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 vsriq_n_p8(a, b, c)                                             \
10613   __extension__                                                         \
10614     ({                                                                  \
10615        poly8x16_t b_ = (b);                                             \
10616        poly8x16_t a_ = (a);                                             \
10617        poly8x16_t result;                                               \
10618        __asm__ ("sri %0.16b,%2.16b,%3"                                  \
10619                 : "=w"(result)                                          \
10620                 : "0"(a_), "w"(b_), "i"(c)                              \
10621                 : /* No clobbers */);                                   \
10622        result;                                                          \
10623      })
10624
10625 #define vsriq_n_p16(a, b, c)                                            \
10626   __extension__                                                         \
10627     ({                                                                  \
10628        poly16x8_t b_ = (b);                                             \
10629        poly16x8_t a_ = (a);                                             \
10630        poly16x8_t result;                                               \
10631        __asm__ ("sri %0.8h,%2.8h,%3"                                    \
10632                 : "=w"(result)                                          \
10633                 : "0"(a_), "w"(b_), "i"(c)                              \
10634                 : /* No clobbers */);                                   \
10635        result;                                                          \
10636      })
10637
10638 __extension__ extern __inline uint8x8_t
10639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10640 vtst_p8 (poly8x8_t a, poly8x8_t b)
10641 {
10642   uint8x8_t result;
10643   __asm__ ("cmtst %0.8b, %1.8b, %2.8b"
10644            : "=w"(result)
10645            : "w"(a), "w"(b)
10646            : /* No clobbers */);
10647   return result;
10648 }
10649
10650 __extension__ extern __inline uint16x4_t
10651 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10652 vtst_p16 (poly16x4_t a, poly16x4_t b)
10653 {
10654   uint16x4_t result;
10655   __asm__ ("cmtst %0.4h, %1.4h, %2.4h"
10656            : "=w"(result)
10657            : "w"(a), "w"(b)
10658            : /* No clobbers */);
10659   return result;
10660 }
10661
10662 __extension__ extern __inline uint8x16_t
10663 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10664 vtstq_p8 (poly8x16_t a, poly8x16_t b)
10665 {
10666   uint8x16_t result;
10667   __asm__ ("cmtst %0.16b, %1.16b, %2.16b"
10668            : "=w"(result)
10669            : "w"(a), "w"(b)
10670            : /* No clobbers */);
10671   return result;
10672 }
10673
10674 __extension__ extern __inline uint16x8_t
10675 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
10676 vtstq_p16 (poly16x8_t a, poly16x8_t b)
10677 {
10678   uint16x8_t result;
10679   __asm__ ("cmtst %0.8h, %1.8h, %2.8h"
10680            : "=w"(result)
10681            : "w"(a), "w"(b)
10682            : /* No clobbers */);
10683   return result;
10684 }
10685
10686 /* End of temporary inline asm implementations.  */
10687
10688 /* Start of temporary inline asm for vldn, vstn and friends.  */
10689
10690 /* Create struct element types for duplicating loads.
10691
10692    Create 2 element structures of:
10693
10694    +------+----+----+----+----+
10695    |      | 8  | 16 | 32 | 64 |
10696    +------+----+----+----+----+
10697    |int   | Y  | Y  | N  | N  |
10698    +------+----+----+----+----+
10699    |uint  | Y  | Y  | N  | N  |
10700    +------+----+----+----+----+
10701    |float | -  | Y  | N  | N  |
10702    +------+----+----+----+----+
10703    |poly  | Y  | Y  | -  | -  |
10704    +------+----+----+----+----+
10705
10706    Create 3 element structures of:
10707
10708    +------+----+----+----+----+
10709    |      | 8  | 16 | 32 | 64 |
10710    +------+----+----+----+----+
10711    |int   | Y  | Y  | Y  | Y  |
10712    +------+----+----+----+----+
10713    |uint  | Y  | Y  | Y  | Y  |
10714    +------+----+----+----+----+
10715    |float | -  | Y  | Y  | Y  |
10716    +------+----+----+----+----+
10717    |poly  | Y  | Y  | -  | -  |
10718    +------+----+----+----+----+
10719
10720    Create 4 element structures of:
10721
10722    +------+----+----+----+----+
10723    |      | 8  | 16 | 32 | 64 |
10724    +------+----+----+----+----+
10725    |int   | Y  | N  | N  | Y  |
10726    +------+----+----+----+----+
10727    |uint  | Y  | N  | N  | Y  |
10728    +------+----+----+----+----+
10729    |float | -  | N  | N  | Y  |
10730    +------+----+----+----+----+
10731    |poly  | Y  | N  | -  | -  |
10732    +------+----+----+----+----+
10733
10734   This is required for casting memory reference.  */
10735 #define __STRUCTN(t, sz, nelem)                 \
10736   typedef struct t ## sz ## x ## nelem ## _t {  \
10737     t ## sz ## _t val[nelem];                   \
10738   }  t ## sz ## x ## nelem ## _t;
10739
10740 /* 2-element structs.  */
10741 __STRUCTN (int, 8, 2)
10742 __STRUCTN (int, 16, 2)
10743 __STRUCTN (uint, 8, 2)
10744 __STRUCTN (uint, 16, 2)
10745 __STRUCTN (float, 16, 2)
10746 __STRUCTN (poly, 8, 2)
10747 __STRUCTN (poly, 16, 2)
10748 /* 3-element structs.  */
10749 __STRUCTN (int, 8, 3)
10750 __STRUCTN (int, 16, 3)
10751 __STRUCTN (int, 32, 3)
10752 __STRUCTN (int, 64, 3)
10753 __STRUCTN (uint, 8, 3)
10754 __STRUCTN (uint, 16, 3)
10755 __STRUCTN (uint, 32, 3)
10756 __STRUCTN (uint, 64, 3)
10757 __STRUCTN (float, 16, 3)
10758 __STRUCTN (float, 32, 3)
10759 __STRUCTN (float, 64, 3)
10760 __STRUCTN (poly, 8, 3)
10761 __STRUCTN (poly, 16, 3)
10762 /* 4-element structs.  */
10763 __STRUCTN (int, 8, 4)
10764 __STRUCTN (int, 64, 4)
10765 __STRUCTN (uint, 8, 4)
10766 __STRUCTN (uint, 64, 4)
10767 __STRUCTN (poly, 8, 4)
10768 __STRUCTN (float, 64, 4)
10769 #undef __STRUCTN
10770
10771
10772 #define __ST2_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10773                         qmode, ptr_mode, funcsuffix, signedtype)             \
10774 __extension__ extern __inline void                                           \
10775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10776 vst2_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10777                           intype __b, const int __c)                         \
10778 {                                                                            \
10779   __builtin_aarch64_simd_oi __o;                                             \
10780   largetype __temp;                                                          \
10781   __temp.val[0]                                                              \
10782     = vcombine_##funcsuffix (__b.val[0],                                     \
10783                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10784   __temp.val[1]                                                              \
10785     = vcombine_##funcsuffix (__b.val[1],                                     \
10786                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10787   __o = __builtin_aarch64_set_qregoi##qmode (__o,                            \
10788                                              (signedtype) __temp.val[0], 0); \
10789   __o = __builtin_aarch64_set_qregoi##qmode (__o,                            \
10790                                              (signedtype) __temp.val[1], 1); \
10791   __builtin_aarch64_st2_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
10792                                      __ptr, __o, __c);                       \
10793 }
10794
10795 __ST2_LANE_FUNC (float16x4x2_t, float16x8x2_t, float16_t, v4hf, v8hf, hf, f16,
10796                  float16x8_t)
10797 __ST2_LANE_FUNC (float32x2x2_t, float32x4x2_t, float32_t, v2sf, v4sf, sf, f32,
10798                  float32x4_t)
10799 __ST2_LANE_FUNC (float64x1x2_t, float64x2x2_t, float64_t, df, v2df, df, f64,
10800                  float64x2_t)
10801 __ST2_LANE_FUNC (poly8x8x2_t, poly8x16x2_t, poly8_t, v8qi, v16qi, qi, p8,
10802                  int8x16_t)
10803 __ST2_LANE_FUNC (poly16x4x2_t, poly16x8x2_t, poly16_t, v4hi, v8hi, hi, p16,
10804                  int16x8_t)
10805 __ST2_LANE_FUNC (poly64x1x2_t, poly64x2x2_t, poly64_t, di, v2di_ssps, di, p64,
10806                  poly64x2_t)
10807 __ST2_LANE_FUNC (int8x8x2_t, int8x16x2_t, int8_t, v8qi, v16qi, qi, s8,
10808                  int8x16_t)
10809 __ST2_LANE_FUNC (int16x4x2_t, int16x8x2_t, int16_t, v4hi, v8hi, hi, s16,
10810                  int16x8_t)
10811 __ST2_LANE_FUNC (int32x2x2_t, int32x4x2_t, int32_t, v2si, v4si, si, s32,
10812                  int32x4_t)
10813 __ST2_LANE_FUNC (int64x1x2_t, int64x2x2_t, int64_t, di, v2di, di, s64,
10814                  int64x2_t)
10815 __ST2_LANE_FUNC (uint8x8x2_t, uint8x16x2_t, uint8_t, v8qi, v16qi, qi, u8,
10816                  int8x16_t)
10817 __ST2_LANE_FUNC (uint16x4x2_t, uint16x8x2_t, uint16_t, v4hi, v8hi, hi, u16,
10818                  int16x8_t)
10819 __ST2_LANE_FUNC (uint32x2x2_t, uint32x4x2_t, uint32_t, v2si, v4si, si, u32,
10820                  int32x4_t)
10821 __ST2_LANE_FUNC (uint64x1x2_t, uint64x2x2_t, uint64_t, di, v2di, di, u64,
10822                  int64x2_t)
10823
10824 #undef __ST2_LANE_FUNC
10825 #define __ST2_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
10826 __extension__ extern __inline void                                          \
10827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10828 vst2q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
10829                            intype __b, const int __c)                       \
10830 {                                                                           \
10831   union { intype __i;                                                       \
10832           __builtin_aarch64_simd_oi __o; } __temp = { __b };                \
10833   __builtin_aarch64_st2_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
10834                                     __ptr, __temp.__o, __c);                \
10835 }
10836
10837 __ST2_LANE_FUNC (float16x8x2_t, float16_t, v8hf, hf, f16)
10838 __ST2_LANE_FUNC (float32x4x2_t, float32_t, v4sf, sf, f32)
10839 __ST2_LANE_FUNC (float64x2x2_t, float64_t, v2df, df, f64)
10840 __ST2_LANE_FUNC (poly8x16x2_t, poly8_t, v16qi, qi, p8)
10841 __ST2_LANE_FUNC (poly16x8x2_t, poly16_t, v8hi, hi, p16)
10842 __ST2_LANE_FUNC (poly64x2x2_t, poly64_t, v2di, di, p64)
10843 __ST2_LANE_FUNC (int8x16x2_t, int8_t, v16qi, qi, s8)
10844 __ST2_LANE_FUNC (int16x8x2_t, int16_t, v8hi, hi, s16)
10845 __ST2_LANE_FUNC (int32x4x2_t, int32_t, v4si, si, s32)
10846 __ST2_LANE_FUNC (int64x2x2_t, int64_t, v2di, di, s64)
10847 __ST2_LANE_FUNC (uint8x16x2_t, uint8_t, v16qi, qi, u8)
10848 __ST2_LANE_FUNC (uint16x8x2_t, uint16_t, v8hi, hi, u16)
10849 __ST2_LANE_FUNC (uint32x4x2_t, uint32_t, v4si, si, u32)
10850 __ST2_LANE_FUNC (uint64x2x2_t, uint64_t, v2di, di, u64)
10851
10852 #define __ST3_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10853                         qmode, ptr_mode, funcsuffix, signedtype)             \
10854 __extension__ extern __inline void                                           \
10855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10856 vst3_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10857                           intype __b, const int __c)                         \
10858 {                                                                            \
10859   __builtin_aarch64_simd_ci __o;                                             \
10860   largetype __temp;                                                          \
10861   __temp.val[0]                                                              \
10862     = vcombine_##funcsuffix (__b.val[0],                                     \
10863                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10864   __temp.val[1]                                                              \
10865     = vcombine_##funcsuffix (__b.val[1],                                     \
10866                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10867   __temp.val[2]                                                              \
10868     = vcombine_##funcsuffix (__b.val[2],                                     \
10869                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10870   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
10871                                              (signedtype) __temp.val[0], 0); \
10872   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
10873                                              (signedtype) __temp.val[1], 1); \
10874   __o = __builtin_aarch64_set_qregci##qmode (__o,                            \
10875                                              (signedtype) __temp.val[2], 2); \
10876   __builtin_aarch64_st3_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
10877                                      __ptr, __o, __c);                       \
10878 }
10879
10880 __ST3_LANE_FUNC (float16x4x3_t, float16x8x3_t, float16_t, v4hf, v8hf, hf, f16,
10881                  float16x8_t)
10882 __ST3_LANE_FUNC (float32x2x3_t, float32x4x3_t, float32_t, v2sf, v4sf, sf, f32,
10883                  float32x4_t)
10884 __ST3_LANE_FUNC (float64x1x3_t, float64x2x3_t, float64_t, df, v2df, df, f64,
10885                  float64x2_t)
10886 __ST3_LANE_FUNC (poly8x8x3_t, poly8x16x3_t, poly8_t, v8qi, v16qi, qi, p8,
10887                  int8x16_t)
10888 __ST3_LANE_FUNC (poly16x4x3_t, poly16x8x3_t, poly16_t, v4hi, v8hi, hi, p16,
10889                  int16x8_t)
10890 __ST3_LANE_FUNC (poly64x1x3_t, poly64x2x3_t, poly64_t, di, v2di_ssps, di, p64,
10891                  poly64x2_t)
10892 __ST3_LANE_FUNC (int8x8x3_t, int8x16x3_t, int8_t, v8qi, v16qi, qi, s8,
10893                  int8x16_t)
10894 __ST3_LANE_FUNC (int16x4x3_t, int16x8x3_t, int16_t, v4hi, v8hi, hi, s16,
10895                  int16x8_t)
10896 __ST3_LANE_FUNC (int32x2x3_t, int32x4x3_t, int32_t, v2si, v4si, si, s32,
10897                  int32x4_t)
10898 __ST3_LANE_FUNC (int64x1x3_t, int64x2x3_t, int64_t, di, v2di, di, s64,
10899                  int64x2_t)
10900 __ST3_LANE_FUNC (uint8x8x3_t, uint8x16x3_t, uint8_t, v8qi, v16qi, qi, u8,
10901                  int8x16_t)
10902 __ST3_LANE_FUNC (uint16x4x3_t, uint16x8x3_t, uint16_t, v4hi, v8hi, hi, u16,
10903                  int16x8_t)
10904 __ST3_LANE_FUNC (uint32x2x3_t, uint32x4x3_t, uint32_t, v2si, v4si, si, u32,
10905                  int32x4_t)
10906 __ST3_LANE_FUNC (uint64x1x3_t, uint64x2x3_t, uint64_t, di, v2di, di, u64,
10907                  int64x2_t)
10908
10909 #undef __ST3_LANE_FUNC
10910 #define __ST3_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
10911 __extension__ extern __inline void                                          \
10912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10913 vst3q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
10914                            intype __b, const int __c)                       \
10915 {                                                                           \
10916   union { intype __i;                                                       \
10917           __builtin_aarch64_simd_ci __o; } __temp = { __b };                \
10918   __builtin_aarch64_st3_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
10919                                     __ptr, __temp.__o, __c);                \
10920 }
10921
10922 __ST3_LANE_FUNC (float16x8x3_t, float16_t, v8hf, hf, f16)
10923 __ST3_LANE_FUNC (float32x4x3_t, float32_t, v4sf, sf, f32)
10924 __ST3_LANE_FUNC (float64x2x3_t, float64_t, v2df, df, f64)
10925 __ST3_LANE_FUNC (poly8x16x3_t, poly8_t, v16qi, qi, p8)
10926 __ST3_LANE_FUNC (poly16x8x3_t, poly16_t, v8hi, hi, p16)
10927 __ST3_LANE_FUNC (poly64x2x3_t, poly64_t, v2di, di, p64)
10928 __ST3_LANE_FUNC (int8x16x3_t, int8_t, v16qi, qi, s8)
10929 __ST3_LANE_FUNC (int16x8x3_t, int16_t, v8hi, hi, s16)
10930 __ST3_LANE_FUNC (int32x4x3_t, int32_t, v4si, si, s32)
10931 __ST3_LANE_FUNC (int64x2x3_t, int64_t, v2di, di, s64)
10932 __ST3_LANE_FUNC (uint8x16x3_t, uint8_t, v16qi, qi, u8)
10933 __ST3_LANE_FUNC (uint16x8x3_t, uint16_t, v8hi, hi, u16)
10934 __ST3_LANE_FUNC (uint32x4x3_t, uint32_t, v4si, si, u32)
10935 __ST3_LANE_FUNC (uint64x2x3_t, uint64_t, v2di, di, u64)
10936
10937 #define __ST4_LANE_FUNC(intype, largetype, ptrtype, mode,                    \
10938                         qmode, ptr_mode, funcsuffix, signedtype)             \
10939 __extension__ extern __inline void                                           \
10940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
10941 vst4_lane_ ## funcsuffix (ptrtype *__ptr,                                    \
10942                           intype __b, const int __c)                         \
10943 {                                                                            \
10944   __builtin_aarch64_simd_xi __o;                                             \
10945   largetype __temp;                                                          \
10946   __temp.val[0]                                                              \
10947     = vcombine_##funcsuffix (__b.val[0],                                     \
10948                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10949   __temp.val[1]                                                              \
10950     = vcombine_##funcsuffix (__b.val[1],                                     \
10951                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10952   __temp.val[2]                                                              \
10953     = vcombine_##funcsuffix (__b.val[2],                                     \
10954                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10955   __temp.val[3]                                                              \
10956     = vcombine_##funcsuffix (__b.val[3],                                     \
10957                              vcreate_##funcsuffix (__AARCH64_UINT64_C (0))); \
10958   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10959                                              (signedtype) __temp.val[0], 0); \
10960   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10961                                              (signedtype) __temp.val[1], 1); \
10962   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10963                                              (signedtype) __temp.val[2], 2); \
10964   __o = __builtin_aarch64_set_qregxi##qmode (__o,                            \
10965                                              (signedtype) __temp.val[3], 3); \
10966   __builtin_aarch64_st4_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *)  \
10967                                      __ptr, __o, __c);                       \
10968 }
10969
10970 __ST4_LANE_FUNC (float16x4x4_t, float16x8x4_t, float16_t, v4hf, v8hf, hf, f16,
10971                  float16x8_t)
10972 __ST4_LANE_FUNC (float32x2x4_t, float32x4x4_t, float32_t, v2sf, v4sf, sf, f32,
10973                  float32x4_t)
10974 __ST4_LANE_FUNC (float64x1x4_t, float64x2x4_t, float64_t, df, v2df, df, f64,
10975                  float64x2_t)
10976 __ST4_LANE_FUNC (poly8x8x4_t, poly8x16x4_t, poly8_t, v8qi, v16qi, qi, p8,
10977                  int8x16_t)
10978 __ST4_LANE_FUNC (poly16x4x4_t, poly16x8x4_t, poly16_t, v4hi, v8hi, hi, p16,
10979                  int16x8_t)
10980 __ST4_LANE_FUNC (poly64x1x4_t, poly64x2x4_t, poly64_t, di, v2di_ssps, di, p64,
10981                  poly64x2_t)
10982 __ST4_LANE_FUNC (int8x8x4_t, int8x16x4_t, int8_t, v8qi, v16qi, qi, s8,
10983                  int8x16_t)
10984 __ST4_LANE_FUNC (int16x4x4_t, int16x8x4_t, int16_t, v4hi, v8hi, hi, s16,
10985                  int16x8_t)
10986 __ST4_LANE_FUNC (int32x2x4_t, int32x4x4_t, int32_t, v2si, v4si, si, s32,
10987                  int32x4_t)
10988 __ST4_LANE_FUNC (int64x1x4_t, int64x2x4_t, int64_t, di, v2di, di, s64,
10989                  int64x2_t)
10990 __ST4_LANE_FUNC (uint8x8x4_t, uint8x16x4_t, uint8_t, v8qi, v16qi, qi, u8,
10991                  int8x16_t)
10992 __ST4_LANE_FUNC (uint16x4x4_t, uint16x8x4_t, uint16_t, v4hi, v8hi, hi, u16,
10993                  int16x8_t)
10994 __ST4_LANE_FUNC (uint32x2x4_t, uint32x4x4_t, uint32_t, v2si, v4si, si, u32,
10995                  int32x4_t)
10996 __ST4_LANE_FUNC (uint64x1x4_t, uint64x2x4_t, uint64_t, di, v2di, di, u64,
10997                  int64x2_t)
10998
10999 #undef __ST4_LANE_FUNC
11000 #define __ST4_LANE_FUNC(intype, ptrtype, mode, ptr_mode, funcsuffix)        \
11001 __extension__ extern __inline void                                          \
11002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
11003 vst4q_lane_ ## funcsuffix (ptrtype *__ptr,                                  \
11004                            intype __b, const int __c)                       \
11005 {                                                                           \
11006   union { intype __i;                                                       \
11007           __builtin_aarch64_simd_xi __o; } __temp = { __b };                \
11008   __builtin_aarch64_st4_lane##mode ((__builtin_aarch64_simd_ ## ptr_mode *) \
11009                                     __ptr, __temp.__o, __c);                \
11010 }
11011
11012 __ST4_LANE_FUNC (float16x8x4_t, float16_t, v8hf, hf, f16)
11013 __ST4_LANE_FUNC (float32x4x4_t, float32_t, v4sf, sf, f32)
11014 __ST4_LANE_FUNC (float64x2x4_t, float64_t, v2df, df, f64)
11015 __ST4_LANE_FUNC (poly8x16x4_t, poly8_t, v16qi, qi, p8)
11016 __ST4_LANE_FUNC (poly16x8x4_t, poly16_t, v8hi, hi, p16)
11017 __ST4_LANE_FUNC (poly64x2x4_t, poly64_t, v2di, di, p64)
11018 __ST4_LANE_FUNC (int8x16x4_t, int8_t, v16qi, qi, s8)
11019 __ST4_LANE_FUNC (int16x8x4_t, int16_t, v8hi, hi, s16)
11020 __ST4_LANE_FUNC (int32x4x4_t, int32_t, v4si, si, s32)
11021 __ST4_LANE_FUNC (int64x2x4_t, int64_t, v2di, di, s64)
11022 __ST4_LANE_FUNC (uint8x16x4_t, uint8_t, v16qi, qi, u8)
11023 __ST4_LANE_FUNC (uint16x8x4_t, uint16_t, v8hi, hi, u16)
11024 __ST4_LANE_FUNC (uint32x4x4_t, uint32_t, v4si, si, u32)
11025 __ST4_LANE_FUNC (uint64x2x4_t, uint64_t, v2di, di, u64)
11026
11027 __extension__ extern __inline int64_t
11028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11029 vaddlv_s32 (int32x2_t a)
11030 {
11031   int64_t result;
11032   __asm__ ("saddlp %0.1d, %1.2s" : "=w"(result) : "w"(a) : );
11033   return result;
11034 }
11035
11036 __extension__ extern __inline uint64_t
11037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11038 vaddlv_u32 (uint32x2_t a)
11039 {
11040   uint64_t result;
11041   __asm__ ("uaddlp %0.1d, %1.2s" : "=w"(result) : "w"(a) : );
11042   return result;
11043 }
11044
11045 __extension__ extern __inline int16x4_t
11046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11047 vqdmulh_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __c)
11048 {
11049   return __builtin_aarch64_sqdmulh_laneqv4hi (__a, __b, __c);
11050 }
11051
11052 __extension__ extern __inline int32x2_t
11053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11054 vqdmulh_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __c)
11055 {
11056   return __builtin_aarch64_sqdmulh_laneqv2si (__a, __b, __c);
11057 }
11058
11059 __extension__ extern __inline int16x8_t
11060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11061 vqdmulhq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __c)
11062 {
11063   return __builtin_aarch64_sqdmulh_laneqv8hi (__a, __b, __c);
11064 }
11065
11066 __extension__ extern __inline int32x4_t
11067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11068 vqdmulhq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __c)
11069 {
11070   return __builtin_aarch64_sqdmulh_laneqv4si (__a, __b, __c);
11071 }
11072
11073 __extension__ extern __inline int16x4_t
11074 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11075 vqrdmulh_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __c)
11076 {
11077   return  __builtin_aarch64_sqrdmulh_laneqv4hi (__a, __b, __c);
11078 }
11079
11080 __extension__ extern __inline int32x2_t
11081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11082 vqrdmulh_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __c)
11083 {
11084   return __builtin_aarch64_sqrdmulh_laneqv2si (__a, __b, __c);
11085 }
11086
11087 __extension__ extern __inline int16x8_t
11088 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11089 vqrdmulhq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __c)
11090 {
11091   return __builtin_aarch64_sqrdmulh_laneqv8hi (__a, __b, __c);
11092 }
11093
11094 __extension__ extern __inline int32x4_t
11095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11096 vqrdmulhq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __c)
11097 {
11098   return __builtin_aarch64_sqrdmulh_laneqv4si (__a, __b, __c);
11099 }
11100
11101 /* Table intrinsics.  */
11102
11103 __extension__ extern __inline poly8x8_t
11104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11105 vqtbl1_p8 (poly8x16_t a, uint8x8_t b)
11106 {
11107   poly8x8_t result;
11108   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11109            : "=w"(result)
