;
6537   __asm__ ("sabal %0.4s,%2.4h,%3.4h"
6538            : "=w"(result)
6539            : "0"(a), "w"(b), "w"(c)
6540            : /* No clobbers */);
6541   return result;
6542 }
6543
6544 __extension__ extern __inline int64x2_t
6545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6546 vabal_s32 (int64x2_t a, int32x2_t b, int32x2_t c)
6547 {
6548   int64x2_t result;
6549   __asm__ ("sabal %0.2d,%2.2s,%3.2s"
6550            : "=w"(result)
6551            : "0"(a), "w"(b), "w"(c)
6552            : /* No clobbers */);
6553   return result;
6554 }
6555
6556 __extension__ extern __inline uint16x8_t
6557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6558 vabal_u8 (uint16x8_t a, uint8x8_t b, uint8x8_t c)
6559 {
6560   uint16x8_t result;
6561   __asm__ ("uabal %0.8h,%2.8b,%3.8b"
6562            : "=w"(result)
6563            : "0"(a), "w"(b), "w"(c)
6564            : /* No clobbers */);
6565   return result;
6566 }
6567
6568 __extension__ extern __inline uint32x4_t
6569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6570 vabal_u16 (uint32x4_t a, uint16x4_t b, uint16x4_t c)
6571 {
6572   uint32x4_t result;
6573   __asm__ ("uabal %0.4s,%2.4h,%3.4h"
6574            : "=w"(result)
6575            : "0"(a), "w"(b), "w"(c)
6576            : /* No clobbers */);
6577   return result;
6578 }
6579
6580 __extension__ extern __inline uint64x2_t
6581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6582 vabal_u32 (uint64x2_t a, uint32x2_t b, uint32x2_t c)
6583 {
6584   uint64x2_t result;
6585   __asm__ ("uabal %0.2d,%2.2s,%3.2s"
6586            : "=w"(result)
6587            : "0"(a), "w"(b), "w"(c)
6588            : /* No clobbers */);
6589   return result;
6590 }
6591
6592 __extension__ extern __inline int8x16_t
6593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6594 vabaq_s8 (int8x16_t a, int8x16_t b, int8x16_t c)
6595 {
6596   int8x16_t result;
6597   __asm__ ("saba %0.16b,%2.16b,%3.16b"
6598            : "=w"(result)
6599            : "0"(a), "w"(b), "w"(c)
6600            : /* No clobbers */);
6601   return result;
6602 }
6603
6604 __extension__ extern __inline int16x8_t
6605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6606 vabaq_s16 (int16x8_t a, int16x8_t b, int16x8_t c)
6607 {
6608   int16x8_t result;
6609   __asm__ ("saba %0.8h,%2.8h,%3.8h"
6610            : "=w"(result)
6611            : "0"(a), "w"(b), "w"(c)
6612            : /* No clobbers */);
6613   return result;
6614 }
6615
6616 __extension__ extern __inline int32x4_t
6617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6618 vabaq_s32 (int32x4_t a, int32x4_t b, int32x4_t c)
6619 {
6620   int32x4_t result;
6621   __asm__ ("saba %0.4s,%2.4s,%3.4s"
6622            : "=w"(result)
6623            : "0"(a), "w"(b), "w"(c)
6624            : /* No clobbers */);
6625   return result;
6626 }
6627
6628 __extension__ extern __inline uint8x16_t
6629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6630 vabaq_u8 (uint8x16_t a, uint8x16_t b, uint8x16_t c)
6631 {
6632   uint8x16_t result;
6633   __asm__ ("uaba %0.16b,%2.16b,%3.16b"
6634            : "=w"(result)
6635            : "0"(a), "w"(b), "w"(c)
6636            : /* No clobbers */);
6637   return result;
6638 }
6639
6640 __extension__ extern __inline uint16x8_t
6641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6642 vabaq_u16 (uint16x8_t a, uint16x8_t b, uint16x8_t c)
6643 {
6644   uint16x8_t result;
6645   __asm__ ("uaba %0.8h,%2.8h,%3.8h"
6646            : "=w"(result)
6647            : "0"(a), "w"(b), "w"(c)
6648            : /* No clobbers */);
6649   return result;
6650 }
6651
6652 __extension__ extern __inline uint32x4_t
6653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6654 vabaq_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t c)
6655 {
6656   uint32x4_t result;
6657   __asm__ ("uaba %0.4s,%2.4s,%3.4s"
6658            : "=w"(result)
6659            : "0"(a), "w"(b), "w"(c)
6660            : /* No clobbers */);
6661   return result;
6662 }
6663
6664 __extension__ extern __inline int8x8_t
6665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6666 vabd_s8 (int8x8_t a, int8x8_t b)
6667 {
6668   int8x8_t result;
6669   __asm__ ("sabd %0.8b, %1.8b, %2.8b"
6670            : "=w"(result)
6671            : "w"(a), "w"(b)
6672            : /* No clobbers */);
6673   return result;
6674 }
6675
6676 __extension__ extern __inline int16x4_t
6677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6678 vabd_s16 (int16x4_t a, int16x4_t b)
6679 {
6680   int16x4_t result;
6681   __asm__ ("sabd %0.4h, %1.4h, %2.4h"
6682            : "=w"(result)
6683            : "w"(a), "w"(b)
6684            : /* No clobbers */);
6685   return result;
6686 }
6687
6688 __extension__ extern __inline int32x2_t
6689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6690 vabd_s32 (int32x2_t a, int32x2_t b)
6691 {
6692   int32x2_t result;
6693   __asm__ ("sabd %0.2s, %1.2s, %2.2s"
6694            : "=w"(result)
6695            : "w"(a), "w"(b)
6696            : /* No clobbers */);
6697   return result;
6698 }
6699
6700 __extension__ extern __inline uint8x8_t
6701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6702 vabd_u8 (uint8x8_t a, uint8x8_t b)
6703 {
6704   uint8x8_t result;
6705   __asm__ ("uabd %0.8b, %1.8b, %2.8b"
6706            : "=w"(result)
6707            : "w"(a), "w"(b)
6708            : /* No clobbers */);
6709   return result;
6710 }
6711
6712 __extension__ extern __inline uint16x4_t
6713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6714 vabd_u16 (uint16x4_t a, uint16x4_t b)
6715 {
6716   uint16x4_t result;
6717   __asm__ ("uabd %0.4h, %1.4h, %2.4h"
6718            : "=w"(result)
6719            : "w"(a), "w"(b)
6720            : /* No clobbers */);
6721   return result;
6722 }
6723
6724 __extension__ extern __inline uint32x2_t
6725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6726 vabd_u32 (uint32x2_t a, uint32x2_t b)
6727 {
6728   uint32x2_t result;
6729   __asm__ ("uabd %0.2s, %1.2s, %2.2s"
6730            : "=w"(result)
6731            : "w"(a), "w"(b)
6732            : /* No clobbers */);
6733   return result;
6734 }
6735
6736 __extension__ extern __inline int16x8_t
6737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6738 vabdl_high_s8 (int8x16_t a, int8x16_t b)
6739 {
6740   int16x8_t result;
6741   __asm__ ("sabdl2 %0.8h,%1.16b,%2.16b"
6742            : "=w"(result)
6743            : "w"(a), "w"(b)
6744            : /* No clobbers */);
6745   return result;
6746 }
6747
6748 __extension__ extern __inline int32x4_t
6749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6750 vabdl_high_s16 (int16x8_t a, int16x8_t b)
6751 {
6752   int32x4_t result;
6753   __asm__ ("sabdl2 %0.4s,%1.8h,%2.8h"
6754            : "=w"(result)
6755            : "w"(a), "w"(b)
6756            : /* No clobbers */);
6757   return result;
6758 }
6759
6760 __extension__ extern __inline int64x2_t
6761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6762 vabdl_high_s32 (int32x4_t a, int32x4_t b)
6763 {
6764   int64x2_t result;
6765   __asm__ ("sabdl2 %0.2d,%1.4s,%2.4s"
6766            : "=w"(result)
6767            : "w"(a), "w"(b)
6768            : /* No clobbers */);
6769   return result;
6770 }
6771
6772 __extension__ extern __inline uint16x8_t
6773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6774 vabdl_high_u8 (uint8x16_t a, uint8x16_t b)
6775 {
6776   uint16x8_t result;
6777   __asm__ ("uabdl2 %0.8h,%1.16b,%2.16b"
6778            : "=w"(result)
6779            : "w"(a), "w"(b)
6780            : /* No clobbers */);
6781   return result;
6782 }
6783
6784 __extension__ extern __inline uint32x4_t
6785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6786 vabdl_high_u16 (uint16x8_t a, uint16x8_t b)
6787 {
6788   uint32x4_t result;
6789   __asm__ ("uabdl2 %0.4s,%1.8h,%2.8h"
6790            : "=w"(result)
6791            : "w"(a), "w"(b)
6792            : /* No clobbers */);
6793   return result;
6794 }
6795
6796 __extension__ extern __inline uint64x2_t
6797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6798 vabdl_high_u32 (uint32x4_t a, uint32x4_t b)
6799 {
6800   uint64x2_t result;
6801   __asm__ ("uabdl2 %0.2d,%1.4s,%2.4s"
6802            : "=w"(result)
6803            : "w"(a), "w"(b)
6804            : /* No clobbers */);
6805   return result;
6806 }
6807
6808 __extension__ extern __inline int16x8_t
6809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6810 vabdl_s8 (int8x8_t a, int8x8_t b)
6811 {
6812   int16x8_t result;
6813   __asm__ ("sabdl %0.8h, %1.8b, %2.8b"
6814            : "=w"(result)
6815            : "w"(a), "w"(b)
6816            : /* No clobbers */);
6817   return result;
6818 }
6819
6820 __extension__ extern __inline int32x4_t
6821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6822 vabdl_s16 (int16x4_t a, int16x4_t b)
6823 {
6824   int32x4_t result;
6825   __asm__ ("sabdl %0.4s, %1.4h, %2.4h"
6826            : "=w"(result)
6827            : "w"(a), "w"(b)
6828            : /* No clobbers */);
6829   return result;
6830 }
6831
6832 __extension__ extern __inline int64x2_t
6833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6834 vabdl_s32 (int32x2_t a, int32x2_t b)
6835 {
6836   int64x2_t result;
6837   __asm__ ("sabdl %0.2d, %1.2s, %2.2s"
6838            : "=w"(result)
6839            : "w"(a), "w"(b)
6840            : /* No clobbers */);
6841   return result;
6842 }
6843
6844 __extension__ extern __inline uint16x8_t
6845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6846 vabdl_u8 (uint8x8_t a, uint8x8_t b)
6847 {
6848   uint16x8_t result;
6849   __asm__ ("uabdl %0.8h, %1.8b, %2.8b"
6850            : "=w"(result)
6851            : "w"(a), "w"(b)
6852            : /* No clobbers */);
6853   return result;
6854 }
6855
6856 __extension__ extern __inline uint32x4_t
6857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6858 vabdl_u16 (uint16x4_t a, uint16x4_t b)
6859 {
6860   uint32x4_t result;
6861   __asm__ ("uabdl %0.4s, %1.4h, %2.4h"
6862            : "=w"(result)
6863            : "w"(a), "w"(b)
6864            : /* No clobbers */);
6865   return result;
6866 }
6867
6868 __extension__ extern __inline uint64x2_t
6869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6870 vabdl_u32 (uint32x2_t a, uint32x2_t b)
6871 {
6872   uint64x2_t result;
6873   __asm__ ("uabdl %0.2d, %1.2s, %2.2s"
6874            : "=w"(result)
6875            : "w"(a), "w"(b)
6876            : /* No clobbers */);
6877   return result;
6878 }
6879
6880 __extension__ extern __inline int8x16_t
6881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6882 vabdq_s8 (int8x16_t a, int8x16_t b)
6883 {
6884   int8x16_t result;
6885   __asm__ ("sabd %0.16b, %1.16b, %2.16b"
6886            : "=w"(result)
6887            : "w"(a), "w"(b)
6888            : /* No clobbers */);
6889   return result;
6890 }
6891
6892 __extension__ extern __inline int16x8_t
6893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6894 vabdq_s16 (int16x8_t a, int16x8_t b)
6895 {
6896   int16x8_t result;
6897   __asm__ ("sabd %0.8h, %1.8h, %2.8h"
6898            : "=w"(result)
6899            : "w"(a), "w"(b)
6900            : /* No clobbers */);
6901   return result;
6902 }
6903
6904 __extension__ extern __inline int32x4_t
6905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6906 vabdq_s32 (int32x4_t a, int32x4_t b)
6907 {
6908   int32x4_t result;
6909   __asm__ ("sabd %0.4s, %1.4s, %2.4s"
6910            : "=w"(result)
6911            : "w"(a), "w"(b)
6912            : /* No clobbers */);
6913   return result;
6914 }
6915
6916 __extension__ extern __inline uint8x16_t
6917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6918 vabdq_u8 (uint8x16_t a, uint8x16_t b)
6919 {
6920   uint8x16_t result;
6921   __asm__ ("uabd %0.16b, %1.16b, %2.16b"
6922            : "=w"(result)
6923            : "w"(a), "w"(b)
6924            : /* No clobbers */);
6925   return result;
6926 }
6927
6928 __extension__ extern __inline uint16x8_t
6929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6930 vabdq_u16 (uint16x8_t a, uint16x8_t b)
6931 {
6932   uint16x8_t result;
6933   __asm__ ("uabd %0.8h, %1.8h, %2.8h"
6934            : "=w"(result)
6935            : "w"(a), "w"(b)
6936            : /* No clobbers */);
6937   return result;
6938 }
6939
6940 __extension__ extern __inline uint32x4_t
6941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6942 vabdq_u32 (uint32x4_t a, uint32x4_t b)
6943 {
6944   uint32x4_t result;
6945   __asm__ ("uabd %0.4s, %1.4s, %2.4s"
6946            : "=w"(result)
6947            : "w"(a), "w"(b)
6948            : /* No clobbers */);
6949   return result;
6950 }
6951
6952 __extension__ extern __inline int16_t
6953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6954 vaddlv_s8 (int8x8_t a)
6955 {
6956   int16_t result;
6957   __asm__ ("saddlv %h0,%1.8b"
6958            : "=w"(result)
6959            : "w"(a)
6960            : /* No clobbers */);
6961   return result;
6962 }
6963
6964 __extension__ extern __inline int32_t
6965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6966 vaddlv_s16 (int16x4_t a)
6967 {
6968   int32_t result;
6969   __asm__ ("saddlv %s0,%1.4h"
6970            : "=w"(result)
6971            : "w"(a)
6972            : /* No clobbers */);
6973   return result;
6974 }
6975
6976 __extension__ extern __inline uint16_t
6977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6978 vaddlv_u8 (uint8x8_t a)
6979 {
6980   uint16_t result;
6981   __asm__ ("uaddlv %h0,%1.8b"
6982            : "=w"(result)
6983            : "w"(a)
6984            : /* No clobbers */);
6985   return result;
6986 }
6987
6988 __extension__ extern __inline uint32_t
6989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
6990 vaddlv_u16 (uint16x4_t a)
6991 {
6992   uint32_t result;
6993   __asm__ ("uaddlv %s0,%1.4h"
6994            : "=w"(result)
6995            : "w"(a)
6996            : /* No clobbers */);
6997   return result;
6998 }
6999
7000 __extension__ extern __inline int16_t
7001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7002 vaddlvq_s8 (int8x16_t a)
7003 {
7004   int16_t result;
7005   __asm__ ("saddlv %h0,%1.16b"
7006            : "=w"(result)
7007            : "w"(a)
7008            : /* No clobbers */);
7009   return result;
7010 }
7011
7012 __extension__ extern __inline int32_t
7013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7014 vaddlvq_s16 (int16x8_t a)
7015 {
7016   int32_t result;
7017   __asm__ ("saddlv %s0,%1.8h"
7018            : "=w"(result)
7019            : "w"(a)
7020            : /* No clobbers */);
7021   return result;
7022 }
7023
7024 __extension__ extern __inline int64_t
7025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7026 vaddlvq_s32 (int32x4_t a)
7027 {
7028   int64_t result;
7029   __asm__ ("saddlv %d0,%1.4s"
7030            : "=w"(result)
7031            : "w"(a)
7032            : /* No clobbers */);
7033   return result;
7034 }
7035
7036 __extension__ extern __inline uint16_t
7037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7038 vaddlvq_u8 (uint8x16_t a)
7039 {
7040   uint16_t result;
7041   __asm__ ("uaddlv %h0,%1.16b"
7042            : "=w"(result)
7043            : "w"(a)
7044            : /* No clobbers */);
7045   return result;
7046 }
7047
7048 __extension__ extern __inline uint32_t
7049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7050 vaddlvq_u16 (uint16x8_t a)
7051 {
7052   uint32_t result;
7053   __asm__ ("uaddlv %s0,%1.8h"
7054            : "=w"(result)
7055            : "w"(a)
7056            : /* No clobbers */);
7057   return result;
7058 }
7059
7060 __extension__ extern __inline uint64_t
7061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7062 vaddlvq_u32 (uint32x4_t a)
7063 {
7064   uint64_t result;
7065   __asm__ ("uaddlv %d0,%1.4s"
7066            : "=w"(result)
7067            : "w"(a)
7068            : /* No clobbers */);
7069   return result;
7070 }
7071
7072 __extension__ extern __inline float32x2_t
7073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7074 vcvtx_f32_f64 (float64x2_t a)
7075 {
7076   float32x2_t result;
7077   __asm__ ("fcvtxn %0.2s,%1.2d"
7078            : "=w"(result)
7079            : "w"(a)
7080            : /* No clobbers */);
7081   return result;
7082 }
7083
7084 __extension__ extern __inline float32x4_t
7085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7086 vcvtx_high_f32_f64 (float32x2_t a, float64x2_t b)
7087 {
7088   float32x4_t result;
7089   __asm__ ("fcvtxn2 %0.4s,%1.2d"
7090            : "=w"(result)
7091            : "w" (b), "0"(a)
7092            : /* No clobbers */);
7093   return result;
7094 }
7095
7096 __extension__ extern __inline float32_t
7097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7098 vcvtxd_f32_f64 (float64_t a)
7099 {
7100   float32_t result;
7101   __asm__ ("fcvtxn %s0,%d1"
7102            : "=w"(result)
7103            : "w"(a)
7104            : /* No clobbers */);
7105   return result;
7106 }
7107
7108 __extension__ extern __inline float32x2_t
7109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7110 vmla_n_f32 (float32x2_t a, float32x2_t b, float32_t c)
7111 {
7112   float32x2_t result;
7113   float32x2_t t1;
7114   __asm__ ("fmul %1.2s, %3.2s, %4.s[0]; fadd %0.2s, %0.2s, %1.2s"
7115            : "=w"(result), "=w"(t1)
7116            : "0"(a), "w"(b), "w"(c)
7117            : /* No clobbers */);
7118   return result;
7119 }
7120
7121 __extension__ extern __inline int16x4_t
7122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7123 vmla_n_s16 (int16x4_t a, int16x4_t b, int16_t c)
7124 {
7125   int16x4_t result;
7126   __asm__ ("mla %0.4h,%2.4h,%3.h[0]"
7127            : "=w"(result)
7128            : "0"(a), "w"(b), "x"(c)
7129            : /* No clobbers */);
7130   return result;
7131 }
7132
7133 __extension__ extern __inline int32x2_t
7134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7135 vmla_n_s32 (int32x2_t a, int32x2_t b, int32_t c)
7136 {
7137   int32x2_t result;
7138   __asm__ ("mla %0.2s,%2.2s,%3.s[0]"
7139            : "=w"(result)
7140            : "0"(a), "w"(b), "w"(c)
7141            : /* No clobbers */);
7142   return result;
7143 }
7144
7145 __extension__ extern __inline uint16x4_t
7146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7147 vmla_n_u16 (uint16x4_t a, uint16x4_t b, uint16_t c)
7148 {
7149   uint16x4_t result;
7150   __asm__ ("mla %0.4h,%2.4h,%3.h[0]"
7151            : "=w"(result)
7152            : "0"(a), "w"(b), "x"(c)
7153            : /* No clobbers */);
7154   return result;
7155 }
7156
7157 __extension__ extern __inline uint32x2_t
7158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7159 vmla_n_u32 (uint32x2_t a, uint32x2_t b, uint32_t c)
7160 {
7161   uint32x2_t result;
7162   __asm__ ("mla %0.2s,%2.2s,%3.s[0]"
7163            : "=w"(result)
7164            : "0"(a), "w"(b), "w"(c)
7165            : /* No clobbers */);
7166   return result;
7167 }
7168
7169 __extension__ extern __inline int8x8_t
7170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7171 vmla_s8 (int8x8_t a, int8x8_t b, int8x8_t c)
7172 {
7173   int8x8_t result;
7174   __asm__ ("mla %0.8b, %2.8b, %3.8b"
7175            : "=w"(result)
7176            : "0"(a), "w"(b), "w"(c)
7177            : /* No clobbers */);
7178   return result;
7179 }
7180
7181 __extension__ extern __inline int16x4_t
7182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7183 vmla_s16 (int16x4_t a, int16x4_t b, int16x4_t c)
7184 {
7185   int16x4_t result;
7186   __asm__ ("mla %0.4h, %2.4h, %3.4h"
7187            : "=w"(result)
7188            : "0"(a), "w"(b), "w"(c)
7189            : /* No clobbers */);
7190   return result;
7191 }
7192
7193 __extension__ extern __inline int32x2_t
7194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7195 vmla_s32 (int32x2_t a, int32x2_t b, int32x2_t c)
7196 {
7197   int32x2_t result;
7198   __asm__ ("mla %0.2s, %2.2s, %3.2s"
7199            : "=w"(result)
7200            : "0"(a), "w"(b), "w"(c)
7201            : /* No clobbers */);
7202   return result;
7203 }
7204
7205 __extension__ extern __inline uint8x8_t
7206 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7207 vmla_u8 (uint8x8_t a, uint8x8_t b, uint8x8_t c)
7208 {
7209   uint8x8_t result;
7210   __asm__ ("mla %0.8b, %2.8b, %3.8b"
7211            : "=w"(result)
7212            : "0"(a), "w"(b), "w"(c)
7213            : /* No clobbers */);
7214   return result;
7215 }
7216
7217 __extension__ extern __inline uint16x4_t
7218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7219 vmla_u16 (uint16x4_t a, uint16x4_t b, uint16x4_t c)
7220 {
7221   uint16x4_t result;
7222   __asm__ ("mla %0.4h, %2.4h, %3.4h"
7223            : "=w"(result)
7224            : "0"(a), "w"(b), "w"(c)
7225            : /* No clobbers */);
7226   return result;
7227 }
7228
7229 __extension__ extern __inline uint32x2_t
7230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7231 vmla_u32 (uint32x2_t a, uint32x2_t b, uint32x2_t c)
7232 {
7233   uint32x2_t result;
7234   __asm__ ("mla %0.2s, %2.2s, %3.2s"
7235            : "=w"(result)
7236            : "0"(a), "w"(b), "w"(c)
7237            : /* No clobbers */);
7238   return result;
7239 }
7240
7241 #define vmlal_high_lane_s16(a, b, c, d)                                 \
7242   __extension__                                                         \
7243     ({                                                                  \
7244        int16x4_t c_ = (c);                                              \
7245        int16x8_t b_ = (b);                                              \
7246        int32x4_t a_ = (a);                                              \
7247        int32x4_t result;                                                \
7248        __asm__ ("smlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7249                 : "=w"(result)                                          \
7250                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7251                 : /* No clobbers */);                                   \
7252        result;                                                          \
7253      })
7254
7255 #define vmlal_high_lane_s32(a, b, c, d)                                 \
7256   __extension__                                                         \
7257     ({                                                                  \
7258        int32x2_t c_ = (c);                                              \
7259        int32x4_t b_ = (b);                                              \
7260        int64x2_t a_ = (a);                                              \
7261        int64x2_t result;                                                \
7262        __asm__ ("smlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7263                 : "=w"(result)                                          \
7264                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7265                 : /* No clobbers */);                                   \
7266        result;                                                          \
7267      })
7268
7269 #define vmlal_high_lane_u16(a, b, c, d)                                 \
7270   __extension__                                                         \
7271     ({                                                                  \
7272        uint16x4_t c_ = (c);                                             \
7273        uint16x8_t b_ = (b);                                             \
7274        uint32x4_t a_ = (a);                                             \
7275        uint32x4_t result;                                               \
7276        __asm__ ("umlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7277                 : "=w"(result)                                          \
7278                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7279                 : /* No clobbers */);                                   \
7280        result;                                                          \
7281      })
7282
7283 #define vmlal_high_lane_u32(a, b, c, d)                                 \
7284   __extension__                                                         \
7285     ({                                                                  \
7286        uint32x2_t c_ = (c);                                             \
7287        uint32x4_t b_ = (b);                                             \
7288        uint64x2_t a_ = (a);                                             \
7289        uint64x2_t result;                                               \
7290        __asm__ ("umlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7291                 : "=w"(result)                                          \
7292                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7293                 : /* No clobbers */);                                   \
7294        result;                                                          \
7295      })
7296
7297 #define vmlal_high_laneq_s16(a, b, c, d)                                \
7298   __extension__                                                         \
7299     ({                                                                  \
7300        int16x8_t c_ = (c);                                              \
7301        int16x8_t b_ = (b);                                              \
7302        int32x4_t a_ = (a);                                              \
7303        int32x4_t result;                                                \
7304        __asm__ ("smlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7305                 : "=w"(result)                                          \
7306                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7307                 : /* No clobbers */);                                   \
7308        result;                                                          \
7309      })
7310
7311 #define vmlal_high_laneq_s32(a, b, c, d)                                \
7312   __extension__                                                         \
7313     ({                                                                  \
7314        int32x4_t c_ = (c);                                              \
7315        int32x4_t b_ = (b);                                              \
7316        int64x2_t a_ = (a);                                              \
7317        int64x2_t result;                                                \
7318        __asm__ ("smlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7319                 : "=w"(result)                                          \
7320                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7321                 : /* No clobbers */);                                   \
7322        result;                                                          \
7323      })
7324
7325 #define vmlal_high_laneq_u16(a, b, c, d)                                \
7326   __extension__                                                         \
7327     ({                                                                  \
7328        uint16x8_t c_ = (c);                                             \
7329        uint16x8_t b_ = (b);                                             \
7330        uint32x4_t a_ = (a);                                             \
7331        uint32x4_t result;                                               \
7332        __asm__ ("umlal2 %0.4s, %2.8h, %3.h[%4]"                         \
7333                 : "=w"(result)                                          \
7334                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7335                 : /* No clobbers */);                                   \
7336        result;                                                          \
7337      })
7338
7339 #define vmlal_high_laneq_u32(a, b, c, d)                                \
7340   __extension__                                                         \
7341     ({                                                                  \
7342        uint32x4_t c_ = (c);                                             \
7343        uint32x4_t b_ = (b);                                             \
7344        uint64x2_t a_ = (a);                                             \
7345        uint64x2_t result;                                               \
7346        __asm__ ("umlal2 %0.2d, %2.4s, %3.s[%4]"                         \
7347                 : "=w"(result)                                          \
7348                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7349                 : /* No clobbers */);                                   \
7350        result;                                                          \
7351      })
7352
7353 __extension__ extern __inline int32x4_t
7354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7355 vmlal_high_n_s16 (int32x4_t a, int16x8_t b, int16_t c)
7356 {
7357   int32x4_t result;
7358   __asm__ ("smlal2 %0.4s,%2.8h,%3.h[0]"
7359            : "=w"(result)
7360            : "0"(a), "w"(b), "x"(c)
7361            : /* No clobbers */);
7362   return result;
7363 }
7364
7365 __extension__ extern __inline int64x2_t
7366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7367 vmlal_high_n_s32 (int64x2_t a, int32x4_t b, int32_t c)
7368 {
7369   int64x2_t result;
7370   __asm__ ("smlal2 %0.2d,%2.4s,%3.s[0]"
7371            : "=w"(result)
7372            : "0"(a), "w"(b), "w"(c)
7373            : /* No clobbers */);
7374   return result;
7375 }
7376
7377 __extension__ extern __inline uint32x4_t
7378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7379 vmlal_high_n_u16 (uint32x4_t a, uint16x8_t b, uint16_t c)
7380 {
7381   uint32x4_t result;
7382   __asm__ ("umlal2 %0.4s,%2.8h,%3.h[0]"
7383            : "=w"(result)
7384            : "0"(a), "w"(b), "x"(c)
7385            : /* No clobbers */);
7386   return result;
7387 }
7388
7389 __extension__ extern __inline uint64x2_t
7390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7391 vmlal_high_n_u32 (uint64x2_t a, uint32x4_t b, uint32_t c)
7392 {
7393   uint64x2_t result;
7394   __asm__ ("umlal2 %0.2d,%2.4s,%3.s[0]"
7395            : "=w"(result)
7396            : "0"(a), "w"(b), "w"(c)
7397            : /* No clobbers */);
7398   return result;
7399 }
7400
7401 __extension__ extern __inline int16x8_t
7402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7403 vmlal_high_s8 (int16x8_t a, int8x16_t b, int8x16_t c)
7404 {
7405   int16x8_t result;
7406   __asm__ ("smlal2 %0.8h,%2.16b,%3.16b"
7407            : "=w"(result)
7408            : "0"(a), "w"(b), "w"(c)
7409            : /* No clobbers */);
7410   return result;
7411 }
7412
7413 __extension__ extern __inline int32x4_t
7414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7415 vmlal_high_s16 (int32x4_t a, int16x8_t b, int16x8_t c)
7416 {
7417   int32x4_t result;
7418   __asm__ ("smlal2 %0.4s,%2.8h,%3.8h"
7419            : "=w"(result)
7420            : "0"(a), "w"(b), "w"(c)
7421            : /* No clobbers */);
7422   return result;
7423 }
7424
7425 __extension__ extern __inline int64x2_t
7426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7427 vmlal_high_s32 (int64x2_t a, int32x4_t b, int32x4_t c)
7428 {
7429   int64x2_t result;
7430   __asm__ ("smlal2 %0.2d,%2.4s,%3.4s"
7431            : "=w"(result)
7432            : "0"(a), "w"(b), "w"(c)
7433            : /* No clobbers */);
7434   return result;
7435 }
7436
7437 __extension__ extern __inline uint16x8_t
7438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7439 vmlal_high_u8 (uint16x8_t a, uint8x16_t b, uint8x16_t c)
7440 {
7441   uint16x8_t result;
7442   __asm__ ("umlal2 %0.8h,%2.16b,%3.16b"
7443            : "=w"(result)
7444            : "0"(a), "w"(b), "w"(c)
7445            : /* No clobbers */);
7446   return result;
7447 }
7448
7449 __extension__ extern __inline uint32x4_t
7450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7451 vmlal_high_u16 (uint32x4_t a, uint16x8_t b, uint16x8_t c)
7452 {
7453   uint32x4_t result;
7454   __asm__ ("umlal2 %0.4s,%2.8h,%3.8h"
7455            : "=w"(result)
7456            : "0"(a), "w"(b), "w"(c)
7457            : /* No clobbers */);
7458   return result;
7459 }
7460
7461 __extension__ extern __inline uint64x2_t
7462 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7463 vmlal_high_u32 (uint64x2_t a, uint32x4_t b, uint32x4_t c)
7464 {
7465   uint64x2_t result;
7466   __asm__ ("umlal2 %0.2d,%2.4s,%3.4s"
7467            : "=w"(result)
7468            : "0"(a), "w"(b), "w"(c)
7469            : /* No clobbers */);
7470   return result;
7471 }
7472
7473 #define vmlal_lane_s16(a, b, c, d)                                      \
7474   __extension__                                                         \
7475     ({                                                                  \
7476        int16x4_t c_ = (c);                                              \
7477        int16x4_t b_ = (b);                                              \
7478        int32x4_t a_ = (a);                                              \
7479        int32x4_t result;                                                \
7480        __asm__ ("smlal %0.4s,%2.4h,%3.h[%4]"                            \
7481                 : "=w"(result)                                          \
7482                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7483                 : /* No clobbers */);                                   \
7484        result;                                                          \
7485      })
7486
7487 #define vmlal_lane_s32(a, b, c, d)                                      \
7488   __extension__                                                         \
7489     ({                                                                  \
7490        int32x2_t c_ = (c);                                              \
7491        int32x2_t b_ = (b);                                              \
7492        int64x2_t a_ = (a);                                              \
7493        int64x2_t result;                                                \
7494        __asm__ ("smlal %0.2d,%2.2s,%3.s[%4]"                            \
7495                 : "=w"(result)                                          \
7496                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7497                 : /* No clobbers */);                                   \
7498        result;                                                          \
7499      })
7500
7501 #define vmlal_lane_u16(a, b, c, d)                                      \
7502   __extension__                                                         \
7503     ({                                                                  \
7504        uint16x4_t c_ = (c);                                             \
7505        uint16x4_t b_ = (b);                                             \
7506        uint32x4_t a_ = (a);                                             \
7507        uint32x4_t result;                                               \
7508        __asm__ ("umlal %0.4s,%2.4h,%3.h[%4]"                            \
7509                 : "=w"(result)                                          \
7510                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7511                 : /* No clobbers */);                                   \
7512        result;                                                          \
7513      })
7514
7515 #define vmlal_lane_u32(a, b, c, d)                                      \
7516   __extension__                                                         \
7517     ({                                                                  \
7518        uint32x2_t c_ = (c);                                             \
7519        uint32x2_t b_ = (b);                                             \
7520        uint64x2_t a_ = (a);                                             \
7521        uint64x2_t result;                                               \
7522        __asm__ ("umlal %0.2d, %2.2s, %3.s[%4]"                          \
7523                 : "=w"(result)                                          \
7524                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7525                 : /* No clobbers */);                                   \
7526        result;                                                          \
7527      })
7528
7529 #define vmlal_laneq_s16(a, b, c, d)                                     \
7530   __extension__                                                         \
7531     ({                                                                  \
7532        int16x8_t c_ = (c);                                              \
7533        int16x4_t b_ = (b);                                              \
7534        int32x4_t a_ = (a);                                              \
7535        int32x4_t result;                                                \
7536        __asm__ ("smlal %0.4s, %2.4h, %3.h[%4]"                          \
7537                 : "=w"(result)                                          \
7538                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7539                 : /* No clobbers */);                                   \
7540        result;                                                          \
7541      })
7542
7543 #define vmlal_laneq_s32(a, b, c, d)                                     \
7544   __extension__                                                         \
7545     ({                                                                  \
7546        int32x4_t c_ = (c);                                              \
7547        int32x2_t b_ = (b);                                              \
7548        int64x2_t a_ = (a);                                              \
7549        int64x2_t result;                                                \
7550        __asm__ ("smlal %0.2d, %2.2s, %3.s[%4]"                          \
7551                 : "=w"(result)                                          \
7552                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7553                 : /* No clobbers */);                                   \
7554        result;                                                          \
7555      })
7556
7557 #define vmlal_laneq_u16(a, b, c, d)                                     \
7558   __extension__                                                         \
7559     ({                                                                  \
7560        uint16x8_t c_ = (c);                                             \
7561        uint16x4_t b_ = (b);                                             \
7562        uint32x4_t a_ = (a);                                             \
7563        uint32x4_t result;                                               \
7564        __asm__ ("umlal %0.4s, %2.4h, %3.h[%4]"                          \
7565                 : "=w"(result)                                          \
7566                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7567                 : /* No clobbers */);                                   \
7568        result;                                                          \
7569      })
7570
7571 #define vmlal_laneq_u32(a, b, c, d)                                     \
7572   __extension__                                                         \
7573     ({                                                                  \
7574        uint32x4_t c_ = (c);                                             \
7575        uint32x2_t b_ = (b);                                             \
7576        uint64x2_t a_ = (a);                                             \
7577        uint64x2_t result;                                               \
7578        __asm__ ("umlal %0.2d, %2.2s, %3.s[%4]"                          \
7579                 : "=w"(result)                                          \
7580                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7581                 : /* No clobbers */);                                   \
7582        result;                                                          \
7583      })
7584
7585 __extension__ extern __inline int32x4_t
7586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7587 vmlal_n_s16 (int32x4_t a, int16x4_t b, int16_t c)
7588 {
7589   int32x4_t result;
7590   __asm__ ("smlal %0.4s,%2.4h,%3.h[0]"
7591            : "=w"(result)
7592            : "0"(a), "w"(b), "x"(c)
7593            : /* No clobbers */);
7594   return result;
7595 }
7596
7597 __extension__ extern __inline int64x2_t
7598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7599 vmlal_n_s32 (int64x2_t a, int32x2_t b, int32_t c)
7600 {
7601   int64x2_t result;
7602   __asm__ ("smlal %0.2d,%2.2s,%3.s[0]"
7603            : "=w"(result)
7604            : "0"(a), "w"(b), "w"(c)
7605            : /* No clobbers */);
7606   return result;
7607 }
7608
7609 __extension__ extern __inline uint32x4_t
7610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7611 vmlal_n_u16 (uint32x4_t a, uint16x4_t b, uint16_t c)
7612 {
7613   uint32x4_t result;
7614   __asm__ ("umlal %0.4s,%2.4h,%3.h[0]"
7615            : "=w"(result)
7616            : "0"(a), "w"(b), "x"(c)
7617            : /* No clobbers */);
7618   return result;
7619 }
7620
7621 __extension__ extern __inline uint64x2_t
7622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7623 vmlal_n_u32 (uint64x2_t a, uint32x2_t b, uint32_t c)
7624 {
7625   uint64x2_t result;
7626   __asm__ ("umlal %0.2d,%2.2s,%3.s[0]"
7627            : "=w"(result)
7628            : "0"(a), "w"(b), "w"(c)
7629            : /* No clobbers */);
7630   return result;
7631 }
7632
7633 __extension__ extern __inline int16x8_t
7634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7635 vmlal_s8 (int16x8_t a, int8x8_t b, int8x8_t c)
7636 {
7637   int16x8_t result;
7638   __asm__ ("smlal %0.8h,%2.8b,%3.8b"
7639            : "=w"(result)
7640            : "0"(a), "w"(b), "w"(c)
7641            : /* No clobbers */);
7642   return result;
7643 }
7644
7645 __extension__ extern __inline int32x4_t
7646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7647 vmlal_s16 (int32x4_t a, int16x4_t b, int16x4_t c)
7648 {
7649   int32x4_t result;
7650   __asm__ ("smlal %0.4s,%2.4h,%3.4h"
7651            : "=w"(result)
7652            : "0"(a), "w"(b), "w"(c)
7653            : /* No clobbers */);
7654   return result;
7655 }
7656
7657 __extension__ extern __inline int64x2_t
7658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7659 vmlal_s32 (int64x2_t a, int32x2_t b, int32x2_t c)
7660 {
7661   int64x2_t result;
7662   __asm__ ("smlal %0.2d,%2.2s,%3.2s"
7663            : "=w"(result)
7664            : "0"(a), "w"(b), "w"(c)
7665            : /* No clobbers */);
7666   return result;
7667 }
7668
7669 __extension__ extern __inline uint16x8_t
7670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7671 vmlal_u8 (uint16x8_t a, uint8x8_t b, uint8x8_t c)
7672 {
7673   uint16x8_t result;
7674   __asm__ ("umlal %0.8h,%2.8b,%3.8b"
7675            : "=w"(result)
7676            : "0"(a), "w"(b), "w"(c)
7677            : /* No clobbers */);
7678   return result;
7679 }
7680
7681 __extension__ extern __inline uint32x4_t
7682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7683 vmlal_u16 (uint32x4_t a, uint16x4_t b, uint16x4_t c)
7684 {
7685   uint32x4_t result;
7686   __asm__ ("umlal %0.4s,%2.4h,%3.4h"
7687            : "=w"(result)
7688            : "0"(a), "w"(b), "w"(c)
7689            : /* No clobbers */);
7690   return result;
7691 }
7692
7693 __extension__ extern __inline uint64x2_t
7694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7695 vmlal_u32 (uint64x2_t a, uint32x2_t b, uint32x2_t c)
7696 {
7697   uint64x2_t result;
7698   __asm__ ("umlal %0.2d,%2.2s,%3.2s"
7699            : "=w"(result)
7700            : "0"(a), "w"(b), "w"(c)
7701            : /* No clobbers */);
7702   return result;
7703 }
7704
7705 __extension__ extern __inline float32x4_t
7706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7707 vmlaq_n_f32 (float32x4_t a, float32x4_t b, float32_t c)
7708 {
7709   float32x4_t result;
7710   float32x4_t t1;
7711   __asm__ ("fmul %1.4s, %3.4s, %4.s[0]; fadd %0.4s, %0.4s, %1.4s"
7712            : "=w"(result), "=w"(t1)
7713            : "0"(a), "w"(b), "w"(c)
7714            : /* No clobbers */);
7715   return result;
7716 }
7717
7718 __extension__ extern __inline int16x8_t
7719 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7720 vmlaq_n_s16 (int16x8_t a, int16x8_t b, int16_t c)
7721 {
7722   int16x8_t result;
7723   __asm__ ("mla %0.8h,%2.8h,%3.h[0]"
7724            : "=w"(result)
7725            : "0"(a), "w"(b), "x"(c)
7726            : /* No clobbers */);
7727   return result;
7728 }
7729
7730 __extension__ extern __inline int32x4_t
7731 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7732 vmlaq_n_s32 (int32x4_t a, int32x4_t b, int32_t c)
7733 {
7734   int32x4_t result;
7735   __asm__ ("mla %0.4s,%2.4s,%3.s[0]"
7736            : "=w"(result)
7737            : "0"(a), "w"(b), "w"(c)
7738            : /* No clobbers */);
7739   return result;
7740 }
7741
7742 __extension__ extern __inline uint16x8_t
7743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7744 vmlaq_n_u16 (uint16x8_t a, uint16x8_t b, uint16_t c)
7745 {
7746   uint16x8_t result;
7747   __asm__ ("mla %0.8h,%2.8h,%3.h[0]"
7748            : "=w"(result)
7749            : "0"(a), "w"(b), "x"(c)
7750            : /* No clobbers */);
7751   return result;
7752 }
7753
7754 __extension__ extern __inline uint32x4_t
7755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7756 vmlaq_n_u32 (uint32x4_t a, uint32x4_t b, uint32_t c)
7757 {
7758   uint32x4_t result;
7759   __asm__ ("mla %0.4s,%2.4s,%3.s[0]"
7760            : "=w"(result)
7761            : "0"(a), "w"(b), "w"(c)
7762            : /* No clobbers */);
7763   return result;
7764 }
7765
7766 __extension__ extern __inline int8x16_t
7767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7768 vmlaq_s8 (int8x16_t a, int8x16_t b, int8x16_t c)
7769 {
7770   int8x16_t result;
7771   __asm__ ("mla %0.16b, %2.16b, %3.16b"
7772            : "=w"(result)
7773            : "0"(a), "w"(b), "w"(c)
7774            : /* No clobbers */);
7775   return result;
7776 }
7777
7778 __extension__ extern __inline int16x8_t
7779 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7780 vmlaq_s16 (int16x8_t a, int16x8_t b, int16x8_t c)
7781 {
7782   int16x8_t result;
7783   __asm__ ("mla %0.8h, %2.8h, %3.8h"
7784            : "=w"(result)
7785            : "0"(a), "w"(b), "w"(c)
7786            : /* No clobbers */);
7787   return result;
7788 }
7789
7790 __extension__ extern __inline int32x4_t
7791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7792 vmlaq_s32 (int32x4_t a, int32x4_t b, int32x4_t c)
7793 {
7794   int32x4_t result;
7795   __asm__ ("mla %0.4s, %2.4s, %3.4s"
7796            : "=w"(result)
7797            : "0"(a), "w"(b), "w"(c)
7798            : /* No clobbers */);
7799   return result;
7800 }
7801
7802 __extension__ extern __inline uint8x16_t
7803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7804 vmlaq_u8 (uint8x16_t a, uint8x16_t b, uint8x16_t c)
7805 {
7806   uint8x16_t result;
7807   __asm__ ("mla %0.16b, %2.16b, %3.16b"
7808            : "=w"(result)
7809            : "0"(a), "w"(b), "w"(c)
7810            : /* No clobbers */);
7811   return result;
7812 }
7813
7814 __extension__ extern __inline uint16x8_t
7815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7816 vmlaq_u16 (uint16x8_t a, uint16x8_t b, uint16x8_t c)
7817 {
7818   uint16x8_t result;
7819   __asm__ ("mla %0.8h, %2.8h, %3.8h"
7820            : "=w"(result)
7821            : "0"(a), "w"(b), "w"(c)
7822            : /* No clobbers */);
7823   return result;
7824 }
7825
7826 __extension__ extern __inline uint32x4_t
7827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7828 vmlaq_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t c)
7829 {
7830   uint32x4_t result;
7831   __asm__ ("mla %0.4s, %2.4s, %3.4s"
7832            : "=w"(result)
7833            : "0"(a), "w"(b), "w"(c)
7834            : /* No clobbers */);
7835   return result;
7836 }
7837
7838 __extension__ extern __inline float32x2_t
7839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7840 vmls_n_f32 (float32x2_t a, float32x2_t b, float32_t c)
7841 {
7842   float32x2_t result;
7843   float32x2_t t1;
7844   __asm__ ("fmul %1.2s, %3.2s, %4.s[0]; fsub %0.2s, %0.2s, %1.2s"
7845            : "=w"(result), "=w"(t1)
7846            : "0"(a), "w"(b), "w"(c)
7847            : /* No clobbers */);
7848   return result;
7849 }
7850
7851 __extension__ extern __inline int16x4_t
7852 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7853 vmls_n_s16 (int16x4_t a, int16x4_t b, int16_t c)
7854 {
7855   int16x4_t result;
7856   __asm__ ("mls %0.4h, %2.4h, %3.h[0]"
7857            : "=w"(result)
7858            : "0"(a), "w"(b), "x"(c)
7859            : /* No clobbers */);
7860   return result;
7861 }
7862
7863 __extension__ extern __inline int32x2_t
7864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7865 vmls_n_s32 (int32x2_t a, int32x2_t b, int32_t c)
7866 {
7867   int32x2_t result;
7868   __asm__ ("mls %0.2s, %2.2s, %3.s[0]"
7869            : "=w"(result)
7870            : "0"(a), "w"(b), "w"(c)
7871            : /* No clobbers */);
7872   return result;
7873 }
7874
7875 __extension__ extern __inline uint16x4_t
7876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7877 vmls_n_u16 (uint16x4_t a, uint16x4_t b, uint16_t c)
7878 {
7879   uint16x4_t result;
7880   __asm__ ("mls %0.4h, %2.4h, %3.h[0]"
7881            : "=w"(result)
7882            : "0"(a), "w"(b), "x"(c)
7883            : /* No clobbers */);
7884   return result;
7885 }
7886
7887 __extension__ extern __inline uint32x2_t
7888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7889 vmls_n_u32 (uint32x2_t a, uint32x2_t b, uint32_t c)
7890 {
7891   uint32x2_t result;
7892   __asm__ ("mls %0.2s, %2.2s, %3.s[0]"
7893            : "=w"(result)
7894            : "0"(a), "w"(b), "w"(c)
7895            : /* No clobbers */);
7896   return result;
7897 }
7898
7899 __extension__ extern __inline int8x8_t
7900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7901 vmls_s8 (int8x8_t a, int8x8_t b, int8x8_t c)
7902 {
7903   int8x8_t result;
7904   __asm__ ("mls %0.8b,%2.8b,%3.8b"
7905            : "=w"(result)
7906            : "0"(a), "w"(b), "w"(c)
7907            : /* No clobbers */);
7908   return result;
7909 }
7910
7911 __extension__ extern __inline int16x4_t
7912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7913 vmls_s16 (int16x4_t a, int16x4_t b, int16x4_t c)
7914 {
7915   int16x4_t result;
7916   __asm__ ("mls %0.4h,%2.4h,%3.4h"
7917            : "=w"(result)
7918            : "0"(a), "w"(b), "w"(c)
7919            : /* No clobbers */);
7920   return result;
7921 }
7922
7923 __extension__ extern __inline int32x2_t
7924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7925 vmls_s32 (int32x2_t a, int32x2_t b, int32x2_t c)
7926 {
7927   int32x2_t result;
7928   __asm__ ("mls %0.2s,%2.2s,%3.2s"
7929            : "=w"(result)
7930            : "0"(a), "w"(b), "w"(c)
7931            : /* No clobbers */);
7932   return result;
7933 }
7934
7935 __extension__ extern __inline uint8x8_t
7936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7937 vmls_u8 (uint8x8_t a, uint8x8_t b, uint8x8_t c)
7938 {
7939   uint8x8_t result;
7940   __asm__ ("mls %0.8b,%2.8b,%3.8b"
7941            : "=w"(result)
7942            : "0"(a), "w"(b), "w"(c)
7943            : /* No clobbers */);
7944   return result;
7945 }
7946
7947 __extension__ extern __inline uint16x4_t
7948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7949 vmls_u16 (uint16x4_t a, uint16x4_t b, uint16x4_t c)
7950 {
7951   uint16x4_t result;
7952   __asm__ ("mls %0.4h,%2.4h,%3.4h"
7953            : "=w"(result)
7954            : "0"(a), "w"(b), "w"(c)
7955            : /* No clobbers */);
7956   return result;
7957 }
7958
7959 __extension__ extern __inline uint32x2_t
7960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
7961 vmls_u32 (uint32x2_t a, uint32x2_t b, uint32x2_t c)
7962 {
7963   uint32x2_t result;
7964   __asm__ ("mls %0.2s,%2.2s,%3.2s"
7965            : "=w"(result)
7966            : "0"(a), "w"(b), "w"(c)
7967            : /* No clobbers */);
7968   return result;
7969 }
7970
7971 #define vmlsl_high_lane_s16(a, b, c, d)                                 \
7972   __extension__                                                         \
7973     ({                                                                  \
7974        int16x4_t c_ = (c);                                              \
7975        int16x8_t b_ = (b);                                              \
7976        int32x4_t a_ = (a);                                              \
7977        int32x4_t result;                                                \
7978        __asm__ ("smlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
7979                 : "=w"(result)                                          \
7980                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
7981                 : /* No clobbers */);                                   \
7982        result;                                                          \
7983      })
7984
7985 #define vmlsl_high_lane_s32(a, b, c, d)                                 \
7986   __extension__                                                         \
7987     ({                                                                  \
7988        int32x2_t c_ = (c);                                              \
7989        int32x4_t b_ = (b);                                              \
7990        int64x2_t a_ = (a);                                              \
7991        int64x2_t result;                                                \
7992        __asm__ ("smlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
7993                 : "=w"(result)                                          \
7994                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
7995                 : /* No clobbers */);                                   \
7996        result;                                                          \
7997      })
7998
7999 #define vmlsl_high_lane_u16(a, b, c, d)                                 \
8000   __extension__                                                         \
8001     ({                                                                  \
8002        uint16x4_t c_ = (c);                                             \
8003        uint16x8_t b_ = (b);                                             \
8004        uint32x4_t a_ = (a);                                             \
8005        uint32x4_t result;                                               \
8006        __asm__ ("umlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8007                 : "=w"(result)                                          \
8008                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8009                 : /* No clobbers */);                                   \
8010        result;                                                          \
8011      })
8012
8013 #define vmlsl_high_lane_u32(a, b, c, d)                                 \
8014   __extension__                                                         \
8015     ({                                                                  \
8016        uint32x2_t c_ = (c);                                             \
8017        uint32x4_t b_ = (b);                                             \
8018        uint64x2_t a_ = (a);                                             \
8019        uint64x2_t result;                                               \
8020        __asm__ ("umlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8021                 : "=w"(result)                                          \
8022                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8023                 : /* No clobbers */);                                   \
8024        result;                                                          \
8025      })
8026
8027 #define vmlsl_high_laneq_s16(a, b, c, d)                                \
8028   __extension__                                                         \
8029     ({                                                                  \
8030        int16x8_t c_ = (c);                                              \
8031        int16x8_t b_ = (b);                                              \
8032        int32x4_t a_ = (a);                                              \
8033        int32x4_t result;                                                \
8034        __asm__ ("smlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8035                 : "=w"(result)                                          \
8036                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8037                 : /* No clobbers */);                                   \
8038        result;                                                          \
8039      })
8040
8041 #define vmlsl_high_laneq_s32(a, b, c, d)                                \
8042   __extension__                                                         \
8043     ({                                                                  \
8044        int32x4_t c_ = (c);                                              \
8045        int32x4_t b_ = (b);                                              \
8046        int64x2_t a_ = (a);                                              \
8047        int64x2_t result;                                                \
8048        __asm__ ("smlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8049                 : "=w"(result)                                          \
8050                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8051                 : /* No clobbers */);                                   \
8052        result;                                                          \
8053      })
8054
8055 #define vmlsl_high_laneq_u16(a, b, c, d)                                \
8056   __extension__                                                         \
8057     ({                                                                  \
8058        uint16x8_t c_ = (c);                                             \
8059        uint16x8_t b_ = (b);                                             \
8060        uint32x4_t a_ = (a);                                             \
8061        uint32x4_t result;                                               \
8062        __asm__ ("umlsl2 %0.4s, %2.8h, %3.h[%4]"                         \
8063                 : "=w"(result)                                          \
8064                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8065                 : /* No clobbers */);                                   \
8066        result;                                                          \
8067      })
8068
8069 #define vmlsl_high_laneq_u32(a, b, c, d)                                \
8070   __extension__                                                         \
8071     ({                                                                  \
8072        uint32x4_t c_ = (c);                                             \
8073        uint32x4_t b_ = (b);                                             \
8074        uint64x2_t a_ = (a);                                             \
8075        uint64x2_t result;                                               \
8076        __asm__ ("umlsl2 %0.2d, %2.4s, %3.s[%4]"                         \
8077                 : "=w"(result)                                          \
8078                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8079                 : /* No clobbers */);                                   \
8080        result;                                                          \
8081      })
8082
8083 __extension__ extern __inline int32x4_t
8084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8085 vmlsl_high_n_s16 (int32x4_t a, int16x8_t b, int16_t c)
8086 {
8087   int32x4_t result;
8088   __asm__ ("smlsl2 %0.4s, %2.8h, %3.h[0]"
8089            : "=w"(result)
8090            : "0"(a), "w"(b), "x"(c)
8091            : /* No clobbers */);
8092   return result;
8093 }
8094
8095 __extension__ extern __inline int64x2_t
8096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8097 vmlsl_high_n_s32 (int64x2_t a, int32x4_t b, int32_t c)
8098 {
8099   int64x2_t result;
8100   __asm__ ("smlsl2 %0.2d, %2.4s, %3.s[0]"
8101            : "=w"(result)
8102            : "0"(a), "w"(b), "w"(c)
8103            : /* No clobbers */);
8104   return result;
8105 }
8106
8107 __extension__ extern __inline uint32x4_t
8108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8109 vmlsl_high_n_u16 (uint32x4_t a, uint16x8_t b, uint16_t c)
8110 {
8111   uint32x4_t result;
8112   __asm__ ("umlsl2 %0.4s, %2.8h, %3.h[0]"
8113            : "=w"(result)
8114            : "0"(a), "w"(b), "x"(c)
8115            : /* No clobbers */);
8116   return result;
8117 }
8118
8119 __extension__ extern __inline uint64x2_t
8120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8121 vmlsl_high_n_u32 (uint64x2_t a, uint32x4_t b, uint32_t c)
8122 {
8123   uint64x2_t result;
8124   __asm__ ("umlsl2 %0.2d, %2.4s, %3.s[0]"
8125            : "=w"(result)
8126            : "0"(a), "w"(b), "w"(c)
8127            : /* No clobbers */);
8128   return result;
8129 }
8130
8131 __extension__ extern __inline int16x8_t
8132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8133 vmlsl_high_s8 (int16x8_t a, int8x16_t b, int8x16_t c)
8134 {
8135   int16x8_t result;
8136   __asm__ ("smlsl2 %0.8h,%2.16b,%3.16b"
8137            : "=w"(result)
8138            : "0"(a), "w"(b), "w"(c)
8139            : /* No clobbers */);
8140   return result;
8141 }
8142
8143 __extension__ extern __inline int32x4_t
8144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8145 vmlsl_high_s16 (int32x4_t a, int16x8_t b, int16x8_t c)
8146 {
8147   int32x4_t result;
8148   __asm__ ("smlsl2 %0.4s,%2.8h,%3.8h"
8149            : "=w"(result)
8150            : "0"(a), "w"(b), "w"(c)
8151            : /* No clobbers */);
8152   return result;
8153 }
8154
8155 __extension__ extern __inline int64x2_t
8156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8157 vmlsl_high_s32 (int64x2_t a, int32x4_t b, int32x4_t c)
8158 {
8159   int64x2_t result;
8160   __asm__ ("smlsl2 %0.2d,%2.4s,%3.4s"
8161            : "=w"(result)
8162            : "0"(a), "w"(b), "w"(c)
8163            : /* No clobbers */);
8164   return result;
8165 }
8166
8167 __extension__ extern __inline uint16x8_t
8168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8169 vmlsl_high_u8 (uint16x8_t a, uint8x16_t b, uint8x16_t c)
8170 {
8171   uint16x8_t result;
8172   __asm__ ("umlsl2 %0.8h,%2.16b,%3.16b"
8173            : "=w"(result)
8174            : "0"(a), "w"(b), "w"(c)
8175            : /* No clobbers */);
8176   return result;
8177 }
8178
8179 __extension__ extern __inline uint32x4_t
8180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8181 vmlsl_high_u16 (uint32x4_t a, uint16x8_t b, uint16x8_t c)
8182 {
8183   uint32x4_t result;
8184   __asm__ ("umlsl2 %0.4s,%2.8h,%3.8h"
8185            : "=w"(result)
8186            : "0"(a), "w"(b), "w"(c)
8187            : /* No clobbers */);
8188   return result;
8189 }
8190
8191 __extension__ extern __inline uint64x2_t
8192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8193 vmlsl_high_u32 (uint64x2_t a, uint32x4_t b, uint32x4_t c)
8194 {
8195   uint64x2_t result;
8196   __asm__ ("umlsl2 %0.2d,%2.4s,%3.4s"
8197            : "=w"(result)
8198            : "0"(a), "w"(b), "w"(c)
8199            : /* No clobbers */);
8200   return result;
8201 }
8202
8203 #define vmlsl_lane_s16(a, b, c, d)                                      \
8204   __extension__                                                         \
8205     ({                                                                  \
8206        int16x4_t c_ = (c);                                              \
8207        int16x4_t b_ = (b);                                              \
8208        int32x4_t a_ = (a);                                              \
8209        int32x4_t result;                                                \
8210        __asm__ ("smlsl %0.4s, %2.4h, %3.h[%4]"                          \
8211                 : "=w"(result)                                          \
8212                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8213                 : /* No clobbers */);                                   \
8214        result;                                                          \
8215      })
8216
8217 #define vmlsl_lane_s32(a, b, c, d)                                      \
8218   __extension__                                                         \
8219     ({                                                                  \
8220        int32x2_t c_ = (c);                                              \
8221        int32x2_t b_ = (b);                                              \
8222        int64x2_t a_ = (a);                                              \
8223        int64x2_t result;                                                \
8224        __asm__ ("smlsl %0.2d, %2.2s, %3.s[%4]"                          \
8225                 : "=w"(result)                                          \
8226                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8227                 : /* No clobbers */);                                   \
8228        result;                                                          \
8229      })
8230
8231 #define vmlsl_lane_u16(a, b, c, d)                                      \
8232   __extension__                                                         \
8233     ({                                                                  \
8234        uint16x4_t c_ = (c);                                             \
8235        uint16x4_t b_ = (b);                                             \
8236        uint32x4_t a_ = (a);                                             \
8237        uint32x4_t result;                                               \
8238        __asm__ ("umlsl %0.4s, %2.4h, %3.h[%4]"                          \
8239                 : "=w"(result)                                          \
8240                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8241                 : /* No clobbers */);                                   \
8242        result;                                                          \
8243      })
8244
8245 #define vmlsl_lane_u32(a, b, c, d)                                      \
8246   __extension__                                                         \
8247     ({                                                                  \
8248        uint32x2_t c_ = (c);                                             \
8249        uint32x2_t b_ = (b);                                             \
8250        uint64x2_t a_ = (a);                                             \
8251        uint64x2_t result;                                               \
8252        __asm__ ("umlsl %0.2d, %2.2s, %3.s[%4]"                          \
8253                 : "=w"(result)                                          \
8254                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8255                 : /* No clobbers */);                                   \
8256        result;                                                          \
8257      })
8258
8259 #define vmlsl_laneq_s16(a, b, c, d)                                     \
8260   __extension__                                                         \
8261     ({                                                                  \
8262        int16x8_t c_ = (c);                                              \
8263        int16x4_t b_ = (b);                                              \
8264        int32x4_t a_ = (a);                                              \
8265        int32x4_t result;                                                \
8266        __asm__ ("smlsl %0.4s, %2.4h, %3.h[%4]"                          \
8267                 : "=w"(result)                                          \
8268                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8269                 : /* No clobbers */);                                   \
8270        result;                                                          \
8271      })
8272
8273 #define vmlsl_laneq_s32(a, b, c, d)                                     \
8274   __extension__                                                         \
8275     ({                                                                  \
8276        int32x4_t c_ = (c);                                              \
8277        int32x2_t b_ = (b);                                              \
8278        int64x2_t a_ = (a);                                              \
8279        int64x2_t result;                                                \
8280        __asm__ ("smlsl %0.2d, %2.2s, %3.s[%4]"                          \
8281                 : "=w"(result)                                          \
8282                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8283                 : /* No clobbers */);                                   \
8284        result;                                                          \
8285      })
8286
8287 #define vmlsl_laneq_u16(a, b, c, d)                                     \
8288   __extension__                                                         \
8289     ({                                                                  \
8290        uint16x8_t c_ = (c);                                             \
8291        uint16x4_t b_ = (b);                                             \
8292        uint32x4_t a_ = (a);                                             \
8293        uint32x4_t result;                                               \
8294        __asm__ ("umlsl %0.4s, %2.4h, %3.h[%4]"                          \
8295                 : "=w"(result)                                          \
8296                 : "0"(a_), "w"(b_), "x"(c_), "i"(d)                     \
8297                 : /* No clobbers */);                                   \
8298        result;                                                          \
8299      })
8300
8301 #define vmlsl_laneq_u32(a, b, c, d)                                     \
8302   __extension__                                                         \
8303     ({                                                                  \
8304        uint32x4_t c_ = (c);                                             \
8305        uint32x2_t b_ = (b);                                             \
8306        uint64x2_t a_ = (a);                                             \
8307        uint64x2_t result;                                               \
8308        __asm__ ("umlsl %0.2d, %2.2s, %3.s[%4]"                          \
8309                 : "=w"(result)                                          \
8310                 : "0"(a_), "w"(b_), "w"(c_), "i"(d)                     \
8311                 : /* No clobbers */);                                   \
8312        result;                                                          \
8313      })
8314
8315 __extension__ extern __inline int32x4_t
8316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8317 vmlsl_n_s16 (int32x4_t a, int16x4_t b, int16_t c)
8318 {
8319   int32x4_t result;
8320   __asm__ ("smlsl %0.4s, %2.4h, %3.h[0]"
8321            : "=w"(result)
8322            : "0"(a), "w"(b), "x"(c)
8323            : /* No clobbers */);
8324   return result;
8325 }
8326
8327 __extension__ extern __inline int64x2_t
8328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8329 vmlsl_n_s32 (int64x2_t a, int32x2_t b, int32_t c)
8330 {
8331   int64x2_t result;
8332   __asm__ ("smlsl %0.2d, %2.2s, %3.s[0]"
8333            : "=w"(result)
8334            : "0"(a), "w"(b), "w"(c)
8335            : /* No clobbers */);
8336   return result;
8337 }
8338
8339 __extension__ extern __inline uint32x4_t
8340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8341 vmlsl_n_u16 (uint32x4_t a, uint16x4_t b, uint16_t c)
8342 {
8343   uint32x4_t result;
8344   __asm__ ("umlsl %0.4s, %2.4h, %3.h[0]"
8345            : "=w"(result)
8346            : "0"(a), "w"(b), "x"(c)
8347            : /* No clobbers */);
8348   return result;
8349 }
8350
8351 __extension__ extern __inline uint64x2_t
8352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8353 vmlsl_n_u32 (uint64x2_t a, uint32x2_t b, uint32_t c)
8354 {
8355   uint64x2_t result;
8356   __asm__ ("umlsl %0.2d, %2.2s, %3.s[0]"
8357            : "=w"(result)
8358            : "0"(a), "w"(b), "w"(c)
8359            : /* No clobbers */);
8360   return result;
8361 }
8362
8363 __extension__ extern __inline int16x8_t
8364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8365 vmlsl_s8 (int16x8_t a, int8x8_t b, int8x8_t c)
8366 {
8367   int16x8_t result;
8368   __asm__ ("smlsl %0.8h, %2.8b, %3.8b"
8369            : "=w"(result)
8370            : "0"(a), "w"(b), "w"(c)
8371            : /* No clobbers */);
8372   return result;
8373 }
8374
8375 __extension__ extern __inline int32x4_t
8376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8377 vmlsl_s16 (int32x4_t a, int16x4_t b, int16x4_t c)
8378 {
8379   int32x4_t result;
8380   __asm__ ("smlsl %0.4s, %2.4h, %3.4h"
8381            : "=w"(result)
8382            : "0"(a), "w"(b), "w"(c)
8383            : /* No clobbers */);
8384   return result;
8385 }
8386
8387 __extension__ extern __inline int64x2_t
8388 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8389 vmlsl_s32 (int64x2_t a, int32x2_t b, int32x2_t c)
8390 {
8391   int64x2_t result;
8392   __asm__ ("smlsl %0.2d, %2.2s, %3.2s"
8393            : "=w"(result)
8394            : "0"(a), "w"(b), "w"(c)
8395            : /* No clobbers */);
8396   return result;
8397 }
8398
8399 __extension__ extern __inline uint16x8_t
8400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8401 vmlsl_u8 (uint16x8_t a, uint8x8_t b, uint8x8_t c)
8402 {
8403   uint16x8_t result;
8404   __asm__ ("umlsl %0.8h, %2.8b, %3.8b"
8405            : "=w"(result)
8406            : "0"(a), "w"(b), "w"(c)
8407            : /* No clobbers */);
8408   return result;
8409 }
8410
8411 __extension__ extern __inline uint32x4_t
8412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8413 vmlsl_u16 (uint32x4_t a, uint16x4_t b, uint16x4_t c)
8414 {
8415   uint32x4_t result;
8416   __asm__ ("umlsl %0.4s, %2.4h, %3.4h"
8417            : "=w"(result)
8418            : "0"(a), "w"(b), "w"(c)
8419            : /* No clobbers */);
8420   return result;
8421 }
8422
8423 __extension__ extern __inline uint64x2_t
8424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8425 vmlsl_u32 (uint64x2_t a, uint32x2_t b, uint32x2_t c)
8426 {
8427   uint64x2_t result;
8428   __asm__ ("umlsl %0.2d, %2.2s, %3.2s"
8429            : "=w"(result)
8430            : "0"(a), "w"(b), "w"(c)
8431            : /* No clobbers */);
8432   return result;
8433 }
8434
8435 __extension__ extern __inline float32x4_t
8436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8437 vmlsq_n_f32 (float32x4_t a, float32x4_t b, float32_t c)
8438 {
8439   float32x4_t result;
8440   float32x4_t t1;
8441   __asm__ ("fmul %1.4s, %3.4s, %4.s[0]; fsub %0.4s, %0.4s, %1.4s"
8442            : "=w"(result), "=w"(t1)
8443            : "0"(a), "w"(b), "w"(c)
8444            : /* No clobbers */);
8445   return result;
8446 }
8447
8448 __extension__ extern __inline int16x8_t
8449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8450 vmlsq_n_s16 (int16x8_t a, int16x8_t b, int16_t c)
8451 {
8452   int16x8_t result;
8453   __asm__ ("mls %0.8h, %2.8h, %3.h[0]"
8454            : "=w"(result)
8455            : "0"(a), "w"(b), "x"(c)
8456            : /* No clobbers */);
8457   return result;
8458 }
8459
8460 __extension__ extern __inline int32x4_t
8461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8462 vmlsq_n_s32 (int32x4_t a, int32x4_t b, int32_t c)
8463 {
8464   int32x4_t result;
8465   __asm__ ("mls %0.4s, %2.4s, %3.s[0]"
8466            : "=w"(result)
8467            : "0"(a), "w"(b), "w"(c)
8468            : /* No clobbers */);
8469   return result;
8470 }
8471
8472 __extension__ extern __inline uint16x8_t
8473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8474 vmlsq_n_u16 (uint16x8_t a, uint16x8_t b, uint16_t c)
8475 {
8476   uint16x8_t result;
8477   __asm__ ("mls %0.8h, %2.8h, %3.h[0]"
8478            : "=w"(result)
8479            : "0"(a), "w"(b), "x"(c)
8480            : /* No clobbers */);
8481   return result;
8482 }
8483
8484 __extension__ extern __inline uint32x4_t
8485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8486 vmlsq_n_u32 (uint32x4_t a, uint32x4_t b, uint32_t c)
8487 {
8488   uint32x4_t result;
8489   __asm__ ("mls %0.4s, %2.4s, %3.s[0]"
8490            : "=w"(result)
8491            : "0"(a), "w"(b), "w"(c)
8492            : /* No clobbers */);
8493   return result;
8494 }
8495
8496 __extension__ extern __inline int8x16_t
8497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8498 vmlsq_s8 (int8x16_t a, int8x16_t b, int8x16_t c)
8499 {
8500   int8x16_t result;
8501   __asm__ ("mls %0.16b,%2.16b,%3.16b"
8502            : "=w"(result)
8503            : "0"(a), "w"(b), "w"(c)
8504            : /* No clobbers */);
8505   return result;
8506 }
8507
8508 __extension__ extern __inline int16x8_t
8509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8510 vmlsq_s16 (int16x8_t a, int16x8_t b, int16x8_t c)
8511 {
8512   int16x8_t result;
8513   __asm__ ("mls %0.8h,%2.8h,%3.8h"
8514            : "=w"(result)
8515            : "0"(a), "w"(b), "w"(c)
8516            : /* No clobbers */);
8517   return result;
8518 }
8519
8520 __extension__ extern __inline int32x4_t
8521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8522 vmlsq_s32 (int32x4_t a, int32x4_t b, int32x4_t c)
8523 {
8524   int32x4_t result;
8525   __asm__ ("mls %0.4s,%2.4s,%3.4s"
8526            : "=w"(result)
8527            : "0"(a), "w"(b), "w"(c)
8528            : /* No clobbers */);
8529   return result;
8530 }
8531
8532 __extension__ extern __inline uint8x16_t
8533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8534 vmlsq_u8 (uint8x16_t a, uint8x16_t b, uint8x16_t c)
8535 {
8536   uint8x16_t result;
8537   __asm__ ("mls %0.16b,%2.16b,%3.16b"
8538            : "=w"(result)
8539            : "0"(a), "w"(b), "w"(c)
8540            : /* No clobbers */);
8541   return result;
8542 }
8543
8544 __extension__ extern __inline uint16x8_t
8545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8546 vmlsq_u16 (uint16x8_t a, uint16x8_t b, uint16x8_t c)
8547 {
8548   uint16x8_t result;
8549   __asm__ ("mls %0.8h,%2.8h,%3.8h"
8550            : "=w"(result)
8551            : "0"(a), "w"(b), "w"(c)
8552            : /* No clobbers */);
8553   return result;
8554 }
8555
8556 __extension__ extern __inline uint32x4_t
8557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8558 vmlsq_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t c)
8559 {
8560   uint32x4_t result;
8561   __asm__ ("mls %0.4s,%2.4s,%3.4s"
8562            : "=w"(result)
8563            : "0"(a), "w"(b), "w"(c)
8564            : /* No clobbers */);
8565   return result;
8566 }
8567
8568 __extension__ extern __inline int16x8_t
8569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8570 vmovl_high_s8 (int8x16_t a)
8571 {
8572   int16x8_t result;
8573   __asm__ ("sshll2 %0.8h,%1.16b,#0"
8574            : "=w"(result)
8575            : "w"(a)
8576            : /* No clobbers */);
8577   return result;
8578 }
8579
8580 __extension__ extern __inline int32x4_t
8581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8582 vmovl_high_s16 (int16x8_t a)
8583 {
8584   int32x4_t result;
8585   __asm__ ("sshll2 %0.4s,%1.8h,#0"
8586            : "=w"(result)
8587            : "w"(a)
8588            : /* No clobbers */);
8589   return result;
8590 }
8591
8592 __extension__ extern __inline int64x2_t
8593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8594 vmovl_high_s32 (int32x4_t a)
8595 {
8596   int64x2_t result;
8597   __asm__ ("sshll2 %0.2d,%1.4s,#0"
8598            : "=w"(result)
8599            : "w"(a)
8600            : /* No clobbers */);
8601   return result;
8602 }
8603
8604 __extension__ extern __inline uint16x8_t
8605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8606 vmovl_high_u8 (uint8x16_t a)
8607 {
8608   uint16x8_t result;
8609   __asm__ ("ushll2 %0.8h,%1.16b,#0"
8610            : "=w"(result)
8611            : "w"(a)
8612            : /* No clobbers */);
8613   return result;
8614 }
8615
8616 __extension__ extern __inline uint32x4_t
8617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8618 vmovl_high_u16 (uint16x8_t a)
8619 {
8620   uint32x4_t result;
8621   __asm__ ("ushll2 %0.4s,%1.8h,#0"
8622            : "=w"(result)
8623            : "w"(a)
8624            : /* No clobbers */);
8625   return result;
8626 }
8627
8628 __extension__ extern __inline uint64x2_t
8629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8630 vmovl_high_u32 (uint32x4_t a)
8631 {
8632   uint64x2_t result;
8633   __asm__ ("ushll2 %0.2d,%1.4s,#0"
8634            : "=w"(result)
8635            : "w"(a)
8636            : /* No clobbers */);
8637   return result;
8638 }
8639
8640 __extension__ extern __inline int16x8_t
8641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8642 vmovl_s8 (int8x8_t a)
8643 {
8644   int16x8_t result;
8645   __asm__ ("sshll %0.8h,%1.8b,#0"
8646            : "=w"(result)
8647            : "w"(a)
8648            : /* No clobbers */);
8649   return result;
8650 }
8651
8652 __extension__ extern __inline int32x4_t
8653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8654 vmovl_s16 (int16x4_t a)
8655 {
8656   int32x4_t result;
8657   __asm__ ("sshll %0.4s,%1.4h,#0"
8658            : "=w"(result)
8659            : "w"(a)
8660            : /* No clobbers */);
8661   return result;
8662 }
8663
8664 __extension__ extern __inline int64x2_t
8665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8666 vmovl_s32 (int32x2_t a)
8667 {
8668   int64x2_t result;
8669   __asm__ ("sshll %0.2d,%1.2s,#0"
8670            : "=w"(result)
8671            : "w"(a)
8672            : /* No clobbers */);
8673   return result;
8674 }
8675
8676 __extension__ extern __inline uint16x8_t
8677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8678 vmovl_u8 (uint8x8_t a)
8679 {
8680   uint16x8_t result;
8681   __asm__ ("ushll %0.8h,%1.8b,#0"
8682            : "=w"(result)
8683            : "w"(a)
8684            : /* No clobbers */);
8685   return result;
8686 }
8687
8688 __extension__ extern __inline uint32x4_t
8689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8690 vmovl_u16 (uint16x4_t a)
8691 {
8692   uint32x4_t result;
8693   __asm__ ("ushll %0.4s,%1.4h,#0"
8694            : "=w"(result)
8695            : "w"(a)
8696            : /* No clobbers */);
8697   return result;
8698 }
8699
8700 __extension__ extern __inline uint64x2_t
8701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8702 vmovl_u32 (uint32x2_t a)
8703 {
8704   uint64x2_t result;
8705   __asm__ ("ushll %0.2d,%1.2s,#0"
8706            : "=w"(result)
8707            : "w"(a)
8708            : /* No clobbers */);
8709   return result;
8710 }
8711
8712 __extension__ extern __inline int8x16_t
8713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8714 vmovn_high_s16 (int8x8_t a, int16x8_t b)
8715 {
8716   int8x16_t result = vcombine_s8 (a, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
8717   __asm__ ("xtn2 %0.16b,%1.8h"
8718            : "+w"(result)
8719            : "w"(b)
8720            : /* No clobbers */);
8721   return result;
8722 }
8723
8724 __extension__ extern __inline int16x8_t
8725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8726 vmovn_high_s32 (int16x4_t a, int32x4_t b)
8727 {
8728   int16x8_t result = vcombine_s16 (a, vcreate_s16 (__AARCH64_UINT64_C (0x0)));
8729   __asm__ ("xtn2 %0.8h,%1.4s"
8730            : "+w"(result)
8731            : "w"(b)
8732            : /* No clobbers */);
8733   return result;
8734 }
8735
8736 __extension__ extern __inline int32x4_t
8737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8738 vmovn_high_s64 (int32x2_t a, int64x2_t b)
8739 {
8740   int32x4_t result = vcombine_s32 (a, vcreate_s32 (__AARCH64_UINT64_C (0x0)));
8741   __asm__ ("xtn2 %0.4s,%1.2d"
8742            : "+w"(result)
8743            : "w"(b)
8744            : /* No clobbers */);
8745   return result;
8746 }
8747
8748 __extension__ extern __inline uint8x16_t
8749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8750 vmovn_high_u16 (uint8x8_t a, uint16x8_t b)
8751 {
8752   uint8x16_t result = vcombine_u8 (a, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
8753   __asm__ ("xtn2 %0.16b,%1.8h"
8754            : "+w"(result)
8755            : "w"(b)
8756            : /* No clobbers */);
8757   return result;
8758 }
8759
8760 __extension__ extern __inline uint16x8_t
8761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8762 vmovn_high_u32 (uint16x4_t a, uint32x4_t b)
8763 {
8764   uint16x8_t result = vcombine_u16 (a, vcreate_u16 (__AARCH64_UINT64_C (0x0)));
8765   __asm__ ("xtn2 %0.8h,%1.4s"
8766            : "+w"(result)
8767            : "w"(b)
8768            : /* No clobbers */);
8769   return result;
8770 }
8771
8772 __extension__ extern __inline uint32x4_t
8773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8774 vmovn_high_u64 (uint32x2_t a, uint64x2_t b)
8775 {
8776   uint32x4_t result = vcombine_u32 (a, vcreate_u32 (__AARCH64_UINT64_C (0x0)));
8777   __asm__ ("xtn2 %0.4s,%1.2d"
8778            : "+w"(result)
8779            : "w"(b)
8780            : /* No clobbers */);
8781   return result;
8782 }
8783
8784 __extension__ extern __inline int8x8_t
8785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8786 vmovn_s16 (int16x8_t a)
8787 {
8788   int8x8_t result;
8789   __asm__ ("xtn %0.8b,%1.8h"
8790            : "=w"(result)
8791            : "w"(a)
8792            : /* No clobbers */);
8793   return result;
8794 }
8795
8796 __extension__ extern __inline int16x4_t
8797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8798 vmovn_s32 (int32x4_t a)
8799 {
8800   int16x4_t result;
8801   __asm__ ("xtn %0.4h,%1.4s"
8802            : "=w"(result)
8803            : "w"(a)
8804            : /* No clobbers */);
8805   return result;
8806 }
8807
8808 __extension__ extern __inline int32x2_t
8809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8810 vmovn_s64 (int64x2_t a)
8811 {
8812   int32x2_t result;
8813   __asm__ ("xtn %0.2s,%1.2d"
8814            : "=w"(result)
8815            : "w"(a)
8816            : /* No clobbers */);
8817   return result;
8818 }
8819
8820 __extension__ extern __inline uint8x8_t
8821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8822 vmovn_u16 (uint16x8_t a)
8823 {
8824   uint8x8_t result;
8825   __asm__ ("xtn %0.8b,%1.8h"
8826            : "=w"(result)
8827            : "w"(a)
8828            : /* No clobbers */);
8829   return result;
8830 }
8831
8832 __extension__ extern __inline uint16x4_t
8833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8834 vmovn_u32 (uint32x4_t a)
8835 {
8836   uint16x4_t result;
8837   __asm__ ("xtn %0.4h,%1.4s"
8838            : "=w"(result)
8839            : "w"(a)
8840            : /* No clobbers */);
8841   return result;
8842 }
8843
8844 __extension__ extern __inline uint32x2_t
8845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8846 vmovn_u64 (uint64x2_t a)
8847 {
8848   uint32x2_t result;
8849   __asm__ ("xtn %0.2s,%1.2d"
8850            : "=w"(result)
8851            : "w"(a)
8852            : /* No clobbers */);
8853   return result;
8854 }
8855
8856 #define vmull_high_lane_s16(a, b, c)                                    \
8857   __extension__                                                         \
8858     ({                                                                  \
8859        int16x4_t b_ = (b);                                              \
8860        int16x8_t a_ = (a);                                              \
8861        int32x4_t result;                                                \
8862        __asm__ ("smull2 %0.4s, %1.8h, %2.h[%3]"                         \
8863                 : "=w"(result)                                          \
8864                 : "w"(a_), "x"(b_), "i"(c)                              \
8865                 : /* No clobbers */);                                   \
8866        result;                                                          \
8867      })
8868
8869 #define vmull_high_lane_s32(a, b, c)                                    \
8870   __extension__                                                         \
8871     ({                                                                  \
8872        int32x2_t b_ = (b);                                              \
8873        int32x4_t a_ = (a);                                              \
8874        int64x2_t result;                                                \
8875        __asm__ ("smull2 %0.2d, %1.4s, %2.s[%3]"                         \
8876                 : "=w"(result)                                          \
8877                 : "w"(a_), "w"(b_), "i"(c)                              \
8878                 : /* No clobbers */);                                   \
8879        result;                                                          \
8880      })
8881
8882 #define vmull_high_lane_u16(a, b, c)                                    \
8883   __extension__                                                         \
8884     ({                                                                  \
8885        uint16x4_t b_ = (b);                                             \
8886        uint16x8_t a_ = (a);                                             \
8887        uint32x4_t result;                                               \
8888        __asm__ ("umull2 %0.4s, %1.8h, %2.h[%3]"                         \
8889                 : "=w"(result)                                          \
8890                 : "w"(a_), "x"(b_), "i"(c)                              \
8891                 : /* No clobbers */);                                   \
8892        result;                                                          \
8893      })
8894
8895 #define vmull_high_lane_u32(a, b, c)                                    \
8896   __extension__                                                         \
8897     ({                                                                  \
8898        uint32x2_t b_ = (b);                                             \
8899        uint32x4_t a_ = (a);                                             \
8900        uint64x2_t result;                                               \
8901        __asm__ ("umull2 %0.2d, %1.4s, %2.s[%3]"                         \
8902                 : "=w"(result)                                          \
8903                 : "w"(a_), "w"(b_), "i"(c)                              \
8904                 : /* No clobbers */);                                   \
8905        result;                                                          \
8906      })
8907
8908 #define vmull_high_laneq_s16(a, b, c)                                   \
8909   __extension__                                                         \
8910     ({                                                                  \
8911        int16x8_t b_ = (b);                                              \
8912        int16x8_t a_ = (a);                                              \
8913        int32x4_t result;                                                \
8914        __asm__ ("smull2 %0.4s, %1.8h, %2.h[%3]"                         \
8915                 : "=w"(result)                                          \
8916                 : "w"(a_), "x"(b_), "i"(c)                              \
8917                 : /* No clobbers */);                                   \
8918        result;                                                          \
8919      })
8920
8921 #define vmull_high_laneq_s32(a, b, c)                                   \
8922   __extension__                                                         \
8923     ({                                                                  \
8924        int32x4_t b_ = (b);                                              \
8925        int32x4_t a_ = (a);                                              \
8926        int64x2_t result;                                                \
8927        __asm__ ("smull2 %0.2d, %1.4s, %2.s[%3]"                         \
8928                 : "=w"(result)                                          \
8929                 : "w"(a_), "w"(b_), "i"(c)                              \
8930                 : /* No clobbers */);                                   \
8931        result;                                                          \
8932      })
8933
8934 #define vmull_high_laneq_u16(a, b, c)                                   \
8935   __extension__                                                         \
8936     ({                                                                  \
8937        uint16x8_t b_ = (b);                                             \
8938        uint16x8_t a_ = (a);                                             \
8939        uint32x4_t result;                                               \
8940        __asm__ ("umull2 %0.4s, %1.8h, %2.h[%3]"                         \
8941                 : "=w"(result)                                          \
8942                 : "w"(a_), "x"(b_), "i"(c)                              \
8943                 : /* No clobbers */);                                   \
8944        result;                                                          \
8945      })
8946
8947 #define vmull_high_laneq_u32(a, b, c)                                   \
8948   __extension__                                                         \
8949     ({                                                                  \
8950        uint32x4_t b_ = (b);                                             \
8951        uint32x4_t a_ = (a);                                             \
8952        uint64x2_t result;                                               \
8953        __asm__ ("umull2 %0.2d, %1.4s, %2.s[%3]"                         \
8954                 : "=w"(result)                                          \
8955                 : "w"(a_), "w"(b_), "i"(c)                              \
8956                 : /* No clobbers */);                                   \
8957        result;                                                          \
8958      })
8959
8960 __extension__ extern __inline int32x4_t
8961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8962 vmull_high_n_s16 (int16x8_t a, int16_t b)
8963 {
8964   int32x4_t result;
8965   __asm__ ("smull2 %0.4s,%1.8h,%2.h[0]"
8966            : "=w"(result)
8967            : "w"(a), "x"(b)
8968            : /* No clobbers */);
8969   return result;
8970 }
8971
8972 __extension__ extern __inline int64x2_t
8973 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8974 vmull_high_n_s32 (int32x4_t a, int32_t b)
8975 {
8976   int64x2_t result;
8977   __asm__ ("smull2 %0.2d,%1.4s,%2.s[0]"
8978            : "=w"(result)
8979            : "w"(a), "w"(b)
8980            : /* No clobbers */);
8981   return result;
8982 }
8983
8984 __extension__ extern __inline uint32x4_t
8985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8986 vmull_high_n_u16 (uint16x8_t a, uint16_t b)
8987 {
8988   uint32x4_t result;
8989   __asm__ ("umull2 %0.4s,%1.8h,%2.h[0]"
8990            : "=w"(result)
8991            : "w"(a), "x"(b)
8992            : /* No clobbers */);
8993   return result;
8994 }
8995
8996 __extension__ extern __inline uint64x2_t
8997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
8998 vmull_high_n_u32 (uint32x4_t a, uint32_t b)
8999 {
9000   uint64x2_t result;
9001   __asm__ ("umull2 %0.2d,%1.4s,%2.s[0]"
9002            : "=w"(result)
9003            : "w"(a), "w"(b)
9004            : /* No clobbers */);
9005   return result;
9006 }
9007
9008 __extension__ extern __inline poly16x8_t
9009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9010 vmull_high_p8 (poly8x16_t a, poly8x16_t b)
9011 {
9012   poly16x8_t result;
9013   __asm__ ("pmull2 %0.8h,%1.16b,%2.16b"
9014            : "=w"(result)
9015            : "w"(a), "w"(b)
9016            : /* No clobbers */);
9017   return result;
9018 }
9019
9020 __extension__ extern __inline int16x8_t
9021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9022 vmull_high_s8 (int8x16_t a, int8x16_t b)
9023 {
9024   int16x8_t result;
9025   __asm__ ("smull2 %0.8h,%1.16b,%2.16b"
9026            : "=w"(result)
9027            : "w"(a), "w"(b)
9028            : /* No clobbers */);
9029   return result;
9030 }
9031
9032 __extension__ extern __inline int32x4_t
9033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9034 vmull_high_s16 (int16x8_t a, int16x8_t b)
9035 {
9036   int32x4_t result;
9037   __asm__ ("smull2 %0.4s,%1.8h,%2.8h"
9038            : "=w"(result)
9039            : "w"(a), "w"(b)
9040            : /* No clobbers */);
9041   return result;
9042 }
9043
9044 __extension__ extern __inline int64x2_t
9045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9046 vmull_high_s32 (int32x4_t a, int32x4_t b)
9047 {
9048   int64x2_t result;
9049   __asm__ ("smull2 %0.2d,%1.4s,%2.4s"
9050            : "=w"(result)
9051            : "w"(a), "w"(b)
9052            : /* No clobbers */);
9053   return result;
9054 }
9055
9056 __extension__ extern __inline uint16x8_t
9057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9058 vmull_high_u8 (uint8x16_t a, uint8x16_t b)
9059 {
9060   uint16x8_t result;
9061   __asm__ ("umull2 %0.8h,%1.16b,%2.16b"
9062            : "=w"(result)
9063            : "w"(a), "w"(b)
9064            : /* No clobbers */);
9065   return result;
9066 }
9067
9068 __extension__ extern __inline uint32x4_t
9069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9070 vmull_high_u16 (uint16x8_t a, uint16x8_t b)
9071 {
9072   uint32x4_t result;
9073   __asm__ ("umull2 %0.4s,%1.8h,%2.8h"
9074            : "=w"(result)
9075            : "w"(a), "w"(b)
9076            : /* No clobbers */);
9077   return result;
9078 }
9079
9080 __extension__ extern __inline uint64x2_t
9081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9082 vmull_high_u32 (uint32x4_t a, uint32x4_t b)
9083 {
9084   uint64x2_t result;
9085   __asm__ ("umull2 %0.2d,%1.4s,%2.4s"
9086            : "=w"(result)
9087            : "w"(a), "w"(b)
9088            : /* No clobbers */);
9089   return result;
9090 }
9091
9092 #define vmull_lane_s16(a, b, c)                                         \
9093   __extension__                                                         \
9094     ({                                                                  \
9095        int16x4_t b_ = (b);                                              \
9096        int16x4_t a_ = (a);                                              \
9097        int32x4_t result;                                                \
9098        __asm__ ("smull %0.4s,%1.4h,%2.h[%3]"                            \
9099                 : "=w"(result)                                          \
9100                 : "w"(a_), "x"(b_), "i"(c)                              \
9101                 : /* No clobbers */);                                   \
9102        result;                                                          \
9103      })
9104
9105 #define vmull_lane_s32(a, b, c)                                         \
9106   __extension__                                                         \
9107     ({                                                                  \
9108        int32x2_t b_ = (b);                                              \
9109        int32x2_t a_ = (a);                                              \
9110        int64x2_t result;                                                \
9111        __asm__ ("smull %0.2d,%1.2s,%2.s[%3]"                            \
9112                 : "=w"(result)                                          \
9113                 : "w"(a_), "w"(b_), "i"(c)                              \
9114                 : /* No clobbers */);                                   \
9115        result;                                                          \
9116      })
9117
9118 #define vmull_lane_u16(a, b, c)                                         \
9119   __extension__                                                         \
9120     ({                                                                  \
9121        uint16x4_t b_ = (b);                                             \
9122        uint16x4_t a_ = (a);                                             \
9123        uint32x4_t result;                                               \
9124        __asm__ ("umull %0.4s,%1.4h,%2.h[%3]"                            \
9125                 : "=w"(result)                                          \
9126                 : "w"(a_), "x"(b_), "i"(c)                              \
9127                 : /* No clobbers */);                                   \
9128        result;                                                          \
9129      })
9130
9131 #define vmull_lane_u32(a, b, c)                                         \
9132   __extension__                                                         \
9133     ({                                                                  \
9134        uint32x2_t b_ = (b);                                             \
9135        uint32x2_t a_ = (a);                                             \
9136        uint64x2_t result;                                               \
9137        __asm__ ("umull %0.2d, %1.2s, %2.s[%3]"                          \
9138                 : "=w"(result)                                          \
9139                 : "w"(a_), "w"(b_), "i"(c)                              \
9140                 : /* No clobbers */);                                   \
9141        result;                                                          \
9142      })
9143
9144 #define vmull_laneq_s16(a, b, c)                                        \
9145   __extension__                                                         \
9146     ({                                                                  \
9147        int16x8_t b_ = (b);                                              \
9148        int16x4_t a_ = (a);                                              \
9149        int32x4_t result;                                                \
9150        __asm__ ("smull %0.4s, %1.4h, %2.h[%3]"                          \
9151                 : "=w"(result)                                          \
9152                 : "w"(a_), "x"(b_), "i"(c)                              \
9153                 : /* No clobbers */);                                   \
9154        result;                                                          \
9155      })
9156
9157 #define vmull_laneq_s32(a, b, c)                                        \
9158   __extension__                                                         \
9159     ({                                                                  \
9160        int32x4_t b_ = (b);                                              \
9161        int32x2_t a_ = (a);                                              \
9162        int64x2_t result;                                                \
9163        __asm__ ("smull %0.2d, %1.2s, %2.s[%3]"                          \
9164                 : "=w"(result)                                          \
9165                 : "w"(a_), "w"(b_), "i"(c)                              \
9166                 : /* No clobbers */);                                   \
9167        result;                                                          \
9168      })
9169
9170 #define vmull_laneq_u16(a, b, c)                                        \
9171   __extension__                                                         \
9172     ({                                                                  \
9173        uint16x8_t b_ = (b);                                             \
9174        uint16x4_t a_ = (a);                                             \
9175        uint32x4_t result;                                               \
9176        __asm__ ("umull %0.4s, %1.4h, %2.h[%3]"                          \
9177                 : "=w"(result)                                          \
9178                 : "w"(a_), "x"(b_), "i"(c)                              \
9179                 : /* No clobbers */);                                   \
9180        result;                                                          \
9181      })
9182
9183 #define vmull_laneq_u32(a, b, c)                                        \
9184   __extension__                                                         \
9185     ({                                                                  \
9186        uint32x4_t b_ = (b);                                             \
9187        uint32x2_t a_ = (a);                                             \
9188        uint64x2_t result;                                               \
9189        __asm__ ("umull %0.2d, %1.2s, %2.s[%3]"                          \
9190                 : "=w"(result)                                          \
9191                 : "w"(a_), "w"(b_), "i"(c)                              \
9192                 : /* No clobbers */);                                   \
9193        result;                                                          \
9194      })
9195
9196 __extension__ extern __inline int32x4_t
9197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9198 vmull_n_s16 (int16x4_t a, int16_t b)
9199 {
9200   int32x4_t result;
9201   __asm__ ("smull %0.4s,%1.4h,%2.h[0]"
9202            : "=w"(result)
9203            : "w"(a), "x"(b)
9204            : /* No clobbers */);
9205   return result;
9206 }
9207
9208 __extension__ extern __inline int64x2_t
9209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9210 vmull_n_s32 (int32x2_t a, int32_t b)
9211 {
9212   int64x2_t result;
9213   __asm__ ("smull %0.2d,%1.2s,%2.s[0]"
9214            : "=w"(result)
9215            : "w"(a), "w"(b)
9216            : /* No clobbers */);
9217   return result;
9218 }
9219
9220 __extension__ extern __inline uint32x4_t
9221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9222 vmull_n_u16 (uint16x4_t a, uint16_t b)
9223 {
9224   uint32x4_t result;
9225   __asm__ ("umull %0.4s,%1.4h,%2.h[0]"
9226            : "=w"(result)
9227            : "w"(a), "x"(b)
9228            : /* No clobbers */);
9229   return result;
9230 }
9231
9232 __extension__ extern __inline uint64x2_t
9233 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9234 vmull_n_u32 (uint32x2_t a, uint32_t b)
9235 {
9236   uint64x2_t result;
9237   __asm__ ("umull %0.2d,%1.2s,%2.s[0]"
9238            : "=w"(result)
9239            : "w"(a), "w"(b)
9240            : /* No clobbers */);
9241   return result;
9242 }
9243
9244 __extension__ extern __inline poly16x8_t
9245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9246 vmull_p8 (poly8x8_t a, poly8x8_t b)
9247 {
9248   poly16x8_t result;
9249   __asm__ ("pmull %0.8h, %1.8b, %2.8b"
9250            : "=w"(result)
9251            : "w"(a), "w"(b)
9252            : /* No clobbers */);
9253   return result;
9254 }
9255
9256 __extension__ extern __inline int16x8_t
9257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9258 vmull_s8 (int8x8_t a, int8x8_t b)
9259 {
9260   int16x8_t result;
9261   __asm__ ("smull %0.8h, %1.8b, %2.8b"
9262            : "=w"(result)
9263            : "w"(a), "w"(b)
9264            : /* No clobbers */);
9265   return result;
9266 }
9267
9268 __extension__ extern __inline int32x4_t
9269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9270 vmull_s16 (int16x4_t a, int16x4_t b)
9271 {
9272   int32x4_t result;
9273   __asm__ ("smull %0.4s, %1.4h, %2.4h"
9274            : "=w"(result)
9275            : "w"(a), "w"(b)
9276            : /* No clobbers */);
9277   return result;
9278 }
9279
9280 __extension__ extern __inline int64x2_t
9281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9282 vmull_s32 (int32x2_t a, int32x2_t b)
9283 {
9284   int64x2_t result;
9285   __asm__ ("smull %0.2d, %1.2s, %2.2s"
9286            : "=w"(result)
9287            : "w"(a), "w"(b)
9288            : /* No clobbers */);
9289   return result;
9290 }
9291
9292 __extension__ extern __inline uint16x8_t
9293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9294 vmull_u8 (uint8x8_t a, uint8x8_t b)
9295 {
9296   uint16x8_t result;
9297   __asm__ ("umull %0.8h, %1.8b, %2.8b"
9298            : "=w"(result)
9299            : "w"(a), "w"(b)
9300            : /* No clobbers */);
9301   return result;
9302 }
9303
9304 __extension__ extern __inline uint32x4_t
9305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9306 vmull_u16 (uint16x4_t a, uint16x4_t b)
9307 {
9308   uint32x4_t result;
9309   __asm__ ("umull %0.4s, %1.4h, %2.4h"
9310            : "=w"(result)
9311            : "w"(a), "w"(b)
9312            : /* No clobbers */);
9313   return result;
9314 }
9315
9316 __extension__ extern __inline uint64x2_t
9317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9318 vmull_u32 (uint32x2_t a, uint32x2_t b)
9319 {
9320   uint64x2_t result;
9321   __asm__ ("umull %0.2d, %1.2s, %2.2s"
9322            : "=w"(result)
9323            : "w"(a), "w"(b)
9324            : /* No clobbers */);
9325   return result;
9326 }
9327
9328 __extension__ extern __inline int16x4_t
9329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9330 vpadal_s8 (int16x4_t a, int8x8_t b)
9331 {
9332   int16x4_t result;
9333   __asm__ ("sadalp %0.4h,%2.8b"
9334            : "=w"(result)
9335            : "0"(a), "w"(b)
9336            : /* No clobbers */);
9337   return result;
9338 }
9339
9340 __extension__ extern __inline int32x2_t
9341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9342 vpadal_s16 (int32x2_t a, int16x4_t b)
9343 {
9344   int32x2_t result;
9345   __asm__ ("sadalp %0.2s,%2.4h"
9346            : "=w"(result)
9347            : "0"(a), "w"(b)
9348            : /* No clobbers */);
9349   return result;
9350 }
9351
9352 __extension__ extern __inline int64x1_t
9353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9354 vpadal_s32 (int64x1_t a, int32x2_t b)
9355 {
9356   int64x1_t result;
9357   __asm__ ("sadalp %0.1d,%2.2s"
9358            : "=w"(result)
9359            : "0"(a), "w"(b)
9360            : /* No clobbers */);
9361   return result;
9362 }
9363
9364 __extension__ extern __inline uint16x4_t
9365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9366 vpadal_u8 (uint16x4_t a, uint8x8_t b)
9367 {
9368   uint16x4_t result;
9369   __asm__ ("uadalp %0.4h,%2.8b"
9370            : "=w"(result)
9371            : "0"(a), "w"(b)
9372            : /* No clobbers */);
9373   return result;
9374 }
9375
9376 __extension__ extern __inline uint32x2_t
9377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9378 vpadal_u16 (uint32x2_t a, uint16x4_t b)
9379 {
9380   uint32x2_t result;
9381   __asm__ ("uadalp %0.2s,%2.4h"
9382            : "=w"(result)
9383            : "0"(a), "w"(b)
9384            : /* No clobbers */);
9385   return result;
9386 }
9387
9388 __extension__ extern __inline uint64x1_t
9389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9390 vpadal_u32 (uint64x1_t a, uint32x2_t b)
9391 {
9392   uint64x1_t result;
9393   __asm__ ("uadalp %0.1d,%2.2s"
9394            : "=w"(result)
9395            : "0"(a), "w"(b)
9396            : /* No clobbers */);
9397   return result;
9398 }
9399
9400 __extension__ extern __inline int16x8_t
9401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9402 vpadalq_s8 (int16x8_t a, int8x16_t b)
9403 {
9404   int16x8_t result;
9405   __asm__ ("sadalp %0.8h,%2.16b"
9406            : "=w"(result)
9407            : "0"(a), "w"(b)
9408            : /* No clobbers */);
9409   return result;
9410 }
9411
9412 __extension__ extern __inline int32x4_t
9413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9414 vpadalq_s16 (int32x4_t a, int16x8_t b)
9415 {
9416   int32x4_t result;
9417   __asm__ ("sadalp %0.4s,%2.8h"
9418            : "=w"(result)
9419            : "0"(a), "w"(b)
9420            : /* No clobbers */);
9421   return result;
9422 }
9423
9424 __extension__ extern __inline int64x2_t
9425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9426 vpadalq_s32 (int64x2_t a, int32x4_t b)
9427 {
9428   int64x2_t result;
9429   __asm__ ("sadalp %0.2d,%2.4s"
9430            : "=w"(result)
9431            : "0"(a), "w"(b)
9432            : /* No clobbers */);
9433   return result;
9434 }
9435
9436 __extension__ extern __inline uint16x8_t
9437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9438 vpadalq_u8 (uint16x8_t a, uint8x16_t b)
9439 {
9440   uint16x8_t result;
9441   __asm__ ("uadalp %0.8h,%2.16b"
9442            : "=w"(result)
9443            : "0"(a), "w"(b)
9444            : /* No clobbers */);
9445   return result;
9446 }
9447
9448 __extension__ extern __inline uint32x4_t
9449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9450 vpadalq_u16 (uint32x4_t a, uint16x8_t b)
9451 {
9452   uint32x4_t result;
9453   __asm__ ("uadalp %0.4s,%2.8h"
9454            : "=w"(result)
9455            : "0"(a), "w"(b)
9456            : /* No clobbers */);
9457   return result;
9458 }
9459
9460 __extension__ extern __inline uint64x2_t
9461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9462 vpadalq_u32 (uint64x2_t a, uint32x4_t b)
9463 {
9464   uint64x2_t result;
9465   __asm__ ("uadalp %0.2d,%2.4s"
9466            : "=w"(result)
9467            : "0"(a), "w"(b)
9468            : /* No clobbers */);
9469   return result;
9470 }
9471
9472 __extension__ extern __inline int16x4_t
9473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9474 vpaddl_s8 (int8x8_t a)
9475 {
9476   int16x4_t result;
9477   __asm__ ("saddlp %0.4h,%1.8b"
9478            : "=w"(result)
9479            : "w"(a)
9480            : /* No clobbers */);
9481   return result;
9482 }
9483
9484 __extension__ extern __inline int32x2_t
9485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9486 vpaddl_s16 (int16x4_t a)
9487 {
9488   int32x2_t result;
9489   __asm__ ("saddlp %0.2s,%1.4h"
9490            : "=w"(result)
9491            : "w"(a)
9492            : /* No clobbers */);
9493   return result;
9494 }
9495
9496 __extension__ extern __inline int64x1_t
9497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9498 vpaddl_s32 (int32x2_t a)
9499 {
9500   int64x1_t result;
9501   __asm__ ("saddlp %0.1d,%1.2s"
9502            : "=w"(result)
9503            : "w"(a)
9504            : /* No clobbers */);
9505   return result;
9506 }
9507
9508 __extension__ extern __inline uint16x4_t
9509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9510 vpaddl_u8 (uint8x8_t a)
9511 {
9512   uint16x4_t result;
9513   __asm__ ("uaddlp %0.4h,%1.8b"
9514            : "=w"(result)
9515            : "w"(a)
9516            : /* No clobbers */);
9517   return result;
9518 }
9519
9520 __extension__ extern __inline uint32x2_t
9521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9522 vpaddl_u16 (uint16x4_t a)
9523 {
9524   uint32x2_t result;
9525   __asm__ ("uaddlp %0.2s,%1.4h"
9526            : "=w"(result)
9527            : "w"(a)
9528            : /* No clobbers */);
9529   return result;
9530 }
9531
9532 __extension__ extern __inline uint64x1_t
9533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9534 vpaddl_u32 (uint32x2_t a)
9535 {
9536   uint64x1_t result;
9537   __asm__ ("uaddlp %0.1d,%1.2s"
9538            : "=w"(result)
9539            : "w"(a)
9540            : /* No clobbers */);
9541   return result;
9542 }
9543
9544 __extension__ extern __inline int16x8_t
9545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9546 vpaddlq_s8 (int8x16_t a)
9547 {
9548   int16x8_t result;
9549   __asm__ ("saddlp %0.8h,%1.16b"
9550            : "=w"(result)
9551            : "w"(a)
9552            : /* No clobbers */);
9553   return result;
9554 }
9555
9556 __extension__ extern __inline int32x4_t
9557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9558 vpaddlq_s16 (int16x8_t a)
9559 {
9560   int32x4_t result;
9561   __asm__ ("saddlp %0.4s,%1.8h"
9562            : "=w"(result)
9563            : "w"(a)
9564            : /* No clobbers */);
9565   return result;
9566 }
9567
9568 __extension__ extern __inline int64x2_t
9569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9570 vpaddlq_s32 (int32x4_t a)
9571 {
9572   int64x2_t result;
9573   __asm__ ("saddlp %0.2d,%1.4s"
9574            : "=w"(result)
9575            : "w"(a)
9576            : /* No clobbers */);
9577   return result;
9578 }
9579
9580 __extension__ extern __inline uint16x8_t
9581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9582 vpaddlq_u8 (uint8x16_t a)
9583 {
9584   uint16x8_t result;
9585   __asm__ ("uaddlp %0.8h,%1.16b"
9586            : "=w"(result)
9587            : "w"(a)
9588            : /* No clobbers */);
9589   return result;
9590 }
9591
9592 __extension__ extern __inline uint32x4_t
9593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9594 vpaddlq_u16 (uint16x8_t a)
9595 {
9596   uint32x4_t result;
9597   __asm__ ("uaddlp %0.4s,%1.8h"
9598            : "=w"(result)
9599            : "w"(a)
9600            : /* No clobbers */);
9601   return result;
9602 }
9603
9604 __extension__ extern __inline uint64x2_t
9605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9606 vpaddlq_u32 (uint32x4_t a)
9607 {
9608   uint64x2_t result;
9609   __asm__ ("uaddlp %0.2d,%1.4s"
9610            : "=w"(result)
9611            : "w"(a)
9612            : /* No clobbers */);
9613   return result;
9614 }
9615
9616 __extension__ extern __inline int8x16_t
9617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9618 vpaddq_s8 (int8x16_t a, int8x16_t b)
9619 {
9620   int8x16_t result;
9621   __asm__ ("addp %0.16b,%1.16b,%2.16b"
9622            : "=w"(result)
9623            : "w"(a), "w"(b)
9624            : /* No clobbers */);
9625   return result;
9626 }
9627
9628 __extension__ extern __inline int16x8_t
9629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9630 vpaddq_s16 (int16x8_t a, int16x8_t b)
9631 {
9632   int16x8_t result;
9633   __asm__ ("addp %0.8h,%1.8h,%2.8h"
9634            : "=w"(result)
9635            : "w"(a), "w"(b)
9636            : /* No clobbers */);
9637   return result;
9638 }
9639
9640 __extension__ extern __inline int32x4_t
9641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9642 vpaddq_s32 (int32x4_t a, int32x4_t b)
9643 {
9644   int32x4_t result;
9645   __asm__ ("addp %0.4s,%1.4s,%2.4s"
9646            : "=w"(result)
9647            : "w"(a), "w"(b)
9648            : /* No clobbers */);
9649   return result;
9650 }
9651
9652 __extension__ extern __inline int64x2_t
9653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9654 vpaddq_s64 (int64x2_t a, int64x2_t b)
9655 {
9656   int64x2_t result;
9657   __asm__ ("addp %0.2d,%1.2d,%2.2d"
9658            : "=w"(result)
9659            : "w"(a), "w"(b)
9660            : /* No clobbers */);
9661   return result;
9662 }
9663
9664 __extension__ extern __inline uint8x16_t
9665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9666 vpaddq_u8 (uint8x16_t a, uint8x16_t b)
9667 {
9668   uint8x16_t result;
9669   __asm__ ("addp %0.16b,%1.16b,%2.16b"
9670            : "=w"(result)
9671            : "w"(a), "w"(b)
9672            : /* No clobbers */);
9673   return result;
9674 }
9675
9676 __extension__ extern __inline uint16x8_t
9677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9678 vpaddq_u16 (uint16x8_t a, uint16x8_t b)
9679 {
9680   uint16x8_t result;
9681   __asm__ ("addp %0.8h,%1.8h,%2.8h"
9682            : "=w"(result)
9683            : "w"(a), "w"(b)
9684            : /* No clobbers */);
9685   return result;
9686 }
9687
9688 __extension__ extern __inline uint32x4_t
9689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9690 vpaddq_u32 (uint32x4_t a, uint32x4_t b)
9691 {
9692   uint32x4_t result;
9693   __asm__ ("addp %0.4s,%1.4s,%2.4s"
9694            : "=w"(result)
9695            : "w"(a), "w"(b)
9696            : /* No clobbers */);
9697   return result;
9698 }
9699
9700 __extension__ extern __inline uint64x2_t
9701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9702 vpaddq_u64 (uint64x2_t a, uint64x2_t b)
9703 {
9704   uint64x2_t result;
9705   __asm__ ("addp %0.2d,%1.2d,%2.2d"
9706            : "=w"(result)
9707            : "w"(a), "w"(b)
9708            : /* No clobbers */);
9709   return result;
9710 }
9711
9712 __extension__ extern __inline int16x4_t
9713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9714 vqdmulh_n_s16 (int16x4_t a, int16_t b)
9715 {
9716   int16x4_t result;
9717   __asm__ ("sqdmulh %0.4h,%1.4h,%2.h[0]"
9718            : "=w"(result)
9719            : "w"(a), "x"(b)
9720            : /* No clobbers */);
9721   return result;
9722 }
9723
9724 __extension__ extern __inline int32x2_t
9725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9726 vqdmulh_n_s32 (int32x2_t a, int32_t b)
9727 {
9728   int32x2_t result;
9729   __asm__ ("sqdmulh %0.2s,%1.2s,%2.s[0]"
9730            : "=w"(result)
9731            : "w"(a), "w"(b)
9732            : /* No clobbers */);
9733   return result;
9734 }
9735
9736 __extension__ extern __inline int16x8_t
9737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9738 vqdmulhq_n_s16 (int16x8_t a, int16_t b)
9739 {
9740   int16x8_t result;
9741   __asm__ ("sqdmulh %0.8h,%1.8h,%2.h[0]"
9742            : "=w"(result)
9743            : "w"(a), "x"(b)
9744            : /* No clobbers */);
9745   return result;
9746 }
9747
9748 __extension__ extern __inline int32x4_t
9749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9750 vqdmulhq_n_s32 (int32x4_t a, int32_t b)
9751 {
9752   int32x4_t result;
9753   __asm__ ("sqdmulh %0.4s,%1.4s,%2.s[0]"
9754            : "=w"(result)
9755            : "w"(a), "w"(b)
9756            : /* No clobbers */);
9757   return result;
9758 }
9759
9760 __extension__ extern __inline int8x16_t
9761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9762 vqmovn_high_s16 (int8x8_t a, int16x8_t b)
9763 {
9764   int8x16_t result = vcombine_s8 (a, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
9765   __asm__ ("sqxtn2 %0.16b, %1.8h"
9766            : "+w"(result)
9767            : "w"(b)
9768            : /* No clobbers */);
9769   return result;
9770 }
9771
9772 __extension__ extern __inline int16x8_t
9773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9774 vqmovn_high_s32 (int16x4_t a, int32x4_t b)
9775 {
9776   int16x8_t result = vcombine_s16 (a, vcreate_s16 (__AARCH64_UINT64_C (0x0)));
9777   __asm__ ("sqxtn2 %0.8h, %1.4s"
9778            : "+w"(result)
9779            : "w"(b)
9780            : /* No clobbers */);
9781   return result;
9782 }
9783
9784 __extension__ extern __inline int32x4_t
9785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9786 vqmovn_high_s64 (int32x2_t a, int64x2_t b)
9787 {
9788   int32x4_t result = vcombine_s32 (a, vcreate_s32 (__AARCH64_UINT64_C (0x0)));
9789   __asm__ ("sqxtn2 %0.4s, %1.2d"
9790            : "+w"(result)
9791            : "w"(b)
9792            : /* No clobbers */);
9793   return result;
9794 }
9795
9796 __extension__ extern __inline uint8x16_t
9797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9798 vqmovn_high_u16 (uint8x8_t a, uint16x8_t b)
9799 {
9800   uint8x16_t result = vcombine_u8 (a, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
9801   __asm__ ("uqxtn2 %0.16b, %1.8h"
9802            : "+w"(result)
9803            : "w"(b)
9804            : /* No clobbers */);
9805   return result;
9806 }
9807
9808 __extension__ extern __inline uint16x8_t
9809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9810 vqmovn_high_u32 (uint16x4_t a, uint32x4_t b)
9811 {
9812   uint16x8_t result = vcombine_u16 (a, vcreate_u16 (__AARCH64_UINT64_C (0x0)));
9813   __asm__ ("uqxtn2 %0.8h, %1.4s"
9814            : "+w"(result)
9815            : "w"(b)
9816            : /* No clobbers */);
9817   return result;
9818 }
9819
9820 __extension__ extern __inline uint32x4_t
9821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9822 vqmovn_high_u64 (uint32x2_t a, uint64x2_t b)
9823 {
9824   uint32x4_t result = vcombine_u32 (a, vcreate_u32 (__AARCH64_UINT64_C (0x0)));
9825   __asm__ ("uqxtn2 %0.4s, %1.2d"
9826            : "+w"(result)
9827            : "w"(b)
9828            : /* No clobbers */);
9829   return result;
9830 }
9831
9832 __extension__ extern __inline uint8x16_t
9833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9834 vqmovun_high_s16 (uint8x8_t a, int16x8_t b)
9835 {
9836   uint8x16_t result = vcombine_u8 (a, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
9837   __asm__ ("sqxtun2 %0.16b, %1.8h"
9838            : "+w"(result)
9839            : "w"(b)
9840            : /* No clobbers */);
9841   return result;
9842 }
9843
9844 __extension__ extern __inline uint16x8_t
9845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9846 vqmovun_high_s32 (uint16x4_t a, int32x4_t b)
9847 {
9848   uint16x8_t result = vcombine_u16 (a, vcreate_u16 (__AARCH64_UINT64_C (0x0)));
9849   __asm__ ("sqxtun2 %0.8h, %1.4s"
9850            : "+w"(result)
9851            : "w"(b)
9852            : /* No clobbers */);
9853   return result;
9854 }
9855
9856 __extension__ extern __inline uint32x4_t
9857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9858 vqmovun_high_s64 (uint32x2_t a, int64x2_t b)
9859 {
9860   uint32x4_t result = vcombine_u32 (a, vcreate_u32 (__AARCH64_UINT64_C (0x0)));
9861   __asm__ ("sqxtun2 %0.4s, %1.2d"
9862            : "+w"(result)
9863            : "w"(b)
9864            : /* No clobbers */);
9865   return result;
9866 }
9867
9868 __extension__ extern __inline int16x4_t
9869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9870 vqrdmulh_n_s16 (int16x4_t a, int16_t b)
9871 {
9872   int16x4_t result;
9873   __asm__ ("sqrdmulh %0.4h,%1.4h,%2.h[0]"
9874            : "=w"(result)
9875            : "w"(a), "x"(b)
9876            : /* No clobbers */);
9877   return result;
9878 }
9879
9880 __extension__ extern __inline int32x2_t
9881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9882 vqrdmulh_n_s32 (int32x2_t a, int32_t b)
9883 {
9884   int32x2_t result;
9885   __asm__ ("sqrdmulh %0.2s,%1.2s,%2.s[0]"
9886            : "=w"(result)
9887            : "w"(a), "w"(b)
9888            : /* No clobbers */);
9889   return result;
9890 }
9891
9892 __extension__ extern __inline int16x8_t
9893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9894 vqrdmulhq_n_s16 (int16x8_t a, int16_t b)
9895 {
9896   int16x8_t result;
9897   __asm__ ("sqrdmulh %0.8h,%1.8h,%2.h[0]"
9898            : "=w"(result)
9899            : "w"(a), "x"(b)
9900            : /* No clobbers */);
9901   return result;
9902 }
9903
9904 __extension__ extern __inline int32x4_t
9905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
9906 vqrdmulhq_n_s32 (int32x4_t a, int32_t b)
9907 {
9908   int32x4_t result;
9909   __asm__ ("sqrdmulh %0.4s,%1.4s,%2.s[0]"
9910            : "=w"(result)
9911            : "w"(a), "w"(b)
9912            : /* No clobbers */);
9913   return result;
9914 }
9915
9916 #define vqrshrn_high_n_s16(a, b, c)                                     \
9917   __extension__                                                         \
9918     ({                                                                  \
9919        int16x8_t b_ = (b);                                              \
9920        int8x8_t a_ = (a);                                               \
9921        int8x16_t result = vcombine_s8                                   \
9922                             (a_, vcreate_s8                             \
9923                                    (__AARCH64_UINT64_C (0x0)));         \
9924        __asm__ ("sqrshrn2 %0.16b, %1.8h, #%2"                           \
9925                 : "+w"(result)                                          \
9926                 : "w"(b_), "i"(c)                                       \
9927                 : /* No clobbers */);                                   \
9928        result;                                                          \
9929      })
9930
9931 #define vqrshrn_high_n_s32(a, b, c)                                     \
9932   __extension__                                                         \
9933     ({                                                                  \
9934        int32x4_t b_ = (b);                                              \
9935        int16x4_t a_ = (a);                                              \
9936        int16x8_t result = vcombine_s16                                  \
9937                             (a_, vcreate_s16                            \
9938                                    (__AARCH64_UINT64_C (0x0)));         \
9939        __asm__ ("sqrshrn2 %0.8h, %1.4s, #%2"                            \
9940                 : "+w"(result)                                          \
9941                 : "w"(b_), "i"(c)                                       \
9942                 : /* No clobbers */);                                   \
9943        result;                                                          \
9944      })
9945
9946 #define vqrshrn_high_n_s64(a, b, c)                                     \
9947   __extension__                                                         \
9948     ({                                                                  \
9949        int64x2_t b_ = (b);                                              \
9950        int32x2_t a_ = (a);                                              \
9951        int32x4_t result = vcombine_s32                                  \
9952                             (a_, vcreate_s32                            \
9953                                    (__AARCH64_UINT64_C (0x0)));         \
9954        __asm__ ("sqrshrn2 %0.4s, %1.2d, #%2"                            \
9955                 : "+w"(result)                                          \
9956                 : "w"(b_), "i"(c)                                       \
9957                 : /* No clobbers */);                                   \
9958        result;                                                          \
9959      })
9960
9961 #define vqrshrn_high_n_u16(a, b, c)                                     \
9962   __extension__                                                         \
9963     ({                                                                  \
9964        uint16x8_t b_ = (b);                                             \
9965        uint8x8_t a_ = (a);                                              \
9966        uint8x16_t result = vcombine_u8                                  \
9967                              (a_, vcreate_u8                            \
9968                                     (__AARCH64_UINT64_C (0x0)));        \
9969        __asm__ ("uqrshrn2 %0.16b, %1.8h, #%2"                           \
9970                 : "+w"(result)                                          \
9971                 : "w"(b_), "i"(c)                                       \
9972                 : /* No clobbers */);                                   \
9973        result;                                                          \
9974      })
9975
9976 #define vqrshrn_high_n_u32(a, b, c)                                     \
9977   __extension__                                                         \
9978     ({                                                                  \
9979        uint32x4_t b_ = (b);                                             \
9980        uint16x4_t a_ = (a);                                             \
9981        uint16x8_t result = vcombine_u16                                 \
9982                              (a_, vcreate_u16                           \
9983                                     (__AARCH64_UINT64_C (0x0)));        \
9984        __asm__ ("uqrshrn2 %0.8h, %1.4s, #%2"                            \
9985                 : "+w"(result)                                          \
9986                 : "w"(b_), "i"(c)                                       \
9987                 : /* No clobbers */);                                   \
9988        result;                                                          \
9989      })
9990
9991 #define vqrshrn_high_n_u64(a, b, c)                                     \
9992   __extension__                                                         \
9993     ({                                                                  \
9994        uint64x2_t b_ = (b);                                             \
9995        uint32x2_t a_ = (a);                                             \
9996        uint32x4_t result = vcombine_u32                                 \
9997                              (a_, vcreate_u32                           \
9998                                     (__AARCH64_UINT64_C (0x0)));        \
9999        __asm__ ("uqrshrn2 %0.4s, %1.2d, #%2"                            \
10000                 : "+w"(result)                                          \
10001                 : "w"(b_), "i"(c)                                       \
10002                 : /* No clobbers */);                                   \
10003        result;                                                          \
10004      })
10005
10006 #define vqrshrun_high_n_s16(a, b, c)                                    \
10007   __extension__                                                         \
10008     ({                                                                  \
10009        int16x8_t b_ = (b);                                              \
10010        uint8x8_t a_ = (a);                                              \
10011        uint8x16_t result = vcombine_u8                                  \
10012                              (a_, vcreate_u8                            \
10013                                     (__AARCH64_UINT64_C (0x0)));        \
10014        __asm__ ("sqrshrun2 %0.16b, %1.8h, #%2"                          \
10015                 : "+w"(result)                                          \
10016                 : "w"(b_), "i"(c)                                       \
10017                 : /* No clobbers */);                                   \
10018        result;                                                          \
10019      })
10020
10021 #define vqrshrun_high_n_s32(a, b, c)                                    \
10022   __extension__                                                         \
10023     ({                                                                  \
10024        int32x4_t b_ = (b);                                              \
10025        uint16x4_t a_ = (a);                                             \
10026        uint16x8_t result = vcombine_u16                                 \
10027                              (a_, vcreate_u16                           \
10028                                     (__AARCH64_UINT64_C (0x0)));        \
10029        __asm__ ("sqrshrun2 %0.8h, %1.4s, #%2"                           \
10030                 : "+w"(result)                                          \
10031                 : "w"(b_), "i"(c)                                       \
10032                 : /* No clobbers */);                                   \
10033        result;                                                          \
10034      })
10035
10036 #define vqrshrun_high_n_s64(a, b, c)                                    \
10037   __extension__                                                         \
10038     ({                                                                  \
10039        int64x2_t b_ = (b);                                              \
10040        uint32x2_t a_ = (a);                                             \
10041        uint32x4_t result = vcombine_u32                                 \
10042                              (a_, vcreate_u32                           \
10043                                     (__AARCH64_UINT64_C (0x0)));        \
10044        __asm__ ("sqrshrun2 %0.4s, %1.2d, #%2"                           \
10045                 : "+w"(result)                                          \
10046                 : "w"(b_), "i"(c)                                       \
10047                 : /* No clobbers */);                                   \
10048        result;                                                          \
10049      })
10050
10051 #define vqshrn_high_n_s16(a, b, c)                                      \
10052   __extension__                                                         \
10053     ({                                                                  \
10054        int16x8_t b_ = (b);                                              \
10055        int8x8_t a_ = (a);                                               \
10056        int8x16_t result = vcombine_s8                                   \
10057                             (a_, vcreate_s8                             \
10058                                    (__AARCH64_UINT64_C (0x0)));         \
10059        __asm__ ("sqshrn2 %0.16b, %1.8h, #%2"                            \
10060                 : "+w"(result)                                          \
10061                 : "w"(b_), "i"(c)                                       \
10062                 : /* No clobbers */);                                   \
10063        result;                                                          \
10064      })
10065
10066 #define vqshrn_high_n_s32(a, b, c)                                      \
10067   __extension__                                                         \
10068     ({                                                                  \
10069        int32x4_t b_ = (b);                                              \
10070        int16x4_t a_ = (a);                                              \
10071        int16x8_t result = vcombine_s16                                  \
10072                             (a_, vcreate_s16                            \
10073                                    (__AARCH64_UINT64_C (0x0)));         \
10074        __asm__ ("sqshrn2 %0.8h, %1.4s, #%2"                             \
10075                 : "+w"(result)                                          \
10076                 : "w"(b_), "i"(c)                                       \
10077                 : /* No clobbers */);                                   \
10078        result;                                                          \
10079      })
10080
10081 #define vqshrn_high_n_s64(a, b, c)                                      \
10082   __extension__                                                         \
10083     ({                                                                  \
10084        int64x2_t b_ = (b);                                              \
10085        int32x2_t a_ = (a);                                              \
10086        int32x4_t result = vcombine_s32                                  \
10087                             (a_, vcreate_s32                            \
10088                                    (__AARCH64_UINT64_C (0x0)));         \
10089        __asm__ ("sqshrn2 %0.4s, %1.2d, #%2"                             \
10090                 : "+w"(result)                                          \
10091                 : "w"(b_), "i"(c)                                       \
10092                 : /* No clobbers */);                                   \
10093        result;                                                          \
10094      })
10095
10096 #define vqshrn_high_n_u16(a, b, c)                                      \
10097   __extension__                                                         \
10098     ({                                                                  \
10099        uint16x8_t b_ = (b);                                             \
10100        uint8x8_t a_ = (a);                                              \
10101        uint8x16_t result = vcombine_u8                                  \
10102                              (a_, vcreate_u8                            \
10103                                     (__AARCH64_UINT64_C (0x0)));        \
10104        __asm__ ("uqshrn2 %0.16b, %1.8h, #%2"                            \
10105                 : "+w"(result)                                          \
10106                 : "w"(b_), "i"(c)                                       \
10107                 : /* No clobbers */);                                   \
10108        result;                                                          \
10109      })
10110
10111 #define vqshrn_high_n_u32(a, b, c)                                      \
10112   __extension__                                                         \
10113     ({                                                                  \
10114        uint32x4_t b_ = (b);                                             \
10115        uint16x4_t a_ = (a);                                             \
10116        uint16x8_t result = vcombine_u16                                 \
10117                              (a_, vcreate_u16                           \
10118                                     (__AARCH64_UINT64_C (0x0)));        \
10119        __asm__ ("uqshrn2 %0.8h, %1.4s, #%2"                             \
10120                 : "+w"(result)                                          \
10121                 : "w"(b_), "i"(c)                                       \
10122                 : /* No clobbers */);                                   \
10123        result;                                                          \
10124      })
10125
10126 #define vqshrn_high_n_u64(a, b, c)                                      \
10127   __extension__                                                         \
10128     ({                                                                  \
10129        uint64x2_t b_ = (b);                                             \
10130        uint32x2_t a_ = (a);                                             \
10131        uint32x4_t result = vcombine_u32                                 \
10132                              (a_, vcreate_u32                           \
10133                                     (__AARCH64_UINT64_C (0x0)));        \
10134        __asm__ ("uqshrn2 %0.4s, %1.2d, #%2"                             \
10135                 : "+w"(result)                                          \
10136                 : "w"(b_), "i"(c)                                       \
10137                 : /* No clobbers */);                                   \
10138        result;                                                          \
10139      })
10140
10141 #define vqshrun_high_n_s16(a, b, c)                                     \
10142   __extension__                                                         \
10143     ({                                                                  \
10144        int16x8_t b_ = (b);                                              \
10145        uint8x8_t a_ = (a);                                              \
10146        uint8x16_t result = vcombine_u8                                  \
10147                              (a_, vcreate_u8                            \
10148                                     (__AARCH64_UINT64_C (0x0)));        \
10149        __asm__ ("sqshrun2 %0.16b, %1.8h, #%2"                           \
10150                 : "+w"(result)                                          \
10151                 : "w"(b_), "i"(c)                                       \
10152                 : /* No clobbers */);                                   \
10153        result;                                                          \
10154      })
10155
10156 #define vqshrun_high_n_s32(a, b, c)                                     \
10157   __extension__                                                         \
10158     ({                                                                  \
10159        int32x4_t b_ = (b);                                              \
10160        uint16x4_t a_ = (a);                                             \
10161        uint16x8_t result = vcombine_u16                                 \
10162                              (a_, vcreate_u16                           \
10163                                     (__AARCH64_UINT64_C (0x0)));        \
10164        __asm__ ("sqshrun2 %0.8h, %1.4s, #%2"                            \
10165                 : "+w"(result)                                          \
10166                 : "w"(b_), "i"(c)                                       \
10167                 : /* No clobbers */);                                   \
10168        result;                                                          \
10169      })
10170
10171 #define vqshrun_high_n_s64(a, b, c)                                     \
10172   __extension__                                                         \
10173     ({                                                                  \
10174        int64x2_t b_ = (b);                                              \
10175        uint32x2_t a_ = (a);                                             \
10176        uint32x4_t result = vcombine_u32                                 \
10177                              (a_, vcreate_u32                           \
10178                                     (__AARCH64_UINT64_C (0x0)));        \
10179        __asm__ ("sqshrun2 %0.4s, %1.2d, #%2"                            \
10180                 : "+w"(result)                                          \
10181                 : "w"(b_), "i"(c)                                       \
10182                 : /* No clobbers */);                                   \
10183        result;                                                          \
10184      })
10185
10186 #define vrshrn_high_n_s16(a, b, c)                                      \
10187   __extension__                                                         \
10188     ({                                                                  \
10189        int16x8_t b_ = (b);                                              \
10190        int8x8_t a_ = (a);                                               \
10191        int8x16_t result = vcombine_s8                                   \
10192                             (a_, vcreate_s8                             \
10193                                    (__AARCH64_UINT64_C (0x0)));         \
10194        __asm__ ("rshrn2 %0.16b,%1.8h,#%2"                               \
10195                 : "+w"(result)                                          \
10196                 : "w"(b_), "i"(c)                                       \
10197                 : /* No clobbers */);                                   \
10198        result;                                                          \
10199      })
10200
10201 #define vrshrn_high_n_s32(a, b, c)                                      \
10202   __extension__                                                         \
10203     ({                                                                  \
10204        int32x4_t b_ = (b);                                              \
10205        int16x4_t a_ = (a);                                              \
10206        int16x8_t result = vcombine_s16                                  \
10207                             (a_, vcreate_s16                            \
10208                                    (__AARCH64_UINT64_C (0x0)));         \
10209        __asm__ ("rshrn2 %0.8h,%1.4s,#%2"                                \
10210                 : "+w"(result)                                          \
10211                 : "w"(b_), "i"(c)                                       \
10212                 : /* No clobbers */);                                   \
10213        result;                                                          \
10214      })
10215
10216 #define vrshrn_high_n_s64(a, b, c)                                      \
10217   __extension__                                                         \
10218     ({                                                                  \
10219        int64x2_t b_ = (b);                                              \
10220        int32x2_t a_ = (a);                                              \
10221        int32x4_t result = vcombine_s32                                  \
10222                             (a_, vcreate_s32                            \
10223                                    (__AARCH64_UINT64_C (0x0)));         \
10224        __asm__ ("rshrn2 %0.4s,%1.2d,#%2"                                \
10225                 : "+w"(result)                                          \
10226                 : "w"(b_), "i"(c)                                       \
10227                 : /* No clobbers */);                                   \
10228        result;                                                          \
10229      })
10230
10231 #define vrshrn_high_n_u16(a, b, c)                                      \
10232   __extension__                                                         \
10233     ({                                                                  \
10234        uint16x8_t b_ = (b);                                             \
10235        uint8x8_t a_ = (a);                                              \
10236        uint8x16_t result = vcombine_u8                                  \
10237                             (a_, vcreate_u8                             \
10238                                    (__AARCH64_UINT64_C (0x0)));         \
10239        __asm__ ("rshrn2 %0.16b,%1.8h,#%2"                               \
10240                 : "+w"(result)                                          \
10241                 : "w"(b_), "i"(c)                                       \
10242                 : /* No clobbers */);                                   \
10243        result;                                                          \
10244      })
10245
10246 #define vrshrn_high_n_u32(a, b, c)                                      \
10247   __extension__                                                         \
10248     ({                                                                  \
10249        uint32x4_t b_ = (b);                                             \
10250        uint16x4_t a_ = (a);                                             \
10251        uint16x8_t result = vcombine_u16                                 \
10252                             (a_, vcreate_u16                            \
10253                                    (__AARCH64_UINT64_C (0x0)));         \
10254        __asm__ ("rshrn2 %0.8h,%1.4s,#%2"                                \
10255                 : "+w"(result)                                          \
10256                 : "w"(b_), "i"(c)                                       \
10257                 : /* No clobbers */);                                   \
10258        result;                                                          \
10259      })
10260
10261 #define vrshrn_high_n_u64(a, b, c)                                      \
10262   __extension__                                                         \
10263     ({                                                                  \
10264        uint64x2_t b_ = (b);                                             \
10265        uint32x2_t a_ = (a);                                             \
10266        uint32x4_t result = vcombine_u32                                 \
10267                             (a_, vcreate_u32                            \
10268                                    (__AARCH64_UINT64_C (0x0)));         \
10269        __asm__ ("rshrn2 %0.4s,%1.2d,#%2"                                \
10270                 : "+w"(result)                                          \
10271                 : "w"(b_), "i"(c)                                       \
10272                 : /* No clobbers */);                                   \
10273        result;                                                          \
10274      })
10275
10276 #define vrshrn_n_s16(a, b)                                              \
10277   __extension__                                                         \
10278     ({                                                                  \
10279        int16x8_t a_ = (a);                                              \
10280        int8x8_t result;                                                 \
10281        __asm__ ("rshrn %0.8b,%1.8h,%2"                                  \
10282                 : "=w"(result)                                          \
10283                 : "w"(a_), "i"(b)                                       \
10284                 : /* No clobbers */);                                   \
10285        result;                                                          \
10286      })
10287
10288 #define vrshrn_n_s32(a, b)                                              \
10289   __extension__                                                         \
10290     ({                                                                  \
10291        int32x4_t a_ = (a);                                              \
10292        int16x4_t result;                                                \
10293        __asm__ ("rshrn %0.4h,%1.4s,%2"                                  \
10294                 : "=w"(result)                                          \
10295                 : "w"(a_), "i"(b)                                       \
10296                 : /* No clobbers */);                                   \
10297        result;                                                          \
10298      })
10299
10300 #define vrshrn_n_s64(a, b)                                              \
10301   __extension__                                                         \
10302     ({                                                                  \
10303        int64x2_t a_ = (a);                                              \
10304        int32x2_t result;                                                \
10305        __asm__ ("rshrn %0.2s,%1.2d,%2"                                  \
10306                 : "=w"(result)                                          \
10307                 : "w"(a_), "i"(b)                                       \
10308                 : /* No clobbers */);                                   \
10309        result;                                                          \
10310      })
10311
10312 #define vrshrn_n_u16(a, b)                                              \
10313   __extension__                                                         \
10314     ({                                                                  \
10315        uint16x8_t a_ = (a);                                             \
10316        uint8x8_t result;                                                \
10317        __asm__ ("rshrn %0.8b,%1.8h,%2"                                  \
10318                 : "=w"(result)                                          \
10319                 : "w"(a_), "i"(b)                                       \
10320                 : /* No clobbers */);                                   \
10321        result;                                                          \
10322      })
10323
10324 #define vrshrn_n_u32(a, b)                                              \
10325   __extension__                                                         \
10326     ({                                                                  \
10327        uint32x4_t a_ = (a);                                             \
10328        uint16x4_t result;                                               \
10329        __asm__ ("rshrn %0.4h,%1.4s,%2"                                  \
10330                 : "=w"(result)                                          \
10331                 : "w"(a_), "i"(b)                                       \
10332                 : /* No clobbers */);                                   \
10333        result;                                                          \
10334      })
10335
10336 #define vrshrn_n_u64(a, b)                                              \
10337   __extension__                                                         \
10338     ({                                                                  \
10339        uint64x2_t a_ = (a);                                             \
10340        uint32x2_t result;                                               \
10341        __asm__ ("rshrn %0.2s,%1.2d,%2"                                  \
10342                 : "=w"(result)                                          \
10343                 : "w"(a_), "i"(b)                                       \
10344                 : /* No clobbers */);                                   \
10345        result;                                                          \
10346      })
10347
10348 __extension__ extern __inline uint32x2_t
10349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10350 vrsqrte_u32 (uint32x2_t a)
10351 {
10352   uint32x2_t result;
10353   __asm__ ("ursqrte %0.2s,%1.2s"
10354            : "=w"(result)
10355            : "w"(a)
10356            : /* No clobbers */);
10357   return result;
10358 }
10359
10360 __extension__ extern __inline uint32x4_t
10361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10362 vrsqrteq_u32 (uint32x4_t a)
10363 {
10364   uint32x4_t result;
10365   __asm__ ("ursqrte %0.4s,%1.4s"
10366            : "=w"(result)
10367            : "w"(a)
10368            : /* No clobbers */);
10369   return result;
10370 }
10371
10372 #define vshrn_high_n_s16(a, b, c)                                       \
10373   __extension__                                                         \
10374     ({                                                                  \
10375        int16x8_t b_ = (b);                                              \
10376        int8x8_t a_ = (a);                                               \
10377        int8x16_t result = vcombine_s8                                   \
10378                             (a_, vcreate_s8                             \
10379                                    (__AARCH64_UINT64_C (0x0)));         \
10380        __asm__ ("shrn2 %0.16b,%1.8h,#%2"                                \
10381                 : "+w"(result)                                          \
10382                 : "w"(b_), "i"(c)                                       \
10383                 : /* No clobbers */);                                   \
10384        result;                                                          \
10385      })
10386
10387 #define vshrn_high_n_s32(a, b, c)                                       \
10388   __extension__                                                         \
10389     ({                                                                  \
10390        int32x4_t b_ = (b);                                              \
10391        int16x4_t a_ = (a);                                              \
10392        int16x8_t result = vcombine_s16                                  \
10393                             (a_, vcreate_s16                            \
10394                                    (__AARCH64_UINT64_C (0x0)));         \
10395        __asm__ ("shrn2 %0.8h,%1.4s,#%2"                                 \
10396                 : "+w"(result)                                          \
10397                 : "w"(b_), "i"(c)                                       \
10398                 : /* No clobbers */);                                   \
10399        result;                                                          \
10400      })
10401
10402 #define vshrn_high_n_s64(a, b, c)                                       \
10403   __extension__                                                         \
10404     ({                                                                  \
10405        int64x2_t b_ = (b);                                              \
10406        int32x2_t a_ = (a);                                              \
10407        int32x4_t result = vcombine_s32                                  \
10408                             (a_, vcreate_s32                            \
10409                                    (__AARCH64_UINT64_C (0x0)));         \
10410        __asm__ ("shrn2 %0.4s,%1.2d,#%2"                                 \
10411                 : "+w"(result)                                          \
10412                 : "w"(b_), "i"(c)                                       \
10413                 : /* No clobbers */);                                   \
10414        result;                                                          \
10415      })
10416
10417 #define vshrn_high_n_u16(a, b, c)                                       \
10418   __extension__                                                         \
10419     ({                                                                  \
10420        uint16x8_t b_ = (b);                                             \
10421        uint8x8_t a_ = (a);                                              \
10422        uint8x16_t result = vcombine_u8                                  \
10423                             (a_, vcreate_u8                             \
10424                                    (__AARCH64_UINT64_C (0x0)));         \
10425        __asm__ ("shrn2 %0.16b,%1.8h,#%2"                                \
10426                 : "+w"(result)                                          \
10427                 : "w"(b_), "i"(c)                                       \
10428                 : /* No clobbers */);                                   \
10429        result;                                                          \
10430      })
10431
10432 #define vshrn_high_n_u32(a, b, c)                                       \
10433   __extension__                                                         \
10434     ({                                                                  \
10435        uint32x4_t b_ = (b);                                             \
10436        uint16x4_t a_ = (a);                                             \
10437        uint16x8_t result = vcombine_u16                                 \
10438                             (a_, vcreate_u16                            \
10439                                    (__AARCH64_UINT64_C (0x0)));         \
10440        __asm__ ("shrn2 %0.8h,%1.4s,#%2"                                 \
10441                 : "+w"(result)                                          \
10442                 : "w"(b_), "i"(c)                                       \
10443                 : /* No clobbers */);                                   \
10444        result;                                                          \
10445      })
10446
10447 #define vshrn_high_n_u64(a, b, c)                                       \
10448   __extension__                                                         \
10449     ({                                                                  \
10450        uint64x2_t b_ = (b);                                             \
10451        uint32x2_t a_ = (a);                                             \
10452        uint32x4_t result = vcombine_u32                                 \
10453                             (a_, vcreate_u32                            \
10454                                    (__AARCH64_UINT64_C (0x0)));         \
10455        __asm__ ("shrn2 %0.4s,%1.2d,#%2"                                 \
10456                 : "+w"(result)                                          \
10457                 : "w"(b_), "i"(c)                                       \
10458                 : /* No clobbers */);                                   \
10459        result;                                                          \
10460      })
10461
10462 #define vshrn_n_s16(a, b)                                               \
10463   __extension__                                                         \
10464     ({                                                                  \
10465        int16x8_t a_ = (a);                                              \
10466        int8x8_t result;                                                 \
10467        __asm__ ("shrn %0.8b,%1.8h,%2"                                   \
10468                 : "=w"(result)                                          \
10469                 : "w"(a_), "i"(b)                                       \
10470                 : /* No clobbers */);                                   \
10471        result;                                                          \
10472      })
10473
10474 #define vshrn_n_s32(a, b)                                               \
10475   __extension__                                                         \
10476     ({                                                                  \
10477        int32x4_t a_ = (a);                                              \
10478        int16x4_t result;                                                \
10479        __asm__ ("shrn %0.4h,%1.4s,%2"                                   \
10480                 : "=w"(result)                                          \
10481                 : "w"(a_), "i"(b)                                       \
10482                 : /* No clobbers */);                                   \
10483        result;                                                          \
10484      })
10485
10486 #define vshrn_n_s64(a, b)                                               \
10487   __extension__                                                         \
10488     ({                                                                  \
10489        int64x2_t a_ = (a);                                              \
10490        int32x2_t result;                                                \
10491        __asm__ ("shrn %0.2s,%1.2d,%2"                                   \
10492                 : "=w"(result)                                          \
10493                 : "w"(a_), "i"(b)                                       \
10494                 : /* No clobbers */);                                   \
10495        result;                                                          \
10496      })
10497
10498 #define vshrn_n_u16(a, b)                                               \
10499   __extension__                                                         \
10500     ({                                                                  \
10501        uint16x8_t a_ = (a);                                             \
10502        uint8x8_t result;                                                \
10503        __asm__ ("shrn %0.8b,%1.8h,%2"                                   \
10504                 : "=w"(result)                                          \
10505                 : "w"(a_), "i"(b)                                       \
10506                 : /* No clobbers */);                                   \
10507        result;                                                          \
10508      })
10509
10510 #define vshrn_n_u32(a, b)                                               \
10511   __extension__                                                         \
10512     ({                                                                  \
10513        uint32x4_t a_ = (a);                                             \
10514        uint16x4_t result;                                               \
10515        __asm__ ("shrn %0.4h,%1.4s,%2"                                   \
10516                 : "=w"(result)                                          \
10517                 : "w"(a_), "i"(b)                                       \
10518                 : /* No clobbers */);                                   \
10519        result;                                                          \
10520      })
10521
10522 #define vshrn_n_u64(a, b)                                               \
10523   __extension__                                                         \
10524     ({                                                                  \
10525        uint64x2_t a_ = (a);                                             \
10526        uint32x2_t result;                                               \
10527        __asm__ ("shrn %0.2s,%1.2d,%2"                                   \
10528                 : "=w"(result)                                          \
10529                 : "w"(a_), "i"(b)                                       \
10530                 : /* No clobbers */);                                   \
10531        result;                                                          \
10532      })
10533
10534 #define vsli_n_p8(a, b, c)                                              \
10535   __extension__                                                         \
10536     ({                                                                  \
10537        poly8x8_t b_ = (b);                                              \
10538        poly8x8_t a_ = (a);                                              \
10539        poly8x8_t result;                                                \
10540        __asm__ ("sli %0.8b,%2.8b,%3"                                    \
10541                 : "=w"(result)                                          \
10542                 : "0"(a_), "w"(b_), "i"(c)                              \
10543                 : /* No clobbers */);                                   \
10544        result;                                                          \
10545      })
10546
10547 #define vsli_n_p16(a, b, c)                                             \
10548   __extension__                                                         \
10549     ({                                                                  \
10550        poly16x4_t b_ = (b);                                             \
10551        poly16x4_t a_ = (a);                                             \
10552        poly16x4_t result;                                               \
10553        __asm__ ("sli %0.4h,%2.4h,%3"                                    \
10554                 : "=w"(result)                                          \
10555                 : "0"(a_), "w"(b_), "i"(c)                              \
10556                 : /* No clobbers */);                                   \
10557        result;                                                          \
10558      })
10559
10560 #define vsliq_n_p8(a, b, c)                                             \
10561   __extension__                                                         \
10562     ({                                                                  \
10563        poly8x16_t b_ = (b);                                             \
10564        poly8x16_t a_ = (a);                                             \
10565        poly8x16_t result;                                               \
10566        __asm__ ("sli %0.16b,%2.16b,%3"                                  \
10567                 : "=w"(result)                                          \
10568                 : "0"(a_), "w"(b_), "i"(c)                              \
10569                 : /* No clobbers */);                                   \
10570        result;                                                          \
10571      })
10572
10573 #define vsliq_n_p16(a, b, c)                                            \
10574   __extension__                                                         \
10575     ({                                                                  \
10576        poly16x8_t b_ = (b);                                             \
10577        poly16x8_t a_ = (a);                                             \
10578        poly16x8_t result;                                               \
10579        __asm__ ("sli %0.8h,%2.8h,%3"                                    \
10580                 : "=w"(result)                                          \
10581                 : "0"(a_), "w"(b_), "i"(c)                              \
10582                 : /* No clobbers */);                                   \
10583        result;                                                          \
10584      })
10585
10586 #define vsri_n_p8(a, b, c)                                              \
10587   __extension__                                                         \
10588     ({                                                                  \
10589        poly8x8_t b_ = (b);                                              \
10590        poly8x8_t a_ = (a);                                              \
10591        poly8x8_t result;                                                \
10592        __asm__ ("sri %0.8b,%2.8b,%3"                                    \
10593                 : "=w"(result)                                          \
10594                 : "0"(a_), "w"(b_), "i"(c)                              \
10595                 : /* No clobbers */);                                   \
10596        result;                                                          \
10597      })
10598
10599 #define vsri_n_p16(a, b, c)                                             \
10600   __extension__                                                         \
10601     ({                                                                  \
10602        poly16x4_t b_ = (b);                                             \
10603        poly16x4_t a_ = (a);                                             \
10604        poly16x4_t result;                                               \
10605        __asm__ ("sri %0.4h,%2.4h,%3"                                    \
10606                 : "=w"(result)                                          \
10607                 : "0"(a_), "w"(b_), "i"(c)                              \
10608                 : /* No clobbers */);                                   \
10609        result;                                                          \
10610      })
10611
10612 #define vsri_n_p64(a, b, c)                                             \
10613   __extension__                                                         \
10614     ({                                                                  \
10615        poly64x1_t b_ = (b);                                             \
10616        poly64x1_t a_ = (a);                                             \
10617        poly64x1_t result;                                               \
10618        __asm__ ("sri %d0,%d2,%3"                                        \
10619                 : "=w"(result)                                          \
10620                 : "0"(a_), "w"(b_), "i"(c)                              \
10621                 : /* No clobbers.  */);                                 \
10622        result;                                                          \
10623      })
10624
10625 #define vsriq_n_p8(a, b, c)                                             \
10626   __extension__                                                         \
10627     ({                                                                  \
10628        poly8x16_t b_ = (b);                                             \
10629        poly8x16_t a_ = (a);                                             \
10630        poly8x16_t result;                                               \
10631        __asm__ ("sri %0.16b,%2.16b,%3"                                  \
10632                 : "=w"(result)                                          \
10633                 : "0"(a_), "w"(b_), "i"(c)                              \
10634                 : /* No clobbers */);                                   \
10635        result;                                                          \
10636      })
10637
10638 #define vsriq_n_p16(a, b, c)                                            \
10639   __extension__                                                         \
10640     ({                                                                  \
10641        poly16x8_t b_ = (b);                                             \
10642        poly16x8_t a_ = (a);                                             \
10643        poly16x8_t result;                                               \
10644        __asm__ ("sri %0.8h,%2.8h,%3"                                    \
10645                 : "=w"(result)                                          \
10646                 : "0"(a_), "w"(b_), "i"(c)                              \
10647                 : /* No clobbers */);                                   \
10648        result;                                                          \
10649      })
10650
10651 #define vsriq_n_p64(a, b, c)                                            \
10652   __extension__                                                         \
10653     ({                                                                  \
10654        poly64x2_t b_ = (b);                                             \
10655        poly64x2_t a_ = (a);                                             \
10656        poly64x2_t result;                                               \
10657        __asm__ ("sri %0.2d,%2.2d,%3"                                    \
10658                 : "=w"(result)                                          \
10659                 : "0"(a_), "w"(b_), "i"(c)                              \
10660                 : /* No clobbers.  */);                                 \
10661        result;                                                          \
10662      })
10663
10664 __extension__ extern __inline uint8x8_t
10665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10666 vtst_p8 (poly8x8_t a, poly8x8_t b)
10667 {
10668   uint8x8_t result;
10669   __asm__ ("cmtst %0.8b, %1.8b, %2.8b"
10670            : "=w"(result)
10671            : "w"(a), "w"(b)
10672            : /* No clobbers */);
10673   return result;
10674 }
10675
10676 __extension__ extern __inline uint16x4_t
10677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10678 vtst_p16 (poly16x4_t a, poly16x4_t b)
10679 {
10680   uint16x4_t result;
10681   __asm__ ("cmtst %0.4h, %1.4h, %2.4h"
10682            : "=w"(result)
10683            : "w"(a), "w"(b)
10684            : /* No clobbers */);
10685   return result;
10686 }
10687
10688 __extension__ extern __inline uint8x16_t
10689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10690 vtstq_p8 (poly8x16_t a, poly8x16_t b)
10691 {
10692   uint8x16_t result;
10693   __asm__ ("cmtst %0.16b, %1.16b, %2.16b"
10694            : "=w"(result)
10695            : "w"(a), "w"(b)
10696            : /* No clobbers */);
10697   return result;
10698 }
10699
10700 __extension__ extern __inline uint16x8_t
10701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10702 vtstq_p16 (poly16x8_t a, poly16x8_t b)
10703 {
10704   uint16x8_t result;
10705   __asm__ ("cmtst %0.8h, %1.8h, %2.8h"
10706            : "=w"(result)
10707            : "w"(a), "w"(b)
10708            : /* No clobbers */);
10709   return result;
10710 }
10711
10712 /* End of temporary inline asm implementations.  */
10713
10714 /* Start of temporary inline asm for vldn, vstn and friends.  */
10715
10716 /* Create struct element types for duplicating loads.
10717
10718    Create 2 element structures of:
10719
10720    +------+----+----+----+----+
10721    |      | 8  | 16 | 32 | 64 |
10722    +------+----+----+----+----+
10723    |int   | Y  | Y  | N  | N  |
10724    +------+----+----+----+----+
10725    |uint  | Y  | Y  | N  | N  |
10726    +------+----+----+----+----+
10727    |float | -  | Y  | N  | N  |
10728    +------+----+----+----+----+
10729    |poly  | Y  | Y  | -  | -  |
10730    +------+----+----+----+----+
10731
10732    Create 3 element structures of:
10733
10734    +------+----+----+----+----+
10735    |      | 8  | 16 | 32 | 64 |
10736    +------+----+----+----+----+
10737    |int   | Y  | Y  | Y  | Y  |
10738    +------+----+----+----+----+
10739    |uint  | Y  | Y  | Y  | Y  |
10740    +------+----+----+----+----+
10741    |float | -  | Y  | Y  | Y  |
10742    +------+----+----+----+----+
10743    |poly  | Y  | Y  | -  | -  |
10744    +------+----+----+----+----+
10745
10746    Create 4 element structures of:
10747
10748    +------+----+----+----+----+
10749    |      | 8  | 16 | 32 | 64 |
10750    +------+----+----+----+----+
10751    |int   | Y  | N  | N  | Y  |
10752    +------+----+----+----+----+
10753    |uint  | Y  | N  | N  | Y  |
10754    +------+----+----+----+----+
10755    |float | -  | N  | N  | Y  |
10756    +------+----+----+----+----+
10757    |poly  | Y  | N  | -  | -  |
10758    +------+----+----+----+----+
10759
10760   This is required for casting memory reference.  */
10761 #define __STRUCTN(t, sz, nelem)                 \
10762   typedef struct t ## sz ## x ## nelem ## _t {  \
10763     t ## sz ## _t val[nelem];                   \
10764   }  t ## sz ## x ## nelem ## _t;
10765
10766 /* 2-element structs.  */
10767 __STRUCTN (int, 8, 2)
10768 __STRUCTN (int, 16, 2)
10769 __STRUCTN (uint, 8, 2)
10770 __STRUCTN (uint, 16, 2)
10771 __STRUCTN (float, 16, 2)
10772 __STRUCTN (poly, 8, 2)
10773 __STRUCTN (poly, 16, 2)
10774 /* 3-element structs.  */
10775 __STRUCTN (int, 8, 3)
10776 __STRUCTN (int, 16, 3)
10777 __STRUCTN (int, 32, 3)
10778 __STRUCTN (int, 64, 3)
10779 __STRUCTN (uint, 8, 3)
10780 __STRUCTN (uint, 16, 3)
10781 __STRUCTN (uint, 32, 3)
10782 __STRUCTN (uint, 64, 3)
10783 __STRUCTN (float, 16, 3)
10784 __STRUCTN (float, 32, 3)
10785 __STRUCTN (float, 64, 3)
10786 __STRUCTN (poly, 8, 3)
10787 __STRUCTN (poly, 16, 3)
10788 /* 4-element structs.  */
10789 __STRUCTN (int, 8, 4)
10790 __STRUCTN (int, 64, 4)
10791 __STRUCTN (uint, 8, 4)
10792 __STRUCTN (uint, 64, 4)
10793 __STRUCTN (poly, 8, 4)
10794 __STRUCTN (float, 64, 4)
10795 #undef __STRUCTN
10796
10797
10798 #define __ST2_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10799                         qmode, ptr_mode, funcsuffix, signedtype)             \
10800 __extension__ extern __inline void                                           \
10801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10802 vst2_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10803                           intype __b, const int __c)                         \
10804 {                                                                            \
10805   __builtin_aarch64_simd_oi __o;                                             \
10806   largetype __temp;                                                          \
10807   __temp.val[0]                                                              \
10808     = vcombine_##funcsuffix (__b.val[0],                                     \
10809                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10810   __temp.val[1]                                                              \
10811     = vcombine_##funcsuffix (__b.val[1],                                     \
10812                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10813   __o = __builtin_aarch64_set_qregoi##qmode (__o,                            \
10814                                              (signedtype) __temp.val[0], 0); \
10815   __o = __builtin_aarch64_set_qregoi##qmode (__o,                            \
10816                                              (signedtype) __temp.val[1], 1); \
10817   __builtin_aarch64_st2_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
10818                                      __ptr, __o, __c);                       \
10819 }
10820
10821 __ST2_LANE_FUNC (float16x4x2_t, float16x8x2_t, float16_t, v4hf, v8hf, hf, f16,
10822                  float16x8_t)
10823 __ST2_LANE_FUNC (float32x2x2_t, float32x4x2_t, float32_t, v2sf, v4sf, sf, f32,
10824                  float32x4_t)
10825 __ST2_LANE_FUNC (float64x1x2_t, float64x2x2_t, float64_t, df, v2df, df, f64,
10826                  float64x2_t)
10827 __ST2_LANE_FUNC (poly8x8x2_t, poly8x16x2_t, poly8_t, v8qi, v16qi, qi, p8,
10828                  int8x16_t)
10829 __ST2_LANE_FUNC (poly16x4x2_t, poly16x8x2_t, poly16_t, v4hi, v8hi, hi, p16,
10830                  int16x8_t)
10831 __ST2_LANE_FUNC (poly64x1x2_t, poly64x2x2_t, poly64_t, di, v2di_ssps, di, p64,
10832                  poly64x2_t)
10833 __ST2_LANE_FUNC (int8x8x2_t, int8x16x2_t, int8_t, v8qi, v16qi, qi, s8,
10834                  int8x16_t)
10835 __ST2_LANE_FUNC (int16x4x2_t, int16x8x2_t, int16_t, v4hi, v8hi, hi, s16,
10836                  int16x8_t)
10837 __ST2_LANE_FUNC (int32x2x2_t, int32x4x2_t, int32_t, v2si, v4si, si, s32,
10838                  int32x4_t)
10839 __ST2_LANE_FUNC (int64x1x2_t, int64x2x2_t, int64_t, di, v2di, di, s64,
10840                  int64x2_t)
10841 __ST2_LANE_FUNC (uint8x8x2_t, uint8x16x2_t, uint8_t, v8qi, v16qi, qi, u8,
10842                  int8x16_t)
10843 __ST2_LANE_FUNC (uint16x4x2_t, uint16x8x2_t, uint16_t, v4hi, v8hi, hi, u16,
10844                  int16x8_t)
10845 __ST2_LANE_FUNC (uint32x2x2_t, uint32x4x2_t, uint32_t, v2si, v4si, si, u32,
10846                  int32x4_t)
10847 __ST2_LANE_FUNC (uint64x1x2_t, uint64x2x2_t, uint64_t, di, v2di, di, u64,
10848                  int64x2_t)
10849
10850 #undef __ST2_LANE_FUNC
10851 #define __ST2_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
10852 __extension__ extern __inline void                                          \
10853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10854 vst2q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
10855                            intype __b, const int __c)                       \
10856 {                                                                           \
10857   union { intype __i;                                                       \
10858           __builtin_aarch64_simd_oi __o; } __temp = { __b };                \
10859   __builtin_aarch64_st2_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
10860                                     __ptr, __temp.__o, __c);                \
10861 }
10862
10863 __ST2_LANE_FUNC (float16x8x2_t, float16_t, v8hf, hf, f16)
10864 __ST2_LANE_FUNC (float32x4x2_t, float32_t, v4sf, sf, f32)
10865 __ST2_LANE_FUNC (float64x2x2_t, float64_t, v2df, df, f64)
10866 __ST2_LANE_FUNC (poly8x16x2_t, poly8_t, v16qi, qi, p8)
10867 __ST2_LANE_FUNC (poly16x8x2_t, poly16_t, v8hi, hi, p16)
10868 __ST2_LANE_FUNC (poly64x2x2_t, poly64_t, v2di, di, p64)
10869 __ST2_LANE_FUNC (int8x16x2_t, int8_t, v16qi, qi, s8)
10870 __ST2_LANE_FUNC (int16x8x2_t, int16_t, v8hi, hi, s16)
10871 __ST2_LANE_FUNC (int32x4x2_t, int32_t, v4si, si, s32)
10872 __ST2_LANE_FUNC (int64x2x2_t, int64_t, v2di, di, s64)
10873 __ST2_LANE_FUNC (uint8x16x2_t, uint8_t, v16qi, qi, u8)
10874 __ST2_LANE_FUNC (uint16x8x2_t, uint16_t, v8hi, hi, u16)
10875 __ST2_LANE_FUNC (uint32x4x2_t, uint32_t, v4si, si, u32)
10876 __ST2_LANE_FUNC (uint64x2x2_t, uint64_t, v2di, di, u64)
10877
10878 #define __ST3_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10879                         qmode, ptr_mode, funcsuffix, signedtype)             \
10880 __extension__ extern __inline void                                           \
10881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10882 vst3_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10883                           intype __b, const int __c)                         \
10884 {                                                                            \
10885   __builtin_aarch64_simd_ci __o;                                             \
10886   largetype __temp;                                                          \
10887   __temp.val[0]                                                              \
10888     = vcombine_##funcsuffix (__b.val[0],                                     \
10889                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10890   __temp.val[1]                                                              \
10891     = vcombine_##funcsuffix (__b.val[1],                                     \
10892                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10893   __temp.val[2]                                                              \
10894     = vcombine_##funcsuffix (__b.val[2],                                     \
10895                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10896   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
10897                                              (signedtype) __temp.val[0], 0); \
10898   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
10899                                              (signedtype) __temp.val[1], 1); \
10900   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
10901                                              (signedtype) __temp.val[2], 2); \
10902   __builtin_aarch64_st3_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
10903                                      __ptr, __o, __c);                       \
10904 }
10905
10906 __ST3_LANE_FUNC (float16x4x3_t, float16x8x3_t, float16_t, v4hf, v8hf, hf, f16,
10907                  float16x8_t)
10908 __ST3_LANE_FUNC (float32x2x3_t, float32x4x3_t, float32_t, v2sf, v4sf, sf, f32,
10909                  float32x4_t)
10910 __ST3_LANE_FUNC (float64x1x3_t, float64x2x3_t, float64_t, df, v2df, df, f64,
10911                  float64x2_t)
10912 __ST3_LANE_FUNC (poly8x8x3_t, poly8x16x3_t, poly8_t, v8qi, v16qi, qi, p8,
10913                  int8x16_t)
10914 __ST3_LANE_FUNC (poly16x4x3_t, poly16x8x3_t, poly16_t, v4hi, v8hi, hi, p16,
10915                  int16x8_t)
10916 __ST3_LANE_FUNC (poly64x1x3_t, poly64x2x3_t, poly64_t, di, v2di_ssps, di, p64,
10917                  poly64x2_t)
10918 __ST3_LANE_FUNC (int8x8x3_t, int8x16x3_t, int8_t, v8qi, v16qi, qi, s8,
10919                  int8x16_t)
10920 __ST3_LANE_FUNC (int16x4x3_t, int16x8x3_t, int16_t, v4hi, v8hi, hi, s16,
10921                  int16x8_t)
10922 __ST3_LANE_FUNC (int32x2x3_t, int32x4x3_t, int32_t, v2si, v4si, si, s32,
10923                  int32x4_t)
10924 __ST3_LANE_FUNC (int64x1x3_t, int64x2x3_t, int64_t, di, v2di, di, s64,
10925                  int64x2_t)
10926 __ST3_LANE_FUNC (uint8x8x3_t, uint8x16x3_t, uint8_t, v8qi, v16qi, qi, u8,
10927                  int8x16_t)
10928 __ST3_LANE_FUNC (uint16x4x3_t, uint16x8x3_t, uint16_t, v4hi, v8hi, hi, u16,
10929                  int16x8_t)
10930 __ST3_LANE_FUNC (uint32x2x3_t, uint32x4x3_t, uint32_t, v2si, v4si, si, u32,
10931                  int32x4_t)
10932 __ST3_LANE_FUNC (uint64x1x3_t, uint64x2x3_t, uint64_t, di, v2di, di, u64,
10933                  int64x2_t)
10934
10935 #undef __ST3_LANE_FUNC
10936 #define __ST3_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
10937 __extension__ extern __inline void                                          \
10938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10939 vst3q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
10940                            intype __b, const int __c)                       \
10941 {                                                                           \
10942   union { intype __i;                                                       \
10943           __builtin_aarch64_simd_ci __o; } __temp = { __b };                \
10944   __builtin_aarch64_st3_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
10945                                     __ptr, __temp.__o, __c);                \
10946 }
10947
10948 __ST3_LANE_FUNC (float16x8x3_t, float16_t, v8hf, hf, f16)
10949 __ST3_LANE_FUNC (float32x4x3_t, float32_t, v4sf, sf, f32)
10950 __ST3_LANE_FUNC (float64x2x3_t, float64_t, v2df, df, f64)
10951 __ST3_LANE_FUNC (poly8x16x3_t, poly8_t, v16qi, qi, p8)
10952 __ST3_LANE_FUNC (poly16x8x3_t, poly16_t, v8hi, hi, p16)
10953 __ST3_LANE_FUNC (poly64x2x3_t, poly64_t, v2di, di, p64)
10954 __ST3_LANE_FUNC (int8x16x3_t, int8_t, v16qi, qi, s8)
10955 __ST3_LANE_FUNC (int16x8x3_t, int16_t, v8hi, hi, s16)
10956 __ST3_LANE_FUNC (int32x4x3_t, int32_t, v4si, si, s32)
10957 __ST3_LANE_FUNC (int64x2x3_t, int64_t, v2di, di, s64)
10958 __ST3_LANE_FUNC (uint8x16x3_t, uint8_t, v16qi, qi, u8)
10959 __ST3_LANE_FUNC (uint16x8x3_t, uint16_t, v8hi, hi, u16)
10960 __ST3_LANE_FUNC (uint32x4x3_t, uint32_t, v4si, si, u32)
10961 __ST3_LANE_FUNC (uint64x2x3_t, uint64_t, v2di, di, u64)
10962
10963 #define __ST4_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10964                         qmode, ptr_mode, funcsuffix, signedtype)             \
10965 __extension__ extern __inline void                                           \
10966 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10967 vst4_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10968                           intype __b, const int __c)                         \
10969 {                                                                            \
10970   __builtin_aarch64_simd_xi __o;                                             \
10971   largetype __temp;                                                          \
10972   __temp.val[0]                                                              \
10973     = vcombine_##funcsuffix (__b.val[0],                                     \
10974                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10975   __temp.val[1]                                                              \
10976     = vcombine_##funcsuffix (__b.val[1],                                     \
10977                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10978   __temp.val[2]                                                              \
10979     = vcombine_##funcsuffix (__b.val[2],                                     \
10980                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10981   __temp.val[3]                                                              \
10982     = vcombine_##funcsuffix (__b.val[3],                                     \
10983                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10984   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10985                                              (signedtype) __temp.val[0], 0); \
10986   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10987                                              (signedtype) __temp.val[1], 1); \
10988   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10989                                              (signedtype) __temp.val[2], 2); \
10990   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10991                                              (signedtype) __temp.val[3], 3); \
10992   __builtin_aarch64_st4_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
10993                                      __ptr, __o, __c);                       \
10994 }
10995
10996 __ST4_LANE_FUNC (float16x4x4_t, float16x8x4_t, float16_t, v4hf, v8hf, hf, f16,
10997                  float16x8_t)
10998 __ST4_LANE_FUNC (float32x2x4_t, float32x4x4_t, float32_t, v2sf, v4sf, sf, f32,
10999                  float32x4_t)
11000 __ST4_LANE_FUNC (float64x1x4_t, float64x2x4_t, float64_t, df, v2df, df, f64,
11001                  float64x2_t)
11002 __ST4_LANE_FUNC (poly8x8x4_t, poly8x16x4_t, poly8_t, v8qi, v16qi, qi, p8,
11003                  int8x16_t)
11004 __ST4_LANE_FUNC (poly16x4x4_t, poly16x8x4_t, poly16_t, v4hi, v8hi, hi, p16,
11005                  int16x8_t)
11006 __ST4_LANE_FUNC (poly64x1x4_t, poly64x2x4_t, poly64_t, di, v2di_ssps, di, p64,
11007                  poly64x2_t)
11008 __ST4_LANE_FUNC (int8x8x4_t, int8x16x4_t, int8_t, v8qi, v16qi, qi, s8,
11009                  int8x16_t)
11010 __ST4_LANE_FUNC (int16x4x4_t, int16x8x4_t, int16_t, v4hi, v8hi, hi, s16,
11011                  int16x8_t)
11012 __ST4_LANE_FUNC (int32x2x4_t, int32x4x4_t, int32_t, v2si, v4si, si, s32,
11013                  int32x4_t)
11014 __ST4_LANE_FUNC (int64x1x4_t, int64x2x4_t, int64_t, di, v2di, di, s64,
11015                  int64x2_t)
11016 __ST4_LANE_FUNC (uint8x8x4_t, uint8x16x4_t, uint8_t, v8qi, v16qi, qi, u8,
11017                  int8x16_t)
11018 __ST4_LANE_FUNC (uint16x4x4_t, uint16x8x4_t, uint16_t, v4hi, v8hi, hi, u16,
11019                  int16x8_t)
11020 __ST4_LANE_FUNC (uint32x2x4_t, uint32x4x4_t, uint32_t, v2si, v4si, si, u32,
11021                  int32x4_t)
11022 __ST4_LANE_FUNC (uint64x1x4_t, uint64x2x4_t, uint64_t, di, v2di, di, u64,
11023                  int64x2_t)
11024
11025 #undef __ST4_LANE_FUNC
11026 #define __ST4_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
11027 __extension__ extern __inline void                                          \
11028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11029 vst4q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
11030                            intype __b, const int __c)                       \
11031 {                                                                           \
11032   union { intype __i;                                                       \
11033           __builtin_aarch64_simd_xi __o; } __temp = { __b };                \
11034   __builtin_aarch64_st4_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
11035                                     __ptr, __temp.__o, __c);                \
11036 }
11037
11038 __ST4_LANE_FUNC (float16x8x4_t, float16_t, v8hf, hf, f16)
11039 __ST4_LANE_FUNC (float32x4x4_t, float32_t, v4sf, sf, f32)
11040 __ST4_LANE_FUNC (float64x2x4_t, float64_t, v2df, df, f64)
11041 __ST4_LANE_FUNC (poly8x16x4_t, poly8_t, v16qi, qi, p8)
11042 __ST4_LANE_FUNC (poly16x8x4_t, poly16_t, v8hi, hi, p16)
11043 __ST4_LANE_FUNC (poly64x2x4_t, poly64_t, v2di, di, p64)
11044 __ST4_LANE_FUNC (int8x16x4_t, int8_t, v16qi, qi, s8)
11045 __ST4_LANE_FUNC (int16x8x4_t, int16_t, v8hi, hi, s16)
11046 __ST4_LANE_FUNC (int32x4x4_t, int32_t, v4si, si, s32)
11047 __ST4_LANE_FUNC (int64x2x4_t, int64_t, v2di, di, s64)
11048 __ST4_LANE_FUNC (uint8x16x4_t, uint8_t, v16qi, qi, u8)
11049 __ST4_LANE_FUNC (uint16x8x4_t, uint16_t, v8hi, hi, u16)
11050 __ST4_LANE_FUNC (uint32x4x4_t, uint32_t, v4si, si, u32)
11051 __ST4_LANE_FUNC (uint64x2x4_t, uint64_t, v2di, di, u64)
11052
11053 __extension__ extern __inline int64_t
11054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11055 vaddlv_s32 (int32x2_t a)
11056 {
11057   int64_t result;
11058   __asm__ ("saddlp %0.1d, %1.2s" : "=w"(result) : "w"(a) : );
11059   return result;
11060 }
11061
11062 __extension__ extern __inline uint64_t
11063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11064 vaddlv_u32 (uint32x2_t a)
11065 {
11066   uint64_t result;
11067   __asm__ ("uaddlp %0.1d, %1.2s" : "=w"(result) : "w"(a) : );
11068   return result;
11069 }
11070
11071 __extension__ extern __inline int16x4_t
11072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11073 vqdmulh_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __c)
11074 {
11075   return __builtin_aarch64_sqdmulh_laneqv4hi (__a, __b, __c);
11076 }
11077
11078 __extension__ extern __inline int32x2_t
11079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11080 vqdmulh_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __c)
11081 {
11082   return __builtin_aarch64_sqdmulh_laneqv2si (__a, __b, __c);
11083 }
11084
11085 __extension__ extern __inline int16x8_t
11086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11087 vqdmulhq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __c)
11088 {
11089   return __builtin_aarch64_sqdmulh_laneqv8hi (__a, __b, __c);
11090 }
11091
11092 __extension__ extern __inline int32x4_t
11093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11094 vqdmulhq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __c)
11095 {
11096   return __builtin_aarch64_sqdmulh_laneqv4si (__a, __b, __c);
11097 }
11098
11099 __extension__ extern __inline int16x4_t
11100 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11101 vqrdmulh_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __c)
11102 {
11103   return  __builtin_aarch64_sqrdmulh_laneqv4hi (__a, __b, __c);
11104 }
11105
11106 __extension__ extern __inline int32x2_t
11107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11108 vqrdmulh_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __c)
11109 {
11110   return __builtin_aarch64_sqrdmulh_laneqv2si (__a, __b, __c);
11111 }
11112
11113 __extension__ extern __inline int16x8_t
11114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11115 vqrdmulhq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __c)
11116 {
11117   return __builtin_aarch64_sqrdmulh_laneqv8hi (__a, __b, __c);
11118 }
11119
11120 __extension__ extern __inline int32x4_t
11121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11122 vqrdmulhq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __c)
11123 {
11124   return __builtin_aarch64_sqrdmulh_laneqv4si (__a, __b, __c);
11125 }
11126
11127 /* Table intrinsics.  */
11128
11129 __extension__ extern __inline poly8x8_t
11130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11131 vqtbl1_p8 (poly8x16_t a, uint8x8_t b)
11132 {
11133   poly8x8_t result;
11134   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11135            : "=w"(result)
11136            : "w"(a), "w"(b)
11137            : /* No clobbers */);
11138   return result;
11139 }
11140
11141 __extension__ extern __inline int8x8_t
11142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11143 vqtbl1_s8 (int8x16_t a, uint8x8_t b)
11144 {
11145   int8x8_t result;
11146   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11147            : "=w"(result)
11148            : "w"(a), "w"(b)
11149            : /* No clobbers */);
11150   return result;
11151 }
11152
11153 __extension__ extern __inline uint8x8_t
11154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11155 vqtbl1_u8 (uint8x16_t a, uint8x8_t b)
11156 {
11157   uint8x8_t result;
11158   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11159            : "=w"(result)
11160            : "w"(a), "w"(b)
11161            : /* No clobbers */);
11162   return result;
11163 }
11164
11165 __extension__ extern __inline poly8x16_t
11166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11167 vqtbl1q_p8 (poly8x16_t a, uint8x16_t b)
11168 {
11169   poly8x16_t result;
11170   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11171            : "=w"(result)
11172            : "w"(a), "w"(b)
11173            : /* No clobbers */);
11174   return result;
11175 }
11176
11177 __extension__ extern __inline int8x16_t
11178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11179 vqtbl1q_s8 (int8x16_t a, uint8x16_t b)
11180 {
11181   int8x16_t result;
11182   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11183            : "=w"(result)
11184            : "w"(a), "w"(b)
11185            : /* No clobbers */);
11186   return result;
11187 }
11188
11189 __extension__ extern __inline uint8x16_t
11190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11191 vqtbl1q_u8 (uint8x16_t a, uint8x16_t b)
11192 {
11193   uint8x16_t result;
11194   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11195            : "=w"(result)
11196            : "w"(a), "w"(b)
11197            : /* No clobbers */);
11198   return result;
11199 }
11200
11201 __extension__ extern __inline int8x8_t
11202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11203 vqtbx1_s8 (int8x8_t r, int8x16_t tab, uint8x8_t idx)
11204 {
11205   int8x8_t result = r;
11206   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11207            : "+w"(result)
11208            : "w"(tab), "w"(idx)
11209            : /* No clobbers */);
11210   return result;
11211 }
11212
11213 __extension__ extern __inline uint8x8_t
11214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11215 vqtbx1_u8 (uint8x8_t r, uint8x16_t tab, uint8x8_t idx)
11216 {
11217   uint8x8_t result = r;
11218   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11219            : "+w"(result)
11220            : "w"(tab), "w"(idx)
11221            : /* No clobbers */);
11222   return result;
11223 }
11224
11225 __extension__ extern __inline poly8x8_t
11226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11227 vqtbx1_p8 (poly8x8_t r, poly8x16_t tab, uint8x8_t idx)
11228 {
11229   poly8x8_t result = r;
11230   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11231            : "+w"(result)
11232            : "w"(tab), "w"(idx)
11233            : /* No clobbers */);
11234   return result;
11235 }
11236
11237 __extension__ extern __inline int8x16_t
11238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11239 vqtbx1q_s8 (int8x16_t r, int8x16_t tab, uint8x16_t idx)
11240 {
11241   int8x16_t result = r;
11242   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11243            : "+w"(result)
11244            : "w"(tab), "w"(idx)
11245            : /* No clobbers */);
11246   return result;
11247 }
11248
11249 __extension__ extern __inline uint8x16_t
11250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11251 vqtbx1q_u8 (uint8x16_t r, uint8x16_t tab, uint8x16_t idx)
11252 {
11253   uint8x16_t result = r;
11254   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11255            : "+w"(result)
11256            : "w"(tab), "w"(idx)
11257            : /* No clobbers */);
11258   return result;
11259 }
11260
11261 __extension__ extern __inline poly8x16_t
11262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11263 vqtbx1q_p8 (poly8x16_t r, poly8x16_t tab, uint8x16_t idx)
11264 {
11265   poly8x16_t result = r;
11266   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11267            : "+w"(result)
11268            : "w"(tab), "w"(idx)
11269            : /* No clobbers */);
11270   return result;
11271 }
11272
11273 /* V7 legacy table intrinsics.  */
11274
11275 __extension__ extern __inline int8x8_t
11276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11277 vtbl1_s8 (int8x8_t tab, int8x8_t idx)
11278 {
11279   int8x8_t result;
11280   int8x16_t temp = vcombine_s8 (tab, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
11281   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11282            : "=w"(result)
11283            : "w"(temp), "w"(idx)
11284            : /* No clobbers */);
11285   return result;
11286 }
11287
11288 __extension__ extern __inline uint8x8_t
11289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11290 vtbl1_u8 (uint8x8_t tab, uint8x8_t idx)
11291 {
11292   uint8x8_t result;
11293   uint8x16_t temp = vcombine_u8 (tab, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
11294   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11295            : "=w"(result)
11296            : "w"(temp), "w"(idx)
11297            : /* No clobbers */);
11298   return result;
11299 }
11300
11301 __extension__ extern __inline poly8x8_t
11302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11303 vtbl1_p8 (poly8x8_t tab, uint8x8_t idx)
11304 {
11305   poly8x8_t result;
11306   poly8x16_t temp = vcombine_p8 (tab, vcreate_p8 (__AARCH64_UINT64_C (0x0)));
11307   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11308            : "=w"(result)
11309            : "w"(temp), "w"(idx)
11310            : /* No clobbers */);
11311   return result;
11312 }
11313
11314 __extension__ extern __inline int8x8_t
11315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11316 vtbl2_s8 (int8x8x2_t tab, int8x8_t idx)
11317 {
11318   int8x8_t result;
11319   int8x16_t temp = vcombine_s8 (tab.val[0], tab.val[1]);
11320   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11321            : "=w"(result)
11322            : "w"(temp), "w"(idx)
11323            : /* No clobbers */);
11324   return result;
11325 }
11326
11327 __extension__ extern __inline uint8x8_t
11328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11329 vtbl2_u8 (uint8x8x2_t tab, uint8x8_t idx)
11330 {
11331   uint8x8_t result;
11332   uint8x16_t temp = vcombine_u8 (tab.val[0], tab.val[1]);
11333   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11334            : "=w"(result)
11335            : "w"(temp), "w"(idx)
11336            : /* No clobbers */);
11337   return result;
11338 }
11339
11340 __extension__ extern __inline poly8x8_t
11341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11342 vtbl2_p8 (poly8x8x2_t tab, uint8x8_t idx)
11343 {
11344   poly8x8_t result;
11345   poly8x16_t temp = vcombine_p8 (tab.val[0], tab.val[1]);
11346   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11347            : "=w"(result)
11348            : "w"(temp), "w"(idx)
11349            : /* No clobbers */);
11350   return result;
11351 }
11352
11353 __extension__ extern __inline int8x8_t
11354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11355 vtbl3_s8 (int8x8x3_t tab, int8x8_t idx)
11356 {
11357   int8x8_t result;
11358   int8x16x2_t temp;
11359   __builtin_aarch64_simd_oi __o;
11360   temp.val[0] = vcombine_s8 (tab.val[0], tab.val[1]);
11361   temp.val[1] = vcombine_s8 (tab.val[2], vcreate_s8 (__AARCH64_UINT64_C (0x0)));
11362   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11363                                            (int8x16_t) temp.val[0], 0);
11364   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11365                                            (int8x16_t) temp.val[1], 1);
11366   result = __builtin_aarch64_tbl3v8qi (__o, idx);
11367   return result;
11368 }
11369
11370 __extension__ extern __inline uint8x8_t
11371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11372 vtbl3_u8 (uint8x8x3_t tab, uint8x8_t idx)
11373 {
11374   uint8x8_t result;
11375   uint8x16x2_t temp;
11376   __builtin_aarch64_simd_oi __o;
11377   temp.val[0] = vcombine_u8 (tab.val[0], tab.val[1]);
11378   temp.val[1] = vcombine_u8 (tab.val[2], vcreate_u8 (__AARCH64_UINT64_C (0x0)));
11379   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11380                                            (int8x16_t) temp.val[0], 0);
11381   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11382                                            (int8x16_t) temp.val[1], 1);
11383   result = (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11384   return result;
11385 }
11386
11387 __extension__ extern __inline poly8x8_t
11388 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11389 vtbl3_p8 (poly8x8x3_t tab, uint8x8_t idx)
11390 {
11391   poly8x8_t result;
11392   poly8x16x2_t temp;
11393   __builtin_aarch64_simd_oi __o;
11394   temp.val[0] = vcombine_p8 (tab.val[0], tab.val[1]);
11395   temp.val[1] = vcombine_p8 (tab.val[2], vcreate_p8 (__AARCH64_UINT64_C (0x0)));
11396   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11397                                            (int8x16_t) temp.val[0], 0);
11398   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11399                                            (int8x16_t) temp.val[1], 1);
11400   result = (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11401   return result;
11402 }
11403
11404 __extension__ extern __inline int8x8_t
11405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11406 vtbl4_s8 (int8x8x4_t tab, int8x8_t idx)
11407 {
11408   int8x8_t result;
11409   int8x16x2_t temp;
11410   __builtin_aarch64_simd_oi __o;
11411   temp.val[0] = vcombine_s8 (tab.val[0], tab.val[1]);
11412   temp.val[1] = vcombine_s8 (tab.val[2], tab.val[3]);
11413   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11414                                            (int8x16_t) temp.val[0], 0);
11415   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11416                                            (int8x16_t) temp.val[1], 1);
11417   result = __builtin_aarch64_tbl3v8qi (__o, idx);
11418   return result;
11419 }
11420
11421 __extension__ extern __inline uint8x8_t
11422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11423 vtbl4_u8 (uint8x8x4_t tab, uint8x8_t idx)
11424 {
11425   uint8x8_t result;
11426   uint8x16x2_t temp;
11427   __builtin_aarch64_simd_oi __o;
11428   temp.val[0] = vcombine_u8 (tab.val[0], tab.val[1]);
11429   temp.val[1] = vcombine_u8 (tab.val[2], tab.val[3]);
11430   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11431                                            (int8x16_t) temp.val[0], 0);
11432   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11433                                            (int8x16_t) temp.val[1], 1);
11434   result = (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11435   return result;
11436 }
11437
11438 __extension__ extern __inline poly8x8_t
11439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11440 vtbl4_p8 (poly8x8x4_t tab, uint8x8_t idx)
11441 {
11442   poly8x8_t result;
11443   poly8x16x2_t temp;
11444   __builtin_aarch64_simd_oi __o;
11445   temp.val[0] = vcombine_p8 (tab.val[0], tab.val[1]);
11446   temp.val[1] = vcombine_p8 (tab.val[2], tab.val[3]);
11447   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11448                                            (int8x16_t) temp.val[0], 0);
11449   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11450                                            (int8x16_t) temp.val[1], 1);
11451   result = (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11452   return result;
11453 }
11454
11455 __extension__ extern __inline int8x8_t
11456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11457 vtbx2_s8 (int8x8_t r, int8x8x2_t tab, int8x8_t idx)
11458 {
11459   int8x8_t result = r;
11460   int8x16_t temp = vcombine_s8 (tab.val[0], tab.val[1]);
11461   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11462            : "+w"(result)
11463            : "w"(temp), "w"(idx)
11464            : /* No clobbers */);
11465   return result;
11466 }
11467
11468 __extension__ extern __inline uint8x8_t
11469 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11470 vtbx2_u8 (uint8x8_t r, uint8x8x2_t tab, uint8x8_t idx)
11471 {
11472   uint8x8_t result = r;
11473   uint8x16_t temp = vcombine_u8 (tab.val[0], tab.val[1]);
11474   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11475            : "+w"(result)
11476            : "w"(temp), "w"(idx)
11477            : /* No clobbers */);
11478   return result;
11479 }
11480
11481 __extension__ extern __inline poly8x8_t
11482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11483 vtbx2_p8 (poly8x8_t r, poly8x8x2_t tab, uint8x8_t idx)
11484 {
11485   poly8x8_t result = r;
11486   poly8x16_t temp = vcombine_p8 (tab.val[0], tab.val[1]);
11487   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11488            : "+w"(result)
11489            : "w"(temp), "w"(idx)
11490            : /* No clobbers */);
11491   return result;
11492 }
11493
11494 /* End of temporary inline asm.  */
11495
11496 /* Start of optimal implementations in approved order.  */
11497
11498 /* vabd.  */
11499
11500 __extension__ extern __inline float32_t
11501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11502 vabds_f32 (float32_t __a, float32_t __b)
11503 {
11504   return __builtin_aarch64_fabdsf (__a, __b);
11505 }
11506
11507 __extension__ extern __inline float64_t
11508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11509 vabdd_f64 (float64_t __a, float64_t __b)
11510 {
11511   return __builtin_aarch64_fabddf (__a, __b);
11512 }
11513
11514 __extension__ extern __inline float32x2_t
11515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11516 vabd_f32 (float32x2_t __a, float32x2_t __b)
11517 {
11518   return __builtin_aarch64_fabdv2sf (__a, __b);
11519 }
11520
11521 __extension__ extern __inline float64x1_t
11522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11523 vabd_f64 (float64x1_t __a, float64x1_t __b)
11524 {
11525   return (float64x1_t) {vabdd_f64 (vget_lane_f64 (__a, 0),
11526                                    vget_lane_f64 (__b, 0))};
11527 }
11528
11529 __extension__ extern __inline float32x4_t
11530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11531 vabdq_f32 (float32x4_t __a, float32x4_t __b)
11532 {
11533   return __builtin_aarch64_fabdv4sf (__a, __b);
11534 }
11535
11536 __extension__ extern __inline float64x2_t
11537 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11538 vabdq_f64 (float64x2_t __a, float64x2_t __b)
11539 {
11540   return __builtin_aarch64_fabdv2df (__a, __b);
11541 }
11542
11543 /* vabs  */
11544
11545 __extension__ extern __inline float32x2_t
11546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11547 vabs_f32 (float32x2_t __a)
11548 {
11549   return __builtin_aarch64_absv2sf (__a);
11550 }
11551
11552 __extension__ extern __inline float64x1_t
11553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11554 vabs_f64 (float64x1_t __a)
11555 {
11556   return (float64x1_t) {__builtin_fabs (__a[0])};
11557 }
11558
11559 __extension__ extern __inline int8x8_t
11560 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11561 vabs_s8 (int8x8_t __a)
11562 {
11563   return __builtin_aarch64_absv8qi (__a);
11564 }
11565
11566 __extension__ extern __inline int16x4_t
11567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11568 vabs_s16 (int16x4_t __a)
11569 {
11570   return __builtin_aarch64_absv4hi (__a);
11571 }
11572
11573 __extension__ extern __inline int32x2_t
11574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11575 vabs_s32 (int32x2_t __a)
11576 {
11577   return __builtin_aarch64_absv2si (__a);
11578 }
11579
11580 __extension__ extern __inline int64x1_t
11581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11582 vabs_s64 (int64x1_t __a)
11583 {
11584   return (int64x1_t) {__builtin_aarch64_absdi (__a[0])};
11585 }
11586
11587 __extension__ extern __inline float32x4_t
11588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11589 vabsq_f32 (float32x4_t __a)
11590 {
11591   return __builtin_aarch64_absv4sf (__a);
11592 }
11593
11594 __extension__ extern __inline float64x2_t
11595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11596 vabsq_f64 (float64x2_t __a)
11597 {
11598   return __builtin_aarch64_absv2df (__a);
11599 }
11600
11601 __extension__ extern __inline int8x16_t
11602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11603 vabsq_s8 (int8x16_t __a)
11604 {
11605   return __builtin_aarch64_absv16qi (__a);
11606 }
11607
11608 __extension__ extern __inline int16x8_t
11609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11610 vabsq_s16 (int16x8_t __a)
11611 {
11612   return __builtin_aarch64_absv8hi (__a);
11613 }
11614
11615 __extension__ extern __inline int32x4_t
11616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11617 vabsq_s32 (int32x4_t __a)
11618 {
11619   return __builtin_aarch64_absv4si (__a);
11620 }
11621
11622 __extension__ extern __inline int64x2_t
11623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11624 vabsq_s64 (int64x2_t __a)
11625 {
11626   return __builtin_aarch64_absv2di (__a);
11627 }
11628
11629 /* vadd */
11630
11631 __extension__ extern __inline int64_t
11632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11633 vaddd_s64 (int64_t __a, int64_t __b)
11634 {
11635   return __a + __b;
11636 }
11637
11638 __extension__ extern __inline uint64_t
11639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11640 vaddd_u64 (uint64_t __a, uint64_t __b)
11641 {
11642   return __a + __b;
11643 }
11644
11645 /* vaddv */
11646
11647 __extension__ extern __inline int8_t
11648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11649 vaddv_s8 (int8x8_t __a)
11650 {
11651   return __builtin_aarch64_reduc_plus_scal_v8qi (__a);
11652 }
11653
11654 __extension__ extern __inline int16_t
11655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11656 vaddv_s16 (int16x4_t __a)
11657 {
11658   return __builtin_aarch64_reduc_plus_scal_v4hi (__a);
11659 }
11660
11661 __extension__ extern __inline int32_t
11662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11663 vaddv_s32 (int32x2_t __a)
11664 {
11665   return __builtin_aarch64_reduc_plus_scal_v2si (__a);
11666 }
11667
11668 __extension__ extern __inline uint8_t
11669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11670 vaddv_u8 (uint8x8_t __a)
11671 {
11672   return (uint8_t) __builtin_aarch64_reduc_plus_scal_v8qi ((int8x8_t) __a);
11673 }
11674
11675 __extension__ extern __inline uint16_t
11676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11677 vaddv_u16 (uint16x4_t __a)
11678 {
11679   return (uint16_t) __builtin_aarch64_reduc_plus_scal_v4hi ((int16x4_t) __a);
11680 }
11681
11682 __extension__ extern __inline uint32_t
11683 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11684 vaddv_u32 (uint32x2_t __a)
11685 {
11686   return (int32_t) __builtin_aarch64_reduc_plus_scal_v2si ((int32x2_t) __a);
11687 }
11688
11689 __extension__ extern __inline int8_t
11690 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11691 vaddvq_s8 (int8x16_t __a)
11692 {
11693   return __builtin_aarch64_reduc_plus_scal_v16qi (__a);
11694 }
11695
11696 __extension__ extern __inline int16_t
11697 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11698 vaddvq_s16 (int16x8_t __a)
11699 {
11700   return __builtin_aarch64_reduc_plus_scal_v8hi (__a);
11701 }
11702
11703 __extension__ extern __inline int32_t
11704 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11705 vaddvq_s32 (int32x4_t __a)
11706 {
11707   return __builtin_aarch64_reduc_plus_scal_v4si (__a);
11708 }
11709
11710 __extension__ extern __inline int64_t
11711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11712 vaddvq_s64 (int64x2_t __a)
11713 {
11714   return __builtin_aarch64_reduc_plus_scal_v2di (__a);
11715 }
11716
11717 __extension__ extern __inline uint8_t
11718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11719 vaddvq_u8 (uint8x16_t __a)
11720 {
11721   return (uint8_t) __builtin_aarch64_reduc_plus_scal_v16qi ((int8x16_t) __a);
11722 }
11723
11724 __extension__ extern __inline uint16_t
11725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11726 vaddvq_u16 (uint16x8_t __a)
11727 {
11728   return (uint16_t) __builtin_aarch64_reduc_plus_scal_v8hi ((int16x8_t) __a);
11729 }
11730
11731 __extension__ extern __inline uint32_t
11732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11733 vaddvq_u32 (uint32x4_t __a)
11734 {
11735   return (uint32_t) __builtin_aarch64_reduc_plus_scal_v4si ((int32x4_t) __a);
11736 }
11737
11738 __extension__ extern __inline uint64_t
11739 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11740 vaddvq_u64 (uint64x2_t __a)
11741 {
11742   return (uint64_t) __builtin_aarch64_reduc_plus_scal_v2di ((int64x2_t) __a);
11743 }
11744
11745 __extension__ extern __inline float32_t
11746 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11747 vaddv_f32 (float32x2_t __a)
11748 {
11749   return __builtin_aarch64_reduc_plus_scal_v2sf (__a);
11750 }
11751
11752 __extension__ extern __inline float32_t
11753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11754 vaddvq_f32 (float32x4_t __a)
11755 {
11756   return __builtin_aarch64_reduc_plus_scal_v4sf (__a);
11757 }
11758
11759 __extension__ extern __inline float64_t
11760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11761 vaddvq_f64 (float64x2_t __a)
11762 {
11763   return __builtin_aarch64_reduc_plus_scal_v2df (__a);
11764 }
11765
11766 /* vbsl  */
11767
11768 __extension__ extern __inline float16x4_t
11769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11770 vbsl_f16 (uint16x4_t __a, float16x4_t __b, float16x4_t __c)
11771 {
11772   return __builtin_aarch64_simd_bslv4hf_suss (__a, __b, __c);
11773 }
11774
11775 __extension__ extern __inline float32x2_t
11776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11777 vbsl_f32 (uint32x2_t __a, float32x2_t __b, float32x2_t __c)
11778 {
11779   return __builtin_aarch64_simd_bslv2sf_suss (__a, __b, __c);
11780 }
11781
11782 __extension__ extern __inline float64x1_t
11783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11784 vbsl_f64 (uint64x1_t __a, float64x1_t __b, float64x1_t __c)
11785 {
11786   return (float64x1_t)
11787     { __builtin_aarch64_simd_bsldf_suss (__a[0], __b[0], __c[0]) };
11788 }
11789
11790 __extension__ extern __inline poly8x8_t
11791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11792 vbsl_p8 (uint8x8_t __a, poly8x8_t __b, poly8x8_t __c)
11793 {
11794   return __builtin_aarch64_simd_bslv8qi_pupp (__a, __b, __c);
11795 }
11796
11797 __extension__ extern __inline poly16x4_t
11798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11799 vbsl_p16 (uint16x4_t __a, poly16x4_t __b, poly16x4_t __c)
11800 {
11801   return __builtin_aarch64_simd_bslv4hi_pupp (__a, __b, __c);
11802 }
11803 __extension__ extern __inline poly64x1_t
11804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11805 vbsl_p64 (uint64x1_t __a, poly64x1_t __b, poly64x1_t __c)
11806 {
11807   return (poly64x1_t)
11808       {__builtin_aarch64_simd_bsldi_pupp (__a[0], __b[0], __c[0])};
11809 }
11810
11811 __extension__ extern __inline int8x8_t
11812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11813 vbsl_s8 (uint8x8_t __a, int8x8_t __b, int8x8_t __c)
11814 {
11815   return __builtin_aarch64_simd_bslv8qi_suss (__a, __b, __c);
11816 }
11817
11818 __extension__ extern __inline int16x4_t
11819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11820 vbsl_s16 (uint16x4_t __a, int16x4_t __b, int16x4_t __c)
11821 {
11822   return __builtin_aarch64_simd_bslv4hi_suss (__a, __b, __c);
11823 }
11824
11825 __extension__ extern __inline int32x2_t
11826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11827 vbsl_s32 (uint32x2_t __a, int32x2_t __b, int32x2_t __c)
11828 {
11829   return __builtin_aarch64_simd_bslv2si_suss (__a, __b, __c);
11830 }
11831
11832 __extension__ extern __inline int64x1_t
11833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11834 vbsl_s64 (uint64x1_t __a, int64x1_t __b, int64x1_t __c)
11835 {
11836   return (int64x1_t)
11837       {__builtin_aarch64_simd_bsldi_suss (__a[0], __b[0], __c[0])};
11838 }
11839
11840 __extension__ extern __inline uint8x8_t
11841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11842 vbsl_u8 (uint8x8_t __a, uint8x8_t __b, uint8x8_t __c)
11843 {
11844   return __builtin_aarch64_simd_bslv8qi_uuuu (__a, __b, __c);
11845 }
11846
11847 __extension__ extern __inline uint16x4_t
11848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11849 vbsl_u16 (uint16x4_t __a, uint16x4_t __b, uint16x4_t __c)
11850 {
11851   return __builtin_aarch64_simd_bslv4hi_uuuu (__a, __b, __c);
11852 }
11853
11854 __extension__ extern __inline uint32x2_t
11855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11856 vbsl_u32 (uint32x2_t __a, uint32x2_t __b, uint32x2_t __c)
11857 {
11858   return __builtin_aarch64_simd_bslv2si_uuuu (__a, __b, __c);
11859 }
11860
11861 __extension__ extern __inline uint64x1_t
11862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11863 vbsl_u64 (uint64x1_t __a, uint64x1_t __b, uint64x1_t __c)
11864 {
11865   return (uint64x1_t)
11866       {__builtin_aarch64_simd_bsldi_uuuu (__a[0], __b[0], __c[0])};
11867 }
11868
11869 __extension__ extern __inline float16x8_t
11870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11871 vbslq_f16 (uint16x8_t __a, float16x8_t __b, float16x8_t __c)
11872 {
11873   return __builtin_aarch64_simd_bslv8hf_suss (__a, __b, __c);
11874 }
11875
11876 __extension__ extern __inline float32x4_t
11877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11878 vbslq_f32 (uint32x4_t __a, float32x4_t __b, float32x4_t __c)
11879 {
11880   return __builtin_aarch64_simd_bslv4sf_suss (__a, __b, __c);
11881 }
11882
11883 __extension__ extern __inline float64x2_t
11884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11885 vbslq_f64 (uint64x2_t __a, float64x2_t __b, float64x2_t __c)
11886 {
11887   return __builtin_aarch64_simd_bslv2df_suss (__a, __b, __c);
11888 }
11889
11890 __extension__ extern __inline poly8x16_t
11891 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11892 vbslq_p8 (uint8x16_t __a, poly8x16_t __b, poly8x16_t __c)
11893 {
11894   return __builtin_aarch64_simd_bslv16qi_pupp (__a, __b, __c);
11895 }
11896
11897 __extension__ extern __inline poly16x8_t
11898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11899 vbslq_p16 (uint16x8_t __a, poly16x8_t __b, poly16x8_t __c)
11900 {
11901   return __builtin_aarch64_simd_bslv8hi_pupp (__a, __b, __c);
11902 }
11903
11904 __extension__ extern __inline int8x16_t
11905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11906 vbslq_s8 (uint8x16_t __a, int8x16_t __b, int8x16_t __c)
11907 {
11908   return __builtin_aarch64_simd_bslv16qi_suss (__a, __b, __c);
11909 }
11910
11911 __extension__ extern __inline int16x8_t
11912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11913 vbslq_s16 (uint16x8_t __a, int16x8_t __b, int16x8_t __c)