11110            : "w"(a), "w"(b)
11111            : /* No clobbers */);
11112   return result;
11113 }
11114
11115 __extension__ extern __inline int8x8_t
11116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11117 vqtbl1_s8 (int8x16_t a, uint8x8_t b)
11118 {
11119   int8x8_t result;
11120   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11121            : "=w"(result)
11122            : "w"(a), "w"(b)
11123            : /* No clobbers */);
11124   return result;
11125 }
11126
11127 __extension__ extern __inline uint8x8_t
11128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11129 vqtbl1_u8 (uint8x16_t a, uint8x8_t b)
11130 {
11131   uint8x8_t result;
11132   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11133            : "=w"(result)
11134            : "w"(a), "w"(b)
11135            : /* No clobbers */);
11136   return result;
11137 }
11138
11139 __extension__ extern __inline poly8x16_t
11140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11141 vqtbl1q_p8 (poly8x16_t a, uint8x16_t b)
11142 {
11143   poly8x16_t result;
11144   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11145            : "=w"(result)
11146            : "w"(a), "w"(b)
11147            : /* No clobbers */);
11148   return result;
11149 }
11150
11151 __extension__ extern __inline int8x16_t
11152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11153 vqtbl1q_s8 (int8x16_t a, uint8x16_t b)
11154 {
11155   int8x16_t result;
11156   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11157            : "=w"(result)
11158            : "w"(a), "w"(b)
11159            : /* No clobbers */);
11160   return result;
11161 }
11162
11163 __extension__ extern __inline uint8x16_t
11164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11165 vqtbl1q_u8 (uint8x16_t a, uint8x16_t b)
11166 {
11167   uint8x16_t result;
11168   __asm__ ("tbl %0.16b, {%1.16b}, %2.16b"
11169            : "=w"(result)
11170            : "w"(a), "w"(b)
11171            : /* No clobbers */);
11172   return result;
11173 }
11174
11175 __extension__ extern __inline int8x8_t
11176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11177 vqtbx1_s8 (int8x8_t r, int8x16_t tab, uint8x8_t idx)
11178 {
11179   int8x8_t result = r;
11180   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11181            : "+w"(result)
11182            : "w"(tab), "w"(idx)
11183            : /* No clobbers */);
11184   return result;
11185 }
11186
11187 __extension__ extern __inline uint8x8_t
11188 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11189 vqtbx1_u8 (uint8x8_t r, uint8x16_t tab, uint8x8_t idx)
11190 {
11191   uint8x8_t result = r;
11192   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11193            : "+w"(result)
11194            : "w"(tab), "w"(idx)
11195            : /* No clobbers */);
11196   return result;
11197 }
11198
11199 __extension__ extern __inline poly8x8_t
11200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11201 vqtbx1_p8 (poly8x8_t r, poly8x16_t tab, uint8x8_t idx)
11202 {
11203   poly8x8_t result = r;
11204   __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
11205            : "+w"(result)
11206            : "w"(tab), "w"(idx)
11207            : /* No clobbers */);
11208   return result;
11209 }
11210
11211 __extension__ extern __inline int8x16_t
11212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11213 vqtbx1q_s8 (int8x16_t r, int8x16_t tab, uint8x16_t idx)
11214 {
11215   int8x16_t result = r;
11216   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11217            : "+w"(result)
11218            : "w"(tab), "w"(idx)
11219            : /* No clobbers */);
11220   return result;
11221 }
11222
11223 __extension__ extern __inline uint8x16_t
11224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11225 vqtbx1q_u8 (uint8x16_t r, uint8x16_t tab, uint8x16_t idx)
11226 {
11227   uint8x16_t result = r;
11228   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11229            : "+w"(result)
11230            : "w"(tab), "w"(idx)
11231            : /* No clobbers */);
11232   return result;
11233 }
11234
11235 __extension__ extern __inline poly8x16_t
11236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11237 vqtbx1q_p8 (poly8x16_t r, poly8x16_t tab, uint8x16_t idx)
11238 {
11239   poly8x16_t result = r;
11240   __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
11241            : "+w"(result)
11242            : "w"(tab), "w"(idx)
11243            : /* No clobbers */);
11244   return result;
11245 }
11246
11247 /* V7 legacy table intrinsics.  */
11248
11249 __extension__ extern __inline int8x8_t
11250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11251 vtbl1_s8 (int8x8_t tab, int8x8_t idx)
11252 {
11253   int8x8_t result;
11254   int8x16_t temp = vcombine_s8 (tab, vcreate_s8 (__AARCH64_UINT64_C (0x0)));
11255   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11256            : "=w"(result)
11257            : "w"(temp), "w"(idx)
11258            : /* No clobbers */);
11259   return result;
11260 }
11261
11262 __extension__ extern __inline uint8x8_t
11263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11264 vtbl1_u8 (uint8x8_t tab, uint8x8_t idx)
11265 {
11266   uint8x8_t result;
11267   uint8x16_t temp = vcombine_u8 (tab, vcreate_u8 (__AARCH64_UINT64_C (0x0)));
11268   __asm__ ("tbl %0.8b, {%1.16b}, %2.8b"
11269            : "=w"(result)
11270            : "w"(temp), "w"(idx)
11271            : /* No clobbers */);
11272   return result;
11273 }
11274
11275 __extension__ extern __inline poly8x8_t
11276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11277 vtbl1_p8 (poly8x8_t tab, uint8x8_t idx)
11278 {
11279   poly8x8_t result;
11280   poly8x16_t temp = vcombine_p8 (tab, vcreate_p8 (__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 int8x8_t
11289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11290 vtbl2_s8 (int8x8x2_t tab, int8x8_t idx)
11291 {
11292   int8x8_t result;
11293   int8x16_t temp = vcombine_s8 (tab.val[0], tab.val[1]);
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 uint8x8_t
11302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11303 vtbl2_u8 (uint8x8x2_t tab, uint8x8_t idx)
11304 {
11305   uint8x8_t result;
11306   uint8x16_t temp = vcombine_u8 (tab.val[0], tab.val[1]);
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 poly8x8_t
11315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11316 vtbl2_p8 (poly8x8x2_t tab, uint8x8_t idx)
11317 {
11318   poly8x8_t result;
11319   poly8x16_t temp = vcombine_p8 (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 int8x8_t
11328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11329 vtbl3_s8 (int8x8x3_t tab, int8x8_t idx)
11330 {
11331   int8x8_t result;
11332   int8x16x2_t temp;
11333   __builtin_aarch64_simd_oi __o;
11334   temp.val[0] = vcombine_s8 (tab.val[0], tab.val[1]);
11335   temp.val[1] = vcombine_s8 (tab.val[2], vcreate_s8 (__AARCH64_UINT64_C (0x0)));
11336   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11337                                            (int8x16_t) temp.val[0], 0);
11338   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11339                                            (int8x16_t) temp.val[1], 1);
11340   result = __builtin_aarch64_tbl3v8qi (__o, idx);
11341   return result;
11342 }
11343
11344 __extension__ extern __inline uint8x8_t
11345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11346 vtbl3_u8 (uint8x8x3_t tab, uint8x8_t idx)
11347 {
11348   uint8x8_t result;
11349   uint8x16x2_t temp;
11350   __builtin_aarch64_simd_oi __o;
11351   temp.val[0] = vcombine_u8 (tab.val[0], tab.val[1]);
11352   temp.val[1] = vcombine_u8 (tab.val[2], vcreate_u8 (__AARCH64_UINT64_C (0x0)));
11353   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11354                                            (int8x16_t) temp.val[0], 0);
11355   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11356                                            (int8x16_t) temp.val[1], 1);
11357   result = (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11358   return result;
11359 }
11360
11361 __extension__ extern __inline poly8x8_t
11362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11363 vtbl3_p8 (poly8x8x3_t tab, uint8x8_t idx)
11364 {
11365   poly8x8_t result;
11366   poly8x16x2_t temp;
11367   __builtin_aarch64_simd_oi __o;
11368   temp.val[0] = vcombine_p8 (tab.val[0], tab.val[1]);
11369   temp.val[1] = vcombine_p8 (tab.val[2], vcreate_p8 (__AARCH64_UINT64_C (0x0)));
11370   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11371                                            (int8x16_t) temp.val[0], 0);
11372   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11373                                            (int8x16_t) temp.val[1], 1);
11374   result = (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11375   return result;
11376 }
11377
11378 __extension__ extern __inline int8x8_t
11379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11380 vtbl4_s8 (int8x8x4_t tab, int8x8_t idx)
11381 {
11382   int8x8_t result;
11383   int8x16x2_t temp;
11384   __builtin_aarch64_simd_oi __o;
11385   temp.val[0] = vcombine_s8 (tab.val[0], tab.val[1]);
11386   temp.val[1] = vcombine_s8 (tab.val[2], tab.val[3]);
11387   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11388                                            (int8x16_t) temp.val[0], 0);
11389   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11390                                            (int8x16_t) temp.val[1], 1);
11391   result = __builtin_aarch64_tbl3v8qi (__o, idx);
11392   return result;
11393 }
11394
11395 __extension__ extern __inline uint8x8_t
11396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11397 vtbl4_u8 (uint8x8x4_t tab, uint8x8_t idx)
11398 {
11399   uint8x8_t result;
11400   uint8x16x2_t temp;
11401   __builtin_aarch64_simd_oi __o;
11402   temp.val[0] = vcombine_u8 (tab.val[0], tab.val[1]);
11403   temp.val[1] = vcombine_u8 (tab.val[2], tab.val[3]);
11404   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11405                                            (int8x16_t) temp.val[0], 0);
11406   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11407                                            (int8x16_t) temp.val[1], 1);
11408   result = (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11409   return result;
11410 }
11411
11412 __extension__ extern __inline poly8x8_t
11413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11414 vtbl4_p8 (poly8x8x4_t tab, uint8x8_t idx)
11415 {
11416   poly8x8_t result;
11417   poly8x16x2_t temp;
11418   __builtin_aarch64_simd_oi __o;
11419   temp.val[0] = vcombine_p8 (tab.val[0], tab.val[1]);
11420   temp.val[1] = vcombine_p8 (tab.val[2], tab.val[3]);
11421   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11422                                            (int8x16_t) temp.val[0], 0);
11423   __o = __builtin_aarch64_set_qregoiv16qi (__o,
11424                                            (int8x16_t) temp.val[1], 1);
11425   result = (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
11426   return result;
11427 }
11428
11429 __extension__ extern __inline int8x8_t
11430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11431 vtbx2_s8 (int8x8_t r, int8x8x2_t tab, int8x8_t idx)
11432 {
11433   int8x8_t result = r;
11434   int8x16_t temp = vcombine_s8 (tab.val[0], tab.val[1]);
11435   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11436            : "+w"(result)
11437            : "w"(temp), "w"(idx)
11438            : /* No clobbers */);
11439   return result;
11440 }
11441
11442 __extension__ extern __inline uint8x8_t
11443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11444 vtbx2_u8 (uint8x8_t r, uint8x8x2_t tab, uint8x8_t idx)
11445 {
11446   uint8x8_t result = r;
11447   uint8x16_t temp = vcombine_u8 (tab.val[0], tab.val[1]);
11448   __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
11449            : "+w"(result)
11450            : "w"(temp), "w"(idx)
11451            : /* No clobbers */);
11452   return result;
11453 }
11454
11455 __extension__ extern __inline poly8x8_t
11456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11457 vtbx2_p8 (poly8x8_t r, poly8x8x2_t tab, uint8x8_t idx)
11458 {
11459   poly8x8_t result = r;
11460   poly8x16_t temp = vcombine_p8 (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 /* End of temporary inline asm.  */
11469
11470 /* Start of optimal implementations in approved order.  */
11471
11472 /* vabd.  */
11473
11474 __extension__ extern __inline float32_t
11475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11476 vabds_f32 (float32_t __a, float32_t __b)
11477 {
11478   return __builtin_aarch64_fabdsf (__a, __b);
11479 }
11480
11481 __extension__ extern __inline float64_t
11482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11483 vabdd_f64 (float64_t __a, float64_t __b)
11484 {
11485   return __builtin_aarch64_fabddf (__a, __b);
11486 }
11487
11488 __extension__ extern __inline float32x2_t
11489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11490 vabd_f32 (float32x2_t __a, float32x2_t __b)
11491 {
11492   return __builtin_aarch64_fabdv2sf (__a, __b);
11493 }
11494
11495 __extension__ extern __inline float64x1_t
11496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11497 vabd_f64 (float64x1_t __a, float64x1_t __b)
11498 {
11499   return (float64x1_t) {vabdd_f64 (vget_lane_f64 (__a, 0),
11500                                    vget_lane_f64 (__b, 0))};
11501 }
11502
11503 __extension__ extern __inline float32x4_t
11504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11505 vabdq_f32 (float32x4_t __a, float32x4_t __b)
11506 {
11507   return __builtin_aarch64_fabdv4sf (__a, __b);
11508 }
11509
11510 __extension__ extern __inline float64x2_t
11511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11512 vabdq_f64 (float64x2_t __a, float64x2_t __b)
11513 {
11514   return __builtin_aarch64_fabdv2df (__a, __b);
11515 }
11516
11517 /* vabs  */
11518
11519 __extension__ extern __inline float32x2_t
11520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11521 vabs_f32 (float32x2_t __a)
11522 {
11523   return __builtin_aarch64_absv2sf (__a);
11524 }
11525
11526 __extension__ extern __inline float64x1_t
11527 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11528 vabs_f64 (float64x1_t __a)
11529 {
11530   return (float64x1_t) {__builtin_fabs (__a[0])};
11531 }
11532
11533 __extension__ extern __inline int8x8_t
11534 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11535 vabs_s8 (int8x8_t __a)
11536 {
11537   return __builtin_aarch64_absv8qi (__a);
11538 }
11539
11540 __extension__ extern __inline int16x4_t
11541 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11542 vabs_s16 (int16x4_t __a)
11543 {
11544   return __builtin_aarch64_absv4hi (__a);
11545 }
11546
11547 __extension__ extern __inline int32x2_t
11548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11549 vabs_s32 (int32x2_t __a)
11550 {
11551   return __builtin_aarch64_absv2si (__a);
11552 }
11553
11554 __extension__ extern __inline int64x1_t
11555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11556 vabs_s64 (int64x1_t __a)
11557 {
11558   return (int64x1_t) {__builtin_aarch64_absdi (__a[0])};
11559 }
11560
11561 __extension__ extern __inline float32x4_t
11562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11563 vabsq_f32 (float32x4_t __a)
11564 {
11565   return __builtin_aarch64_absv4sf (__a);
11566 }
11567
11568 __extension__ extern __inline float64x2_t
11569 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11570 vabsq_f64 (float64x2_t __a)
11571 {
11572   return __builtin_aarch64_absv2df (__a);
11573 }
11574
11575 __extension__ extern __inline int8x16_t
11576 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11577 vabsq_s8 (int8x16_t __a)
11578 {
11579   return __builtin_aarch64_absv16qi (__a);
11580 }
11581
11582 __extension__ extern __inline int16x8_t
11583 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11584 vabsq_s16 (int16x8_t __a)
11585 {
11586   return __builtin_aarch64_absv8hi (__a);
11587 }
11588
11589 __extension__ extern __inline int32x4_t
11590 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11591 vabsq_s32 (int32x4_t __a)
11592 {
11593   return __builtin_aarch64_absv4si (__a);
11594 }
11595
11596 __extension__ extern __inline int64x2_t
11597 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11598 vabsq_s64 (int64x2_t __a)
11599 {
11600   return __builtin_aarch64_absv2di (__a);
11601 }
11602
11603 /* vadd */
11604
11605 __extension__ extern __inline int64_t
11606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11607 vaddd_s64 (int64_t __a, int64_t __b)
11608 {
11609   return __a + __b;
11610 }
11611
11612 __extension__ extern __inline uint64_t
11613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11614 vaddd_u64 (uint64_t __a, uint64_t __b)
11615 {
11616   return __a + __b;
11617 }
11618
11619 /* vaddv */
11620
11621 __extension__ extern __inline int8_t
11622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11623 vaddv_s8 (int8x8_t __a)
11624 {
11625   return __builtin_aarch64_reduc_plus_scal_v8qi (__a);
11626 }
11627
11628 __extension__ extern __inline int16_t
11629 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11630 vaddv_s16 (int16x4_t __a)
11631 {
11632   return __builtin_aarch64_reduc_plus_scal_v4hi (__a);
11633 }
11634
11635 __extension__ extern __inline int32_t
11636 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11637 vaddv_s32 (int32x2_t __a)
11638 {
11639   return __builtin_aarch64_reduc_plus_scal_v2si (__a);
11640 }
11641
11642 __extension__ extern __inline uint8_t
11643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11644 vaddv_u8 (uint8x8_t __a)
11645 {
11646   return (uint8_t) __builtin_aarch64_reduc_plus_scal_v8qi ((int8x8_t) __a);
11647 }
11648
11649 __extension__ extern __inline uint16_t
11650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11651 vaddv_u16 (uint16x4_t __a)
11652 {
11653   return (uint16_t) __builtin_aarch64_reduc_plus_scal_v4hi ((int16x4_t) __a);
11654 }
11655
11656 __extension__ extern __inline uint32_t
11657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11658 vaddv_u32 (uint32x2_t __a)
11659 {
11660   return (int32_t) __builtin_aarch64_reduc_plus_scal_v2si ((int32x2_t) __a);
11661 }
11662
11663 __extension__ extern __inline int8_t
11664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11665 vaddvq_s8 (int8x16_t __a)
11666 {
11667   return __builtin_aarch64_reduc_plus_scal_v16qi (__a);
11668 }
11669
11670 __extension__ extern __inline int16_t
11671 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11672 vaddvq_s16 (int16x8_t __a)
11673 {
11674   return __builtin_aarch64_reduc_plus_scal_v8hi (__a);
11675 }
11676
11677 __extension__ extern __inline int32_t
11678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11679 vaddvq_s32 (int32x4_t __a)
11680 {
11681   return __builtin_aarch64_reduc_plus_scal_v4si (__a);
11682 }
11683
11684 __extension__ extern __inline int64_t
11685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11686 vaddvq_s64 (int64x2_t __a)
11687 {
11688   return __builtin_aarch64_reduc_plus_scal_v2di (__a);
11689 }
11690
11691 __extension__ extern __inline uint8_t
11692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11693 vaddvq_u8 (uint8x16_t __a)
11694 {
11695   return (uint8_t) __builtin_aarch64_reduc_plus_scal_v16qi ((int8x16_t) __a);
11696 }
11697
11698 __extension__ extern __inline uint16_t
11699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11700 vaddvq_u16 (uint16x8_t __a)
11701 {
11702   return (uint16_t) __builtin_aarch64_reduc_plus_scal_v8hi ((int16x8_t) __a);
11703 }
11704
11705 __extension__ extern __inline uint32_t
11706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11707 vaddvq_u32 (uint32x4_t __a)
11708 {
11709   return (uint32_t) __builtin_aarch64_reduc_plus_scal_v4si ((int32x4_t) __a);
11710 }
11711
11712 __extension__ extern __inline uint64_t
11713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11714 vaddvq_u64 (uint64x2_t __a)
11715 {
11716   return (uint64_t) __builtin_aarch64_reduc_plus_scal_v2di ((int64x2_t) __a);
11717 }
11718
11719 __extension__ extern __inline float32_t
11720 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11721 vaddv_f32 (float32x2_t __a)
11722 {
11723   return __builtin_aarch64_reduc_plus_scal_v2sf (__a);
11724 }
11725
11726 __extension__ extern __inline float32_t
11727 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11728 vaddvq_f32 (float32x4_t __a)
11729 {
11730   return __builtin_aarch64_reduc_plus_scal_v4sf (__a);
11731 }
11732
11733 __extension__ extern __inline float64_t
11734 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11735 vaddvq_f64 (float64x2_t __a)
11736 {
11737   return __builtin_aarch64_reduc_plus_scal_v2df (__a);
11738 }
11739
11740 /* vbsl  */
11741
11742 __extension__ extern __inline float16x4_t
11743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11744 vbsl_f16 (uint16x4_t __a, float16x4_t __b, float16x4_t __c)
11745 {
11746   return __builtin_aarch64_simd_bslv4hf_suss (__a, __b, __c);
11747 }
11748
11749 __extension__ extern __inline float32x2_t
11750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11751 vbsl_f32 (uint32x2_t __a, float32x2_t __b, float32x2_t __c)
11752 {
11753   return __builtin_aarch64_simd_bslv2sf_suss (__a, __b, __c);
11754 }
11755
11756 __extension__ extern __inline float64x1_t
11757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11758 vbsl_f64 (uint64x1_t __a, float64x1_t __b, float64x1_t __c)
11759 {
11760   return (float64x1_t)
11761     { __builtin_aarch64_simd_bsldf_suss (__a[0], __b[0], __c[0]) };
11762 }
11763
11764 __extension__ extern __inline poly8x8_t
11765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11766 vbsl_p8 (uint8x8_t __a, poly8x8_t __b, poly8x8_t __c)
11767 {
11768   return __builtin_aarch64_simd_bslv8qi_pupp (__a, __b, __c);
11769 }
11770
11771 __extension__ extern __inline poly16x4_t
11772 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11773 vbsl_p16 (uint16x4_t __a, poly16x4_t __b, poly16x4_t __c)
11774 {
11775   return __builtin_aarch64_simd_bslv4hi_pupp (__a, __b, __c);
11776 }
11777
11778 __extension__ extern __inline int8x8_t
11779 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11780 vbsl_s8 (uint8x8_t __a, int8x8_t __b, int8x8_t __c)
11781 {
11782   return __builtin_aarch64_simd_bslv8qi_suss (__a, __b, __c);
11783 }
11784
11785 __extension__ extern __inline int16x4_t
11786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11787 vbsl_s16 (uint16x4_t __a, int16x4_t __b, int16x4_t __c)
11788 {
11789   return __builtin_aarch64_simd_bslv4hi_suss (__a, __b, __c);
11790 }
11791
11792 __extension__ extern __inline int32x2_t
11793 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11794 vbsl_s32 (uint32x2_t __a, int32x2_t __b, int32x2_t __c)
11795 {
11796   return __builtin_aarch64_simd_bslv2si_suss (__a, __b, __c);
11797 }
11798
11799 __extension__ extern __inline int64x1_t
11800 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11801 vbsl_s64 (uint64x1_t __a, int64x1_t __b, int64x1_t __c)
11802 {
11803   return (int64x1_t)
11804       {__builtin_aarch64_simd_bsldi_suss (__a[0], __b[0], __c[0])};
11805 }
11806
11807 __extension__ extern __inline uint8x8_t
11808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11809 vbsl_u8 (uint8x8_t __a, uint8x8_t __b, uint8x8_t __c)
11810 {
11811   return __builtin_aarch64_simd_bslv8qi_uuuu (__a, __b, __c);
11812 }
11813
11814 __extension__ extern __inline uint16x4_t
11815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11816 vbsl_u16 (uint16x4_t __a, uint16x4_t __b, uint16x4_t __c)
11817 {
11818   return __builtin_aarch64_simd_bslv4hi_uuuu (__a, __b, __c);
11819 }
11820
11821 __extension__ extern __inline uint32x2_t
11822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11823 vbsl_u32 (uint32x2_t __a, uint32x2_t __b, uint32x2_t __c)
11824 {
11825   return __builtin_aarch64_simd_bslv2si_uuuu (__a, __b, __c);
11826 }
11827
11828 __extension__ extern __inline uint64x1_t
11829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11830 vbsl_u64 (uint64x1_t __a, uint64x1_t __b, uint64x1_t __c)
11831 {
11832   return (uint64x1_t)
11833       {__builtin_aarch64_simd_bsldi_uuuu (__a[0], __b[0], __c[0])};
11834 }
11835
11836 __extension__ extern __inline float16x8_t
11837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11838 vbslq_f16 (uint16x8_t __a, float16x8_t __b, float16x8_t __c)
11839 {
11840   return __builtin_aarch64_simd_bslv8hf_suss (__a, __b, __c);
11841 }
11842
11843 __extension__ extern __inline float32x4_t
11844 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11845 vbslq_f32 (uint32x4_t __a, float32x4_t __b, float32x4_t __c)
11846 {
11847   return __builtin_aarch64_simd_bslv4sf_suss (__a, __b, __c);
11848 }
11849
11850 __extension__ extern __inline float64x2_t
11851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11852 vbslq_f64 (uint64x2_t __a, float64x2_t __b, float64x2_t __c)
11853 {
11854   return __builtin_aarch64_simd_bslv2df_suss (__a, __b, __c);
11855 }
11856
11857 __extension__ extern __inline poly8x16_t
11858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11859 vbslq_p8 (uint8x16_t __a, poly8x16_t __b, poly8x16_t __c)
11860 {
11861   return __builtin_aarch64_simd_bslv16qi_pupp (__a, __b, __c);
11862 }
11863
11864 __extension__ extern __inline poly16x8_t