11914 {
11915   return __builtin_aarch64_simd_bslv8hi_suss (__a, __b, __c);
11916 }
11917
11918 __extension__ extern __inline poly64x2_t
11919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11920 vbslq_p64 (uint64x2_t __a, poly64x2_t __b, poly64x2_t __c)
11921 {
11922   return __builtin_aarch64_simd_bslv2di_pupp (__a, __b, __c);
11923 }
11924
11925 __extension__ extern __inline int32x4_t
11926 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11927 vbslq_s32 (uint32x4_t __a, int32x4_t __b, int32x4_t __c)
11928 {
11929   return __builtin_aarch64_simd_bslv4si_suss (__a, __b, __c);
11930 }
11931
11932 __extension__ extern __inline int64x2_t
11933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11934 vbslq_s64 (uint64x2_t __a, int64x2_t __b, int64x2_t __c)
11935 {
11936   return __builtin_aarch64_simd_bslv2di_suss (__a, __b, __c);
11937 }
11938
11939 __extension__ extern __inline uint8x16_t
11940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11941 vbslq_u8 (uint8x16_t __a, uint8x16_t __b, uint8x16_t __c)
11942 {
11943   return __builtin_aarch64_simd_bslv16qi_uuuu (__a, __b, __c);
11944 }
11945
11946 __extension__ extern __inline uint16x8_t
11947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11948 vbslq_u16 (uint16x8_t __a, uint16x8_t __b, uint16x8_t __c)
11949 {
11950   return __builtin_aarch64_simd_bslv8hi_uuuu (__a, __b, __c);
11951 }
11952
11953 __extension__ extern __inline uint32x4_t
11954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11955 vbslq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
11956 {
11957   return __builtin_aarch64_simd_bslv4si_uuuu (__a, __b, __c);
11958 }
11959
11960 __extension__ extern __inline uint64x2_t
11961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11962 vbslq_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
11963 {
11964   return __builtin_aarch64_simd_bslv2di_uuuu (__a, __b, __c);
11965 }
11966
11967 /* ARMv8.1-A instrinsics.  */
11968 #pragma GCC push_options
11969 #pragma GCC target ("arch=armv8.1-a")
11970
11971 __extension__ extern __inline int16x4_t
11972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11973 vqrdmlah_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c)
11974 {
11975   return __builtin_aarch64_sqrdmlahv4hi (__a, __b, __c);
11976 }
11977
11978 __extension__ extern __inline int32x2_t
11979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11980 vqrdmlah_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c)
11981 {
11982   return __builtin_aarch64_sqrdmlahv2si (__a, __b, __c);
11983 }
11984
11985 __extension__ extern __inline int16x8_t
11986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11987 vqrdmlahq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
11988 {
11989   return __builtin_aarch64_sqrdmlahv8hi (__a, __b, __c);
11990 }
11991
11992 __extension__ extern __inline int32x4_t
11993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11994 vqrdmlahq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
11995 {
11996   return __builtin_aarch64_sqrdmlahv4si (__a, __b, __c);
11997 }
11998
11999 __extension__ extern __inline int16x4_t
12000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12001 vqrdmlsh_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c)
12002 {
12003   return __builtin_aarch64_sqrdmlshv4hi (__a, __b, __c);
12004 }
12005
12006 __extension__ extern __inline int32x2_t
12007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12008 vqrdmlsh_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c)
12009 {
12010   return __builtin_aarch64_sqrdmlshv2si (__a, __b, __c);
12011 }
12012
12013 __extension__ extern __inline int16x8_t
12014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12015 vqrdmlshq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
12016 {
12017   return __builtin_aarch64_sqrdmlshv8hi (__a, __b, __c);
12018 }
12019
12020 __extension__ extern __inline int32x4_t
12021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12022 vqrdmlshq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
12023 {
12024   return __builtin_aarch64_sqrdmlshv4si (__a, __b, __c);
12025 }
12026
12027 __extension__ extern __inline int16x4_t
12028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12029 vqrdmlah_laneq_s16 (int16x4_t __a, int16x4_t __b, int16x8_t __c, const int __d)
12030 {
12031   return  __builtin_aarch64_sqrdmlah_laneqv4hi (__a, __b, __c, __d);
12032 }
12033
12034 __extension__ extern __inline int32x2_t
12035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12036 vqrdmlah_laneq_s32 (int32x2_t __a, int32x2_t __b, int32x4_t __c, const int __d)
12037 {
12038   return __builtin_aarch64_sqrdmlah_laneqv2si (__a, __b, __c, __d);
12039 }
12040
12041 __extension__ extern __inline int16x8_t
12042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12043 vqrdmlahq_laneq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c, const int __d)
12044 {
12045   return __builtin_aarch64_sqrdmlah_laneqv8hi (__a, __b, __c, __d);
12046 }
12047
12048 __extension__ extern __inline int32x4_t
12049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12050 vqrdmlahq_laneq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c, const int __d)
12051 {
12052   return __builtin_aarch64_sqrdmlah_laneqv4si (__a, __b, __c, __d);
12053 }
12054
12055 __extension__ extern __inline int16x4_t
12056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12057 vqrdmlsh_laneq_s16 (int16x4_t __a, int16x4_t __b, int16x8_t __c, const int __d)
12058 {
12059   return  __builtin_aarch64_sqrdmlsh_laneqv4hi (__a, __b, __c, __d);
12060 }
12061
12062 __extension__ extern __inline int32x2_t
12063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12064 vqrdmlsh_laneq_s32 (int32x2_t __a, int32x2_t __b, int32x4_t __c, const int __d)
12065 {
12066   return __builtin_aarch64_sqrdmlsh_laneqv2si (__a, __b, __c, __d);
12067 }
12068
12069 __extension__ extern __inline int16x8_t
12070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12071 vqrdmlshq_laneq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c, const int __d)
12072 {
12073   return __builtin_aarch64_sqrdmlsh_laneqv8hi (__a, __b, __c, __d);
12074 }
12075
12076 __extension__ extern __inline int32x4_t
12077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12078 vqrdmlshq_laneq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c, const int __d)
12079 {
12080   return __builtin_aarch64_sqrdmlsh_laneqv4si (__a, __b, __c, __d);
12081 }
12082
12083 __extension__ extern __inline int16x4_t
12084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12085 vqrdmlah_lane_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c, const int __d)
12086 {
12087   return  __builtin_aarch64_sqrdmlah_lanev4hi (__a, __b, __c, __d);
12088 }
12089
12090 __extension__ extern __inline int32x2_t
12091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12092 vqrdmlah_lane_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c, const int __d)
12093 {
12094   return __builtin_aarch64_sqrdmlah_lanev2si (__a, __b, __c, __d);
12095 }
12096
12097 __extension__ extern __inline int16x8_t
12098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12099 vqrdmlahq_lane_s16 (int16x8_t __a, int16x8_t __b, int16x4_t __c, const int __d)
12100 {
12101   return __builtin_aarch64_sqrdmlah_lanev8hi (__a, __b, __c, __d);
12102 }
12103
12104 __extension__ extern __inline int32x4_t
12105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12106 vqrdmlahq_lane_s32 (int32x4_t __a, int32x4_t __b, int32x2_t __c, const int __d)
12107 {
12108   return __builtin_aarch64_sqrdmlah_lanev4si (__a, __b, __c, __d);
12109 }
12110
12111 __extension__ extern __inline int16_t
12112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12113 vqrdmlahh_s16 (int16_t __a, int16_t __b, int16_t __c)
12114 {
12115   return (int16_t) __builtin_aarch64_sqrdmlahhi (__a, __b, __c);
12116 }
12117
12118 __extension__ extern __inline int16_t
12119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12120 vqrdmlahh_lane_s16 (int16_t __a, int16_t __b, int16x4_t __c, const int __d)
12121 {
12122   return __builtin_aarch64_sqrdmlah_lanehi (__a, __b, __c, __d);
12123 }
12124
12125 __extension__ extern __inline int16_t
12126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12127 vqrdmlahh_laneq_s16 (int16_t __a, int16_t __b, int16x8_t __c, const int __d)
12128 {
12129   return __builtin_aarch64_sqrdmlah_laneqhi (__a, __b, __c, __d);
12130 }
12131
12132 __extension__ extern __inline int32_t
12133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12134 vqrdmlahs_s32 (int32_t __a, int32_t __b, int32_t __c)
12135 {
12136   return (int32_t) __builtin_aarch64_sqrdmlahsi (__a, __b, __c);
12137 }
12138
12139 __extension__ extern __inline int32_t
12140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12141 vqrdmlahs_lane_s32 (int32_t __a, int32_t __b, int32x2_t __c, const int __d)
12142 {
12143   return __builtin_aarch64_sqrdmlah_lanesi (__a, __b, __c, __d);
12144 }
12145
12146 __extension__ extern __inline int32_t
12147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12148 vqrdmlahs_laneq_s32 (int32_t __a, int32_t __b, int32x4_t __c, const int __d)
12149 {
12150   return __builtin_aarch64_sqrdmlah_laneqsi (__a, __b, __c, __d);
12151 }
12152
12153 __extension__ extern __inline int16x4_t
12154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12155 vqrdmlsh_lane_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c, const int __d)
12156 {
12157   return  __builtin_aarch64_sqrdmlsh_lanev4hi (__a, __b, __c, __d);
12158 }
12159
12160 __extension__ extern __inline int32x2_t
12161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12162 vqrdmlsh_lane_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c, const int __d)
12163 {
12164   return __builtin_aarch64_sqrdmlsh_lanev2si (__a, __b, __c, __d);
12165 }
12166
12167 __extension__ extern __inline int16x8_t
12168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12169 vqrdmlshq_lane_s16 (int16x8_t __a, int16x8_t __b, int16x4_t __c, const int __d)
12170 {
12171   return __builtin_aarch64_sqrdmlsh_lanev8hi (__a, __b, __c, __d);
12172 }
12173
12174 __extension__ extern __inline int32x4_t
12175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12176 vqrdmlshq_lane_s32 (int32x4_t __a, int32x4_t __b, int32x2_t __c, const int __d)
12177 {
12178   return __builtin_aarch64_sqrdmlsh_lanev4si (__a, __b, __c, __d);
12179 }
12180
12181 __extension__ extern __inline int16_t
12182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12183 vqrdmlshh_s16 (int16_t __a, int16_t __b, int16_t __c)
12184 {
12185   return (int16_t) __builtin_aarch64_sqrdmlshhi (__a, __b, __c);
12186 }
12187
12188 __extension__ extern __inline int16_t
12189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12190 vqrdmlshh_lane_s16 (int16_t __a, int16_t __b, int16x4_t __c, const int __d)
12191 {
12192   return __builtin_aarch64_sqrdmlsh_lanehi (__a, __b, __c, __d);
12193 }
12194
12195 __extension__ extern __inline int16_t
12196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12197 vqrdmlshh_laneq_s16 (int16_t __a, int16_t __b, int16x8_t __c, const int __d)
12198 {
12199   return __builtin_aarch64_sqrdmlsh_laneqhi (__a, __b, __c, __d);
12200 }
12201
12202 __extension__ extern __inline int32_t
12203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12204 vqrdmlshs_s32 (int32_t __a, int32_t __b, int32_t __c)
12205 {
12206   return (int32_t) __builtin_aarch64_sqrdmlshsi (__a, __b, __c);
12207 }
12208
12209 __extension__ extern __inline int32_t
12210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12211 vqrdmlshs_lane_s32 (int32_t __a, int32_t __b, int32x2_t __c, const int __d)
12212 {
12213   return __builtin_aarch64_sqrdmlsh_lanesi (__a, __b, __c, __d);
12214 }
12215
12216 __extension__ extern __inline int32_t
12217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12218 vqrdmlshs_laneq_s32 (int32_t __a, int32_t __b, int32x4_t __c, const int __d)
12219 {
12220   return __builtin_aarch64_sqrdmlsh_laneqsi (__a, __b, __c, __d);
12221 }
12222 #pragma GCC pop_options
12223
12224 #pragma GCC push_options
12225 #pragma GCC target ("+nothing+crypto")
12226 /* vaes  */
12227
12228 __extension__ extern __inline uint8x16_t
12229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12230 vaeseq_u8 (uint8x16_t data, uint8x16_t key)
12231 {
12232   return __builtin_aarch64_crypto_aesev16qi_uuu (data, key);
12233 }
12234
12235 __extension__ extern __inline uint8x16_t
12236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12237 vaesdq_u8 (uint8x16_t data, uint8x16_t key)
12238 {
12239   return __builtin_aarch64_crypto_aesdv16qi_uuu (data, key);
12240 }
12241
12242 __extension__ extern __inline uint8x16_t
12243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12244 vaesmcq_u8 (uint8x16_t data)
12245 {
12246   return __builtin_aarch64_crypto_aesmcv16qi_uu (data);
12247 }
12248
12249 __extension__ extern __inline uint8x16_t
12250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12251 vaesimcq_u8 (uint8x16_t data)
12252 {
12253   return __builtin_aarch64_crypto_aesimcv16qi_uu (data);
12254 }
12255 #pragma GCC pop_options
12256
12257 /* vcage  */
12258
12259 __extension__ extern __inline uint64x1_t
12260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12261 vcage_f64 (float64x1_t __a, float64x1_t __b)
12262 {
12263   return vabs_f64 (__a) >= vabs_f64 (__b);
12264 }
12265
12266 __extension__ extern __inline uint32_t
12267 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12268 vcages_f32 (float32_t __a, float32_t __b)
12269 {
12270   return __builtin_fabsf (__a) >= __builtin_fabsf (__b) ? -1 : 0;
12271 }
12272
12273 __extension__ extern __inline uint32x2_t
12274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12275 vcage_f32 (float32x2_t __a, float32x2_t __b)
12276 {
12277   return vabs_f32 (__a) >= vabs_f32 (__b);
12278 }
12279
12280 __extension__ extern __inline uint32x4_t
12281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12282 vcageq_f32 (float32x4_t __a, float32x4_t __b)
12283 {
12284   return vabsq_f32 (__a) >= vabsq_f32 (__b);
12285 }
12286
12287 __extension__ extern __inline uint64_t
12288 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12289 vcaged_f64 (float64_t __a, float64_t __b)
12290 {
12291   return __builtin_fabs (__a) >= __builtin_fabs (__b) ? -1 : 0;
12292 }
12293
12294 __extension__ extern __inline uint64x2_t
12295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12296 vcageq_f64 (float64x2_t __a, float64x2_t __b)
12297 {
12298   return vabsq_f64 (__a) >= vabsq_f64 (__b);
12299 }
12300
12301 /* vcagt  */
12302
12303 __extension__ extern __inline uint32_t
12304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12305 vcagts_f32 (float32_t __a, float32_t __b)
12306 {
12307   return __builtin_fabsf (__a) > __builtin_fabsf (__b) ? -1 : 0;
12308 }
12309
12310 __extension__ extern __inline uint32x2_t
12311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12312 vcagt_f32 (float32x2_t __a, float32x2_t __b)
12313 {
12314   return vabs_f32 (__a) > vabs_f32 (__b);
12315 }
12316
12317 __extension__ extern __inline uint64x1_t
12318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12319 vcagt_f64 (float64x1_t __a, float64x1_t __b)
12320 {
12321   return vabs_f64 (__a) > vabs_f64 (__b);
12322 }
12323
12324 __extension__ extern __inline uint32x4_t
12325 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12326 vcagtq_f32 (float32x4_t __a, float32x4_t __b)
12327 {
12328   return vabsq_f32 (__a) > vabsq_f32 (__b);
12329 }
12330
12331 __extension__ extern __inline uint64_t
12332 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12333 vcagtd_f64 (float64_t __a, float64_t __b)
12334 {
12335   return __builtin_fabs (__a) > __builtin_fabs (__b) ? -1 : 0;
12336 }
12337
12338 __extension__ extern __inline uint64x2_t
12339 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12340 vcagtq_f64 (float64x2_t __a, float64x2_t __b)
12341 {
12342   return vabsq_f64 (__a) > vabsq_f64 (__b);
12343 }
12344
12345 /* vcale  */
12346
12347 __extension__ extern __inline uint32x2_t
12348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12349 vcale_f32 (float32x2_t __a, float32x2_t __b)
12350 {
12351   return vabs_f32 (__a) <= vabs_f32 (__b);
12352 }
12353
12354 __extension__ extern __inline uint64x1_t
12355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12356 vcale_f64 (float64x1_t __a, float64x1_t __b)
12357 {
12358   return vabs_f64 (__a) <= vabs_f64 (__b);
12359 }
12360
12361 __extension__ extern __inline uint64_t
12362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12363 vcaled_f64 (float64_t __a, float64_t __b)
12364 {
12365   return __builtin_fabs (__a) <= __builtin_fabs (__b) ? -1 : 0;
12366 }
12367
12368 __extension__ extern __inline uint32_t
12369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12370 vcales_f32 (float32_t __a, float32_t __b)
12371 {
12372   return __builtin_fabsf (__a) <= __builtin_fabsf (__b) ? -1 : 0;
12373 }
12374
12375 __extension__ extern __inline uint32x4_t
12376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12377 vcaleq_f32 (float32x4_t __a, float32x4_t __b)
12378 {
12379   return vabsq_f32 (__a) <= vabsq_f32 (__b);
12380 }
12381
12382 __extension__ extern __inline uint64x2_t
12383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12384 vcaleq_f64 (float64x2_t __a, float64x2_t __b)
12385 {
12386   return vabsq_f64 (__a) <= vabsq_f64 (__b);
12387 }
12388
12389 /* vcalt  */
12390
12391 __extension__ extern __inline uint32x2_t
12392 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12393 vcalt_f32 (float32x2_t __a, float32x2_t __b)
12394 {
12395   return vabs_f32 (__a) < vabs_f32 (__b);
12396 }
12397
12398 __extension__ extern __inline uint64x1_t
12399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12400 vcalt_f64 (float64x1_t __a, float64x1_t __b)
12401 {
12402   return vabs_f64 (__a) < vabs_f64 (__b);
12403 }
12404
12405 __extension__ extern __inline uint64_t
12406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12407 vcaltd_f64 (float64_t __a, float64_t __b)
12408 {
12409   return __builtin_fabs (__a) < __builtin_fabs (__b) ? -1 : 0;
12410 }
12411
12412 __extension__ extern __inline uint32x4_t
12413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12414 vcaltq_f32 (float32x4_t __a, float32x4_t __b)
12415 {
12416   return vabsq_f32 (__a) < vabsq_f32 (__b);
12417 }
12418
12419 __extension__ extern __inline uint64x2_t
12420 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12421 vcaltq_f64 (float64x2_t __a, float64x2_t __b)
12422 {
12423   return vabsq_f64 (__a) < vabsq_f64 (__b);
12424 }
12425
12426 __extension__ extern __inline uint32_t
12427 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12428 vcalts_f32 (float32_t __a, float32_t __b)
12429 {
12430   return __builtin_fabsf (__a) < __builtin_fabsf (__b) ? -1 : 0;
12431 }
12432
12433 /* vceq - vector.  */
12434
12435 __extension__ extern __inline uint32x2_t
12436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12437 vceq_f32 (float32x2_t __a, float32x2_t __b)
12438 {
12439   return (uint32x2_t) (__a == __b);
12440 }
12441
12442 __extension__ extern __inline uint64x1_t
12443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12444 vceq_f64 (float64x1_t __a, float64x1_t __b)
12445 {
12446   return (uint64x1_t) (__a == __b);
12447 }
12448
12449 __extension__ extern __inline uint8x8_t
12450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12451 vceq_p8 (poly8x8_t __a, poly8x8_t __b)
12452 {
12453   return (uint8x8_t) (__a == __b);
12454 }
12455
12456 __extension__ extern __inline uint64x1_t
12457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12458 vceq_p64 (poly64x1_t __a, poly64x1_t __b)
12459 {
12460   return (uint64x1_t) (__a == __b);
12461 }
12462
12463 __extension__ extern __inline uint8x8_t
12464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12465 vceq_s8 (int8x8_t __a, int8x8_t __b)
12466 {
12467   return (uint8x8_t) (__a == __b);
12468 }
12469
12470 __extension__ extern __inline uint16x4_t
12471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12472 vceq_s16 (int16x4_t __a, int16x4_t __b)
12473 {
12474   return (uint16x4_t) (__a == __b);
12475 }
12476
12477 __extension__ extern __inline uint32x2_t
12478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12479 vceq_s32 (int32x2_t __a, int32x2_t __b)
12480 {
12481   return (uint32x2_t) (__a == __b);
12482 }
12483
12484 __extension__ extern __inline uint64x1_t
12485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12486 vceq_s64 (int64x1_t __a, int64x1_t __b)
12487 {
12488   return (uint64x1_t) (__a == __b);
12489 }
12490
12491 __extension__ extern __inline uint8x8_t
12492 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12493 vceq_u8 (uint8x8_t __a, uint8x8_t __b)
12494 {
12495   return (__a == __b);
12496 }
12497
12498 __extension__ extern __inline uint16x4_t
12499 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12500 vceq_u16 (uint16x4_t __a, uint16x4_t __b)
12501 {
12502   return (__a == __b);
12503 }
12504
12505 __extension__ extern __inline uint32x2_t
12506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12507 vceq_u32 (uint32x2_t __a, uint32x2_t __b)
12508 {
12509   return (__a == __b);
12510 }
12511
12512 __extension__ extern __inline uint64x1_t
12513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12514 vceq_u64 (uint64x1_t __a, uint64x1_t __b)
12515 {
12516   return (__a == __b);
12517 }
12518
12519 __extension__ extern __inline uint32x4_t
12520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12521 vceqq_f32 (float32x4_t __a, float32x4_t __b)
12522 {
12523   return (uint32x4_t) (__a == __b);
12524 }
12525
12526 __extension__ extern __inline uint64x2_t
12527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12528 vceqq_f64 (float64x2_t __a, float64x2_t __b)
12529 {
12530   return (uint64x2_t) (__a == __b);
12531 }
12532
12533 __extension__ extern __inline uint8x16_t
12534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12535 vceqq_p8 (poly8x16_t __a, poly8x16_t __b)
12536 {
12537   return (uint8x16_t) (__a == __b);
12538 }
12539
12540 __extension__ extern __inline uint8x16_t
12541 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12542 vceqq_s8 (int8x16_t __a, int8x16_t __b)
12543 {
12544   return (uint8x16_t) (__a == __b);
12545 }
12546
12547 __extension__ extern __inline uint16x8_t
12548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12549 vceqq_s16 (int16x8_t __a, int16x8_t __b)
12550 {
12551   return (uint16x8_t) (__a == __b);
12552 }
12553
12554 __extension__ extern __inline uint32x4_t
12555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12556 vceqq_s32 (int32x4_t __a, int32x4_t __b)
12557 {
12558   return (uint32x4_t) (__a == __b);
12559 }
12560
12561 __extension__ extern __inline uint64x2_t
12562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12563 vceqq_s64 (int64x2_t __a, int64x2_t __b)
12564 {
12565   return (uint64x2_t) (__a == __b);
12566 }
12567
12568 __extension__ extern __inline uint8x16_t
12569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12570 vceqq_u8 (uint8x16_t __a, uint8x16_t __b)
12571 {
12572   return (__a == __b);
12573 }
12574
12575 __extension__ extern __inline uint16x8_t
12576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12577 vceqq_u16 (uint16x8_t __a, uint16x8_t __b)
12578 {
12579   return (__a == __b);
12580 }
12581
12582 __extension__ extern __inline uint32x4_t
12583 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12584 vceqq_u32 (uint32x4_t __a, uint32x4_t __b)
12585 {
12586   return (__a == __b);
12587 }
12588
12589 __extension__ extern __inline uint64x2_t
12590 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12591 vceqq_u64 (uint64x2_t __a, uint64x2_t __b)
12592 {
12593   return (__a == __b);
12594 }
12595
12596 /* vceq - scalar.  */
12597
12598 __extension__ extern __inline uint32_t
12599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12600 vceqs_f32 (float32_t __a, float32_t __b)
12601 {
12602   return __a == __b ? -1 : 0;
12603 }
12604
12605 __extension__ extern __inline uint64_t
12606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12607 vceqd_s64 (int64_t __a, int64_t __b)
12608 {
12609   return __a == __b ? -1ll : 0ll;
12610 }
12611
12612 __extension__ extern __inline uint64_t
12613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12614 vceqd_u64 (uint64_t __a, uint64_t __b)
12615 {
12616   return __a == __b ? -1ll : 0ll;
12617 }
12618
12619 __extension__ extern __inline uint64_t
12620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12621 vceqd_f64 (float64_t __a, float64_t __b)
12622 {
12623   return __a == __b ? -1ll : 0ll;
12624 }
12625
12626 /* vceqz - vector.  */
12627
12628 __extension__ extern __inline uint32x2_t
12629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12630 vceqz_f32 (float32x2_t __a)
12631 {
12632   return (uint32x2_t) (__a == 0.0f);
12633 }
12634
12635 __extension__ extern __inline uint64x1_t
12636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12637 vceqz_f64 (float64x1_t __a)
12638 {
12639   return (uint64x1_t) (__a == (float64x1_t) {0.0});
12640 }
12641
12642 __extension__ extern __inline uint8x8_t
12643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12644 vceqz_p8 (poly8x8_t __a)
12645 {
12646   return (uint8x8_t) (__a == 0);
12647 }
12648
12649 __extension__ extern __inline uint8x8_t
12650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12651 vceqz_s8 (int8x8_t __a)
12652 {
12653   return (uint8x8_t) (__a == 0);
12654 }
12655
12656 __extension__ extern __inline uint16x4_t
12657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12658 vceqz_s16 (int16x4_t __a)
12659 {
12660   return (uint16x4_t) (__a == 0);
12661 }
12662
12663 __extension__ extern __inline uint32x2_t
12664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12665 vceqz_s32 (int32x2_t __a)
12666 {
12667   return (uint32x2_t) (__a == 0);
12668 }
12669
12670 __extension__ extern __inline uint64x1_t
12671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12672 vceqz_s64 (int64x1_t __a)
12673 {
12674   return (uint64x1_t) (__a == __AARCH64_INT64_C (0));
12675 }
12676
12677 __extension__ extern __inline uint8x8_t
12678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12679 vceqz_u8 (uint8x8_t __a)
12680 {
12681   return (__a == 0);
12682 }
12683
12684 __extension__ extern __inline uint16x4_t
12685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12686 vceqz_u16 (uint16x4_t __a)
12687 {
12688   return (__a == 0);
12689 }
12690
12691 __extension__ extern __inline uint32x2_t
12692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12693 vceqz_u32 (uint32x2_t __a)
12694 {
12695   return (__a == 0);
12696 }
12697
12698 __extension__ extern __inline uint64x1_t
12699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12700 vceqz_u64 (uint64x1_t __a)
12701 {
12702   return (__a == __AARCH64_UINT64_C (0));
12703 }
12704
12705 __extension__ extern __inline uint32x4_t
12706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12707 vceqzq_f32 (float32x4_t __a)
12708 {
12709   return (uint32x4_t) (__a == 0.0f);
12710 }
12711
12712 __extension__ extern __inline uint64x2_t
12713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12714 vceqzq_f64 (float64x2_t __a)
12715 {
12716   return (uint64x2_t) (__a == 0.0f);
12717 }
12718
12719 __extension__ extern __inline uint8x16_t
12720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12721 vceqzq_p8 (poly8x16_t __a)
12722 {
12723   return (uint8x16_t) (__a == 0);
12724 }
12725
12726 __extension__ extern __inline uint8x16_t
12727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12728 vceqzq_s8 (int8x16_t __a)
12729 {
12730   return (uint8x16_t) (__a == 0);
12731 }
12732
12733 __extension__ extern __inline uint16x8_t
12734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12735 vceqzq_s16 (int16x8_t __a)
12736 {
12737   return (uint16x8_t) (__a == 0);
12738 }
12739
12740 __extension__ extern __inline uint32x4_t
12741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12742 vceqzq_s32 (int32x4_t __a)
12743 {
12744   return (uint32x4_t) (__a == 0);
12745 }
12746
12747 __extension__ extern __inline uint64x2_t
12748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12749 vceqzq_s64 (int64x2_t __a)
12750 {
12751   return (uint64x2_t) (__a == __AARCH64_INT64_C (0));
12752 }
12753
12754 __extension__ extern __inline uint8x16_t
12755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12756 vceqzq_u8 (uint8x16_t __a)
12757 {
12758   return (__a == 0);
12759 }
12760
12761 __extension__ extern __inline uint16x8_t
12762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12763 vceqzq_u16 (uint16x8_t __a)
12764 {
12765   return (__a == 0);
12766 }
12767
12768 __extension__ extern __inline uint32x4_t
12769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12770 vceqzq_u32 (uint32x4_t __a)
12771 {
12772   return (__a == 0);
12773 }
12774
12775 __extension__ extern __inline uint64x2_t
12776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12777 vceqzq_u64 (uint64x2_t __a)
12778 {
12779   return (__a == __AARCH64_UINT64_C (0));
12780 }
12781
12782 /* vceqz - scalar.  */
12783
12784 __extension__ extern __inline uint32_t
12785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12786 vceqzs_f32 (float32_t __a)
12787 {
12788   return __a == 0.0f ? -1 : 0;
12789 }
12790
12791 __extension__ extern __inline uint64_t
12792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12793 vceqzd_s64 (int64_t __a)
12794 {
12795   return __a == 0 ? -1ll : 0ll;
12796 }
12797
12798 __extension__ extern __inline uint64_t
12799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12800 vceqzd_u64 (uint64_t __a)
12801 {
12802   return __a == 0 ? -1ll : 0ll;
12803 }
12804
12805 __extension__ extern __inline uint64_t
12806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12807 vceqzd_f64 (float64_t __a)
12808 {
12809   return __a == 0.0 ? -1ll : 0ll;
12810 }
12811
12812 /* vcge - vector.  */
12813
12814 __extension__ extern __inline uint32x2_t
12815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12816 vcge_f32 (float32x2_t __a, float32x2_t __b)
12817 {
12818   return (uint32x2_t) (__a >= __b);
12819 }
12820
12821 __extension__ extern __inline uint64x1_t
12822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12823 vcge_f64 (float64x1_t __a, float64x1_t __b)
12824 {
12825   return (uint64x1_t) (__a >= __b);
12826 }
12827
12828 __extension__ extern __inline uint8x8_t
12829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12830 vcge_s8 (int8x8_t __a, int8x8_t __b)
12831 {
12832   return (uint8x8_t) (__a >= __b);
12833 }
12834
12835 __extension__ extern __inline uint16x4_t
12836 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12837 vcge_s16 (int16x4_t __a, int16x4_t __b)
12838 {
12839   return (uint16x4_t) (__a >= __b);
12840 }
12841
12842 __extension__ extern __inline uint32x2_t
12843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12844 vcge_s32 (int32x2_t __a, int32x2_t __b)
12845 {
12846   return (uint32x2_t) (__a >= __b);
12847 }
12848
12849 __extension__ extern __inline uint64x1_t
12850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12851 vcge_s64 (int64x1_t __a, int64x1_t __b)
12852 {
12853   return (uint64x1_t) (__a >= __b);
12854 }
12855
12856 __extension__ extern __inline uint8x8_t
12857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12858 vcge_u8 (uint8x8_t __a, uint8x8_t __b)
12859 {
12860   return (__a >= __b);
12861 }
12862
12863 __extension__ extern __inline uint16x4_t
12864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12865 vcge_u16 (uint16x4_t __a, uint16x4_t __b)
12866 {
12867   return (__a >= __b);
12868 }
12869
12870 __extension__ extern __inline uint32x2_t
12871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12872 vcge_u32 (uint32x2_t __a, uint32x2_t __b)
12873 {
12874   return (__a >= __b);
12875 }
12876
12877 __extension__ extern __inline uint64x1_t
12878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12879 vcge_u64 (uint64x1_t __a, uint64x1_t __b)
12880 {
12881   return (__a >= __b);
12882 }
12883
12884 __extension__ extern __inline uint32x4_t
12885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12886 vcgeq_f32 (float32x4_t __a, float32x4_t __b)
12887 {
12888   return (uint32x4_t) (__a >= __b);
12889 }
12890
12891 __extension__ extern __inline uint64x2_t
12892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12893 vcgeq_f64 (float64x2_t __a, float64x2_t __b)
12894 {
12895   return (uint64x2_t) (__a >= __b);
12896 }
12897
12898 __extension__ extern __inline uint8x16_t
12899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12900 vcgeq_s8 (int8x16_t __a, int8x16_t __b)
12901 {
12902   return (uint8x16_t) (__a >= __b);
12903 }
12904
12905 __extension__ extern __inline uint16x8_t
12906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12907 vcgeq_s16 (int16x8_t __a, int16x8_t __b)
12908 {
12909   return (uint16x8_t) (__a >= __b);
12910 }
12911
12912 __extension__ extern __inline uint32x4_t
12913 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12914 vcgeq_s32 (int32x4_t __a, int32x4_t __b)
12915 {
12916   return (uint32x4_t) (__a >= __b);
12917 }
12918
12919 __extension__ extern __inline uint64x2_t
12920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12921 vcgeq_s64 (int64x2_t __a, int64x2_t __b)
12922 {
12923   return (uint64x2_t) (__a >= __b);
12924 }
12925
12926 __extension__ extern __inline uint8x16_t
12927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12928 vcgeq_u8 (uint8x16_t __a, uint8x16_t __b)
12929 {
12930   return (__a >= __b);
12931 }
12932
12933 __extension__ extern __inline uint16x8_t
12934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12935 vcgeq_u16 (uint16x8_t __a, uint16x8_t __b)
12936 {
12937   return (__a >= __b);
12938 }
12939
12940 __extension__ extern __inline uint32x4_t
12941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12942 vcgeq_u32 (uint32x4_t __a, uint32x4_t __b)
12943 {
12944   return (__a >= __b);
12945 }
12946
12947 __extension__ extern __inline uint64x2_t
12948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12949 vcgeq_u64 (uint64x2_t __a, uint64x2_t __b)
12950 {
12951   return (__a >= __b);
12952 }
12953
12954 /* vcge - scalar.  */
12955
12956 __extension__ extern __inline uint32_t
12957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12958 vcges_f32 (float32_t __a, float32_t __b)
12959 {
12960   return __a >= __b ? -1 : 0;
12961 }
12962
12963 __extension__ extern __inline uint64_t
12964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12965 vcged_s64 (int64_t __a, int64_t __b)
12966 {
12967   return __a >= __b ? -1ll : 0ll;
12968 }
12969
12970 __extension__ extern __inline uint64_t
12971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12972 vcged_u64 (uint64_t __a, uint64_t __b)
12973 {
12974   return __a >= __b ? -1ll : 0ll;
12975 }
12976
12977 __extension__ extern __inline uint64_t
12978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12979 vcged_f64 (float64_t __a, float64_t __b)
12980 {
12981   return __a >= __b ? -1ll : 0ll;
12982 }
12983
12984 /* vcgez - vector.  */
12985
12986 __extension__ extern __inline uint32x2_t
12987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12988 vcgez_f32 (float32x2_t __a)
12989 {
12990   return (uint32x2_t) (__a >= 0.0f);
12991 }
12992
12993 __extension__ extern __inline uint64x1_t
12994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12995 vcgez_f64 (float64x1_t __a)
12996 {
12997   return (uint64x1_t) (__a[0] >= (float64x1_t) {0.0});
12998 }
12999
13000 __extension__ extern __inline uint8x8_t
13001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13002 vcgez_s8 (int8x8_t __a)
13003 {
13004   return (uint8x8_t) (__a >= 0);
13005 }
13006
13007 __extension__ extern __inline uint16x4_t
13008 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13009 vcgez_s16 (int16x4_t __a)
13010 {
13011   return (uint16x4_t) (__a >= 0);
13012 }
13013
13014 __extension__ extern __inline uint32x2_t
13015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13016 vcgez_s32 (int32x2_t __a)
13017 {
13018   return (uint32x2_t) (__a >= 0);
13019 }
13020
13021 __extension__ extern __inline uint64x1_t
13022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13023 vcgez_s64 (int64x1_t __a)
13024 {
13025   return (uint64x1_t) (__a >= __AARCH64_INT64_C (0));
13026 }
13027
13028 __extension__ extern __inline uint32x4_t
13029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13030 vcgezq_f32 (float32x4_t __a)
13031 {
13032   return (uint32x4_t) (__a >= 0.0f);
13033 }
13034
13035 __extension__ extern __inline uint64x2_t
13036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13037 vcgezq_f64 (float64x2_t __a)
13038 {
13039   return (uint64x2_t) (__a >= 0.0);
13040 }
13041
13042 __extension__ extern __inline uint8x16_t
13043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13044 vcgezq_s8 (int8x16_t __a)
13045 {
13046   return (uint8x16_t) (__a >= 0);
13047 }
13048
13049 __extension__ extern __inline uint16x8_t
13050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13051 vcgezq_s16 (int16x8_t __a)
13052 {
13053   return (uint16x8_t) (__a >= 0);
13054 }
13055
13056 __extension__ extern __inline uint32x4_t
13057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13058 vcgezq_s32 (int32x4_t __a)
13059 {
13060   return (uint32x4_t) (__a >= 0);
13061 }
13062
13063 __extension__ extern __inline uint64x2_t
13064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13065 vcgezq_s64 (int64x2_t __a)
13066 {
13067   return (uint64x2_t) (__a >= __AARCH64_INT64_C (0));
13068 }
13069
13070 /* vcgez - scalar.  */
13071
13072 __extension__ extern __inline uint32_t
13073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13074 vcgezs_f32 (float32_t __a)
13075 {
13076   return __a >= 0.0f ? -1 : 0;
13077 }
13078
13079 __extension__ extern __inline uint64_t
13080 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13081 vcgezd_s64 (int64_t __a)
13082 {
13083   return __a >= 0 ? -1ll : 0ll;
13084 }
13085
13086 __extension__ extern __inline uint64_t
13087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13088 vcgezd_f64 (float64_t __a)
13089 {
13090   return __a >= 0.0 ? -1ll : 0ll;
13091 }
13092
13093 /* vcgt - vector.  */
13094
13095 __extension__ extern __inline uint32x2_t
13096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13097 vcgt_f32 (float32x2_t __a, float32x2_t __b)
13098 {
13099   return (uint32x2_t) (__a > __b);
13100 }
13101
13102 __extension__ extern __inline uint64x1_t
13103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13104 vcgt_f64 (float64x1_t __a, float64x1_t __b)
13105 {
13106   return (uint64x1_t) (__a > __b);
13107 }
13108
13109 __extension__ extern __inline uint8x8_t
13110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13111 vcgt_s8 (int8x8_t __a, int8x8_t __b)
13112 {
13113   return (uint8x8_t) (__a > __b);
13114 }
13115
13116 __extension__ extern __inline uint16x4_t
13117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13118 vcgt_s16 (int16x4_t __a, int16x4_t __b)
13119 {
13120   return (uint16x4_t) (__a > __b);
13121 }
13122
13123 __extension__ extern __inline uint32x2_t
13124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13125 vcgt_s32 (int32x2_t __a, int32x2_t __b)
13126 {
13127   return (uint32x2_t) (__a > __b);
13128 }
13129
13130 __extension__ extern __inline uint64x1_t
13131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13132 vcgt_s64 (int64x1_t __a, int64x1_t __b)
13133 {
13134   return (uint64x1_t) (__a > __b);
13135 }
13136
13137 __extension__ extern __inline uint8x8_t
13138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13139 vcgt_u8 (uint8x8_t __a, uint8x8_t __b)
13140 {
13141   return (__a > __b);
13142 }
13143
13144 __extension__ extern __inline uint16x4_t
13145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13146 vcgt_u16 (uint16x4_t __a, uint16x4_t __b)
13147 {
13148   return (__a > __b);
13149 }
13150
13151 __extension__ extern __inline uint32x2_t
13152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13153 vcgt_u32 (uint32x2_t __a, uint32x2_t __b)
13154 {
13155   return (__a > __b);
13156 }
13157
13158 __extension__ extern __inline uint64x1_t
13159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13160 vcgt_u64 (uint64x1_t __a, uint64x1_t __b)
13161 {
13162   return (__a > __b);
13163 }
13164
13165 __extension__ extern __inline uint32x4_t
13166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13167 vcgtq_f32 (float32x4_t __a, float32x4_t __b)
13168 {
13169   return (uint32x4_t) (__a > __b);
13170 }
13171
13172 __extension__ extern __inline uint64x2_t
13173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13174 vcgtq_f64 (float64x2_t __a, float64x2_t __b)
13175 {
13176   return (uint64x2_t) (__a > __b);
13177 }
13178
13179 __extension__ extern __inline uint8x16_t
13180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13181 vcgtq_s8 (int8x16_t __a, int8x16_t __b)
13182 {
13183   return (uint8x16_t) (__a > __b);
13184 }
13185
13186 __extension__ extern __inline uint16x8_t
13187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13188 vcgtq_s16 (int16x8_t __a, int16x8_t __b)
13189 {
13190   return (uint16x8_t) (__a > __b);
13191 }
13192
13193 __extension__ extern __inline uint32x4_t
13194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13195 vcgtq_s32 (int32x4_t __a, int32x4_t __b)
13196 {
13197   return (uint32x4_t) (__a > __b);
13198 }
13199
13200 __extension__ extern __inline uint64x2_t
13201 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13202 vcgtq_s64 (int64x2_t __a, int64x2_t __b)
13203 {
13204   return (uint64x2_t) (__a > __b);
13205 }
13206
13207 __extension__ extern __inline uint8x16_t
13208 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13209 vcgtq_u8 (uint8x16_t __a, uint8x16_t __b)
13210 {
13211   return (__a > __b);
13212 }
13213
13214 __extension__ extern __inline uint16x8_t
13215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13216 vcgtq_u16 (uint16x8_t __a, uint16x8_t __b)
13217 {
13218   return (__a > __b);
13219 }
13220
13221 __extension__ extern __inline uint32x4_t
13222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13223 vcgtq_u32 (uint32x4_t __a, uint32x4_t __b)
13224 {
13225   return (__a > __b);
13226 }
13227
13228 __extension__ extern __inline uint64x2_t
13229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13230 vcgtq_u64 (uint64x2_t __a, uint64x2_t __b)
13231 {
13232   return (__a > __b);
13233 }
13234
13235 /* vcgt - scalar.  */
13236
13237 __extension__ extern __inline uint32_t
13238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13239 vcgts_f32 (float32_t __a, float32_t __b)
13240 {
13241   return __a > __b ? -1 : 0;
13242 }
13243
13244 __extension__ extern __inline uint64_t
13245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13246 vcgtd_s64 (int64_t __a, int64_t __b)
13247 {
13248   return __a > __b ? -1ll : 0ll;
13249 }
13250
13251 __extension__ extern __inline uint64_t
13252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13253 vcgtd_u64 (uint64_t __a, uint64_t __b)
13254 {
13255   return __a > __b ? -1ll : 0ll;
13256 }
13257
13258 __extension__ extern __inline uint64_t
13259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13260 vcgtd_f64 (float64_t __a, float64_t __b)
13261 {
13262   return __a > __b ? -1ll : 0ll;
13263 }
13264
13265 /* vcgtz - vector.  */
13266
13267 __extension__ extern __inline uint32x2_t
13268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13269 vcgtz_f32 (float32x2_t __a)
13270 {
13271   return (uint32x2_t) (__a > 0.0f);
13272 }
13273
13274 __extension__ extern __inline uint64x1_t
13275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13276 vcgtz_f64 (float64x1_t __a)
13277 {
13278   return (uint64x1_t) (__a > (float64x1_t) {0.0});
13279 }
13280
13281 __extension__ extern __inline uint8x8_t
13282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13283 vcgtz_s8 (int8x8_t __a)
13284 {
13285   return (uint8x8_t) (__a > 0);
13286 }
13287
13288 __extension__ extern __inline uint16x4_t
13289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13290 vcgtz_s16 (int16x4_t __a)
13291 {
13292   return (uint16x4_t) (__a > 0);
13293 }
13294
13295 __extension__ extern __inline uint32x2_t
13296 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13297 vcgtz_s32 (int32x2_t __a)
13298 {
13299   return (uint32x2_t) (__a > 0);
13300 }
13301
13302 __extension__ extern __inline uint64x1_t
13303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13304 vcgtz_s64 (int64x1_t __a)
13305 {
13306   return (uint64x1_t) (__a > __AARCH64_INT64_C (0));
13307 }
13308
13309 __extension__ extern __inline uint32x4_t
13310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13311 vcgtzq_f32 (float32x4_t __a)
13312 {
13313   return (uint32x4_t) (__a > 0.0f);
13314 }
13315
13316 __extension__ extern __inline uint64x2_t
13317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13318 vcgtzq_f64 (float64x2_t __a)
13319 {
13320     return (uint64x2_t) (__a > 0.0);
13321 }
13322
13323 __extension__ extern __inline uint8x16_t
13324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13325 vcgtzq_s8 (int8x16_t __a)
13326 {
13327   return (uint8x16_t) (__a > 0);
13328 }
13329
13330 __extension__ extern __inline uint16x8_t
13331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13332 vcgtzq_s16 (int16x8_t __a)
13333 {
13334   return (uint16x8_t) (__a > 0);
13335 }
13336
13337 __extension__ extern __inline uint32x4_t
13338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13339 vcgtzq_s32 (int32x4_t __a)
13340 {
13341   return (uint32x4_t) (__a > 0);
13342 }
13343
13344 __extension__ extern __inline uint64x2_t
13345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13346 vcgtzq_s64 (int64x2_t __a)
13347 {
13348   return (uint64x2_t) (__a > __AARCH64_INT64_C (0));
13349 }
13350
13351 /* vcgtz - scalar.  */
13352
13353 __extension__ extern __inline uint32_t
13354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13355 vcgtzs_f32 (float32_t __a)
13356 {
13357   return __a > 0.0f ? -1 : 0;
13358 }
13359
13360 __extension__ extern __inline uint64_t
13361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13362 vcgtzd_s64 (int64_t __a)
13363 {
13364   return __a > 0 ? -1ll : 0ll;
13365 }
13366
13367 __extension__ extern __inline uint64_t
13368 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13369 vcgtzd_f64 (float64_t __a)
13370 {
13371   return __a > 0.0 ? -1ll : 0ll;
13372 }
13373
13374 /* vcle - vector.  */
13375
13376 __extension__ extern __inline uint32x2_t
13377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13378 vcle_f32 (float32x2_t __a, float32x2_t __b)
13379 {
13380   return (uint32x2_t) (__a <= __b);
13381 }
13382
13383 __extension__ extern __inline uint64x1_t
13384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13385 vcle_f64 (float64x1_t __a, float64x1_t __b)
13386 {
13387   return (uint64x1_t) (__a <= __b);
13388 }
13389
13390 __extension__ extern __inline uint8x8_t
13391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13392 vcle_s8 (int8x8_t __a, int8x8_t __b)
13393 {
13394   return (uint8x8_t) (__a <= __b);
13395 }
13396
13397 __extension__ extern __inline uint16x4_t
13398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13399 vcle_s16 (int16x4_t __a, int16x4_t __b)
13400 {
13401   return (uint16x4_t) (__a <= __b);
13402 }
13403
13404 __extension__ extern __inline uint32x2_t
13405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13406 vcle_s32 (int32x2_t __a, int32x2_t __b)
13407 {
13408   return (uint32x2_t) (__a <= __b);
13409 }
13410
13411 __extension__ extern __inline uint64x1_t
13412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13413 vcle_s64 (int64x1_t __a, int64x1_t __b)
13414 {
13415   return (uint64x1_t) (__a <= __b);
13416 }
13417
13418 __extension__ extern __inline uint8x8_t
13419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13420 vcle_u8 (uint8x8_t __a, uint8x8_t __b)
13421 {
13422   return (__a <= __b);
13423 }
13424
13425 __extension__ extern __inline uint16x4_t
13426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13427 vcle_u16 (uint16x4_t __a, uint16x4_t __b)
13428 {
13429   return (__a <= __b);
13430 }
13431
13432 __extension__ extern __inline uint32x2_t
13433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13434 vcle_u32 (uint32x2_t __a, uint32x2_t __b)
13435 {
13436   return (__a <= __b);
13437 }
13438
13439 __extension__ extern __inline uint64x1_t
13440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13441 vcle_u64 (uint64x1_t __a, uint64x1_t __b)
13442 {
13443   return (__a <= __b);
13444 }
13445
13446 __extension__ extern __inline uint32x4_t
13447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13448 vcleq_f32 (float32x4_t __a, float32x4_t __b)
13449 {
13450   return (uint32x4_t) (__a <= __b);
13451 }
13452
13453 __extension__ extern __inline uint64x2_t
13454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13455 vcleq_f64 (float64x2_t __a, float64x2_t __b)
13456 {
13457   return (uint64x2_t) (__a <= __b);
13458 }
13459
13460 __extension__ extern __inline uint8x16_t
13461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13462 vcleq_s8 (int8x16_t __a, int8x16_t __b)
13463 {
13464   return (uint8x16_t) (__a <= __b);
13465 }
13466
13467 __extension__ extern __inline uint16x8_t
13468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13469 vcleq_s16 (int16x8_t __a, int16x8_t __b)
13470 {
13471   return (uint16x8_t) (__a <= __b);
13472 }
13473
13474 __extension__ extern __inline uint32x4_t
13475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13476 vcleq_s32 (int32x4_t __a, int32x4_t __b)
13477 {
13478   return (uint32x4_t) (__a <= __b);
13479 }
13480
13481 __extension__ extern __inline uint64x2_t
13482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13483 vcleq_s64 (int64x2_t __a, int64x2_t __b)
13484 {
13485   return (uint64x2_t) (__a <= __b);
13486 }
13487
13488 __extension__ extern __inline uint8x16_t
13489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13490 vcleq_u8 (uint8x16_t __a, uint8x16_t __b)
13491 {
13492   return (__a <= __b);
13493 }
13494
13495 __extension__ extern __inline uint16x8_t
13496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13497 vcleq_u16 (uint16x8_t __a, uint16x8_t __b)
13498 {
13499   return (__a <= __b);
13500 }
13501
13502 __extension__ extern __inline uint32x4_t
13503 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13504 vcleq_u32 (uint32x4_t __a, uint32x4_t __b)
13505 {
13506   return (__a <= __b);
13507 }
13508
13509 __extension__ extern __inline uint64x2_t
13510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13511 vcleq_u64 (uint64x2_t __a, uint64x2_t __b)
13512 {
13513   return (__a <= __b);
13514 }
13515
13516 /* vcle - scalar.  */
13517
13518 __extension__ extern __inline uint32_t
13519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13520 vcles_f32 (float32_t __a, float32_t __b)
13521 {
13522   return __a <= __b ? -1 : 0;
13523 }
13524
13525 __extension__ extern __inline uint64_t
13526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13527 vcled_s64 (int64_t __a, int64_t __b)
13528 {
13529   return __a <= __b ? -1ll : 0ll;
13530 }
13531
13532 __extension__ extern __inline uint64_t
13533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13534 vcled_u64 (uint64_t __a, uint64_t __b)
13535 {
13536   return __a <= __b ? -1ll : 0ll;
13537 }
13538
13539 __extension__ extern __inline uint64_t
13540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13541 vcled_f64 (float64_t __a, float64_t __b)
13542 {
13543   return __a <= __b ? -1ll : 0ll;
13544 }
13545
13546 /* vclez - vector.  */
13547
13548 __extension__ extern __inline uint32x2_t
13549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13550 vclez_f32 (float32x2_t __a)
13551 {
13552   return (uint32x2_t) (__a <= 0.0f);
13553 }
13554
13555 __extension__ extern __inline uint64x1_t
13556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13557 vclez_f64 (float64x1_t __a)
13558 {
13559   return (uint64x1_t) (__a <= (float64x1_t) {0.0});
13560 }
13561
13562 __extension__ extern __inline uint8x8_t
13563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13564 vclez_s8 (int8x8_t __a)
13565 {
13566   return (uint8x8_t) (__a <= 0);
13567 }
13568
13569 __extension__ extern __inline uint16x4_t
13570 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13571 vclez_s16 (int16x4_t __a)
13572 {
13573   return (uint16x4_t) (__a <= 0);
13574 }
13575
13576 __extension__ extern __inline uint32x2_t
13577 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13578 vclez_s32 (int32x2_t __a)
13579 {
13580   return (uint32x2_t) (__a <= 0);
13581 }
13582
13583 __extension__ extern __inline uint64x1_t
13584 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13585 vclez_s64 (int64x1_t __a)
13586 {
13587   return (uint64x1_t) (__a <= __AARCH64_INT64_C (0));
13588 }
13589
13590 __extension__ extern __inline uint32x4_t
13591 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13592 vclezq_f32 (float32x4_t __a)
13593 {
13594   return (uint32x4_t) (__a <= 0.0f);
13595 }
13596
13597 __extension__ extern __inline uint64x2_t
13598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13599 vclezq_f64 (float64x2_t __a)
13600 {
13601   return (uint64x2_t) (__a <= 0.0);
13602 }
13603
13604 __extension__ extern __inline uint8x16_t
13605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13606 vclezq_s8 (int8x16_t __a)
13607 {
13608   return (uint8x16_t) (__a <= 0);
13609 }
13610
13611 __extension__ extern __inline uint16x8_t
13612 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13613 vclezq_s16 (int16x8_t __a)
13614 {
13615   return (uint16x8_t) (__a <= 0);
13616 }
13617
13618 __extension__ extern __inline uint32x4_t
13619 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13620 vclezq_s32 (int32x4_t __a)
13621 {
13622   return (uint32x4_t) (__a <= 0);
13623 }
13624
13625 __extension__ extern __inline uint64x2_t
13626 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13627 vclezq_s64 (int64x2_t __a)
13628 {
13629   return (uint64x2_t) (__a <= __AARCH64_INT64_C (0));
13630 }
13631
13632 /* vclez - scalar.  */
13633
13634 __extension__ extern __inline uint32_t
13635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13636 vclezs_f32 (float32_t __a)
13637 {
13638   return __a <= 0.0f ? -1 : 0;
13639 }
13640
13641 __extension__ extern __inline uint64_t
13642 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13643 vclezd_s64 (int64_t __a)
13644 {
13645   return __a <= 0 ? -1ll : 0ll;
13646 }
13647
13648 __extension__ extern __inline uint64_t
13649 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13650 vclezd_f64 (float64_t __a)
13651 {
13652   return __a <= 0.0 ? -1ll : 0ll;
13653 }
13654
13655 /* vclt - vector.  */
13656
13657 __extension__ extern __inline uint32x2_t
13658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13659 vclt_f32 (float32x2_t __a, float32x2_t __b)
13660 {
13661   return (uint32x2_t) (__a < __b);
13662 }
13663
13664 __extension__ extern __inline uint64x1_t
13665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13666 vclt_f64 (float64x1_t __a, float64x1_t __b)
13667 {
13668   return (uint64x1_t) (__a < __b);
13669 }
13670
13671 __extension__ extern __inline uint8x8_t
13672 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13673 vclt_s8 (int8x8_t __a, int8x8_t __b)
13674 {
13675   return (uint8x8_t) (__a < __b);
13676 }
13677
13678 __extension__ extern __inline uint16x4_t
13679 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13680 vclt_s16 (int16x4_t __a, int16x4_t __b)
13681 {
13682   return (uint16x4_t) (__a < __b);
13683 }
13684
13685 __extension__ extern __inline uint32x2_t
13686 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13687 vclt_s32 (int32x2_t __a, int32x2_t __b)
13688 {
13689   return (uint32x2_t) (__a < __b);
13690 }
13691
13692 __extension__ extern __inline uint64x1_t
13693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13694 vclt_s64 (int64x1_t __a, int64x1_t __b)
13695 {
13696   return (uint64x1_t) (__a < __b);
13697 }
13698
13699 __extension__ extern __inline uint8x8_t
13700 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13701 vclt_u8 (uint8x8_t __a, uint8x8_t __b)
13702 {
13703   return (__a < __b);
13704 }
13705
13706 __extension__ extern __inline uint16x4_t
13707 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13708 vclt_u16 (uint16x4_t __a, uint16x4_t __b)
13709 {
13710   return (__a < __b);
13711 }
13712
13713 __extension__ extern __inline uint32x2_t
13714 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13715 vclt_u32 (uint32x2_t __a, uint32x2_t __b)
13716 {
13717   return (__a < __b);
13718 }
13719
13720 __extension__ extern __inline uint64x1_t
13721 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13722 vclt_u64 (uint64x1_t __a, uint64x1_t __b)
13723 {
13724   return (__a < __b);
13725 }
13726
13727 __extension__ extern __inline uint32x4_t
13728 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13729 vcltq_f32 (float32x4_t __a, float32x4_t __b)
13730 {
13731   return (uint32x4_t) (__a < __b);
13732 }
13733
13734 __extension__ extern __inline uint64x2_t
13735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13736 vcltq_f64 (float64x2_t __a, float64x2_t __b)
13737 {
13738   return (uint64x2_t) (__a < __b);
13739 }
13740
13741 __extension__ extern __inline uint8x16_t
13742 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13743 vcltq_s8 (int8x16_t __a, int8x16_t __b)
13744 {
13745   return (uint8x16_t) (__a < __b);
13746 }
13747
13748 __extension__ extern __inline uint16x8_t
13749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13750 vcltq_s16 (int16x8_t __a, int16x8_t __b)
13751 {
13752   return (uint16x8_t) (__a < __b);
13753 }
13754
13755 __extension__ extern __inline uint32x4_t
13756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13757 vcltq_s32 (int32x4_t __a, int32x4_t __b)
13758 {
13759   return (uint32x4_t) (__a < __b);
13760 }
13761
13762 __extension__ extern __inline uint64x2_t
13763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13764 vcltq_s64 (int64x2_t __a, int64x2_t __b)
13765 {
13766   return (uint64x2_t) (__a < __b);
13767 }
13768
13769 __extension__ extern __inline uint8x16_t
13770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13771 vcltq_u8 (uint8x16_t __a, uint8x16_t __b)
13772 {
13773   return (__a < __b);
13774 }
13775
13776 __extension__ extern __inline uint16x8_t
13777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13778 vcltq_u16 (uint16x8_t __a, uint16x8_t __b)
13779 {
13780   return (__a < __b);
13781 }
13782
13783 __extension__ extern __inline uint32x4_t
13784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13785 vcltq_u32 (uint32x4_t __a, uint32x4_t __b)
13786 {
13787   return (__a < __b);
13788 }
13789
13790 __extension__ extern __inline uint64x2_t
13791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13792 vcltq_u64 (uint64x2_t __a, uint64x2_t __b)
13793 {
13794   return (__a < __b);
13795 }
13796
13797 /* vclt - scalar.  */
13798
13799 __extension__ extern __inline uint32_t
13800 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13801 vclts_f32 (float32_t __a, float32_t __b)
13802 {
13803   return __a < __b ? -1 : 0;
13804 }
13805
13806 __extension__ extern __inline uint64_t
13807 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13808 vcltd_s64 (int64_t __a, int64_t __b)
13809 {
13810   return __a < __b ? -1ll : 0ll;
13811 }
13812
13813 __extension__ extern __inline uint64_t
13814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13815 vcltd_u64 (uint64_t __a, uint64_t __b)
13816 {
13817   return __a < __b ? -1ll : 0ll;
13818 }
13819
13820 __extension__ extern __inline uint64_t
13821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13822 vcltd_f64 (float64_t __a, float64_t __b)
13823 {
13824   return __a < __b ? -1ll : 0ll;
13825 }
13826
13827 /* vcltz - vector.  */
13828
13829 __extension__ extern __inline uint32x2_t
13830 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13831 vcltz_f32 (float32x2_t __a)
13832 {
13833   return (uint32x2_t) (__a < 0.0f);
13834 }
13835
13836 __extension__ extern __inline uint64x1_t
13837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13838 vcltz_f64 (float64x1_t __a)
13839 {
13840   return (uint64x1_t) (__a < (float64x1_t) {0.0});
13841 }
13842
13843 __extension__ extern __inline uint8x8_t
13844 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13845 vcltz_s8 (int8x8_t __a)
13846 {
13847   return (uint8x8_t) (__a < 0);
13848 }
13849
13850 __extension__ extern __inline uint16x4_t
13851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13852 vcltz_s16 (int16x4_t __a)
13853 {
13854   return (uint16x4_t) (__a < 0);
13855 }
13856
13857 __extension__ extern __inline uint32x2_t
13858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13859 vcltz_s32 (int32x2_t __a)
13860 {
13861   return (uint32x2_t) (__a < 0);
13862 }
13863
13864 __extension__ extern __inline uint64x1_t
13865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13866 vcltz_s64 (int64x1_t __a)
13867 {
13868   return (uint64x1_t) (__a < __AARCH64_INT64_C (0));
13869 }
13870
13871 __extension__ extern __inline uint32x4_t
13872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13873 vcltzq_f32 (float32x4_t __a)
13874 {
13875   return (uint32x4_t) (__a < 0.0f);
13876 }
13877
13878 __extension__ extern __inline uint64x2_t
13879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13880 vcltzq_f64 (float64x2_t __a)
13881 {
13882   return (uint64x2_t) (__a < 0.0);
13883 }
13884
13885 __extension__ extern __inline uint8x16_t
13886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13887 vcltzq_s8 (int8x16_t __a)
13888 {
13889   return (uint8x16_t) (__a < 0);
13890 }
13891
13892 __extension__ extern __inline uint16x8_t
13893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13894 vcltzq_s16 (int16x8_t __a)
13895 {
13896   return (uint16x8_t) (__a < 0);
13897 }
13898
13899 __extension__ extern __inline uint32x4_t
13900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13901 vcltzq_s32 (int32x4_t __a)
13902 {
13903   return (uint32x4_t) (__a < 0);
13904 }
13905
13906 __extension__ extern __inline uint64x2_t
13907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13908 vcltzq_s64 (int64x2_t __a)
13909 {
13910   return (uint64x2_t) (__a < __AARCH64_INT64_C (0));
13911 }
13912
13913 /* vcltz - scalar.  */
13914
13915 __extension__ extern __inline uint32_t
13916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13917 vcltzs_f32 (float32_t __a)
13918 {
13919   return __a < 0.0f ? -1 : 0;
13920 }
13921
13922 __extension__ extern __inline uint64_t
13923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13924 vcltzd_s64 (int64_t __a)
13925 {
13926   return __a < 0 ? -1ll : 0ll;
13927 }
13928
13929 __extension__ extern __inline uint64_t
13930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13931 vcltzd_f64 (float64_t __a)
13932 {
13933   return __a < 0.0 ? -1ll : 0ll;
13934 }
13935
13936 /* vcls.  */
13937
13938 __extension__ extern __inline int8x8_t
13939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13940 vcls_s8 (int8x8_t __a)
13941 {
13942   return __builtin_aarch64_clrsbv8qi (__a);
13943 }
13944
13945 __extension__ extern __inline int16x4_t
13946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13947 vcls_s16 (int16x4_t __a)
13948 {
13949   return __builtin_aarch64_clrsbv4hi (__a);
13950 }
13951
13952 __extension__ extern __inline int32x2_t
13953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13954 vcls_s32 (int32x2_t __a)
13955 {
13956   return __builtin_aarch64_clrsbv2si (__a);
13957 }
13958
13959 __extension__ extern __inline int8x16_t
13960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13961 vclsq_s8 (int8x16_t __a)
13962 {
13963   return __builtin_aarch64_clrsbv16qi (__a);
13964 }
13965
13966 __extension__ extern __inline int16x8_t
13967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13968 vclsq_s16 (int16x8_t __a)
13969 {
13970   return __builtin_aarch64_clrsbv8hi (__a);
13971 }
13972
13973 __extension__ extern __inline int32x4_t
13974 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13975 vclsq_s32 (int32x4_t __a)
13976 {
13977   return __builtin_aarch64_clrsbv4si (__a);
13978 }
13979
13980 /* vclz.  */
13981
13982 __extension__ extern __inline int8x8_t
13983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13984 vclz_s8 (int8x8_t __a)
13985 {
13986   return __builtin_aarch64_clzv8qi (__a);
13987 }
13988
13989 __extension__ extern __inline int16x4_t
13990 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13991 vclz_s16 (int16x4_t __a)
13992 {
13993   return __builtin_aarch64_clzv4hi (__a);
13994 }
13995
13996 __extension__ extern __inline int32x2_t
13997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13998 vclz_s32 (int32x2_t __a)
13999 {
14000   return __builtin_aarch64_clzv2si (__a);
14001 }
14002
14003 __extension__ extern __inline uint8x8_t
14004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14005 vclz_u8 (uint8x8_t __a)
14006 {
14007   return (uint8x8_t)__builtin_aarch64_clzv8qi ((int8x8_t)__a);
14008 }
14009
14010 __extension__ extern __inline uint16x4_t
14011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14012 vclz_u16 (uint16x4_t __a)
14013 {
14014   return (uint16x4_t)__builtin_aarch64_clzv4hi ((int16x4_t)__a);
14015 }
14016
14017 __extension__ extern __inline uint32x2_t
14018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14019 vclz_u32 (uint32x2_t __a)
14020 {
14021   return (uint32x2_t)__builtin_aarch64_clzv2si ((int32x2_t)__a);
14022 }
14023
14024 __extension__ extern __inline int8x16_t
14025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14026 vclzq_s8 (int8x16_t __a)
14027 {
14028   return __builtin_aarch64_clzv16qi (__a);
14029 }
14030
14031 __extension__ extern __inline int16x8_t
14032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14033 vclzq_s16 (int16x8_t __a)
14034 {
14035   return __builtin_aarch64_clzv8hi (__a);
14036 }
14037
14038 __extension__ extern __inline int32x4_t
14039 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14040 vclzq_s32 (int32x4_t __a)
14041 {
14042   return __builtin_aarch64_clzv4si (__a);
14043 }
14044
14045 __extension__ extern __inline uint8x16_t
14046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14047 vclzq_u8 (uint8x16_t __a)
14048 {
14049   return (uint8x16_t)__builtin_aarch64_clzv16qi ((int8x16_t)__a);
14050 }
14051
14052 __extension__ extern __inline uint16x8_t
14053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14054 vclzq_u16 (uint16x8_t __a)
14055 {
14056   return (uint16x8_t)__builtin_aarch64_clzv8hi ((int16x8_t)__a);
14057 }
14058
14059 __extension__ extern __inline uint32x4_t
14060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14061 vclzq_u32 (uint32x4_t __a)
14062 {
14063   return (uint32x4_t)__builtin_aarch64_clzv4si ((int32x4_t)__a);
14064 }
14065
14066 /* vcnt.  */
14067
14068 __extension__ extern __inline poly8x8_t
14069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14070 vcnt_p8 (poly8x8_t __a)
14071 {
14072   return (poly8x8_t) __builtin_aarch64_popcountv8qi ((int8x8_t) __a);
14073 }
14074
14075 __extension__ extern __inline int8x8_t
14076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14077 vcnt_s8 (int8x8_t __a)
14078 {
14079   return __builtin_aarch64_popcountv8qi (__a);
14080 }
14081
14082 __extension__ extern __inline uint8x8_t
14083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14084 vcnt_u8 (uint8x8_t __a)
14085 {
14086   return (uint8x8_t) __builtin_aarch64_popcountv8qi ((int8x8_t) __a);
14087 }
14088
14089 __extension__ extern __inline poly8x16_t
14090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14091 vcntq_p8 (poly8x16_t __a)
14092 {
14093   return (poly8x16_t) __builtin_aarch64_popcountv16qi ((int8x16_t) __a);
14094 }
14095
14096 __extension__ extern __inline int8x16_t
14097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14098 vcntq_s8 (int8x16_t __a)
14099 {
14100   return __builtin_aarch64_popcountv16qi (__a);
14101 }
14102
14103 __extension__ extern __inline uint8x16_t
14104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14105 vcntq_u8 (uint8x16_t __a)
14106 {
14107   return (uint8x16_t) __builtin_aarch64_popcountv16qi ((int8x16_t) __a);
14108 }
14109
14110 /* vcopy_lane.  */
14111
14112 __extension__ extern __inline float32x2_t
14113 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14114 vcopy_lane_f32 (float32x2_t __a, const int __lane1,
14115                 float32x2_t __b, const int __lane2)
14116 {
14117   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14118                                   __a, __lane1);
14119 }
14120
14121 __extension__ extern __inline float64x1_t
14122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14123 vcopy_lane_f64 (float64x1_t __a, const int __lane1,
14124                 float64x1_t __b, const int __lane2)
14125 {
14126   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14127                                   __a, __lane1);
14128 }
14129
14130 __extension__ extern __inline poly8x8_t
14131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14132 vcopy_lane_p8 (poly8x8_t __a, const int __lane1,
14133                poly8x8_t __b, const int __lane2)
14134 {
14135   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14136                                  __a, __lane1);
14137 }
14138
14139 __extension__ extern __inline poly16x4_t
14140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14141 vcopy_lane_p16 (poly16x4_t __a, const int __lane1,
14142                 poly16x4_t __b, const int __lane2)
14143 {
14144   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14145                                   __a, __lane1);
14146 }
14147
14148 __extension__ extern __inline poly64x1_t
14149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14150 vcopy_lane_p64 (poly64x1_t __a, const int __lane1,
14151                 poly64x1_t __b, const int __lane2)
14152 {
14153   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14154                                   __a, __lane1);
14155 }
14156
14157 __extension__ extern __inline int8x8_t
14158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14159 vcopy_lane_s8 (int8x8_t __a, const int __lane1,
14160                int8x8_t __b, const int __lane2)
14161 {
14162   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14163                                  __a, __lane1);
14164 }
14165
14166 __extension__ extern __inline int16x4_t
14167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14168 vcopy_lane_s16 (int16x4_t __a, const int __lane1,
14169                 int16x4_t __b, const int __lane2)
14170 {
14171   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14172                                   __a, __lane1);
14173 }
14174
14175 __extension__ extern __inline int32x2_t
14176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14177 vcopy_lane_s32 (int32x2_t __a, const int __lane1,
14178                 int32x2_t __b, const int __lane2)
14179 {
14180   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14181                                   __a, __lane1);
14182 }
14183
14184 __extension__ extern __inline int64x1_t
14185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14186 vcopy_lane_s64 (int64x1_t __a, const int __lane1,
14187                 int64x1_t __b, const int __lane2)
14188 {
14189   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14190                                   __a, __lane1);
14191 }
14192
14193 __extension__ extern __inline uint8x8_t
14194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14195 vcopy_lane_u8 (uint8x8_t __a, const int __lane1,
14196                uint8x8_t __b, const int __lane2)
14197 {
14198   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14199                                  __a, __lane1);
14200 }
14201
14202 __extension__ extern __inline uint16x4_t
14203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14204 vcopy_lane_u16 (uint16x4_t __a, const int __lane1,
14205                 uint16x4_t __b, const int __lane2)
14206 {
14207   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14208                                   __a, __lane1);
14209 }
14210
14211 __extension__ extern __inline uint32x2_t
14212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14213 vcopy_lane_u32 (uint32x2_t __a, const int __lane1,
14214                 uint32x2_t __b, const int __lane2)
14215 {
14216   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14217                                   __a, __lane1);
14218 }
14219
14220 __extension__ extern __inline uint64x1_t
14221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14222 vcopy_lane_u64 (uint64x1_t __a, const int __lane1,
14223                 uint64x1_t __b, const int __lane2)
14224 {
14225   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14226                                   __a, __lane1);
14227 }
14228
14229 /* vcopy_laneq.  */
14230
14231 __extension__ extern __inline float32x2_t
14232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14233 vcopy_laneq_f32 (float32x2_t __a, const int __lane1,
14234                  float32x4_t __b, const int __lane2)
14235 {
14236   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14237                                   __a, __lane1);
14238 }
14239
14240 __extension__ extern __inline float64x1_t
14241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14242 vcopy_laneq_f64 (float64x1_t __a, const int __lane1,
14243                  float64x2_t __b, const int __lane2)
14244 {
14245   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14246                                   __a, __lane1);
14247 }
14248
14249 __extension__ extern __inline poly8x8_t
14250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14251 vcopy_laneq_p8 (poly8x8_t __a, const int __lane1,
14252                 poly8x16_t __b, const int __lane2)
14253 {
14254   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14255                                  __a, __lane1);
14256 }
14257
14258 __extension__ extern __inline poly16x4_t
14259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14260 vcopy_laneq_p16 (poly16x4_t __a, const int __lane1,
14261                  poly16x8_t __b, const int __lane2)
14262 {
14263   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14264                                   __a, __lane1);
14265 }
14266
14267 __extension__ extern __inline poly64x1_t
14268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14269 vcopy_laneq_p64 (poly64x1_t __a, const int __lane1,
14270                  poly64x2_t __b, const int __lane2)
14271 {
14272   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14273                                   __a, __lane1);
14274 }
14275
14276 __extension__ extern __inline int8x8_t
14277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14278 vcopy_laneq_s8 (int8x8_t __a, const int __lane1,
14279                 int8x16_t __b, const int __lane2)
14280 {
14281   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14282                                  __a, __lane1);
14283 }
14284
14285 __extension__ extern __inline int16x4_t
14286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14287 vcopy_laneq_s16 (int16x4_t __a, const int __lane1,
14288                  int16x8_t __b, const int __lane2)
14289 {
14290   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14291                                   __a, __lane1);
14292 }
14293
14294 __extension__ extern __inline int32x2_t
14295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14296 vcopy_laneq_s32 (int32x2_t __a, const int __lane1,
14297                  int32x4_t __b, const int __lane2)
14298 {
14299   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14300                                   __a, __lane1);
14301 }
14302
14303 __extension__ extern __inline int64x1_t
14304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14305 vcopy_laneq_s64 (int64x1_t __a, const int __lane1,
14306                  int64x2_t __b, const int __lane2)
14307 {
14308   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14309                                   __a, __lane1);
14310 }
14311
14312 __extension__ extern __inline uint8x8_t
14313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14314 vcopy_laneq_u8 (uint8x8_t __a, const int __lane1,
14315                 uint8x16_t __b, const int __lane2)
14316 {
14317   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14318                                  __a, __lane1);
14319 }
14320
14321 __extension__ extern __inline uint16x4_t
14322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14323 vcopy_laneq_u16 (uint16x4_t __a, const int __lane1,
14324                  uint16x8_t __b, const int __lane2)
14325 {
14326   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14327                                   __a, __lane1);
14328 }
14329
14330 __extension__ extern __inline uint32x2_t
14331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14332 vcopy_laneq_u32 (uint32x2_t __a, const int __lane1,
14333                  uint32x4_t __b, const int __lane2)
14334 {
14335   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14336                                   __a, __lane1);
14337 }
14338
14339 __extension__ extern __inline uint64x1_t
14340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14341 vcopy_laneq_u64 (uint64x1_t __a, const int __lane1,
14342                  uint64x2_t __b, const int __lane2)
14343 {
14344   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14345                                   __a, __lane1);
14346 }
14347
14348 /* vcopyq_lane.  */
14349
14350 __extension__ extern __inline float32x4_t
14351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14352 vcopyq_lane_f32 (float32x4_t __a, const int __lane1,
14353                  float32x2_t __b, const int __lane2)
14354 {
14355   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14356                                    __a, __lane1);
14357 }
14358
14359 __extension__ extern __inline float64x2_t
14360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14361 vcopyq_lane_f64 (float64x2_t __a, const int __lane1,
14362                  float64x1_t __b, const int __lane2)
14363 {
14364   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14365                                    __a, __lane1);
14366 }
14367
14368 __extension__ extern __inline poly8x16_t
14369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14370 vcopyq_lane_p8 (poly8x16_t __a, const int __lane1,
14371                 poly8x8_t __b, const int __lane2)
14372 {
14373   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14374                                   __a, __lane1);
14375 }
14376
14377 __extension__ extern __inline poly16x8_t
14378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14379 vcopyq_lane_p16 (poly16x8_t __a, const int __lane1,
14380                  poly16x4_t __b, const int __lane2)
14381 {
14382   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14383                                    __a, __lane1);
14384 }
14385
14386 __extension__ extern __inline poly64x2_t
14387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14388 vcopyq_lane_p64 (poly64x2_t __a, const int __lane1,
14389                  poly64x1_t __b, const int __lane2)
14390 {
14391   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14392                                    __a, __lane1);
14393 }
14394
14395 __extension__ extern __inline int8x16_t
14396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14397 vcopyq_lane_s8 (int8x16_t __a, const int __lane1,
14398                 int8x8_t __b, const int __lane2)
14399 {
14400   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14401                                   __a, __lane1);
14402 }
14403
14404 __extension__ extern __inline int16x8_t
14405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14406 vcopyq_lane_s16 (int16x8_t __a, const int __lane1,
14407                  int16x4_t __b, const int __lane2)
14408 {
14409   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14410                                    __a, __lane1);
14411 }
14412
14413 __extension__ extern __inline int32x4_t
14414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14415 vcopyq_lane_s32 (int32x4_t __a, const int __lane1,
14416                  int32x2_t __b, const int __lane2)
14417 {
14418   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14419                                    __a, __lane1);
14420 }
14421
14422 __extension__ extern __inline int64x2_t
14423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14424 vcopyq_lane_s64 (int64x2_t __a, const int __lane1,
14425                  int64x1_t __b, const int __lane2)
14426 {
14427   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14428                                    __a, __lane1);
14429 }
14430
14431 __extension__ extern __inline uint8x16_t
14432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14433 vcopyq_lane_u8 (uint8x16_t __a, const int __lane1,
14434                 uint8x8_t __b, const int __lane2)
14435 {
14436   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14437                                   __a, __lane1);
14438 }
14439
14440 __extension__ extern __inline uint16x8_t
14441 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14442 vcopyq_lane_u16 (uint16x8_t __a, const int __lane1,
14443                  uint16x4_t __b, const int __lane2)
14444 {
14445   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14446                                    __a, __lane1);
14447 }
14448
14449 __extension__ extern __inline uint32x4_t
14450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14451 vcopyq_lane_u32 (uint32x4_t __a, const int __lane1,
14452                  uint32x2_t __b, const int __lane2)
14453 {
14454   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14455                                    __a, __lane1);
14456 }
14457
14458 __extension__ extern __inline uint64x2_t
14459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14460 vcopyq_lane_u64 (uint64x2_t __a, const int __lane1,
14461                  uint64x1_t __b, const int __lane2)
14462 {
14463   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14464                                    __a, __lane1);
14465 }
14466
14467 /* vcopyq_laneq.  */
14468
14469 __extension__ extern __inline float32x4_t
14470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14471 vcopyq_laneq_f32 (float32x4_t __a, const int __lane1,
14472                   float32x4_t __b, const int __lane2)
14473 {
14474   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14475                                    __a, __lane1);
14476 }
14477
14478 __extension__ extern __inline float64x2_t
14479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14480 vcopyq_laneq_f64 (float64x2_t __a, const int __lane1,
14481                   float64x2_t __b, const int __lane2)
14482 {
14483   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14484                                    __a, __lane1);
14485 }
14486
14487 __extension__ extern __inline poly8x16_t
14488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14489 vcopyq_laneq_p8 (poly8x16_t __a, const int __lane1,
14490                  poly8x16_t __b, const int __lane2)
14491 {
14492   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14493                                   __a, __lane1);
14494 }
14495
14496 __extension__ extern __inline poly16x8_t
14497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14498 vcopyq_laneq_p16 (poly16x8_t __a, const int __lane1,
14499                   poly16x8_t __b, const int __lane2)
14500 {
14501   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14502                                    __a, __lane1);
14503 }
14504
14505 __extension__ extern __inline poly64x2_t
14506 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14507 vcopyq_laneq_p64 (poly64x2_t __a, const int __lane1,
14508                   poly64x2_t __b, const int __lane2)
14509 {
14510   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14511                                    __a, __lane1);
14512 }
14513
14514 __extension__ extern __inline int8x16_t
14515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14516 vcopyq_laneq_s8 (int8x16_t __a, const int __lane1,
14517                  int8x16_t __b, const int __lane2)
14518 {
14519   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14520                                   __a, __lane1);
14521 }
14522
14523 __extension__ extern __inline int16x8_t
14524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14525 vcopyq_laneq_s16 (int16x8_t __a, const int __lane1,
14526                   int16x8_t __b, const int __lane2)
14527 {
14528   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14529                                    __a, __lane1);
14530 }
14531
14532 __extension__ extern __inline int32x4_t
14533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14534 vcopyq_laneq_s32 (int32x4_t __a, const int __lane1,
14535                   int32x4_t __b, const int __lane2)
14536 {
14537   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14538                                    __a, __lane1);
14539 }
14540
14541 __extension__ extern __inline int64x2_t
14542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14543 vcopyq_laneq_s64 (int64x2_t __a, const int __lane1,
14544                   int64x2_t __b, const int __lane2)
14545 {
14546   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14547                                    __a, __lane1);
14548 }
14549
14550 __extension__ extern __inline uint8x16_t
14551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14552 vcopyq_laneq_u8 (uint8x16_t __a, const int __lane1,
14553                  uint8x16_t __b, const int __lane2)
14554 {
14555   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14556                                   __a, __lane1);
14557 }
14558
14559 __extension__ extern __inline uint16x8_t
14560 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14561 vcopyq_laneq_u16 (uint16x8_t __a, const int __lane1,
14562                   uint16x8_t __b, const int __lane2)
14563 {
14564   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14565                                    __a, __lane1);
14566 }
14567
14568 __extension__ extern __inline uint32x4_t
14569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14570 vcopyq_laneq_u32 (uint32x4_t __a, const int __lane1,
14571                   uint32x4_t __b, const int __lane2)
14572 {
14573   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14574                                    __a, __lane1);
14575 }
14576
14577 __extension__ extern __inline uint64x2_t
14578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14579 vcopyq_laneq_u64 (uint64x2_t __a, const int __lane1,
14580                   uint64x2_t __b, const int __lane2)
14581 {
14582   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14583                                    __a, __lane1);
14584 }
14585
14586 /* vcvt (double -> float).  */
14587
14588 __extension__ extern __inline float16x4_t
14589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14590 vcvt_f16_f32 (float32x4_t __a)
14591 {
14592   return __builtin_aarch64_float_truncate_lo_v4hf (__a);
14593 }
14594
14595 __extension__ extern __inline float16x8_t
14596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14597 vcvt_high_f16_f32 (float16x4_t __a, float32x4_t __b)
14598 {
14599   return __builtin_aarch64_float_truncate_hi_v8hf (__a, __b);
14600 }
14601
14602 __extension__ extern __inline float32x2_t
14603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14604 vcvt_f32_f64 (float64x2_t __a)
14605 {
14606   return __builtin_aarch64_float_truncate_lo_v2sf (__a);
14607 }
14608
14609 __extension__ extern __inline float32x4_t
14610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14611 vcvt_high_f32_f64 (float32x2_t __a, float64x2_t __b)
14612 {
14613   return __builtin_aarch64_float_truncate_hi_v4sf (__a, __b);
14614 }
14615
14616 /* vcvt (float -> double).  */
14617
14618 __extension__ extern __inline float32x4_t
14619 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14620 vcvt_f32_f16 (float16x4_t __a)
14621 {
14622   return __builtin_aarch64_float_extend_lo_v4sf (__a);
14623 }
14624
14625 __extension__ extern __inline float64x2_t
14626 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14627 vcvt_f64_f32 (float32x2_t __a)
14628 {
14629
14630   return __builtin_aarch64_float_extend_lo_v2df (__a);
14631 }
14632
14633 __extension__ extern __inline float32x4_t
14634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14635 vcvt_high_f32_f16 (float16x8_t __a)
14636 {
14637   return __builtin_aarch64_vec_unpacks_hi_v8hf (__a);
14638 }
14639
14640 __extension__ extern __inline float64x2_t
14641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14642 vcvt_high_f64_f32 (float32x4_t __a)
14643 {
14644   return __builtin_aarch64_vec_unpacks_hi_v4sf (__a);
14645 }
14646
14647 /* vcvt (<u>fixed-point -> float).  */
14648
14649 __extension__ extern __inline float64_t
14650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14651 vcvtd_n_f64_s64 (int64_t __a, const int __b)
14652 {
14653   return __builtin_aarch64_scvtfdi (__a, __b);
14654 }
14655
14656 __extension__ extern __inline float64_t
14657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14658 vcvtd_n_f64_u64 (uint64_t __a, const int __b)
14659 {
14660   return __builtin_aarch64_ucvtfdi_sus (__a, __b);
14661 }
14662
14663 __extension__ extern __inline float32_t
14664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14665 vcvts_n_f32_s32 (int32_t __a, const int __b)
14666 {
14667   return __builtin_aarch64_scvtfsi (__a, __b);
14668 }
14669
14670 __extension__ extern __inline float32_t
14671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14672 vcvts_n_f32_u32 (uint32_t __a, const int __b)
14673 {
14674   return __builtin_aarch64_ucvtfsi_sus (__a, __b);
14675 }
14676
14677 __extension__ extern __inline float32x2_t
14678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14679 vcvt_n_f32_s32 (int32x2_t __a, const int __b)
14680 {
14681   return __builtin_aarch64_scvtfv2si (__a, __b);
14682 }
14683
14684 __extension__ extern __inline float32x2_t
14685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14686 vcvt_n_f32_u32 (uint32x2_t __a, const int __b)
14687 {
14688   return __builtin_aarch64_ucvtfv2si_sus (__a, __b);
14689 }
14690
14691 __extension__ extern __inline float64x1_t
14692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14693 vcvt_n_f64_s64 (int64x1_t __a, const int __b)
14694 {
14695   return (float64x1_t)
14696     { __builtin_aarch64_scvtfdi (vget_lane_s64 (__a, 0), __b) };
14697 }
14698
14699 __extension__ extern __inline float64x1_t
14700 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14701 vcvt_n_f64_u64 (uint64x1_t __a, const int __b)
14702 {
14703   return (float64x1_t)
14704     { __builtin_aarch64_ucvtfdi_sus (vget_lane_u64 (__a, 0), __b) };
14705 }
14706
14707 __extension__ extern __inline float32x4_t
14708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14709 vcvtq_n_f32_s32 (int32x4_t __a, const int __b)
14710 {
14711   return __builtin_aarch64_scvtfv4si (__a, __b);
14712 }
14713
14714 __extension__ extern __inline float32x4_t
14715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14716 vcvtq_n_f32_u32 (uint32x4_t __a, const int __b)
14717 {
14718   return __builtin_aarch64_ucvtfv4si_sus (__a, __b);
14719 }
14720
14721 __extension__ extern __inline float64x2_t
14722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14723 vcvtq_n_f64_s64 (int64x2_t __a, const int __b)
14724 {
14725   return __builtin_aarch64_scvtfv2di (__a, __b);
14726 }
14727
14728 __extension__ extern __inline float64x2_t
14729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14730 vcvtq_n_f64_u64 (uint64x2_t __a, const int __b)
14731 {
14732   return __builtin_aarch64_ucvtfv2di_sus (__a, __b);
14733 }
14734
14735 /* vcvt (float -> <u>fixed-point).  */
14736
14737 __extension__ extern __inline int64_t
14738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14739 vcvtd_n_s64_f64 (float64_t __a, const int __b)
14740 {
14741   return __builtin_aarch64_fcvtzsdf (__a, __b);
14742 }
14743
14744 __extension__ extern __inline uint64_t
14745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14746 vcvtd_n_u64_f64 (float64_t __a, const int __b)
14747 {
14748   return __builtin_aarch64_fcvtzudf_uss (__a, __b);
14749 }
14750
14751 __extension__ extern __inline int32_t
14752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14753 vcvts_n_s32_f32 (float32_t __a, const int __b)
14754 {
14755   return __builtin_aarch64_fcvtzssf (__a, __b);
14756 }
14757
14758 __extension__ extern __inline uint32_t
14759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14760 vcvts_n_u32_f32 (float32_t __a, const int __b)
14761 {
14762   return __builtin_aarch64_fcvtzusf_uss (__a, __b);
14763 }
14764
14765 __extension__ extern __inline int32x2_t
14766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14767 vcvt_n_s32_f32 (float32x2_t __a, const int __b)
14768 {
14769   return __builtin_aarch64_fcvtzsv2sf (__a, __b);
14770 }
14771
14772 __extension__ extern __inline uint32x2_t
14773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14774 vcvt_n_u32_f32 (float32x2_t __a, const int __b)
14775 {
14776   return __builtin_aarch64_fcvtzuv2sf_uss (__a, __b);
14777 }
14778
14779 __extension__ extern __inline int64x1_t
14780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14781 vcvt_n_s64_f64 (float64x1_t __a, const int __b)
14782 {
14783   return (int64x1_t)
14784     { __builtin_aarch64_fcvtzsdf (vget_lane_f64 (__a, 0), __b) };
14785 }
14786
14787 __extension__ extern __inline uint64x1_t
14788 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14789 vcvt_n_u64_f64 (float64x1_t __a, const int __b)
14790 {
14791   return (uint64x1_t)
14792     { __builtin_aarch64_fcvtzudf_uss (vget_lane_f64 (__a, 0), __b) };
14793 }
14794
14795 __extension__ extern __inline int32x4_t
14796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14797 vcvtq_n_s32_f32 (float32x4_t __a, const int __b)
14798 {
14799   return __builtin_aarch64_fcvtzsv4sf (__a, __b);
14800 }
14801
14802 __extension__ extern __inline uint32x4_t
14803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14804 vcvtq_n_u32_f32 (float32x4_t __a, const int __b)
14805 {
14806   return __builtin_aarch64_fcvtzuv4sf_uss (__a, __b);
14807 }
14808
14809 __extension__ extern __inline int64x2_t
14810 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14811 vcvtq_n_s64_f64 (float64x2_t __a, const int __b)
14812 {
14813   return __builtin_aarch64_fcvtzsv2df (__a, __b);
14814 }
14815
14816 __extension__ extern __inline uint64x2_t
14817 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14818 vcvtq_n_u64_f64 (float64x2_t __a, const int __b)
14819 {
14820   return __builtin_aarch64_fcvtzuv2df_uss (__a, __b);
14821 }
14822
14823 /* vcvt  (<u>int -> float)  */
14824
14825 __extension__ extern __inline float64_t
14826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14827 vcvtd_f64_s64 (int64_t __a)
14828 {
14829   return (float64_t) __a;
14830 }
14831
14832 __extension__ extern __inline float64_t
14833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14834 vcvtd_f64_u64 (uint64_t __a)
14835 {
14836   return (float64_t) __a;
14837 }
14838
14839 __extension__ extern __inline float32_t
14840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14841 vcvts_f32_s32 (int32_t __a)
14842 {
14843   return (float32_t) __a;
14844 }
14845
14846 __extension__ extern __inline float32_t
14847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14848 vcvts_f32_u32 (uint32_t __a)
14849 {
14850   return (float32_t) __a;
14851 }
14852
14853 __extension__ extern __inline float32x2_t
14854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14855 vcvt_f32_s32 (int32x2_t __a)
14856 {
14857   return __builtin_aarch64_floatv2siv2sf (__a);
14858 }
14859
14860 __extension__ extern __inline float32x2_t
14861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14862 vcvt_f32_u32 (uint32x2_t __a)
14863 {
14864   return __builtin_aarch64_floatunsv2siv2sf ((int32x2_t) __a);
14865 }
14866
14867 __extension__ extern __inline float64x1_t
14868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14869 vcvt_f64_s64 (int64x1_t __a)
14870 {
14871   return (float64x1_t) { vget_lane_s64 (__a, 0) };
14872 }
14873
14874 __extension__ extern __inline float64x1_t
14875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14876 vcvt_f64_u64 (uint64x1_t __a)
14877 {
14878   return (float64x1_t) { vget_lane_u64 (__a, 0) };
14879 }
14880
14881 __extension__ extern __inline float32x4_t
14882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14883 vcvtq_f32_s32 (int32x4_t __a)
14884 {
14885   return __builtin_aarch64_floatv4siv4sf (__a);
14886 }
14887
14888 __extension__ extern __inline float32x4_t
14889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14890 vcvtq_f32_u32 (uint32x4_t __a)
14891 {
14892   return __builtin_aarch64_floatunsv4siv4sf ((int32x4_t) __a);
14893 }
14894
14895 __extension__ extern __inline float64x2_t
14896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14897 vcvtq_f64_s64 (int64x2_t __a)
14898 {
14899   return __builtin_aarch64_floatv2div2df (__a);
14900 }
14901
14902 __extension__ extern __inline float64x2_t
14903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14904 vcvtq_f64_u64 (uint64x2_t __a)
14905 {
14906   return __builtin_aarch64_floatunsv2div2df ((int64x2_t) __a);
14907 }
14908
14909 /* vcvt (float -> <u>int)  */
14910
14911 __extension__ extern __inline int64_t
14912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14913 vcvtd_s64_f64 (float64_t __a)
14914 {
14915   return (int64_t) __a;
14916 }
14917
14918 __extension__ extern __inline uint64_t
14919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14920 vcvtd_u64_f64 (float64_t __a)
14921 {
14922   return (uint64_t) __a;
14923 }
14924
14925 __extension__ extern __inline int32_t
14926 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14927 vcvts_s32_f32 (float32_t __a)
14928 {
14929   return (int32_t) __a;
14930 }
14931
14932 __extension__ extern __inline uint32_t
14933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14934 vcvts_u32_f32 (float32_t __a)
14935 {
14936   return (uint32_t) __a;
14937 }
14938
14939 __extension__ extern __inline int32x2_t
14940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14941 vcvt_s32_f32 (float32x2_t __a)
14942 {
14943   return __builtin_aarch64_lbtruncv2sfv2si (__a);
14944 }
14945
14946 __extension__ extern __inline uint32x2_t
14947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14948 vcvt_u32_f32 (float32x2_t __a)
14949 {
14950   return __builtin_aarch64_lbtruncuv2sfv2si_us (__a);
14951 }
14952
14953 __extension__ extern __inline int32x4_t
14954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14955 vcvtq_s32_f32 (float32x4_t __a)
14956 {
14957   return __builtin_aarch64_lbtruncv4sfv4si (__a);
14958 }
14959
14960 __extension__ extern __inline uint32x4_t
14961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14962 vcvtq_u32_f32 (float32x4_t __a)
14963 {
14964   return __builtin_aarch64_lbtruncuv4sfv4si_us (__a);
14965 }
14966
14967 __extension__ extern __inline int64x1_t
14968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14969 vcvt_s64_f64 (float64x1_t __a)
14970 {
14971   return (int64x1_t) {vcvtd_s64_f64 (__a[0])};
14972 }
14973
14974 __extension__ extern __inline uint64x1_t
14975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14976 vcvt_u64_f64 (float64x1_t __a)
14977 {
14978   return (uint64x1_t) {vcvtd_u64_f64 (__a[0])};
14979 }
14980
14981 __extension__ extern __inline int64x2_t
14982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14983 vcvtq_s64_f64 (float64x2_t __a)
14984 {
14985   return __builtin_aarch64_lbtruncv2dfv2di (__a);
14986 }
14987
14988 __extension__ extern __inline uint64x2_t
14989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14990 vcvtq_u64_f64 (float64x2_t __a)
14991 {
14992   return __builtin_aarch64_lbtruncuv2dfv2di_us (__a);
14993 }
14994
14995 /* vcvta  */
14996
14997 __extension__ extern __inline int64_t
14998 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14999 vcvtad_s64_f64 (float64_t __a)
15000 {
15001   return __builtin_aarch64_lrounddfdi (__a);
15002 }
15003
15004 __extension__ extern __inline uint64_t
15005 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15006 vcvtad_u64_f64 (float64_t __a)
15007 {
15008   return __builtin_aarch64_lroundudfdi_us (__a);
15009 }
15010
15011 __extension__ extern __inline int32_t
15012 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15013 vcvtas_s32_f32 (float32_t __a)
15014 {
15015   return __builtin_aarch64_lroundsfsi (__a);
15016 }
15017
15018 __extension__ extern __inline uint32_t
15019 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15020 vcvtas_u32_f32 (float32_t __a)
15021 {
15022   return __builtin_aarch64_lroundusfsi_us (__a);
15023 }
15024
15025 __extension__ extern __inline int32x2_t
15026 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15027 vcvta_s32_f32 (float32x2_t __a)
15028 {
15029   return __builtin_aarch64_lroundv2sfv2si (__a);
15030 }
15031
15032 __extension__ extern __inline uint32x2_t
15033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15034 vcvta_u32_f32 (float32x2_t __a)
15035 {
15036   return __builtin_aarch64_lrounduv2sfv2si_us (__a);
15037 }
15038
15039 __extension__ extern __inline int32x4_t
15040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15041 vcvtaq_s32_f32 (float32x4_t __a)
15042 {
15043   return __builtin_aarch64_lroundv4sfv4si (__a);
15044 }
15045
15046 __extension__ extern __inline uint32x4_t
15047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15048 vcvtaq_u32_f32 (float32x4_t __a)
15049 {
15050   return __builtin_aarch64_lrounduv4sfv4si_us (__a);
15051 }
15052
15053 __extension__ extern __inline int64x1_t
15054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15055 vcvta_s64_f64 (float64x1_t __a)
15056 {
15057   return (int64x1_t) {vcvtad_s64_f64 (__a[0])};
15058 }
15059
15060 __extension__ extern __inline uint64x1_t
15061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15062 vcvta_u64_f64 (float64x1_t __a)
15063 {
15064   return (uint64x1_t) {vcvtad_u64_f64 (__a[0])};
15065 }
15066
15067 __extension__ extern __inline int64x2_t
15068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15069 vcvtaq_s64_f64 (float64x2_t __a)
15070 {
15071   return __builtin_aarch64_lroundv2dfv2di (__a);
15072 }
15073
15074 __extension__ extern __inline uint64x2_t
15075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15076 vcvtaq_u64_f64 (float64x2_t __a)
15077 {
15078   return __builtin_aarch64_lrounduv2dfv2di_us (__a);
15079 }
15080
15081 /* vcvtm  */
15082
15083 __extension__ extern __inline int64_t
15084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15085 vcvtmd_s64_f64 (float64_t __a)
15086 {
15087   return __builtin_llfloor (__a);
15088 }
15089
15090 __extension__ extern __inline uint64_t
15091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15092 vcvtmd_u64_f64 (float64_t __a)
15093 {
15094   return __builtin_aarch64_lfloorudfdi_us (__a);
15095 }
15096
15097 __extension__ extern __inline int32_t
15098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15099 vcvtms_s32_f32 (float32_t __a)
15100 {
15101   return __builtin_ifloorf (__a);
15102 }
15103
15104 __extension__ extern __inline uint32_t
15105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15106 vcvtms_u32_f32 (float32_t __a)
15107 {
15108   return __builtin_aarch64_lfloorusfsi_us (__a);
15109 }
15110
15111 __extension__ extern __inline int32x2_t
15112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15113 vcvtm_s32_f32 (float32x2_t __a)
15114 {
15115   return __builtin_aarch64_lfloorv2sfv2si (__a);
15116 }
15117
15118 __extension__ extern __inline uint32x2_t
15119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15120 vcvtm_u32_f32 (float32x2_t __a)
15121 {
15122   return __builtin_aarch64_lflooruv2sfv2si_us (__a);
15123 }
15124
15125 __extension__ extern __inline int32x4_t
15126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15127 vcvtmq_s32_f32 (float32x4_t __a)
15128 {
15129   return __builtin_aarch64_lfloorv4sfv4si (__a);
15130 }
15131
15132 __extension__ extern __inline uint32x4_t
15133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15134 vcvtmq_u32_f32 (float32x4_t __a)
15135 {
15136   return __builtin_aarch64_lflooruv4sfv4si_us (__a);
15137 }
15138
15139 __extension__ extern __inline int64x1_t
15140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15141 vcvtm_s64_f64 (float64x1_t __a)
15142 {
15143   return (int64x1_t) {vcvtmd_s64_f64 (__a[0])};
15144 }
15145
15146 __extension__ extern __inline uint64x1_t
15147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15148 vcvtm_u64_f64 (float64x1_t __a)
15149 {
15150   return (uint64x1_t) {vcvtmd_u64_f64 (__a[0])};
15151 }
15152
15153 __extension__ extern __inline int64x2_t
15154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15155 vcvtmq_s64_f64 (float64x2_t __a)
15156 {
15157   return __builtin_aarch64_lfloorv2dfv2di (__a);
15158 }
15159
15160 __extension__ extern __inline uint64x2_t
15161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15162 vcvtmq_u64_f64 (float64x2_t __a)
15163 {
15164   return __builtin_aarch64_lflooruv2dfv2di_us (__a);
15165 }
15166
15167 /* vcvtn  */
15168
15169 __extension__ extern __inline int64_t
15170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15171 vcvtnd_s64_f64 (float64_t __a)
15172 {
15173   return __builtin_aarch64_lfrintndfdi (__a);
15174 }
15175
15176 __extension__ extern __inline uint64_t
15177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15178 vcvtnd_u64_f64 (float64_t __a)
15179 {
15180   return __builtin_aarch64_lfrintnudfdi_us (__a);
15181 }
15182
15183 __extension__ extern __inline int32_t
15184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15185 vcvtns_s32_f32 (float32_t __a)
15186 {
15187   return __builtin_aarch64_lfrintnsfsi (__a);
15188 }
15189
15190 __extension__ extern __inline uint32_t
15191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15192 vcvtns_u32_f32 (float32_t __a)
15193 {
15194   return __builtin_aarch64_lfrintnusfsi_us (__a);
15195 }
15196
15197 __extension__ extern __inline int32x2_t
15198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15199 vcvtn_s32_f32 (float32x2_t __a)
15200 {
15201   return __builtin_aarch64_lfrintnv2sfv2si (__a);
15202 }
15203
15204 __extension__ extern __inline uint32x2_t
15205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15206 vcvtn_u32_f32 (float32x2_t __a)
15207 {
15208   return __builtin_aarch64_lfrintnuv2sfv2si_us (__a);
15209 }
15210
15211 __extension__ extern __inline int32x4_t
15212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15213 vcvtnq_s32_f32 (float32x4_t __a)
15214 {
15215   return __builtin_aarch64_lfrintnv4sfv4si (__a);
15216 }
15217
15218 __extension__ extern __inline uint32x4_t
15219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15220 vcvtnq_u32_f32 (float32x4_t __a)
15221 {
15222   return __builtin_aarch64_lfrintnuv4sfv4si_us (__a);
15223 }
15224
15225 __extension__ extern __inline int64x1_t
15226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15227 vcvtn_s64_f64 (float64x1_t __a)
15228 {
15229   return (int64x1_t) {vcvtnd_s64_f64 (__a[0])};
15230 }
15231
15232 __extension__ extern __inline uint64x1_t
15233 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15234 vcvtn_u64_f64 (float64x1_t __a)
15235 {
15236   return (uint64x1_t) {vcvtnd_u64_f64 (__a[0])};
15237 }
15238
15239 __extension__ extern __inline int64x2_t
15240 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15241 vcvtnq_s64_f64 (float64x2_t __a)
15242 {
15243   return __builtin_aarch64_lfrintnv2dfv2di (__a);
15244 }
15245
15246 __extension__ extern __inline uint64x2_t
15247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15248 vcvtnq_u64_f64 (float64x2_t __a)
15249 {
15250   return __builtin_aarch64_lfrintnuv2dfv2di_us (__a);
15251 }
15252
15253 /* vcvtp  */
15254
15255 __extension__ extern __inline int64_t
15256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15257 vcvtpd_s64_f64 (float64_t __a)
15258 {
15259   return __builtin_llceil (__a);
15260 }
15261
15262 __extension__ extern __inline uint64_t
15263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15264 vcvtpd_u64_f64 (float64_t __a)
15265 {
15266   return __builtin_aarch64_lceiludfdi_us (__a);
15267 }
15268
15269 __extension__ extern __inline int32_t
15270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15271 vcvtps_s32_f32 (float32_t __a)
15272 {
15273   return __builtin_iceilf (__a);
15274 }
15275
15276 __extension__ extern __inline uint32_t
15277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15278 vcvtps_u32_f32 (float32_t __a)
15279 {
15280   return __builtin_aarch64_lceilusfsi_us (__a);
15281 }
15282
15283 __extension__ extern __inline int32x2_t
15284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15285 vcvtp_s32_f32 (float32x2_t __a)
15286 {
15287   return __builtin_aarch64_lceilv2sfv2si (__a);
15288 }
15289
15290 __extension__ extern __inline uint32x2_t
15291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15292 vcvtp_u32_f32 (float32x2_t __a)
15293 {
15294   return __builtin_aarch64_lceiluv2sfv2si_us (__a);
15295 }
15296
15297 __extension__ extern __inline int32x4_t
15298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15299 vcvtpq_s32_f32 (float32x4_t __a)
15300 {
15301   return __builtin_aarch64_lceilv4sfv4si (__a);
15302 }
15303
15304 __extension__ extern __inline uint32x4_t
15305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15306 vcvtpq_u32_f32 (float32x4_t __a)
15307 {
15308   return __builtin_aarch64_lceiluv4sfv4si_us (__a);
15309 }
15310
15311 __extension__ extern __inline int64x1_t
15312 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15313 vcvtp_s64_f64 (float64x1_t __a)
15314 {
15315   return (int64x1_t) {vcvtpd_s64_f64 (__a[0])};
15316 }
15317
15318 __extension__ extern __inline uint64x1_t
15319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15320 vcvtp_u64_f64 (float64x1_t __a)
15321 {
15322   return (uint64x1_t) {vcvtpd_u64_f64 (__a[0])};
15323 }
15324
15325 __extension__ extern __inline int64x2_t
15326 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15327 vcvtpq_s64_f64 (float64x2_t __a)
15328 {
15329   return __builtin_aarch64_lceilv2dfv2di (__a);
15330 }
15331
15332 __extension__ extern __inline uint64x2_t
15333 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15334 vcvtpq_u64_f64 (float64x2_t __a)
15335 {
15336   return __builtin_aarch64_lceiluv2dfv2di_us (__a);
15337 }
15338
15339 /* vdup_n  */
15340
15341 __extension__ extern __inline float16x4_t
15342 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15343 vdup_n_f16 (float16_t __a)
15344 {
15345   return (float16x4_t) {__a, __a, __a, __a};
15346 }
15347
15348 __extension__ extern __inline float32x2_t
15349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15350 vdup_n_f32 (float32_t __a)
15351 {
15352   return (float32x2_t) {__a, __a};
15353 }
15354
15355 __extension__ extern __inline float64x1_t
15356 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15357 vdup_n_f64 (float64_t __a)
15358 {
15359   return (float64x1_t) {__a};
15360 }
15361
15362 __extension__ extern __inline poly8x8_t
15363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15364 vdup_n_p8 (poly8_t __a)
15365 {
15366   return (poly8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15367 }
15368
15369 __extension__ extern __inline poly16x4_t
15370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15371 vdup_n_p16 (poly16_t __a)
15372 {
15373   return (poly16x4_t) {__a, __a, __a, __a};
15374 }
15375
15376 __extension__ extern __inline poly64x1_t
15377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15378 vdup_n_p64 (poly64_t __a)
15379 {
15380   return (poly64x1_t) {__a};
15381 }
15382
15383 __extension__ extern __inline int8x8_t
15384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15385 vdup_n_s8 (int8_t __a)
15386 {
15387   return (int8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15388 }
15389
15390 __extension__ extern __inline int16x4_t
15391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15392 vdup_n_s16 (int16_t __a)
15393 {
15394   return (int16x4_t) {__a, __a, __a, __a};
15395 }
15396
15397 __extension__ extern __inline int32x2_t
15398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15399 vdup_n_s32 (int32_t __a)
15400 {
15401   return (int32x2_t) {__a, __a};
15402 }
15403
15404 __extension__ extern __inline int64x1_t
15405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15406 vdup_n_s64 (int64_t __a)
15407 {
15408   return (int64x1_t) {__a};
15409 }
15410
15411 __extension__ extern __inline uint8x8_t
15412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15413 vdup_n_u8 (uint8_t __a)
15414 {
15415   return (uint8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15416 }
15417
15418 __extension__ extern __inline uint16x4_t
15419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15420 vdup_n_u16 (uint16_t __a)
15421 {
15422   return (uint16x4_t) {__a, __a, __a, __a};
15423 }
15424
15425 __extension__ extern __inline uint32x2_t
15426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15427 vdup_n_u32 (uint32_t __a)
15428 {
15429   return (uint32x2_t) {__a, __a};
15430 }
15431
15432 __extension__ extern __inline uint64x1_t
15433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15434 vdup_n_u64 (uint64_t __a)
15435 {
15436   return (uint64x1_t) {__a};
15437 }
15438
15439 /* vdupq_n  */
15440
15441 __extension__ extern __inline float16x8_t
15442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15443 vdupq_n_f16 (float16_t __a)
15444 {
15445   return (float16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15446 }
15447
15448 __extension__ extern __inline float32x4_t
15449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15450 vdupq_n_f32 (float32_t __a)
15451 {
15452   return (float32x4_t) {__a, __a, __a, __a};
15453 }
15454
15455 __extension__ extern __inline float64x2_t
15456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15457 vdupq_n_f64 (float64_t __a)
15458 {
15459   return (float64x2_t) {__a, __a};
15460 }
15461
15462 __extension__ extern __inline poly8x16_t
15463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15464 vdupq_n_p8 (uint32_t __a)
15465 {
15466   return (poly8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15467                        __a, __a, __a, __a, __a, __a, __a, __a};
15468 }
15469
15470 __extension__ extern __inline poly16x8_t
15471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15472 vdupq_n_p16 (uint32_t __a)
15473 {
15474   return (poly16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15475 }
15476
15477 __extension__ extern __inline poly64x2_t
15478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15479 vdupq_n_p64 (uint64_t __a)
15480 {
15481   return (poly64x2_t) {__a, __a};
15482 }
15483
15484 __extension__ extern __inline int8x16_t
15485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15486 vdupq_n_s8 (int32_t __a)