11865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11866 vbslq_p16 (uint16x8_t __a, poly16x8_t __b, poly16x8_t __c)
11867 {
11868   return __builtin_aarch64_simd_bslv8hi_pupp (__a, __b, __c);
11869 }
11870
11871 __extension__ extern __inline int8x16_t
11872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11873 vbslq_s8 (uint8x16_t __a, int8x16_t __b, int8x16_t __c)
11874 {
11875   return __builtin_aarch64_simd_bslv16qi_suss (__a, __b, __c);
11876 }
11877
11878 __extension__ extern __inline int16x8_t
11879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11880 vbslq_s16 (uint16x8_t __a, int16x8_t __b, int16x8_t __c)
11881 {
11882   return __builtin_aarch64_simd_bslv8hi_suss (__a, __b, __c);
11883 }
11884
11885 __extension__ extern __inline int32x4_t
11886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11887 vbslq_s32 (uint32x4_t __a, int32x4_t __b, int32x4_t __c)
11888 {
11889   return __builtin_aarch64_simd_bslv4si_suss (__a, __b, __c);
11890 }
11891
11892 __extension__ extern __inline int64x2_t
11893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11894 vbslq_s64 (uint64x2_t __a, int64x2_t __b, int64x2_t __c)
11895 {
11896   return __builtin_aarch64_simd_bslv2di_suss (__a, __b, __c);
11897 }
11898
11899 __extension__ extern __inline uint8x16_t
11900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11901 vbslq_u8 (uint8x16_t __a, uint8x16_t __b, uint8x16_t __c)
11902 {
11903   return __builtin_aarch64_simd_bslv16qi_uuuu (__a, __b, __c);
11904 }
11905
11906 __extension__ extern __inline uint16x8_t
11907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11908 vbslq_u16 (uint16x8_t __a, uint16x8_t __b, uint16x8_t __c)
11909 {
11910   return __builtin_aarch64_simd_bslv8hi_uuuu (__a, __b, __c);
11911 }
11912
11913 __extension__ extern __inline uint32x4_t
11914 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11915 vbslq_u32 (uint32x4_t __a, uint32x4_t __b, uint32x4_t __c)
11916 {
11917   return __builtin_aarch64_simd_bslv4si_uuuu (__a, __b, __c);
11918 }
11919
11920 __extension__ extern __inline uint64x2_t
11921 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11922 vbslq_u64 (uint64x2_t __a, uint64x2_t __b, uint64x2_t __c)
11923 {
11924   return __builtin_aarch64_simd_bslv2di_uuuu (__a, __b, __c);
11925 }
11926
11927 /* ARMv8.1-A instrinsics.  */
11928 #pragma GCC push_options
11929 #pragma GCC target ("arch=armv8.1-a")
11930
11931 __extension__ extern __inline int16x4_t
11932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11933 vqrdmlah_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c)
11934 {
11935   return __builtin_aarch64_sqrdmlahv4hi (__a, __b, __c);
11936 }
11937
11938 __extension__ extern __inline int32x2_t
11939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11940 vqrdmlah_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c)
11941 {
11942   return __builtin_aarch64_sqrdmlahv2si (__a, __b, __c);
11943 }
11944
11945 __extension__ extern __inline int16x8_t
11946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11947 vqrdmlahq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
11948 {
11949   return __builtin_aarch64_sqrdmlahv8hi (__a, __b, __c);
11950 }
11951
11952 __extension__ extern __inline int32x4_t
11953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11954 vqrdmlahq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
11955 {
11956   return __builtin_aarch64_sqrdmlahv4si (__a, __b, __c);
11957 }
11958
11959 __extension__ extern __inline int16x4_t
11960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11961 vqrdmlsh_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c)
11962 {
11963   return __builtin_aarch64_sqrdmlshv4hi (__a, __b, __c);
11964 }
11965
11966 __extension__ extern __inline int32x2_t
11967 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11968 vqrdmlsh_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c)
11969 {
11970   return __builtin_aarch64_sqrdmlshv2si (__a, __b, __c);
11971 }
11972
11973 __extension__ extern __inline int16x8_t
11974 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11975 vqrdmlshq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c)
11976 {
11977   return __builtin_aarch64_sqrdmlshv8hi (__a, __b, __c);
11978 }
11979
11980 __extension__ extern __inline int32x4_t
11981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11982 vqrdmlshq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c)
11983 {
11984   return __builtin_aarch64_sqrdmlshv4si (__a, __b, __c);
11985 }
11986
11987 __extension__ extern __inline int16x4_t
11988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11989 vqrdmlah_laneq_s16 (int16x4_t __a, int16x4_t __b, int16x8_t __c, const int __d)
11990 {
11991   return  __builtin_aarch64_sqrdmlah_laneqv4hi (__a, __b, __c, __d);
11992 }
11993
11994 __extension__ extern __inline int32x2_t
11995 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
11996 vqrdmlah_laneq_s32 (int32x2_t __a, int32x2_t __b, int32x4_t __c, const int __d)
11997 {
11998   return __builtin_aarch64_sqrdmlah_laneqv2si (__a, __b, __c, __d);
11999 }
12000
12001 __extension__ extern __inline int16x8_t
12002 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12003 vqrdmlahq_laneq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c, const int __d)
12004 {
12005   return __builtin_aarch64_sqrdmlah_laneqv8hi (__a, __b, __c, __d);
12006 }
12007
12008 __extension__ extern __inline int32x4_t
12009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12010 vqrdmlahq_laneq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c, const int __d)
12011 {
12012   return __builtin_aarch64_sqrdmlah_laneqv4si (__a, __b, __c, __d);
12013 }
12014
12015 __extension__ extern __inline int16x4_t
12016 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12017 vqrdmlsh_laneq_s16 (int16x4_t __a, int16x4_t __b, int16x8_t __c, const int __d)
12018 {
12019   return  __builtin_aarch64_sqrdmlsh_laneqv4hi (__a, __b, __c, __d);
12020 }
12021
12022 __extension__ extern __inline int32x2_t
12023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12024 vqrdmlsh_laneq_s32 (int32x2_t __a, int32x2_t __b, int32x4_t __c, const int __d)
12025 {
12026   return __builtin_aarch64_sqrdmlsh_laneqv2si (__a, __b, __c, __d);
12027 }
12028
12029 __extension__ extern __inline int16x8_t
12030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12031 vqrdmlshq_laneq_s16 (int16x8_t __a, int16x8_t __b, int16x8_t __c, const int __d)
12032 {
12033   return __builtin_aarch64_sqrdmlsh_laneqv8hi (__a, __b, __c, __d);
12034 }
12035
12036 __extension__ extern __inline int32x4_t
12037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12038 vqrdmlshq_laneq_s32 (int32x4_t __a, int32x4_t __b, int32x4_t __c, const int __d)
12039 {
12040   return __builtin_aarch64_sqrdmlsh_laneqv4si (__a, __b, __c, __d);
12041 }
12042
12043 __extension__ extern __inline int16x4_t
12044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12045 vqrdmlah_lane_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c, const int __d)
12046 {
12047   return  __builtin_aarch64_sqrdmlah_lanev4hi (__a, __b, __c, __d);
12048 }
12049
12050 __extension__ extern __inline int32x2_t
12051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12052 vqrdmlah_lane_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c, const int __d)
12053 {
12054   return __builtin_aarch64_sqrdmlah_lanev2si (__a, __b, __c, __d);
12055 }
12056
12057 __extension__ extern __inline int16x8_t
12058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12059 vqrdmlahq_lane_s16 (int16x8_t __a, int16x8_t __b, int16x4_t __c, const int __d)
12060 {
12061   return __builtin_aarch64_sqrdmlah_lanev8hi (__a, __b, __c, __d);
12062 }
12063
12064 __extension__ extern __inline int32x4_t
12065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12066 vqrdmlahq_lane_s32 (int32x4_t __a, int32x4_t __b, int32x2_t __c, const int __d)
12067 {
12068   return __builtin_aarch64_sqrdmlah_lanev4si (__a, __b, __c, __d);
12069 }
12070
12071 __extension__ extern __inline int16_t
12072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12073 vqrdmlahh_s16 (int16_t __a, int16_t __b, int16_t __c)
12074 {
12075   return (int16_t) __builtin_aarch64_sqrdmlahhi (__a, __b, __c);
12076 }
12077
12078 __extension__ extern __inline int16_t
12079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12080 vqrdmlahh_lane_s16 (int16_t __a, int16_t __b, int16x4_t __c, const int __d)
12081 {
12082   return __builtin_aarch64_sqrdmlah_lanehi (__a, __b, __c, __d);
12083 }
12084
12085 __extension__ extern __inline int16_t
12086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12087 vqrdmlahh_laneq_s16 (int16_t __a, int16_t __b, int16x8_t __c, const int __d)
12088 {
12089   return __builtin_aarch64_sqrdmlah_laneqhi (__a, __b, __c, __d);
12090 }
12091
12092 __extension__ extern __inline int32_t
12093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12094 vqrdmlahs_s32 (int32_t __a, int32_t __b, int32_t __c)
12095 {
12096   return (int32_t) __builtin_aarch64_sqrdmlahsi (__a, __b, __c);
12097 }
12098
12099 __extension__ extern __inline int32_t
12100 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12101 vqrdmlahs_lane_s32 (int32_t __a, int32_t __b, int32x2_t __c, const int __d)
12102 {
12103   return __builtin_aarch64_sqrdmlah_lanesi (__a, __b, __c, __d);
12104 }
12105
12106 __extension__ extern __inline int32_t
12107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12108 vqrdmlahs_laneq_s32 (int32_t __a, int32_t __b, int32x4_t __c, const int __d)
12109 {
12110   return __builtin_aarch64_sqrdmlah_laneqsi (__a, __b, __c, __d);
12111 }
12112
12113 __extension__ extern __inline int16x4_t
12114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12115 vqrdmlsh_lane_s16 (int16x4_t __a, int16x4_t __b, int16x4_t __c, const int __d)
12116 {
12117   return  __builtin_aarch64_sqrdmlsh_lanev4hi (__a, __b, __c, __d);
12118 }
12119
12120 __extension__ extern __inline int32x2_t
12121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12122 vqrdmlsh_lane_s32 (int32x2_t __a, int32x2_t __b, int32x2_t __c, const int __d)
12123 {
12124   return __builtin_aarch64_sqrdmlsh_lanev2si (__a, __b, __c, __d);
12125 }
12126
12127 __extension__ extern __inline int16x8_t
12128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12129 vqrdmlshq_lane_s16 (int16x8_t __a, int16x8_t __b, int16x4_t __c, const int __d)
12130 {
12131   return __builtin_aarch64_sqrdmlsh_lanev8hi (__a, __b, __c, __d);
12132 }
12133
12134 __extension__ extern __inline int32x4_t
12135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12136 vqrdmlshq_lane_s32 (int32x4_t __a, int32x4_t __b, int32x2_t __c, const int __d)
12137 {
12138   return __builtin_aarch64_sqrdmlsh_lanev4si (__a, __b, __c, __d);
12139 }
12140
12141 __extension__ extern __inline int16_t
12142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12143 vqrdmlshh_s16 (int16_t __a, int16_t __b, int16_t __c)
12144 {
12145   return (int16_t) __builtin_aarch64_sqrdmlshhi (__a, __b, __c);
12146 }
12147
12148 __extension__ extern __inline int16_t
12149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12150 vqrdmlshh_lane_s16 (int16_t __a, int16_t __b, int16x4_t __c, const int __d)
12151 {
12152   return __builtin_aarch64_sqrdmlsh_lanehi (__a, __b, __c, __d);
12153 }
12154
12155 __extension__ extern __inline int16_t
12156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12157 vqrdmlshh_laneq_s16 (int16_t __a, int16_t __b, int16x8_t __c, const int __d)
12158 {
12159   return __builtin_aarch64_sqrdmlsh_laneqhi (__a, __b, __c, __d);
12160 }
12161
12162 __extension__ extern __inline int32_t
12163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12164 vqrdmlshs_s32 (int32_t __a, int32_t __b, int32_t __c)
12165 {
12166   return (int32_t) __builtin_aarch64_sqrdmlshsi (__a, __b, __c);
12167 }
12168
12169 __extension__ extern __inline int32_t
12170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12171 vqrdmlshs_lane_s32 (int32_t __a, int32_t __b, int32x2_t __c, const int __d)
12172 {
12173   return __builtin_aarch64_sqrdmlsh_lanesi (__a, __b, __c, __d);
12174 }
12175
12176 __extension__ extern __inline int32_t
12177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12178 vqrdmlshs_laneq_s32 (int32_t __a, int32_t __b, int32x4_t __c, const int __d)
12179 {
12180   return __builtin_aarch64_sqrdmlsh_laneqsi (__a, __b, __c, __d);
12181 }
12182 #pragma GCC pop_options
12183
12184 #pragma GCC push_options
12185 #pragma GCC target ("+nothing+crypto")
12186 /* vaes  */
12187
12188 __extension__ extern __inline uint8x16_t
12189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12190 vaeseq_u8 (uint8x16_t data, uint8x16_t key)
12191 {
12192   return __builtin_aarch64_crypto_aesev16qi_uuu (data, key);
12193 }
12194
12195 __extension__ extern __inline uint8x16_t
12196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12197 vaesdq_u8 (uint8x16_t data, uint8x16_t key)
12198 {
12199   return __builtin_aarch64_crypto_aesdv16qi_uuu (data, key);
12200 }
12201
12202 __extension__ extern __inline uint8x16_t
12203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12204 vaesmcq_u8 (uint8x16_t data)
12205 {
12206   return __builtin_aarch64_crypto_aesmcv16qi_uu (data);
12207 }
12208
12209 __extension__ extern __inline uint8x16_t
12210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12211 vaesimcq_u8 (uint8x16_t data)
12212 {
12213   return __builtin_aarch64_crypto_aesimcv16qi_uu (data);
12214 }
12215 #pragma GCC pop_options
12216
12217 /* vcage  */
12218
12219 __extension__ extern __inline uint64x1_t
12220 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12221 vcage_f64 (float64x1_t __a, float64x1_t __b)
12222 {
12223   return vabs_f64 (__a) >= vabs_f64 (__b);
12224 }
12225
12226 __extension__ extern __inline uint32_t
12227 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12228 vcages_f32 (float32_t __a, float32_t __b)
12229 {
12230   return __builtin_fabsf (__a) >= __builtin_fabsf (__b) ? -1 : 0;
12231 }
12232
12233 __extension__ extern __inline uint32x2_t
12234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12235 vcage_f32 (float32x2_t __a, float32x2_t __b)
12236 {
12237   return vabs_f32 (__a) >= vabs_f32 (__b);
12238 }
12239
12240 __extension__ extern __inline uint32x4_t
12241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12242 vcageq_f32 (float32x4_t __a, float32x4_t __b)
12243 {
12244   return vabsq_f32 (__a) >= vabsq_f32 (__b);
12245 }
12246
12247 __extension__ extern __inline uint64_t
12248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12249 vcaged_f64 (float64_t __a, float64_t __b)
12250 {
12251   return __builtin_fabs (__a) >= __builtin_fabs (__b) ? -1 : 0;
12252 }
12253
12254 __extension__ extern __inline uint64x2_t
12255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12256 vcageq_f64 (float64x2_t __a, float64x2_t __b)
12257 {
12258   return vabsq_f64 (__a) >= vabsq_f64 (__b);
12259 }
12260
12261 /* vcagt  */
12262
12263 __extension__ extern __inline uint32_t
12264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12265 vcagts_f32 (float32_t __a, float32_t __b)
12266 {
12267   return __builtin_fabsf (__a) > __builtin_fabsf (__b) ? -1 : 0;
12268 }
12269
12270 __extension__ extern __inline uint32x2_t
12271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12272 vcagt_f32 (float32x2_t __a, float32x2_t __b)
12273 {
12274   return vabs_f32 (__a) > vabs_f32 (__b);
12275 }
12276
12277 __extension__ extern __inline uint64x1_t
12278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12279 vcagt_f64 (float64x1_t __a, float64x1_t __b)
12280 {
12281   return vabs_f64 (__a) > vabs_f64 (__b);
12282 }
12283
12284 __extension__ extern __inline uint32x4_t
12285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12286 vcagtq_f32 (float32x4_t __a, float32x4_t __b)
12287 {
12288   return vabsq_f32 (__a) > vabsq_f32 (__b);
12289 }
12290
12291 __extension__ extern __inline uint64_t
12292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12293 vcagtd_f64 (float64_t __a, float64_t __b)
12294 {
12295   return __builtin_fabs (__a) > __builtin_fabs (__b) ? -1 : 0;
12296 }
12297
12298 __extension__ extern __inline uint64x2_t
12299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12300 vcagtq_f64 (float64x2_t __a, float64x2_t __b)
12301 {
12302   return vabsq_f64 (__a) > vabsq_f64 (__b);
12303 }
12304
12305 /* vcale  */
12306
12307 __extension__ extern __inline uint32x2_t
12308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12309 vcale_f32 (float32x2_t __a, float32x2_t __b)
12310 {
12311   return vabs_f32 (__a) <= vabs_f32 (__b);
12312 }
12313
12314 __extension__ extern __inline uint64x1_t
12315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12316 vcale_f64 (float64x1_t __a, float64x1_t __b)
12317 {
12318   return vabs_f64 (__a) <= vabs_f64 (__b);
12319 }
12320
12321 __extension__ extern __inline uint64_t
12322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12323 vcaled_f64 (float64_t __a, float64_t __b)
12324 {
12325   return __builtin_fabs (__a) <= __builtin_fabs (__b) ? -1 : 0;
12326 }
12327
12328 __extension__ extern __inline uint32_t
12329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12330 vcales_f32 (float32_t __a, float32_t __b)
12331 {
12332   return __builtin_fabsf (__a) <= __builtin_fabsf (__b) ? -1 : 0;
12333 }
12334
12335 __extension__ extern __inline uint32x4_t
12336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12337 vcaleq_f32 (float32x4_t __a, float32x4_t __b)
12338 {
12339   return vabsq_f32 (__a) <= vabsq_f32 (__b);
12340 }
12341
12342 __extension__ extern __inline uint64x2_t
12343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12344 vcaleq_f64 (float64x2_t __a, float64x2_t __b)
12345 {
12346   return vabsq_f64 (__a) <= vabsq_f64 (__b);
12347 }
12348
12349 /* vcalt  */
12350
12351 __extension__ extern __inline uint32x2_t
12352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12353 vcalt_f32 (float32x2_t __a, float32x2_t __b)
12354 {
12355   return vabs_f32 (__a) < vabs_f32 (__b);
12356 }
12357
12358 __extension__ extern __inline uint64x1_t
12359 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12360 vcalt_f64 (float64x1_t __a, float64x1_t __b)
12361 {
12362   return vabs_f64 (__a) < vabs_f64 (__b);
12363 }
12364
12365 __extension__ extern __inline uint64_t
12366 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12367 vcaltd_f64 (float64_t __a, float64_t __b)
12368 {
12369   return __builtin_fabs (__a) < __builtin_fabs (__b) ? -1 : 0;
12370 }
12371
12372 __extension__ extern __inline uint32x4_t
12373 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12374 vcaltq_f32 (float32x4_t __a, float32x4_t __b)
12375 {
12376   return vabsq_f32 (__a) < vabsq_f32 (__b);
12377 }
12378
12379 __extension__ extern __inline uint64x2_t
12380 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12381 vcaltq_f64 (float64x2_t __a, float64x2_t __b)
12382 {
12383   return vabsq_f64 (__a) < vabsq_f64 (__b);
12384 }
12385
12386 __extension__ extern __inline uint32_t
12387 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12388 vcalts_f32 (float32_t __a, float32_t __b)
12389 {
12390   return __builtin_fabsf (__a) < __builtin_fabsf (__b) ? -1 : 0;
12391 }
12392
12393 /* vceq - vector.  */
12394
12395 __extension__ extern __inline uint32x2_t
12396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12397 vceq_f32 (float32x2_t __a, float32x2_t __b)
12398 {
12399   return (uint32x2_t) (__a == __b);
12400 }
12401
12402 __extension__ extern __inline uint64x1_t
12403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12404 vceq_f64 (float64x1_t __a, float64x1_t __b)
12405 {
12406   return (uint64x1_t) (__a == __b);
12407 }
12408
12409 __extension__ extern __inline uint8x8_t
12410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12411 vceq_p8 (poly8x8_t __a, poly8x8_t __b)
12412 {
12413   return (uint8x8_t) (__a == __b);
12414 }
12415
12416 __extension__ extern __inline uint8x8_t
12417 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12418 vceq_s8 (int8x8_t __a, int8x8_t __b)
12419 {
12420   return (uint8x8_t) (__a == __b);
12421 }
12422
12423 __extension__ extern __inline uint16x4_t
12424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12425 vceq_s16 (int16x4_t __a, int16x4_t __b)
12426 {
12427   return (uint16x4_t) (__a == __b);
12428 }
12429
12430 __extension__ extern __inline uint32x2_t
12431 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12432 vceq_s32 (int32x2_t __a, int32x2_t __b)
12433 {
12434   return (uint32x2_t) (__a == __b);
12435 }
12436
12437 __extension__ extern __inline uint64x1_t
12438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12439 vceq_s64 (int64x1_t __a, int64x1_t __b)
12440 {
12441   return (uint64x1_t) (__a == __b);
12442 }
12443
12444 __extension__ extern __inline uint8x8_t
12445 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12446 vceq_u8 (uint8x8_t __a, uint8x8_t __b)
12447 {
12448   return (__a == __b);
12449 }
12450
12451 __extension__ extern __inline uint16x4_t
12452 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12453 vceq_u16 (uint16x4_t __a, uint16x4_t __b)
12454 {
12455   return (__a == __b);
12456 }
12457
12458 __extension__ extern __inline uint32x2_t
12459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12460 vceq_u32 (uint32x2_t __a, uint32x2_t __b)
12461 {
12462   return (__a == __b);
12463 }
12464
12465 __extension__ extern __inline uint64x1_t
12466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12467 vceq_u64 (uint64x1_t __a, uint64x1_t __b)
12468 {
12469   return (__a == __b);
12470 }
12471
12472 __extension__ extern __inline uint32x4_t
12473 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12474 vceqq_f32 (float32x4_t __a, float32x4_t __b)
12475 {
12476   return (uint32x4_t) (__a == __b);
12477 }
12478
12479 __extension__ extern __inline uint64x2_t
12480 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12481 vceqq_f64 (float64x2_t __a, float64x2_t __b)
12482 {
12483   return (uint64x2_t) (__a == __b);
12484 }
12485
12486 __extension__ extern __inline uint8x16_t
12487 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12488 vceqq_p8 (poly8x16_t __a, poly8x16_t __b)
12489 {
12490   return (uint8x16_t) (__a == __b);
12491 }
12492
12493 __extension__ extern __inline uint8x16_t
12494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12495 vceqq_s8 (int8x16_t __a, int8x16_t __b)
12496 {
12497   return (uint8x16_t) (__a == __b);
12498 }
12499
12500 __extension__ extern __inline uint16x8_t
12501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12502 vceqq_s16 (int16x8_t __a, int16x8_t __b)
12503 {
12504   return (uint16x8_t) (__a == __b);
12505 }
12506
12507 __extension__ extern __inline uint32x4_t
12508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12509 vceqq_s32 (int32x4_t __a, int32x4_t __b)
12510 {
12511   return (uint32x4_t) (__a == __b);
12512 }
12513
12514 __extension__ extern __inline uint64x2_t
12515 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12516 vceqq_s64 (int64x2_t __a, int64x2_t __b)
12517 {
12518   return (uint64x2_t) (__a == __b);
12519 }
12520
12521 __extension__ extern __inline uint8x16_t
12522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12523 vceqq_u8 (uint8x16_t __a, uint8x16_t __b)
12524 {
12525   return (__a == __b);
12526 }
12527
12528 __extension__ extern __inline uint16x8_t
12529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12530 vceqq_u16 (uint16x8_t __a, uint16x8_t __b)
12531 {
12532   return (__a == __b);
12533 }
12534
12535 __extension__ extern __inline uint32x4_t
12536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12537 vceqq_u32 (uint32x4_t __a, uint32x4_t __b)
12538 {
12539   return (__a == __b);
12540 }
12541
12542 __extension__ extern __inline uint64x2_t
12543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12544 vceqq_u64 (uint64x2_t __a, uint64x2_t __b)
12545 {
12546   return (__a == __b);
12547 }
12548
12549 /* vceq - scalar.  */
12550
12551 __extension__ extern __inline uint32_t
12552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12553 vceqs_f32 (float32_t __a, float32_t __b)
12554 {
12555   return __a == __b ? -1 : 0;
12556 }
12557
12558 __extension__ extern __inline uint64_t
12559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12560 vceqd_s64 (int64_t __a, int64_t __b)
12561 {
12562   return __a == __b ? -1ll : 0ll;
12563 }
12564
12565 __extension__ extern __inline uint64_t
12566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12567 vceqd_u64 (uint64_t __a, uint64_t __b)
12568 {
12569   return __a == __b ? -1ll : 0ll;
12570 }
12571
12572 __extension__ extern __inline uint64_t
12573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12574 vceqd_f64 (float64_t __a, float64_t __b)
12575 {
12576   return __a == __b ? -1ll : 0ll;
12577 }
12578
12579 /* vceqz - vector.  */
12580
12581 __extension__ extern __inline uint32x2_t
12582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12583 vceqz_f32 (float32x2_t __a)
12584 {
12585   return (uint32x2_t) (__a == 0.0f);
12586 }
12587
12588 __extension__ extern __inline uint64x1_t
12589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12590 vceqz_f64 (float64x1_t __a)
12591 {
12592   return (uint64x1_t) (__a == (float64x1_t) {0.0});
12593 }
12594
12595 __extension__ extern __inline uint8x8_t
12596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12597 vceqz_p8 (poly8x8_t __a)
12598 {
12599   return (uint8x8_t) (__a == 0);
12600 }
12601
12602 __extension__ extern __inline uint8x8_t