15487 {
15488   return (int8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15489                       __a, __a, __a, __a, __a, __a, __a, __a};
15490 }
15491
15492 __extension__ extern __inline int16x8_t
15493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15494 vdupq_n_s16 (int32_t __a)
15495 {
15496   return (int16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15497 }
15498
15499 __extension__ extern __inline int32x4_t
15500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15501 vdupq_n_s32 (int32_t __a)
15502 {
15503   return (int32x4_t) {__a, __a, __a, __a};
15504 }
15505
15506 __extension__ extern __inline int64x2_t
15507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15508 vdupq_n_s64 (int64_t __a)
15509 {
15510   return (int64x2_t) {__a, __a};
15511 }
15512
15513 __extension__ extern __inline uint8x16_t
15514 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15515 vdupq_n_u8 (uint32_t __a)
15516 {
15517   return (uint8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15518                        __a, __a, __a, __a, __a, __a, __a, __a};
15519 }
15520
15521 __extension__ extern __inline uint16x8_t
15522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15523 vdupq_n_u16 (uint32_t __a)
15524 {
15525   return (uint16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15526 }
15527
15528 __extension__ extern __inline uint32x4_t
15529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15530 vdupq_n_u32 (uint32_t __a)
15531 {
15532   return (uint32x4_t) {__a, __a, __a, __a};
15533 }
15534
15535 __extension__ extern __inline uint64x2_t
15536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15537 vdupq_n_u64 (uint64_t __a)
15538 {
15539   return (uint64x2_t) {__a, __a};
15540 }
15541
15542 /* vdup_lane  */
15543
15544 __extension__ extern __inline float16x4_t
15545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15546 vdup_lane_f16 (float16x4_t __a, const int __b)
15547 {
15548   return __aarch64_vdup_lane_f16 (__a, __b);
15549 }
15550
15551 __extension__ extern __inline float32x2_t
15552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15553 vdup_lane_f32 (float32x2_t __a, const int __b)
15554 {
15555   return __aarch64_vdup_lane_f32 (__a, __b);
15556 }
15557
15558 __extension__ extern __inline float64x1_t
15559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15560 vdup_lane_f64 (float64x1_t __a, const int __b)
15561 {
15562   return __aarch64_vdup_lane_f64 (__a, __b);
15563 }
15564
15565 __extension__ extern __inline poly8x8_t
15566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15567 vdup_lane_p8 (poly8x8_t __a, const int __b)
15568 {
15569   return __aarch64_vdup_lane_p8 (__a, __b);
15570 }
15571
15572 __extension__ extern __inline poly16x4_t
15573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15574 vdup_lane_p16 (poly16x4_t __a, const int __b)
15575 {
15576   return __aarch64_vdup_lane_p16 (__a, __b);
15577 }
15578
15579 __extension__ extern __inline poly64x1_t
15580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15581 vdup_lane_p64 (poly64x1_t __a, const int __b)
15582 {
15583   return __aarch64_vdup_lane_p64 (__a, __b);
15584 }
15585
15586 __extension__ extern __inline int8x8_t
15587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15588 vdup_lane_s8 (int8x8_t __a, const int __b)
15589 {
15590   return __aarch64_vdup_lane_s8 (__a, __b);
15591 }
15592
15593 __extension__ extern __inline int16x4_t
15594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15595 vdup_lane_s16 (int16x4_t __a, const int __b)
15596 {
15597   return __aarch64_vdup_lane_s16 (__a, __b);
15598 }
15599
15600 __extension__ extern __inline int32x2_t
15601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15602 vdup_lane_s32 (int32x2_t __a, const int __b)
15603 {
15604   return __aarch64_vdup_lane_s32 (__a, __b);
15605 }
15606
15607 __extension__ extern __inline int64x1_t
15608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15609 vdup_lane_s64 (int64x1_t __a, const int __b)
15610 {
15611   return __aarch64_vdup_lane_s64 (__a, __b);
15612 }
15613
15614 __extension__ extern __inline uint8x8_t
15615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15616 vdup_lane_u8 (uint8x8_t __a, const int __b)
15617 {
15618   return __aarch64_vdup_lane_u8 (__a, __b);
15619 }
15620
15621 __extension__ extern __inline uint16x4_t
15622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15623 vdup_lane_u16 (uint16x4_t __a, const int __b)
15624 {
15625   return __aarch64_vdup_lane_u16 (__a, __b);
15626 }
15627
15628 __extension__ extern __inline uint32x2_t
15629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15630 vdup_lane_u32 (uint32x2_t __a, const int __b)
15631 {
15632   return __aarch64_vdup_lane_u32 (__a, __b);
15633 }
15634
15635 __extension__ extern __inline uint64x1_t
15636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15637 vdup_lane_u64 (uint64x1_t __a, const int __b)
15638 {
15639   return __aarch64_vdup_lane_u64 (__a, __b);
15640 }
15641
15642 /* vdup_laneq  */
15643
15644 __extension__ extern __inline float16x4_t
15645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15646 vdup_laneq_f16 (float16x8_t __a, const int __b)
15647 {
15648   return __aarch64_vdup_laneq_f16 (__a, __b);
15649 }
15650
15651 __extension__ extern __inline float32x2_t
15652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15653 vdup_laneq_f32 (float32x4_t __a, const int __b)
15654 {
15655   return __aarch64_vdup_laneq_f32 (__a, __b);
15656 }
15657
15658 __extension__ extern __inline float64x1_t
15659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15660 vdup_laneq_f64 (float64x2_t __a, const int __b)
15661 {
15662   return __aarch64_vdup_laneq_f64 (__a, __b);
15663 }
15664
15665 __extension__ extern __inline poly8x8_t
15666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15667 vdup_laneq_p8 (poly8x16_t __a, const int __b)
15668 {
15669   return __aarch64_vdup_laneq_p8 (__a, __b);
15670 }
15671
15672 __extension__ extern __inline poly16x4_t
15673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15674 vdup_laneq_p16 (poly16x8_t __a, const int __b)
15675 {
15676   return __aarch64_vdup_laneq_p16 (__a, __b);
15677 }
15678
15679 __extension__ extern __inline poly64x1_t
15680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15681 vdup_laneq_p64 (poly64x2_t __a, const int __b)
15682 {
15683   return __aarch64_vdup_laneq_p64 (__a, __b);
15684 }
15685
15686 __extension__ extern __inline int8x8_t
15687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15688 vdup_laneq_s8 (int8x16_t __a, const int __b)
15689 {
15690   return __aarch64_vdup_laneq_s8 (__a, __b);
15691 }
15692
15693 __extension__ extern __inline int16x4_t
15694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15695 vdup_laneq_s16 (int16x8_t __a, const int __b)
15696 {
15697   return __aarch64_vdup_laneq_s16 (__a, __b);
15698 }
15699
15700 __extension__ extern __inline int32x2_t
15701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15702 vdup_laneq_s32 (int32x4_t __a, const int __b)
15703 {
15704   return __aarch64_vdup_laneq_s32 (__a, __b);
15705 }
15706
15707 __extension__ extern __inline int64x1_t
15708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15709 vdup_laneq_s64 (int64x2_t __a, const int __b)
15710 {
15711   return __aarch64_vdup_laneq_s64 (__a, __b);
15712 }
15713
15714 __extension__ extern __inline uint8x8_t
15715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15716 vdup_laneq_u8 (uint8x16_t __a, const int __b)
15717 {
15718   return __aarch64_vdup_laneq_u8 (__a, __b);
15719 }
15720
15721 __extension__ extern __inline uint16x4_t
15722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15723 vdup_laneq_u16 (uint16x8_t __a, const int __b)
15724 {
15725   return __aarch64_vdup_laneq_u16 (__a, __b);
15726 }
15727
15728 __extension__ extern __inline uint32x2_t
15729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15730 vdup_laneq_u32 (uint32x4_t __a, const int __b)
15731 {
15732   return __aarch64_vdup_laneq_u32 (__a, __b);
15733 }
15734
15735 __extension__ extern __inline uint64x1_t
15736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15737 vdup_laneq_u64 (uint64x2_t __a, const int __b)
15738 {
15739   return __aarch64_vdup_laneq_u64 (__a, __b);
15740 }
15741
15742 /* vdupq_lane  */
15743
15744 __extension__ extern __inline float16x8_t
15745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15746 vdupq_lane_f16 (float16x4_t __a, const int __b)
15747 {
15748   return __aarch64_vdupq_lane_f16 (__a, __b);
15749 }
15750
15751 __extension__ extern __inline float32x4_t
15752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15753 vdupq_lane_f32 (float32x2_t __a, const int __b)
15754 {
15755   return __aarch64_vdupq_lane_f32 (__a, __b);
15756 }
15757
15758 __extension__ extern __inline float64x2_t
15759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15760 vdupq_lane_f64 (float64x1_t __a, const int __b)
15761 {
15762   return __aarch64_vdupq_lane_f64 (__a, __b);
15763 }
15764
15765 __extension__ extern __inline poly8x16_t
15766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15767 vdupq_lane_p8 (poly8x8_t __a, const int __b)
15768 {
15769   return __aarch64_vdupq_lane_p8 (__a, __b);
15770 }
15771
15772 __extension__ extern __inline poly16x8_t
15773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15774 vdupq_lane_p16 (poly16x4_t __a, const int __b)
15775 {
15776   return __aarch64_vdupq_lane_p16 (__a, __b);
15777 }
15778
15779 __extension__ extern __inline poly64x2_t
15780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15781 vdupq_lane_p64 (poly64x1_t __a, const int __b)
15782 {
15783   return __aarch64_vdupq_lane_p64 (__a, __b);
15784 }
15785
15786 __extension__ extern __inline int8x16_t
15787 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15788 vdupq_lane_s8 (int8x8_t __a, const int __b)
15789 {
15790   return __aarch64_vdupq_lane_s8 (__a, __b);
15791 }
15792
15793 __extension__ extern __inline int16x8_t
15794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15795 vdupq_lane_s16 (int16x4_t __a, const int __b)
15796 {
15797   return __aarch64_vdupq_lane_s16 (__a, __b);
15798 }
15799
15800 __extension__ extern __inline int32x4_t
15801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15802 vdupq_lane_s32 (int32x2_t __a, const int __b)
15803 {
15804   return __aarch64_vdupq_lane_s32 (__a, __b);
15805 }
15806
15807 __extension__ extern __inline int64x2_t
15808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15809 vdupq_lane_s64 (int64x1_t __a, const int __b)
15810 {
15811   return __aarch64_vdupq_lane_s64 (__a, __b);
15812 }
15813
15814 __extension__ extern __inline uint8x16_t
15815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15816 vdupq_lane_u8 (uint8x8_t __a, const int __b)
15817 {
15818   return __aarch64_vdupq_lane_u8 (__a, __b);
15819 }
15820
15821 __extension__ extern __inline uint16x8_t
15822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15823 vdupq_lane_u16 (uint16x4_t __a, const int __b)
15824 {
15825   return __aarch64_vdupq_lane_u16 (__a, __b);
15826 }
15827
15828 __extension__ extern __inline uint32x4_t
15829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15830 vdupq_lane_u32 (uint32x2_t __a, const int __b)
15831 {
15832   return __aarch64_vdupq_lane_u32 (__a, __b);
15833 }
15834
15835 __extension__ extern __inline uint64x2_t
15836 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15837 vdupq_lane_u64 (uint64x1_t __a, const int __b)
15838 {
15839   return __aarch64_vdupq_lane_u64 (__a, __b);
15840 }
15841
15842 /* vdupq_laneq  */
15843
15844 __extension__ extern __inline float16x8_t
15845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15846 vdupq_laneq_f16 (float16x8_t __a, const int __b)
15847 {
15848   return __aarch64_vdupq_laneq_f16 (__a, __b);
15849 }
15850
15851 __extension__ extern __inline float32x4_t
15852 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15853 vdupq_laneq_f32 (float32x4_t __a, const int __b)
15854 {
15855   return __aarch64_vdupq_laneq_f32 (__a, __b);
15856 }
15857
15858 __extension__ extern __inline float64x2_t
15859 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15860 vdupq_laneq_f64 (float64x2_t __a, const int __b)
15861 {
15862   return __aarch64_vdupq_laneq_f64 (__a, __b);
15863 }
15864
15865 __extension__ extern __inline poly8x16_t
15866 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15867 vdupq_laneq_p8 (poly8x16_t __a, const int __b)
15868 {
15869   return __aarch64_vdupq_laneq_p8 (__a, __b);
15870 }
15871
15872 __extension__ extern __inline poly16x8_t
15873 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15874 vdupq_laneq_p16 (poly16x8_t __a, const int __b)
15875 {
15876   return __aarch64_vdupq_laneq_p16 (__a, __b);
15877 }
15878
15879 __extension__ extern __inline poly64x2_t
15880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15881 vdupq_laneq_p64 (poly64x2_t __a, const int __b)
15882 {
15883   return __aarch64_vdupq_laneq_p64 (__a, __b);
15884 }
15885
15886 __extension__ extern __inline int8x16_t
15887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15888 vdupq_laneq_s8 (int8x16_t __a, const int __b)
15889 {
15890   return __aarch64_vdupq_laneq_s8 (__a, __b);
15891 }
15892
15893 __extension__ extern __inline int16x8_t
15894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15895 vdupq_laneq_s16 (int16x8_t __a, const int __b)
15896 {
15897   return __aarch64_vdupq_laneq_s16 (__a, __b);
15898 }
15899
15900 __extension__ extern __inline int32x4_t
15901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15902 vdupq_laneq_s32 (int32x4_t __a, const int __b)
15903 {
15904   return __aarch64_vdupq_laneq_s32 (__a, __b);
15905 }
15906
15907 __extension__ extern __inline int64x2_t
15908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15909 vdupq_laneq_s64 (int64x2_t __a, const int __b)
15910 {
15911   return __aarch64_vdupq_laneq_s64 (__a, __b);
15912 }
15913
15914 __extension__ extern __inline uint8x16_t
15915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15916 vdupq_laneq_u8 (uint8x16_t __a, const int __b)
15917 {
15918   return __aarch64_vdupq_laneq_u8 (__a, __b);
15919 }
15920
15921 __extension__ extern __inline uint16x8_t
15922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15923 vdupq_laneq_u16 (uint16x8_t __a, const int __b)
15924 {
15925   return __aarch64_vdupq_laneq_u16 (__a, __b);
15926 }
15927
15928 __extension__ extern __inline uint32x4_t
15929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15930 vdupq_laneq_u32 (uint32x4_t __a, const int __b)
15931 {
15932   return __aarch64_vdupq_laneq_u32 (__a, __b);
15933 }
15934
15935 __extension__ extern __inline uint64x2_t
15936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15937 vdupq_laneq_u64 (uint64x2_t __a, const int __b)
15938 {
15939   return __aarch64_vdupq_laneq_u64 (__a, __b);
15940 }
15941
15942 /* vdupb_lane  */
15943 __extension__ extern __inline poly8_t
15944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15945 vdupb_lane_p8 (poly8x8_t __a, const int __b)
15946 {
15947   return __aarch64_vget_lane_any (__a, __b);
15948 }
15949
15950 __extension__ extern __inline int8_t
15951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15952 vdupb_lane_s8 (int8x8_t __a, const int __b)
15953 {
15954   return __aarch64_vget_lane_any (__a, __b);
15955 }
15956
15957 __extension__ extern __inline uint8_t
15958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15959 vdupb_lane_u8 (uint8x8_t __a, const int __b)
15960 {
15961   return __aarch64_vget_lane_any (__a, __b);
15962 }
15963
15964 /* vduph_lane  */
15965
15966 __extension__ extern __inline float16_t
15967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15968 vduph_lane_f16 (float16x4_t __a, const int __b)
15969 {
15970   return __aarch64_vget_lane_any (__a, __b);
15971 }
15972
15973 __extension__ extern __inline poly16_t
15974 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15975 vduph_lane_p16 (poly16x4_t __a, const int __b)
15976 {
15977   return __aarch64_vget_lane_any (__a, __b);
15978 }
15979
15980 __extension__ extern __inline int16_t
15981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15982 vduph_lane_s16 (int16x4_t __a, const int __b)
15983 {
15984   return __aarch64_vget_lane_any (__a, __b);
15985 }
15986
15987 __extension__ extern __inline uint16_t
15988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15989 vduph_lane_u16 (uint16x4_t __a, const int __b)
15990 {
15991   return __aarch64_vget_lane_any (__a, __b);
15992 }
15993
15994 /* vdups_lane  */
15995
15996 __extension__ extern __inline float32_t
15997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15998 vdups_lane_f32 (float32x2_t __a, const int __b)
15999 {
16000   return __aarch64_vget_lane_any (__a, __b);
16001 }
16002
16003 __extension__ extern __inline int32_t
16004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16005 vdups_lane_s32 (int32x2_t __a, const int __b)
16006 {
16007   return __aarch64_vget_lane_any (__a, __b);
16008 }
16009
16010 __extension__ extern __inline uint32_t
16011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16012 vdups_lane_u32 (uint32x2_t __a, const int __b)
16013 {
16014   return __aarch64_vget_lane_any (__a, __b);
16015 }
16016
16017 /* vdupd_lane  */
16018 __extension__ extern __inline float64_t
16019 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16020 vdupd_lane_f64 (float64x1_t __a, const int __b)
16021 {
16022   __AARCH64_LANE_CHECK (__a, __b);
16023   return __a[0];
16024 }
16025
16026 __extension__ extern __inline int64_t
16027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16028 vdupd_lane_s64 (int64x1_t __a, const int __b)
16029 {
16030   __AARCH64_LANE_CHECK (__a, __b);
16031   return __a[0];
16032 }
16033
16034 __extension__ extern __inline uint64_t
16035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16036 vdupd_lane_u64 (uint64x1_t __a, const int __b)
16037 {
16038   __AARCH64_LANE_CHECK (__a, __b);
16039   return __a[0];
16040 }
16041
16042 /* vdupb_laneq  */
16043 __extension__ extern __inline poly8_t
16044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16045 vdupb_laneq_p8 (poly8x16_t __a, const int __b)
16046 {
16047   return __aarch64_vget_lane_any (__a, __b);
16048 }
16049
16050 __extension__ extern __inline int8_t
16051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16052 vdupb_laneq_s8 (int8x16_t __a, const int __b)
16053 {
16054   return __aarch64_vget_lane_any (__a, __b);
16055 }
16056
16057 __extension__ extern __inline uint8_t
16058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16059 vdupb_laneq_u8 (uint8x16_t __a, const int __b)
16060 {
16061   return __aarch64_vget_lane_any (__a, __b);
16062 }
16063
16064 /* vduph_laneq  */
16065
16066 __extension__ extern __inline float16_t
16067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16068 vduph_laneq_f16 (float16x8_t __a, const int __b)
16069 {
16070   return __aarch64_vget_lane_any (__a, __b);
16071 }
16072
16073 __extension__ extern __inline poly16_t
16074 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16075 vduph_laneq_p16 (poly16x8_t __a, const int __b)
16076 {
16077   return __aarch64_vget_lane_any (__a, __b);
16078 }
16079
16080 __extension__ extern __inline int16_t
16081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16082 vduph_laneq_s16 (int16x8_t __a, const int __b)
16083 {
16084   return __aarch64_vget_lane_any (__a, __b);
16085 }
16086
16087 __extension__ extern __inline uint16_t
16088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16089 vduph_laneq_u16 (uint16x8_t __a, const int __b)
16090 {
16091   return __aarch64_vget_lane_any (__a, __b);
16092 }
16093
16094 /* vdups_laneq  */
16095
16096 __extension__ extern __inline float32_t
16097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16098 vdups_laneq_f32 (float32x4_t __a, const int __b)
16099 {
16100   return __aarch64_vget_lane_any (__a, __b);
16101 }
16102
16103 __extension__ extern __inline int32_t
16104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16105 vdups_laneq_s32 (int32x4_t __a, const int __b)
16106 {
16107   return __aarch64_vget_lane_any (__a, __b);
16108 }
16109
16110 __extension__ extern __inline uint32_t
16111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16112 vdups_laneq_u32 (uint32x4_t __a, const int __b)
16113 {
16114   return __aarch64_vget_lane_any (__a, __b);
16115 }
16116
16117 /* vdupd_laneq  */
16118 __extension__ extern __inline float64_t
16119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16120 vdupd_laneq_f64 (float64x2_t __a, const int __b)
16121 {
16122   return __aarch64_vget_lane_any (__a, __b);
16123 }
16124
16125 __extension__ extern __inline int64_t
16126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16127 vdupd_laneq_s64 (int64x2_t __a, const int __b)
16128 {
16129   return __aarch64_vget_lane_any (__a, __b);
16130 }
16131
16132 __extension__ extern __inline uint64_t
16133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16134 vdupd_laneq_u64 (uint64x2_t __a, const int __b)
16135 {
16136   return __aarch64_vget_lane_any (__a, __b);
16137 }
16138
16139 /* vext  */
16140
16141 __extension__ extern __inline float16x4_t
16142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16143 vext_f16 (float16x4_t __a, float16x4_t __b, __const int __c)
16144 {
16145   __AARCH64_LANE_CHECK (__a, __c);
16146 #ifdef __AARCH64EB__
16147   return __builtin_shuffle (__b, __a,
16148                             (uint16x4_t) {4 - __c, 5 - __c, 6 - __c, 7 - __c});
16149 #else
16150   return __builtin_shuffle (__a, __b,
16151                             (uint16x4_t) {__c, __c + 1, __c + 2, __c + 3});
16152 #endif
16153 }
16154
16155 __extension__ extern __inline float32x2_t
16156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16157 vext_f32 (float32x2_t __a, float32x2_t __b, __const int __c)
16158 {
16159   __AARCH64_LANE_CHECK (__a, __c);
16160 #ifdef __AARCH64EB__
16161   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16162 #else
16163   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16164 #endif
16165 }
16166
16167 __extension__ extern __inline float64x1_t
16168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16169 vext_f64 (float64x1_t __a, float64x1_t __b, __const int __c)
16170 {
16171   __AARCH64_LANE_CHECK (__a, __c);
16172   /* The only possible index to the assembler instruction returns element 0.  */
16173   return __a;
16174 }
16175 __extension__ extern __inline poly8x8_t
16176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16177 vext_p8 (poly8x8_t __a, poly8x8_t __b, __const int __c)
16178 {
16179   __AARCH64_LANE_CHECK (__a, __c);
16180 #ifdef __AARCH64EB__
16181   return __builtin_shuffle (__b, __a, (uint8x8_t)
16182       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16183 #else
16184   return __builtin_shuffle (__a, __b,
16185       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16186 #endif
16187 }
16188
16189 __extension__ extern __inline poly16x4_t
16190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16191 vext_p16 (poly16x4_t __a, poly16x4_t __b, __const int __c)
16192 {
16193   __AARCH64_LANE_CHECK (__a, __c);
16194 #ifdef __AARCH64EB__
16195   return __builtin_shuffle (__b, __a,
16196       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16197 #else
16198   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16199 #endif
16200 }
16201
16202 __extension__ extern __inline poly64x1_t
16203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16204 vext_p64 (poly64x1_t __a, poly64x1_t __b, __const int __c)
16205 {
16206   __AARCH64_LANE_CHECK (__a, __c);
16207   /* The only possible index to the assembler instruction returns element 0.  */
16208   return __a;
16209 }
16210
16211 __extension__ extern __inline int8x8_t
16212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16213 vext_s8 (int8x8_t __a, int8x8_t __b, __const int __c)
16214 {
16215   __AARCH64_LANE_CHECK (__a, __c);
16216 #ifdef __AARCH64EB__
16217   return __builtin_shuffle (__b, __a, (uint8x8_t)
16218       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16219 #else
16220   return __builtin_shuffle (__a, __b,
16221       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16222 #endif
16223 }
16224
16225 __extension__ extern __inline int16x4_t
16226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16227 vext_s16 (int16x4_t __a, int16x4_t __b, __const int __c)
16228 {
16229   __AARCH64_LANE_CHECK (__a, __c);
16230 #ifdef __AARCH64EB__
16231   return __builtin_shuffle (__b, __a,
16232       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16233 #else
16234   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16235 #endif
16236 }
16237
16238 __extension__ extern __inline int32x2_t
16239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16240 vext_s32 (int32x2_t __a, int32x2_t __b, __const int __c)
16241 {
16242   __AARCH64_LANE_CHECK (__a, __c);
16243 #ifdef __AARCH64EB__
16244   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16245 #else
16246   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16247 #endif
16248 }
16249
16250 __extension__ extern __inline int64x1_t
16251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16252 vext_s64 (int64x1_t __a, int64x1_t __b, __const int __c)
16253 {
16254   __AARCH64_LANE_CHECK (__a, __c);
16255   /* The only possible index to the assembler instruction returns element 0.  */
16256   return __a;
16257 }
16258
16259 __extension__ extern __inline uint8x8_t
16260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16261 vext_u8 (uint8x8_t __a, uint8x8_t __b, __const int __c)
16262 {
16263   __AARCH64_LANE_CHECK (__a, __c);
16264 #ifdef __AARCH64EB__
16265   return __builtin_shuffle (__b, __a, (uint8x8_t)
16266       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16267 #else
16268   return __builtin_shuffle (__a, __b,
16269       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16270 #endif
16271 }
16272
16273 __extension__ extern __inline uint16x4_t
16274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16275 vext_u16 (uint16x4_t __a, uint16x4_t __b, __const int __c)
16276 {
16277   __AARCH64_LANE_CHECK (__a, __c);
16278 #ifdef __AARCH64EB__
16279   return __builtin_shuffle (__b, __a,
16280       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16281 #else
16282   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16283 #endif
16284 }
16285
16286 __extension__ extern __inline uint32x2_t
16287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16288 vext_u32 (uint32x2_t __a, uint32x2_t __b, __const int __c)
16289 {
16290   __AARCH64_LANE_CHECK (__a, __c);
16291 #ifdef __AARCH64EB__
16292   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16293 #else
16294   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16295 #endif
16296 }
16297
16298 __extension__ extern __inline uint64x1_t
16299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16300 vext_u64 (uint64x1_t __a, uint64x1_t __b, __const int __c)
16301 {
16302   __AARCH64_LANE_CHECK (__a, __c);
16303   /* The only possible index to the assembler instruction returns element 0.  */
16304   return __a;
16305 }
16306
16307 __extension__ extern __inline float16x8_t
16308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16309 vextq_f16 (float16x8_t __a, float16x8_t __b, __const int __c)
16310 {
16311   __AARCH64_LANE_CHECK (__a, __c);
16312 #ifdef __AARCH64EB__
16313   return __builtin_shuffle (__b, __a,
16314                             (uint16x8_t) {8 - __c, 9 - __c, 10 - __c, 11 - __c,
16315                                           12 - __c, 13 - __c, 14 - __c,
16316                                           15 - __c});
16317 #else
16318   return __builtin_shuffle (__a, __b,
16319                             (uint16x8_t) {__c, __c + 1, __c + 2, __c + 3,
16320                                           __c + 4, __c + 5, __c + 6, __c + 7});
16321 #endif
16322 }
16323
16324 __extension__ extern __inline float32x4_t
16325 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16326 vextq_f32 (float32x4_t __a, float32x4_t __b, __const int __c)
16327 {
16328   __AARCH64_LANE_CHECK (__a, __c);
16329 #ifdef __AARCH64EB__
16330   return __builtin_shuffle (__b, __a,
16331       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16332 #else
16333   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16334 #endif
16335 }
16336
16337 __extension__ extern __inline float64x2_t
16338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16339 vextq_f64 (float64x2_t __a, float64x2_t __b, __const int __c)
16340 {
16341   __AARCH64_LANE_CHECK (__a, __c);
16342 #ifdef __AARCH64EB__
16343   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16344 #else
16345   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16346 #endif
16347 }
16348
16349 __extension__ extern __inline poly8x16_t
16350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16351 vextq_p8 (poly8x16_t __a, poly8x16_t __b, __const int __c)
16352 {
16353   __AARCH64_LANE_CHECK (__a, __c);
16354 #ifdef __AARCH64EB__
16355   return __builtin_shuffle (__b, __a, (uint8x16_t)
16356       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16357        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16358 #else
16359   return __builtin_shuffle (__a, __b, (uint8x16_t)
16360       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16361        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16362 #endif
16363 }
16364
16365 __extension__ extern __inline poly16x8_t
16366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16367 vextq_p16 (poly16x8_t __a, poly16x8_t __b, __const int __c)
16368 {
16369   __AARCH64_LANE_CHECK (__a, __c);
16370 #ifdef __AARCH64EB__
16371   return __builtin_shuffle (__b, __a, (uint16x8_t)
16372       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16373 #else
16374   return __builtin_shuffle (__a, __b,
16375       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16376 #endif
16377 }
16378
16379 __extension__ extern __inline poly64x2_t
16380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16381 vextq_p64 (poly64x2_t __a, poly64x2_t __b, __const int __c)
16382 {
16383   __AARCH64_LANE_CHECK (__a, __c);
16384 #ifdef __AARCH64EB__
16385   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16386 #else
16387   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16388 #endif
16389 }
16390
16391 __extension__ extern __inline int8x16_t
16392 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16393 vextq_s8 (int8x16_t __a, int8x16_t __b, __const int __c)
16394 {
16395   __AARCH64_LANE_CHECK (__a, __c);
16396 #ifdef __AARCH64EB__
16397   return __builtin_shuffle (__b, __a, (uint8x16_t)
16398       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16399        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16400 #else
16401   return __builtin_shuffle (__a, __b, (uint8x16_t)
16402       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16403        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16404 #endif
16405 }
16406
16407 __extension__ extern __inline int16x8_t
16408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16409 vextq_s16 (int16x8_t __a, int16x8_t __b, __const int __c)
16410 {
16411   __AARCH64_LANE_CHECK (__a, __c);
16412 #ifdef __AARCH64EB__
16413   return __builtin_shuffle (__b, __a, (uint16x8_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       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16418 #endif
16419 }
16420
16421 __extension__ extern __inline int32x4_t
16422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16423 vextq_s32 (int32x4_t __a, int32x4_t __b, __const int __c)
16424 {
16425   __AARCH64_LANE_CHECK (__a, __c);
16426 #ifdef __AARCH64EB__
16427   return __builtin_shuffle (__b, __a,
16428       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16429 #else
16430   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16431 #endif
16432 }
16433
16434 __extension__ extern __inline int64x2_t
16435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16436 vextq_s64 (int64x2_t __a, int64x2_t __b, __const int __c)
16437 {
16438   __AARCH64_LANE_CHECK (__a, __c);
16439 #ifdef __AARCH64EB__
16440   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16441 #else
16442   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16443 #endif
16444 }
16445
16446 __extension__ extern __inline uint8x16_t
16447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16448 vextq_u8 (uint8x16_t __a, uint8x16_t __b, __const int __c)
16449 {
16450   __AARCH64_LANE_CHECK (__a, __c);
16451 #ifdef __AARCH64EB__
16452   return __builtin_shuffle (__b, __a, (uint8x16_t)
16453       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16454        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16455 #else
16456   return __builtin_shuffle (__a, __b, (uint8x16_t)
16457       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16458        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16459 #endif
16460 }
16461
16462 __extension__ extern __inline uint16x8_t
16463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16464 vextq_u16 (uint16x8_t __a, uint16x8_t __b, __const int __c)
16465 {
16466   __AARCH64_LANE_CHECK (__a, __c);
16467 #ifdef __AARCH64EB__
16468   return __builtin_shuffle (__b, __a, (uint16x8_t)
16469       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16470 #else
16471   return __builtin_shuffle (__a, __b,
16472       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16473 #endif
16474 }
16475
16476 __extension__ extern __inline uint32x4_t
16477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16478 vextq_u32 (uint32x4_t __a, uint32x4_t __b, __const int __c)
16479 {
16480   __AARCH64_LANE_CHECK (__a, __c);
16481 #ifdef __AARCH64EB__
16482   return __builtin_shuffle (__b, __a,
16483       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16484 #else
16485   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16486 #endif
16487 }
16488
16489 __extension__ extern __inline uint64x2_t
16490 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16491 vextq_u64 (uint64x2_t __a, uint64x2_t __b, __const int __c)
16492 {
16493   __AARCH64_LANE_CHECK (__a, __c);
16494 #ifdef __AARCH64EB__
16495   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16496 #else
16497   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16498 #endif
16499 }
16500
16501 /* vfma  */
16502
16503 __extension__ extern __inline float64x1_t
16504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16505 vfma_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
16506 {
16507   return (float64x1_t) {__builtin_fma (__b[0], __c[0], __a[0])};
16508 }
16509
16510 __extension__ extern __inline float32x2_t
16511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16512 vfma_f32 (float32x2_t __a, float32x2_t __b, float32x2_t __c)
16513 {
16514   return __builtin_aarch64_fmav2sf (__b, __c, __a);
16515 }
16516
16517 __extension__ extern __inline float32x4_t
16518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16519 vfmaq_f32 (float32x4_t __a, float32x4_t __b, float32x4_t __c)
16520 {
16521   return __builtin_aarch64_fmav4sf (__b, __c, __a);
16522 }
16523
16524 __extension__ extern __inline float64x2_t
16525 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16526 vfmaq_f64 (float64x2_t __a, float64x2_t __b, float64x2_t __c)
16527 {
16528   return __builtin_aarch64_fmav2df (__b, __c, __a);
16529 }
16530
16531 __extension__ extern __inline float32x2_t
16532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16533 vfma_n_f32 (float32x2_t __a, float32x2_t __b, float32_t __c)
16534 {
16535   return __builtin_aarch64_fmav2sf (__b, vdup_n_f32 (__c), __a);
16536 }
16537
16538 __extension__ extern __inline float64x1_t
16539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16540 vfma_n_f64 (float64x1_t __a, float64x1_t __b, float64_t __c)
16541 {
16542   return (float64x1_t) {__b[0] * __c + __a[0]};
16543 }
16544
16545 __extension__ extern __inline float32x4_t
16546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16547 vfmaq_n_f32 (float32x4_t __a, float32x4_t __b, float32_t __c)
16548 {
16549   return __builtin_aarch64_fmav4sf (__b, vdupq_n_f32 (__c), __a);
16550 }
16551
16552 __extension__ extern __inline float64x2_t
16553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16554 vfmaq_n_f64 (float64x2_t __a, float64x2_t __b, float64_t __c)
16555 {
16556   return __builtin_aarch64_fmav2df (__b, vdupq_n_f64 (__c), __a);
16557 }
16558
16559 /* vfma_lane  */
16560
16561 __extension__ extern __inline float32x2_t
16562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16563 vfma_lane_f32 (float32x2_t __a, float32x2_t __b,
16564                float32x2_t __c, const int __lane)
16565 {
16566   return __builtin_aarch64_fmav2sf (__b,
16567                                     __aarch64_vdup_lane_f32 (__c, __lane),
16568                                     __a);
16569 }
16570
16571 __extension__ extern __inline float64x1_t
16572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16573 vfma_lane_f64 (float64x1_t __a, float64x1_t __b,
16574                float64x1_t __c, const int __lane)
16575 {
16576   return (float64x1_t) {__builtin_fma (__b[0], __c[0], __a[0])};
16577 }
16578
16579 __extension__ extern __inline float64_t
16580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16581 vfmad_lane_f64 (float64_t __a, float64_t __b,
16582                 float64x1_t __c, const int __lane)
16583 {
16584   return __builtin_fma (__b, __c[0], __a);
16585 }
16586
16587 __extension__ extern __inline float32_t
16588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16589 vfmas_lane_f32 (float32_t __a, float32_t __b,
16590                 float32x2_t __c, const int __lane)
16591 {
16592   return __builtin_fmaf (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16593 }
16594
16595 /* vfma_laneq  */
16596
16597 __extension__ extern __inline float32x2_t
16598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16599 vfma_laneq_f32 (float32x2_t __a, float32x2_t __b,
16600                 float32x4_t __c, const int __lane)
16601 {
16602   return __builtin_aarch64_fmav2sf (__b,
16603                                     __aarch64_vdup_laneq_f32 (__c, __lane),
16604                                     __a);
16605 }
16606
16607 __extension__ extern __inline float64x1_t
16608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16609 vfma_laneq_f64 (float64x1_t __a, float64x1_t __b,
16610                 float64x2_t __c, const int __lane)
16611 {
16612   float64_t __c0 = __aarch64_vget_lane_any (__c, __lane);
16613   return (float64x1_t) {__builtin_fma (__b[0], __c0, __a[0])};
16614 }
16615
16616 __extension__ extern __inline float64_t
16617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16618 vfmad_laneq_f64 (float64_t __a, float64_t __b,
16619                  float64x2_t __c, const int __lane)
16620 {
16621   return __builtin_fma (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16622 }
16623
16624 __extension__ extern __inline float32_t
16625 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16626 vfmas_laneq_f32 (float32_t __a, float32_t __b,
16627                  float32x4_t __c, const int __lane)
16628 {
16629   return __builtin_fmaf (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16630 }
16631
16632 /* vfmaq_lane  */
16633
16634 __extension__ extern __inline float32x4_t
16635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16636 vfmaq_lane_f32 (float32x4_t __a, float32x4_t __b,
16637                 float32x2_t __c, const int __lane)
16638 {
16639   return __builtin_aarch64_fmav4sf (__b,
16640                                     __aarch64_vdupq_lane_f32 (__c, __lane),
16641                                     __a);
16642 }
16643
16644 __extension__ extern __inline float64x2_t
16645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16646 vfmaq_lane_f64 (float64x2_t __a, float64x2_t __b,
16647                 float64x1_t __c, const int __lane)
16648 {
16649   return __builtin_aarch64_fmav2df (__b, vdupq_n_f64 (__c[0]), __a);
16650 }
16651
16652 /* vfmaq_laneq  */
16653
16654 __extension__ extern __inline float32x4_t
16655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16656 vfmaq_laneq_f32 (float32x4_t __a, float32x4_t __b,
16657                  float32x4_t __c, const int __lane)
16658 {
16659   return __builtin_aarch64_fmav4sf (__b,
16660                                     __aarch64_vdupq_laneq_f32 (__c, __lane),
16661                                     __a);
16662 }
16663
16664 __extension__ extern __inline float64x2_t
16665 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16666 vfmaq_laneq_f64 (float64x2_t __a, float64x2_t __b,
16667                  float64x2_t __c, const int __lane)
16668 {
16669   return __builtin_aarch64_fmav2df (__b,
16670                                     __aarch64_vdupq_laneq_f64 (__c, __lane),
16671                                     __a);
16672 }
16673
16674 /* vfms  */
16675
16676 __extension__ extern __inline float64x1_t
16677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16678 vfms_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
16679 {
16680   return (float64x1_t) {__builtin_fma (-__b[0], __c[0], __a[0])};
16681 }
16682
16683 __extension__ extern __inline float32x2_t
16684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16685 vfms_f32 (float32x2_t __a, float32x2_t __b, float32x2_t __c)
16686 {
16687   return __builtin_aarch64_fmav2sf (-__b, __c, __a);
16688 }
16689
16690 __extension__ extern __inline float32x4_t
16691 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16692 vfmsq_f32 (float32x4_t __a, float32x4_t __b, float32x4_t __c)
16693 {
16694   return __builtin_aarch64_fmav4sf (-__b, __c, __a);
16695 }
16696
16697 __extension__ extern __inline float64x2_t
16698 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16699 vfmsq_f64 (float64x2_t __a, float64x2_t __b, float64x2_t __c)
16700 {
16701   return __builtin_aarch64_fmav2df (-__b, __c, __a);
16702 }
16703
16704 __extension__ extern __inline float32x2_t
16705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16706 vfms_n_f32 (float32x2_t __a, float32x2_t __b, float32_t __c)
16707 {
16708   return __builtin_aarch64_fmav2sf (-__b, vdup_n_f32 (__c), __a);
16709 }
16710
16711 __extension__ extern __inline float64x1_t
16712 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16713 vfms_n_f64 (float64x1_t __a, float64x1_t __b, float64_t __c)
16714 {
16715   return (float64x1_t) {-__b[0] * __c + __a[0]};
16716 }
16717
16718 __extension__ extern __inline float32x4_t
16719 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16720 vfmsq_n_f32 (float32x4_t __a, float32x4_t __b, float32_t __c)
16721 {
16722   return __builtin_aarch64_fmav4sf (-__b, vdupq_n_f32 (__c), __a);
16723 }
16724
16725 __extension__ extern __inline float64x2_t
16726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16727 vfmsq_n_f64 (float64x2_t __a, float64x2_t __b, float64_t __c)
16728 {
16729   return __builtin_aarch64_fmav2df (-__b, vdupq_n_f64 (__c), __a);
16730 }
16731
16732 /* vfms_lane  */
16733
16734 __extension__ extern __inline float32x2_t
16735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16736 vfms_lane_f32 (float32x2_t __a, float32x2_t __b,
16737                float32x2_t __c, const int __lane)
16738 {
16739   return __builtin_aarch64_fmav2sf (-__b,
16740                                     __aarch64_vdup_lane_f32 (__c, __lane),
16741                                     __a);
16742 }
16743
16744 __extension__ extern __inline float64x1_t
16745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16746 vfms_lane_f64 (float64x1_t __a, float64x1_t __b,
16747                float64x1_t __c, const int __lane)
16748 {
16749   return (float64x1_t) {__builtin_fma (-__b[0], __c[0], __a[0])};
16750 }
16751
16752 __extension__ extern __inline float64_t
16753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16754 vfmsd_lane_f64 (float64_t __a, float64_t __b,
16755                 float64x1_t __c, const int __lane)
16756 {
16757   return __builtin_fma (-__b, __c[0], __a);
16758 }
16759
16760 __extension__ extern __inline float32_t
16761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16762 vfmss_lane_f32 (float32_t __a, float32_t __b,
16763                 float32x2_t __c, const int __lane)
16764 {
16765   return __builtin_fmaf (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16766 }
16767
16768 /* vfms_laneq  */
16769
16770 __extension__ extern __inline float32x2_t
16771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16772 vfms_laneq_f32 (float32x2_t __a, float32x2_t __b,
16773                 float32x4_t __c, const int __lane)
16774 {
16775   return __builtin_aarch64_fmav2sf (-__b,
16776                                     __aarch64_vdup_laneq_f32 (__c, __lane),
16777                                     __a);
16778 }
16779
16780 __extension__ extern __inline float64x1_t
16781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16782 vfms_laneq_f64 (float64x1_t __a, float64x1_t __b,
16783                 float64x2_t __c, const int __lane)
16784 {
16785   float64_t __c0 = __aarch64_vget_lane_any (__c, __lane);
16786   return (float64x1_t) {__builtin_fma (-__b[0], __c0, __a[0])};
16787 }
16788
16789 __extension__ extern __inline float64_t
16790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16791 vfmsd_laneq_f64 (float64_t __a, float64_t __b,
16792                  float64x2_t __c, const int __lane)
16793 {
16794   return __builtin_fma (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16795 }
16796
16797 __extension__ extern __inline float32_t
16798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16799 vfmss_laneq_f32 (float32_t __a, float32_t __b,
16800                  float32x4_t __c, const int __lane)
16801 {
16802   return __builtin_fmaf (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16803 }
16804
16805 /* vfmsq_lane  */
16806
16807 __extension__ extern __inline float32x4_t
16808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16809 vfmsq_lane_f32 (float32x4_t __a, float32x4_t __b,
16810                 float32x2_t __c, const int __lane)
16811 {
16812   return __builtin_aarch64_fmav4sf (-__b,
16813                                     __aarch64_vdupq_lane_f32 (__c, __lane),
16814                                     __a);
16815 }
16816
16817 __extension__ extern __inline float64x2_t
16818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16819 vfmsq_lane_f64 (float64x2_t __a, float64x2_t __b,
16820                 float64x1_t __c, const int __lane)
16821 {
16822   return __builtin_aarch64_fmav2df (-__b, vdupq_n_f64 (__c[0]), __a);
16823 }
16824
16825 /* vfmsq_laneq  */
16826
16827 __extension__ extern __inline float32x4_t
16828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16829 vfmsq_laneq_f32 (float32x4_t __a, float32x4_t __b,
16830                  float32x4_t __c, const int __lane)
16831 {
16832   return __builtin_aarch64_fmav4sf (-__b,
16833                                     __aarch64_vdupq_laneq_f32 (__c, __lane),
16834                                     __a);
16835 }
16836
16837 __extension__ extern __inline float64x2_t
16838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16839 vfmsq_laneq_f64 (float64x2_t __a, float64x2_t __b,
16840                  float64x2_t __c, const int __lane)
16841 {
16842   return __builtin_aarch64_fmav2df (-__b,
16843                                     __aarch64_vdupq_laneq_f64 (__c, __lane),
16844                                     __a);
16845 }
16846
16847 /* vld1 */
16848
16849 __extension__ extern __inline float16x4_t
16850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16851 vld1_f16 (const float16_t *__a)
16852 {
16853   return __builtin_aarch64_ld1v4hf (__a);
16854 }
16855
16856 __extension__ extern __inline float32x2_t
16857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16858 vld1_f32 (const float32_t *a)
16859 {
16860   return __builtin_aarch64_ld1v2sf ((const __builtin_aarch64_simd_sf *) a);
16861 }
16862
16863 __extension__ extern __inline float64x1_t
16864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16865 vld1_f64 (const float64_t *a)
16866 {
16867   return (float64x1_t) {*a};
16868 }
16869
16870 __extension__ extern __inline poly8x8_t
16871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16872 vld1_p8 (const poly8_t *a)
16873 {
16874   return (poly8x8_t)
16875     __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
16876 }
16877
16878 __extension__ extern __inline poly16x4_t
16879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16880 vld1_p16 (const poly16_t *a)
16881 {
16882   return (poly16x4_t)
16883     __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
16884 }
16885
16886 __extension__ extern __inline poly64x1_t
16887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16888 vld1_p64 (const poly64_t *a)
16889 {
16890   return (poly64x1_t) {*a};
16891 }
16892
16893 __extension__ extern __inline int8x8_t
16894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16895 vld1_s8 (const int8_t *a)
16896 {
16897   return __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
16898 }
16899
16900 __extension__ extern __inline int16x4_t
16901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16902 vld1_s16 (const int16_t *a)
16903 {
16904   return __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
16905 }
16906
16907 __extension__ extern __inline int32x2_t
16908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16909 vld1_s32 (const int32_t *a)
16910 {
16911   return __builtin_aarch64_ld1v2si ((const __builtin_aarch64_simd_si *) a);
16912 }
16913
16914 __extension__ extern __inline int64x1_t
16915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16916 vld1_s64 (const int64_t *a)
16917 {
16918   return (int64x1_t) {*a};
16919 }
16920
16921 __extension__ extern __inline uint8x8_t
16922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16923 vld1_u8 (const uint8_t *a)
16924 {
16925   return (uint8x8_t)
16926     __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
16927 }
16928
16929 __extension__ extern __inline uint16x4_t
16930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16931 vld1_u16 (const uint16_t *a)
16932 {
16933   return (uint16x4_t)
16934     __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
16935 }
16936
16937 __extension__ extern __inline uint32x2_t
16938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16939 vld1_u32 (const uint32_t *a)
16940 {
16941   return (uint32x2_t)
16942     __builtin_aarch64_ld1v2si ((const __builtin_aarch64_simd_si *) a);
16943 }
16944
16945 __extension__ extern __inline uint64x1_t
16946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16947 vld1_u64 (const uint64_t *a)
16948 {
16949   return (uint64x1_t) {*a};
16950 }
16951
16952 /* vld1q */
16953
16954 __extension__ extern __inline float16x8_t
16955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16956 vld1q_f16 (const float16_t *__a)
16957 {
16958   return __builtin_aarch64_ld1v8hf (__a);
16959 }
16960
16961 __extension__ extern __inline float32x4_t
16962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16963 vld1q_f32 (const float32_t *a)
16964 {
16965   return __builtin_aarch64_ld1v4sf ((const __builtin_aarch64_simd_sf *) a);
16966 }
16967
16968 __extension__ extern __inline float64x2_t
16969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16970 vld1q_f64 (const float64_t *a)
16971 {
16972   return __builtin_aarch64_ld1v2df ((const __builtin_aarch64_simd_df *) a);
16973 }
16974
16975 __extension__ extern __inline poly8x16_t
16976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16977 vld1q_p8 (const poly8_t *a)
16978 {
16979   return (poly8x16_t)
16980     __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
16981 }
16982
16983 __extension__ extern __inline poly16x8_t
16984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16985 vld1q_p16 (const poly16_t *a)
16986 {
16987   return (poly16x8_t)
16988     __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
16989 }
16990
16991 __extension__ extern __inline poly64x2_t
16992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16993 vld1q_p64 (const poly64_t *a)
16994 {
16995   return (poly64x2_t)
16996     __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
16997 }
16998
16999 __extension__ extern __inline int8x16_t
17000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17001 vld1q_s8 (const int8_t *a)
17002 {
17003   return __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
17004 }
17005
17006 __extension__ extern __inline int16x8_t
17007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17008 vld1q_s16 (const int16_t *a)
17009 {
17010   return __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
17011 }
17012
17013 __extension__ extern __inline int32x4_t
17014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17015 vld1q_s32 (const int32_t *a)
17016 {
17017   return __builtin_aarch64_ld1v4si ((const __builtin_aarch64_simd_si *) a);
17018 }
17019
17020 __extension__ extern __inline int64x2_t
17021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17022 vld1q_s64 (const int64_t *a)
17023 {
17024   return __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
17025 }
17026
17027 __extension__ extern __inline uint8x16_t
17028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17029 vld1q_u8 (const uint8_t *a)
17030 {
17031   return (uint8x16_t)
17032     __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
17033 }
17034
17035 __extension__ extern __inline uint16x8_t
17036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17037 vld1q_u16 (const uint16_t *a)
17038 {
17039   return (uint16x8_t)
17040     __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
17041 }
17042
17043 __extension__ extern __inline uint32x4_t
17044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17045 vld1q_u32 (const uint32_t *a)
17046 {
17047   return (uint32x4_t)
17048     __builtin_aarch64_ld1v4si ((const __builtin_aarch64_simd_si *) a);
17049 }
17050
17051 __extension__ extern __inline uint64x2_t
17052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17053 vld1q_u64 (const uint64_t *a)
17054 {
17055   return (uint64x2_t)
17056     __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
17057 }
17058
17059 /* vld1_dup  */
17060
17061 __extension__ extern __inline float16x4_t
17062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17063 vld1_dup_f16 (const float16_t* __a)
17064 {
17065   return vdup_n_f16 (*__a);
17066 }
17067
17068 __extension__ extern __inline float32x2_t
17069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17070 vld1_dup_f32 (const float32_t* __a)
17071 {
17072   return vdup_n_f32 (*__a);
17073 }
17074
17075 __extension__ extern __inline float64x1_t
17076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17077 vld1_dup_f64 (const float64_t* __a)
17078 {
17079   return vdup_n_f64 (*__a);
17080 }
17081
17082 __extension__ extern __inline poly8x8_t
17083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17084 vld1_dup_p8 (const poly8_t* __a)
17085 {
17086   return vdup_n_p8 (*__a);
17087 }
17088
17089 __extension__ extern __inline poly16x4_t
17090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17091 vld1_dup_p16 (const poly16_t* __a)
17092 {
17093   return vdup_n_p16 (*__a);
17094 }
17095
17096 __extension__ extern __inline poly64x1_t
17097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17098 vld1_dup_p64 (const poly64_t* __a)
17099 {
17100   return vdup_n_p64 (*__a);
17101 }
17102
17103 __extension__ extern __inline int8x8_t
17104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17105 vld1_dup_s8 (const int8_t* __a)
17106 {
17107   return vdup_n_s8 (*__a);
17108 }
17109
17110 __extension__ extern __inline int16x4_t
17111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17112 vld1_dup_s16 (const int16_t* __a)
17113 {
17114   return vdup_n_s16 (*__a);
17115 }
17116
17117 __extension__ extern __inline int32x2_t
17118 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17119 vld1_dup_s32 (const int32_t* __a)
17120 {
17121   return vdup_n_s32 (*__a);
17122 }
17123
17124 __extension__ extern __inline int64x1_t
17125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17126 vld1_dup_s64 (const int64_t* __a)
17127 {
17128   return vdup_n_s64 (*__a);
17129 }
17130
17131 __extension__ extern __inline uint8x8_t
17132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17133 vld1_dup_u8 (const uint8_t* __a)
17134 {
17135   return vdup_n_u8 (*__a);
17136 }
17137
17138 __extension__ extern __inline uint16x4_t
17139 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17140 vld1_dup_u16 (const uint16_t* __a)
17141 {
17142   return vdup_n_u16 (*__a);
17143 }
17144
17145 __extension__ extern __inline uint32x2_t
17146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17147 vld1_dup_u32 (const uint32_t* __a)
17148 {
17149   return vdup_n_u32 (*__a);
17150 }
17151
17152 __extension__ extern __inline uint64x1_t
17153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17154 vld1_dup_u64 (const uint64_t* __a)
17155 {
17156   return vdup_n_u64 (*__a);
17157 }
17158
17159 /* vld1q_dup  */
17160
17161 __extension__ extern __inline float16x8_t
17162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17163 vld1q_dup_f16 (const float16_t* __a)
17164 {
17165   return vdupq_n_f16 (*__a);
17166 }
17167
17168 __extension__ extern __inline float32x4_t
17169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17170 vld1q_dup_f32 (const float32_t* __a)
17171 {
17172   return vdupq_n_f32 (*__a);
17173 }
17174
17175 __extension__ extern __inline float64x2_t
17176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17177 vld1q_dup_f64 (const float64_t* __a)
17178 {
17179   return vdupq_n_f64 (*__a);
17180 }
17181
17182 __extension__ extern __inline poly8x16_t
17183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17184 vld1q_dup_p8 (const poly8_t* __a)
17185 {
17186   return vdupq_n_p8 (*__a);
17187 }
17188
17189 __extension__ extern __inline poly16x8_t
17190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17191 vld1q_dup_p16 (const poly16_t* __a)
17192 {
17193   return vdupq_n_p16 (*__a);
17194 }
17195
17196 __extension__ extern __inline poly64x2_t
17197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17198 vld1q_dup_p64 (const poly64_t* __a)
17199 {
17200   return vdupq_n_p64 (*__a);
17201 }
17202
17203  __extension__ extern __inline int8x16_t
17204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17205 vld1q_dup_s8 (const int8_t* __a)
17206 {
17207   return vdupq_n_s8 (*__a);
17208 }
17209
17210 __extension__ extern __inline int16x8_t
17211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17212 vld1q_dup_s16 (const int16_t* __a)
17213 {
17214   return vdupq_n_s16 (*__a);
17215 }
17216
17217 __extension__ extern __inline int32x4_t
17218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17219 vld1q_dup_s32 (const int32_t* __a)
17220 {
17221   return vdupq_n_s32 (*__a);
17222 }
17223
17224 __extension__ extern __inline int64x2_t
17225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17226 vld1q_dup_s64 (const int64_t* __a)
17227 {
17228   return vdupq_n_s64 (*__a);
17229 }
17230
17231 __extension__ extern __inline uint8x16_t
17232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17233 vld1q_dup_u8 (const uint8_t* __a)
17234 {
17235   return vdupq_n_u8 (*__a);
17236 }
17237
17238 __extension__ extern __inline uint16x8_t
17239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17240 vld1q_dup_u16 (const uint16_t* __a)
17241 {
17242   return vdupq_n_u16 (*__a);
17243 }
17244
17245 __extension__ extern __inline uint32x4_t
17246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17247 vld1q_dup_u32 (const uint32_t* __a)
17248 {
17249   return vdupq_n_u32 (*__a);
17250 }
17251
17252 __extension__ extern __inline uint64x2_t
17253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17254 vld1q_dup_u64 (const uint64_t* __a)
17255 {
17256   return vdupq_n_u64 (*__a);
17257 }
17258
17259 /* vld1_lane  */
17260
17261 __extension__ extern __inline float16x4_t
17262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17263 vld1_lane_f16 (const float16_t *__src, float16x4_t __vec, const int __lane)
17264 {
17265   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17266 }
17267
17268 __extension__ extern __inline float32x2_t
17269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17270 vld1_lane_f32 (const float32_t *__src, float32x2_t __vec, const int __lane)
17271 {
17272   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17273 }
17274
17275 __extension__ extern __inline float64x1_t
17276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17277 vld1_lane_f64 (const float64_t *__src, float64x1_t __vec, const int __lane)
17278 {
17279   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17280 }
17281
17282 __extension__ extern __inline poly8x8_t
17283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17284 vld1_lane_p8 (const poly8_t *__src, poly8x8_t __vec, const int __lane)
17285 {
17286   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17287 }
17288
17289 __extension__ extern __inline poly16x4_t
17290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17291 vld1_lane_p16 (const poly16_t *__src, poly16x4_t __vec, const int __lane)
17292 {
17293   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17294 }
17295
17296 __extension__ extern __inline poly64x1_t
17297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17298 vld1_lane_p64 (const poly64_t *__src, poly64x1_t __vec, const int __lane)
17299 {
17300   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17301 }
17302
17303 __extension__ extern __inline int8x8_t
17304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17305 vld1_lane_s8 (const int8_t *__src, int8x8_t __vec, const int __lane)
17306 {
17307   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17308 }
17309
17310 __extension__ extern __inline int16x4_t
17311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17312 vld1_lane_s16 (const int16_t *__src, int16x4_t __vec, const int __lane)
17313 {
17314   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17315 }
17316
17317 __extension__ extern __inline int32x2_t
17318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17319 vld1_lane_s32 (const int32_t *__src, int32x2_t __vec, const int __lane)
17320 {
17321   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17322 }
17323
17324 __extension__ extern __inline int64x1_t
17325 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17326 vld1_lane_s64 (const int64_t *__src, int64x1_t __vec, const int __lane)
17327 {
17328   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17329 }
17330
17331 __extension__ extern __inline uint8x8_t
17332 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17333 vld1_lane_u8 (const uint8_t *__src, uint8x8_t __vec, const int __lane)
17334 {
17335   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17336 }
17337
17338 __extension__ extern __inline uint16x4_t
17339 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17340 vld1_lane_u16 (const uint16_t *__src, uint16x4_t __vec, const int __lane)
17341 {
17342   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17343 }
17344
17345 __extension__ extern __inline uint32x2_t
17346 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17347 vld1_lane_u32 (const uint32_t *__src, uint32x2_t __vec, const int __lane)
17348 {
17349   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17350 }
17351
17352 __extension__ extern __inline uint64x1_t
17353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17354 vld1_lane_u64 (const uint64_t *__src, uint64x1_t __vec, const int __lane)
17355 {
17356   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17357 }
17358
17359 /* vld1q_lane  */
17360
17361 __extension__ extern __inline float16x8_t
17362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17363 vld1q_lane_f16 (const float16_t *__src, float16x8_t __vec, const int __lane)
17364 {
17365   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17366 }
17367
17368 __extension__ extern __inline float32x4_t
17369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17370 vld1q_lane_f32 (const float32_t *__src, float32x4_t __vec, const int __lane)
17371 {
17372   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17373 }
17374
17375 __extension__ extern __inline float64x2_t
17376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17377 vld1q_lane_f64 (const float64_t *__src, float64x2_t __vec, const int __lane)
17378 {
17379   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17380 }
17381
17382 __extension__ extern __inline poly8x16_t
17383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17384 vld1q_lane_p8 (const poly8_t *__src, poly8x16_t __vec, const int __lane)
17385 {
17386   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17387 }
17388
17389 __extension__ extern __inline poly16x8_t
17390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17391 vld1q_lane_p16 (const poly16_t *__src, poly16x8_t __vec, const int __lane)
17392 {
17393   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17394 }
17395
17396 __extension__ extern __inline poly64x2_t
17397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17398 vld1q_lane_p64 (const poly64_t *__src, poly64x2_t __vec, const int __lane)
17399 {
17400   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17401 }
17402
17403 __extension__ extern __inline int8x16_t
17404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17405 vld1q_lane_s8 (const int8_t *__src, int8x16_t __vec, const int __lane)
17406 {
17407   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17408 }
17409
17410 __extension__ extern __inline int16x8_t
17411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17412 vld1q_lane_s16 (const int16_t *__src, int16x8_t __vec, const int __lane)
17413 {
17414   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17415 }
17416
17417 __extension__ extern __inline int32x4_t
17418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17419 vld1q_lane_s32 (const int32_t *__src, int32x4_t __vec, const int __lane)
17420 {
17421   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17422 }
17423
17424 __extension__ extern __inline int64x2_t
17425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17426 vld1q_lane_s64 (const int64_t *__src, int64x2_t __vec, const int __lane)
17427 {
17428   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17429 }
17430
17431 __extension__ extern __inline uint8x16_t
17432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17433 vld1q_lane_u8 (const uint8_t *__src, uint8x16_t __vec, const int __lane)
17434 {
17435   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17436 }
17437
17438 __extension__ extern __inline uint16x8_t
17439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17440 vld1q_lane_u16 (const uint16_t *__src, uint16x8_t __vec, const int __lane)
17441 {
17442   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17443 }
17444
17445 __extension__ extern __inline uint32x4_t
17446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17447 vld1q_lane_u32 (const uint32_t *__src, uint32x4_t __vec, const int __lane)
17448 {
17449   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17450 }
17451
17452 __extension__ extern __inline uint64x2_t
17453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17454 vld1q_lane_u64 (const uint64_t *__src, uint64x2_t __vec, const int __lane)
17455 {
17456   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17457 }
17458
17459 /* vldn */
17460
17461 __extension__ extern __inline int64x1x2_t
17462 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17463 vld2_s64 (const int64_t * __a)
17464 {
17465   int64x1x2_t ret;
17466   __builtin_aarch64_simd_oi __o;
17467   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
17468   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17469   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17470   return ret;
17471 }
17472
17473 __extension__ extern __inline uint64x1x2_t
17474 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17475 vld2_u64 (const uint64_t * __a)
17476 {
17477   uint64x1x2_t ret;
17478   __builtin_aarch64_simd_oi __o;
17479   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
17480   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17481   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17482   return ret;
17483 }
17484
17485 __extension__ extern __inline float64x1x2_t
17486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17487 vld2_f64 (const float64_t * __a)
17488 {
17489   float64x1x2_t ret;
17490   __builtin_aarch64_simd_oi __o;
17491   __o = __builtin_aarch64_ld2df ((const __builtin_aarch64_simd_df *) __a);
17492   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
17493   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
17494   return ret;
17495 }
17496
17497 __extension__ extern __inline int8x8x2_t
17498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17499 vld2_s8 (const int8_t * __a)
17500 {
17501   int8x8x2_t ret;
17502   __builtin_aarch64_simd_oi __o;
17503   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17504   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17505   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17506   return ret;
17507 }
17508
17509 __extension__ extern __inline poly8x8x2_t
17510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17511 vld2_p8 (const poly8_t * __a)
17512 {
17513   poly8x8x2_t ret;
17514   __builtin_aarch64_simd_oi __o;
17515   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17516   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17517   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17518   return ret;
17519 }
17520
17521 __extension__ extern __inline poly64x1x2_t
17522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17523 vld2_p64 (const poly64_t * __a)
17524 {
17525   poly64x1x2_t ret;
17526   __builtin_aarch64_simd_oi __o;
17527   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
17528   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 0);
17529   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 1);
17530   return ret;
17531 }
17532
17533 __extension__ extern __inline int16x4x2_t
17534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17535 vld2_s16 (const int16_t * __a)
17536 {
17537   int16x4x2_t ret;
17538   __builtin_aarch64_simd_oi __o;
17539   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17540   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17541   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17542   return ret;
17543 }
17544
17545 __extension__ extern __inline poly16x4x2_t
17546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17547 vld2_p16 (const poly16_t * __a)
17548 {
17549   poly16x4x2_t ret;
17550   __builtin_aarch64_simd_oi __o;
17551   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17552   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17553   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17554   return ret;
17555 }
17556
17557 __extension__ extern __inline int32x2x2_t
17558 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17559 vld2_s32 (const int32_t * __a)
17560 {
17561   int32x2x2_t ret;
17562   __builtin_aarch64_simd_oi __o;
17563   __o = __builtin_aarch64_ld2v2si ((const __builtin_aarch64_simd_si *) __a);
17564   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
17565   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
17566   return ret;
17567 }
17568
17569 __extension__ extern __inline uint8x8x2_t
17570 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17571 vld2_u8 (const uint8_t * __a)
17572 {
17573   uint8x8x2_t ret;
17574   __builtin_aarch64_simd_oi __o;
17575   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17576   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17577   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17578   return ret;
17579 }
17580
17581 __extension__ extern __inline uint16x4x2_t
17582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17583 vld2_u16 (const uint16_t * __a)
17584 {
17585   uint16x4x2_t ret;
17586   __builtin_aarch64_simd_oi __o;
17587   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17588   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17589   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17590   return ret;
17591 }
17592
17593 __extension__ extern __inline uint32x2x2_t
17594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17595 vld2_u32 (const uint32_t * __a)
17596 {
17597   uint32x2x2_t ret;
17598   __builtin_aarch64_simd_oi __o;
17599   __o = __builtin_aarch64_ld2v2si ((const __builtin_aarch64_simd_si *) __a);
17600   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
17601   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
17602   return ret;
17603 }
17604
17605 __extension__ extern __inline float16x4x2_t
17606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17607 vld2_f16 (const float16_t * __a)
17608 {
17609   float16x4x2_t ret;
17610   __builtin_aarch64_simd_oi __o;
17611   __o = __builtin_aarch64_ld2v4hf (__a);
17612   ret.val[0] = __builtin_aarch64_get_dregoiv4hf (__o, 0);
17613   ret.val[1] = __builtin_aarch64_get_dregoiv4hf (__o, 1);
17614   return ret;
17615 }
17616
17617 __extension__ extern __inline float32x2x2_t
17618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17619 vld2_f32 (const float32_t * __a)
17620 {
17621   float32x2x2_t ret;
17622   __builtin_aarch64_simd_oi __o;
17623   __o = __builtin_aarch64_ld2v2sf ((const __builtin_aarch64_simd_sf *) __a);
17624   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
17625   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
17626   return ret;
17627 }
17628
17629 __extension__ extern __inline int8x16x2_t
17630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17631 vld2q_s8 (const int8_t * __a)
17632 {
17633   int8x16x2_t ret;
17634   __builtin_aarch64_simd_oi __o;
17635   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17636   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17637   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17638   return ret;
17639 }
17640
17641 __extension__ extern __inline poly8x16x2_t
17642 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17643 vld2q_p8 (const poly8_t * __a)
17644 {
17645   poly8x16x2_t ret;
17646   __builtin_aarch64_simd_oi __o;
17647   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17648   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17649   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17650   return ret;
17651 }
17652
17653 __extension__ extern __inline int16x8x2_t
17654 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17655 vld2q_s16 (const int16_t * __a)
17656 {
17657   int16x8x2_t ret;
17658   __builtin_aarch64_simd_oi __o;
17659   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17660   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17661   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17662   return ret;
17663 }
17664
17665 __extension__ extern __inline poly16x8x2_t
17666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17667 vld2q_p16 (const poly16_t * __a)
17668 {
17669   poly16x8x2_t ret;
17670   __builtin_aarch64_simd_oi __o;
17671   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17672   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17673   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17674   return ret;
17675 }
17676
17677 __extension__ extern __inline poly64x2x2_t
17678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17679 vld2q_p64 (const poly64_t * __a)
17680 {
17681   poly64x2x2_t ret;
17682   __builtin_aarch64_simd_oi __o;
17683   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
17684   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 0);
17685   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 1);
17686   return ret;
17687 }
17688
17689 __extension__ extern __inline int32x4x2_t
17690 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17691 vld2q_s32 (const int32_t * __a)
17692 {
17693   int32x4x2_t ret;
17694   __builtin_aarch64_simd_oi __o;
17695   __o = __builtin_aarch64_ld2v4si ((const __builtin_aarch64_simd_si *) __a);
17696   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
17697   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
17698   return ret;
17699 }
17700
17701 __extension__ extern __inline int64x2x2_t
17702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17703 vld2q_s64 (const int64_t * __a)
17704 {
17705   int64x2x2_t ret;
17706   __builtin_aarch64_simd_oi __o;
17707   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
17708   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17709   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17710   return ret;
17711 }
17712
17713 __extension__ extern __inline uint8x16x2_t
17714 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17715 vld2q_u8 (const uint8_t * __a)
17716 {
17717   uint8x16x2_t ret;
17718   __builtin_aarch64_simd_oi __o;
17719   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17720   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17721   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17722   return ret;
17723 }
17724
17725 __extension__ extern __inline uint16x8x2_t
17726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17727 vld2q_u16 (const uint16_t * __a)
17728 {
17729   uint16x8x2_t ret;
17730   __builtin_aarch64_simd_oi __o;
17731   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17732   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17733   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17734   return ret;
17735 }
17736
17737 __extension__ extern __inline uint32x4x2_t
17738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17739 vld2q_u32 (const uint32_t * __a)
17740 {
17741   uint32x4x2_t ret;
17742   __builtin_aarch64_simd_oi __o;
17743   __o = __builtin_aarch64_ld2v4si ((const __builtin_aarch64_simd_si *) __a);
17744   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
17745   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
17746   return ret;
17747 }
17748
17749 __extension__ extern __inline uint64x2x2_t
17750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17751 vld2q_u64 (const uint64_t * __a)
17752 {
17753   uint64x2x2_t ret;
17754   __builtin_aarch64_simd_oi __o;
17755   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
17756   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17757   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17758   return ret;
17759 }
17760
17761 __extension__ extern __inline float16x8x2_t
17762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17763 vld2q_f16 (const float16_t * __a)
17764 {
17765   float16x8x2_t ret;
17766   __builtin_aarch64_simd_oi __o;
17767   __o = __builtin_aarch64_ld2v8hf (__a);
17768   ret.val[0] = __builtin_aarch64_get_qregoiv8hf (__o, 0);
17769   ret.val[1] = __builtin_aarch64_get_qregoiv8hf (__o, 1);
17770   return ret;
17771 }
17772
17773 __extension__ extern __inline float32x4x2_t
17774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17775 vld2q_f32 (const float32_t * __a)
17776 {
17777   float32x4x2_t ret;
17778   __builtin_aarch64_simd_oi __o;
17779   __o = __builtin_aarch64_ld2v4sf ((const __builtin_aarch64_simd_sf *) __a);
17780   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
17781   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
17782   return ret;
17783 }
17784
17785 __extension__ extern __inline float64x2x2_t
17786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17787 vld2q_f64 (const float64_t * __a)
17788 {
17789   float64x2x2_t ret;
17790   __builtin_aarch64_simd_oi __o;
17791   __o = __builtin_aarch64_ld2v2df ((const __builtin_aarch64_simd_df *) __a);
17792   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
17793   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
17794   return ret;
17795 }
17796
17797 __extension__ extern __inline int64x1x3_t
17798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17799 vld3_s64 (const int64_t * __a)
17800 {
17801   int64x1x3_t ret;
17802   __builtin_aarch64_simd_ci __o;
17803   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
17804   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17805   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17806   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17807   return ret;
17808 }
17809
17810 __extension__ extern __inline uint64x1x3_t
17811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17812 vld3_u64 (const uint64_t * __a)
17813 {
17814   uint64x1x3_t ret;
17815   __builtin_aarch64_simd_ci __o;
17816   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
17817   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17818   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17819   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17820   return ret;
17821 }
17822
17823 __extension__ extern __inline float64x1x3_t
17824 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17825 vld3_f64 (const float64_t * __a)
17826 {
17827   float64x1x3_t ret;
17828   __builtin_aarch64_simd_ci __o;
17829   __o = __builtin_aarch64_ld3df ((const __builtin_aarch64_simd_df *) __a);
17830   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 0)};
17831   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 1)};
17832   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 2)};
17833   return ret;
17834 }
17835
17836 __extension__ extern __inline int8x8x3_t
17837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17838 vld3_s8 (const int8_t * __a)
17839 {
17840   int8x8x3_t ret;
17841   __builtin_aarch64_simd_ci __o;
17842   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17843   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
17844   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
17845   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
17846   return ret;
17847 }
17848
17849 __extension__ extern __inline poly8x8x3_t
17850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17851 vld3_p8 (const poly8_t * __a)
17852 {
17853   poly8x8x3_t ret;
17854   __builtin_aarch64_simd_ci __o;
17855   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17856   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
17857   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
17858   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
17859   return ret;
17860 }
17861
17862 __extension__ extern __inline int16x4x3_t
17863 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17864 vld3_s16 (const int16_t * __a)
17865 {
17866   int16x4x3_t ret;
17867   __builtin_aarch64_simd_ci __o;
17868   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17869   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
17870   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
17871   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
17872   return ret;
17873 }
17874
17875 __extension__ extern __inline poly16x4x3_t
17876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17877 vld3_p16 (const poly16_t * __a)
17878 {
17879   poly16x4x3_t ret;
17880   __builtin_aarch64_simd_ci __o;
17881   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17882   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
17883   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
17884   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
17885   return ret;
17886 }
17887
17888 __extension__ extern __inline int32x2x3_t
17889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17890 vld3_s32 (const int32_t * __a)
17891 {
17892   int32x2x3_t ret;
17893   __builtin_aarch64_simd_ci __o;
17894   __o = __builtin_aarch64_ld3v2si ((const __builtin_aarch64_simd_si *) __a);
17895   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
17896   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
17897   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
17898   return ret;
17899 }
17900
17901 __extension__ extern __inline uint8x8x3_t
17902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17903 vld3_u8 (const uint8_t * __a)
17904 {
17905   uint8x8x3_t ret;
17906   __builtin_aarch64_simd_ci __o;
17907   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17908   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
17909   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
17910   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
17911   return ret;
17912 }
17913
17914 __extension__ extern __inline uint16x4x3_t
17915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17916 vld3_u16 (const uint16_t * __a)
17917 {
17918   uint16x4x3_t ret;
17919   __builtin_aarch64_simd_ci __o;
17920   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17921   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
17922   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
17923   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
17924   return ret;
17925 }
17926
17927 __extension__ extern __inline uint32x2x3_t
17928 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17929 vld3_u32 (const uint32_t * __a)
17930 {
17931   uint32x2x3_t ret;
17932   __builtin_aarch64_simd_ci __o;
17933   __o = __builtin_aarch64_ld3v2si ((const __builtin_aarch64_simd_si *) __a);
17934   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
17935   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
17936   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
17937   return ret;
17938 }
17939
17940 __extension__ extern __inline float16x4x3_t
17941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17942 vld3_f16 (const float16_t * __a)
17943 {
17944   float16x4x3_t ret;
17945   __builtin_aarch64_simd_ci __o;
17946   __o = __builtin_aarch64_ld3v4hf (__a);
17947   ret.val[0] = __builtin_aarch64_get_dregciv4hf (__o, 0);
17948   ret.val[1] = __builtin_aarch64_get_dregciv4hf (__o, 1);
17949   ret.val[2] = __builtin_aarch64_get_dregciv4hf (__o, 2);
17950   return ret;
17951 }
17952
17953 __extension__ extern __inline float32x2x3_t
17954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17955 vld3_f32 (const float32_t * __a)
17956 {
17957   float32x2x3_t ret;
17958   __builtin_aarch64_simd_ci __o;
17959   __o = __builtin_aarch64_ld3v2sf ((const __builtin_aarch64_simd_sf *) __a);
17960   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 0);
17961   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 1);
17962   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 2);
17963   return ret;
17964 }
17965
17966 __extension__ extern __inline poly64x1x3_t
17967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17968 vld3_p64 (const poly64_t * __a)
17969 {
17970   poly64x1x3_t ret;
17971   __builtin_aarch64_simd_ci __o;
17972   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
17973   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 0);
17974   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 1);
17975   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 2);
17976   return ret;
17977 }
17978
17979 __extension__ extern __inline int8x16x3_t
17980 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17981 vld3q_s8 (const int8_t * __a)
17982 {
17983   int8x16x3_t ret;
17984   __builtin_aarch64_simd_ci __o;
17985   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17986   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
17987   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
17988   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
17989   return ret;
17990 }
17991
17992 __extension__ extern __inline poly8x16x3_t
17993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17994 vld3q_p8 (const poly8_t * __a)
17995 {
17996   poly8x16x3_t ret;
17997   __builtin_aarch64_simd_ci __o;
17998   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17999   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
18000   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
18001   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
18002   return ret;
18003 }
18004
18005 __extension__ extern __inline int16x8x3_t
18006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18007 vld3q_s16 (const int16_t * __a)
18008 {
18009   int16x8x3_t ret;
18010   __builtin_aarch64_simd_ci __o;
18011   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18012   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18013   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18014   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18015   return ret;
18016 }
18017
18018 __extension__ extern __inline poly16x8x3_t
18019 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18020 vld3q_p16 (const poly16_t * __a)
18021 {
18022   poly16x8x3_t ret;
18023   __builtin_aarch64_simd_ci __o;
18024   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18025   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18026   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18027   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18028   return ret;
18029 }
18030
18031 __extension__ extern __inline int32x4x3_t
18032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18033 vld3q_s32 (const int32_t * __a)
18034 {
18035   int32x4x3_t ret;
18036   __builtin_aarch64_simd_ci __o;
18037   __o = __builtin_aarch64_ld3v4si ((const __builtin_aarch64_simd_si *) __a);
18038   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
18039   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
18040   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
18041   return ret;
18042 }
18043
18044 __extension__ extern __inline int64x2x3_t
18045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18046 vld3q_s64 (const int64_t * __a)
18047 {
18048   int64x2x3_t ret;
18049   __builtin_aarch64_simd_ci __o;
18050   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
18051   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
18052   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
18053   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
18054   return ret;
18055 }
18056
18057 __extension__ extern __inline uint8x16x3_t
18058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18059 vld3q_u8 (const uint8_t * __a)
18060 {
18061   uint8x16x3_t ret;
18062   __builtin_aarch64_simd_ci __o;
18063   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
18064   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
18065   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
18066   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
18067   return ret;
18068 }
18069
18070 __extension__ extern __inline uint16x8x3_t
18071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18072 vld3q_u16 (const uint16_t * __a)
18073 {
18074   uint16x8x3_t ret;
18075   __builtin_aarch64_simd_ci __o;
18076   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18077   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18078   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18079   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18080   return ret;
18081 }
18082
18083 __extension__ extern __inline uint32x4x3_t
18084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18085 vld3q_u32 (const uint32_t * __a)
18086 {
18087   uint32x4x3_t ret;
18088   __builtin_aarch64_simd_ci __o;
18089   __o = __builtin_aarch64_ld3v4si ((const __builtin_aarch64_simd_si *) __a);
18090   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
18091   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
18092   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
18093   return ret;
18094 }
18095
18096 __extension__ extern __inline uint64x2x3_t
18097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18098 vld3q_u64 (const uint64_t * __a)
18099 {
18100   uint64x2x3_t ret;
18101   __builtin_aarch64_simd_ci __o;
18102   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
18103   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
18104   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
18105   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
18106   return ret;
18107 }
18108
18109 __extension__ extern __inline float16x8x3_t
18110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18111 vld3q_f16 (const float16_t * __a)
18112 {
18113   float16x8x3_t ret;
18114   __builtin_aarch64_simd_ci __o;
18115   __o = __builtin_aarch64_ld3v8hf (__a);
18116   ret.val[0] = __builtin_aarch64_get_qregciv8hf (__o, 0);
18117   ret.val[1] = __builtin_aarch64_get_qregciv8hf (__o, 1);
18118   ret.val[2] = __builtin_aarch64_get_qregciv8hf (__o, 2);
18119   return ret;
18120 }
18121
18122 __extension__ extern __inline float32x4x3_t
18123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18124 vld3q_f32 (const float32_t * __a)
18125 {
18126   float32x4x3_t ret;
18127   __builtin_aarch64_simd_ci __o;
18128   __o = __builtin_aarch64_ld3v4sf ((const __builtin_aarch64_simd_sf *) __a);
18129   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 0);
18130   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 1);
18131   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 2);
18132   return ret;
18133 }
18134
18135 __extension__ extern __inline float64x2x3_t
18136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18137 vld3q_f64 (const float64_t * __a)
18138 {
18139   float64x2x3_t ret;
18140   __builtin_aarch64_simd_ci __o;
18141   __o = __builtin_aarch64_ld3v2df ((const __builtin_aarch64_simd_df *) __a);
18142   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 0);
18143   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 1);
18144   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 2);
18145   return ret;
18146 }
18147
18148 __extension__ extern __inline poly64x2x3_t
18149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18150 vld3q_p64 (const poly64_t * __a)
18151 {
18152   poly64x2x3_t ret;
18153   __builtin_aarch64_simd_ci __o;
18154   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
18155   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 0);
18156   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 1);
18157   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 2);
18158   return ret;
18159 }
18160
18161 __extension__ extern __inline int64x1x4_t
18162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18163 vld4_s64 (const int64_t * __a)
18164 {
18165   int64x1x4_t ret;
18166   __builtin_aarch64_simd_xi __o;
18167   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
18168   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
18169   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
18170   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
18171   ret.val[3] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
18172   return ret;
18173 }
18174
18175 __extension__ extern __inline uint64x1x4_t
18176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18177 vld4_u64 (const uint64_t * __a)
18178 {
18179   uint64x1x4_t ret;
18180   __builtin_aarch64_simd_xi __o;
18181   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
18182   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
18183   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
18184   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
18185   ret.val[3] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
18186   return ret;
18187 }
18188
18189 __extension__ extern __inline float64x1x4_t
18190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18191 vld4_f64 (const float64_t * __a)
18192 {
18193   float64x1x4_t ret;
18194   __builtin_aarch64_simd_xi __o;
18195   __o = __builtin_aarch64_ld4df ((const __builtin_aarch64_simd_df *) __a);
18196   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 0)};
18197   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 1)};
18198   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 2)};
18199   ret.val[3] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 3)};
18200   return ret;
18201 }
18202
18203 __extension__ extern __inline int8x8x4_t
18204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18205 vld4_s8 (const int8_t * __a)
18206 {
18207   int8x8x4_t ret;
18208   __builtin_aarch64_simd_xi __o;
18209   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
18210   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
18211   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
18212   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
18213   ret.val[3] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
18214   return ret;
18215 }
18216
18217 __extension__ extern __inline poly8x8x4_t
18218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18219 vld4_p8 (const poly8_t * __a)
18220 {
18221   poly8x8x4_t ret;
18222   __builtin_aarch64_simd_xi __o;
18223   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
18224   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
18225   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
18226   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
18227   ret.val[3] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
18228   return ret;
18229 }
18230
18231 __extension__ extern __inline int16x4x4_t
18232 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18233 vld4_s16 (const int16_t * __a)
18234 {
18235   int16x4x4_t ret;
18236   __builtin_aarch64_simd_xi __o;
18237   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
18238   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
18239   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
18240   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
18241   ret.val[3] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
18242   return ret;
18243 }
18244
18245 __extension__ extern __inline poly16x4x4_t
18246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18247 vld4_p16 (const poly16_t * __a)
18248 {
18249   poly16x4x4_t ret;
18250   __builtin_aarch64_simd_xi __o;
18251   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
18252   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
18253   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
18254   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
18255   ret.val[3] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
18256   return ret;
18257 }
18258
18259 __extension__ extern __inline int32x2x4_t
18260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18261 vld4_s32 (const int32_t * __a)
18262 {
18263   int32x2x4_t ret;
18264   __builtin_aarch64_simd_xi __o;
18265   __o = __builtin_aarch64_ld4v2si ((const __builtin_aarch64_simd_si *) __a);
18266   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
18267   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
18268   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
18269   ret.val[3] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
18270   return ret;
18271 }
18272
18273 __extension__ extern __inline uint8x8x4_t
18274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18275 vld4_u8 (const uint8_t * __a)
18276 {
18277   uint8x8x4_t ret;
18278   __builtin_aarch64_simd_xi __o;
18279   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
18280   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
18281   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
18282   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
18283   ret.val[3] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
18284   return ret;
18285 }
18286
18287 __extension__ extern __inline uint16x4x4_t
18288 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18289 vld4_u16 (const uint16_t * __a)
18290 {
18291   uint16x4x4_t ret;
18292   __builtin_aarch64_simd_xi __o;
18293   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
18294   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
18295   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
18296   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
18297   ret.val[3] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
18298   return ret;
18299 }
18300
18301 __extension__ extern __inline uint32x2x4_t
18302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18303 vld4_u32 (const uint32_t * __a)
18304 {
18305   uint32x2x4_t ret;
18306   __builtin_aarch64_simd_xi __o;
18307   __o = __builtin_aarch64_ld4v2si ((const __builtin_aarch64_simd_si *) __a);
18308   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
18309   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
18310   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
18311   ret.val[3] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
18312   return ret;
18313 }
18314
18315 __extension__ extern __inline float16x4x4_t
18316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18317 vld4_f16 (const float16_t * __a)
18318 {
18319   float16x4x4_t ret;
18320   __builtin_aarch64_simd_xi __o;
18321   __o = __builtin_aarch64_ld4v4hf (__a);
18322   ret.val[0] = __builtin_aarch64_get_dregxiv4hf (__o, 0);
18323   ret.val[1] = __builtin_aarch64_get_dregxiv4hf (__o, 1);
18324   ret.val[2] = __builtin_aarch64_get_dregxiv4hf (__o, 2);
18325   ret.val[3] = __builtin_aarch64_get_dregxiv4hf (__o, 3);
18326   return ret;
18327 }
18328
18329 __extension__ extern __inline float32x2x4_t
18330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18331 vld4_f32 (const float32_t * __a)
18332 {
18333   float32x2x4_t ret;
18334   __builtin_aarch64_simd_xi __o;
18335   __o = __builtin_aarch64_ld4v2sf ((const __builtin_aarch64_simd_sf *) __a);
18336   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 0);
18337   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 1);
18338   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 2);
18339   ret.val[3] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 3);
18340   return ret;
18341 }
18342
18343 __extension__ extern __inline poly64x1x4_t
18344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18345 vld4_p64 (const poly64_t * __a)
18346 {
18347   poly64x1x4_t  ret;
18348   __builtin_aarch64_simd_xi __o;
18349   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
18350   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 0);
18351   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 1);
18352   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 2);
18353   ret.val[3] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 3);
18354   return ret;
18355 }
18356
18357 __extension__ extern __inline int8x16x4_t
18358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18359 vld4q_s8 (const int8_t * __a)
18360 {
18361   int8x16x4_t ret;
18362   __builtin_aarch64_simd_xi __o;
18363   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
18364   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
18365   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
18366   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
18367   ret.val[3] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
18368   return ret;
18369 }
18370
18371 __extension__ extern __inline poly8x16x4_t
18372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18373 vld4q_p8 (const poly8_t * __a)
18374 {
18375   poly8x16x4_t ret;
18376   __builtin_aarch64_simd_xi __o;
18377   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
18378   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
18379   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
18380   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
18381   ret.val[3] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
18382   return ret;
18383 }
18384
18385 __extension__ extern __inline int16x8x4_t
18386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18387 vld4q_s16 (const int16_t * __a)
18388 {
18389   int16x8x4_t ret;
18390   __builtin_aarch64_simd_xi __o;
18391   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
18392   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
18393   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
18394   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
18395   ret.val[3] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
18396   return ret;
18397 }
18398
18399 __extension__ extern __inline poly16x8x4_t
18400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18401 vld4q_p16 (const poly16_t * __a)
18402 {
18403   poly16x8x4_t ret;
18404   __builtin_aarch64_simd_xi __o;
18405   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
18406   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
18407   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
18408   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
18409   ret.val[3] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
18410   return ret;
18411 }
18412
18413 __extension__ extern __inline int32x4x4_t
18414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18415 vld4q_s32 (const int32_t * __a)
18416 {
18417   int32x4x4_t ret;
18418   __builtin_aarch64_simd_xi __o;
18419   __o = __builtin_aarch64_ld4v4si ((const __builtin_aarch64_simd_si *) __a);
18420   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
18421   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
18422   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
18423   ret.val[3] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
18424   return ret;
18425 }
18426
18427 __extension__ extern __inline int64x2x4_t
18428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18429 vld4q_s64 (const int64_t * __a)
18430 {
18431   int64x2x4_t ret;
18432   __builtin_aarch64_simd_xi __o;
18433   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
18434   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
18435   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
18436   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
18437   ret.val[3] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
18438   return ret;
18439 }
18440
18441 __extension__ extern __inline uint8x16x4_t
18442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18443 vld4q_u8 (const uint8_t * __a)
18444 {
18445   uint8x16x4_t ret;
18446   __builtin_aarch64_simd_xi __o;
18447   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
18448   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
18449   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
18450   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
18451   ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
18452   return ret;
18453 }
18454
18455 __extension__ extern __inline uint16x8x4_t
18456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18457 vld4q_u16 (const uint16_t * __a)
18458 {
18459   uint16x8x4_t ret;
18460   __builtin_aarch64_simd_xi __o;
18461   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
18462   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
18463   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
18464   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
18465   ret.val[3] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
18466   return ret;
18467 }
18468
18469 __extension__ extern __inline uint32x4x4_t
18470 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18471 vld4q_u32 (const uint32_t * __a)
18472 {
18473   uint32x4x4_t ret;
18474   __builtin_aarch64_simd_xi __o;
18475   __o = __builtin_aarch64_ld4v4si ((const __builtin_aarch64_simd_si *) __a);
18476   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
18477   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
18478   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
18479   ret.val[3] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
18480   return ret;
18481 }
18482
18483 __extension__ extern __inline uint64x2x4_t
18484 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18485 vld4q_u64 (const uint64_t * __a)
18486 {
18487   uint64x2x4_t ret;
18488   __builtin_aarch64_simd_xi __o;
18489   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
18490   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
18491   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
18492   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
18493   ret.val[3] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
18494   return ret;
18495 }
18496
18497 __extension__ extern __inline float16x8x4_t
18498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18499 vld4q_f16 (const float16_t * __a)
18500 {
18501   float16x8x4_t ret;
18502   __builtin_aarch64_simd_xi __o;
18503   __o = __builtin_aarch64_ld4v8hf (__a);
18504   ret.val[0] = __builtin_aarch64_get_qregxiv8hf (__o, 0);
18505   ret.val[1] = __builtin_aarch64_get_qregxiv8hf (__o, 1);
18506   ret.val[2] = __builtin_aarch64_get_qregxiv8hf (__o, 2);
18507   ret.val[3] = __builtin_aarch64_get_qregxiv8hf (__o, 3);
18508   return ret;
18509 }
18510
18511 __extension__ extern __inline float32x4x4_t
18512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18513 vld4q_f32 (const float32_t * __a)
18514 {
18515   float32x4x4_t ret;
18516   __builtin_aarch64_simd_xi __o;
18517   __o = __builtin_aarch64_ld4v4sf ((const __builtin_aarch64_simd_sf *) __a);
18518   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 0);
18519   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 1);
18520   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 2);
18521   ret.val[3] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 3);
18522   return ret;
18523 }
18524
18525 __extension__ extern __inline float64x2x4_t
18526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18527 vld4q_f64 (const float64_t * __a)
18528 {
18529   float64x2x4_t ret;
18530   __builtin_aarch64_simd_xi __o;
18531   __o = __builtin_aarch64_ld4v2df ((const __builtin_aarch64_simd_df *) __a);
18532   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 0);
18533   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 1);
18534   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 2);
18535   ret.val[3] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 3);
18536   return ret;
18537 }
18538
18539 __extension__ extern __inline poly64x2x4_t
18540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18541 vld4q_p64 (const poly64_t * __a)
18542 {
18543   poly64x2x4_t  ret;
18544   __builtin_aarch64_simd_xi __o;
18545   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
18546   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 0);
18547   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 1);
18548   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 2);
18549   ret.val[3] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 3);
18550   return ret;
18551 }
18552
18553 /* vldn_dup */
18554
18555 __extension__ extern __inline int8x8x2_t
18556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18557 vld2_dup_s8 (const int8_t * __a)
18558 {
18559   int8x8x2_t ret;
18560   __builtin_aarch64_simd_oi __o;
18561   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18562   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18563   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18564   return ret;
18565 }
18566
18567 __extension__ extern __inline int16x4x2_t
18568 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18569 vld2_dup_s16 (const int16_t * __a)
18570 {
18571   int16x4x2_t ret;
18572   __builtin_aarch64_simd_oi __o;
18573   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18574   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18575   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18576   return ret;
18577 }
18578
18579 __extension__ extern __inline int32x2x2_t
18580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18581 vld2_dup_s32 (const int32_t * __a)
18582 {
18583   int32x2x2_t ret;
18584   __builtin_aarch64_simd_oi __o;
18585   __o = __builtin_aarch64_ld2rv2si ((const __builtin_aarch64_simd_si *) __a);
18586   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
18587   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
18588   return ret;
18589 }
18590
18591 __extension__ extern __inline float16x4x2_t
18592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18593 vld2_dup_f16 (const float16_t * __a)
18594 {
18595   float16x4x2_t ret;
18596   __builtin_aarch64_simd_oi __o;
18597   __o = __builtin_aarch64_ld2rv4hf ((const __builtin_aarch64_simd_hf *) __a);
18598   ret.val[0] = __builtin_aarch64_get_dregoiv4hf (__o, 0);
18599   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregoiv4hf (__o, 1);
18600   return ret;
18601 }
18602
18603 __extension__ extern __inline float32x2x2_t
18604 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18605 vld2_dup_f32 (const float32_t * __a)
18606 {
18607   float32x2x2_t ret;
18608   __builtin_aarch64_simd_oi __o;
18609   __o = __builtin_aarch64_ld2rv2sf ((const __builtin_aarch64_simd_sf *) __a);
18610   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
18611   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
18612   return ret;
18613 }
18614
18615 __extension__ extern __inline float64x1x2_t
18616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18617 vld2_dup_f64 (const float64_t * __a)
18618 {
18619   float64x1x2_t ret;
18620   __builtin_aarch64_simd_oi __o;
18621   __o = __builtin_aarch64_ld2rdf ((const __builtin_aarch64_simd_df *) __a);
18622   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
18623   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
18624   return ret;
18625 }
18626
18627 __extension__ extern __inline uint8x8x2_t
18628 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18629 vld2_dup_u8 (const uint8_t * __a)
18630 {
18631   uint8x8x2_t ret;
18632   __builtin_aarch64_simd_oi __o;
18633   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18634   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18635   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18636   return ret;
18637 }
18638
18639 __extension__ extern __inline uint16x4x2_t
18640 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18641 vld2_dup_u16 (const uint16_t * __a)
18642 {
18643   uint16x4x2_t ret;
18644   __builtin_aarch64_simd_oi __o;
18645   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18646   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18647   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18648   return ret;
18649 }
18650
18651 __extension__ extern __inline uint32x2x2_t
18652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18653 vld2_dup_u32 (const uint32_t * __a)
18654 {
18655   uint32x2x2_t ret;
18656   __builtin_aarch64_simd_oi __o;
18657   __o = __builtin_aarch64_ld2rv2si ((const __builtin_aarch64_simd_si *) __a);
18658   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
18659   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
18660   return ret;
18661 }
18662
18663 __extension__ extern __inline poly8x8x2_t
18664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18665 vld2_dup_p8 (const poly8_t * __a)
18666 {
18667   poly8x8x2_t ret;
18668   __builtin_aarch64_simd_oi __o;
18669   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18670   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18671   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18672   return ret;
18673 }
18674
18675 __extension__ extern __inline poly16x4x2_t
18676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18677 vld2_dup_p16 (const poly16_t * __a)