12603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12604 vceqz_s8 (int8x8_t __a)
12605 {
12606   return (uint8x8_t) (__a == 0);
12607 }
12608
12609 __extension__ extern __inline uint16x4_t
12610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12611 vceqz_s16 (int16x4_t __a)
12612 {
12613   return (uint16x4_t) (__a == 0);
12614 }
12615
12616 __extension__ extern __inline uint32x2_t
12617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12618 vceqz_s32 (int32x2_t __a)
12619 {
12620   return (uint32x2_t) (__a == 0);
12621 }
12622
12623 __extension__ extern __inline uint64x1_t
12624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12625 vceqz_s64 (int64x1_t __a)
12626 {
12627   return (uint64x1_t) (__a == __AARCH64_INT64_C (0));
12628 }
12629
12630 __extension__ extern __inline uint8x8_t
12631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12632 vceqz_u8 (uint8x8_t __a)
12633 {
12634   return (__a == 0);
12635 }
12636
12637 __extension__ extern __inline uint16x4_t
12638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12639 vceqz_u16 (uint16x4_t __a)
12640 {
12641   return (__a == 0);
12642 }
12643
12644 __extension__ extern __inline uint32x2_t
12645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12646 vceqz_u32 (uint32x2_t __a)
12647 {
12648   return (__a == 0);
12649 }
12650
12651 __extension__ extern __inline uint64x1_t
12652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12653 vceqz_u64 (uint64x1_t __a)
12654 {
12655   return (__a == __AARCH64_UINT64_C (0));
12656 }
12657
12658 __extension__ extern __inline uint32x4_t
12659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12660 vceqzq_f32 (float32x4_t __a)
12661 {
12662   return (uint32x4_t) (__a == 0.0f);
12663 }
12664
12665 __extension__ extern __inline uint64x2_t
12666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12667 vceqzq_f64 (float64x2_t __a)
12668 {
12669   return (uint64x2_t) (__a == 0.0f);
12670 }
12671
12672 __extension__ extern __inline uint8x16_t
12673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12674 vceqzq_p8 (poly8x16_t __a)
12675 {
12676   return (uint8x16_t) (__a == 0);
12677 }
12678
12679 __extension__ extern __inline uint8x16_t
12680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12681 vceqzq_s8 (int8x16_t __a)
12682 {
12683   return (uint8x16_t) (__a == 0);
12684 }
12685
12686 __extension__ extern __inline uint16x8_t
12687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12688 vceqzq_s16 (int16x8_t __a)
12689 {
12690   return (uint16x8_t) (__a == 0);
12691 }
12692
12693 __extension__ extern __inline uint32x4_t
12694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12695 vceqzq_s32 (int32x4_t __a)
12696 {
12697   return (uint32x4_t) (__a == 0);
12698 }
12699
12700 __extension__ extern __inline uint64x2_t
12701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12702 vceqzq_s64 (int64x2_t __a)
12703 {
12704   return (uint64x2_t) (__a == __AARCH64_INT64_C (0));
12705 }
12706
12707 __extension__ extern __inline uint8x16_t
12708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12709 vceqzq_u8 (uint8x16_t __a)
12710 {
12711   return (__a == 0);
12712 }
12713
12714 __extension__ extern __inline uint16x8_t
12715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12716 vceqzq_u16 (uint16x8_t __a)
12717 {
12718   return (__a == 0);
12719 }
12720
12721 __extension__ extern __inline uint32x4_t
12722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12723 vceqzq_u32 (uint32x4_t __a)
12724 {
12725   return (__a == 0);
12726 }
12727
12728 __extension__ extern __inline uint64x2_t
12729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12730 vceqzq_u64 (uint64x2_t __a)
12731 {
12732   return (__a == __AARCH64_UINT64_C (0));
12733 }
12734
12735 /* vceqz - scalar.  */
12736
12737 __extension__ extern __inline uint32_t
12738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12739 vceqzs_f32 (float32_t __a)
12740 {
12741   return __a == 0.0f ? -1 : 0;
12742 }
12743
12744 __extension__ extern __inline uint64_t
12745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12746 vceqzd_s64 (int64_t __a)
12747 {
12748   return __a == 0 ? -1ll : 0ll;
12749 }
12750
12751 __extension__ extern __inline uint64_t
12752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12753 vceqzd_u64 (uint64_t __a)
12754 {
12755   return __a == 0 ? -1ll : 0ll;
12756 }
12757
12758 __extension__ extern __inline uint64_t
12759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12760 vceqzd_f64 (float64_t __a)
12761 {
12762   return __a == 0.0 ? -1ll : 0ll;
12763 }
12764
12765 /* vcge - vector.  */
12766
12767 __extension__ extern __inline uint32x2_t
12768 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12769 vcge_f32 (float32x2_t __a, float32x2_t __b)
12770 {
12771   return (uint32x2_t) (__a >= __b);
12772 }
12773
12774 __extension__ extern __inline uint64x1_t
12775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12776 vcge_f64 (float64x1_t __a, float64x1_t __b)
12777 {
12778   return (uint64x1_t) (__a >= __b);
12779 }
12780
12781 __extension__ extern __inline uint8x8_t
12782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12783 vcge_s8 (int8x8_t __a, int8x8_t __b)
12784 {
12785   return (uint8x8_t) (__a >= __b);
12786 }
12787
12788 __extension__ extern __inline uint16x4_t
12789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12790 vcge_s16 (int16x4_t __a, int16x4_t __b)
12791 {
12792   return (uint16x4_t) (__a >= __b);
12793 }
12794
12795 __extension__ extern __inline uint32x2_t
12796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12797 vcge_s32 (int32x2_t __a, int32x2_t __b)
12798 {
12799   return (uint32x2_t) (__a >= __b);
12800 }
12801
12802 __extension__ extern __inline uint64x1_t
12803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12804 vcge_s64 (int64x1_t __a, int64x1_t __b)
12805 {
12806   return (uint64x1_t) (__a >= __b);
12807 }
12808
12809 __extension__ extern __inline uint8x8_t
12810 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12811 vcge_u8 (uint8x8_t __a, uint8x8_t __b)
12812 {
12813   return (__a >= __b);
12814 }
12815
12816 __extension__ extern __inline uint16x4_t
12817 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12818 vcge_u16 (uint16x4_t __a, uint16x4_t __b)
12819 {
12820   return (__a >= __b);
12821 }
12822
12823 __extension__ extern __inline uint32x2_t
12824 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12825 vcge_u32 (uint32x2_t __a, uint32x2_t __b)
12826 {
12827   return (__a >= __b);
12828 }
12829
12830 __extension__ extern __inline uint64x1_t
12831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12832 vcge_u64 (uint64x1_t __a, uint64x1_t __b)
12833 {
12834   return (__a >= __b);
12835 }
12836
12837 __extension__ extern __inline uint32x4_t
12838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12839 vcgeq_f32 (float32x4_t __a, float32x4_t __b)
12840 {
12841   return (uint32x4_t) (__a >= __b);
12842 }
12843
12844 __extension__ extern __inline uint64x2_t
12845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12846 vcgeq_f64 (float64x2_t __a, float64x2_t __b)
12847 {
12848   return (uint64x2_t) (__a >= __b);
12849 }
12850
12851 __extension__ extern __inline uint8x16_t
12852 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12853 vcgeq_s8 (int8x16_t __a, int8x16_t __b)
12854 {
12855   return (uint8x16_t) (__a >= __b);
12856 }
12857
12858 __extension__ extern __inline uint16x8_t
12859 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12860 vcgeq_s16 (int16x8_t __a, int16x8_t __b)
12861 {
12862   return (uint16x8_t) (__a >= __b);
12863 }
12864
12865 __extension__ extern __inline uint32x4_t
12866 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12867 vcgeq_s32 (int32x4_t __a, int32x4_t __b)
12868 {
12869   return (uint32x4_t) (__a >= __b);
12870 }
12871
12872 __extension__ extern __inline uint64x2_t
12873 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12874 vcgeq_s64 (int64x2_t __a, int64x2_t __b)
12875 {
12876   return (uint64x2_t) (__a >= __b);
12877 }
12878
12879 __extension__ extern __inline uint8x16_t
12880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12881 vcgeq_u8 (uint8x16_t __a, uint8x16_t __b)
12882 {
12883   return (__a >= __b);
12884 }
12885
12886 __extension__ extern __inline uint16x8_t
12887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12888 vcgeq_u16 (uint16x8_t __a, uint16x8_t __b)
12889 {
12890   return (__a >= __b);
12891 }
12892
12893 __extension__ extern __inline uint32x4_t
12894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12895 vcgeq_u32 (uint32x4_t __a, uint32x4_t __b)
12896 {
12897   return (__a >= __b);
12898 }
12899
12900 __extension__ extern __inline uint64x2_t
12901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12902 vcgeq_u64 (uint64x2_t __a, uint64x2_t __b)
12903 {
12904   return (__a >= __b);
12905 }
12906
12907 /* vcge - scalar.  */
12908
12909 __extension__ extern __inline uint32_t
12910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12911 vcges_f32 (float32_t __a, float32_t __b)
12912 {
12913   return __a >= __b ? -1 : 0;
12914 }
12915
12916 __extension__ extern __inline uint64_t
12917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12918 vcged_s64 (int64_t __a, int64_t __b)
12919 {
12920   return __a >= __b ? -1ll : 0ll;
12921 }
12922
12923 __extension__ extern __inline uint64_t
12924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12925 vcged_u64 (uint64_t __a, uint64_t __b)
12926 {
12927   return __a >= __b ? -1ll : 0ll;
12928 }
12929
12930 __extension__ extern __inline uint64_t
12931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12932 vcged_f64 (float64_t __a, float64_t __b)
12933 {
12934   return __a >= __b ? -1ll : 0ll;
12935 }
12936
12937 /* vcgez - vector.  */
12938
12939 __extension__ extern __inline uint32x2_t
12940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12941 vcgez_f32 (float32x2_t __a)
12942 {
12943   return (uint32x2_t) (__a >= 0.0f);
12944 }
12945
12946 __extension__ extern __inline uint64x1_t
12947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12948 vcgez_f64 (float64x1_t __a)
12949 {
12950   return (uint64x1_t) (__a[0] >= (float64x1_t) {0.0});
12951 }
12952
12953 __extension__ extern __inline uint8x8_t
12954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12955 vcgez_s8 (int8x8_t __a)
12956 {
12957   return (uint8x8_t) (__a >= 0);
12958 }
12959
12960 __extension__ extern __inline uint16x4_t
12961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12962 vcgez_s16 (int16x4_t __a)
12963 {
12964   return (uint16x4_t) (__a >= 0);
12965 }
12966
12967 __extension__ extern __inline uint32x2_t
12968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12969 vcgez_s32 (int32x2_t __a)
12970 {
12971   return (uint32x2_t) (__a >= 0);
12972 }
12973
12974 __extension__ extern __inline uint64x1_t
12975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12976 vcgez_s64 (int64x1_t __a)
12977 {
12978   return (uint64x1_t) (__a >= __AARCH64_INT64_C (0));
12979 }
12980
12981 __extension__ extern __inline uint32x4_t
12982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12983 vcgezq_f32 (float32x4_t __a)
12984 {
12985   return (uint32x4_t) (__a >= 0.0f);
12986 }
12987
12988 __extension__ extern __inline uint64x2_t
12989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12990 vcgezq_f64 (float64x2_t __a)
12991 {
12992   return (uint64x2_t) (__a >= 0.0);
12993 }
12994
12995 __extension__ extern __inline uint8x16_t
12996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
12997 vcgezq_s8 (int8x16_t __a)
12998 {
12999   return (uint8x16_t) (__a >= 0);
13000 }
13001
13002 __extension__ extern __inline uint16x8_t
13003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13004 vcgezq_s16 (int16x8_t __a)
13005 {
13006   return (uint16x8_t) (__a >= 0);
13007 }
13008
13009 __extension__ extern __inline uint32x4_t
13010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13011 vcgezq_s32 (int32x4_t __a)
13012 {
13013   return (uint32x4_t) (__a >= 0);
13014 }
13015
13016 __extension__ extern __inline uint64x2_t
13017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13018 vcgezq_s64 (int64x2_t __a)
13019 {
13020   return (uint64x2_t) (__a >= __AARCH64_INT64_C (0));
13021 }
13022
13023 /* vcgez - scalar.  */
13024
13025 __extension__ extern __inline uint32_t
13026 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13027 vcgezs_f32 (float32_t __a)
13028 {
13029   return __a >= 0.0f ? -1 : 0;
13030 }
13031
13032 __extension__ extern __inline uint64_t
13033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13034 vcgezd_s64 (int64_t __a)
13035 {
13036   return __a >= 0 ? -1ll : 0ll;
13037 }
13038
13039 __extension__ extern __inline uint64_t
13040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13041 vcgezd_f64 (float64_t __a)
13042 {
13043   return __a >= 0.0 ? -1ll : 0ll;
13044 }
13045
13046 /* vcgt - vector.  */
13047
13048 __extension__ extern __inline uint32x2_t
13049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13050 vcgt_f32 (float32x2_t __a, float32x2_t __b)
13051 {
13052   return (uint32x2_t) (__a > __b);
13053 }
13054
13055 __extension__ extern __inline uint64x1_t
13056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13057 vcgt_f64 (float64x1_t __a, float64x1_t __b)
13058 {
13059   return (uint64x1_t) (__a > __b);
13060 }
13061
13062 __extension__ extern __inline uint8x8_t
13063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13064 vcgt_s8 (int8x8_t __a, int8x8_t __b)
13065 {
13066   return (uint8x8_t) (__a > __b);
13067 }
13068
13069 __extension__ extern __inline uint16x4_t
13070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13071 vcgt_s16 (int16x4_t __a, int16x4_t __b)
13072 {
13073   return (uint16x4_t) (__a > __b);
13074 }
13075
13076 __extension__ extern __inline uint32x2_t
13077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13078 vcgt_s32 (int32x2_t __a, int32x2_t __b)
13079 {
13080   return (uint32x2_t) (__a > __b);
13081 }
13082
13083 __extension__ extern __inline uint64x1_t
13084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13085 vcgt_s64 (int64x1_t __a, int64x1_t __b)
13086 {
13087   return (uint64x1_t) (__a > __b);
13088 }
13089
13090 __extension__ extern __inline uint8x8_t
13091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13092 vcgt_u8 (uint8x8_t __a, uint8x8_t __b)
13093 {
13094   return (__a > __b);
13095 }
13096
13097 __extension__ extern __inline uint16x4_t
13098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13099 vcgt_u16 (uint16x4_t __a, uint16x4_t __b)
13100 {
13101   return (__a > __b);
13102 }
13103
13104 __extension__ extern __inline uint32x2_t
13105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13106 vcgt_u32 (uint32x2_t __a, uint32x2_t __b)
13107 {
13108   return (__a > __b);
13109 }
13110
13111 __extension__ extern __inline uint64x1_t
13112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13113 vcgt_u64 (uint64x1_t __a, uint64x1_t __b)
13114 {
13115   return (__a > __b);
13116 }
13117
13118 __extension__ extern __inline uint32x4_t
13119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13120 vcgtq_f32 (float32x4_t __a, float32x4_t __b)
13121 {
13122   return (uint32x4_t) (__a > __b);
13123 }
13124
13125 __extension__ extern __inline uint64x2_t
13126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13127 vcgtq_f64 (float64x2_t __a, float64x2_t __b)
13128 {
13129   return (uint64x2_t) (__a > __b);
13130 }
13131
13132 __extension__ extern __inline uint8x16_t
13133 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13134 vcgtq_s8 (int8x16_t __a, int8x16_t __b)
13135 {
13136   return (uint8x16_t) (__a > __b);
13137 }
13138
13139 __extension__ extern __inline uint16x8_t
13140 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13141 vcgtq_s16 (int16x8_t __a, int16x8_t __b)
13142 {
13143   return (uint16x8_t) (__a > __b);
13144 }
13145
13146 __extension__ extern __inline uint32x4_t
13147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13148 vcgtq_s32 (int32x4_t __a, int32x4_t __b)
13149 {
13150   return (uint32x4_t) (__a > __b);
13151 }
13152
13153 __extension__ extern __inline uint64x2_t
13154 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13155 vcgtq_s64 (int64x2_t __a, int64x2_t __b)
13156 {
13157   return (uint64x2_t) (__a > __b);
13158 }
13159
13160 __extension__ extern __inline uint8x16_t
13161 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13162 vcgtq_u8 (uint8x16_t __a, uint8x16_t __b)
13163 {
13164   return (__a > __b);
13165 }
13166
13167 __extension__ extern __inline uint16x8_t
13168 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13169 vcgtq_u16 (uint16x8_t __a, uint16x8_t __b)
13170 {
13171   return (__a > __b);
13172 }
13173
13174 __extension__ extern __inline uint32x4_t
13175 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13176 vcgtq_u32 (uint32x4_t __a, uint32x4_t __b)
13177 {
13178   return (__a > __b);
13179 }
13180
13181 __extension__ extern __inline uint64x2_t
13182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13183 vcgtq_u64 (uint64x2_t __a, uint64x2_t __b)
13184 {
13185   return (__a > __b);
13186 }
13187
13188 /* vcgt - scalar.  */
13189
13190 __extension__ extern __inline uint32_t
13191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13192 vcgts_f32 (float32_t __a, float32_t __b)
13193 {
13194   return __a > __b ? -1 : 0;
13195 }
13196
13197 __extension__ extern __inline uint64_t
13198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13199 vcgtd_s64 (int64_t __a, int64_t __b)
13200 {
13201   return __a > __b ? -1ll : 0ll;
13202 }
13203
13204 __extension__ extern __inline uint64_t
13205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13206 vcgtd_u64 (uint64_t __a, uint64_t __b)
13207 {
13208   return __a > __b ? -1ll : 0ll;
13209 }
13210
13211 __extension__ extern __inline uint64_t
13212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13213 vcgtd_f64 (float64_t __a, float64_t __b)
13214 {
13215   return __a > __b ? -1ll : 0ll;
13216 }
13217
13218 /* vcgtz - vector.  */
13219
13220 __extension__ extern __inline uint32x2_t
13221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13222 vcgtz_f32 (float32x2_t __a)
13223 {
13224   return (uint32x2_t) (__a > 0.0f);
13225 }
13226
13227 __extension__ extern __inline uint64x1_t
13228 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13229 vcgtz_f64 (float64x1_t __a)
13230 {
13231   return (uint64x1_t) (__a > (float64x1_t) {0.0});
13232 }
13233
13234 __extension__ extern __inline uint8x8_t
13235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13236 vcgtz_s8 (int8x8_t __a)
13237 {
13238   return (uint8x8_t) (__a > 0);
13239 }
13240
13241 __extension__ extern __inline uint16x4_t
13242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13243 vcgtz_s16 (int16x4_t __a)
13244 {
13245   return (uint16x4_t) (__a > 0);
13246 }
13247
13248 __extension__ extern __inline uint32x2_t
13249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13250 vcgtz_s32 (int32x2_t __a)
13251 {
13252   return (uint32x2_t) (__a > 0);
13253 }
13254
13255 __extension__ extern __inline uint64x1_t
13256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13257 vcgtz_s64 (int64x1_t __a)
13258 {
13259   return (uint64x1_t) (__a > __AARCH64_INT64_C (0));
13260 }
13261
13262 __extension__ extern __inline uint32x4_t
13263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13264 vcgtzq_f32 (float32x4_t __a)
13265 {
13266   return (uint32x4_t) (__a > 0.0f);
13267 }
13268
13269 __extension__ extern __inline uint64x2_t
13270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13271 vcgtzq_f64 (float64x2_t __a)
13272 {
13273     return (uint64x2_t) (__a > 0.0);
13274 }
13275
13276 __extension__ extern __inline uint8x16_t
13277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13278 vcgtzq_s8 (int8x16_t __a)
13279 {
13280   return (uint8x16_t) (__a > 0);
13281 }
13282
13283 __extension__ extern __inline uint16x8_t
13284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13285 vcgtzq_s16 (int16x8_t __a)
13286 {
13287   return (uint16x8_t) (__a > 0);
13288 }
13289
13290 __extension__ extern __inline uint32x4_t
13291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13292 vcgtzq_s32 (int32x4_t __a)
13293 {
13294   return (uint32x4_t) (__a > 0);
13295 }
13296
13297 __extension__ extern __inline uint64x2_t
13298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13299 vcgtzq_s64 (int64x2_t __a)
13300 {
13301   return (uint64x2_t) (__a > __AARCH64_INT64_C (0));
13302 }
13303
13304 /* vcgtz - scalar.  */
13305
13306 __extension__ extern __inline uint32_t
13307 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13308 vcgtzs_f32 (float32_t __a)
13309 {
13310   return __a > 0.0f ? -1 : 0;
13311 }
13312
13313 __extension__ extern __inline uint64_t
13314 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13315 vcgtzd_s64 (int64_t __a)
13316 {
13317   return __a > 0 ? -1ll : 0ll;
13318 }
13319
13320 __extension__ extern __inline uint64_t
13321 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13322 vcgtzd_f64 (float64_t __a)
13323 {
13324   return __a > 0.0 ? -1ll : 0ll;
13325 }
13326
13327 /* vcle - vector.  */
13328
13329 __extension__ extern __inline uint32x2_t
13330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13331 vcle_f32 (float32x2_t __a, float32x2_t __b)
13332 {
13333   return (uint32x2_t) (__a <= __b);
13334 }
13335
13336 __extension__ extern __inline uint64x1_t
13337 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13338 vcle_f64 (float64x1_t __a, float64x1_t __b)
13339 {
13340   return (uint64x1_t) (__a <= __b);
13341 }
13342
13343 __extension__ extern __inline uint8x8_t
13344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13345 vcle_s8 (int8x8_t __a, int8x8_t __b)
13346 {
13347   return (uint8x8_t) (__a <= __b);
13348 }
13349
13350 __extension__ extern __inline uint16x4_t
13351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13352 vcle_s16 (int16x4_t __a, int16x4_t __b)
13353 {
13354   return (uint16x4_t) (__a <= __b);
13355 }
13356
13357 __extension__ extern __inline uint32x2_t
13358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13359 vcle_s32 (int32x2_t __a, int32x2_t __b)
13360 {
13361   return (uint32x2_t) (__a <= __b);
13362 }
13363
13364 __extension__ extern __inline uint64x1_t
13365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13366 vcle_s64 (int64x1_t __a, int64x1_t __b)
13367 {
13368   return (uint64x1_t) (__a <= __b);
13369 }
13370
13371 __extension__ extern __inline uint8x8_t
13372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13373 vcle_u8 (uint8x8_t __a, uint8x8_t __b)
13374 {
13375   return (__a <= __b);
13376 }
13377
13378 __extension__ extern __inline uint16x4_t
13379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13380 vcle_u16 (uint16x4_t __a, uint16x4_t __b)
13381 {
13382   return (__a <= __b);
13383 }
13384
13385 __extension__ extern __inline uint32x2_t
13386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13387 vcle_u32 (uint32x2_t __a, uint32x2_t __b)
13388 {
13389   return (__a <= __b);
13390 }
13391
13392 __extension__ extern __inline uint64x1_t
13393 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13394 vcle_u64 (uint64x1_t __a, uint64x1_t __b)
13395 {
13396   return (__a <= __b);
13397 }
13398
13399 __extension__ extern __inline uint32x4_t
13400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13401 vcleq_f32 (float32x4_t __a, float32x4_t __b)
13402 {
13403   return (uint32x4_t) (__a <= __b);
13404 }
13405
13406 __extension__ extern __inline uint64x2_t
13407 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13408 vcleq_f64 (float64x2_t __a, float64x2_t __b)
13409 {
13410   return (uint64x2_t) (__a <= __b);
13411 }
13412
13413 __extension__ extern __inline uint8x16_t
13414 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13415 vcleq_s8 (int8x16_t __a, int8x16_t __b)
13416 {
13417   return (uint8x16_t) (__a <= __b);
13418 }
13419
13420 __extension__ extern __inline uint16x8_t
13421 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13422 vcleq_s16 (int16x8_t __a, int16x8_t __b)
13423 {
13424   return (uint16x8_t) (__a <= __b);
13425 }
13426
13427 __extension__ extern __inline uint32x4_t
13428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13429 vcleq_s32 (int32x4_t __a, int32x4_t __b)
13430 {
13431   return (uint32x4_t) (__a <= __b);
13432 }
13433
13434 __extension__ extern __inline uint64x2_t
13435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13436 vcleq_s64 (int64x2_t __a, int64x2_t __b)
13437 {
13438   return (uint64x2_t) (__a <= __b);
13439 }
13440
13441 __extension__ extern __inline uint8x16_t
13442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13443 vcleq_u8 (uint8x16_t __a, uint8x16_t __b)
13444 {
13445   return (__a <= __b);
13446 }
13447
13448 __extension__ extern __inline uint16x8_t
13449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13450 vcleq_u16 (uint16x8_t __a, uint16x8_t __b)
13451 {
13452   return (__a <= __b);
13453 }
13454
13455 __extension__ extern __inline uint32x4_t
13456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13457 vcleq_u32 (uint32x4_t __a, uint32x4_t __b)
13458 {
13459   return (__a <= __b);
13460 }
13461
13462 __extension__ extern __inline uint64x2_t
13463 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13464 vcleq_u64 (uint64x2_t __a, uint64x2_t __b)
13465 {
13466   return (__a <= __b);
13467 }
13468
13469 /* vcle - scalar.  */
13470
13471 __extension__ extern __inline uint32_t
13472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13473 vcles_f32 (float32_t __a, float32_t __b)
13474 {
13475   return __a <= __b ? -1 : 0;
13476 }
13477
13478 __extension__ extern __inline uint64_t
13479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13480 vcled_s64 (int64_t __a, int64_t __b)
13481 {
13482   return __a <= __b ? -1ll : 0ll;
13483 }
13484
13485 __extension__ extern __inline uint64_t
13486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13487 vcled_u64 (uint64_t __a, uint64_t __b)
13488 {
13489   return __a <= __b ? -1ll : 0ll;
13490 }
13491
13492 __extension__ extern __inline uint64_t
13493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13494 vcled_f64 (float64_t __a, float64_t __b)
13495 {
13496   return __a <= __b ? -1ll : 0ll;
13497 }
13498
13499 /* vclez - vector.  */
13500
13501 __extension__ extern __inline uint32x2_t
13502 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13503 vclez_f32 (float32x2_t __a)
13504 {
13505   return (uint32x2_t) (__a <= 0.0f);
13506 }
13507
13508 __extension__ extern __inline uint64x1_t
13509 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13510 vclez_f64 (float64x1_t __a)
13511 {
13512   return (uint64x1_t) (__a <= (float64x1_t) {0.0});
13513 }
13514
13515 __extension__ extern __inline uint8x8_t
13516 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13517 vclez_s8 (int8x8_t __a)
13518 {
13519   return (uint8x8_t) (__a <= 0);
13520 }
13521
13522 __extension__ extern __inline uint16x4_t
13523 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13524 vclez_s16 (int16x4_t __a)
13525 {
13526   return (uint16x4_t) (__a <= 0);
13527 }
13528
13529 __extension__ extern __inline uint32x2_t
13530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13531 vclez_s32 (int32x2_t __a)
13532 {
13533   return (uint32x2_t) (__a <= 0);
13534 }
13535
13536 __extension__ extern __inline uint64x1_t
13537 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13538 vclez_s64 (int64x1_t __a)
13539 {
13540   return (uint64x1_t) (__a <= __AARCH64_INT64_C (0));
13541 }
13542
13543 __extension__ extern __inline uint32x4_t
13544 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13545 vclezq_f32 (float32x4_t __a)
13546 {
13547   return (uint32x4_t) (__a <= 0.0f);
13548 }
13549
13550 __extension__ extern __inline uint64x2_t
13551 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13552 vclezq_f64 (float64x2_t __a)
13553 {
13554   return (uint64x2_t) (__a <= 0.0);
13555 }
13556
13557 __extension__ extern __inline uint8x16_t
13558 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13559 vclezq_s8 (int8x16_t __a)
13560 {
13561   return (uint8x16_t) (__a <= 0);
13562 }
13563
13564 __extension__ extern __inline uint16x8_t
13565 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13566 vclezq_s16 (int16x8_t __a)
13567 {
13568   return (uint16x8_t) (__a <= 0);
13569 }
13570
13571 __extension__ extern __inline uint32x4_t
13572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13573 vclezq_s32 (int32x4_t __a)
13574 {
13575   return (uint32x4_t) (__a <= 0);
13576 }
13577
13578 __extension__ extern __inline uint64x2_t
13579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13580 vclezq_s64 (int64x2_t __a)
13581 {
13582   return (uint64x2_t) (__a <= __AARCH64_INT64_C (0));
13583 }
13584
13585 /* vclez - scalar.  */
13586
13587 __extension__ extern __inline uint32_t
13588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13589 vclezs_f32 (float32_t __a)
13590 {
13591   return __a <= 0.0f ? -1 : 0;
13592 }
13593
13594 __extension__ extern __inline uint64_t
13595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13596 vclezd_s64 (int64_t __a)
13597 {
13598   return __a <= 0 ? -1ll : 0ll;
13599 }
13600
13601 __extension__ extern __inline uint64_t
13602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13603 vclezd_f64 (float64_t __a)
13604 {
13605   return __a <= 0.0 ? -1ll : 0ll;
13606 }
13607
13608 /* vclt - vector.  */
13609
13610 __extension__ extern __inline uint32x2_t
13611 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13612 vclt_f32 (float32x2_t __a, float32x2_t __b)
13613 {
13614   return (uint32x2_t) (__a < __b);
13615 }
13616
13617 __extension__ extern __inline uint64x1_t
13618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13619 vclt_f64 (float64x1_t __a, float64x1_t __b)
13620 {
13621   return (uint64x1_t) (__a < __b);
13622 }
13623
13624 __extension__ extern __inline uint8x8_t
13625 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13626 vclt_s8 (int8x8_t __a, int8x8_t __b)
13627 {
13628   return (uint8x8_t) (__a < __b);
13629 }
13630
13631 __extension__ extern __inline uint16x4_t
13632 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13633 vclt_s16 (int16x4_t __a, int16x4_t __b)
13634 {
13635   return (uint16x4_t) (__a < __b);
13636 }
13637
13638 __extension__ extern __inline uint32x2_t
13639 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13640 vclt_s32 (int32x2_t __a, int32x2_t __b)
13641 {
13642   return (uint32x2_t) (__a < __b);
13643 }
13644
13645 __extension__ extern __inline uint64x1_t
13646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13647 vclt_s64 (int64x1_t __a, int64x1_t __b)
13648 {
13649   return (uint64x1_t) (__a < __b);
13650 }
13651
13652 __extension__ extern __inline uint8x8_t
13653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13654 vclt_u8 (uint8x8_t __a, uint8x8_t __b)
13655 {
13656   return (__a < __b);
13657 }
13658
13659 __extension__ extern __inline uint16x4_t
13660 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13661 vclt_u16 (uint16x4_t __a, uint16x4_t __b)
13662 {
13663   return (__a < __b);
13664 }
13665
13666 __extension__ extern __inline uint32x2_t
13667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13668 vclt_u32 (uint32x2_t __a, uint32x2_t __b)
13669 {
13670   return (__a < __b);
13671 }
13672
13673 __extension__ extern __inline uint64x1_t
13674 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13675 vclt_u64 (uint64x1_t __a, uint64x1_t __b)
13676 {
13677   return (__a < __b);
13678 }
13679
13680 __extension__ extern __inline uint32x4_t
13681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13682 vcltq_f32 (float32x4_t __a, float32x4_t __b)
13683 {
13684   return (uint32x4_t) (__a < __b);
13685 }
13686
13687 __extension__ extern __inline uint64x2_t
13688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13689 vcltq_f64 (float64x2_t __a, float64x2_t __b)
13690 {
13691   return (uint64x2_t) (__a < __b);
13692 }
13693
13694 __extension__ extern __inline uint8x16_t
13695 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13696 vcltq_s8 (int8x16_t __a, int8x16_t __b)
13697 {
13698   return (uint8x16_t) (__a < __b);
13699 }
13700
13701 __extension__ extern __inline uint16x8_t
13702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13703 vcltq_s16 (int16x8_t __a, int16x8_t __b)
13704 {
13705   return (uint16x8_t) (__a < __b);
13706 }
13707
13708 __extension__ extern __inline uint32x4_t
13709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13710 vcltq_s32 (int32x4_t __a, int32x4_t __b)
13711 {
13712   return (uint32x4_t) (__a < __b);
13713 }
13714
13715 __extension__ extern __inline uint64x2_t
13716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13717 vcltq_s64 (int64x2_t __a, int64x2_t __b)
13718 {
13719   return (uint64x2_t) (__a < __b);
13720 }
13721
13722 __extension__ extern __inline uint8x16_t
13723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13724 vcltq_u8 (uint8x16_t __a, uint8x16_t __b)
13725 {
13726   return (__a < __b);
13727 }
13728
13729 __extension__ extern __inline uint16x8_t
13730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13731 vcltq_u16 (uint16x8_t __a, uint16x8_t __b)
13732 {
13733   return (__a < __b);
13734 }
13735
13736 __extension__ extern __inline uint32x4_t
13737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13738 vcltq_u32 (uint32x4_t __a, uint32x4_t __b)
13739 {
13740   return (__a < __b);
13741 }
13742
13743 __extension__ extern __inline uint64x2_t
13744 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13745 vcltq_u64 (uint64x2_t __a, uint64x2_t __b)
13746 {
13747   return (__a < __b);
13748 }
13749
13750 /* vclt - scalar.  */
13751
13752 __extension__ extern __inline uint32_t
13753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13754 vclts_f32 (float32_t __a, float32_t __b)
13755 {
13756   return __a < __b ? -1 : 0;
13757 }
13758
13759 __extension__ extern __inline uint64_t
13760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13761 vcltd_s64 (int64_t __a, int64_t __b)
13762 {
13763   return __a < __b ? -1ll : 0ll;
13764 }
13765
13766 __extension__ extern __inline uint64_t
13767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13768 vcltd_u64 (uint64_t __a, uint64_t __b)
13769 {
13770   return __a < __b ? -1ll : 0ll;
13771 }
13772
13773 __extension__ extern __inline uint64_t
13774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13775 vcltd_f64 (float64_t __a, float64_t __b)
13776 {
13777   return __a < __b ? -1ll : 0ll;
13778 }
13779
13780 /* vcltz - vector.  */
13781
13782 __extension__ extern __inline uint32x2_t
13783 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13784 vcltz_f32 (float32x2_t __a)
13785 {
13786   return (uint32x2_t) (__a < 0.0f);
13787 }
13788
13789 __extension__ extern __inline uint64x1_t
13790 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13791 vcltz_f64 (float64x1_t __a)
13792 {
13793   return (uint64x1_t) (__a < (float64x1_t) {0.0});
13794 }
13795
13796 __extension__ extern __inline uint8x8_t
13797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13798 vcltz_s8 (int8x8_t __a)
13799 {
13800   return (uint8x8_t) (__a < 0);
13801 }
13802
13803 __extension__ extern __inline uint16x4_t
13804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13805 vcltz_s16 (int16x4_t __a)
13806 {
13807   return (uint16x4_t) (__a < 0);
13808 }
13809
13810 __extension__ extern __inline uint32x2_t
13811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13812 vcltz_s32 (int32x2_t __a)
13813 {
13814   return (uint32x2_t) (__a < 0);
13815 }
13816
13817 __extension__ extern __inline uint64x1_t
13818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13819 vcltz_s64 (int64x1_t __a)
13820 {
13821   return (uint64x1_t) (__a < __AARCH64_INT64_C (0));
13822 }
13823
13824 __extension__ extern __inline uint32x4_t
13825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13826 vcltzq_f32 (float32x4_t __a)
13827 {
13828   return (uint32x4_t) (__a < 0.0f);
13829 }
13830
13831 __extension__ extern __inline uint64x2_t
13832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13833 vcltzq_f64 (float64x2_t __a)
13834 {
13835   return (uint64x2_t) (__a < 0.0);
13836 }
13837
13838 __extension__ extern __inline uint8x16_t
13839 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13840 vcltzq_s8 (int8x16_t __a)
13841 {
13842   return (uint8x16_t) (__a < 0);
13843 }
13844
13845 __extension__ extern __inline uint16x8_t
13846 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13847 vcltzq_s16 (int16x8_t __a)
13848 {
13849   return (uint16x8_t) (__a < 0);
13850 }
13851
13852 __extension__ extern __inline uint32x4_t
13853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13854 vcltzq_s32 (int32x4_t __a)
13855 {
13856   return (uint32x4_t) (__a < 0);
13857 }
13858
13859 __extension__ extern __inline uint64x2_t
13860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13861 vcltzq_s64 (int64x2_t __a)
13862 {
13863   return (uint64x2_t) (__a < __AARCH64_INT64_C (0));
13864 }
13865
13866 /* vcltz - scalar.  */
13867
13868 __extension__ extern __inline uint32_t
13869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13870 vcltzs_f32 (float32_t __a)
13871 {
13872   return __a < 0.0f ? -1 : 0;
13873 }
13874
13875 __extension__ extern __inline uint64_t
13876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13877 vcltzd_s64 (int64_t __a)
13878 {
13879   return __a < 0 ? -1ll : 0ll;
13880 }
13881
13882 __extension__ extern __inline uint64_t
13883 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13884 vcltzd_f64 (float64_t __a)
13885 {
13886   return __a < 0.0 ? -1ll : 0ll;
13887 }
13888
13889 /* vcls.  */
13890
13891 __extension__ extern __inline int8x8_t
13892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13893 vcls_s8 (int8x8_t __a)
13894 {
13895   return __builtin_aarch64_clrsbv8qi (__a);
13896 }
13897
13898 __extension__ extern __inline int16x4_t
13899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13900 vcls_s16 (int16x4_t __a)
13901 {
13902   return __builtin_aarch64_clrsbv4hi (__a);
13903 }
13904
13905 __extension__ extern __inline int32x2_t
13906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13907 vcls_s32 (int32x2_t __a)
13908 {
13909   return __builtin_aarch64_clrsbv2si (__a);
13910 }
13911
13912 __extension__ extern __inline int8x16_t
13913 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13914 vclsq_s8 (int8x16_t __a)
13915 {
13916   return __builtin_aarch64_clrsbv16qi (__a);
13917 }
13918
13919 __extension__ extern __inline int16x8_t
13920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13921 vclsq_s16 (int16x8_t __a)
13922 {
13923   return __builtin_aarch64_clrsbv8hi (__a);
13924 }
13925
13926 __extension__ extern __inline int32x4_t
13927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13928 vclsq_s32 (int32x4_t __a)
13929 {
13930   return __builtin_aarch64_clrsbv4si (__a);
13931 }
13932
13933 /* vclz.  */
13934
13935 __extension__ extern __inline int8x8_t
13936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13937 vclz_s8 (int8x8_t __a)
13938 {
13939   return __builtin_aarch64_clzv8qi (__a);
13940 }
13941
13942 __extension__ extern __inline int16x4_t
13943 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13944 vclz_s16 (int16x4_t __a)
13945 {
13946   return __builtin_aarch64_clzv4hi (__a);
13947 }
13948
13949 __extension__ extern __inline int32x2_t
13950 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13951 vclz_s32 (int32x2_t __a)
13952 {
13953   return __builtin_aarch64_clzv2si (__a);
13954 }
13955
13956 __extension__ extern __inline uint8x8_t
13957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13958 vclz_u8 (uint8x8_t __a)
13959 {
13960   return (uint8x8_t)__builtin_aarch64_clzv8qi ((int8x8_t)__a);
13961 }
13962
13963 __extension__ extern __inline uint16x4_t
13964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13965 vclz_u16 (uint16x4_t __a)
13966 {
13967   return (uint16x4_t)__builtin_aarch64_clzv4hi ((int16x4_t)__a);
13968 }
13969
13970 __extension__ extern __inline uint32x2_t
13971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13972 vclz_u32 (uint32x2_t __a)
13973 {
13974   return (uint32x2_t)__builtin_aarch64_clzv2si ((int32x2_t)__a);
13975 }
13976
13977 __extension__ extern __inline int8x16_t
13978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13979 vclzq_s8 (int8x16_t __a)
13980 {
13981   return __builtin_aarch64_clzv16qi (__a);
13982 }
13983
13984 __extension__ extern __inline int16x8_t
13985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13986 vclzq_s16 (int16x8_t __a)
13987 {
13988   return __builtin_aarch64_clzv8hi (__a);
13989 }
13990
13991 __extension__ extern __inline int32x4_t
13992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
13993 vclzq_s32 (int32x4_t __a)
13994 {
13995   return __builtin_aarch64_clzv4si (__a);
13996 }
13997
13998 __extension__ extern __inline uint8x16_t
13999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14000 vclzq_u8 (uint8x16_t __a)
14001 {
14002   return (uint8x16_t)__builtin_aarch64_clzv16qi ((int8x16_t)__a);
14003 }
14004
14005 __extension__ extern __inline uint16x8_t
14006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14007 vclzq_u16 (uint16x8_t __a)
14008 {
14009   return (uint16x8_t)__builtin_aarch64_clzv8hi ((int16x8_t)__a);
14010 }
14011
14012 __extension__ extern __inline uint32x4_t
14013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14014 vclzq_u32 (uint32x4_t __a)
14015 {
14016   return (uint32x4_t)__builtin_aarch64_clzv4si ((int32x4_t)__a);
14017 }
14018
14019 /* vcnt.  */
14020
14021 __extension__ extern __inline poly8x8_t
14022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14023 vcnt_p8 (poly8x8_t __a)
14024 {
14025   return (poly8x8_t) __builtin_aarch64_popcountv8qi ((int8x8_t) __a);
14026 }
14027
14028 __extension__ extern __inline int8x8_t
14029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14030 vcnt_s8 (int8x8_t __a)
14031 {
14032   return __builtin_aarch64_popcountv8qi (__a);
14033 }
14034
14035 __extension__ extern __inline uint8x8_t
14036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14037 vcnt_u8 (uint8x8_t __a)
14038 {
14039   return (uint8x8_t) __builtin_aarch64_popcountv8qi ((int8x8_t) __a);
14040 }
14041
14042 __extension__ extern __inline poly8x16_t
14043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14044 vcntq_p8 (poly8x16_t __a)
14045 {
14046   return (poly8x16_t) __builtin_aarch64_popcountv16qi ((int8x16_t) __a);
14047 }
14048
14049 __extension__ extern __inline int8x16_t
14050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14051 vcntq_s8 (int8x16_t __a)
14052 {
14053   return __builtin_aarch64_popcountv16qi (__a);
14054 }
14055
14056 __extension__ extern __inline uint8x16_t
14057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14058 vcntq_u8 (uint8x16_t __a)
14059 {
14060   return (uint8x16_t) __builtin_aarch64_popcountv16qi ((int8x16_t) __a);
14061 }
14062
14063 /* vcopy_lane.  */
14064
14065 __extension__ extern __inline float32x2_t
14066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14067 vcopy_lane_f32 (float32x2_t __a, const int __lane1,
14068                 float32x2_t __b, const int __lane2)
14069 {
14070   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14071                                   __a, __lane1);
14072 }
14073
14074 __extension__ extern __inline float64x1_t
14075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14076 vcopy_lane_f64 (float64x1_t __a, const int __lane1,
14077                 float64x1_t __b, const int __lane2)
14078 {
14079   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14080                                   __a, __lane1);
14081 }
14082
14083 __extension__ extern __inline poly8x8_t
14084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14085 vcopy_lane_p8 (poly8x8_t __a, const int __lane1,
14086                poly8x8_t __b, const int __lane2)
14087 {
14088   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14089                                  __a, __lane1);
14090 }
14091
14092 __extension__ extern __inline poly16x4_t
14093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14094 vcopy_lane_p16 (poly16x4_t __a, const int __lane1,
14095                 poly16x4_t __b, const int __lane2)
14096 {
14097   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14098                                   __a, __lane1);
14099 }
14100
14101 __extension__ extern __inline poly64x1_t
14102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14103 vcopy_lane_p64 (poly64x1_t __a, const int __lane1,
14104                 poly64x1_t __b, const int __lane2)
14105 {
14106   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14107                                   __a, __lane1);
14108 }
14109
14110 __extension__ extern __inline int8x8_t
14111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14112 vcopy_lane_s8 (int8x8_t __a, const int __lane1,
14113                int8x8_t __b, const int __lane2)
14114 {
14115   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14116                                  __a, __lane1);
14117 }
14118
14119 __extension__ extern __inline int16x4_t
14120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14121 vcopy_lane_s16 (int16x4_t __a, const int __lane1,
14122                 int16x4_t __b, const int __lane2)
14123 {
14124   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14125                                   __a, __lane1);
14126 }
14127
14128 __extension__ extern __inline int32x2_t
14129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14130 vcopy_lane_s32 (int32x2_t __a, const int __lane1,
14131                 int32x2_t __b, const int __lane2)
14132 {
14133   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14134                                   __a, __lane1);
14135 }
14136
14137 __extension__ extern __inline int64x1_t
14138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14139 vcopy_lane_s64 (int64x1_t __a, const int __lane1,
14140                 int64x1_t __b, const int __lane2)
14141 {
14142   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14143                                   __a, __lane1);
14144 }
14145
14146 __extension__ extern __inline uint8x8_t
14147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14148 vcopy_lane_u8 (uint8x8_t __a, const int __lane1,
14149                uint8x8_t __b, const int __lane2)
14150 {
14151   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14152                                  __a, __lane1);
14153 }
14154
14155 __extension__ extern __inline uint16x4_t
14156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14157 vcopy_lane_u16 (uint16x4_t __a, const int __lane1,
14158                 uint16x4_t __b, const int __lane2)
14159 {
14160   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14161                                   __a, __lane1);
14162 }
14163
14164 __extension__ extern __inline uint32x2_t
14165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14166 vcopy_lane_u32 (uint32x2_t __a, const int __lane1,
14167                 uint32x2_t __b, const int __lane2)
14168 {
14169   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14170                                   __a, __lane1);
14171 }
14172
14173 __extension__ extern __inline uint64x1_t
14174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14175 vcopy_lane_u64 (uint64x1_t __a, const int __lane1,
14176                 uint64x1_t __b, const int __lane2)
14177 {
14178   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14179                                   __a, __lane1);
14180 }
14181
14182 /* vcopy_laneq.  */
14183
14184 __extension__ extern __inline float32x2_t
14185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14186 vcopy_laneq_f32 (float32x2_t __a, const int __lane1,
14187                  float32x4_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 float64x1_t
14194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14195 vcopy_laneq_f64 (float64x1_t __a, const int __lane1,
14196                  float64x2_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 poly8x8_t
14203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14204 vcopy_laneq_p8 (poly8x8_t __a, const int __lane1,
14205                 poly8x16_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 poly16x4_t
14212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14213 vcopy_laneq_p16 (poly16x4_t __a, const int __lane1,
14214                  poly16x8_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 poly64x1_t
14221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14222 vcopy_laneq_p64 (poly64x1_t __a, const int __lane1,
14223                  poly64x2_t __b, const int __lane2)
14224 {
14225   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14226                                   __a, __lane1);
14227 }
14228
14229 __extension__ extern __inline int8x8_t
14230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14231 vcopy_laneq_s8 (int8x8_t __a, const int __lane1,
14232                 int8x16_t __b, const int __lane2)
14233 {
14234   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14235                                  __a, __lane1);
14236 }
14237
14238 __extension__ extern __inline int16x4_t
14239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14240 vcopy_laneq_s16 (int16x4_t __a, const int __lane1,
14241                  int16x8_t __b, const int __lane2)
14242 {
14243   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14244                                   __a, __lane1);
14245 }
14246
14247 __extension__ extern __inline int32x2_t
14248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14249 vcopy_laneq_s32 (int32x2_t __a, const int __lane1,
14250                  int32x4_t __b, const int __lane2)
14251 {
14252   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14253                                   __a, __lane1);
14254 }
14255
14256 __extension__ extern __inline int64x1_t
14257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14258 vcopy_laneq_s64 (int64x1_t __a, const int __lane1,
14259                  int64x2_t __b, const int __lane2)
14260 {
14261   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14262                                   __a, __lane1);
14263 }
14264
14265 __extension__ extern __inline uint8x8_t
14266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14267 vcopy_laneq_u8 (uint8x8_t __a, const int __lane1,
14268                 uint8x16_t __b, const int __lane2)
14269 {
14270   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14271                                  __a, __lane1);
14272 }
14273
14274 __extension__ extern __inline uint16x4_t
14275 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14276 vcopy_laneq_u16 (uint16x4_t __a, const int __lane1,
14277                  uint16x8_t __b, const int __lane2)
14278 {
14279   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14280                                   __a, __lane1);