18678 {
18679   poly16x4x2_t ret;
18680   __builtin_aarch64_simd_oi __o;
18681   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18682   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18683   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18684   return ret;
18685 }
18686
18687 __extension__ extern __inline poly64x1x2_t
18688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18689 vld2_dup_p64 (const poly64_t * __a)
18690 {
18691   poly64x1x2_t ret;
18692   __builtin_aarch64_simd_oi __o;
18693   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18694   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 0);
18695   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 1);
18696   return ret;
18697 }
18698
18699
18700 __extension__ extern __inline int64x1x2_t
18701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18702 vld2_dup_s64 (const int64_t * __a)
18703 {
18704   int64x1x2_t ret;
18705   __builtin_aarch64_simd_oi __o;
18706   __o = __builtin_aarch64_ld2rdi ((const __builtin_aarch64_simd_di *) __a);
18707   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
18708   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
18709   return ret;
18710 }
18711
18712 __extension__ extern __inline uint64x1x2_t
18713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18714 vld2_dup_u64 (const uint64_t * __a)
18715 {
18716   uint64x1x2_t ret;
18717   __builtin_aarch64_simd_oi __o;
18718   __o = __builtin_aarch64_ld2rdi ((const __builtin_aarch64_simd_di *) __a);
18719   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
18720   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
18721   return ret;
18722 }
18723
18724 __extension__ extern __inline int8x16x2_t
18725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18726 vld2q_dup_s8 (const int8_t * __a)
18727 {
18728   int8x16x2_t ret;
18729   __builtin_aarch64_simd_oi __o;
18730   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
18731   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18732   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18733   return ret;
18734 }
18735
18736 __extension__ extern __inline poly8x16x2_t
18737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18738 vld2q_dup_p8 (const poly8_t * __a)
18739 {
18740   poly8x16x2_t ret;
18741   __builtin_aarch64_simd_oi __o;
18742   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
18743   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18744   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18745   return ret;
18746 }
18747
18748 __extension__ extern __inline int16x8x2_t
18749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18750 vld2q_dup_s16 (const int16_t * __a)
18751 {
18752   int16x8x2_t ret;
18753   __builtin_aarch64_simd_oi __o;
18754   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
18755   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18756   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18757   return ret;
18758 }
18759
18760 __extension__ extern __inline poly16x8x2_t
18761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18762 vld2q_dup_p16 (const poly16_t * __a)
18763 {
18764   poly16x8x2_t ret;
18765   __builtin_aarch64_simd_oi __o;
18766   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
18767   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18768   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18769   return ret;
18770 }
18771
18772 __extension__ extern __inline int32x4x2_t
18773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18774 vld2q_dup_s32 (const int32_t * __a)
18775 {
18776   int32x4x2_t ret;
18777   __builtin_aarch64_simd_oi __o;
18778   __o = __builtin_aarch64_ld2rv4si ((const __builtin_aarch64_simd_si *) __a);
18779   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
18780   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
18781   return ret;
18782 }
18783
18784 __extension__ extern __inline int64x2x2_t
18785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18786 vld2q_dup_s64 (const int64_t * __a)
18787 {
18788   int64x2x2_t ret;
18789   __builtin_aarch64_simd_oi __o;
18790   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18791   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
18792   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
18793   return ret;
18794 }
18795
18796 __extension__ extern __inline uint8x16x2_t
18797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18798 vld2q_dup_u8 (const uint8_t * __a)
18799 {
18800   uint8x16x2_t ret;
18801   __builtin_aarch64_simd_oi __o;
18802   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
18803   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18804   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18805   return ret;
18806 }
18807
18808 __extension__ extern __inline uint16x8x2_t
18809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18810 vld2q_dup_u16 (const uint16_t * __a)
18811 {
18812   uint16x8x2_t ret;
18813   __builtin_aarch64_simd_oi __o;
18814   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
18815   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18816   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18817   return ret;
18818 }
18819
18820 __extension__ extern __inline uint32x4x2_t
18821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18822 vld2q_dup_u32 (const uint32_t * __a)
18823 {
18824   uint32x4x2_t ret;
18825   __builtin_aarch64_simd_oi __o;
18826   __o = __builtin_aarch64_ld2rv4si ((const __builtin_aarch64_simd_si *) __a);
18827   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
18828   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
18829   return ret;
18830 }
18831
18832 __extension__ extern __inline uint64x2x2_t
18833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18834 vld2q_dup_u64 (const uint64_t * __a)
18835 {
18836   uint64x2x2_t ret;
18837   __builtin_aarch64_simd_oi __o;
18838   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18839   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
18840   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
18841   return ret;
18842 }
18843
18844 __extension__ extern __inline float16x8x2_t
18845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18846 vld2q_dup_f16 (const float16_t * __a)
18847 {
18848   float16x8x2_t ret;
18849   __builtin_aarch64_simd_oi __o;
18850   __o = __builtin_aarch64_ld2rv8hf ((const __builtin_aarch64_simd_hf *) __a);
18851   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregoiv8hf (__o, 0);
18852   ret.val[1] = __builtin_aarch64_get_qregoiv8hf (__o, 1);
18853   return ret;
18854 }
18855
18856 __extension__ extern __inline float32x4x2_t
18857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18858 vld2q_dup_f32 (const float32_t * __a)
18859 {
18860   float32x4x2_t ret;
18861   __builtin_aarch64_simd_oi __o;
18862   __o = __builtin_aarch64_ld2rv4sf ((const __builtin_aarch64_simd_sf *) __a);
18863   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
18864   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
18865   return ret;
18866 }
18867
18868 __extension__ extern __inline float64x2x2_t
18869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18870 vld2q_dup_f64 (const float64_t * __a)
18871 {
18872   float64x2x2_t ret;
18873   __builtin_aarch64_simd_oi __o;
18874   __o = __builtin_aarch64_ld2rv2df ((const __builtin_aarch64_simd_df *) __a);
18875   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
18876   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
18877   return ret;
18878 }
18879
18880 __extension__ extern __inline poly64x2x2_t
18881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18882 vld2q_dup_p64 (const poly64_t * __a)
18883 {
18884   poly64x2x2_t ret;
18885   __builtin_aarch64_simd_oi __o;
18886   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18887   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 0);
18888   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 1);
18889   return ret;
18890 }
18891
18892 __extension__ extern __inline int64x1x3_t
18893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18894 vld3_dup_s64 (const int64_t * __a)
18895 {
18896   int64x1x3_t ret;
18897   __builtin_aarch64_simd_ci __o;
18898   __o = __builtin_aarch64_ld3rdi ((const __builtin_aarch64_simd_di *) __a);
18899   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
18900   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
18901   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
18902   return ret;
18903 }
18904
18905 __extension__ extern __inline uint64x1x3_t
18906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18907 vld3_dup_u64 (const uint64_t * __a)
18908 {
18909   uint64x1x3_t ret;
18910   __builtin_aarch64_simd_ci __o;
18911   __o = __builtin_aarch64_ld3rdi ((const __builtin_aarch64_simd_di *) __a);
18912   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
18913   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
18914   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
18915   return ret;
18916 }
18917
18918 __extension__ extern __inline float64x1x3_t
18919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18920 vld3_dup_f64 (const float64_t * __a)
18921 {
18922   float64x1x3_t ret;
18923   __builtin_aarch64_simd_ci __o;
18924   __o = __builtin_aarch64_ld3rdf ((const __builtin_aarch64_simd_df *) __a);
18925   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 0)};
18926   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 1)};
18927   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 2)};
18928   return ret;
18929 }
18930
18931 __extension__ extern __inline int8x8x3_t
18932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18933 vld3_dup_s8 (const int8_t * __a)
18934 {
18935   int8x8x3_t ret;
18936   __builtin_aarch64_simd_ci __o;
18937   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18938   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18939   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18940   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18941   return ret;
18942 }
18943
18944 __extension__ extern __inline poly8x8x3_t
18945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18946 vld3_dup_p8 (const poly8_t * __a)
18947 {
18948   poly8x8x3_t ret;
18949   __builtin_aarch64_simd_ci __o;
18950   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18951   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18952   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18953   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18954   return ret;
18955 }
18956
18957 __extension__ extern __inline int16x4x3_t
18958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18959 vld3_dup_s16 (const int16_t * __a)
18960 {
18961   int16x4x3_t ret;
18962   __builtin_aarch64_simd_ci __o;
18963   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18964   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18965   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18966   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18967   return ret;
18968 }
18969
18970 __extension__ extern __inline poly16x4x3_t
18971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18972 vld3_dup_p16 (const poly16_t * __a)
18973 {
18974   poly16x4x3_t ret;
18975   __builtin_aarch64_simd_ci __o;
18976   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18977   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18978   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18979   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18980   return ret;
18981 }
18982
18983 __extension__ extern __inline int32x2x3_t
18984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18985 vld3_dup_s32 (const int32_t * __a)
18986 {
18987   int32x2x3_t ret;
18988   __builtin_aarch64_simd_ci __o;
18989   __o = __builtin_aarch64_ld3rv2si ((const __builtin_aarch64_simd_si *) __a);
18990   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
18991   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
18992   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
18993   return ret;
18994 }
18995
18996 __extension__ extern __inline uint8x8x3_t
18997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18998 vld3_dup_u8 (const uint8_t * __a)
18999 {
19000   uint8x8x3_t ret;
19001   __builtin_aarch64_simd_ci __o;
19002   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19003   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
19004   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
19005   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
19006   return ret;
19007 }
19008
19009 __extension__ extern __inline uint16x4x3_t
19010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19011 vld3_dup_u16 (const uint16_t * __a)
19012 {
19013   uint16x4x3_t ret;
19014   __builtin_aarch64_simd_ci __o;
19015   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19016   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
19017   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
19018   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
19019   return ret;
19020 }
19021
19022 __extension__ extern __inline uint32x2x3_t
19023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19024 vld3_dup_u32 (const uint32_t * __a)
19025 {
19026   uint32x2x3_t ret;
19027   __builtin_aarch64_simd_ci __o;
19028   __o = __builtin_aarch64_ld3rv2si ((const __builtin_aarch64_simd_si *) __a);
19029   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
19030   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
19031   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
19032   return ret;
19033 }
19034
19035 __extension__ extern __inline float16x4x3_t
19036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19037 vld3_dup_f16 (const float16_t * __a)
19038 {
19039   float16x4x3_t ret;
19040   __builtin_aarch64_simd_ci __o;
19041   __o = __builtin_aarch64_ld3rv4hf ((const __builtin_aarch64_simd_hf *) __a);
19042   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 0);
19043   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 1);
19044   ret.val[2] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 2);
19045   return ret;
19046 }
19047
19048 __extension__ extern __inline float32x2x3_t
19049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19050 vld3_dup_f32 (const float32_t * __a)
19051 {
19052   float32x2x3_t ret;
19053   __builtin_aarch64_simd_ci __o;
19054   __o = __builtin_aarch64_ld3rv2sf ((const __builtin_aarch64_simd_sf *) __a);
19055   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 0);
19056   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 1);
19057   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 2);
19058   return ret;
19059 }
19060
19061 __extension__ extern __inline poly64x1x3_t
19062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19063 vld3_dup_p64 (const poly64_t * __a)
19064 {
19065   poly64x1x3_t ret;
19066   __builtin_aarch64_simd_ci __o;
19067   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19068   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 0);
19069   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 1);
19070   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 2);
19071   return ret;
19072 }
19073
19074 __extension__ extern __inline int8x16x3_t
19075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19076 vld3q_dup_s8 (const int8_t * __a)
19077 {
19078   int8x16x3_t ret;
19079   __builtin_aarch64_simd_ci __o;
19080   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19081   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19082   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19083   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19084   return ret;
19085 }
19086
19087 __extension__ extern __inline poly8x16x3_t
19088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19089 vld3q_dup_p8 (const poly8_t * __a)
19090 {
19091   poly8x16x3_t ret;
19092   __builtin_aarch64_simd_ci __o;
19093   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19094   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19095   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19096   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19097   return ret;
19098 }
19099
19100 __extension__ extern __inline int16x8x3_t
19101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19102 vld3q_dup_s16 (const int16_t * __a)
19103 {
19104   int16x8x3_t ret;
19105   __builtin_aarch64_simd_ci __o;
19106   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19107   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
19108   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
19109   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
19110   return ret;
19111 }
19112
19113 __extension__ extern __inline poly16x8x3_t
19114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19115 vld3q_dup_p16 (const poly16_t * __a)
19116 {
19117   poly16x8x3_t ret;
19118   __builtin_aarch64_simd_ci __o;
19119   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19120   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
19121   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
19122   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
19123   return ret;
19124 }
19125
19126 __extension__ extern __inline int32x4x3_t
19127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19128 vld3q_dup_s32 (const int32_t * __a)
19129 {
19130   int32x4x3_t ret;
19131   __builtin_aarch64_simd_ci __o;
19132   __o = __builtin_aarch64_ld3rv4si ((const __builtin_aarch64_simd_si *) __a);
19133   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
19134   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
19135   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
19136   return ret;
19137 }
19138
19139 __extension__ extern __inline int64x2x3_t
19140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19141 vld3q_dup_s64 (const int64_t * __a)
19142 {
19143   int64x2x3_t ret;
19144   __builtin_aarch64_simd_ci __o;
19145   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19146   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
19147   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
19148   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
19149   return ret;
19150 }
19151
19152 __extension__ extern __inline uint8x16x3_t
19153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19154 vld3q_dup_u8 (const uint8_t * __a)
19155 {
19156   uint8x16x3_t ret;
19157   __builtin_aarch64_simd_ci __o;
19158   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19159   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19160   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19161   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19162   return ret;
19163 }
19164
19165 __extension__ extern __inline uint16x8x3_t
19166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19167 vld3q_dup_u16 (const uint16_t * __a)
19168 {
19169   uint16x8x3_t ret;
19170   __builtin_aarch64_simd_ci __o;
19171   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19172   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
19173   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
19174   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
19175   return ret;
19176 }
19177
19178 __extension__ extern __inline uint32x4x3_t
19179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19180 vld3q_dup_u32 (const uint32_t * __a)
19181 {
19182   uint32x4x3_t ret;
19183   __builtin_aarch64_simd_ci __o;
19184   __o = __builtin_aarch64_ld3rv4si ((const __builtin_aarch64_simd_si *) __a);
19185   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
19186   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
19187   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
19188   return ret;
19189 }
19190
19191 __extension__ extern __inline uint64x2x3_t
19192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19193 vld3q_dup_u64 (const uint64_t * __a)
19194 {
19195   uint64x2x3_t ret;
19196   __builtin_aarch64_simd_ci __o;
19197   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19198   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
19199   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
19200   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
19201   return ret;
19202 }
19203
19204 __extension__ extern __inline float16x8x3_t
19205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19206 vld3q_dup_f16 (const float16_t * __a)
19207 {
19208   float16x8x3_t ret;
19209   __builtin_aarch64_simd_ci __o;
19210   __o = __builtin_aarch64_ld3rv8hf ((const __builtin_aarch64_simd_hf *) __a);
19211   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 0);
19212   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 1);
19213   ret.val[2] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 2);
19214   return ret;
19215 }
19216
19217 __extension__ extern __inline float32x4x3_t
19218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19219 vld3q_dup_f32 (const float32_t * __a)
19220 {
19221   float32x4x3_t ret;
19222   __builtin_aarch64_simd_ci __o;
19223   __o = __builtin_aarch64_ld3rv4sf ((const __builtin_aarch64_simd_sf *) __a);
19224   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 0);
19225   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 1);
19226   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 2);
19227   return ret;
19228 }
19229
19230 __extension__ extern __inline float64x2x3_t
19231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19232 vld3q_dup_f64 (const float64_t * __a)
19233 {
19234   float64x2x3_t ret;
19235   __builtin_aarch64_simd_ci __o;
19236   __o = __builtin_aarch64_ld3rv2df ((const __builtin_aarch64_simd_df *) __a);
19237   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 0);
19238   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 1);
19239   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 2);
19240   return ret;
19241 }
19242
19243 __extension__ extern __inline poly64x2x3_t
19244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19245 vld3q_dup_p64 (const poly64_t * __a)
19246 {
19247   poly64x2x3_t ret;
19248   __builtin_aarch64_simd_ci __o;
19249   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19250   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 0);
19251   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 1);
19252   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 2);
19253   return ret;
19254 }
19255
19256 __extension__ extern __inline int64x1x4_t
19257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19258 vld4_dup_s64 (const int64_t * __a)
19259 {
19260   int64x1x4_t ret;
19261   __builtin_aarch64_simd_xi __o;
19262   __o = __builtin_aarch64_ld4rdi ((const __builtin_aarch64_simd_di *) __a);
19263   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
19264   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
19265   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
19266   ret.val[3] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
19267   return ret;
19268 }
19269
19270 __extension__ extern __inline uint64x1x4_t
19271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19272 vld4_dup_u64 (const uint64_t * __a)
19273 {
19274   uint64x1x4_t ret;
19275   __builtin_aarch64_simd_xi __o;
19276   __o = __builtin_aarch64_ld4rdi ((const __builtin_aarch64_simd_di *) __a);
19277   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
19278   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
19279   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
19280   ret.val[3] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
19281   return ret;
19282 }
19283
19284 __extension__ extern __inline float64x1x4_t
19285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19286 vld4_dup_f64 (const float64_t * __a)
19287 {
19288   float64x1x4_t ret;
19289   __builtin_aarch64_simd_xi __o;
19290   __o = __builtin_aarch64_ld4rdf ((const __builtin_aarch64_simd_df *) __a);
19291   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 0)};
19292   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 1)};
19293   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 2)};
19294   ret.val[3] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 3)};
19295   return ret;
19296 }
19297
19298 __extension__ extern __inline int8x8x4_t
19299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19300 vld4_dup_s8 (const int8_t * __a)
19301 {
19302   int8x8x4_t ret;
19303   __builtin_aarch64_simd_xi __o;
19304   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19305   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19306   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19307   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19308   ret.val[3] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19309   return ret;
19310 }
19311
19312 __extension__ extern __inline poly8x8x4_t
19313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19314 vld4_dup_p8 (const poly8_t * __a)
19315 {
19316   poly8x8x4_t ret;
19317   __builtin_aarch64_simd_xi __o;
19318   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19319   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19320   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19321   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19322   ret.val[3] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19323   return ret;
19324 }
19325
19326 __extension__ extern __inline int16x4x4_t
19327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19328 vld4_dup_s16 (const int16_t * __a)
19329 {
19330   int16x4x4_t ret;
19331   __builtin_aarch64_simd_xi __o;
19332   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19333   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19334   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19335   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19336   ret.val[3] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19337   return ret;
19338 }
19339
19340 __extension__ extern __inline poly16x4x4_t
19341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19342 vld4_dup_p16 (const poly16_t * __a)
19343 {
19344   poly16x4x4_t ret;
19345   __builtin_aarch64_simd_xi __o;
19346   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19347   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19348   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19349   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19350   ret.val[3] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19351   return ret;
19352 }
19353
19354 __extension__ extern __inline int32x2x4_t
19355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19356 vld4_dup_s32 (const int32_t * __a)
19357 {
19358   int32x2x4_t ret;
19359   __builtin_aarch64_simd_xi __o;
19360   __o = __builtin_aarch64_ld4rv2si ((const __builtin_aarch64_simd_si *) __a);
19361   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
19362   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
19363   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
19364   ret.val[3] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
19365   return ret;
19366 }
19367
19368 __extension__ extern __inline uint8x8x4_t
19369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19370 vld4_dup_u8 (const uint8_t * __a)
19371 {
19372   uint8x8x4_t ret;
19373   __builtin_aarch64_simd_xi __o;
19374   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19375   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19376   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19377   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19378   ret.val[3] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19379   return ret;
19380 }
19381
19382 __extension__ extern __inline uint16x4x4_t
19383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19384 vld4_dup_u16 (const uint16_t * __a)
19385 {
19386   uint16x4x4_t ret;
19387   __builtin_aarch64_simd_xi __o;
19388   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19389   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19390   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19391   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19392   ret.val[3] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19393   return ret;
19394 }
19395
19396 __extension__ extern __inline uint32x2x4_t
19397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19398 vld4_dup_u32 (const uint32_t * __a)
19399 {
19400   uint32x2x4_t ret;
19401   __builtin_aarch64_simd_xi __o;
19402   __o = __builtin_aarch64_ld4rv2si ((const __builtin_aarch64_simd_si *) __a);
19403   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
19404   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
19405   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
19406   ret.val[3] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
19407   return ret;
19408 }
19409
19410 __extension__ extern __inline float16x4x4_t
19411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19412 vld4_dup_f16 (const float16_t * __a)
19413 {
19414   float16x4x4_t ret;
19415   __builtin_aarch64_simd_xi __o;
19416   __o = __builtin_aarch64_ld4rv4hf ((const __builtin_aarch64_simd_hf *) __a);
19417   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 0);
19418   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 1);
19419   ret.val[2] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 2);
19420   ret.val[3] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 3);
19421   return ret;
19422 }
19423
19424 __extension__ extern __inline float32x2x4_t
19425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19426 vld4_dup_f32 (const float32_t * __a)
19427 {
19428   float32x2x4_t ret;
19429   __builtin_aarch64_simd_xi __o;
19430   __o = __builtin_aarch64_ld4rv2sf ((const __builtin_aarch64_simd_sf *) __a);
19431   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 0);
19432   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 1);
19433   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 2);
19434   ret.val[3] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 3);
19435   return ret;
19436 }
19437
19438 __extension__ extern __inline poly64x1x4_t
19439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19440 vld4_dup_p64 (const poly64_t * __a)
19441 {
19442   poly64x1x4_t ret;
19443   __builtin_aarch64_simd_xi __o;
19444   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19445   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 0);
19446   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 1);
19447   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 2);
19448   ret.val[3] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 3);
19449   return ret;
19450 }
19451
19452 __extension__ extern __inline int8x16x4_t
19453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19454 vld4q_dup_s8 (const int8_t * __a)
19455 {
19456   int8x16x4_t ret;
19457   __builtin_aarch64_simd_xi __o;
19458   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19459   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19460   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19461   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19462   ret.val[3] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19463   return ret;
19464 }
19465
19466 __extension__ extern __inline poly8x16x4_t
19467 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19468 vld4q_dup_p8 (const poly8_t * __a)
19469 {
19470   poly8x16x4_t ret;
19471   __builtin_aarch64_simd_xi __o;
19472   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19473   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19474   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19475   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19476   ret.val[3] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19477   return ret;
19478 }
19479
19480 __extension__ extern __inline int16x8x4_t
19481 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19482 vld4q_dup_s16 (const int16_t * __a)
19483 {
19484   int16x8x4_t ret;
19485   __builtin_aarch64_simd_xi __o;
19486   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19487   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19488   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19489   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19490   ret.val[3] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19491   return ret;
19492 }
19493
19494 __extension__ extern __inline poly16x8x4_t
19495 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19496 vld4q_dup_p16 (const poly16_t * __a)
19497 {
19498   poly16x8x4_t ret;
19499   __builtin_aarch64_simd_xi __o;
19500   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19501   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19502   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19503   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19504   ret.val[3] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19505   return ret;
19506 }
19507
19508 __extension__ extern __inline int32x4x4_t
19509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19510 vld4q_dup_s32 (const int32_t * __a)
19511 {
19512   int32x4x4_t ret;
19513   __builtin_aarch64_simd_xi __o;
19514   __o = __builtin_aarch64_ld4rv4si ((const __builtin_aarch64_simd_si *) __a);
19515   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
19516   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
19517   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
19518   ret.val[3] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
19519   return ret;
19520 }
19521
19522 __extension__ extern __inline int64x2x4_t
19523 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19524 vld4q_dup_s64 (const int64_t * __a)
19525 {
19526   int64x2x4_t ret;
19527   __builtin_aarch64_simd_xi __o;
19528   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19529   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
19530   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
19531   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
19532   ret.val[3] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
19533   return ret;
19534 }
19535
19536 __extension__ extern __inline uint8x16x4_t
19537 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19538 vld4q_dup_u8 (const uint8_t * __a)
19539 {
19540   uint8x16x4_t ret;
19541   __builtin_aarch64_simd_xi __o;
19542   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19543   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19544   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19545   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19546   ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19547   return ret;
19548 }
19549
19550 __extension__ extern __inline uint16x8x4_t
19551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19552 vld4q_dup_u16 (const uint16_t * __a)
19553 {
19554   uint16x8x4_t ret;
19555   __builtin_aarch64_simd_xi __o;
19556   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19557   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19558   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19559   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19560   ret.val[3] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19561   return ret;
19562 }
19563
19564 __extension__ extern __inline uint32x4x4_t
19565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19566 vld4q_dup_u32 (const uint32_t * __a)
19567 {
19568   uint32x4x4_t ret;
19569   __builtin_aarch64_simd_xi __o;
19570   __o = __builtin_aarch64_ld4rv4si ((const __builtin_aarch64_simd_si *) __a);
19571   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
19572   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
19573   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
19574   ret.val[3] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
19575   return ret;
19576 }
19577
19578 __extension__ extern __inline uint64x2x4_t
19579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19580 vld4q_dup_u64 (const uint64_t * __a)
19581 {
19582   uint64x2x4_t ret;
19583   __builtin_aarch64_simd_xi __o;
19584   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19585   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
19586   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
19587   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
19588   ret.val[3] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
19589   return ret;
19590 }
19591
19592 __extension__ extern __inline float16x8x4_t
19593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19594 vld4q_dup_f16 (const float16_t * __a)
19595 {
19596   float16x8x4_t ret;
19597   __builtin_aarch64_simd_xi __o;
19598   __o = __builtin_aarch64_ld4rv8hf ((const __builtin_aarch64_simd_hf *) __a);
19599   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 0);
19600   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 1);
19601   ret.val[2] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 2);
19602   ret.val[3] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 3);
19603   return ret;
19604 }
19605
19606 __extension__ extern __inline float32x4x4_t
19607 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19608 vld4q_dup_f32 (const float32_t * __a)
19609 {
19610   float32x4x4_t ret;
19611   __builtin_aarch64_simd_xi __o;
19612   __o = __builtin_aarch64_ld4rv4sf ((const __builtin_aarch64_simd_sf *) __a);
19613   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 0);
19614   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 1);
19615   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 2);
19616   ret.val[3] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 3);
19617   return ret;
19618 }
19619
19620 __extension__ extern __inline float64x2x4_t
19621 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19622 vld4q_dup_f64 (const float64_t * __a)
19623 {
19624   float64x2x4_t ret;
19625   __builtin_aarch64_simd_xi __o;
19626   __o = __builtin_aarch64_ld4rv2df ((const __builtin_aarch64_simd_df *) __a);
19627   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 0);
19628   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 1);
19629   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 2);
19630   ret.val[3] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 3);
19631   return ret;
19632 }
19633
19634 __extension__ extern __inline poly64x2x4_t
19635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19636 vld4q_dup_p64 (const poly64_t * __a)
19637 {
19638   poly64x2x4_t ret;
19639   __builtin_aarch64_simd_xi __o;
19640   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19641   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 0);
19642   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 1);
19643   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 2);
19644   ret.val[3] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 3);
19645   return ret;
19646 }
19647
19648 /* vld2_lane */
19649
19650 #define __LD2_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
19651                          qmode, ptrmode, funcsuffix, signedtype)           \
19652 __extension__ extern __inline intype \
19653 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19654 vld2_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
19655 {                                                                          \
19656   __builtin_aarch64_simd_oi __o;                                           \
19657   largetype __temp;                                                        \
19658   __temp.val[0] =                                                          \
19659     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
19660   __temp.val[1] =                                                          \
19661     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
19662   __o = __builtin_aarch64_set_qregoi##qmode (__o,                          \
19663                                             (signedtype) __temp.val[0],    \
19664                                             0);                            \
19665   __o = __builtin_aarch64_set_qregoi##qmode (__o,                          \
19666                                             (signedtype) __temp.val[1],    \
19667                                             1);                            \
19668   __o = __builtin_aarch64_ld2_lane##mode (                                 \
19669           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
19670   __b.val[0] = (vectype) __builtin_aarch64_get_dregoidi (__o, 0);          \
19671   __b.val[1] = (vectype) __builtin_aarch64_get_dregoidi (__o, 1);          \
19672   return __b;                                                              \
19673 }
19674
19675 __LD2_LANE_FUNC (float16x4x2_t, float16x4_t, float16x8x2_t, float16_t, v4hf,
19676                  v8hf, hf, f16, float16x8_t)
19677 __LD2_LANE_FUNC (float32x2x2_t, float32x2_t, float32x4x2_t, float32_t, v2sf, v4sf,
19678                  sf, f32, float32x4_t)
19679 __LD2_LANE_FUNC (float64x1x2_t, float64x1_t, float64x2x2_t, float64_t, df, v2df,
19680                  df, f64, float64x2_t)
19681 __LD2_LANE_FUNC (poly8x8x2_t, poly8x8_t, poly8x16x2_t, poly8_t, v8qi, v16qi, qi, p8,
19682                  int8x16_t)
19683 __LD2_LANE_FUNC (poly16x4x2_t, poly16x4_t, poly16x8x2_t, poly16_t, v4hi, v8hi, hi,
19684                  p16, int16x8_t)
19685 __LD2_LANE_FUNC (poly64x1x2_t, poly64x1_t, poly64x2x2_t, poly64_t, di,
19686                  v2di_ssps, di, p64, poly64x2_t)
19687 __LD2_LANE_FUNC (int8x8x2_t, int8x8_t, int8x16x2_t, int8_t, v8qi, v16qi, qi, s8,
19688                  int8x16_t)
19689 __LD2_LANE_FUNC (int16x4x2_t, int16x4_t, int16x8x2_t, int16_t, v4hi, v8hi, hi, s16,
19690                  int16x8_t)
19691 __LD2_LANE_FUNC (int32x2x2_t, int32x2_t, int32x4x2_t, int32_t, v2si, v4si, si, s32,
19692                  int32x4_t)
19693 __LD2_LANE_FUNC (int64x1x2_t, int64x1_t, int64x2x2_t, int64_t, di, v2di, di, s64,
19694                  int64x2_t)
19695 __LD2_LANE_FUNC (uint8x8x2_t, uint8x8_t, uint8x16x2_t, uint8_t, v8qi, v16qi, qi, u8,
19696                  int8x16_t)
19697 __LD2_LANE_FUNC (uint16x4x2_t, uint16x4_t, uint16x8x2_t, uint16_t, v4hi, v8hi, hi,
19698                  u16, int16x8_t)
19699 __LD2_LANE_FUNC (uint32x2x2_t, uint32x2_t, uint32x4x2_t, uint32_t, v2si, v4si, si,
19700                  u32, int32x4_t)
19701 __LD2_LANE_FUNC (uint64x1x2_t, uint64x1_t, uint64x2x2_t, uint64_t, di, v2di, di,
19702                  u64, int64x2_t)
19703
19704 #undef __LD2_LANE_FUNC
19705
19706 /* vld2q_lane */
19707
19708 #define __LD2_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
19709 __extension__ extern __inline intype \
19710 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19711 vld2q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
19712 {                                                                          \
19713   __builtin_aarch64_simd_oi __o;                                           \
19714   intype ret;                                                              \
19715   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) __b.val[0], 0); \
19716   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) __b.val[1], 1); \
19717   __o = __builtin_aarch64_ld2_lane##mode (                                 \
19718         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
19719   ret.val[0] = (vtype) __builtin_aarch64_get_qregoiv4si (__o, 0);          \
19720   ret.val[1] = (vtype) __builtin_aarch64_get_qregoiv4si (__o, 1);          \
19721   return ret;                                                              \
19722 }
19723
19724 __LD2_LANE_FUNC (float16x8x2_t, float16x8_t, float16_t, v8hf, hf, f16)
19725 __LD2_LANE_FUNC (float32x4x2_t, float32x4_t, float32_t, v4sf, sf, f32)
19726 __LD2_LANE_FUNC (float64x2x2_t, float64x2_t, float64_t, v2df, df, f64)
19727 __LD2_LANE_FUNC (poly8x16x2_t, poly8x16_t, poly8_t, v16qi, qi, p8)
19728 __LD2_LANE_FUNC (poly16x8x2_t, poly16x8_t, poly16_t, v8hi, hi, p16)
19729 __LD2_LANE_FUNC (poly64x2x2_t, poly64x2_t, poly64_t, v2di, di, p64)
19730 __LD2_LANE_FUNC (int8x16x2_t, int8x16_t, int8_t, v16qi, qi, s8)
19731 __LD2_LANE_FUNC (int16x8x2_t, int16x8_t, int16_t, v8hi, hi, s16)
19732 __LD2_LANE_FUNC (int32x4x2_t, int32x4_t, int32_t, v4si, si, s32)
19733 __LD2_LANE_FUNC (int64x2x2_t, int64x2_t, int64_t, v2di, di, s64)
19734 __LD2_LANE_FUNC (uint8x16x2_t, uint8x16_t, uint8_t, v16qi, qi, u8)
19735 __LD2_LANE_FUNC (uint16x8x2_t, uint16x8_t, uint16_t, v8hi, hi, u16)
19736 __LD2_LANE_FUNC (uint32x4x2_t, uint32x4_t, uint32_t, v4si, si, u32)
19737 __LD2_LANE_FUNC (uint64x2x2_t, uint64x2_t, uint64_t, v2di, di, u64)
19738
19739 #undef __LD2_LANE_FUNC
19740
19741 /* vld3_lane */
19742
19743 #define __LD3_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
19744                          qmode, ptrmode, funcsuffix, signedtype)           \
19745 __extension__ extern __inline intype \
19746 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19747 vld3_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
19748 {                                                                          \
19749   __builtin_aarch64_simd_ci __o;                                           \
19750   largetype __temp;                                                        \
19751   __temp.val[0] =                                                          \
19752     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
19753   __temp.val[1] =                                                          \
19754     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
19755   __temp.val[2] =                                                          \
19756     vcombine_##funcsuffix (__b.val[2], vcreate_##funcsuffix (0));          \
19757   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
19758                                             (signedtype) __temp.val[0],    \
19759                                             0);                            \
19760   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
19761                                             (signedtype) __temp.val[1],    \
19762                                             1);                            \
19763   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
19764                                             (signedtype) __temp.val[2],    \
19765                                             2);                            \
19766   __o = __builtin_aarch64_ld3_lane##mode (                                 \
19767           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
19768   __b.val[0] = (vectype) __builtin_aarch64_get_dregcidi (__o, 0);          \
19769   __b.val[1] = (vectype) __builtin_aarch64_get_dregcidi (__o, 1);          \
19770   __b.val[2] = (vectype) __builtin_aarch64_get_dregcidi (__o, 2);          \
19771   return __b;                                                              \
19772 }
19773
19774 __LD3_LANE_FUNC (float16x4x3_t, float16x4_t, float16x8x3_t, float16_t, v4hf,
19775                  v8hf, hf, f16, float16x8_t)
19776 __LD3_LANE_FUNC (float32x2x3_t, float32x2_t, float32x4x3_t, float32_t, v2sf, v4sf,
19777                  sf, f32, float32x4_t)
19778 __LD3_LANE_FUNC (float64x1x3_t, float64x1_t, float64x2x3_t, float64_t, df, v2df,
19779                  df, f64, float64x2_t)
19780 __LD3_LANE_FUNC (poly8x8x3_t, poly8x8_t, poly8x16x3_t, poly8_t, v8qi, v16qi, qi, p8,
19781                  int8x16_t)
19782 __LD3_LANE_FUNC (poly16x4x3_t, poly16x4_t, poly16x8x3_t, poly16_t, v4hi, v8hi, hi,
19783                  p16, int16x8_t)
19784 __LD3_LANE_FUNC (poly64x1x3_t, poly64x1_t, poly64x2x3_t, poly64_t, di,
19785                  v2di_ssps, di, p64, poly64x2_t)
19786 __LD3_LANE_FUNC (int8x8x3_t, int8x8_t, int8x16x3_t, int8_t, v8qi, v16qi, qi, s8,
19787                  int8x16_t)
19788 __LD3_LANE_FUNC (int16x4x3_t, int16x4_t, int16x8x3_t, int16_t, v4hi, v8hi, hi, s16,
19789                  int16x8_t)
19790 __LD3_LANE_FUNC (int32x2x3_t, int32x2_t, int32x4x3_t, int32_t, v2si, v4si, si, s32,
19791                  int32x4_t)
19792 __LD3_LANE_FUNC (int64x1x3_t, int64x1_t, int64x2x3_t, int64_t, di, v2di, di, s64,
19793                  int64x2_t)
19794 __LD3_LANE_FUNC (uint8x8x3_t, uint8x8_t, uint8x16x3_t, uint8_t, v8qi, v16qi, qi, u8,
19795                  int8x16_t)
19796 __LD3_LANE_FUNC (uint16x4x3_t, uint16x4_t, uint16x8x3_t, uint16_t, v4hi, v8hi, hi,
19797                  u16, int16x8_t)
19798 __LD3_LANE_FUNC (uint32x2x3_t, uint32x2_t, uint32x4x3_t, uint32_t, v2si, v4si, si,
19799                  u32, int32x4_t)
19800 __LD3_LANE_FUNC (uint64x1x3_t, uint64x1_t, uint64x2x3_t, uint64_t, di, v2di, di,
19801                  u64, int64x2_t)
19802
19803 #undef __LD3_LANE_FUNC
19804
19805 /* vld3q_lane */
19806
19807 #define __LD3_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
19808 __extension__ extern __inline intype \
19809 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19810 vld3q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
19811 {                                                                          \
19812   __builtin_aarch64_simd_ci __o;                                           \
19813   intype ret;                                                              \
19814   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[0], 0); \
19815   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[1], 1); \
19816   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[2], 2); \
19817   __o = __builtin_aarch64_ld3_lane##mode (                                 \
19818         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
19819   ret.val[0] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 0);          \
19820   ret.val[1] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 1);          \
19821   ret.val[2] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 2);          \
19822   return ret;                                                              \
19823 }
19824
19825 __LD3_LANE_FUNC (float16x8x3_t, float16x8_t, float16_t, v8hf, hf, f16)
19826 __LD3_LANE_FUNC (float32x4x3_t, float32x4_t, float32_t, v4sf, sf, f32)
19827 __LD3_LANE_FUNC (float64x2x3_t, float64x2_t, float64_t, v2df, df, f64)
19828 __LD3_LANE_FUNC (poly8x16x3_t, poly8x16_t, poly8_t, v16qi, qi, p8)
19829 __LD3_LANE_FUNC (poly16x8x3_t, poly16x8_t, poly16_t, v8hi, hi, p16)
19830 __LD3_LANE_FUNC (poly64x2x3_t, poly64x2_t, poly64_t, v2di, di, p64)
19831 __LD3_LANE_FUNC (int8x16x3_t, int8x16_t, int8_t, v16qi, qi, s8)
19832 __LD3_LANE_FUNC (int16x8x3_t, int16x8_t, int16_t, v8hi, hi, s16)
19833 __LD3_LANE_FUNC (int32x4x3_t, int32x4_t, int32_t, v4si, si, s32)
19834 __LD3_LANE_FUNC (int64x2x3_t, int64x2_t, int64_t, v2di, di, s64)
19835 __LD3_LANE_FUNC (uint8x16x3_t, uint8x16_t, uint8_t, v16qi, qi, u8)
19836 __LD3_LANE_FUNC (uint16x8x3_t, uint16x8_t, uint16_t, v8hi, hi, u16)
19837 __LD3_LANE_FUNC (uint32x4x3_t, uint32x4_t, uint32_t, v4si, si, u32)
19838 __LD3_LANE_FUNC (uint64x2x3_t, uint64x2_t, uint64_t, v2di, di, u64)
19839
19840 #undef __LD3_LANE_FUNC
19841
19842 /* vld4_lane */
19843
19844 #define __LD4_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
19845                          qmode, ptrmode, funcsuffix, signedtype)           \
19846 __extension__ extern __inline intype \
19847 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19848 vld4_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
19849 {                                                                          \
19850   __builtin_aarch64_simd_xi __o;                                           \
19851   largetype __temp;                                                        \
19852   __temp.val[0] =                                                          \
19853     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
19854   __temp.val[1] =                                                          \
19855     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
19856   __temp.val[2] =                                                          \
19857     vcombine_##funcsuffix (__b.val[2], vcreate_##funcsuffix (0));          \
19858   __temp.val[3] =                                                          \
19859     vcombine_##funcsuffix (__b.val[3], vcreate_##funcsuffix (0));          \
19860   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19861                                             (signedtype) __temp.val[0],    \
19862                                             0);                            \
19863   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19864                                             (signedtype) __temp.val[1],    \
19865                                             1);                            \
19866   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19867                                             (signedtype) __temp.val[2],    \
19868                                             2);                            \
19869   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19870                                             (signedtype) __temp.val[3],    \
19871                                             3);                            \
19872   __o = __builtin_aarch64_ld4_lane##mode (                                 \
19873           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
19874   __b.val[0] = (vectype) __builtin_aarch64_get_dregxidi (__o, 0);          \
19875   __b.val[1] = (vectype) __builtin_aarch64_get_dregxidi (__o, 1);          \
19876   __b.val[2] = (vectype) __builtin_aarch64_get_dregxidi (__o, 2);          \
19877   __b.val[3] = (vectype) __builtin_aarch64_get_dregxidi (__o, 3);          \
19878   return __b;                                                              \
19879 }
19880
19881 /* vld4q_lane */
19882
19883 __LD4_LANE_FUNC (float16x4x4_t, float16x4_t, float16x8x4_t, float16_t, v4hf,
19884                  v8hf, hf, f16, float16x8_t)
19885 __LD4_LANE_FUNC (float32x2x4_t, float32x2_t, float32x4x4_t, float32_t, v2sf, v4sf,
19886                  sf, f32, float32x4_t)
19887 __LD4_LANE_FUNC (float64x1x4_t, float64x1_t, float64x2x4_t, float64_t, df, v2df,
19888                  df, f64, float64x2_t)
19889 __LD4_LANE_FUNC (poly8x8x4_t, poly8x8_t, poly8x16x4_t, poly8_t, v8qi, v16qi, qi, p8,
19890                  int8x16_t)
19891 __LD4_LANE_FUNC (poly16x4x4_t, poly16x4_t, poly16x8x4_t, poly16_t, v4hi, v8hi, hi,
19892                  p16, int16x8_t)
19893 __LD4_LANE_FUNC (poly64x1x4_t, poly64x1_t, poly64x2x4_t, poly64_t, di,
19894                  v2di_ssps, di, p64, poly64x2_t)
19895 __LD4_LANE_FUNC (int8x8x4_t, int8x8_t, int8x16x4_t, int8_t, v8qi, v16qi, qi, s8,
19896                  int8x16_t)
19897 __LD4_LANE_FUNC (int16x4x4_t, int16x4_t, int16x8x4_t, int16_t, v4hi, v8hi, hi, s16,
19898                  int16x8_t)
19899 __LD4_LANE_FUNC (int32x2x4_t, int32x2_t, int32x4x4_t, int32_t, v2si, v4si, si, s32,
19900                  int32x4_t)
19901 __LD4_LANE_FUNC (int64x1x4_t, int64x1_t, int64x2x4_t, int64_t, di, v2di, di, s64,
19902                  int64x2_t)
19903 __LD4_LANE_FUNC (uint8x8x4_t, uint8x8_t, uint8x16x4_t, uint8_t, v8qi, v16qi, qi, u8,
19904                  int8x16_t)
19905 __LD4_LANE_FUNC (uint16x4x4_t, uint16x4_t, uint16x8x4_t, uint16_t, v4hi, v8hi, hi,
19906                  u16, int16x8_t)
19907 __LD4_LANE_FUNC (uint32x2x4_t, uint32x2_t, uint32x4x4_t, uint32_t, v2si, v4si, si,
19908                  u32, int32x4_t)
19909 __LD4_LANE_FUNC (uint64x1x4_t, uint64x1_t, uint64x2x4_t, uint64_t, di, v2di, di,
19910                  u64, int64x2_t)
19911
19912 #undef __LD4_LANE_FUNC
19913
19914 /* vld4q_lane */
19915
19916 #define __LD4_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
19917 __extension__ extern __inline intype \
19918 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19919 vld4q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
19920 {                                                                          \
19921   __builtin_aarch64_simd_xi __o;                                           \
19922   intype ret;                                                              \
19923   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[0], 0); \
19924   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[1], 1); \
19925   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[2], 2); \
19926   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[3], 3); \
19927   __o = __builtin_aarch64_ld4_lane##mode (                                 \
19928         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
19929   ret.val[0] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 0);          \
19930   ret.val[1] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 1);          \
19931   ret.val[2] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 2);          \
19932   ret.val[3] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 3);          \
19933   return ret;                                                              \
19934 }
19935
19936 __LD4_LANE_FUNC (float16x8x4_t, float16x8_t, float16_t, v8hf, hf, f16)
19937 __LD4_LANE_FUNC (float32x4x4_t, float32x4_t, float32_t, v4sf, sf, f32)
19938 __LD4_LANE_FUNC (float64x2x4_t, float64x2_t, float64_t, v2df, df, f64)
19939 __LD4_LANE_FUNC (poly8x16x4_t, poly8x16_t, poly8_t, v16qi, qi, p8)
19940 __LD4_LANE_FUNC (poly16x8x4_t, poly16x8_t, poly16_t, v8hi, hi, p16)
19941 __LD4_LANE_FUNC (poly64x2x4_t, poly64x2_t, poly64_t, v2di, di, p64)
19942 __LD4_LANE_FUNC (int8x16x4_t, int8x16_t, int8_t, v16qi, qi, s8)
19943 __LD4_LANE_FUNC (int16x8x4_t, int16x8_t, int16_t, v8hi, hi, s16)
19944 __LD4_LANE_FUNC (int32x4x4_t, int32x4_t, int32_t, v4si, si, s32)
19945 __LD4_LANE_FUNC (int64x2x4_t, int64x2_t, int64_t, v2di, di, s64)
19946 __LD4_LANE_FUNC (uint8x16x4_t, uint8x16_t, uint8_t, v16qi, qi, u8)
19947 __LD4_LANE_FUNC (uint16x8x4_t, uint16x8_t, uint16_t, v8hi, hi, u16)
19948 __LD4_LANE_FUNC (uint32x4x4_t, uint32x4_t, uint32_t, v4si, si, u32)
19949 __LD4_LANE_FUNC (uint64x2x4_t, uint64x2_t, uint64_t, v2di, di, u64)
19950
19951 #undef __LD4_LANE_FUNC
19952
19953 /* vmax */
19954
19955 __extension__ extern __inline float32x2_t
19956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19957 vmax_f32 (float32x2_t __a, float32x2_t __b)
19958 {
19959   return __builtin_aarch64_smax_nanv2sf (__a, __b);
19960 }
19961
19962 __extension__ extern __inline float64x1_t
19963 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19964 vmax_f64 (float64x1_t __a, float64x1_t __b)
19965 {
19966     return (float64x1_t)
19967       { __builtin_aarch64_smax_nandf (vget_lane_f64 (__a, 0),
19968                                       vget_lane_f64 (__b, 0)) };
19969 }
19970
19971 __extension__ extern __inline int8x8_t
19972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19973 vmax_s8 (int8x8_t __a, int8x8_t __b)
19974 {
19975   return __builtin_aarch64_smaxv8qi (__a, __b);
19976 }
19977
19978 __extension__ extern __inline int16x4_t
19979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19980 vmax_s16 (int16x4_t __a, int16x4_t __b)
19981 {
19982   return __builtin_aarch64_smaxv4hi (__a, __b);
19983 }
19984
19985 __extension__ extern __inline int32x2_t
19986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19987 vmax_s32 (int32x2_t __a, int32x2_t __b)
19988 {
19989   return __builtin_aarch64_smaxv2si (__a, __b);
19990 }
19991
19992 __extension__ extern __inline uint8x8_t
19993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19994 vmax_u8 (uint8x8_t __a, uint8x8_t __b)
19995 {
19996   return (uint8x8_t) __builtin_aarch64_umaxv8qi ((int8x8_t) __a,
19997                                                  (int8x8_t) __b);
19998 }
19999
20000 __extension__ extern __inline uint16x4_t
20001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20002 vmax_u16 (uint16x4_t __a, uint16x4_t __b)
20003 {
20004   return (uint16x4_t) __builtin_aarch64_umaxv4hi ((int16x4_t) __a,
20005                                                   (int16x4_t) __b);
20006 }
20007
20008 __extension__ extern __inline uint32x2_t
20009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20010 vmax_u32 (uint32x2_t __a, uint32x2_t __b)
20011 {
20012   return (uint32x2_t) __builtin_aarch64_umaxv2si ((int32x2_t) __a,
20013                                                   (int32x2_t) __b);
20014 }
20015
20016 __extension__ extern __inline float32x4_t
20017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20018 vmaxq_f32 (float32x4_t __a, float32x4_t __b)
20019 {
20020   return __builtin_aarch64_smax_nanv4sf (__a, __b);
20021 }
20022
20023 __extension__ extern __inline float64x2_t
20024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20025 vmaxq_f64 (float64x2_t __a, float64x2_t __b)
20026 {
20027   return __builtin_aarch64_smax_nanv2df (__a, __b);
20028 }
20029
20030 __extension__ extern __inline int8x16_t
20031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20032 vmaxq_s8 (int8x16_t __a, int8x16_t __b)
20033 {
20034   return __builtin_aarch64_smaxv16qi (__a, __b);
20035 }
20036
20037 __extension__ extern __inline int16x8_t
20038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20039 vmaxq_s16 (int16x8_t __a, int16x8_t __b)
20040 {
20041   return __builtin_aarch64_smaxv8hi (__a, __b);
20042 }
20043
20044 __extension__ extern __inline int32x4_t
20045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20046 vmaxq_s32 (int32x4_t __a, int32x4_t __b)
20047 {
20048   return __builtin_aarch64_smaxv4si (__a, __b);
20049 }
20050
20051 __extension__ extern __inline uint8x16_t
20052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20053 vmaxq_u8 (uint8x16_t __a, uint8x16_t __b)
20054 {
20055   return (uint8x16_t) __builtin_aarch64_umaxv16qi ((int8x16_t) __a,
20056                                                    (int8x16_t) __b);
20057 }
20058
20059 __extension__ extern __inline uint16x8_t
20060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20061 vmaxq_u16 (uint16x8_t __a, uint16x8_t __b)
20062 {
20063   return (uint16x8_t) __builtin_aarch64_umaxv8hi ((int16x8_t) __a,
20064                                                   (int16x8_t) __b);
20065 }
20066
20067 __extension__ extern __inline uint32x4_t
20068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20069 vmaxq_u32 (uint32x4_t __a, uint32x4_t __b)
20070 {
20071   return (uint32x4_t) __builtin_aarch64_umaxv4si ((int32x4_t) __a,
20072                                                   (int32x4_t) __b);
20073 }
20074 /* vmulx */
20075
20076 __extension__ extern __inline float32x2_t
20077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20078 vmulx_f32 (float32x2_t __a, float32x2_t __b)
20079 {
20080   return __builtin_aarch64_fmulxv2sf (__a, __b);
20081 }
20082
20083 __extension__ extern __inline float32x4_t
20084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20085 vmulxq_f32 (float32x4_t __a, float32x4_t __b)
20086 {
20087   return __builtin_aarch64_fmulxv4sf (__a, __b);
20088 }
20089
20090 __extension__ extern __inline float64x1_t
20091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20092 vmulx_f64 (float64x1_t __a, float64x1_t __b)
20093 {
20094   return (float64x1_t) {__builtin_aarch64_fmulxdf (__a[0], __b[0])};
20095 }
20096
20097 __extension__ extern __inline float64x2_t
20098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20099 vmulxq_f64 (float64x2_t __a, float64x2_t __b)
20100 {
20101   return __builtin_aarch64_fmulxv2df (__a, __b);
20102 }
20103
20104 __extension__ extern __inline float32_t
20105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20106 vmulxs_f32 (float32_t __a, float32_t __b)
20107 {
20108   return __builtin_aarch64_fmulxsf (__a, __b);
20109 }
20110
20111 __extension__ extern __inline float64_t
20112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20113 vmulxd_f64 (float64_t __a, float64_t __b)
20114 {
20115   return __builtin_aarch64_fmulxdf (__a, __b);
20116 }
20117
20118 __extension__ extern __inline float32x2_t
20119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20120 vmulx_lane_f32 (float32x2_t __a, float32x2_t __v, const int __lane)
20121 {
20122   return vmulx_f32 (__a, __aarch64_vdup_lane_f32 (__v, __lane));
20123 }
20124
20125 __extension__ extern __inline float64x1_t
20126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20127 vmulx_lane_f64 (float64x1_t __a, float64x1_t __v, const int __lane)
20128 {
20129   return vmulx_f64 (__a, __aarch64_vdup_lane_f64 (__v, __lane));
20130 }
20131
20132 __extension__ extern __inline float32x4_t
20133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20134 vmulxq_lane_f32 (float32x4_t __a, float32x2_t __v, const int __lane)
20135 {
20136   return vmulxq_f32 (__a, __aarch64_vdupq_lane_f32 (__v, __lane));
20137 }
20138
20139 __extension__ extern __inline float64x2_t
20140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20141 vmulxq_lane_f64 (float64x2_t __a, float64x1_t __v, const int __lane)
20142 {
20143   return vmulxq_f64 (__a, __aarch64_vdupq_lane_f64 (__v, __lane));
20144 }
20145
20146 __extension__ extern __inline float32x2_t
20147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20148 vmulx_laneq_f32 (float32x2_t __a, float32x4_t __v, const int __lane)
20149 {
20150   return vmulx_f32 (__a, __aarch64_vdup_laneq_f32 (__v, __lane));
20151 }
20152
20153 __extension__ extern __inline float64x1_t
20154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20155 vmulx_laneq_f64 (float64x1_t __a, float64x2_t __v, const int __lane)
20156 {
20157   return vmulx_f64 (__a, __aarch64_vdup_laneq_f64 (__v, __lane));
20158 }
20159
20160 __extension__ extern __inline float32x4_t
20161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20162 vmulxq_laneq_f32 (float32x4_t __a, float32x4_t __v, const int __lane)
20163 {
20164   return vmulxq_f32 (__a, __aarch64_vdupq_laneq_f32 (__v, __lane));
20165 }
20166
20167 __extension__ extern __inline float64x2_t
20168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20169 vmulxq_laneq_f64 (float64x2_t __a, float64x2_t __v, const int __lane)
20170 {
20171   return vmulxq_f64 (__a, __aarch64_vdupq_laneq_f64 (__v, __lane));
20172 }
20173
20174 __extension__ extern __inline float32_t
20175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20176 vmulxs_lane_f32 (float32_t __a, float32x2_t __v, const int __lane)
20177 {
20178   return vmulxs_f32 (__a, __aarch64_vget_lane_any (__v, __lane));
20179 }
20180
20181 __extension__ extern __inline float32_t
20182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20183 vmulxs_laneq_f32 (float32_t __a, float32x4_t __v, const int __lane)
20184 {
20185   return vmulxs_f32 (__a, __aarch64_vget_lane_any (__v, __lane));
20186 }
20187
20188 __extension__ extern __inline float64_t
20189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20190 vmulxd_lane_f64 (float64_t __a, float64x1_t __v, const int __lane)
20191 {
20192   return vmulxd_f64 (__a, __aarch64_vget_lane_any (__v, __lane));
20193 }
20194
20195 __extension__ extern __inline float64_t
20196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20197 vmulxd_laneq_f64 (float64_t __a, float64x2_t __v, const int __lane)
20198 {
20199   return vmulxd_f64 (__a, __aarch64_vget_lane_any (__v, __lane));
20200 }
20201
20202 /* vpmax  */
20203
20204 __extension__ extern __inline int8x8_t
20205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20206 vpmax_s8 (int8x8_t a, int8x8_t b)
20207 {
20208   return __builtin_aarch64_smaxpv8qi (a, b);
20209 }
20210
20211 __extension__ extern __inline int16x4_t
20212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20213 vpmax_s16 (int16x4_t a, int16x4_t b)
20214 {
20215   return __builtin_aarch64_smaxpv4hi (a, b);
20216 }
20217
20218 __extension__ extern __inline int32x2_t
20219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20220 vpmax_s32 (int32x2_t a, int32x2_t b)
20221 {
20222   return __builtin_aarch64_smaxpv2si (a, b);
20223 }
20224
20225 __extension__ extern __inline uint8x8_t
20226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20227 vpmax_u8 (uint8x8_t a, uint8x8_t b)
20228 {
20229   return (uint8x8_t) __builtin_aarch64_umaxpv8qi ((int8x8_t) a,
20230                                                   (int8x8_t) b);
20231 }
20232
20233 __extension__ extern __inline uint16x4_t
20234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20235 vpmax_u16 (uint16x4_t a, uint16x4_t b)
20236 {
20237   return (uint16x4_t) __builtin_aarch64_umaxpv4hi ((int16x4_t) a,
20238                                                    (int16x4_t) b);
20239 }
20240
20241 __extension__ extern __inline uint32x2_t
20242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20243 vpmax_u32 (uint32x2_t a, uint32x2_t b)
20244 {
20245   return (uint32x2_t) __builtin_aarch64_umaxpv2si ((int32x2_t) a,
20246                                                    (int32x2_t) b);
20247 }
20248
20249 __extension__ extern __inline int8x16_t
20250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20251 vpmaxq_s8 (int8x16_t a, int8x16_t b)
20252 {
20253   return __builtin_aarch64_smaxpv16qi (a, b);
20254 }
20255
20256 __extension__ extern __inline int16x8_t
20257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20258 vpmaxq_s16 (int16x8_t a, int16x8_t b)
20259 {
20260   return __builtin_aarch64_smaxpv8hi (a, b);
20261 }
20262
20263 __extension__ extern __inline int32x4_t
20264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20265 vpmaxq_s32 (int32x4_t a, int32x4_t b)
20266 {
20267   return __builtin_aarch64_smaxpv4si (a, b);
20268 }
20269
20270 __extension__ extern __inline uint8x16_t
20271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20272 vpmaxq_u8 (uint8x16_t a, uint8x16_t b)
20273 {
20274   return (uint8x16_t) __builtin_aarch64_umaxpv16qi ((int8x16_t) a,
20275                                                     (int8x16_t) b);
20276 }
20277
20278 __extension__ extern __inline uint16x8_t
20279 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20280 vpmaxq_u16 (uint16x8_t a, uint16x8_t b)
20281 {
20282   return (uint16x8_t) __builtin_aarch64_umaxpv8hi ((int16x8_t) a,
20283                                                    (int16x8_t) b);
20284 }
20285
20286 __extension__ extern __inline uint32x4_t
20287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20288 vpmaxq_u32 (uint32x4_t a, uint32x4_t b)
20289 {
20290   return (uint32x4_t) __builtin_aarch64_umaxpv4si ((int32x4_t) a,
20291                                                    (int32x4_t) b);
20292 }
20293
20294 __extension__ extern __inline float32x2_t
20295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20296 vpmax_f32 (float32x2_t a, float32x2_t b)
20297 {
20298   return __builtin_aarch64_smax_nanpv2sf (a, b);
20299 }
20300
20301 __extension__ extern __inline float32x4_t
20302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20303 vpmaxq_f32 (float32x4_t a, float32x4_t b)
20304 {
20305   return __builtin_aarch64_smax_nanpv4sf (a, b);
20306 }
20307
20308 __extension__ extern __inline float64x2_t
20309 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20310 vpmaxq_f64 (float64x2_t a, float64x2_t b)
20311 {
20312   return __builtin_aarch64_smax_nanpv2df (a, b);
20313 }
20314
20315 __extension__ extern __inline float64_t
20316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20317 vpmaxqd_f64 (float64x2_t a)
20318 {
20319   return __builtin_aarch64_reduc_smax_nan_scal_v2df (a);
20320 }
20321
20322 __extension__ extern __inline float32_t
20323 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20324 vpmaxs_f32 (float32x2_t a)
20325 {
20326   return __builtin_aarch64_reduc_smax_nan_scal_v2sf (a);
20327 }
20328
20329 /* vpmaxnm  */
20330
20331 __extension__ extern __inline float32x2_t
20332 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20333 vpmaxnm_f32 (float32x2_t a, float32x2_t b)
20334 {
20335   return __builtin_aarch64_smaxpv2sf (a, b);
20336 }
20337
20338 __extension__ extern __inline float32x4_t
20339 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20340 vpmaxnmq_f32 (float32x4_t a, float32x4_t b)
20341 {
20342   return __builtin_aarch64_smaxpv4sf (a, b);
20343 }
20344
20345 __extension__ extern __inline float64x2_t
20346 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20347 vpmaxnmq_f64 (float64x2_t a, float64x2_t b)
20348 {
20349   return __builtin_aarch64_smaxpv2df (a, b);
20350 }
20351
20352 __extension__ extern __inline float64_t
20353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20354 vpmaxnmqd_f64 (float64x2_t a)
20355 {
20356   return __builtin_aarch64_reduc_smax_scal_v2df (a);
20357 }
20358
20359 __extension__ extern __inline float32_t
20360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20361 vpmaxnms_f32 (float32x2_t a)
20362 {
20363   return __builtin_aarch64_reduc_smax_scal_v2sf (a);
20364 }
20365
20366 /* vpmin  */
20367
20368 __extension__ extern __inline int8x8_t
20369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20370 vpmin_s8 (int8x8_t a, int8x8_t b)
20371 {
20372   return __builtin_aarch64_sminpv8qi (a, b);
20373 }
20374
20375 __extension__ extern __inline int16x4_t
20376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20377 vpmin_s16 (int16x4_t a, int16x4_t b)
20378 {
20379   return __builtin_aarch64_sminpv4hi (a, b);
20380 }
20381
20382 __extension__ extern __inline int32x2_t
20383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20384 vpmin_s32 (int32x2_t a, int32x2_t b)
20385 {
20386   return __builtin_aarch64_sminpv2si (a, b);
20387 }
20388
20389 __extension__ extern __inline uint8x8_t
20390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20391 vpmin_u8 (uint8x8_t a, uint8x8_t b)
20392 {
20393   return (uint8x8_t) __builtin_aarch64_uminpv8qi ((int8x8_t) a,
20394                                                   (int8x8_t) b);
20395 }
20396
20397 __extension__ extern __inline uint16x4_t
20398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20399 vpmin_u16 (uint16x4_t a, uint16x4_t b)
20400 {
20401   return (uint16x4_t) __builtin_aarch64_uminpv4hi ((int16x4_t) a,
20402                                                    (int16x4_t) b);
20403 }
20404
20405 __extension__ extern __inline uint32x2_t
20406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20407 vpmin_u32 (uint32x2_t a, uint32x2_t b)
20408 {
20409   return (uint32x2_t) __builtin_aarch64_uminpv2si ((int32x2_t) a,
20410                                                    (int32x2_t) b);
20411 }
20412
20413 __extension__ extern __inline int8x16_t
20414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20415 vpminq_s8 (int8x16_t a, int8x16_t b)
20416 {
20417   return __builtin_aarch64_sminpv16qi (a, b);
20418 }
20419
20420 __extension__ extern __inline int16x8_t
20421 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20422 vpminq_s16 (int16x8_t a, int16x8_t b)
20423 {
20424   return __builtin_aarch64_sminpv8hi (a, b);
20425 }
20426
20427 __extension__ extern __inline int32x4_t
20428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20429 vpminq_s32 (int32x4_t a, int32x4_t b)
20430 {
20431   return __builtin_aarch64_sminpv4si (a, b);
20432 }
20433
20434 __extension__ extern __inline uint8x16_t
20435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20436 vpminq_u8 (uint8x16_t a, uint8x16_t b)
20437 {
20438   return (uint8x16_t) __builtin_aarch64_uminpv16qi ((int8x16_t) a,
20439                                                     (int8x16_t) b);
20440 }
20441
20442 __extension__ extern __inline uint16x8_t
20443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20444 vpminq_u16 (uint16x8_t a, uint16x8_t b)
20445 {
20446   return (uint16x8_t) __builtin_aarch64_uminpv8hi ((int16x8_t) a,
20447                                                    (int16x8_t) b);
20448 }
20449
20450 __extension__ extern __inline uint32x4_t
20451 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20452 vpminq_u32 (uint32x4_t a, uint32x4_t b)
20453 {
20454   return (uint32x4_t) __builtin_aarch64_uminpv4si ((int32x4_t) a,
20455                                                    (int32x4_t) b);
20456 }
20457
20458 __extension__ extern __inline float32x2_t
20459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20460 vpmin_f32 (float32x2_t a, float32x2_t b)
20461 {
20462   return __builtin_aarch64_smin_nanpv2sf (a, b);
20463 }
20464
20465 __extension__ extern __inline float32x4_t
20466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20467 vpminq_f32 (float32x4_t a, float32x4_t b)
20468 {
20469   return __builtin_aarch64_smin_nanpv4sf (a, b);
20470 }
20471
20472 __extension__ extern __inline float64x2_t
20473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20474 vpminq_f64 (float64x2_t a, float64x2_t b)
20475 {
20476   return __builtin_aarch64_smin_nanpv2df (a, b);
20477 }
20478
20479 __extension__ extern __inline float64_t
20480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20481 vpminqd_f64 (float64x2_t a)
20482 {
20483   return __builtin_aarch64_reduc_smin_nan_scal_v2df (a);
20484 }
20485
20486 __extension__ extern __inline float32_t
20487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20488 vpmins_f32 (float32x2_t a)
20489 {
20490   return __builtin_aarch64_reduc_smin_nan_scal_v2sf (a);
20491 }
20492
20493 /* vpminnm  */
20494
20495 __extension__ extern __inline float32x2_t
20496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20497 vpminnm_f32 (float32x2_t a, float32x2_t b)
20498 {
20499   return __builtin_aarch64_sminpv2sf (a, b);
20500 }
20501
20502 __extension__ extern __inline float32x4_t
20503 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20504 vpminnmq_f32 (float32x4_t a, float32x4_t b)
20505 {
20506   return __builtin_aarch64_sminpv4sf (a, b);
20507 }
20508
20509 __extension__ extern __inline float64x2_t
20510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20511 vpminnmq_f64 (float64x2_t a, float64x2_t b)
20512 {
20513   return __builtin_aarch64_sminpv2df (a, b);
20514 }
20515
20516 __extension__ extern __inline float64_t
20517 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20518 vpminnmqd_f64 (float64x2_t a)
20519 {
20520   return __builtin_aarch64_reduc_smin_scal_v2df (a);
20521 }
20522
20523 __extension__ extern __inline float32_t
20524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20525 vpminnms_f32 (float32x2_t a)
20526 {
20527   return __builtin_aarch64_reduc_smin_scal_v2sf (a);
20528 }
20529
20530 /* vmaxnm  */
20531
20532 __extension__ extern __inline float32x2_t
20533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20534 vmaxnm_f32 (float32x2_t __a, float32x2_t __b)
20535 {
20536   return __builtin_aarch64_fmaxv2sf (__a, __b);
20537 }
20538
20539 __extension__ extern __inline float64x1_t
20540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20541 vmaxnm_f64 (float64x1_t __a, float64x1_t __b)
20542 {
20543   return (float64x1_t)
20544     { __builtin_aarch64_fmaxdf (vget_lane_f64 (__a, 0),
20545                                 vget_lane_f64 (__b, 0)) };
20546 }
20547
20548 __extension__ extern __inline float32x4_t
20549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20550 vmaxnmq_f32 (float32x4_t __a, float32x4_t __b)
20551 {
20552   return __builtin_aarch64_fmaxv4sf (__a, __b);
20553 }
20554
20555 __extension__ extern __inline float64x2_t
20556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20557 vmaxnmq_f64 (float64x2_t __a, float64x2_t __b)
20558 {
20559   return __builtin_aarch64_fmaxv2df (__a, __b);
20560 }
20561
20562 /* vmaxv  */
20563
20564 __extension__ extern __inline float32_t
20565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20566 vmaxv_f32 (float32x2_t __a)
20567 {
20568   return __builtin_aarch64_reduc_smax_nan_scal_v2sf (__a);
20569 }
20570
20571 __extension__ extern __inline int8_t
20572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20573 vmaxv_s8 (int8x8_t __a)
20574 {
20575   return __builtin_aarch64_reduc_smax_scal_v8qi (__a);
20576 }
20577
20578 __extension__ extern __inline int16_t
20579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20580 vmaxv_s16 (int16x4_t __a)
20581 {
20582   return __builtin_aarch64_reduc_smax_scal_v4hi (__a);
20583 }
20584
20585 __extension__ extern __inline int32_t
20586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20587 vmaxv_s32 (int32x2_t __a)
20588 {
20589   return __builtin_aarch64_reduc_smax_scal_v2si (__a);
20590 }
20591
20592 __extension__ extern __inline uint8_t
20593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20594 vmaxv_u8 (uint8x8_t __a)
20595 {
20596   return __builtin_aarch64_reduc_umax_scal_v8qi_uu (__a);
20597 }
20598
20599 __extension__ extern __inline uint16_t
20600 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20601 vmaxv_u16 (uint16x4_t __a)
20602 {
20603   return __builtin_aarch64_reduc_umax_scal_v4hi_uu (__a);
20604 }
20605
20606 __extension__ extern __inline uint32_t
20607 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20608 vmaxv_u32 (uint32x2_t __a)
20609 {
20610   return __builtin_aarch64_reduc_umax_scal_v2si_uu (__a);
20611 }
20612
20613 __extension__ extern __inline float32_t
20614 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20615 vmaxvq_f32 (float32x4_t __a)
20616 {
20617   return __builtin_aarch64_reduc_smax_nan_scal_v4sf (__a);
20618 }
20619
20620 __extension__ extern __inline float64_t
20621 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20622 vmaxvq_f64 (float64x2_t __a)
20623 {
20624   return __builtin_aarch64_reduc_smax_nan_scal_v2df (__a);
20625 }
20626
20627 __extension__ extern __inline int8_t
20628 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20629 vmaxvq_s8 (int8x16_t __a)
20630 {
20631   return __builtin_aarch64_reduc_smax_scal_v16qi (__a);
20632 }
20633
20634 __extension__ extern __inline int16_t
20635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20636 vmaxvq_s16 (int16x8_t __a)
20637 {
20638   return __builtin_aarch64_reduc_smax_scal_v8hi (__a);
20639 }
20640
20641 __extension__ extern __inline int32_t
20642 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20643 vmaxvq_s32 (int32x4_t __a)
20644 {
20645   return __builtin_aarch64_reduc_smax_scal_v4si (__a);
20646 }
20647
20648 __extension__ extern __inline uint8_t
20649 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20650 vmaxvq_u8 (uint8x16_t __a)
20651 {
20652   return __builtin_aarch64_reduc_umax_scal_v16qi_uu (__a);
20653 }
20654
20655 __extension__ extern __inline uint16_t
20656 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20657 vmaxvq_u16 (uint16x8_t __a)
20658 {
20659   return __builtin_aarch64_reduc_umax_scal_v8hi_uu (__a);
20660 }
20661
20662 __extension__ extern __inline uint32_t
20663 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20664 vmaxvq_u32 (uint32x4_t __a)
20665 {
20666   return __builtin_aarch64_reduc_umax_scal_v4si_uu (__a);
20667 }
20668
20669 /* vmaxnmv  */
20670
20671 __extension__ extern __inline float32_t
20672 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20673 vmaxnmv_f32 (float32x2_t __a)
20674 {
20675   return __builtin_aarch64_reduc_smax_scal_v2sf (__a);
20676 }
20677
20678 __extension__ extern __inline float32_t
20679 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20680 vmaxnmvq_f32 (float32x4_t __a)
20681 {
20682   return __builtin_aarch64_reduc_smax_scal_v4sf (__a);
20683 }
20684
20685 __extension__ extern __inline float64_t
20686 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20687 vmaxnmvq_f64 (float64x2_t __a)
20688 {
20689   return __builtin_aarch64_reduc_smax_scal_v2df (__a);
20690 }
20691
20692 /* vmin  */
20693
20694 __extension__ extern __inline float32x2_t
20695 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20696 vmin_f32 (float32x2_t __a, float32x2_t __b)
20697 {
20698   return __builtin_aarch64_smin_nanv2sf (__a, __b);
20699 }
20700
20701 __extension__ extern __inline float64x1_t
20702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20703 vmin_f64 (float64x1_t __a, float64x1_t __b)
20704 {
20705     return (float64x1_t)
20706           { __builtin_aarch64_smin_nandf (vget_lane_f64 (__a, 0),
20707                                           vget_lane_f64 (__b, 0)) };
20708 }
20709
20710 __extension__ extern __inline int8x8_t
20711 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20712 vmin_s8 (int8x8_t __a, int8x8_t __b)
20713 {
20714   return __builtin_aarch64_sminv8qi (__a, __b);
20715 }
20716
20717 __extension__ extern __inline int16x4_t
20718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20719 vmin_s16 (int16x4_t __a, int16x4_t __b)
20720 {
20721   return __builtin_aarch64_sminv4hi (__a, __b);
20722 }
20723
20724 __extension__ extern __inline int32x2_t
20725 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20726 vmin_s32 (int32x2_t __a, int32x2_t __b)
20727 {
20728   return __builtin_aarch64_sminv2si (__a, __b);
20729 }
20730
20731 __extension__ extern __inline uint8x8_t
20732 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20733 vmin_u8 (uint8x8_t __a, uint8x8_t __b)
20734 {
20735   return (uint8x8_t) __builtin_aarch64_uminv8qi ((int8x8_t) __a,
20736                                                  (int8x8_t) __b);
20737 }
20738
20739 __extension__ extern __inline uint16x4_t
20740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20741 vmin_u16 (uint16x4_t __a, uint16x4_t __b)
20742 {
20743   return (uint16x4_t) __builtin_aarch64_uminv4hi ((int16x4_t) __a,
20744                                                   (int16x4_t) __b);
20745 }
20746
20747 __extension__ extern __inline uint32x2_t
20748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20749 vmin_u32 (uint32x2_t __a, uint32x2_t __b)
20750 {
20751   return (uint32x2_t) __builtin_aarch64_uminv2si ((int32x2_t) __a,
20752                                                   (int32x2_t) __b);
20753 }
20754
20755 __extension__ extern __inline float32x4_t
20756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20757 vminq_f32 (float32x4_t __a, float32x4_t __b)
20758 {
20759   return __builtin_aarch64_smin_nanv4sf (__a, __b);
20760 }
20761
20762 __extension__ extern __inline float64x2_t
20763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20764 vminq_f64 (float64x2_t __a, float64x2_t __b)
20765 {
20766   return __builtin_aarch64_smin_nanv2df (__a, __b);
20767 }
20768
20769 __extension__ extern __inline int8x16_t
20770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20771 vminq_s8 (int8x16_t __a, int8x16_t __b)
20772 {
20773   return __builtin_aarch64_sminv16qi (__a, __b);
20774 }
20775
20776 __extension__ extern __inline int16x8_t
20777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20778 vminq_s16 (int16x8_t __a, int16x8_t __b)
20779 {
20780   return __builtin_aarch64_sminv8hi (__a, __b);
20781 }
20782
20783 __extension__ extern __inline int32x4_t
20784 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20785 vminq_s32 (int32x4_t __a, int32x4_t __b)
20786 {
20787   return __builtin_aarch64_sminv4si (__a, __b);
20788 }
20789
20790 __extension__ extern __inline uint8x16_t
20791 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20792 vminq_u8 (uint8x16_t __a, uint8x16_t __b)
20793 {
20794   return (uint8x16_t) __builtin_aarch64_uminv16qi ((int8x16_t) __a,
20795                                                    (int8x16_t) __b);
20796 }
20797
20798 __extension__ extern __inline uint16x8_t
20799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20800 vminq_u16 (uint16x8_t __a, uint16x8_t __b)
20801 {
20802   return (uint16x8_t) __builtin_aarch64_uminv8hi ((int16x8_t) __a,
20803                                                   (int16x8_t) __b);
20804 }
20805
20806 __extension__ extern __inline uint32x4_t
20807 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20808 vminq_u32 (uint32x4_t __a, uint32x4_t __b)
20809 {
20810   return (uint32x4_t) __builtin_aarch64_uminv4si ((int32x4_t) __a,
20811                                                   (int32x4_t) __b);
20812 }
20813
20814 /* vminnm  */
20815
20816 __extension__ extern __inline float32x2_t
20817 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20818 vminnm_f32 (float32x2_t __a, float32x2_t __b)
20819 {
20820   return __builtin_aarch64_fminv2sf (__a, __b);
20821 }
20822
20823 __extension__ extern __inline float64x1_t
20824 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20825 vminnm_f64 (float64x1_t __a, float64x1_t __b)
20826 {
20827   return (float64x1_t)
20828     { __builtin_aarch64_fmindf (vget_lane_f64 (__a, 0),
20829                                 vget_lane_f64 (__b, 0)) };
20830 }
20831
20832 __extension__ extern __inline float32x4_t
20833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20834 vminnmq_f32 (float32x4_t __a, float32x4_t __b)
20835 {
20836   return __builtin_aarch64_fminv4sf (__a, __b);
20837 }
20838
20839 __extension__ extern __inline float64x2_t
20840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20841 vminnmq_f64 (float64x2_t __a, float64x2_t __b)
20842 {
20843   return __builtin_aarch64_fminv2df (__a, __b);
20844 }
20845
20846 /* vminv  */
20847
20848 __extension__ extern __inline float32_t
20849 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20850 vminv_f32 (float32x2_t __a)
20851 {
20852   return __builtin_aarch64_reduc_smin_nan_scal_v2sf (__a);
20853 }
20854
20855 __extension__ extern __inline int8_t
20856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20857 vminv_s8 (int8x8_t __a)
20858 {
20859   return __builtin_aarch64_reduc_smin_scal_v8qi (__a);
20860 }
20861
20862 __extension__ extern __inline int16_t
20863 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20864 vminv_s16 (int16x4_t __a)
20865 {
20866   return __builtin_aarch64_reduc_smin_scal_v4hi (__a);
20867 }
20868
20869 __extension__ extern __inline int32_t
20870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20871 vminv_s32 (int32x2_t __a)
20872 {
20873   return __builtin_aarch64_reduc_smin_scal_v2si (__a);
20874 }
20875
20876 __extension__ extern __inline uint8_t
20877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20878 vminv_u8 (uint8x8_t __a)
20879 {
20880   return __builtin_aarch64_reduc_umin_scal_v8qi_uu (__a);
20881 }
20882
20883 __extension__ extern __inline uint16_t
20884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20885 vminv_u16 (uint16x4_t __a)
20886 {
20887   return __builtin_aarch64_reduc_umin_scal_v4hi_uu (__a);
20888 }
20889
20890 __extension__ extern __inline uint32_t
20891 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20892 vminv_u32 (uint32x2_t __a)
20893 {
20894   return __builtin_aarch64_reduc_umin_scal_v2si_uu (__a);
20895 }
20896
20897 __extension__ extern __inline float32_t
20898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20899 vminvq_f32 (float32x4_t __a)
20900 {
20901   return __builtin_aarch64_reduc_smin_nan_scal_v4sf (__a);
20902 }
20903
20904 __extension__ extern __inline float64_t
20905 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20906 vminvq_f64 (float64x2_t __a)
20907 {
20908   return __builtin_aarch64_reduc_smin_nan_scal_v2df (__a);
20909 }
20910
20911 __extension__ extern __inline int8_t
20912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20913 vminvq_s8 (int8x16_t __a)
20914 {
20915   return __builtin_aarch64_reduc_smin_scal_v16qi (__a);
20916 }
20917
20918 __extension__ extern __inline int16_t
20919 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20920 vminvq_s16 (int16x8_t __a)
20921 {
20922   return __builtin_aarch64_reduc_smin_scal_v8hi (__a);
20923 }
20924
20925 __extension__ extern __inline int32_t
20926 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20927 vminvq_s32 (int32x4_t __a)
20928 {
20929   return __builtin_aarch64_reduc_smin_scal_v4si (__a);
20930 }
20931
20932 __extension__ extern __inline uint8_t
20933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20934 vminvq_u8 (uint8x16_t __a)
20935 {
20936   return __builtin_aarch64_reduc_umin_scal_v16qi_uu (__a);
20937 }
20938
20939 __extension__ extern __inline uint16_t
20940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20941 vminvq_u16 (uint16x8_t __a)
20942 {
20943   return __builtin_aarch64_reduc_umin_scal_v8hi_uu (__a);
20944 }
20945
20946 __extension__ extern __inline uint32_t
20947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20948 vminvq_u32 (uint32x4_t __a)
20949 {
20950   return __builtin_aarch64_reduc_umin_scal_v4si_uu (__a);
20951 }
20952
20953 /* vminnmv  */
20954
20955 __extension__ extern __inline float32_t
20956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20957 vminnmv_f32 (float32x2_t __a)
20958 {
20959   return __builtin_aarch64_reduc_smin_scal_v2sf (__a);
20960 }
20961
20962 __extension__ extern __inline float32_t
20963 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20964 vminnmvq_f32 (float32x4_t __a)
20965 {
20966   return __builtin_aarch64_reduc_smin_scal_v4sf (__a);
20967 }
20968
20969 __extension__ extern __inline float64_t
20970 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20971 vminnmvq_f64 (float64x2_t __a)
20972 {
20973   return __builtin_aarch64_reduc_smin_scal_v2df (__a);
20974 }
20975
20976 /* vmla */
20977
20978 __extension__ extern __inline float32x2_t
20979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20980 vmla_f32 (float32x2_t a, float32x2_t b, float32x2_t c)
20981 {
20982   return a + b * c;
20983 }
20984
20985 __extension__ extern __inline float64x1_t
20986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20987 vmla_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
20988 {
20989   return __a + __b * __c;
20990 }
20991
20992 __extension__ extern __inline float32x4_t
20993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20994 vmlaq_f32 (float32x4_t a, float32x4_t b, float32x4_t c)
20995 {
20996   return a + b * c;
20997 }
20998
20999 __extension__ extern __inline float64x2_t
21000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21001 vmlaq_f64 (float64x2_t a, float64x2_t b, float64x2_t c)
21002 {
21003   return a + b * c;
21004 }
21005
21006 /* vmla_lane  */
21007
21008 __extension__ extern __inline float32x2_t
21009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21010 vmla_lane_f32 (float32x2_t __a, float32x2_t __b,
21011                float32x2_t __c, const int __lane)
21012 {
21013   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21014 }
21015
21016 __extension__ extern __inline int16x4_t
21017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21018 vmla_lane_s16 (int16x4_t __a, int16x4_t __b,
21019                 int16x4_t __c, const int __lane)
21020 {
21021   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21022 }
21023
21024 __extension__ extern __inline int32x2_t
21025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21026 vmla_lane_s32 (int32x2_t __a, int32x2_t __b,
21027                 int32x2_t __c, const int __lane)
21028 {
21029   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21030 }
21031
21032 __extension__ extern __inline uint16x4_t
21033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21034 vmla_lane_u16 (uint16x4_t __a, uint16x4_t __b,
21035                 uint16x4_t __c, const int __lane)
21036 {
21037   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21038 }
21039
21040 __extension__ extern __inline uint32x2_t
21041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21042 vmla_lane_u32 (uint32x2_t __a, uint32x2_t __b,
21043                uint32x2_t __c, const int __lane)
21044 {
21045   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21046 }
21047
21048 /* vmla_laneq  */
21049
21050 __extension__ extern __inline float32x2_t
21051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21052 vmla_laneq_f32 (float32x2_t __a, float32x2_t __b,
21053                 float32x4_t __c, const int __lane)
21054 {
21055   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21056 }
21057
21058 __extension__ extern __inline int16x4_t
21059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21060 vmla_laneq_s16 (int16x4_t __a, int16x4_t __b,
21061                 int16x8_t __c, const int __lane)
21062 {
21063   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21064 }
21065
21066 __extension__ extern __inline int32x2_t
21067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21068 vmla_laneq_s32 (int32x2_t __a, int32x2_t __b,
21069                 int32x4_t __c, const int __lane)
21070 {
21071   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21072 }
21073
21074 __extension__ extern __inline uint16x4_t
21075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21076 vmla_laneq_u16 (uint16x4_t __a, uint16x4_t __b,
21077                 uint16x8_t __c, const int __lane)
21078 {
21079   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21080 }
21081
21082 __extension__ extern __inline uint32x2_t
21083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21084 vmla_laneq_u32 (uint32x2_t __a, uint32x2_t __b,
21085                 uint32x4_t __c, const int __lane)
21086 {
21087   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21088 }
21089
21090 /* vmlaq_lane  */
21091
21092 __extension__ extern __inline float32x4_t
21093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21094 vmlaq_lane_f32 (float32x4_t __a, float32x4_t __b,
21095                 float32x2_t __c, const int __lane)
21096 {
21097   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21098 }
21099
21100 __extension__ extern __inline int16x8_t
21101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21102 vmlaq_lane_s16 (int16x8_t __a, int16x8_t __b,
21103                 int16x4_t __c, const int __lane)
21104 {
21105   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21106 }
21107
21108 __extension__ extern __inline int32x4_t
21109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21110 vmlaq_lane_s32 (int32x4_t __a, int32x4_t __b,
21111                 int32x2_t __c, const int __lane)
21112 {
21113   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21114 }
21115
21116 __extension__ extern __inline uint16x8_t
21117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21118 vmlaq_lane_u16 (uint16x8_t __a, uint16x8_t __b,
21119                 uint16x4_t __c, const int __lane)
21120 {
21121   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21122 }
21123
21124 __extension__ extern __inline uint32x4_t
21125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21126 vmlaq_lane_u32 (uint32x4_t __a, uint32x4_t __b,
21127                 uint32x2_t __c, const int __lane)
21128 {
21129   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21130 }
21131
21132   /* vmlaq_laneq  */
21133
21134 __extension__ extern __inline float32x4_t
21135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21136 vmlaq_laneq_f32 (float32x4_t __a, float32x4_t __b,
21137                  float32x4_t __c, const int __lane)
21138 {
21139   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21140 }
21141
21142 __extension__ extern __inline int16x8_t
21143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21144 vmlaq_laneq_s16 (int16x8_t __a, int16x8_t __b,
21145                 int16x8_t __c, const int __lane)
21146 {
21147   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21148 }
21149
21150 __extension__ extern __inline int32x4_t
21151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21152 vmlaq_laneq_s32 (int32x4_t __a, int32x4_t __b,
21153                 int32x4_t __c, const int __lane)
21154 {
21155   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21156 }
21157
21158 __extension__ extern __inline uint16x8_t
21159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21160 vmlaq_laneq_u16 (uint16x8_t __a, uint16x8_t __b,
21161                 uint16x8_t __c, const int __lane)
21162 {
21163   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21164 }
21165
21166 __extension__ extern __inline uint32x4_t
21167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21168 vmlaq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
21169                 uint32x4_t __c, const int __lane)
21170 {
21171   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21172 }
21173
21174 /* vmls  */
21175
21176 __extension__ extern __inline float32x2_t
21177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21178 vmls_f32 (float32x2_t a, float32x2_t b, float32x2_t c)
21179 {
21180   return a - b * c;
21181 }
21182
21183 __extension__ extern __inline float64x1_t
21184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21185 vmls_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
21186 {
21187   return __a - __b * __c;
21188 }
21189
21190 __extension__ extern __inline float32x4_t
21191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21192 vmlsq_f32 (float32x4_t a, float32x4_t b, float32x4_t c)
21193 {
21194   return a - b * c;
21195 }
21196
21197 __extension__ extern __inline float64x2_t
21198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21199 vmlsq_f64 (float64x2_t a, float64x2_t b, float64x2_t c)
21200 {
21201   return a - b * c;
21202 }
21203
21204 /* vmls_lane  */
21205
21206 __extension__ extern __inline float32x2_t
21207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21208 vmls_lane_f32 (float32x2_t __a, float32x2_t __b,
21209                float32x2_t __c, const int __lane)
21210 {
21211   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21212 }
21213
21214 __extension__ extern __inline int16x4_t
21215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21216 vmls_lane_s16 (int16x4_t __a, int16x4_t __b,
21217                 int16x4_t __c, const int __lane)
21218 {
21219   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21220 }
21221
21222 __extension__ extern __inline int32x2_t
21223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21224 vmls_lane_s32 (int32x2_t __a, int32x2_t __b,
21225                 int32x2_t __c, const int __lane)
21226 {
21227   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21228 }
21229
21230 __extension__ extern __inline uint16x4_t
21231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21232 vmls_lane_u16 (uint16x4_t __a, uint16x4_t __b,
21233                 uint16x4_t __c, const int __lane)
21234 {
21235   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21236 }
21237
21238 __extension__ extern __inline uint32x2_t
21239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21240 vmls_lane_u32 (uint32x2_t __a, uint32x2_t __b,
21241                uint32x2_t __c, const int __lane)
21242 {
21243   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21244 }
21245
21246 /* vmls_laneq  */
21247
21248 __extension__ extern __inline float32x2_t
21249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21250 vmls_laneq_f32 (float32x2_t __a, float32x2_t __b,
21251                float32x4_t __c, const int __lane)
21252 {
21253   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21254 }
21255
21256 __extension__ extern __inline int16x4_t
21257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21258 vmls_laneq_s16 (int16x4_t __a, int16x4_t __b,
21259                 int16x8_t __c, const int __lane)
21260 {
21261   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21262 }
21263
21264 __extension__ extern __inline int32x2_t
21265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21266 vmls_laneq_s32 (int32x2_t __a, int32x2_t __b,
21267                 int32x4_t __c, const int __lane)
21268 {
21269   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21270 }
21271
21272 __extension__ extern __inline uint16x4_t
21273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21274 vmls_laneq_u16 (uint16x4_t __a, uint16x4_t __b,
21275                 uint16x8_t __c, const int __lane)
21276 {
21277   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21278 }
21279
21280 __extension__ extern __inline uint32x2_t
21281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21282 vmls_laneq_u32 (uint32x2_t __a, uint32x2_t __b,
21283                 uint32x4_t __c, const int __lane)
21284 {
21285   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21286 }
21287
21288 /* vmlsq_lane  */
21289
21290 __extension__ extern __inline float32x4_t
21291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21292 vmlsq_lane_f32 (float32x4_t __a, float32x4_t __b,
21293                 float32x2_t __c, const int __lane)
21294 {
21295   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21296 }
21297
21298 __extension__ extern __inline int16x8_t
21299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21300 vmlsq_lane_s16 (int16x8_t __a, int16x8_t __b,
21301                 int16x4_t __c, const int __lane)
21302 {
21303   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21304 }
21305
21306 __extension__ extern __inline int32x4_t
21307 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21308 vmlsq_lane_s32 (int32x4_t __a, int32x4_t __b,
21309                 int32x2_t __c, const int __lane)
21310 {
21311   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21312 }
21313
21314 __extension__ extern __inline uint16x8_t
21315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21316 vmlsq_lane_u16 (uint16x8_t __a, uint16x8_t __b,
21317                 uint16x4_t __c, const int __lane)
21318 {
21319   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21320 }
21321
21322 __extension__ extern __inline uint32x4_t
21323 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21324 vmlsq_lane_u32 (uint32x4_t __a, uint32x4_t __b,
21325                 uint32x2_t __c, const int __lane)
21326 {
21327   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21328 }
21329
21330   /* vmlsq_laneq  */
21331
21332 __extension__ extern __inline float32x4_t
21333 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21334 vmlsq_laneq_f32 (float32x4_t __a, float32x4_t __b,
21335                 float32x4_t __c, const int __lane)
21336 {
21337   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21338 }
21339
21340 __extension__ extern __inline int16x8_t
21341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21342 vmlsq_laneq_s16 (int16x8_t __a, int16x8_t __b,
21343                 int16x8_t __c, const int __lane)
21344 {
21345   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21346 }
21347
21348 __extension__ extern __inline int32x4_t
21349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21350 vmlsq_laneq_s32 (int32x4_t __a, int32x4_t __b,
21351                 int32x4_t __c, const int __lane)
21352 {
21353   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21354 }
21355 __extension__ extern __inline uint16x8_t
21356 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21357 vmlsq_laneq_u16 (uint16x8_t __a, uint16x8_t __b,
21358                 uint16x8_t __c, const int __lane)
21359 {
21360   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21361 }
21362
21363 __extension__ extern __inline uint32x4_t
21364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21365 vmlsq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
21366                 uint32x4_t __c, const int __lane)
21367 {
21368   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21369 }
21370
21371 /* vmov_n_  */
21372
21373 __extension__ extern __inline float16x4_t
21374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21375 vmov_n_f16 (float16_t __a)
21376 {
21377   return vdup_n_f16 (__a);
21378 }
21379
21380 __extension__ extern __inline float32x2_t
21381 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21382 vmov_n_f32 (float32_t __a)
21383 {
21384   return vdup_n_f32 (__a);
21385 }
21386
21387 __extension__ extern __inline float64x1_t
21388 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21389 vmov_n_f64 (float64_t __a)
21390 {
21391   return (float64x1_t) {__a};
21392 }
21393
21394 __extension__ extern __inline poly8x8_t
21395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21396 vmov_n_p8 (poly8_t __a)
21397 {
21398   return vdup_n_p8 (__a);
21399 }
21400
21401 __extension__ extern __inline poly16x4_t
21402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21403 vmov_n_p16 (poly16_t __a)
21404 {
21405   return vdup_n_p16 (__a);
21406 }
21407
21408 __extension__ extern __inline poly64x1_t
21409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21410 vmov_n_p64 (poly64_t __a)
21411 {
21412   return vdup_n_p64 (__a);
21413 }
21414
21415 __extension__ extern __inline int8x8_t
21416 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21417 vmov_n_s8 (int8_t __a)
21418 {
21419   return vdup_n_s8 (__a);
21420 }
21421
21422 __extension__ extern __inline int16x4_t
21423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21424 vmov_n_s16 (int16_t __a)
21425 {
21426   return vdup_n_s16 (__a);
21427 }
21428
21429 __extension__ extern __inline int32x2_t
21430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21431 vmov_n_s32 (int32_t __a)
21432 {
21433   return vdup_n_s32 (__a);
21434 }
21435
21436 __extension__ extern __inline int64x1_t
21437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21438 vmov_n_s64 (int64_t __a)
21439 {
21440   return (int64x1_t) {__a};
21441 }
21442
21443 __extension__ extern __inline uint8x8_t
21444 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21445 vmov_n_u8 (uint8_t __a)
21446 {
21447   return vdup_n_u8 (__a);
21448 }
21449
21450 __extension__ extern __inline uint16x4_t
21451 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21452 vmov_n_u16 (uint16_t __a)
21453 {
21454     return vdup_n_u16 (__a);
21455 }
21456
21457 __extension__ extern __inline uint32x2_t
21458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21459 vmov_n_u32 (uint32_t __a)
21460 {
21461    return vdup_n_u32 (__a);
21462 }
21463
21464 __extension__ extern __inline uint64x1_t
21465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21466 vmov_n_u64 (uint64_t __a)
21467 {
21468   return (uint64x1_t) {__a};
21469 }
21470
21471 __extension__ extern __inline float16x8_t
21472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21473 vmovq_n_f16 (float16_t __a)
21474 {
21475   return vdupq_n_f16 (__a);
21476 }
21477
21478 __extension__ extern __inline float32x4_t
21479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21480 vmovq_n_f32 (float32_t __a)
21481 {
21482   return vdupq_n_f32 (__a);
21483 }
21484
21485 __extension__ extern __inline float64x2_t
21486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21487 vmovq_n_f64 (float64_t __a)
21488 {
21489   return vdupq_n_f64 (__a);
21490 }
21491
21492 __extension__ extern __inline poly8x16_t
21493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21494 vmovq_n_p8 (poly8_t __a)
21495 {
21496   return vdupq_n_p8 (__a);
21497 }
21498
21499 __extension__ extern __inline poly16x8_t
21500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21501 vmovq_n_p16 (poly16_t __a)
21502 {
21503   return vdupq_n_p16 (__a);
21504 }
21505
21506 __extension__ extern __inline poly64x2_t
21507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21508 vmovq_n_p64 (poly64_t __a)
21509 {
21510   return vdupq_n_p64 (__a);
21511 }
21512
21513 __extension__ extern __inline int8x16_t
21514 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21515 vmovq_n_s8 (int8_t __a)
21516 {
21517   return vdupq_n_s8 (__a);
21518 }
21519
21520 __extension__ extern __inline int16x8_t
21521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21522 vmovq_n_s16 (int16_t __a)
21523 {
21524   return vdupq_n_s16 (__a);
21525 }
21526
21527 __extension__ extern __inline int32x4_t
21528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21529 vmovq_n_s32 (int32_t __a)
21530 {
21531   return vdupq_n_s32 (__a);
21532 }
21533
21534 __extension__ extern __inline int64x2_t
21535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21536 vmovq_n_s64 (int64_t __a)
21537 {
21538   return vdupq_n_s64 (__a);
21539 }
21540
21541 __extension__ extern __inline uint8x16_t
21542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21543 vmovq_n_u8 (uint8_t __a)
21544 {
21545   return vdupq_n_u8 (__a);
21546 }
21547
21548 __extension__ extern __inline uint16x8_t
21549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21550 vmovq_n_u16 (uint16_t __a)
21551 {
21552   return vdupq_n_u16 (__a);
21553 }
21554
21555 __extension__ extern __inline uint32x4_t
21556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21557 vmovq_n_u32 (uint32_t __a)
21558 {
21559   return vdupq_n_u32 (__a);
21560 }
21561
21562 __extension__ extern __inline uint64x2_t
21563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21564 vmovq_n_u64 (uint64_t __a)
21565 {
21566   return vdupq_n_u64 (__a);
21567 }
21568
21569 /* vmul_lane  */
21570
21571 __extension__ extern __inline float32x2_t
21572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21573 vmul_lane_f32 (float32x2_t __a, float32x2_t __b, const int __lane)
21574 {
21575   return __a * __aarch64_vget_lane_any (__b, __lane);
21576 }
21577
21578 __extension__ extern __inline float64x1_t
21579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21580 vmul_lane_f64 (float64x1_t __a, float64x1_t __b, const int __lane)
21581 {
21582   return __a * __b;
21583 }
21584
21585 __extension__ extern __inline int16x4_t
21586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21587 vmul_lane_s16 (int16x4_t __a, int16x4_t __b, const int __lane)
21588 {
21589   return __a * __aarch64_vget_lane_any (__b, __lane);
21590 }
21591
21592 __extension__ extern __inline int32x2_t
21593 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21594 vmul_lane_s32 (int32x2_t __a, int32x2_t __b, const int __lane)
21595 {
21596   return __a * __aarch64_vget_lane_any (__b, __lane);
21597 }
21598
21599 __extension__ extern __inline uint16x4_t
21600 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21601 vmul_lane_u16 (uint16x4_t __a, uint16x4_t __b, const int __lane)
21602 {
21603   return __a * __aarch64_vget_lane_any (__b, __lane);
21604 }
21605
21606 __extension__ extern __inline uint32x2_t
21607 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21608 vmul_lane_u32 (uint32x2_t __a, uint32x2_t __b, const int __lane)
21609 {
21610   return __a * __aarch64_vget_lane_any (__b, __lane);
21611 }
21612
21613 /* vmuld_lane  */
21614
21615 __extension__ extern __inline float64_t
21616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21617 vmuld_lane_f64 (float64_t __a, float64x1_t __b, const int __lane)
21618 {
21619   return __a * __aarch64_vget_lane_any (__b, __lane);
21620 }
21621
21622 __extension__ extern __inline float64_t
21623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21624 vmuld_laneq_f64 (float64_t __a, float64x2_t __b, const int __lane)
21625 {
21626   return __a * __aarch64_vget_lane_any (__b, __lane);
21627 }
21628
21629 /* vmuls_lane  */
21630
21631 __extension__ extern __inline float32_t
21632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21633 vmuls_lane_f32 (float32_t __a, float32x2_t __b, const int __lane)
21634 {
21635   return __a * __aarch64_vget_lane_any (__b, __lane);
21636 }
21637
21638 __extension__ extern __inline float32_t
21639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21640 vmuls_laneq_f32 (float32_t __a, float32x4_t __b, const int __lane)
21641 {
21642   return __a * __aarch64_vget_lane_any (__b, __lane);
21643 }
21644
21645 /* vmul_laneq  */
21646
21647 __extension__ extern __inline float32x2_t