14281 }
14282
14283 __extension__ extern __inline uint32x2_t
14284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14285 vcopy_laneq_u32 (uint32x2_t __a, const int __lane1,
14286                  uint32x4_t __b, const int __lane2)
14287 {
14288   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14289                                   __a, __lane1);
14290 }
14291
14292 __extension__ extern __inline uint64x1_t
14293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14294 vcopy_laneq_u64 (uint64x1_t __a, const int __lane1,
14295                  uint64x2_t __b, const int __lane2)
14296 {
14297   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14298                                   __a, __lane1);
14299 }
14300
14301 /* vcopyq_lane.  */
14302
14303 __extension__ extern __inline float32x4_t
14304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14305 vcopyq_lane_f32 (float32x4_t __a, const int __lane1,
14306                  float32x2_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 float64x2_t
14313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14314 vcopyq_lane_f64 (float64x2_t __a, const int __lane1,
14315                  float64x1_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 poly8x16_t
14322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14323 vcopyq_lane_p8 (poly8x16_t __a, const int __lane1,
14324                 poly8x8_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 poly16x8_t
14331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14332 vcopyq_lane_p16 (poly16x8_t __a, const int __lane1,
14333                  poly16x4_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 poly64x2_t
14340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14341 vcopyq_lane_p64 (poly64x2_t __a, const int __lane1,
14342                  poly64x1_t __b, const int __lane2)
14343 {
14344   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14345                                    __a, __lane1);
14346 }
14347
14348 __extension__ extern __inline int8x16_t
14349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14350 vcopyq_lane_s8 (int8x16_t __a, const int __lane1,
14351                 int8x8_t __b, const int __lane2)
14352 {
14353   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14354                                   __a, __lane1);
14355 }
14356
14357 __extension__ extern __inline int16x8_t
14358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14359 vcopyq_lane_s16 (int16x8_t __a, const int __lane1,
14360                  int16x4_t __b, const int __lane2)
14361 {
14362   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14363                                    __a, __lane1);
14364 }
14365
14366 __extension__ extern __inline int32x4_t
14367 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14368 vcopyq_lane_s32 (int32x4_t __a, const int __lane1,
14369                  int32x2_t __b, const int __lane2)
14370 {
14371   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14372                                    __a, __lane1);
14373 }
14374
14375 __extension__ extern __inline int64x2_t
14376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14377 vcopyq_lane_s64 (int64x2_t __a, const int __lane1,
14378                  int64x1_t __b, const int __lane2)
14379 {
14380   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14381                                    __a, __lane1);
14382 }
14383
14384 __extension__ extern __inline uint8x16_t
14385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14386 vcopyq_lane_u8 (uint8x16_t __a, const int __lane1,
14387                 uint8x8_t __b, const int __lane2)
14388 {
14389   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14390                                   __a, __lane1);
14391 }
14392
14393 __extension__ extern __inline uint16x8_t
14394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14395 vcopyq_lane_u16 (uint16x8_t __a, const int __lane1,
14396                  uint16x4_t __b, const int __lane2)
14397 {
14398   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14399                                    __a, __lane1);
14400 }
14401
14402 __extension__ extern __inline uint32x4_t
14403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14404 vcopyq_lane_u32 (uint32x4_t __a, const int __lane1,
14405                  uint32x2_t __b, const int __lane2)
14406 {
14407   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14408                                    __a, __lane1);
14409 }
14410
14411 __extension__ extern __inline uint64x2_t
14412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14413 vcopyq_lane_u64 (uint64x2_t __a, const int __lane1,
14414                  uint64x1_t __b, const int __lane2)
14415 {
14416   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14417                                    __a, __lane1);
14418 }
14419
14420 /* vcopyq_laneq.  */
14421
14422 __extension__ extern __inline float32x4_t
14423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14424 vcopyq_laneq_f32 (float32x4_t __a, const int __lane1,
14425                   float32x4_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 float64x2_t
14432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14433 vcopyq_laneq_f64 (float64x2_t __a, const int __lane1,
14434                   float64x2_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 poly8x16_t
14441 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14442 vcopyq_laneq_p8 (poly8x16_t __a, const int __lane1,
14443                  poly8x16_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 poly16x8_t
14450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14451 vcopyq_laneq_p16 (poly16x8_t __a, const int __lane1,
14452                   poly16x8_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 poly64x2_t
14459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14460 vcopyq_laneq_p64 (poly64x2_t __a, const int __lane1,
14461                   poly64x2_t __b, const int __lane2)
14462 {
14463   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14464                                    __a, __lane1);
14465 }
14466
14467 __extension__ extern __inline int8x16_t
14468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14469 vcopyq_laneq_s8 (int8x16_t __a, const int __lane1,
14470                  int8x16_t __b, const int __lane2)
14471 {
14472   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14473                                   __a, __lane1);
14474 }
14475
14476 __extension__ extern __inline int16x8_t
14477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14478 vcopyq_laneq_s16 (int16x8_t __a, const int __lane1,
14479                   int16x8_t __b, const int __lane2)
14480 {
14481   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14482                                    __a, __lane1);
14483 }
14484
14485 __extension__ extern __inline int32x4_t
14486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14487 vcopyq_laneq_s32 (int32x4_t __a, const int __lane1,
14488                   int32x4_t __b, const int __lane2)
14489 {
14490   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14491                                    __a, __lane1);
14492 }
14493
14494 __extension__ extern __inline int64x2_t
14495 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14496 vcopyq_laneq_s64 (int64x2_t __a, const int __lane1,
14497                   int64x2_t __b, const int __lane2)
14498 {
14499   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14500                                    __a, __lane1);
14501 }
14502
14503 __extension__ extern __inline uint8x16_t
14504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14505 vcopyq_laneq_u8 (uint8x16_t __a, const int __lane1,
14506                  uint8x16_t __b, const int __lane2)
14507 {
14508   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14509                                   __a, __lane1);
14510 }
14511
14512 __extension__ extern __inline uint16x8_t
14513 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14514 vcopyq_laneq_u16 (uint16x8_t __a, const int __lane1,
14515                   uint16x8_t __b, const int __lane2)
14516 {
14517   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14518                                    __a, __lane1);
14519 }
14520
14521 __extension__ extern __inline uint32x4_t
14522 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14523 vcopyq_laneq_u32 (uint32x4_t __a, const int __lane1,
14524                   uint32x4_t __b, const int __lane2)
14525 {
14526   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14527                                    __a, __lane1);
14528 }
14529
14530 __extension__ extern __inline uint64x2_t
14531 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14532 vcopyq_laneq_u64 (uint64x2_t __a, const int __lane1,
14533                   uint64x2_t __b, const int __lane2)
14534 {
14535   return __aarch64_vset_lane_any (__aarch64_vget_lane_any (__b, __lane2),
14536                                    __a, __lane1);
14537 }
14538
14539 /* vcvt (double -> float).  */
14540
14541 __extension__ extern __inline float16x4_t
14542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14543 vcvt_f16_f32 (float32x4_t __a)
14544 {
14545   return __builtin_aarch64_float_truncate_lo_v4hf (__a);
14546 }
14547
14548 __extension__ extern __inline float16x8_t
14549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14550 vcvt_high_f16_f32 (float16x4_t __a, float32x4_t __b)
14551 {
14552   return __builtin_aarch64_float_truncate_hi_v8hf (__a, __b);
14553 }
14554
14555 __extension__ extern __inline float32x2_t
14556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14557 vcvt_f32_f64 (float64x2_t __a)
14558 {
14559   return __builtin_aarch64_float_truncate_lo_v2sf (__a);
14560 }
14561
14562 __extension__ extern __inline float32x4_t
14563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14564 vcvt_high_f32_f64 (float32x2_t __a, float64x2_t __b)
14565 {
14566   return __builtin_aarch64_float_truncate_hi_v4sf (__a, __b);
14567 }
14568
14569 /* vcvt (float -> double).  */
14570
14571 __extension__ extern __inline float32x4_t
14572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14573 vcvt_f32_f16 (float16x4_t __a)
14574 {
14575   return __builtin_aarch64_float_extend_lo_v4sf (__a);
14576 }
14577
14578 __extension__ extern __inline float64x2_t
14579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14580 vcvt_f64_f32 (float32x2_t __a)
14581 {
14582
14583   return __builtin_aarch64_float_extend_lo_v2df (__a);
14584 }
14585
14586 __extension__ extern __inline float32x4_t
14587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14588 vcvt_high_f32_f16 (float16x8_t __a)
14589 {
14590   return __builtin_aarch64_vec_unpacks_hi_v8hf (__a);
14591 }
14592
14593 __extension__ extern __inline float64x2_t
14594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14595 vcvt_high_f64_f32 (float32x4_t __a)
14596 {
14597   return __builtin_aarch64_vec_unpacks_hi_v4sf (__a);
14598 }
14599
14600 /* vcvt (<u>fixed-point -> float).  */
14601
14602 __extension__ extern __inline float64_t
14603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14604 vcvtd_n_f64_s64 (int64_t __a, const int __b)
14605 {
14606   return __builtin_aarch64_scvtfdi (__a, __b);
14607 }
14608
14609 __extension__ extern __inline float64_t
14610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14611 vcvtd_n_f64_u64 (uint64_t __a, const int __b)
14612 {
14613   return __builtin_aarch64_ucvtfdi_sus (__a, __b);
14614 }
14615
14616 __extension__ extern __inline float32_t
14617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14618 vcvts_n_f32_s32 (int32_t __a, const int __b)
14619 {
14620   return __builtin_aarch64_scvtfsi (__a, __b);
14621 }
14622
14623 __extension__ extern __inline float32_t
14624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14625 vcvts_n_f32_u32 (uint32_t __a, const int __b)
14626 {
14627   return __builtin_aarch64_ucvtfsi_sus (__a, __b);
14628 }
14629
14630 __extension__ extern __inline float32x2_t
14631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14632 vcvt_n_f32_s32 (int32x2_t __a, const int __b)
14633 {
14634   return __builtin_aarch64_scvtfv2si (__a, __b);
14635 }
14636
14637 __extension__ extern __inline float32x2_t
14638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14639 vcvt_n_f32_u32 (uint32x2_t __a, const int __b)
14640 {
14641   return __builtin_aarch64_ucvtfv2si_sus (__a, __b);
14642 }
14643
14644 __extension__ extern __inline float64x1_t
14645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14646 vcvt_n_f64_s64 (int64x1_t __a, const int __b)
14647 {
14648   return (float64x1_t)
14649     { __builtin_aarch64_scvtfdi (vget_lane_s64 (__a, 0), __b) };
14650 }
14651
14652 __extension__ extern __inline float64x1_t
14653 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14654 vcvt_n_f64_u64 (uint64x1_t __a, const int __b)
14655 {
14656   return (float64x1_t)
14657     { __builtin_aarch64_ucvtfdi_sus (vget_lane_u64 (__a, 0), __b) };
14658 }
14659
14660 __extension__ extern __inline float32x4_t
14661 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14662 vcvtq_n_f32_s32 (int32x4_t __a, const int __b)
14663 {
14664   return __builtin_aarch64_scvtfv4si (__a, __b);
14665 }
14666
14667 __extension__ extern __inline float32x4_t
14668 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14669 vcvtq_n_f32_u32 (uint32x4_t __a, const int __b)
14670 {
14671   return __builtin_aarch64_ucvtfv4si_sus (__a, __b);
14672 }
14673
14674 __extension__ extern __inline float64x2_t
14675 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14676 vcvtq_n_f64_s64 (int64x2_t __a, const int __b)
14677 {
14678   return __builtin_aarch64_scvtfv2di (__a, __b);
14679 }
14680
14681 __extension__ extern __inline float64x2_t
14682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14683 vcvtq_n_f64_u64 (uint64x2_t __a, const int __b)
14684 {
14685   return __builtin_aarch64_ucvtfv2di_sus (__a, __b);
14686 }
14687
14688 /* vcvt (float -> <u>fixed-point).  */
14689
14690 __extension__ extern __inline int64_t
14691 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14692 vcvtd_n_s64_f64 (float64_t __a, const int __b)
14693 {
14694   return __builtin_aarch64_fcvtzsdf (__a, __b);
14695 }
14696
14697 __extension__ extern __inline uint64_t
14698 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14699 vcvtd_n_u64_f64 (float64_t __a, const int __b)
14700 {
14701   return __builtin_aarch64_fcvtzudf_uss (__a, __b);
14702 }
14703
14704 __extension__ extern __inline int32_t
14705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14706 vcvts_n_s32_f32 (float32_t __a, const int __b)
14707 {
14708   return __builtin_aarch64_fcvtzssf (__a, __b);
14709 }
14710
14711 __extension__ extern __inline uint32_t
14712 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14713 vcvts_n_u32_f32 (float32_t __a, const int __b)
14714 {
14715   return __builtin_aarch64_fcvtzusf_uss (__a, __b);
14716 }
14717
14718 __extension__ extern __inline int32x2_t
14719 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14720 vcvt_n_s32_f32 (float32x2_t __a, const int __b)
14721 {
14722   return __builtin_aarch64_fcvtzsv2sf (__a, __b);
14723 }
14724
14725 __extension__ extern __inline uint32x2_t
14726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14727 vcvt_n_u32_f32 (float32x2_t __a, const int __b)
14728 {
14729   return __builtin_aarch64_fcvtzuv2sf_uss (__a, __b);
14730 }
14731
14732 __extension__ extern __inline int64x1_t
14733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14734 vcvt_n_s64_f64 (float64x1_t __a, const int __b)
14735 {
14736   return (int64x1_t)
14737     { __builtin_aarch64_fcvtzsdf (vget_lane_f64 (__a, 0), __b) };
14738 }
14739
14740 __extension__ extern __inline uint64x1_t
14741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14742 vcvt_n_u64_f64 (float64x1_t __a, const int __b)
14743 {
14744   return (uint64x1_t)
14745     { __builtin_aarch64_fcvtzudf_uss (vget_lane_f64 (__a, 0), __b) };
14746 }
14747
14748 __extension__ extern __inline int32x4_t
14749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14750 vcvtq_n_s32_f32 (float32x4_t __a, const int __b)
14751 {
14752   return __builtin_aarch64_fcvtzsv4sf (__a, __b);
14753 }
14754
14755 __extension__ extern __inline uint32x4_t
14756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14757 vcvtq_n_u32_f32 (float32x4_t __a, const int __b)
14758 {
14759   return __builtin_aarch64_fcvtzuv4sf_uss (__a, __b);
14760 }
14761
14762 __extension__ extern __inline int64x2_t
14763 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14764 vcvtq_n_s64_f64 (float64x2_t __a, const int __b)
14765 {
14766   return __builtin_aarch64_fcvtzsv2df (__a, __b);
14767 }
14768
14769 __extension__ extern __inline uint64x2_t
14770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14771 vcvtq_n_u64_f64 (float64x2_t __a, const int __b)
14772 {
14773   return __builtin_aarch64_fcvtzuv2df_uss (__a, __b);
14774 }
14775
14776 /* vcvt  (<u>int -> float)  */
14777
14778 __extension__ extern __inline float64_t
14779 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14780 vcvtd_f64_s64 (int64_t __a)
14781 {
14782   return (float64_t) __a;
14783 }
14784
14785 __extension__ extern __inline float64_t
14786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14787 vcvtd_f64_u64 (uint64_t __a)
14788 {
14789   return (float64_t) __a;
14790 }
14791
14792 __extension__ extern __inline float32_t
14793 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14794 vcvts_f32_s32 (int32_t __a)
14795 {
14796   return (float32_t) __a;
14797 }
14798
14799 __extension__ extern __inline float32_t
14800 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14801 vcvts_f32_u32 (uint32_t __a)
14802 {
14803   return (float32_t) __a;
14804 }
14805
14806 __extension__ extern __inline float32x2_t
14807 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14808 vcvt_f32_s32 (int32x2_t __a)
14809 {
14810   return __builtin_aarch64_floatv2siv2sf (__a);
14811 }
14812
14813 __extension__ extern __inline float32x2_t
14814 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14815 vcvt_f32_u32 (uint32x2_t __a)
14816 {
14817   return __builtin_aarch64_floatunsv2siv2sf ((int32x2_t) __a);
14818 }
14819
14820 __extension__ extern __inline float64x1_t
14821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14822 vcvt_f64_s64 (int64x1_t __a)
14823 {
14824   return (float64x1_t) { vget_lane_s64 (__a, 0) };
14825 }
14826
14827 __extension__ extern __inline float64x1_t
14828 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14829 vcvt_f64_u64 (uint64x1_t __a)
14830 {
14831   return (float64x1_t) { vget_lane_u64 (__a, 0) };
14832 }
14833
14834 __extension__ extern __inline float32x4_t
14835 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14836 vcvtq_f32_s32 (int32x4_t __a)
14837 {
14838   return __builtin_aarch64_floatv4siv4sf (__a);
14839 }
14840
14841 __extension__ extern __inline float32x4_t
14842 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14843 vcvtq_f32_u32 (uint32x4_t __a)
14844 {
14845   return __builtin_aarch64_floatunsv4siv4sf ((int32x4_t) __a);
14846 }
14847
14848 __extension__ extern __inline float64x2_t
14849 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14850 vcvtq_f64_s64 (int64x2_t __a)
14851 {
14852   return __builtin_aarch64_floatv2div2df (__a);
14853 }
14854
14855 __extension__ extern __inline float64x2_t
14856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14857 vcvtq_f64_u64 (uint64x2_t __a)
14858 {
14859   return __builtin_aarch64_floatunsv2div2df ((int64x2_t) __a);
14860 }
14861
14862 /* vcvt (float -> <u>int)  */
14863
14864 __extension__ extern __inline int64_t
14865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14866 vcvtd_s64_f64 (float64_t __a)
14867 {
14868   return (int64_t) __a;
14869 }
14870
14871 __extension__ extern __inline uint64_t
14872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14873 vcvtd_u64_f64 (float64_t __a)
14874 {
14875   return (uint64_t) __a;
14876 }
14877
14878 __extension__ extern __inline int32_t
14879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14880 vcvts_s32_f32 (float32_t __a)
14881 {
14882   return (int32_t) __a;
14883 }
14884
14885 __extension__ extern __inline uint32_t
14886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14887 vcvts_u32_f32 (float32_t __a)
14888 {
14889   return (uint32_t) __a;
14890 }
14891
14892 __extension__ extern __inline int32x2_t
14893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14894 vcvt_s32_f32 (float32x2_t __a)
14895 {
14896   return __builtin_aarch64_lbtruncv2sfv2si (__a);
14897 }
14898
14899 __extension__ extern __inline uint32x2_t
14900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14901 vcvt_u32_f32 (float32x2_t __a)
14902 {
14903   return __builtin_aarch64_lbtruncuv2sfv2si_us (__a);
14904 }
14905
14906 __extension__ extern __inline int32x4_t
14907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14908 vcvtq_s32_f32 (float32x4_t __a)
14909 {
14910   return __builtin_aarch64_lbtruncv4sfv4si (__a);
14911 }
14912
14913 __extension__ extern __inline uint32x4_t
14914 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14915 vcvtq_u32_f32 (float32x4_t __a)
14916 {
14917   return __builtin_aarch64_lbtruncuv4sfv4si_us (__a);
14918 }
14919
14920 __extension__ extern __inline int64x1_t
14921 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14922 vcvt_s64_f64 (float64x1_t __a)
14923 {
14924   return (int64x1_t) {vcvtd_s64_f64 (__a[0])};
14925 }
14926
14927 __extension__ extern __inline uint64x1_t
14928 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14929 vcvt_u64_f64 (float64x1_t __a)
14930 {
14931   return (uint64x1_t) {vcvtd_u64_f64 (__a[0])};
14932 }
14933
14934 __extension__ extern __inline int64x2_t
14935 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14936 vcvtq_s64_f64 (float64x2_t __a)
14937 {
14938   return __builtin_aarch64_lbtruncv2dfv2di (__a);
14939 }
14940
14941 __extension__ extern __inline uint64x2_t
14942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14943 vcvtq_u64_f64 (float64x2_t __a)
14944 {
14945   return __builtin_aarch64_lbtruncuv2dfv2di_us (__a);
14946 }
14947
14948 /* vcvta  */
14949
14950 __extension__ extern __inline int64_t
14951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14952 vcvtad_s64_f64 (float64_t __a)
14953 {
14954   return __builtin_aarch64_lrounddfdi (__a);
14955 }
14956
14957 __extension__ extern __inline uint64_t
14958 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14959 vcvtad_u64_f64 (float64_t __a)
14960 {
14961   return __builtin_aarch64_lroundudfdi_us (__a);
14962 }
14963
14964 __extension__ extern __inline int32_t
14965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14966 vcvtas_s32_f32 (float32_t __a)
14967 {
14968   return __builtin_aarch64_lroundsfsi (__a);
14969 }
14970
14971 __extension__ extern __inline uint32_t
14972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14973 vcvtas_u32_f32 (float32_t __a)
14974 {
14975   return __builtin_aarch64_lroundusfsi_us (__a);
14976 }
14977
14978 __extension__ extern __inline int32x2_t
14979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14980 vcvta_s32_f32 (float32x2_t __a)
14981 {
14982   return __builtin_aarch64_lroundv2sfv2si (__a);
14983 }
14984
14985 __extension__ extern __inline uint32x2_t
14986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14987 vcvta_u32_f32 (float32x2_t __a)
14988 {
14989   return __builtin_aarch64_lrounduv2sfv2si_us (__a);
14990 }
14991
14992 __extension__ extern __inline int32x4_t
14993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
14994 vcvtaq_s32_f32 (float32x4_t __a)
14995 {
14996   return __builtin_aarch64_lroundv4sfv4si (__a);
14997 }
14998
14999 __extension__ extern __inline uint32x4_t
15000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15001 vcvtaq_u32_f32 (float32x4_t __a)
15002 {
15003   return __builtin_aarch64_lrounduv4sfv4si_us (__a);
15004 }
15005
15006 __extension__ extern __inline int64x1_t
15007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15008 vcvta_s64_f64 (float64x1_t __a)
15009 {
15010   return (int64x1_t) {vcvtad_s64_f64 (__a[0])};
15011 }
15012
15013 __extension__ extern __inline uint64x1_t
15014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15015 vcvta_u64_f64 (float64x1_t __a)
15016 {
15017   return (uint64x1_t) {vcvtad_u64_f64 (__a[0])};
15018 }
15019
15020 __extension__ extern __inline int64x2_t
15021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15022 vcvtaq_s64_f64 (float64x2_t __a)
15023 {
15024   return __builtin_aarch64_lroundv2dfv2di (__a);
15025 }
15026
15027 __extension__ extern __inline uint64x2_t
15028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15029 vcvtaq_u64_f64 (float64x2_t __a)
15030 {
15031   return __builtin_aarch64_lrounduv2dfv2di_us (__a);
15032 }
15033
15034 /* vcvtm  */
15035
15036 __extension__ extern __inline int64_t
15037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15038 vcvtmd_s64_f64 (float64_t __a)
15039 {
15040   return __builtin_llfloor (__a);
15041 }
15042
15043 __extension__ extern __inline uint64_t
15044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15045 vcvtmd_u64_f64 (float64_t __a)
15046 {
15047   return __builtin_aarch64_lfloorudfdi_us (__a);
15048 }
15049
15050 __extension__ extern __inline int32_t
15051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15052 vcvtms_s32_f32 (float32_t __a)
15053 {
15054   return __builtin_ifloorf (__a);
15055 }
15056
15057 __extension__ extern __inline uint32_t
15058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15059 vcvtms_u32_f32 (float32_t __a)
15060 {
15061   return __builtin_aarch64_lfloorusfsi_us (__a);
15062 }
15063
15064 __extension__ extern __inline int32x2_t
15065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15066 vcvtm_s32_f32 (float32x2_t __a)
15067 {
15068   return __builtin_aarch64_lfloorv2sfv2si (__a);
15069 }
15070
15071 __extension__ extern __inline uint32x2_t
15072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15073 vcvtm_u32_f32 (float32x2_t __a)
15074 {
15075   return __builtin_aarch64_lflooruv2sfv2si_us (__a);
15076 }
15077
15078 __extension__ extern __inline int32x4_t
15079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15080 vcvtmq_s32_f32 (float32x4_t __a)
15081 {
15082   return __builtin_aarch64_lfloorv4sfv4si (__a);
15083 }
15084
15085 __extension__ extern __inline uint32x4_t
15086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15087 vcvtmq_u32_f32 (float32x4_t __a)
15088 {
15089   return __builtin_aarch64_lflooruv4sfv4si_us (__a);
15090 }
15091
15092 __extension__ extern __inline int64x1_t
15093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15094 vcvtm_s64_f64 (float64x1_t __a)
15095 {
15096   return (int64x1_t) {vcvtmd_s64_f64 (__a[0])};
15097 }
15098
15099 __extension__ extern __inline uint64x1_t
15100 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15101 vcvtm_u64_f64 (float64x1_t __a)
15102 {
15103   return (uint64x1_t) {vcvtmd_u64_f64 (__a[0])};
15104 }
15105
15106 __extension__ extern __inline int64x2_t
15107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15108 vcvtmq_s64_f64 (float64x2_t __a)
15109 {
15110   return __builtin_aarch64_lfloorv2dfv2di (__a);
15111 }
15112
15113 __extension__ extern __inline uint64x2_t
15114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15115 vcvtmq_u64_f64 (float64x2_t __a)
15116 {
15117   return __builtin_aarch64_lflooruv2dfv2di_us (__a);
15118 }
15119
15120 /* vcvtn  */
15121
15122 __extension__ extern __inline int64_t
15123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15124 vcvtnd_s64_f64 (float64_t __a)
15125 {
15126   return __builtin_aarch64_lfrintndfdi (__a);
15127 }
15128
15129 __extension__ extern __inline uint64_t
15130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15131 vcvtnd_u64_f64 (float64_t __a)
15132 {
15133   return __builtin_aarch64_lfrintnudfdi_us (__a);
15134 }
15135
15136 __extension__ extern __inline int32_t
15137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15138 vcvtns_s32_f32 (float32_t __a)
15139 {
15140   return __builtin_aarch64_lfrintnsfsi (__a);
15141 }
15142
15143 __extension__ extern __inline uint32_t
15144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15145 vcvtns_u32_f32 (float32_t __a)
15146 {
15147   return __builtin_aarch64_lfrintnusfsi_us (__a);
15148 }
15149
15150 __extension__ extern __inline int32x2_t
15151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15152 vcvtn_s32_f32 (float32x2_t __a)
15153 {
15154   return __builtin_aarch64_lfrintnv2sfv2si (__a);
15155 }
15156
15157 __extension__ extern __inline uint32x2_t
15158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15159 vcvtn_u32_f32 (float32x2_t __a)
15160 {
15161   return __builtin_aarch64_lfrintnuv2sfv2si_us (__a);
15162 }
15163
15164 __extension__ extern __inline int32x4_t
15165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15166 vcvtnq_s32_f32 (float32x4_t __a)
15167 {
15168   return __builtin_aarch64_lfrintnv4sfv4si (__a);
15169 }
15170
15171 __extension__ extern __inline uint32x4_t
15172 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15173 vcvtnq_u32_f32 (float32x4_t __a)
15174 {
15175   return __builtin_aarch64_lfrintnuv4sfv4si_us (__a);
15176 }
15177
15178 __extension__ extern __inline int64x1_t
15179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15180 vcvtn_s64_f64 (float64x1_t __a)
15181 {
15182   return (int64x1_t) {vcvtnd_s64_f64 (__a[0])};
15183 }
15184
15185 __extension__ extern __inline uint64x1_t
15186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15187 vcvtn_u64_f64 (float64x1_t __a)
15188 {
15189   return (uint64x1_t) {vcvtnd_u64_f64 (__a[0])};
15190 }
15191
15192 __extension__ extern __inline int64x2_t
15193 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15194 vcvtnq_s64_f64 (float64x2_t __a)
15195 {
15196   return __builtin_aarch64_lfrintnv2dfv2di (__a);
15197 }
15198
15199 __extension__ extern __inline uint64x2_t
15200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15201 vcvtnq_u64_f64 (float64x2_t __a)
15202 {
15203   return __builtin_aarch64_lfrintnuv2dfv2di_us (__a);
15204 }
15205
15206 /* vcvtp  */
15207
15208 __extension__ extern __inline int64_t
15209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15210 vcvtpd_s64_f64 (float64_t __a)
15211 {
15212   return __builtin_llceil (__a);
15213 }
15214
15215 __extension__ extern __inline uint64_t
15216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15217 vcvtpd_u64_f64 (float64_t __a)
15218 {
15219   return __builtin_aarch64_lceiludfdi_us (__a);
15220 }
15221
15222 __extension__ extern __inline int32_t
15223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15224 vcvtps_s32_f32 (float32_t __a)
15225 {
15226   return __builtin_iceilf (__a);
15227 }
15228
15229 __extension__ extern __inline uint32_t
15230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15231 vcvtps_u32_f32 (float32_t __a)
15232 {
15233   return __builtin_aarch64_lceilusfsi_us (__a);
15234 }
15235
15236 __extension__ extern __inline int32x2_t
15237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15238 vcvtp_s32_f32 (float32x2_t __a)
15239 {
15240   return __builtin_aarch64_lceilv2sfv2si (__a);
15241 }
15242
15243 __extension__ extern __inline uint32x2_t
15244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15245 vcvtp_u32_f32 (float32x2_t __a)
15246 {
15247   return __builtin_aarch64_lceiluv2sfv2si_us (__a);
15248 }
15249
15250 __extension__ extern __inline int32x4_t
15251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15252 vcvtpq_s32_f32 (float32x4_t __a)
15253 {
15254   return __builtin_aarch64_lceilv4sfv4si (__a);
15255 }
15256
15257 __extension__ extern __inline uint32x4_t
15258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15259 vcvtpq_u32_f32 (float32x4_t __a)
15260 {
15261   return __builtin_aarch64_lceiluv4sfv4si_us (__a);
15262 }
15263
15264 __extension__ extern __inline int64x1_t
15265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15266 vcvtp_s64_f64 (float64x1_t __a)
15267 {
15268   return (int64x1_t) {vcvtpd_s64_f64 (__a[0])};
15269 }
15270
15271 __extension__ extern __inline uint64x1_t
15272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15273 vcvtp_u64_f64 (float64x1_t __a)
15274 {
15275   return (uint64x1_t) {vcvtpd_u64_f64 (__a[0])};
15276 }
15277
15278 __extension__ extern __inline int64x2_t
15279 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15280 vcvtpq_s64_f64 (float64x2_t __a)
15281 {
15282   return __builtin_aarch64_lceilv2dfv2di (__a);
15283 }
15284
15285 __extension__ extern __inline uint64x2_t
15286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15287 vcvtpq_u64_f64 (float64x2_t __a)
15288 {
15289   return __builtin_aarch64_lceiluv2dfv2di_us (__a);
15290 }
15291
15292 /* vdup_n  */
15293
15294 __extension__ extern __inline float16x4_t
15295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15296 vdup_n_f16 (float16_t __a)
15297 {
15298   return (float16x4_t) {__a, __a, __a, __a};
15299 }
15300
15301 __extension__ extern __inline float32x2_t
15302 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15303 vdup_n_f32 (float32_t __a)
15304 {
15305   return (float32x2_t) {__a, __a};
15306 }
15307
15308 __extension__ extern __inline float64x1_t
15309 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15310 vdup_n_f64 (float64_t __a)
15311 {
15312   return (float64x1_t) {__a};
15313 }
15314
15315 __extension__ extern __inline poly8x8_t
15316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15317 vdup_n_p8 (poly8_t __a)
15318 {
15319   return (poly8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15320 }
15321
15322 __extension__ extern __inline poly16x4_t
15323 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15324 vdup_n_p16 (poly16_t __a)
15325 {
15326   return (poly16x4_t) {__a, __a, __a, __a};
15327 }
15328
15329 __extension__ extern __inline poly64x1_t
15330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15331 vdup_n_p64 (poly64_t __a)
15332 {
15333   return (poly64x1_t) {__a};
15334 }
15335
15336 __extension__ extern __inline int8x8_t
15337 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15338 vdup_n_s8 (int8_t __a)
15339 {
15340   return (int8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15341 }
15342
15343 __extension__ extern __inline int16x4_t
15344 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15345 vdup_n_s16 (int16_t __a)
15346 {
15347   return (int16x4_t) {__a, __a, __a, __a};
15348 }
15349
15350 __extension__ extern __inline int32x2_t
15351 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15352 vdup_n_s32 (int32_t __a)
15353 {
15354   return (int32x2_t) {__a, __a};
15355 }
15356
15357 __extension__ extern __inline int64x1_t
15358 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15359 vdup_n_s64 (int64_t __a)
15360 {
15361   return (int64x1_t) {__a};
15362 }
15363
15364 __extension__ extern __inline uint8x8_t
15365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15366 vdup_n_u8 (uint8_t __a)
15367 {
15368   return (uint8x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15369 }
15370
15371 __extension__ extern __inline uint16x4_t
15372 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15373 vdup_n_u16 (uint16_t __a)
15374 {
15375   return (uint16x4_t) {__a, __a, __a, __a};
15376 }
15377
15378 __extension__ extern __inline uint32x2_t
15379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15380 vdup_n_u32 (uint32_t __a)
15381 {
15382   return (uint32x2_t) {__a, __a};
15383 }
15384
15385 __extension__ extern __inline uint64x1_t
15386 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15387 vdup_n_u64 (uint64_t __a)
15388 {
15389   return (uint64x1_t) {__a};
15390 }
15391
15392 /* vdupq_n  */
15393
15394 __extension__ extern __inline float16x8_t
15395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15396 vdupq_n_f16 (float16_t __a)
15397 {
15398   return (float16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15399 }
15400
15401 __extension__ extern __inline float32x4_t
15402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15403 vdupq_n_f32 (float32_t __a)
15404 {
15405   return (float32x4_t) {__a, __a, __a, __a};
15406 }
15407
15408 __extension__ extern __inline float64x2_t
15409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15410 vdupq_n_f64 (float64_t __a)
15411 {
15412   return (float64x2_t) {__a, __a};
15413 }
15414
15415 __extension__ extern __inline poly8x16_t
15416 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15417 vdupq_n_p8 (uint32_t __a)
15418 {
15419   return (poly8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15420                        __a, __a, __a, __a, __a, __a, __a, __a};
15421 }
15422
15423 __extension__ extern __inline poly16x8_t
15424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15425 vdupq_n_p16 (uint32_t __a)
15426 {
15427   return (poly16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15428 }
15429
15430 __extension__ extern __inline poly64x2_t
15431 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15432 vdupq_n_p64 (uint64_t __a)
15433 {
15434   return (poly64x2_t) {__a, __a};
15435 }
15436
15437 __extension__ extern __inline int8x16_t
15438 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15439 vdupq_n_s8 (int32_t __a)
15440 {
15441   return (int8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15442                       __a, __a, __a, __a, __a, __a, __a, __a};
15443 }
15444
15445 __extension__ extern __inline int16x8_t
15446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15447 vdupq_n_s16 (int32_t __a)
15448 {
15449   return (int16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15450 }
15451
15452 __extension__ extern __inline int32x4_t
15453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15454 vdupq_n_s32 (int32_t __a)
15455 {
15456   return (int32x4_t) {__a, __a, __a, __a};
15457 }
15458
15459 __extension__ extern __inline int64x2_t
15460 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15461 vdupq_n_s64 (int64_t __a)
15462 {
15463   return (int64x2_t) {__a, __a};
15464 }
15465
15466 __extension__ extern __inline uint8x16_t
15467 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15468 vdupq_n_u8 (uint32_t __a)
15469 {
15470   return (uint8x16_t) {__a, __a, __a, __a, __a, __a, __a, __a,
15471                        __a, __a, __a, __a, __a, __a, __a, __a};
15472 }
15473
15474 __extension__ extern __inline uint16x8_t
15475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15476 vdupq_n_u16 (uint32_t __a)
15477 {
15478   return (uint16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
15479 }
15480
15481 __extension__ extern __inline uint32x4_t
15482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15483 vdupq_n_u32 (uint32_t __a)
15484 {
15485   return (uint32x4_t) {__a, __a, __a, __a};
15486 }
15487
15488 __extension__ extern __inline uint64x2_t
15489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15490 vdupq_n_u64 (uint64_t __a)
15491 {
15492   return (uint64x2_t) {__a, __a};
15493 }
15494
15495 /* vdup_lane  */
15496
15497 __extension__ extern __inline float16x4_t
15498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15499 vdup_lane_f16 (float16x4_t __a, const int __b)
15500 {
15501   return __aarch64_vdup_lane_f16 (__a, __b);
15502 }
15503
15504 __extension__ extern __inline float32x2_t
15505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15506 vdup_lane_f32 (float32x2_t __a, const int __b)
15507 {
15508   return __aarch64_vdup_lane_f32 (__a, __b);
15509 }
15510
15511 __extension__ extern __inline float64x1_t
15512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15513 vdup_lane_f64 (float64x1_t __a, const int __b)
15514 {
15515   return __aarch64_vdup_lane_f64 (__a, __b);
15516 }
15517
15518 __extension__ extern __inline poly8x8_t
15519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15520 vdup_lane_p8 (poly8x8_t __a, const int __b)
15521 {
15522   return __aarch64_vdup_lane_p8 (__a, __b);
15523 }
15524
15525 __extension__ extern __inline poly16x4_t
15526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15527 vdup_lane_p16 (poly16x4_t __a, const int __b)
15528 {
15529   return __aarch64_vdup_lane_p16 (__a, __b);
15530 }
15531
15532 __extension__ extern __inline poly64x1_t
15533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15534 vdup_lane_p64 (poly64x1_t __a, const int __b)
15535 {
15536   return __aarch64_vdup_lane_p64 (__a, __b);
15537 }
15538
15539 __extension__ extern __inline int8x8_t
15540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15541 vdup_lane_s8 (int8x8_t __a, const int __b)
15542 {
15543   return __aarch64_vdup_lane_s8 (__a, __b);
15544 }
15545
15546 __extension__ extern __inline int16x4_t
15547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15548 vdup_lane_s16 (int16x4_t __a, const int __b)
15549 {
15550   return __aarch64_vdup_lane_s16 (__a, __b);
15551 }
15552
15553 __extension__ extern __inline int32x2_t
15554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15555 vdup_lane_s32 (int32x2_t __a, const int __b)
15556 {
15557   return __aarch64_vdup_lane_s32 (__a, __b);
15558 }
15559
15560 __extension__ extern __inline int64x1_t
15561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15562 vdup_lane_s64 (int64x1_t __a, const int __b)
15563 {
15564   return __aarch64_vdup_lane_s64 (__a, __b);
15565 }
15566
15567 __extension__ extern __inline uint8x8_t
15568 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15569 vdup_lane_u8 (uint8x8_t __a, const int __b)
15570 {
15571   return __aarch64_vdup_lane_u8 (__a, __b);
15572 }
15573
15574 __extension__ extern __inline uint16x4_t
15575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15576 vdup_lane_u16 (uint16x4_t __a, const int __b)
15577 {
15578   return __aarch64_vdup_lane_u16 (__a, __b);
15579 }
15580
15581 __extension__ extern __inline uint32x2_t
15582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15583 vdup_lane_u32 (uint32x2_t __a, const int __b)
15584 {
15585   return __aarch64_vdup_lane_u32 (__a, __b);
15586 }
15587
15588 __extension__ extern __inline uint64x1_t
15589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15590 vdup_lane_u64 (uint64x1_t __a, const int __b)
15591 {
15592   return __aarch64_vdup_lane_u64 (__a, __b);
15593 }
15594
15595 /* vdup_laneq  */
15596
15597 __extension__ extern __inline float16x4_t
15598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15599 vdup_laneq_f16 (float16x8_t __a, const int __b)
15600 {
15601   return __aarch64_vdup_laneq_f16 (__a, __b);
15602 }
15603
15604 __extension__ extern __inline float32x2_t
15605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15606 vdup_laneq_f32 (float32x4_t __a, const int __b)
15607 {
15608   return __aarch64_vdup_laneq_f32 (__a, __b);
15609 }
15610
15611 __extension__ extern __inline float64x1_t
15612 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15613 vdup_laneq_f64 (float64x2_t __a, const int __b)
15614 {
15615   return __aarch64_vdup_laneq_f64 (__a, __b);
15616 }
15617
15618 __extension__ extern __inline poly8x8_t
15619 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15620 vdup_laneq_p8 (poly8x16_t __a, const int __b)
15621 {
15622   return __aarch64_vdup_laneq_p8 (__a, __b);
15623 }
15624
15625 __extension__ extern __inline poly16x4_t
15626 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15627 vdup_laneq_p16 (poly16x8_t __a, const int __b)
15628 {
15629   return __aarch64_vdup_laneq_p16 (__a, __b);
15630 }
15631
15632 __extension__ extern __inline poly64x1_t
15633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15634 vdup_laneq_p64 (poly64x2_t __a, const int __b)
15635 {
15636   return __aarch64_vdup_laneq_p64 (__a, __b);
15637 }
15638
15639 __extension__ extern __inline int8x8_t
15640 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15641 vdup_laneq_s8 (int8x16_t __a, const int __b)
15642 {
15643   return __aarch64_vdup_laneq_s8 (__a, __b);
15644 }
15645
15646 __extension__ extern __inline int16x4_t
15647 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15648 vdup_laneq_s16 (int16x8_t __a, const int __b)
15649 {
15650   return __aarch64_vdup_laneq_s16 (__a, __b);
15651 }
15652
15653 __extension__ extern __inline int32x2_t
15654 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15655 vdup_laneq_s32 (int32x4_t __a, const int __b)
15656 {
15657   return __aarch64_vdup_laneq_s32 (__a, __b);
15658 }
15659
15660 __extension__ extern __inline int64x1_t
15661 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15662 vdup_laneq_s64 (int64x2_t __a, const int __b)
15663 {
15664   return __aarch64_vdup_laneq_s64 (__a, __b);
15665 }
15666
15667 __extension__ extern __inline uint8x8_t
15668 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15669 vdup_laneq_u8 (uint8x16_t __a, const int __b)
15670 {
15671   return __aarch64_vdup_laneq_u8 (__a, __b);
15672 }
15673
15674 __extension__ extern __inline uint16x4_t
15675 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15676 vdup_laneq_u16 (uint16x8_t __a, const int __b)
15677 {
15678   return __aarch64_vdup_laneq_u16 (__a, __b);
15679 }
15680
15681 __extension__ extern __inline uint32x2_t
15682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15683 vdup_laneq_u32 (uint32x4_t __a, const int __b)
15684 {
15685   return __aarch64_vdup_laneq_u32 (__a, __b);
15686 }
15687
15688 __extension__ extern __inline uint64x1_t
15689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15690 vdup_laneq_u64 (uint64x2_t __a, const int __b)
15691 {
15692   return __aarch64_vdup_laneq_u64 (__a, __b);
15693 }
15694
15695 /* vdupq_lane  */
15696
15697 __extension__ extern __inline float16x8_t
15698 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15699 vdupq_lane_f16 (float16x4_t __a, const int __b)
15700 {
15701   return __aarch64_vdupq_lane_f16 (__a, __b);
15702 }
15703
15704 __extension__ extern __inline float32x4_t
15705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15706 vdupq_lane_f32 (float32x2_t __a, const int __b)
15707 {
15708   return __aarch64_vdupq_lane_f32 (__a, __b);
15709 }
15710
15711 __extension__ extern __inline float64x2_t
15712 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15713 vdupq_lane_f64 (float64x1_t __a, const int __b)
15714 {
15715   return __aarch64_vdupq_lane_f64 (__a, __b);
15716 }
15717
15718 __extension__ extern __inline poly8x16_t
15719 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15720 vdupq_lane_p8 (poly8x8_t __a, const int __b)
15721 {
15722   return __aarch64_vdupq_lane_p8 (__a, __b);
15723 }
15724
15725 __extension__ extern __inline poly16x8_t
15726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15727 vdupq_lane_p16 (poly16x4_t __a, const int __b)
15728 {
15729   return __aarch64_vdupq_lane_p16 (__a, __b);
15730 }
15731
15732 __extension__ extern __inline poly64x2_t
15733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15734 vdupq_lane_p64 (poly64x1_t __a, const int __b)
15735 {
15736   return __aarch64_vdupq_lane_p64 (__a, __b);
15737 }
15738
15739 __extension__ extern __inline int8x16_t
15740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15741 vdupq_lane_s8 (int8x8_t __a, const int __b)
15742 {
15743   return __aarch64_vdupq_lane_s8 (__a, __b);
15744 }
15745
15746 __extension__ extern __inline int16x8_t
15747 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15748 vdupq_lane_s16 (int16x4_t __a, const int __b)
15749 {
15750   return __aarch64_vdupq_lane_s16 (__a, __b);
15751 }
15752
15753 __extension__ extern __inline int32x4_t
15754 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15755 vdupq_lane_s32 (int32x2_t __a, const int __b)
15756 {
15757   return __aarch64_vdupq_lane_s32 (__a, __b);
15758 }
15759
15760 __extension__ extern __inline int64x2_t
15761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15762 vdupq_lane_s64 (int64x1_t __a, const int __b)
15763 {
15764   return __aarch64_vdupq_lane_s64 (__a, __b);
15765 }
15766
15767 __extension__ extern __inline uint8x16_t
15768 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15769 vdupq_lane_u8 (uint8x8_t __a, const int __b)
15770 {
15771   return __aarch64_vdupq_lane_u8 (__a, __b);
15772 }
15773
15774 __extension__ extern __inline uint16x8_t
15775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15776 vdupq_lane_u16 (uint16x4_t __a, const int __b)
15777 {
15778   return __aarch64_vdupq_lane_u16 (__a, __b);
15779 }
15780
15781 __extension__ extern __inline uint32x4_t
15782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15783 vdupq_lane_u32 (uint32x2_t __a, const int __b)
15784 {
15785   return __aarch64_vdupq_lane_u32 (__a, __b);
15786 }
15787
15788 __extension__ extern __inline uint64x2_t
15789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15790 vdupq_lane_u64 (uint64x1_t __a, const int __b)
15791 {
15792   return __aarch64_vdupq_lane_u64 (__a, __b);
15793 }
15794
15795 /* vdupq_laneq  */
15796
15797 __extension__ extern __inline float16x8_t
15798 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15799 vdupq_laneq_f16 (float16x8_t __a, const int __b)
15800 {
15801   return __aarch64_vdupq_laneq_f16 (__a, __b);
15802 }
15803
15804 __extension__ extern __inline float32x4_t
15805 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15806 vdupq_laneq_f32 (float32x4_t __a, const int __b)
15807 {
15808   return __aarch64_vdupq_laneq_f32 (__a, __b);
15809 }
15810
15811 __extension__ extern __inline float64x2_t
15812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15813 vdupq_laneq_f64 (float64x2_t __a, const int __b)
15814 {
15815   return __aarch64_vdupq_laneq_f64 (__a, __b);
15816 }
15817
15818 __extension__ extern __inline poly8x16_t
15819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15820 vdupq_laneq_p8 (poly8x16_t __a, const int __b)
15821 {
15822   return __aarch64_vdupq_laneq_p8 (__a, __b);
15823 }
15824
15825 __extension__ extern __inline poly16x8_t
15826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15827 vdupq_laneq_p16 (poly16x8_t __a, const int __b)
15828 {
15829   return __aarch64_vdupq_laneq_p16 (__a, __b);
15830 }
15831
15832 __extension__ extern __inline poly64x2_t
15833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15834 vdupq_laneq_p64 (poly64x2_t __a, const int __b)
15835 {
15836   return __aarch64_vdupq_laneq_p64 (__a, __b);
15837 }
15838
15839 __extension__ extern __inline int8x16_t
15840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15841 vdupq_laneq_s8 (int8x16_t __a, const int __b)
15842 {
15843   return __aarch64_vdupq_laneq_s8 (__a, __b);
15844 }
15845
15846 __extension__ extern __inline int16x8_t
15847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15848 vdupq_laneq_s16 (int16x8_t __a, const int __b)
15849 {
15850   return __aarch64_vdupq_laneq_s16 (__a, __b);
15851 }
15852
15853 __extension__ extern __inline int32x4_t
15854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15855 vdupq_laneq_s32 (int32x4_t __a, const int __b)