21648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21649 vmul_laneq_f32 (float32x2_t __a, float32x4_t __b, const int __lane)
21650 {
21651   return __a * __aarch64_vget_lane_any (__b, __lane);
21652 }
21653
21654 __extension__ extern __inline float64x1_t
21655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21656 vmul_laneq_f64 (float64x1_t __a, float64x2_t __b, const int __lane)
21657 {
21658   return __a * __aarch64_vget_lane_any (__b, __lane);
21659 }
21660
21661 __extension__ extern __inline int16x4_t
21662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21663 vmul_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __lane)
21664 {
21665   return __a * __aarch64_vget_lane_any (__b, __lane);
21666 }
21667
21668 __extension__ extern __inline int32x2_t
21669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21670 vmul_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __lane)
21671 {
21672   return __a * __aarch64_vget_lane_any (__b, __lane);
21673 }
21674
21675 __extension__ extern __inline uint16x4_t
21676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21677 vmul_laneq_u16 (uint16x4_t __a, uint16x8_t __b, const int __lane)
21678 {
21679   return __a * __aarch64_vget_lane_any (__b, __lane);
21680 }
21681
21682 __extension__ extern __inline uint32x2_t
21683 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21684 vmul_laneq_u32 (uint32x2_t __a, uint32x4_t __b, const int __lane)
21685 {
21686   return __a * __aarch64_vget_lane_any (__b, __lane);
21687 }
21688
21689 /* vmul_n  */
21690
21691 __extension__ extern __inline float64x1_t
21692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21693 vmul_n_f64  (float64x1_t __a, float64_t __b)
21694 {
21695   return (float64x1_t) { vget_lane_f64 (__a, 0) * __b };
21696 }
21697
21698 /* vmulq_lane  */
21699
21700 __extension__ extern __inline float32x4_t
21701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21702 vmulq_lane_f32 (float32x4_t __a, float32x2_t __b, const int __lane)
21703 {
21704   return __a * __aarch64_vget_lane_any (__b, __lane);
21705 }
21706
21707 __extension__ extern __inline float64x2_t
21708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21709 vmulq_lane_f64 (float64x2_t __a, float64x1_t __b, const int __lane)
21710 {
21711   __AARCH64_LANE_CHECK (__a, __lane);
21712   return __a * __b[0];
21713 }
21714
21715 __extension__ extern __inline int16x8_t
21716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21717 vmulq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __lane)
21718 {
21719   return __a * __aarch64_vget_lane_any (__b, __lane);
21720 }
21721
21722 __extension__ extern __inline int32x4_t
21723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21724 vmulq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __lane)
21725 {
21726   return __a * __aarch64_vget_lane_any (__b, __lane);
21727 }
21728
21729 __extension__ extern __inline uint16x8_t
21730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21731 vmulq_lane_u16 (uint16x8_t __a, uint16x4_t __b, const int __lane)
21732 {
21733   return __a * __aarch64_vget_lane_any (__b, __lane);
21734 }
21735
21736 __extension__ extern __inline uint32x4_t
21737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21738 vmulq_lane_u32 (uint32x4_t __a, uint32x2_t __b, const int __lane)
21739 {
21740   return __a * __aarch64_vget_lane_any (__b, __lane);
21741 }
21742
21743 /* vmulq_laneq  */
21744
21745 __extension__ extern __inline float32x4_t
21746 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21747 vmulq_laneq_f32 (float32x4_t __a, float32x4_t __b, const int __lane)
21748 {
21749   return __a * __aarch64_vget_lane_any (__b, __lane);
21750 }
21751
21752 __extension__ extern __inline float64x2_t
21753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21754 vmulq_laneq_f64 (float64x2_t __a, float64x2_t __b, const int __lane)
21755 {
21756   return __a * __aarch64_vget_lane_any (__b, __lane);
21757 }
21758
21759 __extension__ extern __inline int16x8_t
21760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21761 vmulq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __lane)
21762 {
21763   return __a * __aarch64_vget_lane_any (__b, __lane);
21764 }
21765
21766 __extension__ extern __inline int32x4_t
21767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21768 vmulq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __lane)
21769 {
21770   return __a * __aarch64_vget_lane_any (__b, __lane);
21771 }
21772
21773 __extension__ extern __inline uint16x8_t
21774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21775 vmulq_laneq_u16 (uint16x8_t __a, uint16x8_t __b, const int __lane)
21776 {
21777   return __a * __aarch64_vget_lane_any (__b, __lane);
21778 }
21779
21780 __extension__ extern __inline uint32x4_t
21781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21782 vmulq_laneq_u32 (uint32x4_t __a, uint32x4_t __b, const int __lane)
21783 {
21784   return __a * __aarch64_vget_lane_any (__b, __lane);
21785 }
21786
21787 /* vmul_n.  */
21788
21789 __extension__ extern __inline float32x2_t
21790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21791 vmul_n_f32 (float32x2_t __a, float32_t __b)
21792 {
21793   return __a * __b;
21794 }
21795
21796 __extension__ extern __inline float32x4_t
21797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21798 vmulq_n_f32 (float32x4_t __a, float32_t __b)
21799 {
21800   return __a * __b;
21801 }
21802
21803 __extension__ extern __inline float64x2_t
21804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21805 vmulq_n_f64 (float64x2_t __a, float64_t __b)
21806 {
21807   return __a * __b;
21808 }
21809
21810 __extension__ extern __inline int16x4_t
21811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21812 vmul_n_s16 (int16x4_t __a, int16_t __b)
21813 {
21814   return __a * __b;
21815 }
21816
21817 __extension__ extern __inline int16x8_t
21818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21819 vmulq_n_s16 (int16x8_t __a, int16_t __b)
21820 {
21821   return __a * __b;
21822 }
21823
21824 __extension__ extern __inline int32x2_t
21825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21826 vmul_n_s32 (int32x2_t __a, int32_t __b)
21827 {
21828   return __a * __b;
21829 }
21830
21831 __extension__ extern __inline int32x4_t
21832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21833 vmulq_n_s32 (int32x4_t __a, int32_t __b)
21834 {
21835   return __a * __b;
21836 }
21837
21838 __extension__ extern __inline uint16x4_t
21839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21840 vmul_n_u16 (uint16x4_t __a, uint16_t __b)
21841 {
21842   return __a * __b;
21843 }
21844
21845 __extension__ extern __inline uint16x8_t
21846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21847 vmulq_n_u16 (uint16x8_t __a, uint16_t __b)
21848 {
21849   return __a * __b;
21850 }
21851
21852 __extension__ extern __inline uint32x2_t
21853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21854 vmul_n_u32 (uint32x2_t __a, uint32_t __b)
21855 {
21856   return __a * __b;
21857 }
21858
21859 __extension__ extern __inline uint32x4_t
21860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21861 vmulq_n_u32 (uint32x4_t __a, uint32_t __b)
21862 {
21863   return __a * __b;
21864 }
21865
21866 /* vmvn  */
21867
21868 __extension__ extern __inline poly8x8_t
21869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21870 vmvn_p8 (poly8x8_t __a)
21871 {
21872   return (poly8x8_t) ~((int8x8_t) __a);
21873 }
21874
21875 __extension__ extern __inline int8x8_t
21876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21877 vmvn_s8 (int8x8_t __a)
21878 {
21879   return ~__a;
21880 }
21881
21882 __extension__ extern __inline int16x4_t
21883 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21884 vmvn_s16 (int16x4_t __a)
21885 {
21886   return ~__a;
21887 }
21888
21889 __extension__ extern __inline int32x2_t
21890 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21891 vmvn_s32 (int32x2_t __a)
21892 {
21893   return ~__a;
21894 }
21895
21896 __extension__ extern __inline uint8x8_t
21897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21898 vmvn_u8 (uint8x8_t __a)
21899 {
21900   return ~__a;
21901 }
21902
21903 __extension__ extern __inline uint16x4_t
21904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21905 vmvn_u16 (uint16x4_t __a)
21906 {
21907   return ~__a;
21908 }
21909
21910 __extension__ extern __inline uint32x2_t
21911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21912 vmvn_u32 (uint32x2_t __a)
21913 {
21914   return ~__a;
21915 }
21916
21917 __extension__ extern __inline poly8x16_t
21918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21919 vmvnq_p8 (poly8x16_t __a)
21920 {
21921   return (poly8x16_t) ~((int8x16_t) __a);
21922 }
21923
21924 __extension__ extern __inline int8x16_t
21925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21926 vmvnq_s8 (int8x16_t __a)
21927 {
21928   return ~__a;
21929 }
21930
21931 __extension__ extern __inline int16x8_t
21932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21933 vmvnq_s16 (int16x8_t __a)
21934 {
21935   return ~__a;
21936 }
21937
21938 __extension__ extern __inline int32x4_t
21939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21940 vmvnq_s32 (int32x4_t __a)
21941 {
21942   return ~__a;
21943 }
21944
21945 __extension__ extern __inline uint8x16_t
21946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21947 vmvnq_u8 (uint8x16_t __a)
21948 {
21949   return ~__a;
21950 }
21951
21952 __extension__ extern __inline uint16x8_t
21953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21954 vmvnq_u16 (uint16x8_t __a)
21955 {
21956   return ~__a;
21957 }
21958
21959 __extension__ extern __inline uint32x4_t
21960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21961 vmvnq_u32 (uint32x4_t __a)
21962 {
21963   return ~__a;
21964 }
21965
21966 /* vneg  */
21967
21968 __extension__ extern __inline float32x2_t
21969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21970 vneg_f32 (float32x2_t __a)
21971 {
21972   return -__a;
21973 }
21974
21975 __extension__ extern __inline float64x1_t
21976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21977 vneg_f64 (float64x1_t __a)
21978 {
21979   return -__a;
21980 }
21981
21982 __extension__ extern __inline int8x8_t
21983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21984 vneg_s8 (int8x8_t __a)
21985 {
21986   return -__a;
21987 }
21988
21989 __extension__ extern __inline int16x4_t
21990 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21991 vneg_s16 (int16x4_t __a)
21992 {
21993   return -__a;
21994 }
21995
21996 __extension__ extern __inline int32x2_t
21997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21998 vneg_s32 (int32x2_t __a)
21999 {
22000   return -__a;
22001 }
22002
22003 __extension__ extern __inline int64x1_t
22004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22005 vneg_s64 (int64x1_t __a)
22006 {
22007   return -__a;
22008 }
22009
22010 __extension__ extern __inline float32x4_t
22011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22012 vnegq_f32 (float32x4_t __a)
22013 {
22014   return -__a;
22015 }
22016
22017 __extension__ extern __inline float64x2_t
22018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22019 vnegq_f64 (float64x2_t __a)
22020 {
22021   return -__a;
22022 }
22023
22024 __extension__ extern __inline int8x16_t
22025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22026 vnegq_s8 (int8x16_t __a)
22027 {
22028   return -__a;
22029 }
22030
22031 __extension__ extern __inline int16x8_t
22032 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22033 vnegq_s16 (int16x8_t __a)
22034 {
22035   return -__a;
22036 }
22037
22038 __extension__ extern __inline int32x4_t
22039 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22040 vnegq_s32 (int32x4_t __a)
22041 {
22042   return -__a;
22043 }
22044
22045 __extension__ extern __inline int64x2_t
22046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22047 vnegq_s64 (int64x2_t __a)
22048 {
22049   return -__a;
22050 }
22051
22052 /* vpadd  */
22053
22054 __extension__ extern __inline float32x2_t
22055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22056 vpadd_f32 (float32x2_t __a, float32x2_t __b)
22057 {
22058   return __builtin_aarch64_faddpv2sf (__a, __b);
22059 }
22060
22061 __extension__ extern __inline float32x4_t
22062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22063 vpaddq_f32 (float32x4_t __a, float32x4_t __b)
22064 {
22065   return __builtin_aarch64_faddpv4sf (__a, __b);
22066 }
22067
22068 __extension__ extern __inline float64x2_t
22069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22070 vpaddq_f64 (float64x2_t __a, float64x2_t __b)
22071 {
22072   return __builtin_aarch64_faddpv2df (__a, __b);
22073 }
22074
22075 __extension__ extern __inline int8x8_t
22076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22077 vpadd_s8 (int8x8_t __a, int8x8_t __b)
22078 {
22079   return __builtin_aarch64_addpv8qi (__a, __b);
22080 }
22081
22082 __extension__ extern __inline int16x4_t
22083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22084 vpadd_s16 (int16x4_t __a, int16x4_t __b)
22085 {
22086   return __builtin_aarch64_addpv4hi (__a, __b);
22087 }
22088
22089 __extension__ extern __inline int32x2_t
22090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22091 vpadd_s32 (int32x2_t __a, int32x2_t __b)
22092 {
22093   return __builtin_aarch64_addpv2si (__a, __b);
22094 }
22095
22096 __extension__ extern __inline uint8x8_t
22097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22098 vpadd_u8 (uint8x8_t __a, uint8x8_t __b)
22099 {
22100   return (uint8x8_t) __builtin_aarch64_addpv8qi ((int8x8_t) __a,
22101                                                  (int8x8_t) __b);
22102 }
22103
22104 __extension__ extern __inline uint16x4_t
22105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22106 vpadd_u16 (uint16x4_t __a, uint16x4_t __b)
22107 {
22108   return (uint16x4_t) __builtin_aarch64_addpv4hi ((int16x4_t) __a,
22109                                                   (int16x4_t) __b);
22110 }
22111
22112 __extension__ extern __inline uint32x2_t
22113 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22114 vpadd_u32 (uint32x2_t __a, uint32x2_t __b)
22115 {
22116   return (uint32x2_t) __builtin_aarch64_addpv2si ((int32x2_t) __a,
22117                                                   (int32x2_t) __b);
22118 }
22119
22120 __extension__ extern __inline float32_t
22121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22122 vpadds_f32 (float32x2_t __a)
22123 {
22124   return __builtin_aarch64_reduc_plus_scal_v2sf (__a);
22125 }
22126
22127 __extension__ extern __inline float64_t
22128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22129 vpaddd_f64 (float64x2_t __a)
22130 {
22131   return __builtin_aarch64_reduc_plus_scal_v2df (__a);
22132 }
22133
22134 __extension__ extern __inline int64_t
22135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22136 vpaddd_s64 (int64x2_t __a)
22137 {
22138   return __builtin_aarch64_addpdi (__a);
22139 }
22140
22141 __extension__ extern __inline uint64_t
22142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22143 vpaddd_u64 (uint64x2_t __a)
22144 {
22145   return __builtin_aarch64_addpdi ((int64x2_t) __a);
22146 }
22147
22148 /* vqabs */
22149
22150 __extension__ extern __inline int64x2_t
22151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22152 vqabsq_s64 (int64x2_t __a)
22153 {
22154   return (int64x2_t) __builtin_aarch64_sqabsv2di (__a);
22155 }
22156
22157 __extension__ extern __inline int8_t
22158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22159 vqabsb_s8 (int8_t __a)
22160 {
22161   return (int8_t) __builtin_aarch64_sqabsqi (__a);
22162 }
22163
22164 __extension__ extern __inline int16_t
22165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22166 vqabsh_s16 (int16_t __a)
22167 {
22168   return (int16_t) __builtin_aarch64_sqabshi (__a);
22169 }
22170
22171 __extension__ extern __inline int32_t
22172 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22173 vqabss_s32 (int32_t __a)
22174 {
22175   return (int32_t) __builtin_aarch64_sqabssi (__a);
22176 }
22177
22178 __extension__ extern __inline int64_t
22179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22180 vqabsd_s64 (int64_t __a)
22181 {
22182   return __builtin_aarch64_sqabsdi (__a);
22183 }
22184
22185 /* vqadd */
22186
22187 __extension__ extern __inline int8_t
22188 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22189 vqaddb_s8 (int8_t __a, int8_t __b)
22190 {
22191   return (int8_t) __builtin_aarch64_sqaddqi (__a, __b);
22192 }
22193
22194 __extension__ extern __inline int16_t
22195 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22196 vqaddh_s16 (int16_t __a, int16_t __b)
22197 {
22198   return (int16_t) __builtin_aarch64_sqaddhi (__a, __b);
22199 }
22200
22201 __extension__ extern __inline int32_t
22202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22203 vqadds_s32 (int32_t __a, int32_t __b)
22204 {
22205   return (int32_t) __builtin_aarch64_sqaddsi (__a, __b);
22206 }
22207
22208 __extension__ extern __inline int64_t
22209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22210 vqaddd_s64 (int64_t __a, int64_t __b)
22211 {
22212   return __builtin_aarch64_sqadddi (__a, __b);
22213 }
22214
22215 __extension__ extern __inline uint8_t
22216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22217 vqaddb_u8 (uint8_t __a, uint8_t __b)
22218 {
22219   return (uint8_t) __builtin_aarch64_uqaddqi_uuu (__a, __b);
22220 }
22221
22222 __extension__ extern __inline uint16_t
22223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22224 vqaddh_u16 (uint16_t __a, uint16_t __b)
22225 {
22226   return (uint16_t) __builtin_aarch64_uqaddhi_uuu (__a, __b);
22227 }
22228
22229 __extension__ extern __inline uint32_t
22230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22231 vqadds_u32 (uint32_t __a, uint32_t __b)
22232 {
22233   return (uint32_t) __builtin_aarch64_uqaddsi_uuu (__a, __b);
22234 }
22235
22236 __extension__ extern __inline uint64_t
22237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22238 vqaddd_u64 (uint64_t __a, uint64_t __b)
22239 {
22240   return __builtin_aarch64_uqadddi_uuu (__a, __b);
22241 }
22242
22243 /* vqdmlal */
22244
22245 __extension__ extern __inline int32x4_t
22246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22247 vqdmlal_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c)
22248 {
22249   return __builtin_aarch64_sqdmlalv4hi (__a, __b, __c);
22250 }
22251
22252 __extension__ extern __inline int32x4_t
22253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22254 vqdmlal_high_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c)
22255 {
22256   return __builtin_aarch64_sqdmlal2v8hi (__a, __b, __c);
22257 }
22258
22259 __extension__ extern __inline int32x4_t
22260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22261 vqdmlal_high_lane_s16 (int32x4_t __a, int16x8_t __b, int16x4_t __c,
22262                        int const __d)
22263 {
22264   return __builtin_aarch64_sqdmlal2_lanev8hi (__a, __b, __c, __d);
22265 }
22266
22267 __extension__ extern __inline int32x4_t
22268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22269 vqdmlal_high_laneq_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c,
22270                         int const __d)
22271 {
22272   return __builtin_aarch64_sqdmlal2_laneqv8hi (__a, __b, __c, __d);
22273 }
22274
22275 __extension__ extern __inline int32x4_t
22276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22277 vqdmlal_high_n_s16 (int32x4_t __a, int16x8_t __b, int16_t __c)
22278 {
22279   return __builtin_aarch64_sqdmlal2_nv8hi (__a, __b, __c);
22280 }
22281
22282 __extension__ extern __inline int32x4_t
22283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22284 vqdmlal_lane_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c, int const __d)
22285 {
22286   return __builtin_aarch64_sqdmlal_lanev4hi (__a, __b, __c, __d);
22287 }
22288
22289 __extension__ extern __inline int32x4_t
22290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22291 vqdmlal_laneq_s16 (int32x4_t __a, int16x4_t __b, int16x8_t __c, int const __d)
22292 {
22293   return __builtin_aarch64_sqdmlal_laneqv4hi (__a, __b, __c, __d);
22294 }
22295
22296 __extension__ extern __inline int32x4_t
22297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22298 vqdmlal_n_s16 (int32x4_t __a, int16x4_t __b, int16_t __c)
22299 {
22300   return __builtin_aarch64_sqdmlal_nv4hi (__a, __b, __c);
22301 }
22302
22303 __extension__ extern __inline int64x2_t
22304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22305 vqdmlal_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c)
22306 {
22307   return __builtin_aarch64_sqdmlalv2si (__a, __b, __c);
22308 }
22309
22310 __extension__ extern __inline int64x2_t
22311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22312 vqdmlal_high_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c)
22313 {
22314   return __builtin_aarch64_sqdmlal2v4si (__a, __b, __c);
22315 }
22316
22317 __extension__ extern __inline int64x2_t
22318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22319 vqdmlal_high_lane_s32 (int64x2_t __a, int32x4_t __b, int32x2_t __c,
22320                        int const __d)
22321 {
22322   return __builtin_aarch64_sqdmlal2_lanev4si (__a, __b, __c, __d);
22323 }
22324
22325 __extension__ extern __inline int64x2_t
22326 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22327 vqdmlal_high_laneq_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c,
22328                         int const __d)
22329 {
22330   return __builtin_aarch64_sqdmlal2_laneqv4si (__a, __b, __c, __d);
22331 }
22332
22333 __extension__ extern __inline int64x2_t
22334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22335 vqdmlal_high_n_s32 (int64x2_t __a, int32x4_t __b, int32_t __c)
22336 {
22337   return __builtin_aarch64_sqdmlal2_nv4si (__a, __b, __c);
22338 }
22339
22340 __extension__ extern __inline int64x2_t
22341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22342 vqdmlal_lane_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c, int const __d)
22343 {
22344   return __builtin_aarch64_sqdmlal_lanev2si (__a, __b, __c, __d);
22345 }
22346
22347 __extension__ extern __inline int64x2_t
22348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22349 vqdmlal_laneq_s32 (int64x2_t __a, int32x2_t __b, int32x4_t __c, int const __d)
22350 {
22351   return __builtin_aarch64_sqdmlal_laneqv2si (__a, __b, __c, __d);
22352 }
22353
22354 __extension__ extern __inline int64x2_t
22355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22356 vqdmlal_n_s32 (int64x2_t __a, int32x2_t __b, int32_t __c)
22357 {
22358   return __builtin_aarch64_sqdmlal_nv2si (__a, __b, __c);
22359 }
22360
22361 __extension__ extern __inline int32_t
22362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22363 vqdmlalh_s16 (int32_t __a, int16_t __b, int16_t __c)
22364 {
22365   return __builtin_aarch64_sqdmlalhi (__a, __b, __c);
22366 }
22367
22368 __extension__ extern __inline int32_t
22369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22370 vqdmlalh_lane_s16 (int32_t __a, int16_t __b, int16x4_t __c, const int __d)
22371 {
22372   return __builtin_aarch64_sqdmlal_lanehi (__a, __b, __c, __d);
22373 }
22374
22375 __extension__ extern __inline int32_t
22376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22377 vqdmlalh_laneq_s16 (int32_t __a, int16_t __b, int16x8_t __c, const int __d)
22378 {
22379   return __builtin_aarch64_sqdmlal_laneqhi (__a, __b, __c, __d);
22380 }
22381
22382 __extension__ extern __inline int64_t
22383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22384 vqdmlals_s32 (int64_t __a, int32_t __b, int32_t __c)
22385 {
22386   return __builtin_aarch64_sqdmlalsi (__a, __b, __c);
22387 }
22388
22389 __extension__ extern __inline int64_t
22390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22391 vqdmlals_lane_s32 (int64_t __a, int32_t __b, int32x2_t __c, const int __d)
22392 {
22393   return __builtin_aarch64_sqdmlal_lanesi (__a, __b, __c, __d);
22394 }
22395
22396 __extension__ extern __inline int64_t
22397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22398 vqdmlals_laneq_s32 (int64_t __a, int32_t __b, int32x4_t __c, const int __d)
22399 {
22400   return __builtin_aarch64_sqdmlal_laneqsi (__a, __b, __c, __d);
22401 }
22402
22403 /* vqdmlsl */
22404
22405 __extension__ extern __inline int32x4_t
22406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22407 vqdmlsl_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c)
22408 {
22409   return __builtin_aarch64_sqdmlslv4hi (__a, __b, __c);
22410 }
22411
22412 __extension__ extern __inline int32x4_t
22413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22414 vqdmlsl_high_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c)
22415 {
22416   return __builtin_aarch64_sqdmlsl2v8hi (__a, __b, __c);
22417 }
22418
22419 __extension__ extern __inline int32x4_t
22420 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22421 vqdmlsl_high_lane_s16 (int32x4_t __a, int16x8_t __b, int16x4_t __c,
22422                        int const __d)
22423 {
22424   return __builtin_aarch64_sqdmlsl2_lanev8hi (__a, __b, __c, __d);
22425 }
22426
22427 __extension__ extern __inline int32x4_t
22428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22429 vqdmlsl_high_laneq_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c,
22430                         int const __d)
22431 {
22432   return __builtin_aarch64_sqdmlsl2_laneqv8hi (__a, __b, __c, __d);
22433 }
22434
22435 __extension__ extern __inline int32x4_t
22436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22437 vqdmlsl_high_n_s16 (int32x4_t __a, int16x8_t __b, int16_t __c)
22438 {
22439   return __builtin_aarch64_sqdmlsl2_nv8hi (__a, __b, __c);
22440 }
22441
22442 __extension__ extern __inline int32x4_t
22443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22444 vqdmlsl_lane_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c, int const __d)
22445 {
22446   return __builtin_aarch64_sqdmlsl_lanev4hi (__a, __b, __c, __d);
22447 }
22448
22449 __extension__ extern __inline int32x4_t
22450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22451 vqdmlsl_laneq_s16 (int32x4_t __a, int16x4_t __b, int16x8_t __c, int const __d)
22452 {
22453   return __builtin_aarch64_sqdmlsl_laneqv4hi (__a, __b, __c, __d);
22454 }
22455
22456 __extension__ extern __inline int32x4_t
22457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22458 vqdmlsl_n_s16 (int32x4_t __a, int16x4_t __b, int16_t __c)
22459 {
22460   return __builtin_aarch64_sqdmlsl_nv4hi (__a, __b, __c);
22461 }
22462
22463 __extension__ extern __inline int64x2_t
22464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22465 vqdmlsl_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c)
22466 {
22467   return __builtin_aarch64_sqdmlslv2si (__a, __b, __c);
22468 }
22469
22470 __extension__ extern __inline int64x2_t
22471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22472 vqdmlsl_high_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c)
22473 {
22474   return __builtin_aarch64_sqdmlsl2v4si (__a, __b, __c);
22475 }
22476
22477 __extension__ extern __inline int64x2_t
22478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22479 vqdmlsl_high_lane_s32 (int64x2_t __a, int32x4_t __b, int32x2_t __c,
22480                        int const __d)
22481 {
22482   return __builtin_aarch64_sqdmlsl2_lanev4si (__a, __b, __c, __d);
22483 }
22484
22485 __extension__ extern __inline int64x2_t
22486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22487 vqdmlsl_high_laneq_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c,
22488                         int const __d)
22489 {
22490   return __builtin_aarch64_sqdmlsl2_laneqv4si (__a, __b, __c, __d);
22491 }
22492
22493 __extension__ extern __inline int64x2_t
22494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22495 vqdmlsl_high_n_s32 (int64x2_t __a, int32x4_t __b, int32_t __c)
22496 {
22497   return __builtin_aarch64_sqdmlsl2_nv4si (__a, __b, __c);
22498 }
22499
22500 __extension__ extern __inline int64x2_t
22501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22502 vqdmlsl_lane_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c, int const __d)
22503 {
22504   return __builtin_aarch64_sqdmlsl_lanev2si (__a, __b, __c, __d);
22505 }
22506
22507 __extension__ extern __inline int64x2_t
22508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22509 vqdmlsl_laneq_s32 (int64x2_t __a, int32x2_t __b, int32x4_t __c, int const __d)
22510 {
22511   return __builtin_aarch64_sqdmlsl_laneqv2si (__a, __b, __c, __d);
22512 }
22513
22514 __extension__ extern __inline int64x2_t
22515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22516 vqdmlsl_n_s32 (int64x2_t __a, int32x2_t __b, int32_t __c)
22517 {
22518   return __builtin_aarch64_sqdmlsl_nv2si (__a, __b, __c);
22519 }
22520
22521 __extension__ extern __inline int32_t
22522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22523 vqdmlslh_s16 (int32_t __a, int16_t __b, int16_t __c)
22524 {
22525   return __builtin_aarch64_sqdmlslhi (__a, __b, __c);
22526 }
22527
22528 __extension__ extern __inline int32_t
22529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22530 vqdmlslh_lane_s16 (int32_t __a, int16_t __b, int16x4_t __c, const int __d)
22531 {
22532   return __builtin_aarch64_sqdmlsl_lanehi (__a, __b, __c, __d);
22533 }
22534
22535 __extension__ extern __inline int32_t
22536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22537 vqdmlslh_laneq_s16 (int32_t __a, int16_t __b, int16x8_t __c, const int __d)
22538 {
22539   return __builtin_aarch64_sqdmlsl_laneqhi (__a, __b, __c, __d);
22540 }
22541
22542 __extension__ extern __inline int64_t
22543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22544 vqdmlsls_s32 (int64_t __a, int32_t __b, int32_t __c)
22545 {
22546   return __builtin_aarch64_sqdmlslsi (__a, __b, __c);
22547 }
22548
22549 __extension__ extern __inline int64_t
22550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22551 vqdmlsls_lane_s32 (int64_t __a, int32_t __b, int32x2_t __c, const int __d)
22552 {
22553   return __builtin_aarch64_sqdmlsl_lanesi (__a, __b, __c, __d);
22554 }
22555
22556 __extension__ extern __inline int64_t
22557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22558 vqdmlsls_laneq_s32 (int64_t __a, int32_t __b, int32x4_t __c, const int __d)
22559 {
22560   return __builtin_aarch64_sqdmlsl_laneqsi (__a, __b, __c, __d);
22561 }
22562
22563 /* vqdmulh */
22564
22565 __extension__ extern __inline int16x4_t
22566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22567 vqdmulh_lane_s16 (int16x4_t __a, int16x4_t __b, const int __c)
22568 {
22569   return __builtin_aarch64_sqdmulh_lanev4hi (__a, __b, __c);
22570 }
22571
22572 __extension__ extern __inline int32x2_t
22573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22574 vqdmulh_lane_s32 (int32x2_t __a, int32x2_t __b, const int __c)
22575 {
22576   return __builtin_aarch64_sqdmulh_lanev2si (__a, __b, __c);
22577 }
22578
22579 __extension__ extern __inline int16x8_t
22580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22581 vqdmulhq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __c)
22582 {
22583   return __builtin_aarch64_sqdmulh_lanev8hi (__a, __b, __c);
22584 }
22585
22586 __extension__ extern __inline int32x4_t
22587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22588 vqdmulhq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __c)
22589 {
22590   return __builtin_aarch64_sqdmulh_lanev4si (__a, __b, __c);
22591 }
22592
22593 __extension__ extern __inline int16_t
22594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22595 vqdmulhh_s16 (int16_t __a, int16_t __b)
22596 {
22597   return (int16_t) __builtin_aarch64_sqdmulhhi (__a, __b);
22598 }
22599
22600 __extension__ extern __inline int16_t
22601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22602 vqdmulhh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
22603 {
22604   return __builtin_aarch64_sqdmulh_lanehi (__a, __b, __c);
22605 }
22606
22607 __extension__ extern __inline int16_t
22608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22609 vqdmulhh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
22610 {
22611   return __builtin_aarch64_sqdmulh_laneqhi (__a, __b, __c);
22612 }
22613
22614 __extension__ extern __inline int32_t
22615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22616 vqdmulhs_s32 (int32_t __a, int32_t __b)
22617 {
22618   return (int32_t) __builtin_aarch64_sqdmulhsi (__a, __b);
22619 }
22620
22621 __extension__ extern __inline int32_t
22622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22623 vqdmulhs_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
22624 {
22625   return __builtin_aarch64_sqdmulh_lanesi (__a, __b, __c);
22626 }
22627
22628 __extension__ extern __inline int32_t
22629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22630 vqdmulhs_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
22631 {
22632   return __builtin_aarch64_sqdmulh_laneqsi (__a, __b, __c);
22633 }
22634
22635 /* vqdmull */
22636
22637 __extension__ extern __inline int32x4_t
22638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22639 vqdmull_s16 (int16x4_t __a, int16x4_t __b)
22640 {
22641   return __builtin_aarch64_sqdmullv4hi (__a, __b);
22642 }
22643
22644 __extension__ extern __inline int32x4_t
22645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22646 vqdmull_high_s16 (int16x8_t __a, int16x8_t __b)
22647 {
22648   return __builtin_aarch64_sqdmull2v8hi (__a, __b);
22649 }
22650
22651 __extension__ extern __inline int32x4_t
22652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22653 vqdmull_high_lane_s16 (int16x8_t __a, int16x4_t __b, int const __c)
22654 {
22655   return __builtin_aarch64_sqdmull2_lanev8hi (__a, __b,__c);
22656 }
22657
22658 __extension__ extern __inline int32x4_t
22659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22660 vqdmull_high_laneq_s16 (int16x8_t __a, int16x8_t __b, int const __c)
22661 {
22662   return __builtin_aarch64_sqdmull2_laneqv8hi (__a, __b,__c);
22663 }
22664
22665 __extension__ extern __inline int32x4_t
22666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22667 vqdmull_high_n_s16 (int16x8_t __a, int16_t __b)
22668 {
22669   return __builtin_aarch64_sqdmull2_nv8hi (__a, __b);
22670 }
22671
22672 __extension__ extern __inline int32x4_t
22673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22674 vqdmull_lane_s16 (int16x4_t __a, int16x4_t __b, int const __c)
22675 {
22676   return __builtin_aarch64_sqdmull_lanev4hi (__a, __b, __c);
22677 }
22678
22679 __extension__ extern __inline int32x4_t
22680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22681 vqdmull_laneq_s16 (int16x4_t __a, int16x8_t __b, int const __c)
22682 {
22683   return __builtin_aarch64_sqdmull_laneqv4hi (__a, __b, __c);
22684 }
22685
22686 __extension__ extern __inline int32x4_t
22687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22688 vqdmull_n_s16 (int16x4_t __a, int16_t __b)
22689 {
22690   return __builtin_aarch64_sqdmull_nv4hi (__a, __b);
22691 }
22692
22693 __extension__ extern __inline int64x2_t
22694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22695 vqdmull_s32 (int32x2_t __a, int32x2_t __b)
22696 {
22697   return __builtin_aarch64_sqdmullv2si (__a, __b);
22698 }
22699
22700 __extension__ extern __inline int64x2_t
22701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22702 vqdmull_high_s32 (int32x4_t __a, int32x4_t __b)
22703 {
22704   return __builtin_aarch64_sqdmull2v4si (__a, __b);
22705 }
22706
22707 __extension__ extern __inline int64x2_t
22708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22709 vqdmull_high_lane_s32 (int32x4_t __a, int32x2_t __b, int const __c)
22710 {
22711   return __builtin_aarch64_sqdmull2_lanev4si (__a, __b, __c);
22712 }
22713
22714 __extension__ extern __inline int64x2_t
22715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22716 vqdmull_high_laneq_s32 (int32x4_t __a, int32x4_t __b, int const __c)
22717 {
22718   return __builtin_aarch64_sqdmull2_laneqv4si (__a, __b, __c);
22719 }
22720
22721 __extension__ extern __inline int64x2_t
22722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22723 vqdmull_high_n_s32 (int32x4_t __a, int32_t __b)
22724 {
22725   return __builtin_aarch64_sqdmull2_nv4si (__a, __b);
22726 }
22727
22728 __extension__ extern __inline int64x2_t
22729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22730 vqdmull_lane_s32 (int32x2_t __a, int32x2_t __b, int const __c)
22731 {
22732   return __builtin_aarch64_sqdmull_lanev2si (__a, __b, __c);
22733 }
22734
22735 __extension__ extern __inline int64x2_t
22736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22737 vqdmull_laneq_s32 (int32x2_t __a, int32x4_t __b, int const __c)
22738 {
22739   return __builtin_aarch64_sqdmull_laneqv2si (__a, __b, __c);
22740 }
22741
22742 __extension__ extern __inline int64x2_t
22743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22744 vqdmull_n_s32 (int32x2_t __a, int32_t __b)
22745 {
22746   return __builtin_aarch64_sqdmull_nv2si (__a, __b);
22747 }
22748
22749 __extension__ extern __inline int32_t
22750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22751 vqdmullh_s16 (int16_t __a, int16_t __b)
22752 {
22753   return (int32_t) __builtin_aarch64_sqdmullhi (__a, __b);
22754 }
22755
22756 __extension__ extern __inline int32_t
22757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22758 vqdmullh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
22759 {
22760   return __builtin_aarch64_sqdmull_lanehi (__a, __b, __c);
22761 }
22762
22763 __extension__ extern __inline int32_t
22764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22765 vqdmullh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
22766 {
22767   return __builtin_aarch64_sqdmull_laneqhi (__a, __b, __c);
22768 }
22769
22770 __extension__ extern __inline int64_t
22771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22772 vqdmulls_s32 (int32_t __a, int32_t __b)
22773 {
22774   return __builtin_aarch64_sqdmullsi (__a, __b);
22775 }
22776
22777 __extension__ extern __inline int64_t
22778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22779 vqdmulls_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
22780 {
22781   return __builtin_aarch64_sqdmull_lanesi (__a, __b, __c);
22782 }
22783
22784 __extension__ extern __inline int64_t
22785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22786 vqdmulls_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
22787 {
22788   return __builtin_aarch64_sqdmull_laneqsi (__a, __b, __c);
22789 }
22790
22791 /* vqmovn */
22792
22793 __extension__ extern __inline int8x8_t
22794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22795 vqmovn_s16 (int16x8_t __a)
22796 {
22797   return (int8x8_t) __builtin_aarch64_sqmovnv8hi (__a);
22798 }
22799
22800 __extension__ extern __inline int16x4_t
22801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22802 vqmovn_s32 (int32x4_t __a)
22803 {
22804   return (int16x4_t) __builtin_aarch64_sqmovnv4si (__a);
22805 }
22806
22807 __extension__ extern __inline int32x2_t
22808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22809 vqmovn_s64 (int64x2_t __a)
22810 {
22811   return (int32x2_t) __builtin_aarch64_sqmovnv2di (__a);
22812 }
22813
22814 __extension__ extern __inline uint8x8_t
22815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22816 vqmovn_u16 (uint16x8_t __a)
22817 {
22818   return (uint8x8_t) __builtin_aarch64_uqmovnv8hi ((int16x8_t) __a);
22819 }
22820
22821 __extension__ extern __inline uint16x4_t
22822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22823 vqmovn_u32 (uint32x4_t __a)
22824 {
22825   return (uint16x4_t) __builtin_aarch64_uqmovnv4si ((int32x4_t) __a);
22826 }
22827
22828 __extension__ extern __inline uint32x2_t
22829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22830 vqmovn_u64 (uint64x2_t __a)
22831 {
22832   return (uint32x2_t) __builtin_aarch64_uqmovnv2di ((int64x2_t) __a);
22833 }
22834
22835 __extension__ extern __inline int8_t
22836 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22837 vqmovnh_s16 (int16_t __a)
22838 {
22839   return (int8_t) __builtin_aarch64_sqmovnhi (__a);
22840 }
22841
22842 __extension__ extern __inline int16_t
22843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22844 vqmovns_s32 (int32_t __a)
22845 {
22846   return (int16_t) __builtin_aarch64_sqmovnsi (__a);
22847 }
22848
22849 __extension__ extern __inline int32_t
22850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22851 vqmovnd_s64 (int64_t __a)
22852 {
22853   return (int32_t) __builtin_aarch64_sqmovndi (__a);
22854 }
22855
22856 __extension__ extern __inline uint8_t
22857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22858 vqmovnh_u16 (uint16_t __a)
22859 {
22860   return (uint8_t) __builtin_aarch64_uqmovnhi (__a);
22861 }
22862
22863 __extension__ extern __inline uint16_t
22864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22865 vqmovns_u32 (uint32_t __a)
22866 {
22867   return (uint16_t) __builtin_aarch64_uqmovnsi (__a);
22868 }
22869
22870 __extension__ extern __inline uint32_t
22871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22872 vqmovnd_u64 (uint64_t __a)
22873 {
22874   return (uint32_t) __builtin_aarch64_uqmovndi (__a);
22875 }
22876
22877 /* vqmovun */
22878
22879 __extension__ extern __inline uint8x8_t
22880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22881 vqmovun_s16 (int16x8_t __a)
22882 {
22883   return (uint8x8_t) __builtin_aarch64_sqmovunv8hi (__a);
22884 }
22885
22886 __extension__ extern __inline uint16x4_t
22887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22888 vqmovun_s32 (int32x4_t __a)
22889 {
22890   return (uint16x4_t) __builtin_aarch64_sqmovunv4si (__a);
22891 }
22892
22893 __extension__ extern __inline uint32x2_t
22894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22895 vqmovun_s64 (int64x2_t __a)
22896 {
22897   return (uint32x2_t) __builtin_aarch64_sqmovunv2di (__a);
22898 }
22899
22900 __extension__ extern __inline int8_t
22901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22902 vqmovunh_s16 (int16_t __a)
22903 {
22904   return (int8_t) __builtin_aarch64_sqmovunhi (__a);
22905 }
22906
22907 __extension__ extern __inline int16_t
22908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22909 vqmovuns_s32 (int32_t __a)
22910 {
22911   return (int16_t) __builtin_aarch64_sqmovunsi (__a);
22912 }
22913
22914 __extension__ extern __inline int32_t
22915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22916 vqmovund_s64 (int64_t __a)
22917 {
22918   return (int32_t) __builtin_aarch64_sqmovundi (__a);
22919 }
22920
22921 /* vqneg */
22922
22923 __extension__ extern __inline int64x2_t
22924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22925 vqnegq_s64 (int64x2_t __a)
22926 {
22927   return (int64x2_t) __builtin_aarch64_sqnegv2di (__a);
22928 }
22929
22930 __extension__ extern __inline int8_t
22931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22932 vqnegb_s8 (int8_t __a)
22933 {
22934   return (int8_t) __builtin_aarch64_sqnegqi (__a);
22935 }
22936
22937 __extension__ extern __inline int16_t
22938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22939 vqnegh_s16 (int16_t __a)
22940 {
22941   return (int16_t) __builtin_aarch64_sqneghi (__a);
22942 }
22943
22944 __extension__ extern __inline int32_t
22945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22946 vqnegs_s32 (int32_t __a)
22947 {
22948   return (int32_t) __builtin_aarch64_sqnegsi (__a);
22949 }
22950
22951 __extension__ extern __inline int64_t
22952 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22953 vqnegd_s64 (int64_t __a)
22954 {
22955   return __builtin_aarch64_sqnegdi (__a);
22956 }
22957
22958 /* vqrdmulh */
22959
22960 __extension__ extern __inline int16x4_t
22961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22962 vqrdmulh_lane_s16 (int16x4_t __a, int16x4_t __b, const int __c)
22963 {
22964   return  __builtin_aarch64_sqrdmulh_lanev4hi (__a, __b, __c);
22965 }
22966
22967 __extension__ extern __inline int32x2_t
22968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22969 vqrdmulh_lane_s32 (int32x2_t __a, int32x2_t __b, const int __c)
22970 {
22971   return __builtin_aarch64_sqrdmulh_lanev2si (__a, __b, __c);
22972 }
22973
22974 __extension__ extern __inline int16x8_t
22975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22976 vqrdmulhq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __c)
22977 {
22978   return __builtin_aarch64_sqrdmulh_lanev8hi (__a, __b, __c);
22979 }
22980
22981 __extension__ extern __inline int32x4_t
22982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22983 vqrdmulhq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __c)
22984 {
22985   return __builtin_aarch64_sqrdmulh_lanev4si (__a, __b, __c);
22986 }
22987
22988 __extension__ extern __inline int16_t
22989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22990 vqrdmulhh_s16 (int16_t __a, int16_t __b)
22991 {
22992   return (int16_t) __builtin_aarch64_sqrdmulhhi (__a, __b);
22993 }
22994
22995 __extension__ extern __inline int16_t
22996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22997 vqrdmulhh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
22998 {
22999   return __builtin_aarch64_sqrdmulh_lanehi (__a, __b, __c);
23000 }
23001
23002 __extension__ extern __inline int16_t
23003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23004 vqrdmulhh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
23005 {
23006   return __builtin_aarch64_sqrdmulh_laneqhi (__a, __b, __c);
23007 }
23008
23009 __extension__ extern __inline int32_t
23010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23011 vqrdmulhs_s32 (int32_t __a, int32_t __b)
23012 {
23013   return (int32_t) __builtin_aarch64_sqrdmulhsi (__a, __b);
23014 }
23015
23016 __extension__ extern __inline int32_t
23017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23018 vqrdmulhs_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
23019 {
23020   return __builtin_aarch64_sqrdmulh_lanesi (__a, __b, __c);
23021 }
23022
23023 __extension__ extern __inline int32_t
23024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23025 vqrdmulhs_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
23026 {
23027   return __builtin_aarch64_sqrdmulh_laneqsi (__a, __b, __c);
23028 }
23029
23030 /* vqrshl */
23031
23032 __extension__ extern __inline int8x8_t
23033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23034 vqrshl_s8 (int8x8_t __a, int8x8_t __b)
23035 {
23036   return __builtin_aarch64_sqrshlv8qi (__a, __b);
23037 }
23038
23039 __extension__ extern __inline int16x4_t
23040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23041 vqrshl_s16 (int16x4_t __a, int16x4_t __b)
23042 {
23043   return __builtin_aarch64_sqrshlv4hi (__a, __b);
23044 }
23045
23046 __extension__ extern __inline int32x2_t
23047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23048 vqrshl_s32 (int32x2_t __a, int32x2_t __b)
23049 {
23050   return __builtin_aarch64_sqrshlv2si (__a, __b);
23051 }
23052
23053 __extension__ extern __inline int64x1_t
23054 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23055 vqrshl_s64 (int64x1_t __a, int64x1_t __b)
23056 {
23057   return (int64x1_t) {__builtin_aarch64_sqrshldi (__a[0], __b[0])};
23058 }
23059
23060 __extension__ extern __inline uint8x8_t
23061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23062 vqrshl_u8 (uint8x8_t __a, int8x8_t __b)
23063 {
23064   return __builtin_aarch64_uqrshlv8qi_uus ( __a, __b);
23065 }
23066
23067 __extension__ extern __inline uint16x4_t
23068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23069 vqrshl_u16 (uint16x4_t __a, int16x4_t __b)
23070 {
23071   return __builtin_aarch64_uqrshlv4hi_uus ( __a, __b);
23072 }
23073
23074 __extension__ extern __inline uint32x2_t
23075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23076 vqrshl_u32 (uint32x2_t __a, int32x2_t __b)
23077 {
23078   return __builtin_aarch64_uqrshlv2si_uus ( __a, __b);
23079 }
23080
23081 __extension__ extern __inline uint64x1_t
23082 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23083 vqrshl_u64 (uint64x1_t __a, int64x1_t __b)
23084 {
23085   return (uint64x1_t) {__builtin_aarch64_uqrshldi_uus (__a[0], __b[0])};
23086 }
23087
23088 __extension__ extern __inline int8x16_t
23089 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23090 vqrshlq_s8 (int8x16_t __a, int8x16_t __b)
23091 {
23092   return __builtin_aarch64_sqrshlv16qi (__a, __b);
23093 }
23094
23095 __extension__ extern __inline int16x8_t
23096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23097 vqrshlq_s16 (int16x8_t __a, int16x8_t __b)
23098 {
23099   return __builtin_aarch64_sqrshlv8hi (__a, __b);
23100 }
23101
23102 __extension__ extern __inline int32x4_t
23103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23104 vqrshlq_s32 (int32x4_t __a, int32x4_t __b)
23105 {
23106   return __builtin_aarch64_sqrshlv4si (__a, __b);
23107 }
23108
23109 __extension__ extern __inline int64x2_t
23110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23111 vqrshlq_s64 (int64x2_t __a, int64x2_t __b)
23112 {
23113   return __builtin_aarch64_sqrshlv2di (__a, __b);
23114 }
23115
23116 __extension__ extern __inline uint8x16_t
23117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23118 vqrshlq_u8 (uint8x16_t __a, int8x16_t __b)
23119 {
23120   return __builtin_aarch64_uqrshlv16qi_uus ( __a, __b);
23121 }
23122
23123 __extension__ extern __inline uint16x8_t
23124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23125 vqrshlq_u16 (uint16x8_t __a, int16x8_t __b)
23126 {
23127   return __builtin_aarch64_uqrshlv8hi_uus ( __a, __b);
23128 }
23129
23130 __extension__ extern __inline uint32x4_t
23131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23132 vqrshlq_u32 (uint32x4_t __a, int32x4_t __b)
23133 {
23134   return __builtin_aarch64_uqrshlv4si_uus ( __a, __b);
23135 }
23136
23137 __extension__ extern __inline uint64x2_t
23138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23139 vqrshlq_u64 (uint64x2_t __a, int64x2_t __b)
23140 {
23141   return __builtin_aarch64_uqrshlv2di_uus ( __a, __b);
23142 }
23143
23144 __extension__ extern __inline int8_t
23145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23146 vqrshlb_s8 (int8_t __a, int8_t __b)
23147 {
23148   return __builtin_aarch64_sqrshlqi (__a, __b);
23149 }
23150
23151 __extension__ extern __inline int16_t
23152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23153 vqrshlh_s16 (int16_t __a, int16_t __b)
23154 {
23155   return __builtin_aarch64_sqrshlhi (__a, __b);
23156 }
23157
23158 __extension__ extern __inline int32_t
23159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23160 vqrshls_s32 (int32_t __a, int32_t __b)
23161 {
23162   return __builtin_aarch64_sqrshlsi (__a, __b);
23163 }
23164
23165 __extension__ extern __inline int64_t
23166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23167 vqrshld_s64 (int64_t __a, int64_t __b)
23168 {
23169   return __builtin_aarch64_sqrshldi (__a, __b);
23170 }
23171
23172 __extension__ extern __inline uint8_t
23173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23174 vqrshlb_u8 (uint8_t __a, uint8_t __b)
23175 {
23176   return __builtin_aarch64_uqrshlqi_uus (__a, __b);
23177 }
23178
23179 __extension__ extern __inline uint16_t
23180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23181 vqrshlh_u16 (uint16_t __a, uint16_t __b)
23182 {
23183   return __builtin_aarch64_uqrshlhi_uus (__a, __b);
23184 }
23185
23186 __extension__ extern __inline uint32_t
23187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23188 vqrshls_u32 (uint32_t __a, uint32_t __b)
23189 {
23190   return __builtin_aarch64_uqrshlsi_uus (__a, __b);
23191 }
23192
23193 __extension__ extern __inline uint64_t
23194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23195 vqrshld_u64 (uint64_t __a, uint64_t __b)
23196 {
23197   return __builtin_aarch64_uqrshldi_uus (__a, __b);
23198 }
23199
23200 /* vqrshrn */
23201
23202 __extension__ extern __inline int8x8_t
23203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23204 vqrshrn_n_s16 (int16x8_t __a, const int __b)
23205 {
23206   return (int8x8_t) __builtin_aarch64_sqrshrn_nv8hi (__a, __b);
23207 }
23208
23209 __extension__ extern __inline int16x4_t
23210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23211 vqrshrn_n_s32 (int32x4_t __a, const int __b)
23212 {
23213   return (int16x4_t) __builtin_aarch64_sqrshrn_nv4si (__a, __b);
23214 }
23215
23216 __extension__ extern __inline int32x2_t
23217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23218 vqrshrn_n_s64 (int64x2_t __a, const int __b)
23219 {
23220   return (int32x2_t) __builtin_aarch64_sqrshrn_nv2di (__a, __b);
23221 }
23222
23223 __extension__ extern __inline uint8x8_t
23224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23225 vqrshrn_n_u16 (uint16x8_t __a, const int __b)
23226 {
23227   return __builtin_aarch64_uqrshrn_nv8hi_uus ( __a, __b);
23228 }
23229
23230 __extension__ extern __inline uint16x4_t
23231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23232 vqrshrn_n_u32 (uint32x4_t __a, const int __b)
23233 {
23234   return __builtin_aarch64_uqrshrn_nv4si_uus ( __a, __b);
23235 }
23236
23237 __extension__ extern __inline uint32x2_t
23238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23239 vqrshrn_n_u64 (uint64x2_t __a, const int __b)
23240 {
23241   return __builtin_aarch64_uqrshrn_nv2di_uus ( __a, __b);
23242 }
23243
23244 __extension__ extern __inline int8_t
23245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23246 vqrshrnh_n_s16 (int16_t __a, const int __b)
23247 {
23248   return (int8_t) __builtin_aarch64_sqrshrn_nhi (__a, __b);
23249 }
23250
23251 __extension__ extern __inline int16_t
23252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23253 vqrshrns_n_s32 (int32_t __a, const int __b)
23254 {
23255   return (int16_t) __builtin_aarch64_sqrshrn_nsi (__a, __b);
23256 }
23257
23258 __extension__ extern __inline int32_t
23259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23260 vqrshrnd_n_s64 (int64_t __a, const int __b)
23261 {
23262   return (int32_t) __builtin_aarch64_sqrshrn_ndi (__a, __b);
23263 }
23264
23265 __extension__ extern __inline uint8_t
23266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23267 vqrshrnh_n_u16 (uint16_t __a, const int __b)
23268 {
23269   return __builtin_aarch64_uqrshrn_nhi_uus (__a, __b);
23270 }
23271
23272 __extension__ extern __inline uint16_t
23273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23274 vqrshrns_n_u32 (uint32_t __a, const int __b)
23275 {
23276   return __builtin_aarch64_uqrshrn_nsi_uus (__a, __b);
23277 }
23278
23279 __extension__ extern __inline uint32_t
23280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23281 vqrshrnd_n_u64 (uint64_t __a, const int __b)
23282 {
23283   return __builtin_aarch64_uqrshrn_ndi_uus (__a, __b);
23284 }
23285
23286 /* vqrshrun */
23287
23288 __extension__ extern __inline uint8x8_t
23289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23290 vqrshrun_n_s16 (int16x8_t __a, const int __b)
23291 {
23292   return (uint8x8_t) __builtin_aarch64_sqrshrun_nv8hi (__a, __b);
23293 }
23294
23295 __extension__ extern __inline uint16x4_t
23296 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23297 vqrshrun_n_s32 (int32x4_t __a, const int __b)
23298 {
23299   return (uint16x4_t) __builtin_aarch64_sqrshrun_nv4si (__a, __b);
23300 }
23301
23302 __extension__ extern __inline uint32x2_t
23303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23304 vqrshrun_n_s64 (int64x2_t __a, const int __b)
23305 {
23306   return (uint32x2_t) __builtin_aarch64_sqrshrun_nv2di (__a, __b);
23307 }
23308
23309 __extension__ extern __inline int8_t
23310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23311 vqrshrunh_n_s16 (int16_t __a, const int __b)
23312 {
23313   return (int8_t) __builtin_aarch64_sqrshrun_nhi (__a, __b);
23314 }
23315
23316 __extension__ extern __inline int16_t
23317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23318 vqrshruns_n_s32 (int32_t __a, const int __b)
23319 {
23320   return (int16_t) __builtin_aarch64_sqrshrun_nsi (__a, __b);
23321 }
23322
23323 __extension__ extern __inline int32_t
23324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23325 vqrshrund_n_s64 (int64_t __a, const int __b)
23326 {
23327   return (int32_t) __builtin_aarch64_sqrshrun_ndi (__a, __b);
23328 }
23329
23330 /* vqshl */
23331
23332 __extension__ extern __inline int8x8_t
23333 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23334 vqshl_s8 (int8x8_t __a, int8x8_t __b)
23335 {
23336   return __builtin_aarch64_sqshlv8qi (__a, __b);
23337 }
23338
23339 __extension__ extern __inline int16x4_t
23340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23341 vqshl_s16 (int16x4_t __a, int16x4_t __b)
23342 {
23343   return __builtin_aarch64_sqshlv4hi (__a, __b);
23344 }
23345
23346 __extension__ extern __inline int32x2_t
23347 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23348 vqshl_s32 (int32x2_t __a, int32x2_t __b)
23349 {
23350   return __builtin_aarch64_sqshlv2si (__a, __b);
23351 }
23352
23353 __extension__ extern __inline int64x1_t
23354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23355 vqshl_s64 (int64x1_t __a, int64x1_t __b)
23356 {
23357   return (int64x1_t) {__builtin_aarch64_sqshldi (__a[0], __b[0])};
23358 }
23359
23360 __extension__ extern __inline uint8x8_t
23361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23362 vqshl_u8 (uint8x8_t __a, int8x8_t __b)
23363 {
23364   return __builtin_aarch64_uqshlv8qi_uus ( __a, __b);
23365 }
23366
23367 __extension__ extern __inline uint16x4_t
23368 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23369 vqshl_u16 (uint16x4_t __a, int16x4_t __b)
23370 {
23371   return __builtin_aarch64_uqshlv4hi_uus ( __a, __b);
23372 }
23373
23374 __extension__ extern __inline uint32x2_t
23375 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23376 vqshl_u32 (uint32x2_t __a, int32x2_t __b)
23377 {
23378   return __builtin_aarch64_uqshlv2si_uus ( __a, __b);
23379 }
23380
23381 __extension__ extern __inline uint64x1_t
23382 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23383 vqshl_u64 (uint64x1_t __a, int64x1_t __b)
23384 {
23385   return (uint64x1_t) {__builtin_aarch64_uqshldi_uus (__a[0], __b[0])};
23386 }
23387
23388 __extension__ extern __inline int8x16_t
23389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23390 vqshlq_s8 (int8x16_t __a, int8x16_t __b)
23391 {
23392   return __builtin_aarch64_sqshlv16qi (__a, __b);
23393 }
23394
23395 __extension__ extern __inline int16x8_t
23396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23397 vqshlq_s16 (int16x8_t __a, int16x8_t __b)
23398 {
23399   return __builtin_aarch64_sqshlv8hi (__a, __b);
23400 }
23401
23402 __extension__ extern __inline int32x4_t
23403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23404 vqshlq_s32 (int32x4_t __a, int32x4_t __b)
23405 {
23406   return __builtin_aarch64_sqshlv4si (__a, __b);
23407 }
23408
23409 __extension__ extern __inline int64x2_t
23410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23411 vqshlq_s64 (int64x2_t __a, int64x2_t __b)
23412 {
23413   return __builtin_aarch64_sqshlv2di (__a, __b);
23414 }
23415
23416 __extension__ extern __inline uint8x16_t
23417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23418 vqshlq_u8 (uint8x16_t __a, int8x16_t __b)
23419 {
23420   return __builtin_aarch64_uqshlv16qi_uus ( __a, __b);
23421 }
23422
23423 __extension__ extern __inline uint16x8_t
23424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23425 vqshlq_u16 (uint16x8_t __a, int16x8_t __b)
23426 {
23427   return __builtin_aarch64_uqshlv8hi_uus ( __a, __b);
23428 }
23429
23430 __extension__ extern __inline uint32x4_t
23431 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23432 vqshlq_u32 (uint32x4_t __a, int32x4_t __b)
23433 {
23434   return __builtin_aarch64_uqshlv4si_uus ( __a, __b);
23435 }
23436
23437 __extension__ extern __inline uint64x2_t
23438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23439 vqshlq_u64 (uint64x2_t __a, int64x2_t __b)
23440 {
23441   return __builtin_aarch64_uqshlv2di_uus ( __a, __b);
23442 }
23443
23444 __extension__ extern __inline int8_t
23445 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23446 vqshlb_s8 (int8_t __a, int8_t __b)
23447 {
23448   return __builtin_aarch64_sqshlqi (__a, __b);
23449 }
23450
23451 __extension__ extern __inline int16_t
23452 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23453 vqshlh_s16 (int16_t __a, int16_t __b)
23454 {
23455   return __builtin_aarch64_sqshlhi (__a, __b);
23456 }
23457
23458 __extension__ extern __inline int32_t
23459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23460 vqshls_s32 (int32_t __a, int32_t __b)
23461 {
23462   return __builtin_aarch64_sqshlsi (__a, __b);
23463 }
23464
23465 __extension__ extern __inline int64_t
23466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23467 vqshld_s64 (int64_t __a, int64_t __b)
23468 {
23469   return __builtin_aarch64_sqshldi (__a, __b);
23470 }
23471
23472 __extension__ extern __inline uint8_t
23473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23474 vqshlb_u8 (uint8_t __a, uint8_t __b)
23475 {
23476   return __builtin_aarch64_uqshlqi_uus (__a, __b);
23477 }
23478
23479 __extension__ extern __inline uint16_t
23480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23481 vqshlh_u16 (uint16_t __a, uint16_t __b)
23482 {
23483   return __builtin_aarch64_uqshlhi_uus (__a, __b);
23484 }
23485
23486 __extension__ extern __inline uint32_t
23487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23488 vqshls_u32 (uint32_t __a, uint32_t __b)
23489 {
23490   return __builtin_aarch64_uqshlsi_uus (__a, __b);
23491 }
23492
23493 __extension__ extern __inline uint64_t
23494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23495 vqshld_u64 (uint64_t __a, uint64_t __b)
23496 {
23497   return __builtin_aarch64_uqshldi_uus (__a, __b);
23498 }
23499
23500 __extension__ extern __inline int8x8_t
23501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23502 vqshl_n_s8 (int8x8_t __a, const int __b)
23503 {
23504   return (int8x8_t) __builtin_aarch64_sqshl_nv8qi (__a, __b);
23505 }
23506
23507 __extension__ extern __inline int16x4_t
23508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23509 vqshl_n_s16 (int16x4_t __a, const int __b)
23510 {
23511   return (int16x4_t) __builtin_aarch64_sqshl_nv4hi (__a, __b);
23512 }
23513
23514 __extension__ extern __inline int32x2_t
23515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23516 vqshl_n_s32 (int32x2_t __a, const int __b)
23517 {
23518   return (int32x2_t) __builtin_aarch64_sqshl_nv2si (__a, __b);
23519 }
23520
23521 __extension__ extern __inline int64x1_t
23522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23523 vqshl_n_s64 (int64x1_t __a, const int __b)
23524 {
23525   return (int64x1_t) {__builtin_aarch64_sqshl_ndi (__a[0], __b)};
23526 }
23527
23528 __extension__ extern __inline uint8x8_t
23529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23530 vqshl_n_u8 (uint8x8_t __a, const int __b)
23531 {
23532   return __builtin_aarch64_uqshl_nv8qi_uus (__a, __b);
23533 }
23534
23535 __extension__ extern __inline uint16x4_t
23536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23537 vqshl_n_u16 (uint16x4_t __a, const int __b)
23538 {
23539   return __builtin_aarch64_uqshl_nv4hi_uus (__a, __b);
23540 }
23541
23542 __extension__ extern __inline uint32x2_t
23543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23544 vqshl_n_u32 (uint32x2_t __a, const int __b)
23545 {
23546   return __builtin_aarch64_uqshl_nv2si_uus (__a, __b);
23547 }
23548
23549 __extension__ extern __inline uint64x1_t
23550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23551 vqshl_n_u64 (uint64x1_t __a, const int __b)
23552 {
23553   return (uint64x1_t) {__builtin_aarch64_uqshl_ndi_uus (__a[0], __b)};
23554 }
23555
23556 __extension__ extern __inline int8x16_t
23557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23558 vqshlq_n_s8 (int8x16_t __a, const int __b)
23559 {
23560   return (int8x16_t) __builtin_aarch64_sqshl_nv16qi (__a, __b);
23561 }
23562
23563 __extension__ extern __inline int16x8_t
23564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23565 vqshlq_n_s16 (int16x8_t __a, const int __b)
23566 {
23567   return (int16x8_t) __builtin_aarch64_sqshl_nv8hi (__a, __b);
23568 }
23569
23570 __extension__ extern __inline int32x4_t
23571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23572 vqshlq_n_s32 (int32x4_t __a, const int __b)
23573 {
23574   return (int32x4_t) __builtin_aarch64_sqshl_nv4si (__a, __b);
23575 }
23576
23577 __extension__ extern __inline int64x2_t
23578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23579 vqshlq_n_s64 (int64x2_t __a, const int __b)
23580 {
23581   return (int64x2_t) __builtin_aarch64_sqshl_nv2di (__a, __b);
23582 }
23583
23584 __extension__ extern __inline uint8x16_t
23585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23586 vqshlq_n_u8 (uint8x16_t __a, const int __b)
23587 {
23588   return __builtin_aarch64_uqshl_nv16qi_uus (__a, __b);
23589 }
23590
23591 __extension__ extern __inline uint16x8_t
23592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23593 vqshlq_n_u16 (uint16x8_t __a, const int __b)
23594 {
23595   return __builtin_aarch64_uqshl_nv8hi_uus (__a, __b);
23596 }
23597
23598 __extension__ extern __inline uint32x4_t
23599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23600 vqshlq_n_u32 (uint32x4_t __a, const int __b)
23601 {
23602   return __builtin_aarch64_uqshl_nv4si_uus (__a, __b);
23603 }
23604
23605 __extension__ extern __inline uint64x2_t
23606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23607 vqshlq_n_u64 (uint64x2_t __a, const int __b)
23608 {
23609   return __builtin_aarch64_uqshl_nv2di_uus (__a, __b);
23610 }
23611
23612 __extension__ extern __inline int8_t
23613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23614 vqshlb_n_s8 (int8_t __a, const int __b)
23615 {
23616   return (int8_t) __builtin_aarch64_sqshl_nqi (__a, __b);
23617 }
23618
23619 __extension__ extern __inline int16_t
23620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23621 vqshlh_n_s16 (int16_t __a, const int __b)
23622 {
23623   return (int16_t) __builtin_aarch64_sqshl_nhi (__a, __b);
23624 }
23625
23626 __extension__ extern __inline int32_t
23627 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23628 vqshls_n_s32 (int32_t __a, const int __b)
23629 {
23630   return (int32_t) __builtin_aarch64_sqshl_nsi (__a, __b);
23631 }
23632
23633 __extension__ extern __inline int64_t
23634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23635 vqshld_n_s64 (int64_t __a, const int __b)
23636 {
23637   return __builtin_aarch64_sqshl_ndi (__a, __b);
23638 }
23639
23640 __extension__ extern __inline uint8_t
23641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23642 vqshlb_n_u8 (uint8_t __a, const int __b)
23643 {
23644   return __builtin_aarch64_uqshl_nqi_uus (__a, __b);
23645 }
23646
23647 __extension__ extern __inline uint16_t
23648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23649 vqshlh_n_u16 (uint16_t __a, const int __b)
23650 {
23651   return __builtin_aarch64_uqshl_nhi_uus (__a, __b);
23652 }
23653
23654 __extension__ extern __inline uint32_t
23655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23656 vqshls_n_u32 (uint32_t __a, const int __b)
23657 {
23658   return __builtin_aarch64_uqshl_nsi_uus (__a, __b);
23659 }
23660
23661 __extension__ extern __inline uint64_t
23662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23663 vqshld_n_u64 (uint64_t __a, const int __b)
23664 {
23665   return __builtin_aarch64_uqshl_ndi_uus (__a, __b);
23666 }
23667
23668 /* vqshlu */
23669
23670 __extension__ extern __inline uint8x8_t
23671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23672 vqshlu_n_s8 (int8x8_t __a, const int __b)
23673 {
23674   return __builtin_aarch64_sqshlu_nv8qi_uss (__a, __b);
23675 }
23676
23677 __extension__ extern __inline uint16x4_t
23678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23679 vqshlu_n_s16 (int16x4_t __a, const int __b)
23680 {
23681   return __builtin_aarch64_sqshlu_nv4hi_uss (__a, __b);
23682 }
23683
23684 __extension__ extern __inline uint32x2_t
23685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23686 vqshlu_n_s32 (int32x2_t __a, const int __b)
23687 {
23688   return __builtin_aarch64_sqshlu_nv2si_uss (__a, __b);
23689 }
23690
23691 __extension__ extern __inline uint64x1_t
23692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23693 vqshlu_n_s64 (int64x1_t __a, const int __b)
23694 {
23695   return (uint64x1_t) {__builtin_aarch64_sqshlu_ndi_uss (__a[0], __b)};
23696 }
23697
23698 __extension__ extern __inline uint8x16_t
23699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23700 vqshluq_n_s8 (int8x16_t __a, const int __b)
23701 {
23702   return __builtin_aarch64_sqshlu_nv16qi_uss (__a, __b);
23703 }
23704
23705 __extension__ extern __inline uint16x8_t
23706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23707 vqshluq_n_s16 (int16x8_t __a, const int __b)
23708 {
23709   return __builtin_aarch64_sqshlu_nv8hi_uss (__a, __b);
23710 }
23711
23712 __extension__ extern __inline uint32x4_t
23713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23714 vqshluq_n_s32 (int32x4_t __a, const int __b)
23715 {
23716   return __builtin_aarch64_sqshlu_nv4si_uss (__a, __b);
23717 }
23718
23719 __extension__ extern __inline uint64x2_t
23720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23721 vqshluq_n_s64 (int64x2_t __a, const int __b)
23722 {
23723   return __builtin_aarch64_sqshlu_nv2di_uss (__a, __b);
23724 }
23725
23726 __extension__ extern __inline int8_t
23727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23728 vqshlub_n_s8 (int8_t __a, const int __b)
23729 {
23730   return (int8_t) __builtin_aarch64_sqshlu_nqi_uss (__a, __b);
23731 }
23732
23733 __extension__ extern __inline int16_t
23734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23735 vqshluh_n_s16 (int16_t __a, const int __b)
23736 {
23737   return (int16_t) __builtin_aarch64_sqshlu_nhi_uss (__a, __b);
23738 }
23739
23740 __extension__ extern __inline int32_t
23741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23742 vqshlus_n_s32 (int32_t __a, const int __b)
23743 {
23744   return (int32_t) __builtin_aarch64_sqshlu_nsi_uss (__a, __b);
23745 }
23746
23747 __extension__ extern __inline uint64_t
23748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23749 vqshlud_n_s64 (int64_t __a, const int __b)
23750 {
23751   return __builtin_aarch64_sqshlu_ndi_uss (__a, __b);
23752 }
23753
23754 /* vqshrn */
23755
23756 __extension__ extern __inline int8x8_t
23757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23758 vqshrn_n_s16 (int16x8_t __a, const int __b)
23759 {
23760   return (int8x8_t) __builtin_aarch64_sqshrn_nv8hi (__a, __b);
23761 }
23762
23763 __extension__ extern __inline int16x4_t
23764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23765 vqshrn_n_s32 (int32x4_t __a, const int __b)
23766 {
23767   return (int16x4_t) __builtin_aarch64_sqshrn_nv4si (__a, __b);
23768 }
23769
23770 __extension__ extern __inline int32x2_t
23771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23772 vqshrn_n_s64 (int64x2_t __a, const int __b)
23773 {
23774   return (int32x2_t) __builtin_aarch64_sqshrn_nv2di (__a, __b);
23775 }
23776
23777 __extension__ extern __inline uint8x8_t
23778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23779 vqshrn_n_u16 (uint16x8_t __a, const int __b)
23780 {
23781   return __builtin_aarch64_uqshrn_nv8hi_uus ( __a, __b);
23782 }
23783
23784 __extension__ extern __inline uint16x4_t
23785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23786 vqshrn_n_u32 (uint32x4_t __a, const int __b)
23787 {
23788   return __builtin_aarch64_uqshrn_nv4si_uus ( __a, __b);
23789 }
23790
23791 __extension__ extern __inline uint32x2_t
23792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23793 vqshrn_n_u64 (uint64x2_t __a, const int __b)
23794 {
23795   return __builtin_aarch64_uqshrn_nv2di_uus ( __a, __b);
23796 }
23797
23798 __extension__ extern __inline int8_t
23799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23800 vqshrnh_n_s16 (int16_t __a, const int __b)
23801 {
23802   return (int8_t) __builtin_aarch64_sqshrn_nhi (__a, __b);
23803 }
23804
23805 __extension__ extern __inline int16_t
23806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23807 vqshrns_n_s32 (int32_t __a, const int __b)
23808 {
23809   return (int16_t) __builtin_aarch64_sqshrn_nsi (__a, __b);
23810 }
23811
23812 __extension__ extern __inline int32_t
23813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23814 vqshrnd_n_s64 (int64_t __a, const int __b)
23815 {
23816   return (int32_t) __builtin_aarch64_sqshrn_ndi (__a, __b);
23817 }
23818
23819 __extension__ extern __inline uint8_t
23820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23821 vqshrnh_n_u16 (uint16_t __a, const int __b)
23822 {
23823   return __builtin_aarch64_uqshrn_nhi_uus (__a, __b);
23824 }
23825
23826 __extension__ extern __inline uint16_t
23827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23828 vqshrns_n_u32 (uint32_t __a, const int __b)
23829 {
23830   return __builtin_aarch64_uqshrn_nsi_uus (__a, __b);
23831 }
23832
23833 __extension__ extern __inline uint32_t
23834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23835 vqshrnd_n_u64 (uint64_t __a, const int __b)
23836 {
23837   return __builtin_aarch64_uqshrn_ndi_uus (__a, __b);
23838 }
23839
23840 /* vqshrun */
23841
23842 __extension__ extern __inline uint8x8_t
23843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23844 vqshrun_n_s16 (int16x8_t __a, const int __b)
23845 {
23846   return (uint8x8_t) __builtin_aarch64_sqshrun_nv8hi (__a, __b);
23847 }
23848
23849 __extension__ extern __inline uint16x4_t
23850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23851 vqshrun_n_s32 (int32x4_t __a, const int __b)
23852 {
23853   return (uint16x4_t) __builtin_aarch64_sqshrun_nv4si (__a, __b);
23854 }
23855
23856 __extension__ extern __inline uint32x2_t
23857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23858 vqshrun_n_s64 (int64x2_t __a, const int __b)
23859 {
23860   return (uint32x2_t) __builtin_aarch64_sqshrun_nv2di (__a, __b);
23861 }
23862
23863 __extension__ extern __inline int8_t
23864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23865 vqshrunh_n_s16 (int16_t __a, const int __b)
23866 {
23867   return (int8_t) __builtin_aarch64_sqshrun_nhi (__a, __b);
23868 }
23869
23870 __extension__ extern __inline int16_t
23871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23872 vqshruns_n_s32 (int32_t __a, const int __b)
23873 {
23874   return (int16_t) __builtin_aarch64_sqshrun_nsi (__a, __b);
23875 }
23876
23877 __extension__ extern __inline int32_t
23878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23879 vqshrund_n_s64 (int64_t __a, const int __b)
23880 {
23881   return (int32_t) __builtin_aarch64_sqshrun_ndi (__a, __b);
23882 }
23883
23884 /* vqsub */
23885
23886 __extension__ extern __inline int8_t
23887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23888 vqsubb_s8 (int8_t __a, int8_t __b)
23889 {
23890   return (int8_t) __builtin_aarch64_sqsubqi (__a, __b);
23891 }
23892
23893 __extension__ extern __inline int16_t
23894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23895 vqsubh_s16 (int16_t __a, int16_t __b)
23896 {
23897   return (int16_t) __builtin_aarch64_sqsubhi (__a, __b);
23898 }
23899
23900 __extension__ extern __inline int32_t
23901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23902 vqsubs_s32 (int32_t __a, int32_t __b)
23903 {
23904   return (int32_t) __builtin_aarch64_sqsubsi (__a, __b);
23905 }
23906
23907 __extension__ extern __inline int64_t
23908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23909 vqsubd_s64 (int64_t __a, int64_t __b)
23910 {
23911   return __builtin_aarch64_sqsubdi (__a, __b);
23912 }
23913
23914 __extension__ extern __inline uint8_t
23915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23916 vqsubb_u8 (uint8_t __a, uint8_t __b)
23917 {
23918   return (uint8_t) __builtin_aarch64_uqsubqi_uuu (__a, __b);
23919 }
23920
23921 __extension__ extern __inline uint16_t
23922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23923 vqsubh_u16 (uint16_t __a, uint16_t __b)
23924 {
23925   return (uint16_t) __builtin_aarch64_uqsubhi_uuu (__a, __b);
23926 }
23927
23928 __extension__ extern __inline uint32_t
23929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23930 vqsubs_u32 (uint32_t __a, uint32_t __b)
23931 {
23932   return (uint32_t) __builtin_aarch64_uqsubsi_uuu (__a, __b);
23933 }
23934
23935 __extension__ extern __inline uint64_t
23936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23937 vqsubd_u64 (uint64_t __a, uint64_t __b)
23938 {
23939   return __builtin_aarch64_uqsubdi_uuu (__a, __b);
23940 }
23941
23942 /* vqtbl2 */
23943
23944 __extension__ extern __inline int8x8_t
23945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23946 vqtbl2_s8 (int8x16x2_t tab, uint8x8_t idx)
23947 {
23948   __builtin_aarch64_simd_oi __o;
23949   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
23950   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
23951   return __builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
23952 }
23953
23954 __extension__ extern __inline uint8x8_t
23955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23956 vqtbl2_u8 (uint8x16x2_t tab, uint8x8_t idx)
23957 {
23958   __builtin_aarch64_simd_oi __o;
23959   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23960   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23961   return (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
23962 }
23963
23964 __extension__ extern __inline poly8x8_t
23965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23966 vqtbl2_p8 (poly8x16x2_t tab, uint8x8_t idx)
23967 {
23968   __builtin_aarch64_simd_oi __o;
23969   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23970   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23971   return (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
23972 }
23973
23974 __extension__ extern __inline int8x16_t
23975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23976 vqtbl2q_s8 (int8x16x2_t tab, uint8x16_t idx)
23977 {
23978   __builtin_aarch64_simd_oi __o;
23979   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23980   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23981   return __builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
23982 }
23983
23984 __extension__ extern __inline uint8x16_t
23985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23986 vqtbl2q_u8 (uint8x16x2_t tab, uint8x16_t idx)
23987 {
23988   __builtin_aarch64_simd_oi __o;
23989   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23990   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23991   return (uint8x16_t)__builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
23992 }
23993
23994 __extension__ extern __inline poly8x16_t
23995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23996 vqtbl2q_p8 (poly8x16x2_t tab, uint8x16_t idx)
23997 {
23998   __builtin_aarch64_simd_oi __o;
23999   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24000   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24001   return (poly8x16_t)__builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
24002 }
24003
24004 /* vqtbl3 */
24005
24006 __extension__ extern __inline int8x8_t
24007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24008 vqtbl3_s8 (int8x16x3_t tab, uint8x8_t idx)
24009 {
24010   __builtin_aarch64_simd_ci __o;
24011   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24012   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24013   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24014   return __builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
24015 }
24016
24017 __extension__ extern __inline uint8x8_t
24018 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24019 vqtbl3_u8 (uint8x16x3_t tab, uint8x8_t idx)
24020 {
24021   __builtin_aarch64_simd_ci __o;
24022   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24023   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24024   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24025   return (uint8x8_t)__builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
24026 }
24027
24028 __extension__ extern __inline poly8x8_t
24029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24030 vqtbl3_p8 (poly8x16x3_t tab, uint8x8_t idx)
24031 {
24032   __builtin_aarch64_simd_ci __o;
24033   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24034   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24035   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24036   return (poly8x8_t)__builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
24037 }
24038
24039 __extension__ extern __inline int8x16_t
24040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24041 vqtbl3q_s8 (int8x16x3_t tab, uint8x16_t idx)
24042 {
24043   __builtin_aarch64_simd_ci __o;
24044   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24045   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24046   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24047   return __builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24048 }
24049
24050 __extension__ extern __inline uint8x16_t
24051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24052 vqtbl3q_u8 (uint8x16x3_t tab, uint8x16_t idx)
24053 {
24054   __builtin_aarch64_simd_ci __o;
24055   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24056   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24057   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24058   return (uint8x16_t)__builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24059 }
24060
24061 __extension__ extern __inline poly8x16_t
24062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24063 vqtbl3q_p8 (poly8x16x3_t tab, uint8x16_t idx)
24064 {
24065   __builtin_aarch64_simd_ci __o;
24066   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24067   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24068   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24069   return (poly8x16_t)__builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24070 }
24071
24072 /* vqtbl4 */
24073
24074 __extension__ extern __inline int8x8_t
24075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24076 vqtbl4_s8 (int8x16x4_t tab, uint8x8_t idx)
24077 {
24078   __builtin_aarch64_simd_xi __o;
24079   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24080   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24081   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24082   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24083   return __builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24084 }
24085
24086 __extension__ extern __inline uint8x8_t
24087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24088 vqtbl4_u8 (uint8x16x4_t tab, uint8x8_t idx)
24089 {
24090   __builtin_aarch64_simd_xi __o;
24091   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24092   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24093   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24094   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24095   return (uint8x8_t)__builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24096 }
24097
24098 __extension__ extern __inline poly8x8_t
24099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24100 vqtbl4_p8 (poly8x16x4_t tab, uint8x8_t idx)
24101 {
24102   __builtin_aarch64_simd_xi __o;
24103   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24104   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24105   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24106   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24107   return (poly8x8_t)__builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24108 }
24109
24110 __extension__ extern __inline int8x16_t
24111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24112 vqtbl4q_s8 (int8x16x4_t tab, uint8x16_t idx)
24113 {
24114   __builtin_aarch64_simd_xi __o;
24115   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24116   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24117   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24118   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24119   return __builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
24120 }
24121
24122 __extension__ extern __inline uint8x16_t
24123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24124 vqtbl4q_u8 (uint8x16x4_t tab, uint8x16_t idx)
24125 {
24126   __builtin_aarch64_simd_xi __o;
24127   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24128   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24129   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24130   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24131   return (uint8x16_t)__builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
24132 }
24133
24134 __extension__ extern __inline poly8x16_t
24135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24136 vqtbl4q_p8 (poly8x16x4_t tab, uint8x16_t idx)
24137 {
24138   __builtin_aarch64_simd_xi __o;
24139   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24140   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24141   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24142   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24143   return (poly8x16_t)__builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
24144 }
24145
24146
24147 /* vqtbx2 */
24148 __extension__ extern __inline int8x8_t
24149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24150 vqtbx2_s8 (int8x8_t r, int8x16x2_t tab, uint8x8_t idx)
24151 {
24152   __builtin_aarch64_simd_oi __o;
24153   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
24154   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
24155   return __builtin_aarch64_tbx4v8qi (r, __o, (int8x8_t)idx);
24156 }
24157
24158 __extension__ extern __inline uint8x8_t
24159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24160 vqtbx2_u8 (uint8x8_t r, uint8x16x2_t tab, uint8x8_t idx)
24161 {
24162   __builtin_aarch64_simd_oi __o;
24163   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24164   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24165   return (uint8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)r, __o,
24166                                                 (int8x8_t)idx);
24167 }
24168
24169 __extension__ extern __inline poly8x8_t
24170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24171 vqtbx2_p8 (poly8x8_t r, poly8x16x2_t tab, uint8x8_t idx)
24172 {
24173   __builtin_aarch64_simd_oi __o;
24174   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24175   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24176   return (poly8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)r, __o,
24177                                                 (int8x8_t)idx);
24178 }
24179
24180 __extension__ extern __inline int8x16_t
24181 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24182 vqtbx2q_s8 (int8x16_t r, int8x16x2_t tab, uint8x16_t idx)
24183 {
24184   __builtin_aarch64_simd_oi __o;
24185   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
24186   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
24187   return __builtin_aarch64_tbx4v16qi (r, __o, (int8x16_t)idx);
24188 }
24189
24190 __extension__ extern __inline uint8x16_t
24191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24192 vqtbx2q_u8 (uint8x16_t r, uint8x16x2_t tab, uint8x16_t idx)
24193 {
24194   __builtin_aarch64_simd_oi __o;
24195   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24196   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24197   return (uint8x16_t)__builtin_aarch64_tbx4v16qi ((int8x16_t)r, __o,
24198                                                   (int8x16_t)idx);
24199 }
24200
24201 __extension__ extern __inline poly8x16_t
24202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24203 vqtbx2q_p8 (poly8x16_t r, poly8x16x2_t tab, uint8x16_t idx)
24204 {
24205   __builtin_aarch64_simd_oi __o;
24206   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24207   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24208   return (poly8x16_t)__builtin_aarch64_tbx4v16qi ((int8x16_t)r, __o,
24209                                                   (int8x16_t)idx);
24210 }
24211
24212 /* vqtbx3 */
24213 __extension__ extern __inline int8x8_t
24214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24215 vqtbx3_s8 (int8x8_t r, int8x16x3_t tab, uint8x8_t idx)
24216 {
24217   __builtin_aarch64_simd_ci __o;
24218   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[0], 0);
24219   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[1], 1);
24220   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[2], 2);
24221   return __builtin_aarch64_qtbx3v8qi (r, __o, (int8x8_t)idx);
24222 }
24223
24224 __extension__ extern __inline uint8x8_t
24225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24226 vqtbx3_u8 (uint8x8_t r, uint8x16x3_t tab, uint8x8_t idx)
24227 {
24228   __builtin_aarch64_simd_ci __o;
24229   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24230   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24231   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24232   return (uint8x8_t)__builtin_aarch64_qtbx3v8qi ((int8x8_t)r, __o,
24233                                                  (int8x8_t)idx);
24234 }
24235
24236 __extension__ extern __inline poly8x8_t
24237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24238 vqtbx3_p8 (poly8x8_t r, poly8x16x3_t tab, uint8x8_t idx)
24239 {
24240   __builtin_aarch64_simd_ci __o;
24241   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24242   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24243   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24244   return (poly8x8_t)__builtin_aarch64_qtbx3v8qi ((int8x8_t)r, __o,
24245                                                  (int8x8_t)idx);
24246 }
24247
24248 __extension__ extern __inline int8x16_t
24249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24250 vqtbx3q_s8 (int8x16_t r, int8x16x3_t tab, uint8x16_t idx)
24251 {
24252   __builtin_aarch64_simd_ci __o;
24253   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[0], 0);
24254   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[1], 1);
24255   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[2], 2);
24256   return __builtin_aarch64_qtbx3v16qi (r, __o, (int8x16_t)idx);
24257 }
24258
24259 __extension__ extern __inline uint8x16_t
24260 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24261 vqtbx3q_u8 (uint8x16_t r, uint8x16x3_t tab, uint8x16_t idx)
24262 {
24263   __builtin_aarch64_simd_ci __o;
24264   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24265   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24266   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24267   return (uint8x16_t)__builtin_aarch64_qtbx3v16qi ((int8x16_t)r, __o,
24268                                                    (int8x16_t)idx);
24269 }
24270
24271 __extension__ extern __inline poly8x16_t
24272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24273 vqtbx3q_p8 (poly8x16_t r, poly8x16x3_t tab, uint8x16_t idx)
24274 {
24275   __builtin_aarch64_simd_ci __o;
24276   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24277   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24278   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24279   return (poly8x16_t)__builtin_aarch64_qtbx3v16qi ((int8x16_t)r, __o,
24280                                                    (int8x16_t)idx);
24281 }
24282
24283 /* vqtbx4 */
24284
24285 __extension__ extern __inline int8x8_t
24286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24287 vqtbx4_s8 (int8x8_t r, int8x16x4_t tab, uint8x8_t idx)
24288 {
24289   __builtin_aarch64_simd_xi __o;
24290   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[0], 0);
24291   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[1], 1);
24292   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[2], 2);
24293   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[3], 3);
24294   return __builtin_aarch64_qtbx4v8qi (r, __o, (int8x8_t)idx);
24295 }
24296
24297 __extension__ extern __inline uint8x8_t
24298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24299 vqtbx4_u8 (uint8x8_t r, uint8x16x4_t tab, uint8x8_t idx)
24300 {
24301   __builtin_aarch64_simd_xi __o;
24302   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24303   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24304   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24305   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24306   return (uint8x8_t)__builtin_aarch64_qtbx4v8qi ((int8x8_t)r, __o,
24307                                                  (int8x8_t)idx);
24308 }
24309
24310 __extension__ extern __inline poly8x8_t
24311 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24312 vqtbx4_p8 (poly8x8_t r, poly8x16x4_t tab, uint8x8_t idx)
24313 {
24314   __builtin_aarch64_simd_xi __o;
24315   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24316   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24317   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24318   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24319   return (poly8x8_t)__builtin_aarch64_qtbx4v8qi ((int8x8_t)r, __o,
24320                                                  (int8x8_t)idx);
24321 }
24322
24323 __extension__ extern __inline int8x16_t
24324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24325 vqtbx4q_s8 (int8x16_t r, int8x16x4_t tab, uint8x16_t idx)
24326 {
24327   __builtin_aarch64_simd_xi __o;
24328   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[0], 0);
24329   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[1], 1);
24330   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[2], 2);
24331   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[3], 3);
24332   return __builtin_aarch64_qtbx4v16qi (r, __o, (int8x16_t)idx);
24333 }
24334
24335 __extension__ extern __inline uint8x16_t
24336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24337 vqtbx4q_u8 (uint8x16_t r, uint8x16x4_t tab, uint8x16_t idx)
24338 {
24339   __builtin_aarch64_simd_xi __o;
24340   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24341   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24342   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24343   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24344   return (uint8x16_t)__builtin_aarch64_qtbx4v16qi ((int8x16_t)r, __o,
24345                                                    (int8x16_t)idx);
24346 }
24347
24348 __extension__ extern __inline poly8x16_t
24349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24350 vqtbx4q_p8 (poly8x16_t r, poly8x16x4_t tab, uint8x16_t idx)
24351 {
24352   __builtin_aarch64_simd_xi __o;
24353   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24354   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24355   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24356   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24357   return (poly8x16_t)__builtin_aarch64_qtbx4v16qi ((int8x16_t)r, __o,
24358                                                    (int8x16_t)idx);
24359 }
24360
24361 /* vrbit  */
24362
24363 __extension__ extern __inline poly8x8_t
24364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24365 vrbit_p8 (poly8x8_t __a)
24366 {
24367   return (poly8x8_t) __builtin_aarch64_rbitv8qi ((int8x8_t) __a);
24368 }
24369
24370 __extension__ extern __inline int8x8_t
24371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24372 vrbit_s8 (int8x8_t __a)
24373 {
24374   return __builtin_aarch64_rbitv8qi (__a);
24375 }
24376
24377 __extension__ extern __inline uint8x8_t
24378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24379 vrbit_u8 (uint8x8_t __a)
24380 {
24381   return (uint8x8_t) __builtin_aarch64_rbitv8qi ((int8x8_t) __a);
24382 }
24383
24384 __extension__ extern __inline poly8x16_t
24385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24386 vrbitq_p8 (poly8x16_t __a)
24387 {
24388   return (poly8x16_t) __builtin_aarch64_rbitv16qi ((int8x16_t)__a);
24389 }
24390
24391 __extension__ extern __inline int8x16_t
24392 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24393 vrbitq_s8 (int8x16_t __a)
24394 {
24395   return __builtin_aarch64_rbitv16qi (__a);
24396 }
24397
24398 __extension__ extern __inline uint8x16_t
24399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24400 vrbitq_u8 (uint8x16_t __a)
24401 {
24402   return (uint8x16_t) __builtin_aarch64_rbitv16qi ((int8x16_t) __a);
24403 }
24404
24405 /* vrecpe  */
24406
24407 __extension__ extern __inline uint32x2_t
24408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24409 vrecpe_u32 (uint32x2_t __a)
24410 {
24411   return (uint32x2_t) __builtin_aarch64_urecpev2si ((int32x2_t) __a);
24412 }
24413
24414 __extension__ extern __inline uint32x4_t
24415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24416 vrecpeq_u32 (uint32x4_t __a)
24417 {
24418   return (uint32x4_t) __builtin_aarch64_urecpev4si ((int32x4_t) __a);
24419 }
24420
24421 __extension__ extern __inline float32_t
24422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24423 vrecpes_f32 (float32_t __a)
24424 {
24425   return __builtin_aarch64_frecpesf (__a);
24426 }
24427
24428 __extension__ extern __inline float64_t
24429 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24430 vrecped_f64 (float64_t __a)
24431 {
24432   return __builtin_aarch64_frecpedf (__a);
24433 }
24434
24435 __extension__ extern __inline float32x2_t
24436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24437 vrecpe_f32 (float32x2_t __a)
24438 {
24439   return __builtin_aarch64_frecpev2sf (__a);
24440 }
24441
24442 __extension__ extern __inline float64x1_t
24443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24444 vrecpe_f64 (float64x1_t __a)
24445 {
24446   return (float64x1_t) { vrecped_f64 (vget_lane_f64 (__a, 0)) };
24447 }
24448
24449 __extension__ extern __inline float32x4_t
24450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24451 vrecpeq_f32 (float32x4_t __a)
24452 {
24453   return __builtin_aarch64_frecpev4sf (__a);
24454 }
24455
24456 __extension__ extern __inline float64x2_t
24457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24458 vrecpeq_f64 (float64x2_t __a)
24459 {
24460   return __builtin_aarch64_frecpev2df (__a);
24461 }
24462
24463 /* vrecps  */
24464
24465 __extension__ extern __inline float32_t
24466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24467 vrecpss_f32 (float32_t __a, float32_t __b)
24468 {
24469   return __builtin_aarch64_frecpssf (__a, __b);
24470 }
24471
24472 __extension__ extern __inline float64_t
24473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24474 vrecpsd_f64 (float64_t __a, float64_t __b)
24475 {
24476   return __builtin_aarch64_frecpsdf (__a, __b);
24477 }
24478
24479 __extension__ extern __inline float32x2_t
24480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24481 vrecps_f32 (float32x2_t __a, float32x2_t __b)
24482 {
24483   return __builtin_aarch64_frecpsv2sf (__a, __b);
24484 }
24485
24486 __extension__ extern __inline float64x1_t
24487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24488 vrecps_f64 (float64x1_t __a, float64x1_t __b)
24489 {
24490   return (float64x1_t) { vrecpsd_f64 (vget_lane_f64 (__a, 0),
24491                                       vget_lane_f64 (__b, 0)) };
24492 }
24493
24494 __extension__ extern __inline float32x4_t
24495 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24496 vrecpsq_f32 (float32x4_t __a, float32x4_t __b)
24497 {
24498   return __builtin_aarch64_frecpsv4sf (__a, __b);
24499 }
24500
24501 __extension__ extern __inline float64x2_t
24502 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24503 vrecpsq_f64 (float64x2_t __a, float64x2_t __b)
24504 {
24505   return __builtin_aarch64_frecpsv2df (__a, __b);
24506 }
24507
24508 /* vrecpx  */
24509
24510 __extension__ extern __inline float32_t
24511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24512 vrecpxs_f32 (float32_t __a)
24513 {
24514   return __builtin_aarch64_frecpxsf (__a);
24515 }
24516
24517 __extension__ extern __inline float64_t
24518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24519 vrecpxd_f64 (float64_t __a)
24520 {
24521   return __builtin_aarch64_frecpxdf (__a);
24522 }
24523
24524
24525 /* vrev  */
24526
24527 __extension__ extern __inline poly8x8_t
24528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24529 vrev16_p8 (poly8x8_t a)
24530 {
24531   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24532 }
24533
24534 __extension__ extern __inline int8x8_t
24535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24536 vrev16_s8 (int8x8_t a)
24537 {
24538   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24539 }
24540
24541 __extension__ extern __inline uint8x8_t
24542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24543 vrev16_u8 (uint8x8_t a)
24544 {
24545   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24546 }
24547
24548 __extension__ extern __inline poly8x16_t
24549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24550 vrev16q_p8 (poly8x16_t a)
24551 {
24552   return __builtin_shuffle (a,
24553       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
24554 }
24555
24556 __extension__ extern __inline int8x16_t
24557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24558 vrev16q_s8 (int8x16_t a)
24559 {
24560   return __builtin_shuffle (a,
24561       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
24562 }
24563
24564 __extension__ extern __inline uint8x16_t
24565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24566 vrev16q_u8 (uint8x16_t a)
24567 {
24568   return __builtin_shuffle (a,
24569       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
24570 }
24571
24572 __extension__ extern __inline poly8x8_t
24573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24574 vrev32_p8 (poly8x8_t a)
24575 {
24576   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24577 }
24578
24579 __extension__ extern __inline poly16x4_t
24580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24581 vrev32_p16 (poly16x4_t a)
24582 {
24583   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
24584 }
24585
24586 __extension__ extern __inline int8x8_t
24587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24588 vrev32_s8 (int8x8_t a)
24589 {
24590   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24591 }
24592
24593 __extension__ extern __inline int16x4_t
24594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24595 vrev32_s16 (int16x4_t a)
24596 {
24597   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
24598 }
24599
24600 __extension__ extern __inline uint8x8_t
24601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24602 vrev32_u8 (uint8x8_t a)
24603 {
24604   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24605 }
24606
24607 __extension__ extern __inline uint16x4_t
24608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24609 vrev32_u16 (uint16x4_t a)
24610 {
24611   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
24612 }
24613
24614 __extension__ extern __inline poly8x16_t
24615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24616 vrev32q_p8 (poly8x16_t a)
24617 {
24618   return __builtin_shuffle (a,
24619       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
24620 }
24621
24622 __extension__ extern __inline poly16x8_t
24623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24624 vrev32q_p16 (poly16x8_t a)
24625 {
24626   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24627 }
24628
24629 __extension__ extern __inline int8x16_t
24630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24631 vrev32q_s8 (int8x16_t a)
24632 {
24633   return __builtin_shuffle (a,
24634       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
24635 }
24636
24637 __extension__ extern __inline int16x8_t
24638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24639 vrev32q_s16 (int16x8_t a)
24640 {
24641   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24642 }
24643
24644 __extension__ extern __inline uint8x16_t
24645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24646 vrev32q_u8 (uint8x16_t a)
24647 {
24648   return __builtin_shuffle (a,
24649       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
24650 }
24651
24652 __extension__ extern __inline uint16x8_t
24653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24654 vrev32q_u16 (uint16x8_t a)
24655 {
24656   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24657 }
24658
24659 __extension__ extern __inline float16x4_t
24660 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24661 vrev64_f16 (float16x4_t __a)
24662 {
24663   return __builtin_shuffle (__a, (uint16x4_t) { 3, 2, 1, 0 });
24664 }
24665
24666 __extension__ extern __inline float32x2_t
24667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24668 vrev64_f32 (float32x2_t a)
24669 {
24670   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
24671 }
24672
24673 __extension__ extern __inline poly8x8_t
24674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24675 vrev64_p8 (poly8x8_t a)
24676 {
24677   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
24678 }
24679
24680 __extension__ extern __inline poly16x4_t
24681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24682 vrev64_p16 (poly16x4_t a)
24683 {
24684   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
24685 }
24686
24687 __extension__ extern __inline int8x8_t
24688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24689 vrev64_s8 (int8x8_t a)
24690 {
24691   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
24692 }
24693
24694 __extension__ extern __inline int16x4_t
24695 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24696 vrev64_s16 (int16x4_t a)
24697 {
24698   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
24699 }
24700
24701 __extension__ extern __inline int32x2_t
24702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24703 vrev64_s32 (int32x2_t a)
24704 {
24705   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
24706 }
24707
24708 __extension__ extern __inline uint8x8_t
24709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24710 vrev64_u8 (uint8x8_t a)
24711 {
24712   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
24713 }
24714
24715 __extension__ extern __inline uint16x4_t
24716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24717 vrev64_u16 (uint16x4_t a)
24718 {
24719   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
24720 }
24721
24722 __extension__ extern __inline uint32x2_t
24723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24724 vrev64_u32 (uint32x2_t a)
24725 {
24726   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
24727 }
24728
24729 __extension__ extern __inline float16x8_t
24730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24731 vrev64q_f16 (float16x8_t __a)
24732 {
24733   return __builtin_shuffle (__a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24734 }
24735
24736 __extension__ extern __inline float32x4_t
24737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24738 vrev64q_f32 (float32x4_t a)
24739 {
24740   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
24741 }
24742
24743 __extension__ extern __inline poly8x16_t
24744 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24745 vrev64q_p8 (poly8x16_t a)
24746 {
24747   return __builtin_shuffle (a,
24748       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
24749 }
24750
24751 __extension__ extern __inline poly16x8_t
24752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24753 vrev64q_p16 (poly16x8_t a)
24754 {
24755   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24756 }
24757
24758 __extension__ extern __inline int8x16_t
24759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24760 vrev64q_s8 (int8x16_t a)
24761 {
24762   return __builtin_shuffle (a,
24763       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
24764 }
24765
24766 __extension__ extern __inline int16x8_t
24767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24768 vrev64q_s16 (int16x8_t a)
24769 {
24770   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24771 }
24772
24773 __extension__ extern __inline int32x4_t
24774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24775 vrev64q_s32 (int32x4_t a)
24776 {
24777   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
24778 }
24779
24780 __extension__ extern __inline uint8x16_t
24781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24782 vrev64q_u8 (uint8x16_t a)
24783 {
24784   return __builtin_shuffle (a,
24785       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
24786 }
24787
24788 __extension__ extern __inline uint16x8_t
24789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24790 vrev64q_u16 (uint16x8_t a)
24791 {
24792   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24793 }
24794
24795 __extension__ extern __inline uint32x4_t
24796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24797 vrev64q_u32 (uint32x4_t a)
24798 {
24799   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
24800 }
24801
24802 /* vrnd  */
24803
24804 __extension__ extern __inline float32x2_t
24805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24806 vrnd_f32 (float32x2_t __a)
24807 {
24808   return __builtin_aarch64_btruncv2sf (__a);
24809 }
24810
24811 __extension__ extern __inline float64x1_t
24812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24813 vrnd_f64 (float64x1_t __a)
24814 {
24815   return vset_lane_f64 (__builtin_trunc (vget_lane_f64 (__a, 0)), __a, 0);
24816 }
24817
24818 __extension__ extern __inline float32x4_t
24819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24820 vrndq_f32 (float32x4_t __a)
24821 {
24822   return __builtin_aarch64_btruncv4sf (__a);
24823 }
24824
24825 __extension__ extern __inline float64x2_t
24826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24827 vrndq_f64 (float64x2_t __a)
24828 {
24829   return __builtin_aarch64_btruncv2df (__a);
24830 }
24831
24832 /* vrnda  */
24833
24834 __extension__ extern __inline float32x2_t
24835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24836 vrnda_f32 (float32x2_t __a)
24837 {
24838   return __builtin_aarch64_roundv2sf (__a);
24839 }
24840
24841 __extension__ extern __inline float64x1_t
24842 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24843 vrnda_f64 (float64x1_t __a)
24844 {
24845   return vset_lane_f64 (__builtin_round (vget_lane_f64 (__a, 0)), __a, 0);
24846 }
24847
24848 __extension__ extern __inline float32x4_t
24849 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24850 vrndaq_f32 (float32x4_t __a)
24851 {
24852   return __builtin_aarch64_roundv4sf (__a);
24853 }
24854
24855 __extension__ extern __inline float64x2_t
24856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24857 vrndaq_f64 (float64x2_t __a)
24858 {
24859   return __builtin_aarch64_roundv2df (__a);
24860 }
24861
24862 /* vrndi  */
24863
24864 __extension__ extern __inline float32x2_t
24865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24866 vrndi_f32 (float32x2_t __a)
24867 {
24868   return __builtin_aarch64_nearbyintv2sf (__a);
24869 }
24870
24871 __extension__ extern __inline float64x1_t
24872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24873 vrndi_f64 (float64x1_t __a)
24874 {
24875   return vset_lane_f64 (__builtin_nearbyint (vget_lane_f64 (__a, 0)), __a, 0);
24876 }
24877
24878 __extension__ extern __inline float32x4_t
24879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24880 vrndiq_f32 (float32x4_t __a)
24881 {
24882   return __builtin_aarch64_nearbyintv4sf (__a);
24883 }
24884
24885 __extension__ extern __inline float64x2_t
24886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24887 vrndiq_f64 (float64x2_t __a)
24888 {
24889   return __builtin_aarch64_nearbyintv2df (__a);
24890 }
24891
24892 /* vrndm  */
24893
24894 __extension__ extern __inline float32x2_t
24895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24896 vrndm_f32 (float32x2_t __a)
24897 {
24898   return __builtin_aarch64_floorv2sf (__a);
24899 }
24900
24901 __extension__ extern __inline float64x1_t
24902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24903 vrndm_f64 (float64x1_t __a)
24904 {
24905   return vset_lane_f64 (__builtin_floor (vget_lane_f64 (__a, 0)), __a, 0);
24906 }
24907
24908 __extension__ extern __inline float32x4_t
24909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24910 vrndmq_f32 (float32x4_t __a)
24911 {
24912   return __builtin_aarch64_floorv4sf (__a);
24913 }
24914
24915 __extension__ extern __inline float64x2_t
24916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24917 vrndmq_f64 (float64x2_t __a)
24918 {
24919   return __builtin_aarch64_floorv2df (__a);
24920 }
24921
24922 /* vrndn  */
24923
24924 __extension__ extern __inline float32x2_t
24925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24926 vrndn_f32 (float32x2_t __a)
24927 {
24928   return __builtin_aarch64_frintnv2sf (__a);
24929 }
24930
24931 __extension__ extern __inline float64x1_t
24932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24933 vrndn_f64 (float64x1_t __a)
24934 {
24935   return (float64x1_t) {__builtin_aarch64_frintndf (__a[0])};
24936 }
24937
24938 __extension__ extern __inline float32x4_t
24939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24940 vrndnq_f32 (float32x4_t __a)
24941 {
24942   return __builtin_aarch64_frintnv4sf (__a);
24943 }
24944
24945 __extension__ extern __inline float64x2_t
24946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24947 vrndnq_f64 (float64x2_t __a)
24948 {
24949   return __builtin_aarch64_frintnv2df (__a);
24950 }
24951
24952 /* vrndp  */
24953
24954 __extension__ extern __inline float32x2_t
24955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24956 vrndp_f32 (float32x2_t __a)
24957 {
24958   return __builtin_aarch64_ceilv2sf (__a);
24959 }
24960
24961 __extension__ extern __inline float64x1_t
24962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24963 vrndp_f64 (float64x1_t __a)
24964 {
24965   return vset_lane_f64 (__builtin_ceil (vget_lane_f64 (__a, 0)), __a, 0);
24966 }
24967
24968 __extension__ extern __inline float32x4_t
24969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24970 vrndpq_f32 (float32x4_t __a)
24971 {
24972   return __builtin_aarch64_ceilv4sf (__a);
24973 }
24974
24975 __extension__ extern __inline float64x2_t
24976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24977 vrndpq_f64 (float64x2_t __a)
24978 {
24979   return __builtin_aarch64_ceilv2df (__a);
24980 }
24981
24982 /* vrndx  */
24983
24984 __extension__ extern __inline float32x2_t
24985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24986 vrndx_f32 (float32x2_t __a)
24987 {
24988   return __builtin_aarch64_rintv2sf (__a);
24989 }
24990
24991 __extension__ extern __inline float64x1_t
24992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24993 vrndx_f64 (float64x1_t __a)
24994 {
24995   return vset_lane_f64 (__builtin_rint (vget_lane_f64 (__a, 0)), __a, 0);
24996 }
24997
24998 __extension__ extern __inline float32x4_t
24999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25000 vrndxq_f32 (float32x4_t __a)
25001 {
25002   return __builtin_aarch64_rintv4sf (__a);
25003 }
25004
25005 __extension__ extern __inline float64x2_t
25006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25007 vrndxq_f64 (float64x2_t __a)
25008 {
25009   return __builtin_aarch64_rintv2df (__a);
25010 }
25011
25012 /* vrshl */
25013
25014 __extension__ extern __inline int8x8_t
25015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25016 vrshl_s8 (int8x8_t __a, int8x8_t __b)
25017 {
25018   return (int8x8_t) __builtin_aarch64_srshlv8qi (__a, __b);
25019 }
25020
25021 __extension__ extern __inline int16x4_t
25022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25023 vrshl_s16 (int16x4_t __a, int16x4_t __b)
25024 {
25025   return (int16x4_t) __builtin_aarch64_srshlv4hi (__a, __b);
25026 }
25027
25028 __extension__ extern __inline int32x2_t
25029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25030 vrshl_s32 (int32x2_t __a, int32x2_t __b)
25031 {
25032   return (int32x2_t) __builtin_aarch64_srshlv2si (__a, __b);
25033 }
25034
25035 __extension__ extern __inline int64x1_t
25036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25037 vrshl_s64 (int64x1_t __a, int64x1_t __b)
25038 {
25039   return (int64x1_t) {__builtin_aarch64_srshldi (__a[0], __b[0])};
25040 }
25041
25042 __extension__ extern __inline uint8x8_t
25043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25044 vrshl_u8 (uint8x8_t __a, int8x8_t __b)
25045 {
25046   return __builtin_aarch64_urshlv8qi_uus (__a, __b);
25047 }
25048
25049 __extension__ extern __inline uint16x4_t
25050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25051 vrshl_u16 (uint16x4_t __a, int16x4_t __b)
25052 {
25053   return __builtin_aarch64_urshlv4hi_uus (__a, __b);
25054 }
25055
25056 __extension__ extern __inline uint32x2_t
25057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25058 vrshl_u32 (uint32x2_t __a, int32x2_t __b)
25059 {
25060   return __builtin_aarch64_urshlv2si_uus (__a, __b);
25061 }
25062
25063 __extension__ extern __inline uint64x1_t
25064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25065 vrshl_u64 (uint64x1_t __a, int64x1_t __b)
25066 {
25067   return (uint64x1_t) {__builtin_aarch64_urshldi_uus (__a[0], __b[0])};
25068 }
25069
25070 __extension__ extern __inline int8x16_t
25071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25072 vrshlq_s8 (int8x16_t __a, int8x16_t __b)
25073 {
25074   return (int8x16_t) __builtin_aarch64_srshlv16qi (__a, __b);
25075 }
25076
25077 __extension__ extern __inline int16x8_t
25078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25079 vrshlq_s16 (int16x8_t __a, int16x8_t __b)
25080 {
25081   return (int16x8_t) __builtin_aarch64_srshlv8hi (__a, __b);
25082 }
25083
25084 __extension__ extern __inline int32x4_t
25085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25086 vrshlq_s32 (int32x4_t __a, int32x4_t __b)
25087 {
25088   return (int32x4_t) __builtin_aarch64_srshlv4si (__a, __b);
25089 }
25090
25091 __extension__ extern __inline int64x2_t
25092 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25093 vrshlq_s64 (int64x2_t __a, int64x2_t __b)
25094 {
25095   return (int64x2_t) __builtin_aarch64_srshlv2di (__a, __b);
25096 }
25097
25098 __extension__ extern __inline uint8x16_t
25099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25100 vrshlq_u8 (uint8x16_t __a, int8x16_t __b)
25101 {
25102   return __builtin_aarch64_urshlv16qi_uus (__a, __b);
25103 }
25104
25105 __extension__ extern __inline uint16x8_t
25106 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25107 vrshlq_u16 (uint16x8_t __a, int16x8_t __b)
25108 {
25109   return __builtin_aarch64_urshlv8hi_uus (__a, __b);
25110 }
25111
25112 __extension__ extern __inline uint32x4_t
25113 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25114 vrshlq_u32 (uint32x4_t __a, int32x4_t __b)
25115 {
25116   return __builtin_aarch64_urshlv4si_uus (__a, __b);
25117 }
25118
25119 __extension__ extern __inline uint64x2_t
25120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25121 vrshlq_u64 (uint64x2_t __a, int64x2_t __b)
25122 {
25123   return __builtin_aarch64_urshlv2di_uus (__a, __b);
25124 }
25125
25126 __extension__ extern __inline int64_t
25127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25128 vrshld_s64 (int64_t __a, int64_t __b)
25129 {
25130   return __builtin_aarch64_srshldi (__a, __b);
25131 }
25132
25133 __extension__ extern __inline uint64_t
25134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25135 vrshld_u64 (uint64_t __a, int64_t __b)
25136 {
25137   return __builtin_aarch64_urshldi_uus (__a, __b);
25138 }
25139
25140 /* vrshr */
25141
25142 __extension__ extern __inline int8x8_t
25143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25144 vrshr_n_s8 (int8x8_t __a, const int __b)
25145 {
25146   return (int8x8_t) __builtin_aarch64_srshr_nv8qi (__a, __b);
25147 }
25148
25149 __extension__ extern __inline int16x4_t
25150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25151 vrshr_n_s16 (int16x4_t __a, const int __b)
25152 {
25153   return (int16x4_t) __builtin_aarch64_srshr_nv4hi (__a, __b);
25154 }
25155
25156 __extension__ extern __inline int32x2_t
25157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25158 vrshr_n_s32 (int32x2_t __a, const int __b)
25159 {
25160   return (int32x2_t) __builtin_aarch64_srshr_nv2si (__a, __b);
25161 }
25162
25163 __extension__ extern __inline int64x1_t
25164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25165 vrshr_n_s64 (int64x1_t __a, const int __b)
25166 {
25167   return (int64x1_t) {__builtin_aarch64_srshr_ndi (__a[0], __b)};
25168 }
25169
25170 __extension__ extern __inline uint8x8_t
25171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25172 vrshr_n_u8 (uint8x8_t __a, const int __b)
25173 {
25174   return __builtin_aarch64_urshr_nv8qi_uus (__a, __b);
25175 }
25176
25177 __extension__ extern __inline uint16x4_t
25178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25179 vrshr_n_u16 (uint16x4_t __a, const int __b)
25180 {
25181   return __builtin_aarch64_urshr_nv4hi_uus (__a, __b);
25182 }
25183
25184 __extension__ extern __inline uint32x2_t
25185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25186 vrshr_n_u32 (uint32x2_t __a, const int __b)
25187 {
25188   return __builtin_aarch64_urshr_nv2si_uus (__a, __b);
25189 }
25190
25191 __extension__ extern __inline uint64x1_t
25192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25193 vrshr_n_u64 (uint64x1_t __a, const int __b)
25194 {
25195   return (uint64x1_t) {__builtin_aarch64_urshr_ndi_uus (__a[0], __b)};
25196 }
25197
25198 __extension__ extern __inline int8x16_t
25199 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25200 vrshrq_n_s8 (int8x16_t __a, const int __b)
25201 {
25202   return (int8x16_t) __builtin_aarch64_srshr_nv16qi (__a, __b);
25203 }
25204
25205 __extension__ extern __inline int16x8_t
25206 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25207 vrshrq_n_s16 (int16x8_t __a, const int __b)
25208 {
25209   return (int16x8_t) __builtin_aarch64_srshr_nv8hi (__a, __b);
25210 }
25211
25212 __extension__ extern __inline int32x4_t
25213 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25214 vrshrq_n_s32 (int32x4_t __a, const int __b)
25215 {
25216   return (int32x4_t) __builtin_aarch64_srshr_nv4si (__a, __b);
25217 }
25218
25219 __extension__ extern __inline int64x2_t
25220 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25221 vrshrq_n_s64 (int64x2_t __a, const int __b)
25222 {
25223   return (int64x2_t) __builtin_aarch64_srshr_nv2di (__a, __b);
25224 }
25225
25226 __extension__ extern __inline uint8x16_t
25227 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25228 vrshrq_n_u8 (uint8x16_t __a, const int __b)
25229 {
25230   return __builtin_aarch64_urshr_nv16qi_uus (__a, __b);
25231 }
25232
25233 __extension__ extern __inline uint16x8_t
25234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25235 vrshrq_n_u16 (uint16x8_t __a, const int __b)
25236 {
25237   return __builtin_aarch64_urshr_nv8hi_uus (__a, __b);
25238 }
25239
25240 __extension__ extern __inline uint32x4_t
25241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25242 vrshrq_n_u32 (uint32x4_t __a, const int __b)
25243 {
25244   return __builtin_aarch64_urshr_nv4si_uus (__a, __b);
25245 }
25246
25247 __extension__ extern __inline uint64x2_t
25248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25249 vrshrq_n_u64 (uint64x2_t __a, const int __b)
25250 {
25251   return __builtin_aarch64_urshr_nv2di_uus (__a, __b);
25252 }
25253
25254 __extension__ extern __inline int64_t
25255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25256 vrshrd_n_s64 (int64_t __a, const int __b)
25257 {
25258   return __builtin_aarch64_srshr_ndi (__a, __b);
25259 }
25260
25261 __extension__ extern __inline uint64_t
25262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25263 vrshrd_n_u64 (uint64_t __a, const int __b)
25264 {
25265   return __builtin_aarch64_urshr_ndi_uus (__a, __b);
25266 }
25267
25268 /* vrsqrte.  */
25269
25270 __extension__ extern __inline float32_t
25271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25272 vrsqrtes_f32 (float32_t __a)
25273 {
25274   return __builtin_aarch64_rsqrtesf (__a);
25275 }
25276
25277 __extension__ extern __inline float64_t
25278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25279 vrsqrted_f64 (float64_t __a)
25280 {
25281   return __builtin_aarch64_rsqrtedf (__a);
25282 }
25283
25284 __extension__ extern __inline float32x2_t
25285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25286 vrsqrte_f32 (float32x2_t __a)
25287 {
25288   return __builtin_aarch64_rsqrtev2sf (__a);
25289 }
25290
25291 __extension__ extern __inline float64x1_t
25292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25293 vrsqrte_f64 (float64x1_t __a)
25294 {
25295   return (float64x1_t) {vrsqrted_f64 (vget_lane_f64 (__a, 0))};
25296 }
25297
25298 __extension__ extern __inline float32x4_t
25299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25300 vrsqrteq_f32 (float32x4_t __a)
25301 {
25302   return __builtin_aarch64_rsqrtev4sf (__a);
25303 }
25304
25305 __extension__ extern __inline float64x2_t
25306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25307 vrsqrteq_f64 (float64x2_t __a)
25308 {
25309   return __builtin_aarch64_rsqrtev2df (__a);
25310 }
25311
25312 /* vrsqrts.  */
25313
25314 __extension__ extern __inline float32_t
25315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25316 vrsqrtss_f32 (float32_t __a, float32_t __b)
25317 {
25318   return __builtin_aarch64_rsqrtssf (__a, __b);
25319 }
25320
25321 __extension__ extern __inline float64_t
25322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25323 vrsqrtsd_f64 (float64_t __a, float64_t __b)
25324 {
25325   return __builtin_aarch64_rsqrtsdf (__a, __b);
25326 }
25327
25328 __extension__ extern __inline float32x2_t
25329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25330 vrsqrts_f32 (float32x2_t __a, float32x2_t __b)
25331 {
25332   return __builtin_aarch64_rsqrtsv2sf (__a, __b);
25333 }
25334
25335 __extension__ extern __inline float64x1_t
25336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25337 vrsqrts_f64 (float64x1_t __a, float64x1_t __b)
25338 {
25339   return (float64x1_t) {vrsqrtsd_f64 (vget_lane_f64 (__a, 0),
25340                                       vget_lane_f64 (__b, 0))};
25341 }
25342
25343 __extension__ extern __inline float32x4_t
25344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25345 vrsqrtsq_f32 (float32x4_t __a, float32x4_t __b)
25346 {
25347   return __builtin_aarch64_rsqrtsv4sf (__a, __b);
25348 }
25349
25350 __extension__ extern __inline float64x2_t
25351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25352 vrsqrtsq_f64 (float64x2_t __a, float64x2_t __b)
25353 {
25354   return __builtin_aarch64_rsqrtsv2df (__a, __b);
25355 }
25356
25357 /* vrsra */
25358
25359 __extension__ extern __inline int8x8_t
25360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25361 vrsra_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
25362 {
25363   return (int8x8_t) __builtin_aarch64_srsra_nv8qi (__a, __b, __c);
25364 }
25365
25366 __extension__ extern __inline int16x4_t
25367 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25368 vrsra_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
25369 {
25370   return (int16x4_t) __builtin_aarch64_srsra_nv4hi (__a, __b, __c);
25371 }
25372
25373 __extension__ extern __inline int32x2_t
25374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25375 vrsra_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
25376 {
25377   return (int32x2_t) __builtin_aarch64_srsra_nv2si (__a, __b, __c);
25378 }
25379
25380 __extension__ extern __inline int64x1_t
25381 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25382 vrsra_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
25383 {
25384   return (int64x1_t) {__builtin_aarch64_srsra_ndi (__a[0], __b[0], __c)};
25385 }
25386
25387 __extension__ extern __inline uint8x8_t
25388 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25389 vrsra_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
25390 {
25391   return __builtin_aarch64_ursra_nv8qi_uuus (__a, __b, __c);
25392 }
25393
25394 __extension__ extern __inline uint16x4_t
25395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25396 vrsra_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
25397 {
25398   return __builtin_aarch64_ursra_nv4hi_uuus (__a, __b, __c);
25399 }
25400
25401 __extension__ extern __inline uint32x2_t
25402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25403 vrsra_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
25404 {
25405   return __builtin_aarch64_ursra_nv2si_uuus (__a, __b, __c);
25406 }
25407
25408 __extension__ extern __inline uint64x1_t
25409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25410 vrsra_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
25411 {
25412   return (uint64x1_t) {__builtin_aarch64_ursra_ndi_uuus (__a[0], __b[0], __c)};
25413 }
25414
25415 __extension__ extern __inline int8x16_t
25416 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25417 vrsraq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
25418 {
25419   return (int8x16_t) __builtin_aarch64_srsra_nv16qi (__a, __b, __c);
25420 }
25421
25422 __extension__ extern __inline int16x8_t
25423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25424 vrsraq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
25425 {
25426   return (int16x8_t) __builtin_aarch64_srsra_nv8hi (__a, __b, __c);
25427 }
25428
25429 __extension__ extern __inline int32x4_t
25430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25431 vrsraq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
25432 {
25433   return (int32x4_t) __builtin_aarch64_srsra_nv4si (__a, __b, __c);
25434 }
25435
25436 __extension__ extern __inline int64x2_t
25437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25438 vrsraq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
25439 {
25440   return (int64x2_t) __builtin_aarch64_srsra_nv2di (__a, __b, __c);
25441 }
25442
25443 __extension__ extern __inline uint8x16_t
25444 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25445 vrsraq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
25446 {
25447   return __builtin_aarch64_ursra_nv16qi_uuus (__a, __b, __c);
25448 }
25449
25450 __extension__ extern __inline uint16x8_t
25451 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25452 vrsraq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
25453 {
25454   return __builtin_aarch64_ursra_nv8hi_uuus (__a, __b, __c);
25455 }
25456
25457 __extension__ extern __inline uint32x4_t
25458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25459 vrsraq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
25460 {
25461   return __builtin_aarch64_ursra_nv4si_uuus (__a, __b, __c);
25462 }
25463
25464 __extension__ extern __inline uint64x2_t
25465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25466 vrsraq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
25467 {
25468   return __builtin_aarch64_ursra_nv2di_uuus (__a, __b, __c);
25469 }
25470
25471 __extension__ extern __inline int64_t
25472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25473 vrsrad_n_s64 (int64_t __a, int64_t __b, const int __c)
25474 {
25475   return __builtin_aarch64_srsra_ndi (__a, __b, __c);
25476 }
25477
25478 __extension__ extern __inline uint64_t
25479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25480 vrsrad_n_u64 (uint64_t __a, uint64_t __b, const int __c)
25481 {
25482   return __builtin_aarch64_ursra_ndi_uuus (__a, __b, __c);
25483 }
25484
25485 #pragma GCC push_options
25486 #pragma GCC target ("+nothing+crypto")
25487
25488 /* vsha1  */
25489
25490 __extension__ extern __inline uint32x4_t
25491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25492 vsha1cq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
25493 {
25494   return __builtin_aarch64_crypto_sha1cv4si_uuuu (hash_abcd, hash_e, wk);
25495 }
25496
25497 __extension__ extern __inline uint32x4_t
25498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25499 vsha1mq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
25500 {
25501   return __builtin_aarch64_crypto_sha1mv4si_uuuu (hash_abcd, hash_e, wk);
25502 }
25503
25504 __extension__ extern __inline uint32x4_t
25505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25506 vsha1pq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
25507 {
25508   return __builtin_aarch64_crypto_sha1pv4si_uuuu (hash_abcd, hash_e, wk);
25509 }
25510
25511 __extension__ extern __inline uint32_t
25512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25513 vsha1h_u32 (uint32_t hash_e)
25514 {
25515   return __builtin_aarch64_crypto_sha1hsi_uu (hash_e);
25516 }
25517
25518 __extension__ extern __inline uint32x4_t
25519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25520 vsha1su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7, uint32x4_t w8_11)
25521 {
25522   return __builtin_aarch64_crypto_sha1su0v4si_uuuu (w0_3, w4_7, w8_11);
25523 }
25524
25525 __extension__ extern __inline uint32x4_t
25526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25527 vsha1su1q_u32 (uint32x4_t tw0_3, uint32x4_t w12_15)
25528 {
25529   return __builtin_aarch64_crypto_sha1su1v4si_uuu (tw0_3, w12_15);
25530 }
25531
25532 __extension__ extern __inline uint32x4_t
25533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25534 vsha256hq_u32 (uint32x4_t hash_abcd, uint32x4_t hash_efgh, uint32x4_t wk)
25535 {
25536   return __builtin_aarch64_crypto_sha256hv4si_uuuu (hash_abcd, hash_efgh, wk);
25537 }
25538
25539 __extension__ extern __inline uint32x4_t
25540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25541 vsha256h2q_u32 (uint32x4_t hash_efgh, uint32x4_t hash_abcd, uint32x4_t wk)
25542 {
25543   return __builtin_aarch64_crypto_sha256h2v4si_uuuu (hash_efgh, hash_abcd, wk);
25544 }
25545
25546 __extension__ extern __inline uint32x4_t
25547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25548 vsha256su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7)
25549 {
25550   return __builtin_aarch64_crypto_sha256su0v4si_uuu (w0_3, w4_7);
25551 }
25552
25553 __extension__ extern __inline uint32x4_t
25554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25555 vsha256su1q_u32 (uint32x4_t tw0_3, uint32x4_t w8_11, uint32x4_t w12_15)
25556 {
25557   return __builtin_aarch64_crypto_sha256su1v4si_uuuu (tw0_3, w8_11, w12_15);
25558 }
25559
25560 __extension__ extern __inline poly128_t
25561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25562 vmull_p64 (poly64_t a, poly64_t b)
25563 {
25564   return
25565     __builtin_aarch64_crypto_pmulldi_ppp (a, b);
25566 }
25567
25568 __extension__ extern __inline poly128_t
25569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25570 vmull_high_p64 (poly64x2_t a, poly64x2_t b)
25571 {
25572   return __builtin_aarch64_crypto_pmullv2di_ppp (a, b);
25573 }
25574
25575 #pragma GCC pop_options
25576
25577 /* vshl */
25578
25579 __extension__ extern __inline int8x8_t
25580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25581 vshl_n_s8 (int8x8_t __a, const int __b)
25582 {
25583   return (int8x8_t) __builtin_aarch64_ashlv8qi (__a, __b);
25584 }
25585
25586 __extension__ extern __inline int16x4_t
25587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25588 vshl_n_s16 (int16x4_t __a, const int __b)
25589 {
25590   return (int16x4_t) __builtin_aarch64_ashlv4hi (__a, __b);
25591 }
25592
25593 __extension__ extern __inline int32x2_t
25594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25595 vshl_n_s32 (int32x2_t __a, const int __b)
25596 {
25597   return (int32x2_t) __builtin_aarch64_ashlv2si (__a, __b);
25598 }
25599
25600 __extension__ extern __inline int64x1_t
25601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25602 vshl_n_s64 (int64x1_t __a, const int __b)
25603 {
25604   return (int64x1_t) {__builtin_aarch64_ashldi (__a[0], __b)};
25605 }
25606
25607 __extension__ extern __inline uint8x8_t
25608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25609 vshl_n_u8 (uint8x8_t __a, const int __b)
25610 {
25611   return (uint8x8_t) __builtin_aarch64_ashlv8qi ((int8x8_t) __a, __b);
25612 }
25613
25614 __extension__ extern __inline uint16x4_t
25615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25616 vshl_n_u16 (uint16x4_t __a, const int __b)
25617 {
25618   return (uint16x4_t) __builtin_aarch64_ashlv4hi ((int16x4_t) __a, __b);
25619 }
25620
25621 __extension__ extern __inline uint32x2_t
25622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25623 vshl_n_u32 (uint32x2_t __a, const int __b)
25624 {
25625   return (uint32x2_t) __builtin_aarch64_ashlv2si ((int32x2_t) __a, __b);
25626 }
25627
25628 __extension__ extern __inline uint64x1_t
25629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25630 vshl_n_u64 (uint64x1_t __a, const int __b)
25631 {
25632   return (uint64x1_t) {__builtin_aarch64_ashldi ((int64_t) __a[0], __b)};
25633 }
25634
25635 __extension__ extern __inline int8x16_t
25636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25637 vshlq_n_s8 (int8x16_t __a, const int __b)
25638 {
25639   return (int8x16_t) __builtin_aarch64_ashlv16qi (__a, __b);
25640 }
25641
25642 __extension__ extern __inline int16x8_t
25643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25644 vshlq_n_s16 (int16x8_t __a, const int __b)
25645 {
25646   return (int16x8_t) __builtin_aarch64_ashlv8hi (__a, __b);
25647 }
25648
25649 __extension__ extern __inline int32x4_t
25650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25651 vshlq_n_s32 (int32x4_t __a, const int __b)
25652 {
25653   return (int32x4_t) __builtin_aarch64_ashlv4si (__a, __b);
25654 }
25655
25656 __extension__ extern __inline int64x2_t
25657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25658 vshlq_n_s64 (int64x2_t __a, const int __b)
25659 {
25660   return (int64x2_t) __builtin_aarch64_ashlv2di (__a, __b);
25661 }
25662
25663 __extension__ extern __inline uint8x16_t
25664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25665 vshlq_n_u8 (uint8x16_t __a, const int __b)
25666 {
25667   return (uint8x16_t) __builtin_aarch64_ashlv16qi ((int8x16_t) __a, __b);
25668 }
25669
25670 __extension__ extern __inline uint16x8_t
25671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25672 vshlq_n_u16 (uint16x8_t __a, const int __b)
25673 {
25674   return (uint16x8_t) __builtin_aarch64_ashlv8hi ((int16x8_t) __a, __b);
25675 }
25676
25677 __extension__ extern __inline uint32x4_t
25678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25679 vshlq_n_u32 (uint32x4_t __a, const int __b)
25680 {
25681   return (uint32x4_t) __builtin_aarch64_ashlv4si ((int32x4_t) __a, __b);
25682 }
25683
25684 __extension__ extern __inline uint64x2_t
25685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25686 vshlq_n_u64 (uint64x2_t __a, const int __b)
25687 {
25688   return (uint64x2_t) __builtin_aarch64_ashlv2di ((int64x2_t) __a, __b);
25689 }
25690
25691 __extension__ extern __inline int64_t
25692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25693 vshld_n_s64 (int64_t __a, const int __b)
25694 {
25695   return __builtin_aarch64_ashldi (__a, __b);
25696 }
25697
25698 __extension__ extern __inline uint64_t
25699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25700 vshld_n_u64 (uint64_t __a, const int __b)
25701 {
25702   return (uint64_t) __builtin_aarch64_ashldi (__a, __b);
25703 }
25704
25705 __extension__ extern __inline int8x8_t
25706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25707 vshl_s8 (int8x8_t __a, int8x8_t __b)
25708 {
25709   return __builtin_aarch64_sshlv8qi (__a, __b);
25710 }
25711
25712 __extension__ extern __inline int16x4_t
25713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25714 vshl_s16 (int16x4_t __a, int16x4_t __b)
25715 {
25716   return __builtin_aarch64_sshlv4hi (__a, __b);
25717 }
25718
25719 __extension__ extern __inline int32x2_t
25720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25721 vshl_s32 (int32x2_t __a, int32x2_t __b)
25722 {
25723   return __builtin_aarch64_sshlv2si (__a, __b);
25724 }
25725
25726 __extension__ extern __inline int64x1_t
25727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25728 vshl_s64 (int64x1_t __a, int64x1_t __b)
25729 {
25730   return (int64x1_t) {__builtin_aarch64_sshldi (__a[0], __b[0])};
25731 }
25732
25733 __extension__ extern __inline uint8x8_t
25734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25735 vshl_u8 (uint8x8_t __a, int8x8_t __b)
25736 {
25737   return __builtin_aarch64_ushlv8qi_uus (__a, __b);
25738 }
25739
25740 __extension__ extern __inline uint16x4_t
25741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25742 vshl_u16 (uint16x4_t __a, int16x4_t __b)
25743 {
25744   return __builtin_aarch64_ushlv4hi_uus (__a, __b);
25745 }
25746
25747 __extension__ extern __inline uint32x2_t
25748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25749 vshl_u32 (uint32x2_t __a, int32x2_t __b)
25750 {
25751   return __builtin_aarch64_ushlv2si_uus (__a, __b);
25752 }
25753
25754 __extension__ extern __inline uint64x1_t
25755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25756 vshl_u64 (uint64x1_t __a, int64x1_t __b)
25757 {
25758   return (uint64x1_t) {__builtin_aarch64_ushldi_uus (__a[0], __b[0])};
25759 }
25760
25761 __extension__ extern __inline int8x16_t
25762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25763 vshlq_s8 (int8x16_t __a, int8x16_t __b)
25764 {
25765   return __builtin_aarch64_sshlv16qi (__a, __b);
25766 }
25767
25768 __extension__ extern __inline int16x8_t
25769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25770 vshlq_s16 (int16x8_t __a, int16x8_t __b)
25771 {
25772   return __builtin_aarch64_sshlv8hi (__a, __b);
25773 }
25774
25775 __extension__ extern __inline int32x4_t
25776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25777 vshlq_s32 (int32x4_t __a, int32x4_t __b)
25778 {
25779   return __builtin_aarch64_sshlv4si (__a, __b);
25780 }
25781
25782 __extension__ extern __inline int64x2_t
25783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25784 vshlq_s64 (int64x2_t __a, int64x2_t __b)
25785 {
25786   return __builtin_aarch64_sshlv2di (__a, __b);
25787 }
25788
25789 __extension__ extern __inline uint8x16_t
25790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25791 vshlq_u8 (uint8x16_t __a, int8x16_t __b)
25792 {
25793   return __builtin_aarch64_ushlv16qi_uus (__a, __b);
25794 }
25795
25796 __extension__ extern __inline uint16x8_t
25797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25798 vshlq_u16 (uint16x8_t __a, int16x8_t __b)
25799 {
25800   return __builtin_aarch64_ushlv8hi_uus (__a, __b);
25801 }
25802
25803 __extension__ extern __inline uint32x4_t
25804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25805 vshlq_u32 (uint32x4_t __a, int32x4_t __b)
25806 {
25807   return __builtin_aarch64_ushlv4si_uus (__a, __b);
25808 }
25809
25810 __extension__ extern __inline uint64x2_t
25811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25812 vshlq_u64 (uint64x2_t __a, int64x2_t __b)
25813 {
25814   return __builtin_aarch64_ushlv2di_uus (__a, __b);
25815 }
25816
25817 __extension__ extern __inline int64_t
25818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25819 vshld_s64 (int64_t __a, int64_t __b)
25820 {
25821   return __builtin_aarch64_sshldi (__a, __b);
25822 }
25823
25824 __extension__ extern __inline uint64_t
25825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25826 vshld_u64 (uint64_t __a, uint64_t __b)
25827 {
25828   return __builtin_aarch64_ushldi_uus (__a, __b);
25829 }
25830
25831 __extension__ extern __inline int16x8_t
25832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25833 vshll_high_n_s8 (int8x16_t __a, const int __b)
25834 {
25835   return __builtin_aarch64_sshll2_nv16qi (__a, __b);
25836 }
25837
25838 __extension__ extern __inline int32x4_t
25839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25840 vshll_high_n_s16 (int16x8_t __a, const int __b)
25841 {
25842   return __builtin_aarch64_sshll2_nv8hi (__a, __b);
25843 }
25844
25845 __extension__ extern __inline int64x2_t
25846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25847 vshll_high_n_s32 (int32x4_t __a, const int __b)
25848 {
25849   return __builtin_aarch64_sshll2_nv4si (__a, __b);
25850 }
25851
25852 __extension__ extern __inline uint16x8_t
25853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25854 vshll_high_n_u8 (uint8x16_t __a, const int __b)
25855 {
25856   return (uint16x8_t) __builtin_aarch64_ushll2_nv16qi ((int8x16_t) __a, __b);
25857 }
25858
25859 __extension__ extern __inline uint32x4_t
25860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25861 vshll_high_n_u16 (uint16x8_t __a, const int __b)
25862 {
25863   return (uint32x4_t) __builtin_aarch64_ushll2_nv8hi ((int16x8_t) __a, __b);
25864 }
25865
25866 __extension__ extern __inline uint64x2_t
25867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25868 vshll_high_n_u32 (uint32x4_t __a, const int __b)
25869 {
25870   return (uint64x2_t) __builtin_aarch64_ushll2_nv4si ((int32x4_t) __a, __b);
25871 }
25872
25873 __extension__ extern __inline int16x8_t
25874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25875 vshll_n_s8 (int8x8_t __a, const int __b)
25876 {
25877   return __builtin_aarch64_sshll_nv8qi (__a, __b);
25878 }
25879
25880 __extension__ extern __inline int32x4_t
25881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25882 vshll_n_s16 (int16x4_t __a, const int __b)
25883 {
25884   return __builtin_aarch64_sshll_nv4hi (__a, __b);
25885 }
25886
25887 __extension__ extern __inline int64x2_t
25888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25889 vshll_n_s32 (int32x2_t __a, const int __b)
25890 {
25891   return __builtin_aarch64_sshll_nv2si (__a, __b);
25892 }
25893
25894 __extension__ extern __inline uint16x8_t
25895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25896 vshll_n_u8 (uint8x8_t __a, const int __b)
25897 {
25898   return __builtin_aarch64_ushll_nv8qi_uus (__a, __b);
25899 }
25900
25901 __extension__ extern __inline uint32x4_t
25902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25903 vshll_n_u16 (uint16x4_t __a, const int __b)
25904 {
25905   return __builtin_aarch64_ushll_nv4hi_uus (__a, __b);
25906 }
25907
25908 __extension__ extern __inline uint64x2_t
25909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25910 vshll_n_u32 (uint32x2_t __a, const int __b)
25911 {
25912   return __builtin_aarch64_ushll_nv2si_uus (__a, __b);
25913 }
25914
25915 /* vshr */
25916
25917 __extension__ extern __inline int8x8_t
25918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25919 vshr_n_s8 (int8x8_t __a, const int __b)
25920 {
25921   return (int8x8_t) __builtin_aarch64_ashrv8qi (__a, __b);
25922 }
25923
25924 __extension__ extern __inline int16x4_t
25925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25926 vshr_n_s16 (int16x4_t __a, const int __b)
25927 {
25928   return (int16x4_t) __builtin_aarch64_ashrv4hi (__a, __b);
25929 }
25930
25931 __extension__ extern __inline int32x2_t
25932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25933 vshr_n_s32 (int32x2_t __a, const int __b)
25934 {
25935   return (int32x2_t) __builtin_aarch64_ashrv2si (__a, __b);
25936 }
25937
25938 __extension__ extern __inline int64x1_t
25939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25940 vshr_n_s64 (int64x1_t __a, const int __b)
25941 {
25942   return (int64x1_t) {__builtin_aarch64_ashr_simddi (__a[0], __b)};
25943 }
25944
25945 __extension__ extern __inline uint8x8_t
25946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25947 vshr_n_u8 (uint8x8_t __a, const int __b)
25948 {
25949   return (uint8x8_t) __builtin_aarch64_lshrv8qi ((int8x8_t) __a, __b);
25950 }
25951
25952 __extension__ extern __inline uint16x4_t
25953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25954 vshr_n_u16 (uint16x4_t __a, const int __b)
25955 {
25956   return (uint16x4_t) __builtin_aarch64_lshrv4hi ((int16x4_t) __a, __b);
25957 }
25958
25959 __extension__ extern __inline uint32x2_t
25960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25961 vshr_n_u32 (uint32x2_t __a, const int __b)
25962 {
25963   return (uint32x2_t) __builtin_aarch64_lshrv2si ((int32x2_t) __a, __b);
25964 }
25965
25966 __extension__ extern __inline uint64x1_t
25967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25968 vshr_n_u64 (uint64x1_t __a, const int __b)
25969 {
25970   return (uint64x1_t) {__builtin_aarch64_lshr_simddi_uus ( __a[0], __b)};
25971 }
25972
25973 __extension__ extern __inline int8x16_t
25974 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25975 vshrq_n_s8 (int8x16_t __a, const int __b)
25976 {
25977   return (int8x16_t) __builtin_aarch64_ashrv16qi (__a, __b);
25978 }
25979
25980 __extension__ extern __inline int16x8_t
25981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25982 vshrq_n_s16 (int16x8_t __a, const int __b)
25983 {
25984   return (int16x8_t) __builtin_aarch64_ashrv8hi (__a, __b);
25985 }
25986
25987 __extension__ extern __inline int32x4_t
25988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25989 vshrq_n_s32 (int32x4_t __a, const int __b)
25990 {
25991   return (int32x4_t) __builtin_aarch64_ashrv4si (__a, __b);
25992 }
25993
25994 __extension__ extern __inline int64x2_t
25995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25996 vshrq_n_s64 (int64x2_t __a, const int __b)
25997 {
25998   return (int64x2_t) __builtin_aarch64_ashrv2di (__a, __b);
25999 }
26000
26001 __extension__ extern __inline uint8x16_t
26002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26003 vshrq_n_u8 (uint8x16_t __a, const int __b)
26004 {
26005   return (uint8x16_t) __builtin_aarch64_lshrv16qi ((int8x16_t) __a, __b);
26006 }
26007
26008 __extension__ extern __inline uint16x8_t
26009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26010 vshrq_n_u16 (uint16x8_t __a, const int __b)
26011 {
26012   return (uint16x8_t) __builtin_aarch64_lshrv8hi ((int16x8_t) __a, __b);
26013 }
26014
26015 __extension__ extern __inline uint32x4_t
26016 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26017 vshrq_n_u32 (uint32x4_t __a, const int __b)
26018 {
26019   return (uint32x4_t) __builtin_aarch64_lshrv4si ((int32x4_t) __a, __b);
26020 }
26021
26022 __extension__ extern __inline uint64x2_t
26023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26024 vshrq_n_u64 (uint64x2_t __a, const int __b)
26025 {
26026   return (uint64x2_t) __builtin_aarch64_lshrv2di ((int64x2_t) __a, __b);
26027 }
26028
26029 __extension__ extern __inline int64_t
26030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26031 vshrd_n_s64 (int64_t __a, const int __b)
26032 {
26033   return __builtin_aarch64_ashr_simddi (__a, __b);
26034 }
26035
26036 __extension__ extern __inline uint64_t
26037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26038 vshrd_n_u64 (uint64_t __a, const int __b)
26039 {
26040   return __builtin_aarch64_lshr_simddi_uus (__a, __b);
26041 }
26042
26043 /* vsli */
26044
26045 __extension__ extern __inline int8x8_t
26046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26047 vsli_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26048 {
26049   return (int8x8_t) __builtin_aarch64_ssli_nv8qi (__a, __b, __c);
26050 }
26051
26052 __extension__ extern __inline int16x4_t
26053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26054 vsli_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26055 {
26056   return (int16x4_t) __builtin_aarch64_ssli_nv4hi (__a, __b, __c);
26057 }
26058
26059 __extension__ extern __inline int32x2_t
26060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26061 vsli_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26062 {
26063   return (int32x2_t) __builtin_aarch64_ssli_nv2si (__a, __b, __c);
26064 }
26065
26066 __extension__ extern __inline int64x1_t
26067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26068 vsli_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26069 {
26070   return (int64x1_t) {__builtin_aarch64_ssli_ndi (__a[0], __b[0], __c)};
26071 }
26072
26073 __extension__ extern __inline uint8x8_t
26074 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26075 vsli_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26076 {
26077   return __builtin_aarch64_usli_nv8qi_uuus (__a, __b, __c);
26078 }
26079
26080 __extension__ extern __inline uint16x4_t
26081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26082 vsli_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26083 {
26084   return __builtin_aarch64_usli_nv4hi_uuus (__a, __b, __c);
26085 }
26086
26087 __extension__ extern __inline uint32x2_t
26088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26089 vsli_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26090 {
26091   return __builtin_aarch64_usli_nv2si_uuus (__a, __b, __c);
26092 }
26093
26094 __extension__ extern __inline uint64x1_t
26095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26096 vsli_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26097 {
26098   return (uint64x1_t) {__builtin_aarch64_usli_ndi_uuus (__a[0], __b[0], __c)};
26099 }
26100
26101 __extension__ extern __inline poly64x1_t
26102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26103 vsli_n_p64 (poly64x1_t __a, poly64x1_t __b, const int __c)
26104 {
26105   return (poly64x1_t) {__builtin_aarch64_ssli_ndi_ppps (__a[0], __b[0], __c)};
26106 }
26107
26108 __extension__ extern __inline int8x16_t
26109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26110 vsliq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26111 {
26112   return (int8x16_t) __builtin_aarch64_ssli_nv16qi (__a, __b, __c);
26113 }
26114
26115 __extension__ extern __inline int16x8_t
26116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26117 vsliq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26118 {
26119   return (int16x8_t) __builtin_aarch64_ssli_nv8hi (__a, __b, __c);
26120 }
26121
26122 __extension__ extern __inline int32x4_t
26123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26124 vsliq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26125 {
26126   return (int32x4_t) __builtin_aarch64_ssli_nv4si (__a, __b, __c);
26127 }
26128
26129 __extension__ extern __inline int64x2_t
26130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26131 vsliq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26132 {
26133   return (int64x2_t) __builtin_aarch64_ssli_nv2di (__a, __b, __c);
26134 }
26135
26136 __extension__ extern __inline uint8x16_t
26137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26138 vsliq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26139 {
26140   return __builtin_aarch64_usli_nv16qi_uuus (__a, __b, __c);
26141 }
26142
26143 __extension__ extern __inline uint16x8_t
26144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26145 vsliq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26146 {
26147   return __builtin_aarch64_usli_nv8hi_uuus (__a, __b, __c);
26148 }
26149
26150 __extension__ extern __inline uint32x4_t
26151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26152 vsliq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26153 {
26154   return __builtin_aarch64_usli_nv4si_uuus (__a, __b, __c);
26155 }
26156
26157 __extension__ extern __inline uint64x2_t
26158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26159 vsliq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26160 {
26161   return __builtin_aarch64_usli_nv2di_uuus (__a, __b, __c);
26162 }
26163
26164 __extension__ extern __inline poly64x2_t
26165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26166 vsliq_n_p64 (poly64x2_t __a, poly64x2_t __b, const int __c)
26167 {
26168   return __builtin_aarch64_ssli_nv2di_ppps (__a, __b, __c);
26169 }
26170
26171 __extension__ extern __inline int64_t
26172 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26173 vslid_n_s64 (int64_t __a, int64_t __b, const int __c)
26174 {
26175   return __builtin_aarch64_ssli_ndi (__a, __b, __c);
26176 }
26177
26178 __extension__ extern __inline uint64_t
26179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26180 vslid_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26181 {
26182   return __builtin_aarch64_usli_ndi_uuus (__a, __b, __c);
26183 }
26184
26185 /* vsqadd */
26186
26187 __extension__ extern __inline uint8x8_t
26188 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26189 vsqadd_u8 (uint8x8_t __a, int8x8_t __b)
26190 {
26191   return __builtin_aarch64_usqaddv8qi_uus (__a, __b);
26192 }
26193
26194 __extension__ extern __inline uint16x4_t
26195 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26196 vsqadd_u16 (uint16x4_t __a, int16x4_t __b)
26197 {
26198   return __builtin_aarch64_usqaddv4hi_uus (__a, __b);
26199 }
26200
26201 __extension__ extern __inline uint32x2_t
26202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26203 vsqadd_u32 (uint32x2_t __a, int32x2_t __b)
26204 {
26205   return __builtin_aarch64_usqaddv2si_uus (__a, __b);
26206 }
26207
26208 __extension__ extern __inline uint64x1_t
26209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26210 vsqadd_u64 (uint64x1_t __a, int64x1_t __b)
26211 {
26212   return (uint64x1_t) {__builtin_aarch64_usqadddi_uus (__a[0], __b[0])};
26213 }
26214
26215 __extension__ extern __inline uint8x16_t
26216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26217 vsqaddq_u8 (uint8x16_t __a, int8x16_t __b)
26218 {
26219   return __builtin_aarch64_usqaddv16qi_uus (__a, __b);
26220 }
26221
26222 __extension__ extern __inline uint16x8_t
26223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26224 vsqaddq_u16 (uint16x8_t __a, int16x8_t __b)
26225 {
26226   return __builtin_aarch64_usqaddv8hi_uus (__a, __b);
26227 }
26228
26229 __extension__ extern __inline uint32x4_t
26230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26231 vsqaddq_u32 (uint32x4_t __a, int32x4_t __b)
26232 {
26233   return __builtin_aarch64_usqaddv4si_uus (__a, __b);
26234 }
26235
26236 __extension__ extern __inline uint64x2_t
26237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26238 vsqaddq_u64 (uint64x2_t __a, int64x2_t __b)
26239 {
26240   return __builtin_aarch64_usqaddv2di_uus (__a, __b);
26241 }
26242
26243 __extension__ extern __inline uint8_t
26244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26245 vsqaddb_u8 (uint8_t __a, int8_t __b)
26246 {
26247   return __builtin_aarch64_usqaddqi_uus (__a, __b);
26248 }
26249
26250 __extension__ extern __inline uint16_t
26251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26252 vsqaddh_u16 (uint16_t __a, int16_t __b)
26253 {
26254   return __builtin_aarch64_usqaddhi_uus (__a, __b);
26255 }
26256
26257 __extension__ extern __inline uint32_t
26258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26259 vsqadds_u32 (uint32_t __a, int32_t __b)
26260 {
26261   return __builtin_aarch64_usqaddsi_uus (__a, __b);
26262 }
26263
26264 __extension__ extern __inline uint64_t
26265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26266 vsqaddd_u64 (uint64_t __a, int64_t __b)
26267 {
26268   return __builtin_aarch64_usqadddi_uus (__a, __b);
26269 }
26270
26271 /* vsqrt */
26272 __extension__ extern __inline float32x2_t
26273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26274 vsqrt_f32 (float32x2_t a)
26275 {
26276   return __builtin_aarch64_sqrtv2sf (a);
26277 }
26278
26279 __extension__ extern __inline float32x4_t
26280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26281 vsqrtq_f32 (float32x4_t a)
26282 {
26283   return __builtin_aarch64_sqrtv4sf (a);
26284 }
26285
26286 __extension__ extern __inline float64x1_t
26287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26288 vsqrt_f64 (float64x1_t a)
26289 {
26290   return (float64x1_t) { __builtin_aarch64_sqrtdf (a[0]) };
26291 }
26292
26293 __extension__ extern __inline float64x2_t
26294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26295 vsqrtq_f64 (float64x2_t a)
26296 {
26297   return __builtin_aarch64_sqrtv2df (a);
26298 }
26299
26300 /* vsra */
26301
26302 __extension__ extern __inline int8x8_t
26303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26304 vsra_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26305 {
26306   return (int8x8_t) __builtin_aarch64_ssra_nv8qi (__a, __b, __c);
26307 }
26308
26309 __extension__ extern __inline int16x4_t
26310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26311 vsra_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26312 {
26313   return (int16x4_t) __builtin_aarch64_ssra_nv4hi (__a, __b, __c);
26314 }
26315
26316 __extension__ extern __inline int32x2_t
26317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26318 vsra_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26319 {
26320   return (int32x2_t) __builtin_aarch64_ssra_nv2si (__a, __b, __c);
26321 }
26322
26323 __extension__ extern __inline int64x1_t
26324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26325 vsra_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26326 {
26327   return (int64x1_t) {__builtin_aarch64_ssra_ndi (__a[0], __b[0], __c)};
26328 }
26329
26330 __extension__ extern __inline uint8x8_t
26331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26332 vsra_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26333 {
26334   return __builtin_aarch64_usra_nv8qi_uuus (__a, __b, __c);
26335 }
26336
26337 __extension__ extern __inline uint16x4_t
26338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26339 vsra_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26340 {
26341   return __builtin_aarch64_usra_nv4hi_uuus (__a, __b, __c);
26342 }
26343
26344 __extension__ extern __inline uint32x2_t
26345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26346 vsra_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26347 {
26348   return __builtin_aarch64_usra_nv2si_uuus (__a, __b, __c);
26349 }
26350
26351 __extension__ extern __inline uint64x1_t
26352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26353 vsra_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26354 {
26355   return (uint64x1_t) {__builtin_aarch64_usra_ndi_uuus (__a[0], __b[0], __c)};
26356 }
26357
26358 __extension__ extern __inline int8x16_t
26359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26360 vsraq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26361 {
26362   return (int8x16_t) __builtin_aarch64_ssra_nv16qi (__a, __b, __c);
26363 }
26364
26365 __extension__ extern __inline int16x8_t
26366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26367 vsraq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26368 {
26369   return (int16x8_t) __builtin_aarch64_ssra_nv8hi (__a, __b, __c);
26370 }
26371
26372 __extension__ extern __inline int32x4_t
26373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26374 vsraq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26375 {
26376   return (int32x4_t) __builtin_aarch64_ssra_nv4si (__a, __b, __c);
26377 }
26378
26379 __extension__ extern __inline int64x2_t
26380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26381 vsraq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26382 {
26383   return (int64x2_t) __builtin_aarch64_ssra_nv2di (__a, __b, __c);
26384 }
26385
26386 __extension__ extern __inline uint8x16_t
26387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26388 vsraq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26389 {
26390   return __builtin_aarch64_usra_nv16qi_uuus (__a, __b, __c);
26391 }
26392
26393 __extension__ extern __inline uint16x8_t
26394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26395 vsraq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26396 {
26397   return __builtin_aarch64_usra_nv8hi_uuus (__a, __b, __c);
26398 }
26399
26400 __extension__ extern __inline uint32x4_t
26401 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26402 vsraq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26403 {
26404   return __builtin_aarch64_usra_nv4si_uuus (__a, __b, __c);
26405 }
26406
26407 __extension__ extern __inline uint64x2_t
26408 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26409 vsraq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26410 {
26411   return __builtin_aarch64_usra_nv2di_uuus (__a, __b, __c);
26412 }
26413
26414 __extension__ extern __inline int64_t
26415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26416 vsrad_n_s64 (int64_t __a, int64_t __b, const int __c)
26417 {
26418   return __builtin_aarch64_ssra_ndi (__a, __b, __c);
26419 }
26420
26421 __extension__ extern __inline uint64_t
26422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26423 vsrad_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26424 {
26425   return __builtin_aarch64_usra_ndi_uuus (__a, __b, __c);
26426 }
26427
26428 /* vsri */
26429
26430 __extension__ extern __inline int8x8_t
26431 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26432 vsri_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26433 {
26434   return (int8x8_t) __builtin_aarch64_ssri_nv8qi (__a, __b, __c);
26435 }
26436
26437 __extension__ extern __inline int16x4_t
26438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26439 vsri_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26440 {
26441   return (int16x4_t) __builtin_aarch64_ssri_nv4hi (__a, __b, __c);
26442 }
26443
26444 __extension__ extern __inline int32x2_t
26445 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26446 vsri_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26447 {
26448   return (int32x2_t) __builtin_aarch64_ssri_nv2si (__a, __b, __c);
26449 }
26450
26451 __extension__ extern __inline int64x1_t
26452 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26453 vsri_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26454 {
26455   return (int64x1_t) {__builtin_aarch64_ssri_ndi (__a[0], __b[0], __c)};
26456 }
26457
26458 __extension__ extern __inline uint8x8_t
26459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26460 vsri_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26461 {
26462   return __builtin_aarch64_usri_nv8qi_uuus (__a, __b, __c);
26463 }
26464
26465 __extension__ extern __inline uint16x4_t
26466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26467 vsri_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26468 {
26469   return __builtin_aarch64_usri_nv4hi_uuus (__a, __b, __c);
26470 }
26471
26472 __extension__ extern __inline uint32x2_t
26473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26474 vsri_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26475 {
26476   return __builtin_aarch64_usri_nv2si_uuus (__a, __b, __c);
26477 }
26478
26479 __extension__ extern __inline uint64x1_t
26480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26481 vsri_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26482 {
26483   return (uint64x1_t) {__builtin_aarch64_usri_ndi_uuus (__a[0], __b[0], __c)};
26484 }
26485
26486 __extension__ extern __inline int8x16_t
26487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26488 vsriq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26489 {
26490   return (int8x16_t) __builtin_aarch64_ssri_nv16qi (__a, __b, __c);
26491 }
26492
26493 __extension__ extern __inline int16x8_t
26494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26495 vsriq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26496 {
26497   return (int16x8_t) __builtin_aarch64_ssri_nv8hi (__a, __b, __c);
26498 }
26499
26500 __extension__ extern __inline int32x4_t
26501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26502 vsriq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26503 {
26504   return (int32x4_t) __builtin_aarch64_ssri_nv4si (__a, __b, __c);
26505 }
26506
26507 __extension__ extern __inline int64x2_t
26508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26509 vsriq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26510 {
26511   return (int64x2_t) __builtin_aarch64_ssri_nv2di (__a, __b, __c);
26512 }
26513
26514 __extension__ extern __inline uint8x16_t
26515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26516 vsriq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26517 {
26518   return __builtin_aarch64_usri_nv16qi_uuus (__a, __b, __c);
26519 }
26520
26521 __extension__ extern __inline uint16x8_t
26522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26523 vsriq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26524 {
26525   return __builtin_aarch64_usri_nv8hi_uuus (__a, __b, __c);
26526 }
26527
26528 __extension__ extern __inline uint32x4_t
26529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26530 vsriq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26531 {
26532   return __builtin_aarch64_usri_nv4si_uuus (__a, __b, __c);
26533 }
26534
26535 __extension__ extern __inline uint64x2_t
26536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26537 vsriq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26538 {
26539   return __builtin_aarch64_usri_nv2di_uuus (__a, __b, __c);
26540 }
26541
26542 __extension__ extern __inline int64_t
26543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26544 vsrid_n_s64 (int64_t __a, int64_t __b, const int __c)
26545 {
26546   return __builtin_aarch64_ssri_ndi (__a, __b, __c);
26547 }
26548
26549 __extension__ extern __inline uint64_t
26550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26551 vsrid_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26552 {
26553   return __builtin_aarch64_usri_ndi_uuus (__a, __b, __c);
26554 }
26555
26556 /* vst1 */
26557
26558 __extension__ extern __inline void
26559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26560 vst1_f16 (float16_t *__a, float16x4_t __b)
26561 {
26562   __builtin_aarch64_st1v4hf (__a, __b);
26563 }
26564
26565 __extension__ extern __inline void
26566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26567 vst1_f32 (float32_t *a, float32x2_t b)
26568 {
26569   __builtin_aarch64_st1v2sf ((__builtin_aarch64_simd_sf *) a, b);
26570 }
26571
26572 __extension__ extern __inline void
26573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26574 vst1_f64 (float64_t *a, float64x1_t b)
26575 {
26576   *a = b[0];
26577 }
26578
26579 __extension__ extern __inline void
26580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26581 vst1_p8 (poly8_t *a, poly8x8_t b)
26582 {
26583   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a,
26584                              (int8x8_t) b);
26585 }
26586
26587 __extension__ extern __inline void
26588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26589 vst1_p16 (poly16_t *a, poly16x4_t b)
26590 {
26591   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a,
26592                              (int16x4_t) b);
26593 }
26594
26595 __extension__ extern __inline void
26596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26597 vst1_p64 (poly64_t *a, poly64x1_t b)
26598 {
26599   *a = b[0];
26600 }
26601
26602 __extension__ extern __inline void
26603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26604 vst1_s8 (int8_t *a, int8x8_t b)
26605 {
26606   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a, b);
26607 }
26608
26609 __extension__ extern __inline void
26610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26611 vst1_s16 (int16_t *a, int16x4_t b)
26612 {
26613   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a, b);
26614 }
26615
26616 __extension__ extern __inline void
26617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26618 vst1_s32 (int32_t *a, int32x2_t b)
26619 {
26620   __builtin_aarch64_st1v2si ((__builtin_aarch64_simd_si *) a, b);
26621 }
26622
26623 __extension__ extern __inline void
26624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26625 vst1_s64 (int64_t *a, int64x1_t b)
26626 {
26627   *a = b[0];
26628 }
26629
26630 __extension__ extern __inline void
26631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26632 vst1_u8 (uint8_t *a, uint8x8_t b)
26633 {
26634   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a,
26635                              (int8x8_t) b);
26636 }
26637
26638 __extension__ extern __inline void
26639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26640 vst1_u16 (uint16_t *a, uint16x4_t b)
26641 {
26642   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a,
26643                              (int16x4_t) b);
26644 }
26645
26646 __extension__ extern __inline void
26647 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26648 vst1_u32 (uint32_t *a, uint32x2_t b)
26649 {
26650   __builtin_aarch64_st1v2si ((__builtin_aarch64_simd_si *) a,
26651                              (int32x2_t) b);
26652 }
26653
26654 __extension__ extern __inline void
26655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26656 vst1_u64 (uint64_t *a, uint64x1_t b)
26657 {
26658   *a = b[0];
26659 }
26660
26661 /* vst1q */
26662
26663 __extension__ extern __inline void
26664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26665 vst1q_f16 (float16_t *__a, float16x8_t __b)
26666 {
26667   __builtin_aarch64_st1v8hf (__a, __b);
26668 }
26669
26670 __extension__ extern __inline void
26671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26672 vst1q_f32 (float32_t *a, float32x4_t b)
26673 {
26674   __builtin_aarch64_st1v4sf ((__builtin_aarch64_simd_sf *) a, b);
26675 }
26676
26677 __extension__ extern __inline void
26678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26679 vst1q_f64 (float64_t *a, float64x2_t b)
26680 {
26681   __builtin_aarch64_st1v2df ((__builtin_aarch64_simd_df *) a, b);
26682 }
26683
26684 __extension__ extern __inline void
26685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26686 vst1q_p8 (poly8_t *a, poly8x16_t b)
26687 {
26688   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a,
26689                               (int8x16_t) b);
26690 }
26691
26692 __extension__ extern __inline void
26693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26694 vst1q_p16 (poly16_t *a, poly16x8_t b)
26695 {
26696   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a,
26697                              (int16x8_t) b);
26698 }
26699
26700 __extension__ extern __inline void
26701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26702 vst1q_p64 (poly64_t *a, poly64x2_t b)
26703 {
26704   __builtin_aarch64_st1v2di_sp ((__builtin_aarch64_simd_di *) a,
26705                                 (poly64x2_t) b);
26706 }
26707
26708 __extension__ extern __inline void
26709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26710 vst1q_s8 (int8_t *a, int8x16_t b)
26711 {
26712   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a, b);
26713 }
26714
26715 __extension__ extern __inline void
26716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26717 vst1q_s16 (int16_t *a, int16x8_t b)
26718 {
26719   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a, b);
26720 }
26721
26722 __extension__ extern __inline void
26723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26724 vst1q_s32 (int32_t *a, int32x4_t b)
26725 {
26726   __builtin_aarch64_st1v4si ((__builtin_aarch64_simd_si *) a, b);
26727 }
26728
26729 __extension__ extern __inline void
26730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26731 vst1q_s64 (int64_t *a, int64x2_t b)
26732 {
26733   __builtin_aarch64_st1v2di ((__builtin_aarch64_simd_di *) a, b);
26734 }
26735
26736 __extension__ extern __inline void
26737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26738 vst1q_u8 (uint8_t *a, uint8x16_t b)
26739 {
26740   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a,
26741                               (int8x16_t) b);
26742 }
26743
26744 __extension__ extern __inline void
26745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26746 vst1q_u16 (uint16_t *a, uint16x8_t b)
26747 {
26748   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a,
26749                              (int16x8_t) b);
26750 }
26751
26752 __extension__ extern __inline void
26753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26754 vst1q_u32 (uint32_t *a, uint32x4_t b)
26755 {
26756   __builtin_aarch64_st1v4si ((__builtin_aarch64_simd_si *) a,
26757                              (int32x4_t) b);
26758 }
26759
26760 __extension__ extern __inline void
26761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26762 vst1q_u64 (uint64_t *a, uint64x2_t b)
26763 {
26764   __builtin_aarch64_st1v2di ((__builtin_aarch64_simd_di *) a,
26765                              (int64x2_t) b);
26766 }
26767
26768 /* vst1_lane */
26769
26770 __extension__ extern __inline void
26771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26772 vst1_lane_f16 (float16_t *__a, float16x4_t __b, const int __lane)
26773 {
26774   *__a = __aarch64_vget_lane_any (__b, __lane);
26775 }
26776
26777 __extension__ extern __inline void
26778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26779 vst1_lane_f32 (float32_t *__a, float32x2_t __b, const int __lane)
26780 {
26781   *__a = __aarch64_vget_lane_any (__b, __lane);
26782 }
26783
26784 __extension__ extern __inline void
26785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26786 vst1_lane_f64 (float64_t *__a, float64x1_t __b, const int __lane)
26787 {
26788   *__a = __aarch64_vget_lane_any (__b, __lane);
26789 }
26790
26791 __extension__ extern __inline void
26792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26793 vst1_lane_p8 (poly8_t *__a, poly8x8_t __b, const int __lane)
26794 {
26795   *__a = __aarch64_vget_lane_any (__b, __lane);
26796 }
26797
26798 __extension__ extern __inline void
26799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26800 vst1_lane_p16 (poly16_t *__a, poly16x4_t __b, const int __lane)
26801 {
26802   *__a = __aarch64_vget_lane_any (__b, __lane);
26803 }
26804
26805 __extension__ extern __inline void
26806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26807 vst1_lane_p64 (poly64_t *__a, poly64x1_t __b, const int __lane)
26808 {
26809   *__a = __aarch64_vget_lane_any (__b, __lane);
26810 }
26811
26812 __extension__ extern __inline void
26813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26814 vst1_lane_s8 (int8_t *__a, int8x8_t __b, const int __lane)
26815 {
26816   *__a = __aarch64_vget_lane_any (__b, __lane);
26817 }
26818
26819 __extension__ extern __inline void
26820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26821 vst1_lane_s16 (int16_t *__a, int16x4_t __b, const int __lane)
26822 {
26823   *__a = __aarch64_vget_lane_any (__b, __lane);
26824 }
26825
26826 __extension__ extern __inline void
26827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26828 vst1_lane_s32 (int32_t *__a, int32x2_t __b, const int __lane)
26829 {
26830   *__a = __aarch64_vget_lane_any (__b, __lane);
26831 }
26832
26833 __extension__ extern __inline void
26834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26835 vst1_lane_s64 (int64_t *__a, int64x1_t __b, const int __lane)
26836 {
26837   *__a = __aarch64_vget_lane_any (__b, __lane);
26838 }
26839
26840 __extension__ extern __inline void
26841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26842 vst1_lane_u8 (uint8_t *__a, uint8x8_t __b, const int __lane)
26843 {
26844   *__a = __aarch64_vget_lane_any (__b, __lane);
26845 }
26846
26847 __extension__ extern __inline void
26848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26849 vst1_lane_u16 (uint16_t *__a, uint16x4_t __b, const int __lane)
26850 {
26851   *__a = __aarch64_vget_lane_any (__b, __lane);
26852 }
26853
26854 __extension__ extern __inline void
26855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26856 vst1_lane_u32 (uint32_t *__a, uint32x2_t __b, const int __lane)
26857 {
26858   *__a = __aarch64_vget_lane_any (__b, __lane);
26859 }
26860
26861 __extension__ extern __inline void
26862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26863 vst1_lane_u64 (uint64_t *__a, uint64x1_t __b, const int __lane)
26864 {
26865   *__a = __aarch64_vget_lane_any (__b, __lane);
26866 }
26867
26868 /* vst1q_lane */
26869
26870 __extension__ extern __inline void
26871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26872 vst1q_lane_f16 (float16_t *__a, float16x8_t __b, const int __lane)
26873 {
26874   *__a = __aarch64_vget_lane_any (__b, __lane);
26875 }
26876
26877 __extension__ extern __inline void
26878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26879 vst1q_lane_f32 (float32_t *__a, float32x4_t __b, const int __lane)
26880 {
26881   *__a = __aarch64_vget_lane_any (__b, __lane);
26882 }
26883
26884 __extension__ extern __inline void
26885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26886 vst1q_lane_f64 (float64_t *__a, float64x2_t __b, const int __lane)
26887 {
26888   *__a = __aarch64_vget_lane_any (__b, __lane);
26889 }
26890
26891 __extension__ extern __inline void
26892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26893 vst1q_lane_p8 (poly8_t *__a, poly8x16_t __b, const int __lane)
26894 {
26895   *__a = __aarch64_vget_lane_any (__b, __lane);
26896 }
26897
26898 __extension__ extern __inline void
26899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26900 vst1q_lane_p16 (poly16_t *__a, poly16x8_t __b, const int __lane)
26901 {
26902   *__a = __aarch64_vget_lane_any (__b, __lane);
26903 }
26904
26905 __extension__ extern __inline void
26906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26907 vst1q_lane_p64 (poly64_t *__a, poly64x2_t __b, const int __lane)
26908 {
26909   *__a = __aarch64_vget_lane_any (__b, __lane);
26910 }
26911
26912 __extension__ extern __inline void
26913 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26914 vst1q_lane_s8 (int8_t *__a, int8x16_t __b, const int __lane)
26915 {
26916   *__a = __aarch64_vget_lane_any (__b, __lane);
26917 }
26918
26919 __extension__ extern __inline void
26920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26921 vst1q_lane_s16 (int16_t *__a, int16x8_t __b, const int __lane)
26922 {
26923   *__a = __aarch64_vget_lane_any (__b, __lane);
26924 }
26925
26926 __extension__ extern __inline void
26927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26928 vst1q_lane_s32 (int32_t *__a, int32x4_t __b, const int __lane)
26929 {
26930   *__a = __aarch64_vget_lane_any (__b, __lane);
26931 }
26932
26933 __extension__ extern __inline void
26934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26935 vst1q_lane_s64 (int64_t *__a, int64x2_t __b, const int __lane)
26936 {
26937   *__a = __aarch64_vget_lane_any (__b, __lane);
26938 }
26939
26940 __extension__ extern __inline void
26941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26942 vst1q_lane_u8 (uint8_t *__a, uint8x16_t __b, const int __lane)
26943 {
26944   *__a = __aarch64_vget_lane_any (__b, __lane);
26945 }
26946
26947 __extension__ extern __inline void
26948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26949 vst1q_lane_u16 (uint16_t *__a, uint16x8_t __b, const int __lane)
26950 {
26951   *__a = __aarch64_vget_lane_any (__b, __lane);
26952 }
26953
26954 __extension__ extern __inline void
26955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26956 vst1q_lane_u32 (uint32_t *__a, uint32x4_t __b, const int __lane)
26957 {
26958   *__a = __aarch64_vget_lane_any (__b, __lane);
26959 }
26960
26961 __extension__ extern __inline void
26962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26963 vst1q_lane_u64 (uint64_t *__a, uint64x2_t __b, const int __lane)
26964 {
26965   *__a = __aarch64_vget_lane_any (__b, __lane);
26966 }
26967
26968 /* vstn */
26969
26970 __extension__ extern __inline void
26971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26972 vst2_s64 (int64_t * __a, int64x1x2_t val)
26973 {
26974   __builtin_aarch64_simd_oi __o;
26975   int64x2x2_t temp;
26976   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
26977   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
26978   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
26979   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
26980   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
26981 }
26982
26983 __extension__ extern __inline void
26984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26985 vst2_u64 (uint64_t * __a, uint64x1x2_t val)
26986 {
26987   __builtin_aarch64_simd_oi __o;
26988   uint64x2x2_t temp;
26989   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
26990   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
26991   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
26992   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
26993   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
26994 }
26995
26996 __extension__ extern __inline void
26997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26998 vst2_f64 (float64_t * __a, float64x1x2_t val)
26999 {
27000   __builtin_aarch64_simd_oi __o;
27001   float64x2x2_t temp;
27002   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
27003   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
27004   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[0], 0);
27005   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[1], 1);
27006   __builtin_aarch64_st2df ((__builtin_aarch64_simd_df *) __a, __o);
27007 }
27008
27009 __extension__ extern __inline void
27010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27011 vst2_s8 (int8_t * __a, int8x8x2_t val)
27012 {
27013   __builtin_aarch64_simd_oi __o;
27014   int8x16x2_t temp;
27015   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
27016   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
27017   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27018   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27019   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27020 }
27021
27022 __extension__ extern __inline void
27023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27024 vst2_p8 (poly8_t * __a, poly8x8x2_t val)
27025 {
27026   __builtin_aarch64_simd_oi __o;
27027   poly8x16x2_t temp;
27028   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
27029   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
27030   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27031   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27032   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27033 }
27034
27035 __extension__ extern __inline void
27036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27037 vst2_s16 (int16_t * __a, int16x4x2_t val)
27038 {
27039   __builtin_aarch64_simd_oi __o;
27040   int16x8x2_t temp;
27041   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
27042   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
27043   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27044   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27045   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27046 }
27047
27048 __extension__ extern __inline void
27049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27050 vst2_p16 (poly16_t * __a, poly16x4x2_t val)
27051 {
27052   __builtin_aarch64_simd_oi __o;
27053   poly16x8x2_t temp;
27054   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
27055   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
27056   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27057   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27058   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27059 }
27060
27061 __extension__ extern __inline void
27062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27063 vst2_s32 (int32_t * __a, int32x2x2_t val)
27064 {
27065   __builtin_aarch64_simd_oi __o;
27066   int32x4x2_t temp;
27067   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27068   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27069   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
27070   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
27071   __builtin_aarch64_st2v2si ((__builtin_aarch64_simd_si *) __a, __o);
27072 }
27073
27074 __extension__ extern __inline void
27075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27076 vst2_u8 (uint8_t * __a, uint8x8x2_t val)
27077 {
27078   __builtin_aarch64_simd_oi __o;
27079   uint8x16x2_t temp;
27080   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27081   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27082   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27083   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27084   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27085 }
27086
27087 __extension__ extern __inline void
27088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27089 vst2_u16 (uint16_t * __a, uint16x4x2_t val)
27090 {
27091   __builtin_aarch64_simd_oi __o;
27092   uint16x8x2_t temp;
27093   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27094   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27095   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27096   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27097   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27098 }
27099
27100 __extension__ extern __inline void
27101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27102 vst2_u32 (uint32_t * __a, uint32x2x2_t val)
27103 {
27104   __builtin_aarch64_simd_oi __o;
27105   uint32x4x2_t temp;
27106   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
27107   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
27108   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
27109   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
27110   __builtin_aarch64_st2v2si ((__builtin_aarch64_simd_si *) __a, __o);
27111 }
27112
27113 __extension__ extern __inline void
27114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27115 vst2_f16 (float16_t * __a, float16x4x2_t val)
27116 {
27117   __builtin_aarch64_simd_oi __o;
27118   float16x8x2_t temp;
27119   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
27120   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
27121   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[0], 0);
27122   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[1], 1);
27123   __builtin_aarch64_st2v4hf (__a, __o);
27124 }
27125
27126 __extension__ extern __inline void
27127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27128 vst2_f32 (float32_t * __a, float32x2x2_t val)
27129 {
27130   __builtin_aarch64_simd_oi __o;
27131   float32x4x2_t temp;
27132   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
27133   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
27134   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[0], 0);
27135   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[1], 1);
27136   __builtin_aarch64_st2v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
27137 }
27138
27139 __extension__ extern __inline void
27140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27141 vst2_p64 (poly64_t * __a, poly64x1x2_t val)
27142 {
27143   __builtin_aarch64_simd_oi __o;
27144   poly64x2x2_t temp;
27145   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
27146   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
27147   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27148                                                (poly64x2_t) temp.val[0], 0);
27149   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27150                                                (poly64x2_t) temp.val[1], 1);
27151   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
27152 }
27153
27154 __extension__ extern __inline void
27155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27156 vst2q_s8 (int8_t * __a, int8x16x2_t val)
27157 {
27158   __builtin_aarch64_simd_oi __o;
27159   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
27160   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
27161   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27162 }
27163
27164 __extension__ extern __inline void
27165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27166 vst2q_p8 (poly8_t * __a, poly8x16x2_t val)
27167 {
27168   __builtin_aarch64_simd_oi __o;
27169   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
27170   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
27171   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27172 }
27173
27174 __extension__ extern __inline void
27175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27176 vst2q_s16 (int16_t * __a, int16x8x2_t val)
27177 {
27178   __builtin_aarch64_simd_oi __o;
27179   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
27180   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
27181   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27182 }
27183
27184 __extension__ extern __inline void
27185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27186 vst2q_p16 (poly16_t * __a, poly16x8x2_t val)
27187 {
27188   __builtin_aarch64_simd_oi __o;
27189   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
27190   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
27191   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27192 }
27193
27194 __extension__ extern __inline void
27195 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27196 vst2q_s32 (int32_t * __a, int32x4x2_t val)
27197 {
27198   __builtin_aarch64_simd_oi __o;
27199   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
27200   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
27201   __builtin_aarch64_st2v4si ((__builtin_aarch64_simd_si *) __a, __o);
27202 }
27203
27204 __extension__ extern __inline void
27205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27206 vst2q_s64 (int64_t * __a, int64x2x2_t val)
27207 {
27208   __builtin_aarch64_simd_oi __o;
27209   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
27210   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
27211   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
27212 }
27213
27214 __extension__ extern __inline void
27215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27216 vst2q_u8 (uint8_t * __a, uint8x16x2_t val)
27217 {
27218   __builtin_aarch64_simd_oi __o;
27219   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
27220   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
27221   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27222 }
27223
27224 __extension__ extern __inline void
27225 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27226 vst2q_u16 (uint16_t * __a, uint16x8x2_t val)
27227 {
27228   __builtin_aarch64_simd_oi __o;
27229   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
27230   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
27231   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27232 }
27233
27234 __extension__ extern __inline void
27235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27236 vst2q_u32 (uint32_t * __a, uint32x4x2_t val)
27237 {
27238   __builtin_aarch64_simd_oi __o;
27239   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
27240   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
27241   __builtin_aarch64_st2v4si ((__builtin_aarch64_simd_si *) __a, __o);
27242 }
27243
27244 __extension__ extern __inline void
27245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27246 vst2q_u64 (uint64_t * __a, uint64x2x2_t val)
27247 {
27248   __builtin_aarch64_simd_oi __o;
27249   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
27250   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
27251   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
27252 }
27253
27254 __extension__ extern __inline void
27255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27256 vst2q_f16 (float16_t * __a, float16x8x2_t val)
27257 {
27258   __builtin_aarch64_simd_oi __o;
27259   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[0], 0);
27260   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[1], 1);
27261   __builtin_aarch64_st2v8hf (__a, __o);
27262 }
27263
27264 __extension__ extern __inline void
27265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27266 vst2q_f32 (float32_t * __a, float32x4x2_t val)
27267 {
27268   __builtin_aarch64_simd_oi __o;
27269   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[0], 0);
27270   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[1], 1);
27271   __builtin_aarch64_st2v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
27272 }
27273
27274 __extension__ extern __inline void
27275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27276 vst2q_f64 (float64_t * __a, float64x2x2_t val)
27277 {
27278   __builtin_aarch64_simd_oi __o;
27279   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[0], 0);
27280   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[1], 1);
27281   __builtin_aarch64_st2v2df ((__builtin_aarch64_simd_df *) __a, __o);
27282 }
27283
27284 __extension__ extern __inline void
27285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27286 vst2q_p64 (poly64_t * __a, poly64x2x2_t val)
27287 {
27288   __builtin_aarch64_simd_oi __o;
27289   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27290                                                (poly64x2_t) val.val[0], 0);
27291   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27292                                                (poly64x2_t) val.val[1], 1);
27293   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
27294 }
27295
27296 __extension__ extern __inline void
27297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27298 vst3_s64 (int64_t * __a, int64x1x3_t val)
27299 {
27300   __builtin_aarch64_simd_ci __o;
27301   int64x2x3_t temp;
27302   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
27303   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
27304   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
27305   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
27306   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
27307   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
27308   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
27309 }
27310
27311 __extension__ extern __inline void
27312 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27313 vst3_u64 (uint64_t * __a, uint64x1x3_t val)
27314 {
27315   __builtin_aarch64_simd_ci __o;
27316   uint64x2x3_t temp;
27317   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
27318   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
27319   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
27320   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
27321   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
27322   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
27323   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
27324 }
27325
27326 __extension__ extern __inline void
27327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27328 vst3_f64 (float64_t * __a, float64x1x3_t val)
27329 {
27330   __builtin_aarch64_simd_ci __o;
27331   float64x2x3_t temp;
27332   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
27333   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
27334   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
27335   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[0], 0);
27336   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[1], 1);
27337   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[2], 2);
27338   __builtin_aarch64_st3df ((__builtin_aarch64_simd_df *) __a, __o);
27339 }
27340
27341 __extension__ extern __inline void
27342 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27343 vst3_s8 (int8_t * __a, int8x8x3_t val)
27344 {
27345   __builtin_aarch64_simd_ci __o;
27346   int8x16x3_t temp;
27347   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
27348   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
27349   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
27350   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
27351   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
27352   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
27353   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27354 }
27355
27356 __extension__ extern __inline void
27357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27358 vst3_p8 (poly8_t * __a, poly8x8x3_t val)
27359 {
27360   __builtin_aarch64_simd_ci __o;
27361   poly8x16x3_t temp;
27362   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
27363   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
27364   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
27365   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
27366   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
27367   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
27368   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27369 }
27370
27371 __extension__ extern __inline void
27372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27373 vst3_s16 (int16_t * __a, int16x4x3_t val)
27374 {
27375   __builtin_aarch64_simd_ci __o;
27376   int16x8x3_t temp;
27377   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
27378   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
27379   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
27380   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
27381   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
27382   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
27383   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27384 }
27385
27386 __extension__ extern __inline void
27387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27388 vst3_p16 (poly16_t * __a, poly16x4x3_t val)
27389 {
27390   __builtin_aarch64_simd_ci __o;
27391   poly16x8x3_t temp;
27392   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
27393   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
27394   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
27395   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
27396   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
27397   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
27398   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27399 }
27400
27401 __extension__ extern __inline void
27402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27403 vst3_s32 (int32_t * __a, int32x2x3_t val)
27404 {
27405   __builtin_aarch64_simd_ci __o;
27406   int32x4x3_t temp;
27407   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27408   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27409   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
27410   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
27411   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
27412   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
27413   __builtin_aarch64_st3v2si ((__builtin_aarch64_simd_si *) __a, __o);
27414 }
27415
27416 __extension__ extern __inline void
27417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27418 vst3_u8 (uint8_t * __a, uint8x8x3_t val)
27419 {
27420   __builtin_aarch64_simd_ci __o;
27421   uint8x16x3_t temp;
27422   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27423   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27424   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
27425   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
27426   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
27427   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
27428   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27429 }
27430
27431 __extension__ extern __inline void
27432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27433 vst3_u16 (uint16_t * __a, uint16x4x3_t val)
27434 {
27435   __builtin_aarch64_simd_ci __o;
27436   uint16x8x3_t temp;
27437   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27438   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27439   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
27440   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
27441   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
27442   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
27443   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27444 }
27445
27446 __extension__ extern __inline void
27447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27448 vst3_u32 (uint32_t * __a, uint32x2x3_t val)
27449 {
27450   __builtin_aarch64_simd_ci __o;
27451   uint32x4x3_t temp;
27452   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
27453   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
27454   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
27455   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
27456   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
27457   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
27458   __builtin_aarch64_st3v2si ((__builtin_aarch64_simd_si *) __a, __o);
27459 }
27460
27461 __extension__ extern __inline void
27462 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27463 vst3_f16 (float16_t * __a, float16x4x3_t val)
27464 {
27465   __builtin_aarch64_simd_ci __o;
27466   float16x8x3_t temp;
27467   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
27468   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
27469   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
27470   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[0], 0);
27471   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[1], 1);
27472   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[2], 2);
27473   __builtin_aarch64_st3v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
27474 }
27475
27476 __extension__ extern __inline void
27477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27478 vst3_f32 (float32_t * __a, float32x2x3_t val)
27479 {
27480   __builtin_aarch64_simd_ci __o;
27481   float32x4x3_t temp;
27482   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
27483   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
27484   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
27485   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[0], 0);
27486   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[1], 1);
27487   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[2], 2);
27488   __builtin_aarch64_st3v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
27489 }
27490
27491 __extension__ extern __inline void
27492 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27493 vst3_p64 (poly64_t * __a, poly64x1x3_t val)
27494 {
27495   __builtin_aarch64_simd_ci __o;
27496   poly64x2x3_t temp;
27497   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
27498   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
27499   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
27500   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27501                                                (poly64x2_t) temp.val[0], 0);
27502   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27503                                                (poly64x2_t) temp.val[1], 1);
27504   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27505                                                (poly64x2_t) temp.val[2], 2);
27506   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
27507 }
27508
27509 __extension__ extern __inline void
27510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27511 vst3q_s8 (int8_t * __a, int8x16x3_t val)
27512 {
27513   __builtin_aarch64_simd_ci __o;
27514   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
27515   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
27516   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
27517   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27518 }
27519
27520 __extension__ extern __inline void
27521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27522 vst3q_p8 (poly8_t * __a, poly8x16x3_t val)
27523 {
27524   __builtin_aarch64_simd_ci __o;
27525   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
27526   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
27527   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
27528   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27529 }
27530
27531 __extension__ extern __inline void
27532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27533 vst3q_s16 (int16_t * __a, int16x8x3_t val)
27534 {
27535   __builtin_aarch64_simd_ci __o;
27536   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
27537   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
27538   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
27539   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27540 }
27541
27542 __extension__ extern __inline void
27543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27544 vst3q_p16 (poly16_t * __a, poly16x8x3_t val)
27545 {
27546   __builtin_aarch64_simd_ci __o;
27547   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
27548   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
27549   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
27550   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27551 }
27552
27553 __extension__ extern __inline void
27554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27555 vst3q_s32 (int32_t * __a, int32x4x3_t val)
27556 {
27557   __builtin_aarch64_simd_ci __o;
27558   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
27559   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
27560   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
27561   __builtin_aarch64_st3v4si ((__builtin_aarch64_simd_si *) __a, __o);
27562 }
27563
27564 __extension__ extern __inline void
27565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27566 vst3q_s64 (int64_t * __a, int64x2x3_t val)
27567 {
27568   __builtin_aarch64_simd_ci __o;
27569   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
27570   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
27571   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
27572   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
27573 }
27574
27575 __extension__ extern __inline void
27576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27577 vst3q_u8 (uint8_t * __a, uint8x16x3_t val)
27578 {
27579   __builtin_aarch64_simd_ci __o;
27580   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
27581   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
27582   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
27583   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27584 }
27585
27586 __extension__ extern __inline void
27587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27588 vst3q_u16 (uint16_t * __a, uint16x8x3_t val)
27589 {
27590   __builtin_aarch64_simd_ci __o;
27591   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
27592   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
27593   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
27594   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27595 }
27596
27597 __extension__ extern __inline void
27598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27599 vst3q_u32 (uint32_t * __a, uint32x4x3_t val)
27600 {
27601   __builtin_aarch64_simd_ci __o;
27602   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
27603   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
27604   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
27605   __builtin_aarch64_st3v4si ((__builtin_aarch64_simd_si *) __a, __o);
27606 }
27607
27608 __extension__ extern __inline void
27609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27610 vst3q_u64 (uint64_t * __a, uint64x2x3_t val)
27611 {
27612   __builtin_aarch64_simd_ci __o;
27613   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
27614   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
27615   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
27616   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
27617 }
27618
27619 __extension__ extern __inline void
27620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27621 vst3q_f16 (float16_t * __a, float16x8x3_t val)
27622 {
27623   __builtin_aarch64_simd_ci __o;
27624   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[0], 0);
27625   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[1], 1);
27626   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[2], 2);
27627   __builtin_aarch64_st3v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
27628 }
27629
27630 __extension__ extern __inline void
27631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27632 vst3q_f32 (float32_t * __a, float32x4x3_t val)
27633 {
27634   __builtin_aarch64_simd_ci __o;
27635   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[0], 0);
27636   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[1], 1);
27637   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[2], 2);
27638   __builtin_aarch64_st3v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
27639 }
27640
27641 __extension__ extern __inline void
27642 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27643 vst3q_f64 (float64_t * __a, float64x2x3_t val)
27644 {
27645   __builtin_aarch64_simd_ci __o;
27646   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[0], 0);
27647   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[1], 1);
27648   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[2], 2);
27649   __builtin_aarch64_st3v2df ((__builtin_aarch64_simd_df *) __a, __o);
27650 }
27651
27652 __extension__ extern __inline void
27653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27654 vst3q_p64 (poly64_t * __a, poly64x2x3_t val)
27655 {
27656   __builtin_aarch64_simd_ci __o;
27657   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27658                                                (poly64x2_t) val.val[0], 0);
27659   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27660                                                (poly64x2_t) val.val[1], 1);
27661   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27662                                                (poly64x2_t) val.val[2], 2);
27663   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
27664 }
27665
27666 __extension__ extern __inline void
27667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27668 vst4_s64 (int64_t * __a, int64x1x4_t val)
27669 {
27670   __builtin_aarch64_simd_xi __o;
27671   int64x2x4_t temp;
27672   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
27673   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
27674   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
27675   temp.val[3] = vcombine_s64 (val.val[3], vcreate_s64 (__AARCH64_INT64_C (0)));
27676   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[0], 0);
27677   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[1], 1);
27678   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[2], 2);
27679   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[3], 3);
27680   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
27681 }
27682
27683 __extension__ extern __inline void
27684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27685 vst4_u64 (uint64_t * __a, uint64x1x4_t val)
27686 {
27687   __builtin_aarch64_simd_xi __o;
27688   uint64x2x4_t temp;
27689   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
27690   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
27691   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
27692   temp.val[3] = vcombine_u64 (val.val[3], vcreate_u64 (__AARCH64_UINT64_C (0)));
27693   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[0], 0);
27694   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[1], 1);
27695   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[2], 2);
27696   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[3], 3);
27697   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
27698 }
27699
27700 __extension__ extern __inline void
27701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27702 vst4_f64 (float64_t * __a, float64x1x4_t val)
27703 {
27704   __builtin_aarch64_simd_xi __o;
27705   float64x2x4_t temp;
27706   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
27707   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
27708   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
27709   temp.val[3] = vcombine_f64 (val.val[3], vcreate_f64 (__AARCH64_UINT64_C (0)));
27710   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[0], 0);
27711   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[1], 1);
27712   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[2], 2);
27713   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[3], 3);
27714   __builtin_aarch64_st4df ((__builtin_aarch64_simd_df *) __a, __o);
27715 }
27716
27717 __extension__ extern __inline void
27718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27719 vst4_s8 (int8_t * __a, int8x8x4_t val)
27720 {
27721   __builtin_aarch64_simd_xi __o;
27722   int8x16x4_t temp;
27723   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
27724   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
27725   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
27726   temp.val[3] = vcombine_s8 (val.val[3], vcreate_s8 (__AARCH64_INT64_C (0)));
27727   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
27728   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
27729   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
27730   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
27731   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27732 }
27733
27734 __extension__ extern __inline void
27735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27736 vst4_p8 (poly8_t * __a, poly8x8x4_t val)
27737 {
27738   __builtin_aarch64_simd_xi __o;
27739   poly8x16x4_t temp;
27740   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
27741   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
27742   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
27743   temp.val[3] = vcombine_p8 (val.val[3], vcreate_p8 (__AARCH64_UINT64_C (0)));
27744   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
27745   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
27746   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
27747   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
27748   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27749 }
27750
27751 __extension__ extern __inline void
27752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27753 vst4_s16 (int16_t * __a, int16x4x4_t val)
27754 {
27755   __builtin_aarch64_simd_xi __o;
27756   int16x8x4_t temp;
27757   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
27758   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
27759   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
27760   temp.val[3] = vcombine_s16 (val.val[3], vcreate_s16 (__AARCH64_INT64_C (0)));
27761   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
27762   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
27763   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
27764   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
27765   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27766 }
27767
27768 __extension__ extern __inline void
27769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27770 vst4_p16 (poly16_t * __a, poly16x4x4_t val)
27771 {
27772   __builtin_aarch64_simd_xi __o;
27773   poly16x8x4_t temp;
27774   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
27775   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
27776   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
27777   temp.val[3] = vcombine_p16 (val.val[3], vcreate_p16 (__AARCH64_UINT64_C (0)));
27778   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
27779   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
27780   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
27781   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
27782   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27783 }
27784
27785 __extension__ extern __inline void
27786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27787 vst4_s32 (int32_t * __a, int32x2x4_t val)
27788 {
27789   __builtin_aarch64_simd_xi __o;
27790   int32x4x4_t temp;
27791   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27792   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27793   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
27794   temp.val[3] = vcombine_s32 (val.val[3], vcreate_s32 (__AARCH64_INT64_C (0)));
27795   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[0], 0);
27796   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[1], 1);
27797   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[2], 2);
27798   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[3], 3);
27799   __builtin_aarch64_st4v2si ((__builtin_aarch64_simd_si *) __a, __o);
27800 }
27801
27802 __extension__ extern __inline void
27803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27804 vst4_u8 (uint8_t * __a, uint8x8x4_t val)
27805 {
27806   __builtin_aarch64_simd_xi __o;
27807   uint8x16x4_t temp;
27808   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27809   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27810   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
27811   temp.val[3] = vcombine_u8 (val.val[3], vcreate_u8 (__AARCH64_UINT64_C (0)));
27812   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
27813   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
27814   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
27815   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
27816   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27817 }
27818
27819 __extension__ extern __inline void
27820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27821 vst4_u16 (uint16_t * __a, uint16x4x4_t val)
27822 {
27823   __builtin_aarch64_simd_xi __o;
27824   uint16x8x4_t temp;
27825   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27826   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27827   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
27828   temp.val[3] = vcombine_u16 (val.val[3], vcreate_u16 (__AARCH64_UINT64_C (0)));
27829   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
27830   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
27831   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
27832   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
27833   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27834 }
27835
27836 __extension__ extern __inline void
27837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27838 vst4_u32 (uint32_t * __a, uint32x2x4_t val)
27839 {
27840   __builtin_aarch64_simd_xi __o;
27841   uint32x4x4_t temp;
27842   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
27843   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
27844   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
27845   temp.val[3] = vcombine_u32 (val.val[3], vcreate_u32 (__AARCH64_UINT64_C (0)));
27846   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[0], 0);
27847   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[1], 1);
27848   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[2], 2);
27849   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[3], 3);
27850   __builtin_aarch64_st4v2si ((__builtin_aarch64_simd_si *) __a, __o);
27851 }
27852
27853 __extension__ extern __inline void
27854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27855 vst4_f16 (float16_t * __a, float16x4x4_t val)
27856 {
27857   __builtin_aarch64_simd_xi __o;
27858   float16x8x4_t temp;
27859   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
27860   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
27861   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
27862   temp.val[3] = vcombine_f16 (val.val[3], vcreate_f16 (__AARCH64_UINT64_C (0)));
27863   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[0], 0);
27864   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[1], 1);
27865   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[2], 2);
27866   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[3], 3);
27867   __builtin_aarch64_st4v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
27868 }
27869
27870 __extension__ extern __inline void
27871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27872 vst4_f32 (float32_t * __a, float32x2x4_t val)
27873 {
27874   __builtin_aarch64_simd_xi __o;
27875   float32x4x4_t temp;
27876   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
27877   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
27878   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
27879   temp.val[3] = vcombine_f32 (val.val[3], vcreate_f32 (__AARCH64_UINT64_C (0)));
27880   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[0], 0);
27881   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[1], 1);
27882   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[2], 2);
27883   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[3], 3);
27884   __builtin_aarch64_st4v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
27885 }
27886
27887 __extension__ extern __inline void
27888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27889 vst4_p64 (poly64_t * __a, poly64x1x4_t val)
27890 {
27891   __builtin_aarch64_simd_xi __o;
27892   poly64x2x4_t temp;
27893   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
27894   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
27895   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
27896   temp.val[3] = vcombine_p64 (val.val[3], vcreate_p64 (__AARCH64_UINT64_C (0)));
27897   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27898                                                (poly64x2_t) temp.val[0], 0);
27899   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27900                                                (poly64x2_t) temp.val[1], 1);
27901   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27902                                                (poly64x2_t) temp.val[2], 2);
27903   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27904                                                (poly64x2_t) temp.val[3], 3);
27905   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
27906 }
27907
27908 __extension__ extern __inline void
27909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27910 vst4q_s8 (int8_t * __a, int8x16x4_t val)
27911 {
27912   __builtin_aarch64_simd_xi __o;
27913   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
27914   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
27915   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
27916   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
27917   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27918 }
27919
27920 __extension__ extern __inline void
27921 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27922 vst4q_p8 (poly8_t * __a, poly8x16x4_t val)
27923 {
27924   __builtin_aarch64_simd_xi __o;
27925   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
27926   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
27927   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
27928   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
27929   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27930 }
27931
27932 __extension__ extern __inline void
27933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27934 vst4q_s16 (int16_t * __a, int16x8x4_t val)
27935 {
27936   __builtin_aarch64_simd_xi __o;
27937   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
27938   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
27939   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
27940   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
27941   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27942 }
27943
27944 __extension__ extern __inline void
27945 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27946 vst4q_p16 (poly16_t * __a, poly16x8x4_t val)
27947 {
27948   __builtin_aarch64_simd_xi __o;
27949   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
27950   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
27951   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
27952   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
27953   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27954 }
27955
27956 __extension__ extern __inline void
27957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27958 vst4q_s32 (int32_t * __a, int32x4x4_t val)
27959 {
27960   __builtin_aarch64_simd_xi __o;
27961   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[0], 0);
27962   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[1], 1);
27963   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[2], 2);
27964   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[3], 3);
27965   __builtin_aarch64_st4v4si ((__builtin_aarch64_simd_si *) __a, __o);
27966 }
27967
27968 __extension__ extern __inline void
27969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27970 vst4q_s64 (int64_t * __a, int64x2x4_t val)
27971 {
27972   __builtin_aarch64_simd_xi __o;
27973   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[0], 0);
27974   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[1], 1);
27975   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[2], 2);
27976   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[3], 3);
27977   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
27978 }
27979
27980 __extension__ extern __inline void
27981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27982 vst4q_u8 (uint8_t * __a, uint8x16x4_t val)
27983 {
27984   __builtin_aarch64_simd_xi __o;
27985   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
27986   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
27987   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
27988   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
27989   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27990 }
27991
27992 __extension__ extern __inline void
27993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27994 vst4q_u16 (uint16_t * __a, uint16x8x4_t val)
27995 {
27996   __builtin_aarch64_simd_xi __o;
27997   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
27998   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
27999   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
28000   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
28001   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
28002 }
28003
28004 __extension__ extern __inline void
28005 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28006 vst4q_u32 (uint32_t * __a, uint32x4x4_t val)
28007 {
28008   __builtin_aarch64_simd_xi __o;
28009   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[0], 0);
28010   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[1], 1);
28011   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[2], 2);
28012   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[3], 3);
28013   __builtin_aarch64_st4v4si ((__builtin_aarch64_simd_si *) __a, __o);
28014 }
28015
28016 __extension__ extern __inline void
28017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28018 vst4q_u64 (uint64_t * __a, uint64x2x4_t val)
28019 {
28020   __builtin_aarch64_simd_xi __o;
28021   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[0], 0);
28022   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[1], 1);
28023   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[2], 2);
28024   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[3], 3);
28025   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
28026 }
28027
28028 __extension__ extern __inline void
28029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28030 vst4q_f16 (float16_t * __a, float16x8x4_t val)
28031 {
28032   __builtin_aarch64_simd_xi __o;
28033   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[0], 0);
28034   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[1], 1);
28035   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[2], 2);
28036   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[3], 3);
28037   __builtin_aarch64_st4v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
28038 }
28039
28040 __extension__ extern __inline void
28041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28042 vst4q_f32 (float32_t * __a, float32x4x4_t val)
28043 {
28044   __builtin_aarch64_simd_xi __o;
28045   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[0], 0);
28046   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[1], 1);
28047   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[2], 2);
28048   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[3], 3);
28049   __builtin_aarch64_st4v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
28050 }
28051
28052 __extension__ extern __inline void
28053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28054 vst4q_f64 (float64_t * __a, float64x2x4_t val)
28055 {
28056   __builtin_aarch64_simd_xi __o;
28057   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[0], 0);
28058   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[1], 1);
28059   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[2], 2);
28060   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[3], 3);
28061   __builtin_aarch64_st4v2df ((__builtin_aarch64_simd_df *) __a, __o);
28062 }
28063
28064 __extension__ extern __inline void
28065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28066 vst4q_p64 (poly64_t * __a, poly64x2x4_t val)
28067 {
28068   __builtin_aarch64_simd_xi __o;
28069   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28070                                                (poly64x2_t) val.val[0], 0);
28071   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28072                                                (poly64x2_t) val.val[1], 1);
28073   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28074                                                (poly64x2_t) val.val[2], 2);
28075   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28076                                                (poly64x2_t) val.val[3], 3);
28077   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
28078 }
28079
28080 /* vsub */
28081
28082 __extension__ extern __inline int64_t
28083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28084 vsubd_s64 (int64_t __a, int64_t __b)
28085 {
28086   return __a - __b;
28087 }
28088
28089 __extension__ extern __inline uint64_t
28090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28091 vsubd_u64 (uint64_t __a, uint64_t __b)
28092 {
28093   return __a - __b;
28094 }
28095
28096 /* vtbx1  */
28097
28098 __extension__ extern __inline int8x8_t
28099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28100 vtbx1_s8 (int8x8_t __r, int8x8_t __tab, int8x8_t __idx)
28101 {
28102   uint8x8_t __mask = vclt_u8 (vreinterpret_u8_s8 (__idx),
28103                               vmov_n_u8 (8));
28104   int8x8_t __tbl = vtbl1_s8 (__tab, __idx);
28105
28106   return vbsl_s8 (__mask, __tbl, __r);
28107 }
28108
28109 __extension__ extern __inline uint8x8_t
28110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28111 vtbx1_u8 (uint8x8_t __r, uint8x8_t __tab, uint8x8_t __idx)
28112 {
28113   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (8));
28114   uint8x8_t __tbl = vtbl1_u8 (__tab, __idx);
28115
28116   return vbsl_u8 (__mask, __tbl, __r);
28117 }
28118
28119 __extension__ extern __inline poly8x8_t
28120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28121 vtbx1_p8 (poly8x8_t __r, poly8x8_t __tab, uint8x8_t __idx)
28122 {
28123   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (8));
28124   poly8x8_t __tbl = vtbl1_p8 (__tab, __idx);
28125
28126   return vbsl_p8 (__mask, __tbl, __r);
28127 }
28128
28129 /* vtbx3  */
28130
28131 __extension__ extern __inline int8x8_t
28132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28133 vtbx3_s8 (int8x8_t __r, int8x8x3_t __tab, int8x8_t __idx)
28134 {
28135   uint8x8_t __mask = vclt_u8 (vreinterpret_u8_s8 (__idx),
28136                               vmov_n_u8 (24));
28137   int8x8_t __tbl = vtbl3_s8 (__tab, __idx);
28138
28139   return vbsl_s8 (__mask, __tbl, __r);
28140 }
28141
28142 __extension__ extern __inline uint8x8_t
28143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28144 vtbx3_u8 (uint8x8_t __r, uint8x8x3_t __tab, uint8x8_t __idx)
28145 {
28146   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (24));
28147   uint8x8_t __tbl = vtbl3_u8 (__tab, __idx);
28148
28149   return vbsl_u8 (__mask, __tbl, __r);
28150 }
28151
28152 __extension__ extern __inline poly8x8_t
28153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28154 vtbx3_p8 (poly8x8_t __r, poly8x8x3_t __tab, uint8x8_t __idx)
28155 {
28156   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (24));
28157   poly8x8_t __tbl = vtbl3_p8 (__tab, __idx);
28158
28159   return vbsl_p8 (__mask, __tbl, __r);
28160 }
28161
28162 /* vtbx4  */
28163
28164 __extension__ extern __inline int8x8_t
28165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28166 vtbx4_s8 (int8x8_t __r, int8x8x4_t __tab, int8x8_t __idx)
28167 {
28168   int8x8_t result;
28169   int8x16x2_t temp;
28170   __builtin_aarch64_simd_oi __o;
28171   temp.val[0] = vcombine_s8 (__tab.val[0], __tab.val[1]);
28172   temp.val[1] = vcombine_s8 (__tab.val[2], __tab.val[3]);
28173   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28174                                            (int8x16_t) temp.val[0], 0);
28175   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28176                                            (int8x16_t) temp.val[1], 1);
28177   result = __builtin_aarch64_tbx4v8qi (__r, __o, __idx);
28178   return result;
28179 }
28180
28181 __extension__ extern __inline uint8x8_t
28182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28183 vtbx4_u8 (uint8x8_t __r, uint8x8x4_t __tab, uint8x8_t __idx)
28184 {
28185   uint8x8_t result;
28186   uint8x16x2_t temp;
28187   __builtin_aarch64_simd_oi __o;
28188   temp.val[0] = vcombine_u8 (__tab.val[0], __tab.val[1]);
28189   temp.val[1] = vcombine_u8 (__tab.val[2], __tab.val[3]);
28190   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28191                                            (int8x16_t) temp.val[0], 0);
28192   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28193                                            (int8x16_t) temp.val[1], 1);
28194   result = (uint8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)__r, __o,
28195                                                   (int8x8_t)__idx);
28196   return result;
28197 }
28198
28199 __extension__ extern __inline poly8x8_t
28200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28201 vtbx4_p8 (poly8x8_t __r, poly8x8x4_t __tab, uint8x8_t __idx)
28202 {
28203   poly8x8_t result;
28204   poly8x16x2_t temp;
28205   __builtin_aarch64_simd_oi __o;
28206   temp.val[0] = vcombine_p8 (__tab.val[0], __tab.val[1]);
28207   temp.val[1] = vcombine_p8 (__tab.val[2], __tab.val[3]);
28208   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28209                                            (int8x16_t) temp.val[0], 0);
28210   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28211                                            (int8x16_t) temp.val[1], 1);
28212   result = (poly8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)__r, __o,
28213                                                   (int8x8_t)__idx);
28214   return result;
28215 }
28216
28217 /* vtrn */
28218
28219 __extension__ extern __inline float16x4_t
28220 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28221 vtrn1_f16 (float16x4_t __a, float16x4_t __b)
28222 {
28223 #ifdef __AARCH64EB__
28224   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28225 #else
28226   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28227 #endif
28228 }
28229
28230 __extension__ extern __inline float32x2_t
28231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28232 vtrn1_f32 (float32x2_t __a, float32x2_t __b)
28233 {
28234 #ifdef __AARCH64EB__
28235   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
28236 #else
28237   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
28238 #endif
28239 }
28240
28241 __extension__ extern __inline poly8x8_t
28242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28243 vtrn1_p8 (poly8x8_t __a, poly8x8_t __b)
28244 {
28245 #ifdef __AARCH64EB__
28246   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28247 #else
28248   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28249 #endif
28250 }
28251
28252 __extension__ extern __inline poly16x4_t
28253 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28254 vtrn1_p16 (poly16x4_t __a, poly16x4_t __b)
28255 {
28256 #ifdef __AARCH64EB__
28257   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28258 #else
28259   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28260 #endif
28261 }
28262
28263 __extension__ extern __inline int8x8_t
28264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28265 vtrn1_s8 (int8x8_t __a, int8x8_t __b)
28266 {
28267 #ifdef __AARCH64EB__
28268   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28269 #else
28270   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28271 #endif
28272 }
28273
28274 __extension__ extern __inline int16x4_t
28275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28276 vtrn1_s16 (int16x4_t __a, int16x4_t __b)
28277 {
28278 #ifdef __AARCH64EB__
28279   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28280 #else
28281   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28282 #endif
28283 }
28284
28285 __extension__ extern __inline int32x2_t
28286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28287 vtrn1_s32 (int32x2_t __a, int32x2_t __b)
28288 {
28289 #ifdef __AARCH64EB__
28290   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
28291 #else
28292   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
28293 #endif
28294 }
28295
28296 __extension__ extern __inline uint8x8_t
28297 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28298 vtrn1_u8 (uint8x8_t __a, uint8x8_t __b)
28299 {
28300 #ifdef __AARCH64EB__
28301   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28302 #else
28303   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28304 #endif
28305 }
28306
28307 __extension__ extern __inline uint16x4_t
28308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28309 vtrn1_u16 (uint16x4_t __a, uint16x4_t __b)
28310 {
28311 #ifdef __AARCH64EB__
28312   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28313 #else
28314   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28315 #endif
28316 }
28317
28318 __extension__ extern __inline uint32x2_t
28319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28320 vtrn1_u32 (uint32x2_t __a, uint32x2_t __b)
28321 {
28322 #ifdef __AARCH64EB__
28323   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
28324 #else
28325   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
28326 #endif
28327 }
28328
28329 __extension__ extern __inline float16x8_t
28330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28331 vtrn1q_f16 (float16x8_t __a, float16x8_t __b)
28332 {
28333 #ifdef __AARCH64EB__
28334   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28335 #else
28336   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28337 #endif
28338 }
28339
28340 __extension__ extern __inline float32x4_t
28341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28342 vtrn1q_f32 (float32x4_t __a, float32x4_t __b)
28343 {
28344 #ifdef __AARCH64EB__
28345   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
28346 #else
28347   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
28348 #endif
28349 }
28350
28351 __extension__ extern __inline float64x2_t
28352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28353 vtrn1q_f64 (float64x2_t __a, float64x2_t __b)
28354 {
28355 #ifdef __AARCH64EB__
28356   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
28357 #else
28358   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
28359 #endif
28360 }
28361
28362 __extension__ extern __inline poly8x16_t
28363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28364 vtrn1q_p8 (poly8x16_t __a, poly8x16_t __b)
28365 {
28366 #ifdef __AARCH64EB__
28367   return __builtin_shuffle (__a, __b,
28368       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
28369 #else
28370   return __builtin_shuffle (__a, __b,
28371       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
28372 #endif
28373 }
28374
28375 __extension__ extern __inline poly16x8_t
28376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28377 vtrn1q_p16 (poly16x8_t __a, poly16x8_t __b)
28378 {
28379 #ifdef __AARCH64EB__
28380   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28381 #else
28382   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28383 #endif
28384 }
28385
28386 __extension__ extern __inline int8x16_t
28387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28388 vtrn1q_s8 (int8x16_t __a, int8x16_t __b)
28389 {
28390 #ifdef __AARCH64EB__
28391   return __builtin_shuffle (__a, __b,
28392       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
28393 #else
28394   return __builtin_shuffle (__a, __b,
28395       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
28396 #endif
28397 }
28398
28399 __extension__ extern __inline int16x8_t
28400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28401 vtrn1q_s16 (int16x8_t __a, int16x8_t __b)
28402 {
28403 #ifdef __AARCH64EB__
28404   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28405 #else
28406   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28407 #endif
28408 }
28409
28410 __extension__ extern __inline int32x4_t
28411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28412 vtrn1q_s32 (int32x4_t __a, int32x4_t __b)
28413 {
28414 #ifdef __AARCH64EB__
28415   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
28416 #else
28417   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
28418 #endif
28419 }
28420
28421 __extension__ extern __inline int64x2_t
28422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28423 vtrn1q_s64 (int64x2_t __a, int64x2_t __b)
28424 {
28425 #ifdef __AARCH64EB__
28426   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
28427 #else
28428   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
28429 #endif
28430 }
28431
28432 __extension__ extern __inline uint8x16_t
28433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28434 vtrn1q_u8 (uint8x16_t __a, uint8x16_t __b)
28435 {
28436 #ifdef __AARCH64EB__
28437   return __builtin_shuffle (__a, __b,
28438       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
28439 #else
28440   return __builtin_shuffle (__a, __b,
28441       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
28442 #endif
28443 }
28444
28445 __extension__ extern __inline uint16x8_t
28446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28447 vtrn1q_u16 (uint16x8_t __a, uint16x8_t __b)
28448 {
28449 #ifdef __AARCH64EB__
28450   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28451 #else
28452   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28453 #endif
28454 }
28455
28456 __extension__ extern __inline uint32x4_t
28457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28458 vtrn1q_u32 (uint32x4_t __a, uint32x4_t __b)
28459 {
28460 #ifdef __AARCH64EB__
28461   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
28462 #else
28463   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
28464 #endif
28465 }
28466
28467 __extension__ extern __inline uint64x2_t
28468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28469 vtrn1q_u64 (uint64x2_t __a, uint64x2_t __b)
28470 {
28471 #ifdef __AARCH64EB__
28472   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
28473 #else
28474   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
28475 #endif
28476 }
28477
28478 __extension__ extern __inline float16x4_t
28479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28480 vtrn2_f16 (float16x4_t __a, float16x4_t __b)
28481 {
28482 #ifdef __AARCH64EB__
28483   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28484 #else
28485   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28486 #endif
28487 }
28488
28489 __extension__ extern __inline float32x2_t
28490 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28491 vtrn2_f32 (float32x2_t __a, float32x2_t __b)
28492 {
28493 #ifdef __AARCH64EB__
28494   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
28495 #else
28496   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
28497 #endif
28498 }
28499
28500 __extension__ extern __inline poly8x8_t
28501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28502 vtrn2_p8 (poly8x8_t __a, poly8x8_t __b)
28503 {
28504 #ifdef __AARCH64EB__
28505   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28506 #else
28507   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28508 #endif
28509 }
28510
28511 __extension__ extern __inline poly16x4_t
28512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28513 vtrn2_p16 (poly16x4_t __a, poly16x4_t __b)
28514 {
28515 #ifdef __AARCH64EB__
28516   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28517 #else
28518   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28519 #endif
28520 }
28521
28522 __extension__ extern __inline int8x8_t
28523 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28524 vtrn2_s8 (int8x8_t __a, int8x8_t __b)
28525 {
28526 #ifdef __AARCH64EB__
28527   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28528 #else
28529   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28530 #endif
28531 }
28532
28533 __extension__ extern __inline int16x4_t
28534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28535 vtrn2_s16 (int16x4_t __a, int16x4_t __b)
28536 {
28537 #ifdef __AARCH64EB__
28538   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28539 #else
28540   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28541 #endif
28542 }
28543
28544 __extension__ extern __inline int32x2_t
28545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28546 vtrn2_s32 (int32x2_t __a, int32x2_t __b)
28547 {
28548 #ifdef __AARCH64EB__
28549   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
28550 #else
28551   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
28552 #endif
28553 }
28554
28555 __extension__ extern __inline uint8x8_t
28556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28557 vtrn2_u8 (uint8x8_t __a, uint8x8_t __b)
28558 {
28559 #ifdef __AARCH64EB__
28560   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28561 #else
28562   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28563 #endif
28564 }
28565
28566 __extension__ extern __inline uint16x4_t
28567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28568 vtrn2_u16 (uint16x4_t __a, uint16x4_t __b)
28569 {
28570 #ifdef __AARCH64EB__
28571   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28572 #else
28573   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28574 #endif
28575 }
28576
28577 __extension__ extern __inline uint32x2_t
28578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28579 vtrn2_u32 (uint32x2_t __a, uint32x2_t __b)
28580 {
28581 #ifdef __AARCH64EB__
28582   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
28583 #else
28584   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
28585 #endif
28586 }
28587
28588 __extension__ extern __inline float16x8_t
28589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28590 vtrn2q_f16 (float16x8_t __a, float16x8_t __b)
28591 {
28592 #ifdef __AARCH64EB__
28593   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28594 #else
28595   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28596 #endif
28597 }
28598
28599 __extension__ extern __inline float32x4_t
28600 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28601 vtrn2q_f32 (float32x4_t __a, float32x4_t __b)
28602 {
28603 #ifdef __AARCH64EB__
28604   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
28605 #else
28606   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
28607 #endif
28608 }
28609
28610 __extension__ extern __inline float64x2_t
28611 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28612 vtrn2q_f64 (float64x2_t __a, float64x2_t __b)
28613 {
28614 #ifdef __AARCH64EB__
28615   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
28616 #else
28617   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
28618 #endif
28619 }
28620
28621 __extension__ extern __inline poly8x16_t
28622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28623 vtrn2q_p8 (poly8x16_t __a, poly8x16_t __b)
28624 {
28625 #ifdef __AARCH64EB__
28626   return __builtin_shuffle (__a, __b,
28627       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
28628 #else
28629   return __builtin_shuffle (__a, __b,
28630       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
28631 #endif
28632 }
28633
28634 __extension__ extern __inline poly16x8_t
28635 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28636 vtrn2q_p16 (poly16x8_t __a, poly16x8_t __b)
28637 {
28638 #ifdef __AARCH64EB__
28639   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28640 #else
28641   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28642 #endif
28643 }
28644
28645 __extension__ extern __inline int8x16_t
28646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28647 vtrn2q_s8 (int8x16_t __a, int8x16_t __b)
28648 {
28649 #ifdef __AARCH64EB__
28650   return __builtin_shuffle (__a, __b,
28651       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
28652 #else
28653   return __builtin_shuffle (__a, __b,
28654       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
28655 #endif
28656 }
28657
28658 __extension__ extern __inline int16x8_t
28659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28660 vtrn2q_s16 (int16x8_t __a, int16x8_t __b)
28661 {
28662 #ifdef __AARCH64EB__
28663   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28664 #else
28665   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28666 #endif
28667 }
28668
28669 __extension__ extern __inline int32x4_t
28670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28671 vtrn2q_s32 (int32x4_t __a, int32x4_t __b)
28672 {
28673 #ifdef __AARCH64EB__
28674   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
28675 #else
28676   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
28677 #endif
28678 }
28679
28680 __extension__ extern __inline int64x2_t
28681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28682 vtrn2q_s64 (int64x2_t __a, int64x2_t __b)
28683 {
28684 #ifdef __AARCH64EB__
28685   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
28686 #else
28687   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
28688 #endif
28689 }
28690
28691 __extension__ extern __inline uint8x16_t
28692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28693 vtrn2q_u8 (uint8x16_t __a, uint8x16_t __b)
28694 {
28695 #ifdef __AARCH64EB__
28696   return __builtin_shuffle (__a, __b,
28697       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
28698 #else
28699   return __builtin_shuffle (__a, __b,
28700       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
28701 #endif
28702 }
28703
28704 __extension__ extern __inline uint16x8_t
28705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28706 vtrn2q_u16 (uint16x8_t __a, uint16x8_t __b)
28707 {
28708 #ifdef __AARCH64EB__
28709   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28710 #else
28711   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28712 #endif
28713 }
28714
28715 __extension__ extern __inline uint32x4_t
28716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28717 vtrn2q_u32 (uint32x4_t __a, uint32x4_t __b)
28718 {
28719 #ifdef __AARCH64EB__
28720   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
28721 #else
28722   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
28723 #endif
28724 }
28725
28726 __extension__ extern __inline uint64x2_t
28727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28728 vtrn2q_u64 (uint64x2_t __a, uint64x2_t __b)
28729 {
28730 #ifdef __AARCH64EB__
28731   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
28732 #else
28733   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
28734 #endif
28735 }
28736
28737 __extension__ extern __inline float16x4x2_t
28738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28739 vtrn_f16 (float16x4_t __a, float16x4_t __b)
28740 {
28741   return (float16x4x2_t) {vtrn1_f16 (__a, __b), vtrn2_f16 (__a, __b)};
28742 }
28743
28744 __extension__ extern __inline float32x2x2_t
28745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28746 vtrn_f32 (float32x2_t a, float32x2_t b)
28747 {
28748   return (float32x2x2_t) {vtrn1_f32 (a, b), vtrn2_f32 (a, b)};
28749 }
28750
28751 __extension__ extern __inline poly8x8x2_t
28752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28753 vtrn_p8 (poly8x8_t a, poly8x8_t b)
28754 {
28755   return (poly8x8x2_t) {vtrn1_p8 (a, b), vtrn2_p8 (a, b)};
28756 }
28757
28758 __extension__ extern __inline poly16x4x2_t
28759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28760 vtrn_p16 (poly16x4_t a, poly16x4_t b)
28761 {
28762   return (poly16x4x2_t) {vtrn1_p16 (a, b), vtrn2_p16 (a, b)};
28763 }
28764
28765 __extension__ extern __inline int8x8x2_t
28766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28767 vtrn_s8 (int8x8_t a, int8x8_t b)
28768 {
28769   return (int8x8x2_t) {vtrn1_s8 (a, b), vtrn2_s8 (a, b)};
28770 }
28771
28772 __extension__ extern __inline int16x4x2_t
28773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28774 vtrn_s16 (int16x4_t a, int16x4_t b)
28775 {
28776   return (int16x4x2_t) {vtrn1_s16 (a, b), vtrn2_s16 (a, b)};
28777 }
28778
28779 __extension__ extern __inline int32x2x2_t
28780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28781 vtrn_s32 (int32x2_t a, int32x2_t b)
28782 {
28783   return (int32x2x2_t) {vtrn1_s32 (a, b), vtrn2_s32 (a, b)};
28784 }
28785
28786 __extension__ extern __inline uint8x8x2_t
28787 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28788 vtrn_u8 (uint8x8_t a, uint8x8_t b)
28789 {
28790   return (uint8x8x2_t) {vtrn1_u8 (a, b), vtrn2_u8 (a, b)};
28791 }
28792
28793 __extension__ extern __inline uint16x4x2_t
28794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28795 vtrn_u16 (uint16x4_t a, uint16x4_t b)
28796 {
28797   return (uint16x4x2_t) {vtrn1_u16 (a, b), vtrn2_u16 (a, b)};
28798 }
28799
28800 __extension__ extern __inline uint32x2x2_t
28801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28802 vtrn_u32 (uint32x2_t a, uint32x2_t b)
28803 {
28804   return (uint32x2x2_t) {vtrn1_u32 (a, b), vtrn2_u32 (a, b)};
28805 }
28806
28807 __extension__ extern __inline float16x8x2_t
28808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28809 vtrnq_f16 (float16x8_t __a, float16x8_t __b)
28810 {
28811   return (float16x8x2_t) {vtrn1q_f16 (__a, __b), vtrn2q_f16 (__a, __b)};
28812 }
28813
28814 __extension__ extern __inline float32x4x2_t
28815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28816 vtrnq_f32 (float32x4_t a, float32x4_t b)
28817 {
28818   return (float32x4x2_t) {vtrn1q_f32 (a, b), vtrn2q_f32 (a, b)};
28819 }
28820
28821 __extension__ extern __inline poly8x16x2_t
28822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28823 vtrnq_p8 (poly8x16_t a, poly8x16_t b)
28824 {
28825   return (poly8x16x2_t) {vtrn1q_p8 (a, b), vtrn2q_p8 (a, b)};
28826 }
28827
28828 __extension__ extern __inline poly16x8x2_t
28829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28830 vtrnq_p16 (poly16x8_t a, poly16x8_t b)
28831 {
28832   return (poly16x8x2_t) {vtrn1q_p16 (a, b), vtrn2q_p16 (a, b)};
28833 }
28834
28835 __extension__ extern __inline int8x16x2_t
28836 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28837 vtrnq_s8 (int8x16_t a, int8x16_t b)
28838 {
28839   return (int8x16x2_t) {vtrn1q_s8 (a, b), vtrn2q_s8 (a, b)};
28840 }
28841
28842 __extension__ extern __inline int16x8x2_t
28843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28844 vtrnq_s16 (int16x8_t a, int16x8_t b)
28845 {
28846   return (int16x8x2_t) {vtrn1q_s16 (a, b), vtrn2q_s16 (a, b)};
28847 }
28848
28849 __extension__ extern __inline int32x4x2_t
28850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28851 vtrnq_s32 (int32x4_t a, int32x4_t b)
28852 {
28853   return (int32x4x2_t) {vtrn1q_s32 (a, b), vtrn2q_s32 (a, b)};
28854 }
28855
28856 __extension__ extern __inline uint8x16x2_t
28857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28858 vtrnq_u8 (uint8x16_t a, uint8x16_t b)
28859 {
28860   return (uint8x16x2_t) {vtrn1q_u8 (a, b), vtrn2q_u8 (a, b)};
28861 }
28862
28863 __extension__ extern __inline uint16x8x2_t
28864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28865 vtrnq_u16 (uint16x8_t a, uint16x8_t b)
28866 {
28867   return (uint16x8x2_t) {vtrn1q_u16 (a, b), vtrn2q_u16 (a, b)};
28868 }
28869
28870 __extension__ extern __inline uint32x4x2_t
28871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28872 vtrnq_u32 (uint32x4_t a, uint32x4_t b)
28873 {
28874   return (uint32x4x2_t) {vtrn1q_u32 (a, b), vtrn2q_u32 (a, b)};
28875 }
28876
28877 /* vtst */
28878
28879 __extension__ extern __inline uint8x8_t
28880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28881 vtst_s8 (int8x8_t __a, int8x8_t __b)
28882 {
28883   return (uint8x8_t) ((__a & __b) != 0);
28884 }
28885
28886 __extension__ extern __inline uint16x4_t
28887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28888 vtst_s16 (int16x4_t __a, int16x4_t __b)
28889 {
28890   return (uint16x4_t) ((__a & __b) != 0);
28891 }
28892
28893 __extension__ extern __inline uint32x2_t
28894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28895 vtst_s32 (int32x2_t __a, int32x2_t __b)
28896 {
28897   return (uint32x2_t) ((__a & __b) != 0);
28898 }
28899
28900 __extension__ extern __inline uint64x1_t
28901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28902 vtst_s64 (int64x1_t __a, int64x1_t __b)
28903 {
28904   return (uint64x1_t) ((__a & __b) != __AARCH64_INT64_C (0));
28905 }
28906
28907 __extension__ extern __inline uint8x8_t
28908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28909 vtst_u8 (uint8x8_t __a, uint8x8_t __b)
28910 {
28911   return ((__a & __b) != 0);
28912 }
28913
28914 __extension__ extern __inline uint16x4_t
28915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28916 vtst_u16 (uint16x4_t __a, uint16x4_t __b)
28917 {
28918   return ((__a & __b) != 0);
28919 }
28920
28921 __extension__ extern __inline uint32x2_t
28922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28923 vtst_u32 (uint32x2_t __a, uint32x2_t __b)
28924 {
28925   return ((__a & __b) != 0);
28926 }
28927
28928 __extension__ extern __inline uint64x1_t
28929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28930 vtst_u64 (uint64x1_t __a, uint64x1_t __b)
28931 {
28932   return ((__a & __b) != __AARCH64_UINT64_C (0));
28933 }
28934
28935 __extension__ extern __inline uint8x16_t
28936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28937 vtstq_s8 (int8x16_t __a, int8x16_t __b)
28938 {
28939   return (uint8x16_t) ((__a & __b) != 0);
28940 }
28941
28942 __extension__ extern __inline uint16x8_t
28943 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28944 vtstq_s16 (int16x8_t __a, int16x8_t __b)
28945 {
28946   return (uint16x8_t) ((__a & __b) != 0);
28947 }
28948
28949 __extension__ extern __inline uint32x4_t
28950 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28951 vtstq_s32 (int32x4_t __a, int32x4_t __b)
28952 {
28953   return (uint32x4_t) ((__a & __b) != 0);
28954 }
28955
28956 __extension__ extern __inline uint64x2_t
28957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28958 vtstq_s64 (int64x2_t __a, int64x2_t __b)
28959 {
28960   return (uint64x2_t) ((__a & __b) != __AARCH64_INT64_C (0));
28961 }
28962
28963 __extension__ extern __inline uint8x16_t
28964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28965 vtstq_u8 (uint8x16_t __a, uint8x16_t __b)
28966 {
28967   return ((__a & __b) != 0);
28968 }
28969
28970 __extension__ extern __inline uint16x8_t
28971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28972 vtstq_u16 (uint16x8_t __a, uint16x8_t __b)
28973 {
28974   return ((__a & __b) != 0);
28975 }
28976
28977 __extension__ extern __inline uint32x4_t
28978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28979 vtstq_u32 (uint32x4_t __a, uint32x4_t __b)
28980 {
28981   return ((__a & __b) != 0);
28982 }
28983
28984 __extension__ extern __inline uint64x2_t
28985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28986 vtstq_u64 (uint64x2_t __a, uint64x2_t __b)
28987 {
28988   return ((__a & __b) != __AARCH64_UINT64_C (0));
28989 }
28990
28991 __extension__ extern __inline uint64_t
28992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28993 vtstd_s64 (int64_t __a, int64_t __b)
28994 {
28995   return (__a & __b) ? -1ll : 0ll;
28996 }
28997
28998 __extension__ extern __inline uint64_t
28999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29000 vtstd_u64 (uint64_t __a, uint64_t __b)
29001 {
29002   return (__a & __b) ? -1ll : 0ll;
29003 }
29004
29005 /* vuqadd */
29006
29007 __extension__ extern __inline int8x8_t
29008 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29009 vuqadd_s8 (int8x8_t __a, uint8x8_t __b)
29010 {
29011   return __builtin_aarch64_suqaddv8qi_ssu (__a,  __b);
29012 }
29013
29014 __extension__ extern __inline int16x4_t
29015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29016 vuqadd_s16 (int16x4_t __a, uint16x4_t __b)
29017 {
29018   return __builtin_aarch64_suqaddv4hi_ssu (__a,  __b);
29019 }
29020
29021 __extension__ extern __inline int32x2_t
29022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29023 vuqadd_s32 (int32x2_t __a, uint32x2_t __b)
29024 {
29025   return __builtin_aarch64_suqaddv2si_ssu (__a,  __b);
29026 }
29027
29028 __extension__ extern __inline int64x1_t
29029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29030 vuqadd_s64 (int64x1_t __a, uint64x1_t __b)
29031 {
29032   return (int64x1_t) {__builtin_aarch64_suqadddi_ssu (__a[0], __b[0])};
29033 }
29034
29035 __extension__ extern __inline int8x16_t
29036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29037 vuqaddq_s8 (int8x16_t __a, uint8x16_t __b)
29038 {
29039   return __builtin_aarch64_suqaddv16qi_ssu (__a,  __b);
29040 }
29041
29042 __extension__ extern __inline int16x8_t
29043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29044 vuqaddq_s16 (int16x8_t __a, uint16x8_t __b)
29045 {
29046   return __builtin_aarch64_suqaddv8hi_ssu (__a,  __b);
29047 }
29048
29049 __extension__ extern __inline int32x4_t
29050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29051 vuqaddq_s32 (int32x4_t __a, uint32x4_t __b)
29052 {
29053   return __builtin_aarch64_suqaddv4si_ssu (__a,  __b);
29054 }
29055
29056 __extension__ extern __inline int64x2_t
29057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29058 vuqaddq_s64 (int64x2_t __a, uint64x2_t __b)
29059 {
29060   return __builtin_aarch64_suqaddv2di_ssu (__a,  __b);
29061 }
29062
29063 __extension__ extern __inline int8_t
29064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29065 vuqaddb_s8 (int8_t __a, uint8_t __b)
29066 {
29067   return __builtin_aarch64_suqaddqi_ssu (__a,  __b);
29068 }
29069
29070 __extension__ extern __inline int16_t
29071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29072 vuqaddh_s16 (int16_t __a, uint16_t __b)
29073 {
29074   return __builtin_aarch64_suqaddhi_ssu (__a,  __b);
29075 }
29076
29077 __extension__ extern __inline int32_t
29078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29079 vuqadds_s32 (int32_t __a, uint32_t __b)
29080 {
29081   return __builtin_aarch64_suqaddsi_ssu (__a,  __b);
29082 }
29083
29084 __extension__ extern __inline int64_t
29085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29086 vuqaddd_s64 (int64_t __a, uint64_t __b)
29087 {
29088   return __builtin_aarch64_suqadddi_ssu (__a,  __b);
29089 }
29090
29091 #define __DEFINTERLEAVE(op, rettype, intype, funcsuffix, Q)             \
29092   __extension__ extern __inline rettype                                 \
29093   __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
29094   v ## op ## Q ## _ ## funcsuffix (intype a, intype b)                  \
29095   {                                                                     \
29096     return (rettype) {v ## op ## 1 ## Q ## _ ## funcsuffix (a, b),      \
29097                       v ## op ## 2 ## Q ## _ ## funcsuffix (a, b)};     \
29098   }
29099
29100 #define __INTERLEAVE_LIST(op)                                   \
29101   __DEFINTERLEAVE (op, float16x4x2_t, float16x4_t, f16,)        \
29102   __DEFINTERLEAVE (op, float32x2x2_t, float32x2_t, f32,)        \
29103   __DEFINTERLEAVE (op, poly8x8x2_t, poly8x8_t, p8,)             \
29104   __DEFINTERLEAVE (op, poly16x4x2_t, poly16x4_t, p16,)          \
29105   __DEFINTERLEAVE (op, int8x8x2_t, int8x8_t, s8,)               \
29106   __DEFINTERLEAVE (op, int16x4x2_t, int16x4_t, s16,)            \
29107   __DEFINTERLEAVE (op, int32x2x2_t, int32x2_t, s32,)            \
29108   __DEFINTERLEAVE (op, uint8x8x2_t, uint8x8_t, u8,)             \
29109   __DEFINTERLEAVE (op, uint16x4x2_t, uint16x4_t, u16,)          \
29110   __DEFINTERLEAVE (op, uint32x2x2_t, uint32x2_t, u32,)          \
29111   __DEFINTERLEAVE (op, float16x8x2_t, float16x8_t, f16, q)      \
29112   __DEFINTERLEAVE (op, float32x4x2_t, float32x4_t, f32, q)      \
29113   __DEFINTERLEAVE (op, poly8x16x2_t, poly8x16_t, p8, q)         \
29114   __DEFINTERLEAVE (op, poly16x8x2_t, poly16x8_t, p16, q)        \
29115   __DEFINTERLEAVE (op, int8x16x2_t, int8x16_t, s8, q)           \
29116   __DEFINTERLEAVE (op, int16x8x2_t, int16x8_t, s16, q)          \
29117   __DEFINTERLEAVE (op, int32x4x2_t, int32x4_t, s32, q)          \
29118   __DEFINTERLEAVE (op, uint8x16x2_t, uint8x16_t, u8, q)         \
29119   __DEFINTERLEAVE (op, uint16x8x2_t, uint16x8_t, u16, q)        \
29120   __DEFINTERLEAVE (op, uint32x4x2_t, uint32x4_t, u32, q)
29121
29122 /* vuzp */
29123
29124 __extension__ extern __inline float16x4_t
29125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29126 vuzp1_f16 (float16x4_t __a, float16x4_t __b)
29127 {
29128 #ifdef __AARCH64EB__
29129   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29130 #else
29131   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29132 #endif
29133 }
29134
29135 __extension__ extern __inline float32x2_t
29136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29137 vuzp1_f32 (float32x2_t __a, float32x2_t __b)
29138 {
29139 #ifdef __AARCH64EB__
29140   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29141 #else
29142   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29143 #endif
29144 }
29145
29146 __extension__ extern __inline poly8x8_t
29147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29148 vuzp1_p8 (poly8x8_t __a, poly8x8_t __b)
29149 {
29150 #ifdef __AARCH64EB__
29151   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29152 #else
29153   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29154 #endif
29155 }
29156
29157 __extension__ extern __inline poly16x4_t
29158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29159 vuzp1_p16 (poly16x4_t __a, poly16x4_t __b)
29160 {
29161 #ifdef __AARCH64EB__
29162   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29163 #else
29164   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29165 #endif
29166 }
29167
29168 __extension__ extern __inline int8x8_t
29169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29170 vuzp1_s8 (int8x8_t __a, int8x8_t __b)
29171 {
29172 #ifdef __AARCH64EB__
29173   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29174 #else
29175   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29176 #endif
29177 }
29178
29179 __extension__ extern __inline int16x4_t
29180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29181 vuzp1_s16 (int16x4_t __a, int16x4_t __b)
29182 {
29183 #ifdef __AARCH64EB__
29184   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29185 #else
29186   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29187 #endif
29188 }
29189
29190 __extension__ extern __inline int32x2_t
29191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29192 vuzp1_s32 (int32x2_t __a, int32x2_t __b)
29193 {
29194 #ifdef __AARCH64EB__
29195   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29196 #else
29197   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29198 #endif
29199 }
29200
29201 __extension__ extern __inline uint8x8_t
29202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29203 vuzp1_u8 (uint8x8_t __a, uint8x8_t __b)
29204 {
29205 #ifdef __AARCH64EB__
29206   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29207 #else
29208   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29209 #endif
29210 }
29211
29212 __extension__ extern __inline uint16x4_t
29213 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29214 vuzp1_u16 (uint16x4_t __a, uint16x4_t __b)
29215 {
29216 #ifdef __AARCH64EB__
29217   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29218 #else
29219   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29220 #endif
29221 }
29222
29223 __extension__ extern __inline uint32x2_t
29224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29225 vuzp1_u32 (uint32x2_t __a, uint32x2_t __b)
29226 {
29227 #ifdef __AARCH64EB__
29228   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29229 #else
29230   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29231 #endif
29232 }
29233
29234 __extension__ extern __inline float16x8_t
29235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29236 vuzp1q_f16 (float16x8_t __a, float16x8_t __b)
29237 {
29238 #ifdef __AARCH64EB__
29239   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29240 #else
29241   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29242 #endif
29243 }
29244
29245 __extension__ extern __inline float32x4_t
29246 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29247 vuzp1q_f32 (float32x4_t __a, float32x4_t __b)
29248 {
29249 #ifdef __AARCH64EB__
29250   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
29251 #else
29252   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
29253 #endif
29254 }
29255
29256 __extension__ extern __inline float64x2_t
29257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29258 vuzp1q_f64 (float64x2_t __a, float64x2_t __b)
29259 {
29260 #ifdef __AARCH64EB__
29261   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29262 #else
29263   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29264 #endif
29265 }
29266
29267 __extension__ extern __inline poly8x16_t
29268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29269 vuzp1q_p8 (poly8x16_t __a, poly8x16_t __b)
29270 {
29271 #ifdef __AARCH64EB__
29272   return __builtin_shuffle (__a, __b, (uint8x16_t)
29273       {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
29274 #else
29275   return __builtin_shuffle (__a, __b, (uint8x16_t)
29276       {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
29277 #endif
29278 }
29279
29280 __extension__ extern __inline poly16x8_t
29281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29282 vuzp1q_p16 (poly16x8_t __a, poly16x8_t __b)
29283 {
29284 #ifdef __AARCH64EB__
29285   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29286 #else
29287   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29288 #endif
29289 }
29290
29291 __extension__ extern __inline int8x16_t
29292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29293 vuzp1q_s8 (int8x16_t __a, int8x16_t __b)
29294 {
29295 #ifdef __AARCH64EB__
29296   return __builtin_shuffle (__a, __b,
29297       (uint8x16_t) {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
29298 #else
29299   return __builtin_shuffle (__a, __b,
29300       (uint8x16_t) {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
29301 #endif
29302 }
29303
29304 __extension__ extern __inline int16x8_t
29305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29306 vuzp1q_s16 (int16x8_t __a, int16x8_t __b)
29307 {
29308 #ifdef __AARCH64EB__
29309   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29310 #else
29311   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29312 #endif
29313 }
29314
29315 __extension__ extern __inline int32x4_t
29316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29317 vuzp1q_s32 (int32x4_t __a, int32x4_t __b)
29318 {
29319 #ifdef __AARCH64EB__
29320   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
29321 #else
29322   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
29323 #endif
29324 }
29325
29326 __extension__ extern __inline int64x2_t
29327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29328 vuzp1q_s64 (int64x2_t __a, int64x2_t __b)
29329 {
29330 #ifdef __AARCH64EB__
29331   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29332 #else
29333   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29334 #endif
29335 }
29336
29337 __extension__ extern __inline uint8x16_t
29338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29339 vuzp1q_u8 (uint8x16_t __a, uint8x16_t __b)
29340 {
29341 #ifdef __AARCH64EB__
29342   return __builtin_shuffle (__a, __b,
29343       (uint8x16_t) {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
29344 #else
29345   return __builtin_shuffle (__a, __b,
29346       (uint8x16_t) {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
29347 #endif
29348 }
29349
29350 __extension__ extern __inline uint16x8_t
29351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29352 vuzp1q_u16 (uint16x8_t __a, uint16x8_t __b)
29353 {
29354 #ifdef __AARCH64EB__
29355   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29356 #else
29357   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29358 #endif
29359 }
29360
29361 __extension__ extern __inline uint32x4_t
29362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29363 vuzp1q_u32 (uint32x4_t __a, uint32x4_t __b)
29364 {
29365 #ifdef __AARCH64EB__
29366   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
29367 #else
29368   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
29369 #endif
29370 }
29371
29372 __extension__ extern __inline uint64x2_t
29373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29374 vuzp1q_u64 (uint64x2_t __a, uint64x2_t __b)
29375 {
29376 #ifdef __AARCH64EB__
29377   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29378 #else
29379   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29380 #endif
29381 }
29382
29383 __extension__ extern __inline float16x4_t
29384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29385 vuzp2_f16 (float16x4_t __a, float16x4_t __b)
29386 {
29387 #ifdef __AARCH64EB__
29388   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29389 #else
29390   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29391 #endif
29392 }
29393
29394 __extension__ extern __inline float32x2_t
29395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29396 vuzp2_f32 (float32x2_t __a, float32x2_t __b)
29397 {
29398 #ifdef __AARCH64EB__
29399   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29400 #else
29401   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29402 #endif
29403 }
29404
29405 __extension__ extern __inline poly8x8_t
29406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29407 vuzp2_p8 (poly8x8_t __a, poly8x8_t __b)
29408 {
29409 #ifdef __AARCH64EB__
29410   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29411 #else
29412   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29413 #endif
29414 }
29415
29416 __extension__ extern __inline poly16x4_t
29417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29418 vuzp2_p16 (poly16x4_t __a, poly16x4_t __b)
29419 {
29420 #ifdef __AARCH64EB__
29421   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29422 #else
29423   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29424 #endif
29425 }
29426
29427 __extension__ extern __inline int8x8_t
29428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29429 vuzp2_s8 (int8x8_t __a, int8x8_t __b)
29430 {
29431 #ifdef __AARCH64EB__
29432   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29433 #else
29434   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29435 #endif
29436 }
29437
29438 __extension__ extern __inline int16x4_t
29439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29440 vuzp2_s16 (int16x4_t __a, int16x4_t __b)
29441 {
29442 #ifdef __AARCH64EB__
29443   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29444 #else
29445   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29446 #endif
29447 }
29448
29449 __extension__ extern __inline int32x2_t
29450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29451 vuzp2_s32 (int32x2_t __a, int32x2_t __b)
29452 {
29453 #ifdef __AARCH64EB__
29454   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29455 #else
29456   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29457 #endif
29458 }
29459
29460 __extension__ extern __inline uint8x8_t
29461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29462 vuzp2_u8 (uint8x8_t __a, uint8x8_t __b)
29463 {
29464 #ifdef __AARCH64EB__
29465   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29466 #else
29467   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29468 #endif
29469 }
29470
29471 __extension__ extern __inline uint16x4_t
29472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29473 vuzp2_u16 (uint16x4_t __a, uint16x4_t __b)
29474 {
29475 #ifdef __AARCH64EB__
29476   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29477 #else
29478   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29479 #endif
29480 }
29481
29482 __extension__ extern __inline uint32x2_t
29483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29484 vuzp2_u32 (uint32x2_t __a, uint32x2_t __b)
29485 {
29486 #ifdef __AARCH64EB__
29487   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29488 #else
29489   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29490 #endif
29491 }
29492
29493 __extension__ extern __inline float16x8_t
29494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29495 vuzp2q_f16 (float16x8_t __a, float16x8_t __b)
29496 {
29497 #ifdef __AARCH64EB__
29498   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29499 #else
29500   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29501 #endif
29502 }
29503
29504 __extension__ extern __inline float32x4_t
29505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29506 vuzp2q_f32 (float32x4_t __a, float32x4_t __b)
29507 {
29508 #ifdef __AARCH64EB__
29509   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
29510 #else
29511   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
29512 #endif
29513 }
29514
29515 __extension__ extern __inline float64x2_t
29516 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29517 vuzp2q_f64 (float64x2_t __a, float64x2_t __b)
29518 {
29519 #ifdef __AARCH64EB__
29520   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29521 #else
29522   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29523 #endif
29524 }
29525
29526 __extension__ extern __inline poly8x16_t
29527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29528 vuzp2q_p8 (poly8x16_t __a, poly8x16_t __b)
29529 {
29530 #ifdef __AARCH64EB__
29531   return __builtin_shuffle (__a, __b,
29532       (uint8x16_t) {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
29533 #else
29534   return __builtin_shuffle (__a, __b,
29535       (uint8x16_t) {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
29536 #endif
29537 }
29538
29539 __extension__ extern __inline poly16x8_t
29540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29541 vuzp2q_p16 (poly16x8_t __a, poly16x8_t __b)
29542 {
29543 #ifdef __AARCH64EB__
29544   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29545 #else
29546   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29547 #endif
29548 }
29549
29550 __extension__ extern __inline int8x16_t
29551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29552 vuzp2q_s8 (int8x16_t __a, int8x16_t __b)
29553 {
29554 #ifdef __AARCH64EB__
29555   return __builtin_shuffle (__a, __b,
29556       (uint8x16_t) {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
29557 #else
29558   return __builtin_shuffle (__a, __b,
29559       (uint8x16_t) {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
29560 #endif
29561 }
29562
29563 __extension__ extern __inline int16x8_t
29564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29565 vuzp2q_s16 (int16x8_t __a, int16x8_t __b)
29566 {
29567 #ifdef __AARCH64EB__
29568   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29569 #else
29570   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29571 #endif
29572 }
29573
29574 __extension__ extern __inline int32x4_t
29575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29576 vuzp2q_s32 (int32x4_t __a, int32x4_t __b)
29577 {
29578 #ifdef __AARCH64EB__
29579   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
29580 #else
29581   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
29582 #endif
29583 }
29584
29585 __extension__ extern __inline int64x2_t
29586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29587 vuzp2q_s64 (int64x2_t __a, int64x2_t __b)
29588 {
29589 #ifdef __AARCH64EB__
29590   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29591 #else
29592   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29593 #endif
29594 }
29595
29596 __extension__ extern __inline uint8x16_t
29597 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29598 vuzp2q_u8 (uint8x16_t __a, uint8x16_t __b)
29599 {
29600 #ifdef __AARCH64EB__
29601   return __builtin_shuffle (__a, __b, (uint8x16_t)
29602       {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
29603 #else
29604   return __builtin_shuffle (__a, __b, (uint8x16_t)
29605       {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
29606 #endif
29607 }
29608
29609 __extension__ extern __inline uint16x8_t
29610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29611 vuzp2q_u16 (uint16x8_t __a, uint16x8_t __b)
29612 {
29613 #ifdef __AARCH64EB__
29614   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29615 #else
29616   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29617 #endif
29618 }
29619
29620 __extension__ extern __inline uint32x4_t
29621 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29622 vuzp2q_u32 (uint32x4_t __a, uint32x4_t __b)
29623 {
29624 #ifdef __AARCH64EB__
29625   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
29626 #else
29627   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
29628 #endif
29629 }
29630
29631 __extension__ extern __inline uint64x2_t
29632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29633 vuzp2q_u64 (uint64x2_t __a, uint64x2_t __b)
29634 {
29635 #ifdef __AARCH64EB__
29636   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29637 #else
29638   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29639 #endif
29640 }
29641
29642 __INTERLEAVE_LIST (uzp)
29643
29644 /* vzip */
29645
29646 __extension__ extern __inline float16x4_t
29647 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29648 vzip1_f16 (float16x4_t __a, float16x4_t __b)
29649 {
29650 #ifdef __AARCH64EB__
29651   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29652 #else
29653   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29654 #endif
29655 }
29656
29657 __extension__ extern __inline float32x2_t
29658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29659 vzip1_f32 (float32x2_t __a, float32x2_t __b)
29660 {
29661 #ifdef __AARCH64EB__
29662   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29663 #else
29664   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29665 #endif
29666 }
29667
29668 __extension__ extern __inline poly8x8_t
29669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29670 vzip1_p8 (poly8x8_t __a, poly8x8_t __b)
29671 {
29672 #ifdef __AARCH64EB__
29673   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29674 #else
29675   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29676 #endif
29677 }
29678
29679 __extension__ extern __inline poly16x4_t
29680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29681 vzip1_p16 (poly16x4_t __a, poly16x4_t __b)
29682 {
29683 #ifdef __AARCH64EB__
29684   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29685 #else
29686   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29687 #endif
29688 }
29689
29690 __extension__ extern __inline int8x8_t
29691 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29692 vzip1_s8 (int8x8_t __a, int8x8_t __b)
29693 {
29694 #ifdef __AARCH64EB__
29695   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29696 #else
29697   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29698 #endif
29699 }
29700
29701 __extension__ extern __inline int16x4_t
29702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29703 vzip1_s16 (int16x4_t __a, int16x4_t __b)
29704 {
29705 #ifdef __AARCH64EB__
29706   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29707 #else
29708   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29709 #endif
29710 }
29711
29712 __extension__ extern __inline int32x2_t
29713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29714 vzip1_s32 (int32x2_t __a, int32x2_t __b)
29715 {
29716 #ifdef __AARCH64EB__
29717   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29718 #else
29719   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29720 #endif
29721 }
29722
29723 __extension__ extern __inline uint8x8_t
29724 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29725 vzip1_u8 (uint8x8_t __a, uint8x8_t __b)
29726 {
29727 #ifdef __AARCH64EB__
29728   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29729 #else
29730   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29731 #endif
29732 }
29733
29734 __extension__ extern __inline uint16x4_t
29735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29736 vzip1_u16 (uint16x4_t __a, uint16x4_t __b)
29737 {
29738 #ifdef __AARCH64EB__
29739   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29740 #else
29741   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29742 #endif
29743 }
29744
29745 __extension__ extern __inline uint32x2_t
29746 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29747 vzip1_u32 (uint32x2_t __a, uint32x2_t __b)
29748 {
29749 #ifdef __AARCH64EB__
29750   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29751 #else
29752   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29753 #endif
29754 }
29755
29756 __extension__ extern __inline float16x8_t
29757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29758 vzip1q_f16 (float16x8_t __a, float16x8_t __b)
29759 {
29760 #ifdef __AARCH64EB__
29761   return __builtin_shuffle (__a, __b,
29762                             (uint16x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29763 #else
29764   return __builtin_shuffle (__a, __b,
29765                             (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29766 #endif
29767 }
29768
29769 __extension__ extern __inline float32x4_t
29770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29771 vzip1q_f32 (float32x4_t __a, float32x4_t __b)
29772 {
29773 #ifdef __AARCH64EB__
29774   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
29775 #else
29776   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
29777 #endif
29778 }
29779
29780 __extension__ extern __inline float64x2_t
29781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29782 vzip1q_f64 (float64x2_t __a, float64x2_t __b)
29783 {
29784 #ifdef __AARCH64EB__
29785   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29786 #else
29787   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29788 #endif
29789 }
29790
29791 __extension__ extern __inline poly8x16_t
29792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29793 vzip1q_p8 (poly8x16_t __a, poly8x16_t __b)
29794 {
29795 #ifdef __AARCH64EB__
29796   return __builtin_shuffle (__a, __b, (uint8x16_t)
29797       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
29798 #else
29799   return __builtin_shuffle (__a, __b, (uint8x16_t)
29800       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
29801 #endif
29802 }
29803
29804 __extension__ extern __inline poly16x8_t
29805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29806 vzip1q_p16 (poly16x8_t __a, poly16x8_t __b)
29807 {
29808 #ifdef __AARCH64EB__
29809   return __builtin_shuffle (__a, __b, (uint16x8_t)
29810       {12, 4, 13, 5, 14, 6, 15, 7});
29811 #else
29812   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29813 #endif
29814 }
29815
29816 __extension__ extern __inline int8x16_t
29817 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29818 vzip1q_s8 (int8x16_t __a, int8x16_t __b)
29819 {
29820 #ifdef __AARCH64EB__
29821   return __builtin_shuffle (__a, __b, (uint8x16_t)
29822       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
29823 #else
29824   return __builtin_shuffle (__a, __b, (uint8x16_t)
29825       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
29826 #endif
29827 }
29828
29829 __extension__ extern __inline int16x8_t
29830 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29831 vzip1q_s16 (int16x8_t __a, int16x8_t __b)
29832 {
29833 #ifdef __AARCH64EB__
29834   return __builtin_shuffle (__a, __b, (uint16x8_t)
29835       {12, 4, 13, 5, 14, 6, 15, 7});
29836 #else
29837   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29838 #endif
29839 }
29840
29841 __extension__ extern __inline int32x4_t
29842 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29843 vzip1q_s32 (int32x4_t __a, int32x4_t __b)
29844 {
29845 #ifdef __AARCH64EB__
29846   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
29847 #else
29848   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
29849 #endif
29850 }
29851
29852 __extension__ extern __inline int64x2_t
29853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29854 vzip1q_s64 (int64x2_t __a, int64x2_t __b)
29855 {
29856 #ifdef __AARCH64EB__
29857   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29858 #else
29859   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29860 #endif
29861 }
29862
29863 __extension__ extern __inline uint8x16_t
29864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29865 vzip1q_u8 (uint8x16_t __a, uint8x16_t __b)
29866 {
29867 #ifdef __AARCH64EB__
29868   return __builtin_shuffle (__a, __b, (uint8x16_t)
29869       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
29870 #else
29871   return __builtin_shuffle (__a, __b, (uint8x16_t)
29872       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
29873 #endif
29874 }
29875
29876 __extension__ extern __inline uint16x8_t
29877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29878 vzip1q_u16 (uint16x8_t __a, uint16x8_t __b)
29879 {
29880 #ifdef __AARCH64EB__
29881   return __builtin_shuffle (__a, __b, (uint16x8_t)
29882       {12, 4, 13, 5, 14, 6, 15, 7});
29883 #else
29884   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29885 #endif
29886 }
29887
29888 __extension__ extern __inline uint32x4_t
29889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29890 vzip1q_u32 (uint32x4_t __a, uint32x4_t __b)
29891 {
29892 #ifdef __AARCH64EB__
29893   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
29894 #else
29895   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
29896 #endif
29897 }
29898
29899 __extension__ extern __inline uint64x2_t
29900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29901 vzip1q_u64 (uint64x2_t __a, uint64x2_t __b)
29902 {
29903 #ifdef __AARCH64EB__
29904   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29905 #else
29906   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29907 #endif
29908 }
29909
29910 __extension__ extern __inline float16x4_t
29911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29912 vzip2_f16 (float16x4_t __a, float16x4_t __b)
29913 {
29914 #ifdef __AARCH64EB__
29915   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29916 #else
29917   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29918 #endif
29919 }
29920
29921 __extension__ extern __inline float32x2_t
29922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29923 vzip2_f32 (float32x2_t __a, float32x2_t __b)
29924 {
29925 #ifdef __AARCH64EB__
29926   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29927 #else
29928   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29929 #endif
29930 }
29931
29932 __extension__ extern __inline poly8x8_t
29933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29934 vzip2_p8 (poly8x8_t __a, poly8x8_t __b)
29935 {
29936 #ifdef __AARCH64EB__
29937   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29938 #else
29939   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29940 #endif
29941 }
29942
29943 __extension__ extern __inline poly16x4_t
29944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29945 vzip2_p16 (poly16x4_t __a, poly16x4_t __b)
29946 {
29947 #ifdef __AARCH64EB__
29948   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29949 #else
29950   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29951 #endif
29952 }
29953
29954 __extension__ extern __inline int8x8_t
29955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29956 vzip2_s8 (int8x8_t __a, int8x8_t __b)
29957 {
29958 #ifdef __AARCH64EB__
29959   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29960 #else
29961   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29962 #endif
29963 }
29964
29965 __extension__ extern __inline int16x4_t
29966 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29967 vzip2_s16 (int16x4_t __a, int16x4_t __b)
29968 {
29969 #ifdef __AARCH64EB__
29970   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29971 #else
29972   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29973 #endif
29974 }
29975
29976 __extension__ extern __inline int32x2_t
29977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29978 vzip2_s32 (int32x2_t __a, int32x2_t __b)
29979 {
29980 #ifdef __AARCH64EB__
29981   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29982 #else
29983   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29984 #endif
29985 }
29986
29987 __extension__ extern __inline uint8x8_t
29988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29989 vzip2_u8 (uint8x8_t __a, uint8x8_t __b)
29990 {
29991 #ifdef __AARCH64EB__
29992   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29993 #else
29994   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29995 #endif
29996 }
29997
29998 __extension__ extern __inline uint16x4_t
29999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30000 vzip2_u16 (uint16x4_t __a, uint16x4_t __b)
30001 {
30002 #ifdef __AARCH64EB__
30003   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
30004 #else
30005   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
30006 #endif
30007 }
30008
30009 __extension__ extern __inline uint32x2_t
30010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30011 vzip2_u32 (uint32x2_t __a, uint32x2_t __b)
30012 {
30013 #ifdef __AARCH64EB__
30014   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
30015 #else
30016   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
30017 #endif
30018 }
30019
30020 __extension__ extern __inline float16x8_t
30021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30022 vzip2q_f16 (float16x8_t __a, float16x8_t __b)
30023 {
30024 #ifdef __AARCH64EB__
30025   return __builtin_shuffle (__a, __b,
30026                             (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30027 #else
30028   return __builtin_shuffle (__a, __b,
30029                             (uint16x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
30030 #endif
30031 }
30032
30033 __extension__ extern __inline float32x4_t
30034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30035 vzip2q_f32 (float32x4_t __a, float32x4_t __b)
30036 {
30037 #ifdef __AARCH64EB__
30038   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
30039 #else
30040   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
30041 #endif
30042 }
30043
30044 __extension__ extern __inline float64x2_t
30045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30046 vzip2q_f64 (float64x2_t __a, float64x2_t __b)
30047 {
30048 #ifdef __AARCH64EB__
30049   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30050 #else
30051   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30052 #endif
30053 }
30054
30055 __extension__ extern __inline poly8x16_t
30056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30057 vzip2q_p8 (poly8x16_t __a, poly8x16_t __b)
30058 {
30059 #ifdef __AARCH64EB__
30060   return __builtin_shuffle (__a, __b, (uint8x16_t)
30061       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
30062 #else
30063   return __builtin_shuffle (__a, __b, (uint8x16_t)
30064       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
30065 #endif
30066 }
30067
30068 __extension__ extern __inline poly16x8_t
30069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30070 vzip2q_p16 (poly16x8_t __a, poly16x8_t __b)
30071 {
30072 #ifdef __AARCH64EB__
30073   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30074 #else
30075   return __builtin_shuffle (__a, __b, (uint16x8_t)
30076       {4, 12, 5, 13, 6, 14, 7, 15});
30077 #endif
30078 }
30079
30080 __extension__ extern __inline int8x16_t
30081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30082 vzip2q_s8 (int8x16_t __a, int8x16_t __b)
30083 {
30084 #ifdef __AARCH64EB__
30085   return __builtin_shuffle (__a, __b, (uint8x16_t)
30086       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
30087 #else
30088   return __builtin_shuffle (__a, __b, (uint8x16_t)
30089       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
30090 #endif
30091 }
30092
30093 __extension__ extern __inline int16x8_t
30094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30095 vzip2q_s16 (int16x8_t __a, int16x8_t __b)
30096 {
30097 #ifdef __AARCH64EB__
30098   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30099 #else
30100   return __builtin_shuffle (__a, __b, (uint16x8_t)
30101       {4, 12, 5, 13, 6, 14, 7, 15});
30102 #endif
30103 }
30104
30105 __extension__ extern __inline int32x4_t
30106 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30107 vzip2q_s32 (int32x4_t __a, int32x4_t __b)
30108 {
30109 #ifdef __AARCH64EB__
30110   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
30111 #else
30112   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
30113 #endif
30114 }
30115
30116 __extension__ extern __inline int64x2_t
30117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30118 vzip2q_s64 (int64x2_t __a, int64x2_t __b)
30119 {
30120 #ifdef __AARCH64EB__
30121   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30122 #else
30123   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30124 #endif
30125 }
30126
30127 __extension__ extern __inline uint8x16_t
30128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30129 vzip2q_u8 (uint8x16_t __a, uint8x16_t __b)
30130 {
30131 #ifdef __AARCH64EB__
30132   return __builtin_shuffle (__a, __b, (uint8x16_t)
30133       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
30134 #else
30135   return __builtin_shuffle (__a, __b, (uint8x16_t)
30136       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
30137 #endif
30138 }
30139
30140 __extension__ extern __inline uint16x8_t
30141 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30142 vzip2q_u16 (uint16x8_t __a, uint16x8_t __b)
30143 {
30144 #ifdef __AARCH64EB__
30145   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30146 #else
30147   return __builtin_shuffle (__a, __b, (uint16x8_t)
30148       {4, 12, 5, 13, 6, 14, 7, 15});
30149 #endif
30150 }
30151
30152 __extension__ extern __inline uint32x4_t
30153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30154 vzip2q_u32 (uint32x4_t __a, uint32x4_t __b)
30155 {
30156 #ifdef __AARCH64EB__
30157   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
30158 #else
30159   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
30160 #endif
30161 }
30162
30163 __extension__ extern __inline uint64x2_t
30164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30165 vzip2q_u64 (uint64x2_t __a, uint64x2_t __b)
30166 {
30167 #ifdef __AARCH64EB__
30168   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30169 #else
30170   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30171 #endif
30172 }
30173
30174 __INTERLEAVE_LIST (zip)
30175
30176 #undef __INTERLEAVE_LIST
30177 #undef __DEFINTERLEAVE
30178
30179 /* End of optimal implementations in approved order.  */
30180
30181 #pragma GCC pop_options
30182
30183 /* ARMv8.2-A FP16 intrinsics.  */
30184
30185 #include "arm_fp16.h"
30186
30187 #pragma GCC push_options
30188 #pragma GCC target ("arch=armv8.2-a+fp16")
30189
30190 /* ARMv8.2-A FP16 one operand vector intrinsics.  */
30191
30192 __extension__ extern __inline float16x4_t
30193 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30194 vabs_f16 (float16x4_t __a)
30195 {
30196   return __builtin_aarch64_absv4hf (__a);
30197 }
30198
30199 __extension__ extern __inline float16x8_t
30200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30201 vabsq_f16 (float16x8_t __a)
30202 {
30203   return __builtin_aarch64_absv8hf (__a);
30204 }
30205
30206 __extension__ extern __inline uint16x4_t
30207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30208 vceqz_f16 (float16x4_t __a)
30209 {
30210   return __builtin_aarch64_cmeqv4hf_uss (__a, vdup_n_f16 (0.0f));
30211 }
30212
30213 __extension__ extern __inline uint16x8_t
30214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30215 vceqzq_f16 (float16x8_t __a)
30216 {
30217   return __builtin_aarch64_cmeqv8hf_uss (__a, vdupq_n_f16 (0.0f));
30218 }
30219
30220 __extension__ extern __inline uint16x4_t
30221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30222 vcgez_f16 (float16x4_t __a)
30223 {
30224   return __builtin_aarch64_cmgev4hf_uss (__a, vdup_n_f16 (0.0f));
30225 }
30226
30227 __extension__ extern __inline uint16x8_t
30228 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30229 vcgezq_f16 (float16x8_t __a)
30230 {
30231   return __builtin_aarch64_cmgev8hf_uss (__a, vdupq_n_f16 (0.0f));
30232 }
30233
30234 __extension__ extern __inline uint16x4_t
30235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30236 vcgtz_f16 (float16x4_t __a)
30237 {
30238   return __builtin_aarch64_cmgtv4hf_uss (__a, vdup_n_f16 (0.0f));
30239 }
30240
30241 __extension__ extern __inline uint16x8_t
30242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30243 vcgtzq_f16 (float16x8_t __a)
30244 {
30245   return __builtin_aarch64_cmgtv8hf_uss (__a, vdupq_n_f16 (0.0f));
30246 }
30247
30248 __extension__ extern __inline uint16x4_t
30249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30250 vclez_f16 (float16x4_t __a)
30251 {
30252   return __builtin_aarch64_cmlev4hf_uss (__a, vdup_n_f16 (0.0f));
30253 }
30254
30255 __extension__ extern __inline uint16x8_t
30256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30257 vclezq_f16 (float16x8_t __a)
30258 {
30259   return __builtin_aarch64_cmlev8hf_uss (__a, vdupq_n_f16 (0.0f));
30260 }
30261
30262 __extension__ extern __inline uint16x4_t
30263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30264 vcltz_f16 (float16x4_t __a)
30265 {
30266   return __builtin_aarch64_cmltv4hf_uss (__a, vdup_n_f16 (0.0f));
30267 }
30268
30269 __extension__ extern __inline uint16x8_t
30270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30271 vcltzq_f16 (float16x8_t __a)
30272 {
30273   return __builtin_aarch64_cmltv8hf_uss (__a, vdupq_n_f16 (0.0f));
30274 }
30275
30276 __extension__ extern __inline float16x4_t
30277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30278 vcvt_f16_s16 (int16x4_t __a)
30279 {
30280   return __builtin_aarch64_floatv4hiv4hf (__a);
30281 }
30282
30283 __extension__ extern __inline float16x8_t
30284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30285 vcvtq_f16_s16 (int16x8_t __a)
30286 {
30287   return __builtin_aarch64_floatv8hiv8hf (__a);
30288 }
30289
30290 __extension__ extern __inline float16x4_t
30291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30292 vcvt_f16_u16 (uint16x4_t __a)
30293 {
30294   return __builtin_aarch64_floatunsv4hiv4hf ((int16x4_t) __a);
30295 }
30296
30297 __extension__ extern __inline float16x8_t
30298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30299 vcvtq_f16_u16 (uint16x8_t __a)
30300 {
30301   return __builtin_aarch64_floatunsv8hiv8hf ((int16x8_t) __a);
30302 }
30303
30304 __extension__ extern __inline int16x4_t
30305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30306 vcvt_s16_f16 (float16x4_t __a)
30307 {
30308   return __builtin_aarch64_lbtruncv4hfv4hi (__a);
30309 }
30310
30311 __extension__ extern __inline int16x8_t
30312 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30313 vcvtq_s16_f16 (float16x8_t __a)
30314 {
30315   return __builtin_aarch64_lbtruncv8hfv8hi (__a);
30316 }
30317
30318 __extension__ extern __inline uint16x4_t
30319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30320 vcvt_u16_f16 (float16x4_t __a)
30321 {
30322   return __builtin_aarch64_lbtruncuv4hfv4hi_us (__a);
30323 }
30324
30325 __extension__ extern __inline uint16x8_t
30326 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30327 vcvtq_u16_f16 (float16x8_t __a)
30328 {
30329   return __builtin_aarch64_lbtruncuv8hfv8hi_us (__a);
30330 }
30331
30332 __extension__ extern __inline int16x4_t
30333 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30334 vcvta_s16_f16 (float16x4_t __a)
30335 {
30336   return __builtin_aarch64_lroundv4hfv4hi (__a);
30337 }
30338
30339 __extension__ extern __inline int16x8_t
30340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30341 vcvtaq_s16_f16 (float16x8_t __a)
30342 {
30343   return __builtin_aarch64_lroundv8hfv8hi (__a);
30344 }
30345
30346 __extension__ extern __inline uint16x4_t
30347 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30348 vcvta_u16_f16 (float16x4_t __a)
30349 {
30350   return __builtin_aarch64_lrounduv4hfv4hi_us (__a);
30351 }
30352
30353 __extension__ extern __inline uint16x8_t
30354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30355 vcvtaq_u16_f16 (float16x8_t __a)
30356 {
30357   return __builtin_aarch64_lrounduv8hfv8hi_us (__a);
30358 }
30359
30360 __extension__ extern __inline int16x4_t
30361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30362 vcvtm_s16_f16 (float16x4_t __a)
30363 {
30364   return __builtin_aarch64_lfloorv4hfv4hi (__a);
30365 }
30366
30367 __extension__ extern __inline int16x8_t
30368 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30369 vcvtmq_s16_f16 (float16x8_t __a)
30370 {
30371   return __builtin_aarch64_lfloorv8hfv8hi (__a);
30372 }
30373
30374 __extension__ extern __inline uint16x4_t
30375 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30376 vcvtm_u16_f16 (float16x4_t __a)
30377 {
30378   return __builtin_aarch64_lflooruv4hfv4hi_us (__a);
30379 }
30380
30381 __extension__ extern __inline uint16x8_t
30382 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30383 vcvtmq_u16_f16 (float16x8_t __a)
30384 {
30385   return __builtin_aarch64_lflooruv8hfv8hi_us (__a);
30386 }
30387
30388 __extension__ extern __inline int16x4_t
30389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30390 vcvtn_s16_f16 (float16x4_t __a)
30391 {
30392   return __builtin_aarch64_lfrintnv4hfv4hi (__a);
30393 }
30394
30395 __extension__ extern __inline int16x8_t
30396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30397 vcvtnq_s16_f16 (float16x8_t __a)
30398 {
30399   return __builtin_aarch64_lfrintnv8hfv8hi (__a);
30400 }
30401
30402 __extension__ extern __inline uint16x4_t
30403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30404 vcvtn_u16_f16 (float16x4_t __a)
30405 {
30406   return __builtin_aarch64_lfrintnuv4hfv4hi_us (__a);
30407 }
30408
30409 __extension__ extern __inline uint16x8_t
30410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30411 vcvtnq_u16_f16 (float16x8_t __a)
30412 {
30413   return __builtin_aarch64_lfrintnuv8hfv8hi_us (__a);
30414 }
30415
30416 __extension__ extern __inline int16x4_t
30417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30418 vcvtp_s16_f16 (float16x4_t __a)
30419 {
30420   return __builtin_aarch64_lceilv4hfv4hi (__a);
30421 }
30422
30423 __extension__ extern __inline int16x8_t
30424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30425 vcvtpq_s16_f16 (float16x8_t __a)
30426 {
30427   return __builtin_aarch64_lceilv8hfv8hi (__a);
30428 }
30429
30430 __extension__ extern __inline uint16x4_t
30431 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30432 vcvtp_u16_f16 (float16x4_t __a)
30433 {
30434   return __builtin_aarch64_lceiluv4hfv4hi_us (__a);
30435 }
30436
30437 __extension__ extern __inline uint16x8_t
30438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30439 vcvtpq_u16_f16 (float16x8_t __a)
30440 {
30441   return __builtin_aarch64_lceiluv8hfv8hi_us (__a);
30442 }
30443
30444 __extension__ extern __inline float16x4_t
30445 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30446 vneg_f16 (float16x4_t __a)
30447 {
30448   return -__a;
30449 }
30450
30451 __extension__ extern __inline float16x8_t
30452 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30453 vnegq_f16 (float16x8_t __a)
30454 {
30455   return -__a;
30456 }
30457
30458 __extension__ extern __inline float16x4_t
30459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30460 vrecpe_f16 (float16x4_t __a)
30461 {
30462   return __builtin_aarch64_frecpev4hf (__a);
30463 }
30464
30465 __extension__ extern __inline float16x8_t
30466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30467 vrecpeq_f16 (float16x8_t __a)
30468 {
30469   return __builtin_aarch64_frecpev8hf (__a);
30470 }
30471
30472 __extension__ extern __inline float16x4_t
30473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30474 vrnd_f16 (float16x4_t __a)
30475 {
30476   return __builtin_aarch64_btruncv4hf (__a);
30477 }
30478
30479 __extension__ extern __inline float16x8_t
30480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30481 vrndq_f16 (float16x8_t __a)
30482 {
30483   return __builtin_aarch64_btruncv8hf (__a);
30484 }
30485
30486 __extension__ extern __inline float16x4_t
30487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30488 vrnda_f16 (float16x4_t __a)
30489 {
30490   return __builtin_aarch64_roundv4hf (__a);
30491 }
30492
30493 __extension__ extern __inline float16x8_t
30494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30495 vrndaq_f16 (float16x8_t __a)
30496 {
30497   return __builtin_aarch64_roundv8hf (__a);
30498 }
30499
30500 __extension__ extern __inline float16x4_t
30501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30502 vrndi_f16 (float16x4_t __a)
30503 {
30504   return __builtin_aarch64_nearbyintv4hf (__a);
30505 }
30506
30507 __extension__ extern __inline float16x8_t
30508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30509 vrndiq_f16 (float16x8_t __a)
30510 {
30511   return __builtin_aarch64_nearbyintv8hf (__a);
30512 }
30513
30514 __extension__ extern __inline float16x4_t
30515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30516 vrndm_f16 (float16x4_t __a)
30517 {
30518   return __builtin_aarch64_floorv4hf (__a);
30519 }
30520
30521 __extension__ extern __inline float16x8_t
30522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30523 vrndmq_f16 (float16x8_t __a)
30524 {
30525   return __builtin_aarch64_floorv8hf (__a);
30526 }
30527
30528 __extension__ extern __inline float16x4_t
30529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30530 vrndn_f16 (float16x4_t __a)
30531 {
30532   return __builtin_aarch64_frintnv4hf (__a);
30533 }
30534
30535 __extension__ extern __inline float16x8_t
30536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30537 vrndnq_f16 (float16x8_t __a)
30538 {
30539   return __builtin_aarch64_frintnv8hf (__a);
30540 }
30541
30542 __extension__ extern __inline float16x4_t
30543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30544 vrndp_f16 (float16x4_t __a)
30545 {
30546   return __builtin_aarch64_ceilv4hf (__a);
30547 }
30548
30549 __extension__ extern __inline float16x8_t
30550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30551 vrndpq_f16 (float16x8_t __a)
30552 {
30553   return __builtin_aarch64_ceilv8hf (__a);
30554 }
30555
30556 __extension__ extern __inline float16x4_t
30557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30558 vrndx_f16 (float16x4_t __a)
30559 {
30560   return __builtin_aarch64_rintv4hf (__a);
30561 }
30562
30563 __extension__ extern __inline float16x8_t
30564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30565 vrndxq_f16 (float16x8_t __a)
30566 {
30567   return __builtin_aarch64_rintv8hf (__a);
30568 }
30569
30570 __extension__ extern __inline float16x4_t
30571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30572 vrsqrte_f16 (float16x4_t a)
30573 {
30574   return __builtin_aarch64_rsqrtev4hf (a);
30575 }
30576
30577 __extension__ extern __inline float16x8_t
30578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30579 vrsqrteq_f16 (float16x8_t a)
30580 {
30581   return __builtin_aarch64_rsqrtev8hf (a);
30582 }
30583
30584 __extension__ extern __inline float16x4_t
30585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30586 vsqrt_f16 (float16x4_t a)
30587 {
30588   return __builtin_aarch64_sqrtv4hf (a);
30589 }
30590
30591 __extension__ extern __inline float16x8_t
30592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30593 vsqrtq_f16 (float16x8_t a)
30594 {
30595   return __builtin_aarch64_sqrtv8hf (a);
30596 }
30597
30598 /* ARMv8.2-A FP16 two operands vector intrinsics.  */
30599
30600 __extension__ extern __inline float16x4_t
30601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30602 vadd_f16 (float16x4_t __a, float16x4_t __b)
30603 {
30604   return __a + __b;
30605 }
30606
30607 __extension__ extern __inline float16x8_t
30608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30609 vaddq_f16 (float16x8_t __a, float16x8_t __b)
30610 {
30611   return __a + __b;
30612 }
30613
30614 __extension__ extern __inline float16x4_t
30615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30616 vabd_f16 (float16x4_t a, float16x4_t b)
30617 {
30618   return __builtin_aarch64_fabdv4hf (a, b);
30619 }
30620
30621 __extension__ extern __inline float16x8_t
30622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30623 vabdq_f16 (float16x8_t a, float16x8_t b)
30624 {
30625   return __builtin_aarch64_fabdv8hf (a, b);
30626 }
30627
30628 __extension__ extern __inline uint16x4_t
30629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30630 vcage_f16 (float16x4_t __a, float16x4_t __b)
30631 {
30632   return __builtin_aarch64_facgev4hf_uss (__a, __b);
30633 }
30634
30635 __extension__ extern __inline uint16x8_t
30636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30637 vcageq_f16 (float16x8_t __a, float16x8_t __b)
30638 {
30639   return __builtin_aarch64_facgev8hf_uss (__a, __b);
30640 }
30641
30642 __extension__ extern __inline uint16x4_t
30643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30644 vcagt_f16 (float16x4_t __a, float16x4_t __b)
30645 {
30646   return __builtin_aarch64_facgtv4hf_uss (__a, __b);
30647 }
30648
30649 __extension__ extern __inline uint16x8_t
30650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30651 vcagtq_f16 (float16x8_t __a, float16x8_t __b)
30652 {
30653   return __builtin_aarch64_facgtv8hf_uss (__a, __b);
30654 }
30655
30656 __extension__ extern __inline uint16x4_t
30657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30658 vcale_f16 (float16x4_t __a, float16x4_t __b)
30659 {
30660   return __builtin_aarch64_faclev4hf_uss (__a, __b);
30661 }
30662
30663 __extension__ extern __inline uint16x8_t
30664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30665 vcaleq_f16 (float16x8_t __a, float16x8_t __b)
30666 {
30667   return __builtin_aarch64_faclev8hf_uss (__a, __b);
30668 }
30669
30670 __extension__ extern __inline uint16x4_t
30671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30672 vcalt_f16 (float16x4_t __a, float16x4_t __b)
30673 {
30674   return __builtin_aarch64_facltv4hf_uss (__a, __b);
30675 }
30676
30677 __extension__ extern __inline uint16x8_t
30678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30679 vcaltq_f16 (float16x8_t __a, float16x8_t __b)
30680 {
30681   return __builtin_aarch64_facltv8hf_uss (__a, __b);
30682 }
30683
30684 __extension__ extern __inline uint16x4_t
30685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30686 vceq_f16 (float16x4_t __a, float16x4_t __b)
30687 {
30688   return __builtin_aarch64_cmeqv4hf_uss (__a, __b);
30689 }
30690
30691 __extension__ extern __inline uint16x8_t
30692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30693 vceqq_f16 (float16x8_t __a, float16x8_t __b)
30694 {
30695   return __builtin_aarch64_cmeqv8hf_uss (__a, __b);
30696 }
30697
30698 __extension__ extern __inline uint16x4_t
30699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30700 vcge_f16 (float16x4_t __a, float16x4_t __b)
30701 {
30702   return __builtin_aarch64_cmgev4hf_uss (__a, __b);
30703 }
30704
30705 __extension__ extern __inline uint16x8_t
30706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30707 vcgeq_f16 (float16x8_t __a, float16x8_t __b)
30708 {
30709   return __builtin_aarch64_cmgev8hf_uss (__a, __b);
30710 }
30711
30712 __extension__ extern __inline uint16x4_t
30713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30714 vcgt_f16 (float16x4_t __a, float16x4_t __b)
30715 {
30716   return __builtin_aarch64_cmgtv4hf_uss (__a, __b);
30717 }
30718
30719 __extension__ extern __inline uint16x8_t
30720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30721 vcgtq_f16 (float16x8_t __a, float16x8_t __b)
30722 {
30723   return __builtin_aarch64_cmgtv8hf_uss (__a, __b);
30724 }
30725
30726 __extension__ extern __inline uint16x4_t
30727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30728 vcle_f16 (float16x4_t __a, float16x4_t __b)
30729 {
30730   return __builtin_aarch64_cmlev4hf_uss (__a, __b);
30731 }
30732
30733 __extension__ extern __inline uint16x8_t
30734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30735 vcleq_f16 (float16x8_t __a, float16x8_t __b)
30736 {
30737   return __builtin_aarch64_cmlev8hf_uss (__a, __b);
30738 }
30739
30740 __extension__ extern __inline uint16x4_t
30741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30742 vclt_f16 (float16x4_t __a, float16x4_t __b)
30743 {
30744   return __builtin_aarch64_cmltv4hf_uss (__a, __b);
30745 }
30746
30747 __extension__ extern __inline uint16x8_t
30748 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30749 vcltq_f16 (float16x8_t __a, float16x8_t __b)
30750 {
30751   return __builtin_aarch64_cmltv8hf_uss (__a, __b);
30752 }
30753
30754 __extension__ extern __inline float16x4_t
30755 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30756 vcvt_n_f16_s16 (int16x4_t __a, const int __b)
30757 {
30758   return __builtin_aarch64_scvtfv4hi (__a, __b);
30759 }
30760
30761 __extension__ extern __inline float16x8_t
30762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30763 vcvtq_n_f16_s16 (int16x8_t __a, const int __b)
30764 {
30765   return __builtin_aarch64_scvtfv8hi (__a, __b);
30766 }
30767
30768 __extension__ extern __inline float16x4_t
30769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30770 vcvt_n_f16_u16 (uint16x4_t __a, const int __b)
30771 {
30772   return __builtin_aarch64_ucvtfv4hi_sus (__a, __b);
30773 }
30774
30775 __extension__ extern __inline float16x8_t
30776 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30777 vcvtq_n_f16_u16 (uint16x8_t __a, const int __b)
30778 {
30779   return __builtin_aarch64_ucvtfv8hi_sus (__a, __b);
30780 }
30781
30782 __extension__ extern __inline int16x4_t
30783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30784 vcvt_n_s16_f16 (float16x4_t __a, const int __b)
30785 {
30786   return __builtin_aarch64_fcvtzsv4hf (__a, __b);
30787 }
30788
30789 __extension__ extern __inline int16x8_t
30790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30791 vcvtq_n_s16_f16 (float16x8_t __a, const int __b)
30792 {
30793   return __builtin_aarch64_fcvtzsv8hf (__a, __b);
30794 }
30795
30796 __extension__ extern __inline uint16x4_t
30797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30798 vcvt_n_u16_f16 (float16x4_t __a, const int __b)
30799 {
30800   return __builtin_aarch64_fcvtzuv4hf_uss (__a, __b);
30801 }
30802
30803 __extension__ extern __inline uint16x8_t
30804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30805 vcvtq_n_u16_f16 (float16x8_t __a, const int __b)
30806 {
30807   return __builtin_aarch64_fcvtzuv8hf_uss (__a, __b);
30808 }
30809
30810 __extension__ extern __inline float16x4_t
30811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30812 vdiv_f16 (float16x4_t __a, float16x4_t __b)
30813 {
30814   return __a / __b;
30815 }
30816
30817 __extension__ extern __inline float16x8_t
30818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30819 vdivq_f16 (float16x8_t __a, float16x8_t __b)
30820 {
30821   return __a / __b;
30822 }
30823
30824 __extension__ extern __inline float16x4_t
30825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30826 vmax_f16 (float16x4_t __a, float16x4_t __b)
30827 {
30828   return __builtin_aarch64_smax_nanv4hf (__a, __b);
30829 }
30830
30831 __extension__ extern __inline float16x8_t
30832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30833 vmaxq_f16 (float16x8_t __a, float16x8_t __b)
30834 {
30835   return __builtin_aarch64_smax_nanv8hf (__a, __b);
30836 }
30837
30838 __extension__ extern __inline float16x4_t
30839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30840 vmaxnm_f16 (float16x4_t __a, float16x4_t __b)
30841 {
30842   return __builtin_aarch64_fmaxv4hf (__a, __b);
30843 }
30844
30845 __extension__ extern __inline float16x8_t
30846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30847 vmaxnmq_f16 (float16x8_t __a, float16x8_t __b)
30848 {
30849   return __builtin_aarch64_fmaxv8hf (__a, __b);
30850 }
30851
30852 __extension__ extern __inline float16x4_t
30853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30854 vmin_f16 (float16x4_t __a, float16x4_t __b)
30855 {
30856   return __builtin_aarch64_smin_nanv4hf (__a, __b);
30857 }
30858
30859 __extension__ extern __inline float16x8_t
30860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30861 vminq_f16 (float16x8_t __a, float16x8_t __b)
30862 {
30863   return __builtin_aarch64_smin_nanv8hf (__a, __b);
30864 }
30865
30866 __extension__ extern __inline float16x4_t
30867 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30868 vminnm_f16 (float16x4_t __a, float16x4_t __b)
30869 {
30870   return __builtin_aarch64_fminv4hf (__a, __b);
30871 }
30872
30873 __extension__ extern __inline float16x8_t
30874 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30875 vminnmq_f16 (float16x8_t __a, float16x8_t __b)
30876 {
30877   return __builtin_aarch64_fminv8hf (__a, __b);
30878 }
30879
30880 __extension__ extern __inline float16x4_t
30881 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30882 vmul_f16 (float16x4_t __a, float16x4_t __b)
30883 {
30884   return __a * __b;
30885 }
30886
30887 __extension__ extern __inline float16x8_t
30888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30889 vmulq_f16 (float16x8_t __a, float16x8_t __b)
30890 {
30891   return __a * __b;
30892 }
30893
30894 __extension__ extern __inline float16x4_t
30895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30896 vmulx_f16 (float16x4_t __a, float16x4_t __b)
30897 {
30898   return __builtin_aarch64_fmulxv4hf (__a, __b);
30899 }
30900
30901 __extension__ extern __inline float16x8_t
30902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30903 vmulxq_f16 (float16x8_t __a, float16x8_t __b)
30904 {
30905   return __builtin_aarch64_fmulxv8hf (__a, __b);
30906 }
30907
30908 __extension__ extern __inline float16x4_t
30909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30910 vpadd_f16 (float16x4_t a, float16x4_t b)
30911 {
30912   return __builtin_aarch64_faddpv4hf (a, b);
30913 }
30914
30915 __extension__ extern __inline float16x8_t
30916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30917 vpaddq_f16 (float16x8_t a, float16x8_t b)
30918 {
30919   return __builtin_aarch64_faddpv8hf (a, b);
30920 }
30921
30922 __extension__ extern __inline float16x4_t
30923 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30924 vpmax_f16 (float16x4_t a, float16x4_t b)
30925 {
30926   return __builtin_aarch64_smax_nanpv4hf (a, b);
30927 }
30928
30929 __extension__ extern __inline float16x8_t
30930 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30931 vpmaxq_f16 (float16x8_t a, float16x8_t b)
30932 {
30933   return __builtin_aarch64_smax_nanpv8hf (a, b);
30934 }
30935
30936 __extension__ extern __inline float16x4_t
30937 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30938 vpmaxnm_f16 (float16x4_t a, float16x4_t b)
30939 {
30940   return __builtin_aarch64_smaxpv4hf (a, b);
30941 }
30942
30943 __extension__ extern __inline float16x8_t
30944 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30945 vpmaxnmq_f16 (float16x8_t a, float16x8_t b)
30946 {
30947   return __builtin_aarch64_smaxpv8hf (a, b);
30948 }
30949
30950 __extension__ extern __inline float16x4_t
30951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30952 vpmin_f16 (float16x4_t a, float16x4_t b)
30953 {
30954   return __builtin_aarch64_smin_nanpv4hf (a, b);
30955 }
30956
30957 __extension__ extern __inline float16x8_t
30958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30959 vpminq_f16 (float16x8_t a, float16x8_t b)
30960 {
30961   return __builtin_aarch64_smin_nanpv8hf (a, b);
30962 }
30963
30964 __extension__ extern __inline float16x4_t
30965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30966 vpminnm_f16 (float16x4_t a, float16x4_t b)
30967 {
30968   return __builtin_aarch64_sminpv4hf (a, b);
30969 }
30970
30971 __extension__ extern __inline float16x8_t
30972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30973 vpminnmq_f16 (float16x8_t a, float16x8_t b)
30974 {
30975   return __builtin_aarch64_sminpv8hf (a, b);
30976 }
30977
30978 __extension__ extern __inline float16x4_t
30979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30980 vrecps_f16 (float16x4_t __a, float16x4_t __b)
30981 {
30982   return __builtin_aarch64_frecpsv4hf (__a, __b);
30983 }
30984
30985 __extension__ extern __inline float16x8_t
30986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30987 vrecpsq_f16 (float16x8_t __a, float16x8_t __b)
30988 {
30989   return __builtin_aarch64_frecpsv8hf (__a, __b);
30990 }
30991
30992 __extension__ extern __inline float16x4_t
30993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30994 vrsqrts_f16 (float16x4_t a, float16x4_t b)
30995 {
30996   return __builtin_aarch64_rsqrtsv4hf (a, b);
30997 }
30998
30999 __extension__ extern __inline float16x8_t
31000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31001 vrsqrtsq_f16 (float16x8_t a, float16x8_t b)
31002 {
31003   return __builtin_aarch64_rsqrtsv8hf (a, b);
31004 }
31005
31006 __extension__ extern __inline float16x4_t
31007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31008 vsub_f16 (float16x4_t __a, float16x4_t __b)
31009 {
31010   return __a - __b;
31011 }
31012
31013 __extension__ extern __inline float16x8_t
31014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31015 vsubq_f16 (float16x8_t __a, float16x8_t __b)
31016 {
31017   return __a - __b;
31018 }
31019
31020 /* ARMv8.2-A FP16 three operands vector intrinsics.  */
31021
31022 __extension__ extern __inline float16x4_t
31023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31024 vfma_f16 (float16x4_t __a, float16x4_t __b, float16x4_t __c)
31025 {
31026   return __builtin_aarch64_fmav4hf (__b, __c, __a);
31027 }
31028
31029 __extension__ extern __inline float16x8_t
31030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31031 vfmaq_f16 (float16x8_t __a, float16x8_t __b, float16x8_t __c)
31032 {
31033   return __builtin_aarch64_fmav8hf (__b, __c, __a);
31034 }
31035
31036 __extension__ extern __inline float16x4_t
31037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31038 vfms_f16 (float16x4_t __a, float16x4_t __b, float16x4_t __c)
31039 {
31040   return __builtin_aarch64_fnmav4hf (__b, __c, __a);
31041 }
31042
31043 __extension__ extern __inline float16x8_t
31044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31045 vfmsq_f16 (float16x8_t __a, float16x8_t __b, float16x8_t __c)
31046 {
31047   return __builtin_aarch64_fnmav8hf (__b, __c, __a);
31048 }
31049
31050 /* ARMv8.2-A FP16 lane vector intrinsics.  */
31051
31052 __extension__ extern __inline float16_t
31053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31054 vfmah_lane_f16 (float16_t __a, float16_t __b,
31055                 float16x4_t __c, const int __lane)
31056 {
31057   return vfmah_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
31058 }
31059
31060 __extension__ extern __inline float16_t
31061 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31062 vfmah_laneq_f16 (float16_t __a, float16_t __b,
31063                  float16x8_t __c, const int __lane)
31064 {
31065   return vfmah_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
31066 }
31067
31068 __extension__ extern __inline float16x4_t
31069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31070 vfma_lane_f16 (float16x4_t __a, float16x4_t __b,
31071                float16x4_t __c, const int __lane)
31072 {
31073   return vfma_f16 (__a, __b, __aarch64_vdup_lane_f16 (__c, __lane));
31074 }
31075
31076 __extension__ extern __inline float16x8_t
31077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31078 vfmaq_lane_f16 (float16x8_t __a, float16x8_t __b,
31079                 float16x4_t __c, const int __lane)
31080 {
31081   return vfmaq_f16 (__a, __b, __aarch64_vdupq_lane_f16 (__c, __lane));
31082 }
31083
31084 __extension__ extern __inline float16x4_t
31085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31086 vfma_laneq_f16 (float16x4_t __a, float16x4_t __b,
31087                 float16x8_t __c, const int __lane)
31088 {
31089   return vfma_f16 (__a, __b, __aarch64_vdup_laneq_f16 (__c, __lane));
31090 }
31091
31092 __extension__ extern __inline float16x8_t
31093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31094 vfmaq_laneq_f16 (float16x8_t __a, float16x8_t __b,
31095                  float16x8_t __c, const int __lane)
31096 {
31097   return vfmaq_f16 (__a, __b, __aarch64_vdupq_laneq_f16 (__c, __lane));
31098 }
31099
31100 __extension__ extern __inline float16x4_t
31101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31102 vfma_n_f16 (float16x4_t __a, float16x4_t __b, float16_t __c)
31103 {
31104   return vfma_f16 (__a, __b, vdup_n_f16 (__c));
31105 }
31106
31107 __extension__ extern __inline float16x8_t
31108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31109 vfmaq_n_f16 (float16x8_t __a, float16x8_t __b, float16_t __c)
31110 {
31111   return vfmaq_f16 (__a, __b, vdupq_n_f16 (__c));
31112 }
31113
31114 __extension__ extern __inline float16_t
31115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31116 vfmsh_lane_f16 (float16_t __a, float16_t __b,
31117                 float16x4_t __c, const int __lane)
31118 {
31119   return vfmsh_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
31120 }
31121
31122 __extension__ extern __inline float16_t
31123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31124 vfmsh_laneq_f16 (float16_t __a, float16_t __b,
31125                  float16x8_t __c, const int __lane)
31126 {
31127   return vfmsh_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
31128 }
31129
31130 __extension__ extern __inline float16x4_t
31131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31132 vfms_lane_f16 (float16x4_t __a, float16x4_t __b,
31133                float16x4_t __c, const int __lane)
31134 {
31135   return vfms_f16 (__a, __b, __aarch64_vdup_lane_f16 (__c, __lane));
31136 }
31137
31138 __extension__ extern __inline float16x8_t
31139 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31140 vfmsq_lane_f16 (float16x8_t __a, float16x8_t __b,
31141                 float16x4_t __c, const int __lane)
31142 {
31143   return vfmsq_f16 (__a, __b, __aarch64_vdupq_lane_f16 (__c, __lane));
31144 }
31145
31146 __extension__ extern __inline float16x4_t
31147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31148 vfms_laneq_f16 (float16x4_t __a, float16x4_t __b,
31149                 float16x8_t __c, const int __lane)
31150 {
31151   return vfms_f16 (__a, __b, __aarch64_vdup_laneq_f16 (__c, __lane));
31152 }
31153
31154 __extension__ extern __inline float16x8_t
31155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31156 vfmsq_laneq_f16 (float16x8_t __a, float16x8_t __b,
31157                  float16x8_t __c, const int __lane)
31158 {
31159   return vfmsq_f16 (__a, __b, __aarch64_vdupq_laneq_f16 (__c, __lane));
31160 }
31161
31162 __extension__ extern __inline float16x4_t
31163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31164 vfms_n_f16 (float16x4_t __a, float16x4_t __b, float16_t __c)
31165 {
31166   return vfms_f16 (__a, __b, vdup_n_f16 (__c));
31167 }
31168
31169 __extension__ extern __inline float16x8_t
31170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31171 vfmsq_n_f16 (float16x8_t __a, float16x8_t __b, float16_t __c)
31172 {
31173   return vfmsq_f16 (__a, __b, vdupq_n_f16 (__c));
31174 }
31175
31176 __extension__ extern __inline float16_t
31177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31178 vmulh_lane_f16 (float16_t __a, float16x4_t __b, const int __lane)
31179 {
31180   return __a * __aarch64_vget_lane_any (__b, __lane);
31181 }
31182
31183 __extension__ extern __inline float16x4_t
31184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31185 vmul_lane_f16 (float16x4_t __a, float16x4_t __b, const int __lane)
31186 {
31187   return vmul_f16 (__a, vdup_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31188 }
31189
31190 __extension__ extern __inline float16x8_t
31191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31192 vmulq_lane_f16 (float16x8_t __a, float16x4_t __b, const int __lane)
31193 {
31194   return vmulq_f16 (__a, vdupq_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31195 }
31196
31197 __extension__ extern __inline float16_t
31198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31199 vmulh_laneq_f16 (float16_t __a, float16x8_t __b, const int __lane)
31200 {
31201   return __a * __aarch64_vget_lane_any (__b, __lane);
31202 }
31203
31204 __extension__ extern __inline float16x4_t
31205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31206 vmul_laneq_f16 (float16x4_t __a, float16x8_t __b, const int __lane)
31207 {
31208   return vmul_f16 (__a, vdup_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31209 }
31210
31211 __extension__ extern __inline float16x8_t
31212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31213 vmulq_laneq_f16 (float16x8_t __a, float16x8_t __b, const int __lane)
31214 {
31215   return vmulq_f16 (__a, vdupq_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31216 }
31217
31218 __extension__ extern __inline float16x4_t
31219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31220 vmul_n_f16 (float16x4_t __a, float16_t __b)
31221 {
31222   return vmul_lane_f16 (__a, vdup_n_f16 (__b), 0);
31223 }
31224
31225 __extension__ extern __inline float16x8_t
31226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31227 vmulq_n_f16 (float16x8_t __a, float16_t __b)
31228 {
31229   return vmulq_laneq_f16 (__a, vdupq_n_f16 (__b), 0);
31230 }
31231
31232 __extension__ extern __inline float16_t
31233 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31234 vmulxh_lane_f16 (float16_t __a, float16x4_t __b, const int __lane)
31235 {
31236   return vmulxh_f16 (__a, __aarch64_vget_lane_any (__b, __lane));
31237 }
31238
31239 __extension__ extern __inline float16x4_t
31240 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31241 vmulx_lane_f16 (float16x4_t __a, float16x4_t __b, const int __lane)
31242 {
31243   return vmulx_f16 (__a, __aarch64_vdup_lane_f16 (__b, __lane));
31244 }
31245
31246 __extension__ extern __inline float16x8_t
31247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31248 vmulxq_lane_f16 (float16x8_t __a, float16x4_t __b, const int __lane)
31249 {
31250   return vmulxq_f16 (__a, __aarch64_vdupq_lane_f16 (__b, __lane));
31251 }
31252
31253 __extension__ extern __inline float16_t
31254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31255 vmulxh_laneq_f16 (float16_t __a, float16x8_t __b, const int __lane)
31256 {
31257   return vmulxh_f16 (__a, __aarch64_vget_lane_any (__b, __lane));
31258 }
31259
31260 __extension__ extern __inline float16x4_t
31261 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31262 vmulx_laneq_f16 (float16x4_t __a, float16x8_t __b, const int __lane)
31263 {
31264   return vmulx_f16 (__a, __aarch64_vdup_laneq_f16 (__b, __lane));
31265 }
31266
31267 __extension__ extern __inline float16x8_t
31268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31269 vmulxq_laneq_f16 (float16x8_t __a, float16x8_t __b, const int __lane)
31270 {
31271   return vmulxq_f16 (__a, __aarch64_vdupq_laneq_f16 (__b, __lane));
31272 }
31273
31274 __extension__ extern __inline float16x4_t
31275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31276 vmulx_n_f16 (float16x4_t __a, float16_t __b)
31277 {
31278   return vmulx_f16 (__a, vdup_n_f16 (__b));
31279 }
31280
31281 __extension__ extern __inline float16x8_t
31282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31283 vmulxq_n_f16 (float16x8_t __a, float16_t __b)
31284 {
31285   return vmulxq_f16 (__a, vdupq_n_f16 (__b));
31286 }
31287
31288 /* ARMv8.2-A FP16 reduction vector intrinsics.  */
31289
31290 __extension__ extern __inline float16_t
31291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31292 vmaxv_f16 (float16x4_t __a)
31293 {
31294   return __builtin_aarch64_reduc_smax_nan_scal_v4hf (__a);
31295 }
31296
31297 __extension__ extern __inline float16_t
31298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31299 vmaxvq_f16 (float16x8_t __a)
31300 {
31301   return __builtin_aarch64_reduc_smax_nan_scal_v8hf (__a);
31302 }
31303
31304 __extension__ extern __inline float16_t
31305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31306 vminv_f16 (float16x4_t __a)
31307 {
31308   return __builtin_aarch64_reduc_smin_nan_scal_v4hf (__a);
31309 }
31310
31311 __extension__ extern __inline float16_t
31312 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31313 vminvq_f16 (float16x8_t __a)
31314 {
31315   return __builtin_aarch64_reduc_smin_nan_scal_v8hf (__a);
31316 }
31317
31318 __extension__ extern __inline float16_t
31319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31320 vmaxnmv_f16 (float16x4_t __a)
31321 {
31322   return __builtin_aarch64_reduc_smax_scal_v4hf (__a);
31323 }
31324
31325 __extension__ extern __inline float16_t
31326 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31327 vmaxnmvq_f16 (float16x8_t __a)
31328 {
31329   return __builtin_aarch64_reduc_smax_scal_v8hf (__a);
31330 }
31331
31332 __extension__ extern __inline float16_t
31333 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31334 vminnmv_f16 (float16x4_t __a)
31335 {
31336   return __builtin_aarch64_reduc_smin_scal_v4hf (__a);
31337 }
31338
31339 __extension__ extern __inline float16_t
31340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31341 vminnmvq_f16 (float16x8_t __a)
31342 {
31343   return __builtin_aarch64_reduc_smin_scal_v8hf (__a);
31344 }
31345
31346 #pragma GCC pop_options
31347
31348 #undef __aarch64_vget_lane_any
31349
31350 #undef __aarch64_vdup_lane_any
31351 #undef __aarch64_vdup_lane_f16
31352 #undef __aarch64_vdup_lane_f32
31353 #undef __aarch64_vdup_lane_f64
31354 #undef __aarch64_vdup_lane_p8
31355 #undef __aarch64_vdup_lane_p16
31356 #undef __aarch64_vdup_lane_s8
31357 #undef __aarch64_vdup_lane_s16
31358 #undef __aarch64_vdup_lane_s32
31359 #undef __aarch64_vdup_lane_s64
31360 #undef __aarch64_vdup_lane_u8
31361 #undef __aarch64_vdup_lane_u16
31362 #undef __aarch64_vdup_lane_u32
31363 #undef __aarch64_vdup_lane_u64
31364 #undef __aarch64_vdup_laneq_f16
31365 #undef __aarch64_vdup_laneq_f32
31366 #undef __aarch64_vdup_laneq_f64
31367 #undef __aarch64_vdup_laneq_p8
31368 #undef __aarch64_vdup_laneq_p16
31369 #undef __aarch64_vdup_laneq_s8
31370 #undef __aarch64_vdup_laneq_s16
31371 #undef __aarch64_vdup_laneq_s32
31372 #undef __aarch64_vdup_laneq_s64
31373 #undef __aarch64_vdup_laneq_u8
31374 #undef __aarch64_vdup_laneq_u16
31375 #undef __aarch64_vdup_laneq_u32
31376 #undef __aarch64_vdup_laneq_u64
31377 #undef __aarch64_vdupq_lane_f16
31378 #undef __aarch64_vdupq_lane_f32
31379 #undef __aarch64_vdupq_lane_f64
31380 #undef __aarch64_vdupq_lane_p8
31381 #undef __aarch64_vdupq_lane_p16
31382 #undef __aarch64_vdupq_lane_s8
31383 #undef __aarch64_vdupq_lane_s16
31384 #undef __aarch64_vdupq_lane_s32
31385 #undef __aarch64_vdupq_lane_s64
31386 #undef __aarch64_vdupq_lane_u8
31387 #undef __aarch64_vdupq_lane_u16
31388 #undef __aarch64_vdupq_lane_u32
31389 #undef __aarch64_vdupq_lane_u64
31390 #undef __aarch64_vdupq_laneq_f16
31391 #undef __aarch64_vdupq_laneq_f32
31392 #undef __aarch64_vdupq_laneq_f64
31393 #undef __aarch64_vdupq_laneq_p8
31394 #undef __aarch64_vdupq_laneq_p16
31395 #undef __aarch64_vdupq_laneq_s8
31396 #undef __aarch64_vdupq_laneq_s16
31397 #undef __aarch64_vdupq_laneq_s32
31398 #undef __aarch64_vdupq_laneq_s64
31399 #undef __aarch64_vdupq_laneq_u8
31400 #undef __aarch64_vdupq_laneq_u16
31401 #undef __aarch64_vdupq_laneq_u32
31402 #undef __aarch64_vdupq_laneq_u64
31403
31404 #endif