15856 {
15857   return __aarch64_vdupq_laneq_s32 (__a, __b);
15858 }
15859
15860 __extension__ extern __inline int64x2_t
15861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15862 vdupq_laneq_s64 (int64x2_t __a, const int __b)
15863 {
15864   return __aarch64_vdupq_laneq_s64 (__a, __b);
15865 }
15866
15867 __extension__ extern __inline uint8x16_t
15868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15869 vdupq_laneq_u8 (uint8x16_t __a, const int __b)
15870 {
15871   return __aarch64_vdupq_laneq_u8 (__a, __b);
15872 }
15873
15874 __extension__ extern __inline uint16x8_t
15875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15876 vdupq_laneq_u16 (uint16x8_t __a, const int __b)
15877 {
15878   return __aarch64_vdupq_laneq_u16 (__a, __b);
15879 }
15880
15881 __extension__ extern __inline uint32x4_t
15882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15883 vdupq_laneq_u32 (uint32x4_t __a, const int __b)
15884 {
15885   return __aarch64_vdupq_laneq_u32 (__a, __b);
15886 }
15887
15888 __extension__ extern __inline uint64x2_t
15889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15890 vdupq_laneq_u64 (uint64x2_t __a, const int __b)
15891 {
15892   return __aarch64_vdupq_laneq_u64 (__a, __b);
15893 }
15894
15895 /* vdupb_lane  */
15896 __extension__ extern __inline poly8_t
15897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15898 vdupb_lane_p8 (poly8x8_t __a, const int __b)
15899 {
15900   return __aarch64_vget_lane_any (__a, __b);
15901 }
15902
15903 __extension__ extern __inline int8_t
15904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15905 vdupb_lane_s8 (int8x8_t __a, const int __b)
15906 {
15907   return __aarch64_vget_lane_any (__a, __b);
15908 }
15909
15910 __extension__ extern __inline uint8_t
15911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15912 vdupb_lane_u8 (uint8x8_t __a, const int __b)
15913 {
15914   return __aarch64_vget_lane_any (__a, __b);
15915 }
15916
15917 /* vduph_lane  */
15918
15919 __extension__ extern __inline float16_t
15920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15921 vduph_lane_f16 (float16x4_t __a, const int __b)
15922 {
15923   return __aarch64_vget_lane_any (__a, __b);
15924 }
15925
15926 __extension__ extern __inline poly16_t
15927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15928 vduph_lane_p16 (poly16x4_t __a, const int __b)
15929 {
15930   return __aarch64_vget_lane_any (__a, __b);
15931 }
15932
15933 __extension__ extern __inline int16_t
15934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15935 vduph_lane_s16 (int16x4_t __a, const int __b)
15936 {
15937   return __aarch64_vget_lane_any (__a, __b);
15938 }
15939
15940 __extension__ extern __inline uint16_t
15941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15942 vduph_lane_u16 (uint16x4_t __a, const int __b)
15943 {
15944   return __aarch64_vget_lane_any (__a, __b);
15945 }
15946
15947 /* vdups_lane  */
15948
15949 __extension__ extern __inline float32_t
15950 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15951 vdups_lane_f32 (float32x2_t __a, const int __b)
15952 {
15953   return __aarch64_vget_lane_any (__a, __b);
15954 }
15955
15956 __extension__ extern __inline int32_t
15957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15958 vdups_lane_s32 (int32x2_t __a, const int __b)
15959 {
15960   return __aarch64_vget_lane_any (__a, __b);
15961 }
15962
15963 __extension__ extern __inline uint32_t
15964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15965 vdups_lane_u32 (uint32x2_t __a, const int __b)
15966 {
15967   return __aarch64_vget_lane_any (__a, __b);
15968 }
15969
15970 /* vdupd_lane  */
15971 __extension__ extern __inline float64_t
15972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15973 vdupd_lane_f64 (float64x1_t __a, const int __b)
15974 {
15975   __AARCH64_LANE_CHECK (__a, __b);
15976   return __a[0];
15977 }
15978
15979 __extension__ extern __inline int64_t
15980 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15981 vdupd_lane_s64 (int64x1_t __a, const int __b)
15982 {
15983   __AARCH64_LANE_CHECK (__a, __b);
15984   return __a[0];
15985 }
15986
15987 __extension__ extern __inline uint64_t
15988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15989 vdupd_lane_u64 (uint64x1_t __a, const int __b)
15990 {
15991   __AARCH64_LANE_CHECK (__a, __b);
15992   return __a[0];
15993 }
15994
15995 /* vdupb_laneq  */
15996 __extension__ extern __inline poly8_t
15997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
15998 vdupb_laneq_p8 (poly8x16_t __a, const int __b)
15999 {
16000   return __aarch64_vget_lane_any (__a, __b);
16001 }
16002
16003 __extension__ extern __inline int8_t
16004 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16005 vdupb_laneq_s8 (int8x16_t __a, const int __b)
16006 {
16007   return __aarch64_vget_lane_any (__a, __b);
16008 }
16009
16010 __extension__ extern __inline uint8_t
16011 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16012 vdupb_laneq_u8 (uint8x16_t __a, const int __b)
16013 {
16014   return __aarch64_vget_lane_any (__a, __b);
16015 }
16016
16017 /* vduph_laneq  */
16018
16019 __extension__ extern __inline float16_t
16020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16021 vduph_laneq_f16 (float16x8_t __a, const int __b)
16022 {
16023   return __aarch64_vget_lane_any (__a, __b);
16024 }
16025
16026 __extension__ extern __inline poly16_t
16027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16028 vduph_laneq_p16 (poly16x8_t __a, const int __b)
16029 {
16030   return __aarch64_vget_lane_any (__a, __b);
16031 }
16032
16033 __extension__ extern __inline int16_t
16034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16035 vduph_laneq_s16 (int16x8_t __a, const int __b)
16036 {
16037   return __aarch64_vget_lane_any (__a, __b);
16038 }
16039
16040 __extension__ extern __inline uint16_t
16041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16042 vduph_laneq_u16 (uint16x8_t __a, const int __b)
16043 {
16044   return __aarch64_vget_lane_any (__a, __b);
16045 }
16046
16047 /* vdups_laneq  */
16048
16049 __extension__ extern __inline float32_t
16050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16051 vdups_laneq_f32 (float32x4_t __a, const int __b)
16052 {
16053   return __aarch64_vget_lane_any (__a, __b);
16054 }
16055
16056 __extension__ extern __inline int32_t
16057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16058 vdups_laneq_s32 (int32x4_t __a, const int __b)
16059 {
16060   return __aarch64_vget_lane_any (__a, __b);
16061 }
16062
16063 __extension__ extern __inline uint32_t
16064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16065 vdups_laneq_u32 (uint32x4_t __a, const int __b)
16066 {
16067   return __aarch64_vget_lane_any (__a, __b);
16068 }
16069
16070 /* vdupd_laneq  */
16071 __extension__ extern __inline float64_t
16072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16073 vdupd_laneq_f64 (float64x2_t __a, const int __b)
16074 {
16075   return __aarch64_vget_lane_any (__a, __b);
16076 }
16077
16078 __extension__ extern __inline int64_t
16079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16080 vdupd_laneq_s64 (int64x2_t __a, const int __b)
16081 {
16082   return __aarch64_vget_lane_any (__a, __b);
16083 }
16084
16085 __extension__ extern __inline uint64_t
16086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16087 vdupd_laneq_u64 (uint64x2_t __a, const int __b)
16088 {
16089   return __aarch64_vget_lane_any (__a, __b);
16090 }
16091
16092 /* vext  */
16093
16094 __extension__ extern __inline float16x4_t
16095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16096 vext_f16 (float16x4_t __a, float16x4_t __b, __const int __c)
16097 {
16098   __AARCH64_LANE_CHECK (__a, __c);
16099 #ifdef __AARCH64EB__
16100   return __builtin_shuffle (__b, __a,
16101                             (uint16x4_t) {4 - __c, 5 - __c, 6 - __c, 7 - __c});
16102 #else
16103   return __builtin_shuffle (__a, __b,
16104                             (uint16x4_t) {__c, __c + 1, __c + 2, __c + 3});
16105 #endif
16106 }
16107
16108 __extension__ extern __inline float32x2_t
16109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16110 vext_f32 (float32x2_t __a, float32x2_t __b, __const int __c)
16111 {
16112   __AARCH64_LANE_CHECK (__a, __c);
16113 #ifdef __AARCH64EB__
16114   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16115 #else
16116   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16117 #endif
16118 }
16119
16120 __extension__ extern __inline float64x1_t
16121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16122 vext_f64 (float64x1_t __a, float64x1_t __b, __const int __c)
16123 {
16124   __AARCH64_LANE_CHECK (__a, __c);
16125   /* The only possible index to the assembler instruction returns element 0.  */
16126   return __a;
16127 }
16128 __extension__ extern __inline poly8x8_t
16129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16130 vext_p8 (poly8x8_t __a, poly8x8_t __b, __const int __c)
16131 {
16132   __AARCH64_LANE_CHECK (__a, __c);
16133 #ifdef __AARCH64EB__
16134   return __builtin_shuffle (__b, __a, (uint8x8_t)
16135       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16136 #else
16137   return __builtin_shuffle (__a, __b,
16138       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16139 #endif
16140 }
16141
16142 __extension__ extern __inline poly16x4_t
16143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16144 vext_p16 (poly16x4_t __a, poly16x4_t __b, __const int __c)
16145 {
16146   __AARCH64_LANE_CHECK (__a, __c);
16147 #ifdef __AARCH64EB__
16148   return __builtin_shuffle (__b, __a,
16149       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16150 #else
16151   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16152 #endif
16153 }
16154
16155 __extension__ extern __inline int8x8_t
16156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16157 vext_s8 (int8x8_t __a, int8x8_t __b, __const int __c)
16158 {
16159   __AARCH64_LANE_CHECK (__a, __c);
16160 #ifdef __AARCH64EB__
16161   return __builtin_shuffle (__b, __a, (uint8x8_t)
16162       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16163 #else
16164   return __builtin_shuffle (__a, __b,
16165       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16166 #endif
16167 }
16168
16169 __extension__ extern __inline int16x4_t
16170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16171 vext_s16 (int16x4_t __a, int16x4_t __b, __const int __c)
16172 {
16173   __AARCH64_LANE_CHECK (__a, __c);
16174 #ifdef __AARCH64EB__
16175   return __builtin_shuffle (__b, __a,
16176       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16177 #else
16178   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16179 #endif
16180 }
16181
16182 __extension__ extern __inline int32x2_t
16183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16184 vext_s32 (int32x2_t __a, int32x2_t __b, __const int __c)
16185 {
16186   __AARCH64_LANE_CHECK (__a, __c);
16187 #ifdef __AARCH64EB__
16188   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16189 #else
16190   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16191 #endif
16192 }
16193
16194 __extension__ extern __inline int64x1_t
16195 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16196 vext_s64 (int64x1_t __a, int64x1_t __b, __const int __c)
16197 {
16198   __AARCH64_LANE_CHECK (__a, __c);
16199   /* The only possible index to the assembler instruction returns element 0.  */
16200   return __a;
16201 }
16202
16203 __extension__ extern __inline uint8x8_t
16204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16205 vext_u8 (uint8x8_t __a, uint8x8_t __b, __const int __c)
16206 {
16207   __AARCH64_LANE_CHECK (__a, __c);
16208 #ifdef __AARCH64EB__
16209   return __builtin_shuffle (__b, __a, (uint8x8_t)
16210       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16211 #else
16212   return __builtin_shuffle (__a, __b,
16213       (uint8x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16214 #endif
16215 }
16216
16217 __extension__ extern __inline uint16x4_t
16218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16219 vext_u16 (uint16x4_t __a, uint16x4_t __b, __const int __c)
16220 {
16221   __AARCH64_LANE_CHECK (__a, __c);
16222 #ifdef __AARCH64EB__
16223   return __builtin_shuffle (__b, __a,
16224       (uint16x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16225 #else
16226   return __builtin_shuffle (__a, __b, (uint16x4_t) {__c, __c+1, __c+2, __c+3});
16227 #endif
16228 }
16229
16230 __extension__ extern __inline uint32x2_t
16231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16232 vext_u32 (uint32x2_t __a, uint32x2_t __b, __const int __c)
16233 {
16234   __AARCH64_LANE_CHECK (__a, __c);
16235 #ifdef __AARCH64EB__
16236   return __builtin_shuffle (__b, __a, (uint32x2_t) {2-__c, 3-__c});
16237 #else
16238   return __builtin_shuffle (__a, __b, (uint32x2_t) {__c, __c+1});
16239 #endif
16240 }
16241
16242 __extension__ extern __inline uint64x1_t
16243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16244 vext_u64 (uint64x1_t __a, uint64x1_t __b, __const int __c)
16245 {
16246   __AARCH64_LANE_CHECK (__a, __c);
16247   /* The only possible index to the assembler instruction returns element 0.  */
16248   return __a;
16249 }
16250
16251 __extension__ extern __inline float16x8_t
16252 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16253 vextq_f16 (float16x8_t __a, float16x8_t __b, __const int __c)
16254 {
16255   __AARCH64_LANE_CHECK (__a, __c);
16256 #ifdef __AARCH64EB__
16257   return __builtin_shuffle (__b, __a,
16258                             (uint16x8_t) {8 - __c, 9 - __c, 10 - __c, 11 - __c,
16259                                           12 - __c, 13 - __c, 14 - __c,
16260                                           15 - __c});
16261 #else
16262   return __builtin_shuffle (__a, __b,
16263                             (uint16x8_t) {__c, __c + 1, __c + 2, __c + 3,
16264                                           __c + 4, __c + 5, __c + 6, __c + 7});
16265 #endif
16266 }
16267
16268 __extension__ extern __inline float32x4_t
16269 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16270 vextq_f32 (float32x4_t __a, float32x4_t __b, __const int __c)
16271 {
16272   __AARCH64_LANE_CHECK (__a, __c);
16273 #ifdef __AARCH64EB__
16274   return __builtin_shuffle (__b, __a,
16275       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16276 #else
16277   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16278 #endif
16279 }
16280
16281 __extension__ extern __inline float64x2_t
16282 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16283 vextq_f64 (float64x2_t __a, float64x2_t __b, __const int __c)
16284 {
16285   __AARCH64_LANE_CHECK (__a, __c);
16286 #ifdef __AARCH64EB__
16287   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16288 #else
16289   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16290 #endif
16291 }
16292
16293 __extension__ extern __inline poly8x16_t
16294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16295 vextq_p8 (poly8x16_t __a, poly8x16_t __b, __const int __c)
16296 {
16297   __AARCH64_LANE_CHECK (__a, __c);
16298 #ifdef __AARCH64EB__
16299   return __builtin_shuffle (__b, __a, (uint8x16_t)
16300       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16301        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16302 #else
16303   return __builtin_shuffle (__a, __b, (uint8x16_t)
16304       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16305        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16306 #endif
16307 }
16308
16309 __extension__ extern __inline poly16x8_t
16310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16311 vextq_p16 (poly16x8_t __a, poly16x8_t __b, __const int __c)
16312 {
16313   __AARCH64_LANE_CHECK (__a, __c);
16314 #ifdef __AARCH64EB__
16315   return __builtin_shuffle (__b, __a, (uint16x8_t)
16316       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16317 #else
16318   return __builtin_shuffle (__a, __b,
16319       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16320 #endif
16321 }
16322
16323 __extension__ extern __inline int8x16_t
16324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16325 vextq_s8 (int8x16_t __a, int8x16_t __b, __const int __c)
16326 {
16327   __AARCH64_LANE_CHECK (__a, __c);
16328 #ifdef __AARCH64EB__
16329   return __builtin_shuffle (__b, __a, (uint8x16_t)
16330       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16331        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16332 #else
16333   return __builtin_shuffle (__a, __b, (uint8x16_t)
16334       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16335        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16336 #endif
16337 }
16338
16339 __extension__ extern __inline int16x8_t
16340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16341 vextq_s16 (int16x8_t __a, int16x8_t __b, __const int __c)
16342 {
16343   __AARCH64_LANE_CHECK (__a, __c);
16344 #ifdef __AARCH64EB__
16345   return __builtin_shuffle (__b, __a, (uint16x8_t)
16346       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16347 #else
16348   return __builtin_shuffle (__a, __b,
16349       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16350 #endif
16351 }
16352
16353 __extension__ extern __inline int32x4_t
16354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16355 vextq_s32 (int32x4_t __a, int32x4_t __b, __const int __c)
16356 {
16357   __AARCH64_LANE_CHECK (__a, __c);
16358 #ifdef __AARCH64EB__
16359   return __builtin_shuffle (__b, __a,
16360       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16361 #else
16362   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16363 #endif
16364 }
16365
16366 __extension__ extern __inline int64x2_t
16367 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16368 vextq_s64 (int64x2_t __a, int64x2_t __b, __const int __c)
16369 {
16370   __AARCH64_LANE_CHECK (__a, __c);
16371 #ifdef __AARCH64EB__
16372   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16373 #else
16374   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16375 #endif
16376 }
16377
16378 __extension__ extern __inline uint8x16_t
16379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16380 vextq_u8 (uint8x16_t __a, uint8x16_t __b, __const int __c)
16381 {
16382   __AARCH64_LANE_CHECK (__a, __c);
16383 #ifdef __AARCH64EB__
16384   return __builtin_shuffle (__b, __a, (uint8x16_t)
16385       {16-__c, 17-__c, 18-__c, 19-__c, 20-__c, 21-__c, 22-__c, 23-__c,
16386        24-__c, 25-__c, 26-__c, 27-__c, 28-__c, 29-__c, 30-__c, 31-__c});
16387 #else
16388   return __builtin_shuffle (__a, __b, (uint8x16_t)
16389       {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7,
16390        __c+8, __c+9, __c+10, __c+11, __c+12, __c+13, __c+14, __c+15});
16391 #endif
16392 }
16393
16394 __extension__ extern __inline uint16x8_t
16395 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16396 vextq_u16 (uint16x8_t __a, uint16x8_t __b, __const int __c)
16397 {
16398   __AARCH64_LANE_CHECK (__a, __c);
16399 #ifdef __AARCH64EB__
16400   return __builtin_shuffle (__b, __a, (uint16x8_t)
16401       {8-__c, 9-__c, 10-__c, 11-__c, 12-__c, 13-__c, 14-__c, 15-__c});
16402 #else
16403   return __builtin_shuffle (__a, __b,
16404       (uint16x8_t) {__c, __c+1, __c+2, __c+3, __c+4, __c+5, __c+6, __c+7});
16405 #endif
16406 }
16407
16408 __extension__ extern __inline uint32x4_t
16409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16410 vextq_u32 (uint32x4_t __a, uint32x4_t __b, __const int __c)
16411 {
16412   __AARCH64_LANE_CHECK (__a, __c);
16413 #ifdef __AARCH64EB__
16414   return __builtin_shuffle (__b, __a,
16415       (uint32x4_t) {4-__c, 5-__c, 6-__c, 7-__c});
16416 #else
16417   return __builtin_shuffle (__a, __b, (uint32x4_t) {__c, __c+1, __c+2, __c+3});
16418 #endif
16419 }
16420
16421 __extension__ extern __inline uint64x2_t
16422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16423 vextq_u64 (uint64x2_t __a, uint64x2_t __b, __const int __c)
16424 {
16425   __AARCH64_LANE_CHECK (__a, __c);
16426 #ifdef __AARCH64EB__
16427   return __builtin_shuffle (__b, __a, (uint64x2_t) {2-__c, 3-__c});
16428 #else
16429   return __builtin_shuffle (__a, __b, (uint64x2_t) {__c, __c+1});
16430 #endif
16431 }
16432
16433 /* vfma  */
16434
16435 __extension__ extern __inline float64x1_t
16436 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16437 vfma_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
16438 {
16439   return (float64x1_t) {__builtin_fma (__b[0], __c[0], __a[0])};
16440 }
16441
16442 __extension__ extern __inline float32x2_t
16443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16444 vfma_f32 (float32x2_t __a, float32x2_t __b, float32x2_t __c)
16445 {
16446   return __builtin_aarch64_fmav2sf (__b, __c, __a);
16447 }
16448
16449 __extension__ extern __inline float32x4_t
16450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16451 vfmaq_f32 (float32x4_t __a, float32x4_t __b, float32x4_t __c)
16452 {
16453   return __builtin_aarch64_fmav4sf (__b, __c, __a);
16454 }
16455
16456 __extension__ extern __inline float64x2_t
16457 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16458 vfmaq_f64 (float64x2_t __a, float64x2_t __b, float64x2_t __c)
16459 {
16460   return __builtin_aarch64_fmav2df (__b, __c, __a);
16461 }
16462
16463 __extension__ extern __inline float32x2_t
16464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16465 vfma_n_f32 (float32x2_t __a, float32x2_t __b, float32_t __c)
16466 {
16467   return __builtin_aarch64_fmav2sf (__b, vdup_n_f32 (__c), __a);
16468 }
16469
16470 __extension__ extern __inline float64x1_t
16471 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16472 vfma_n_f64 (float64x1_t __a, float64x1_t __b, float64_t __c)
16473 {
16474   return (float64x1_t) {__b[0] * __c + __a[0]};
16475 }
16476
16477 __extension__ extern __inline float32x4_t
16478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16479 vfmaq_n_f32 (float32x4_t __a, float32x4_t __b, float32_t __c)
16480 {
16481   return __builtin_aarch64_fmav4sf (__b, vdupq_n_f32 (__c), __a);
16482 }
16483
16484 __extension__ extern __inline float64x2_t
16485 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16486 vfmaq_n_f64 (float64x2_t __a, float64x2_t __b, float64_t __c)
16487 {
16488   return __builtin_aarch64_fmav2df (__b, vdupq_n_f64 (__c), __a);
16489 }
16490
16491 /* vfma_lane  */
16492
16493 __extension__ extern __inline float32x2_t
16494 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16495 vfma_lane_f32 (float32x2_t __a, float32x2_t __b,
16496                float32x2_t __c, const int __lane)
16497 {
16498   return __builtin_aarch64_fmav2sf (__b,
16499                                     __aarch64_vdup_lane_f32 (__c, __lane),
16500                                     __a);
16501 }
16502
16503 __extension__ extern __inline float64x1_t
16504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16505 vfma_lane_f64 (float64x1_t __a, float64x1_t __b,
16506                float64x1_t __c, const int __lane)
16507 {
16508   return (float64x1_t) {__builtin_fma (__b[0], __c[0], __a[0])};
16509 }
16510
16511 __extension__ extern __inline float64_t
16512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16513 vfmad_lane_f64 (float64_t __a, float64_t __b,
16514                 float64x1_t __c, const int __lane)
16515 {
16516   return __builtin_fma (__b, __c[0], __a);
16517 }
16518
16519 __extension__ extern __inline float32_t
16520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16521 vfmas_lane_f32 (float32_t __a, float32_t __b,
16522                 float32x2_t __c, const int __lane)
16523 {
16524   return __builtin_fmaf (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16525 }
16526
16527 /* vfma_laneq  */
16528
16529 __extension__ extern __inline float32x2_t
16530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16531 vfma_laneq_f32 (float32x2_t __a, float32x2_t __b,
16532                 float32x4_t __c, const int __lane)
16533 {
16534   return __builtin_aarch64_fmav2sf (__b,
16535                                     __aarch64_vdup_laneq_f32 (__c, __lane),
16536                                     __a);
16537 }
16538
16539 __extension__ extern __inline float64x1_t
16540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16541 vfma_laneq_f64 (float64x1_t __a, float64x1_t __b,
16542                 float64x2_t __c, const int __lane)
16543 {
16544   float64_t __c0 = __aarch64_vget_lane_any (__c, __lane);
16545   return (float64x1_t) {__builtin_fma (__b[0], __c0, __a[0])};
16546 }
16547
16548 __extension__ extern __inline float64_t
16549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16550 vfmad_laneq_f64 (float64_t __a, float64_t __b,
16551                  float64x2_t __c, const int __lane)
16552 {
16553   return __builtin_fma (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16554 }
16555
16556 __extension__ extern __inline float32_t
16557 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16558 vfmas_laneq_f32 (float32_t __a, float32_t __b,
16559                  float32x4_t __c, const int __lane)
16560 {
16561   return __builtin_fmaf (__b, __aarch64_vget_lane_any (__c, __lane), __a);
16562 }
16563
16564 /* vfmaq_lane  */
16565
16566 __extension__ extern __inline float32x4_t
16567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16568 vfmaq_lane_f32 (float32x4_t __a, float32x4_t __b,
16569                 float32x2_t __c, const int __lane)
16570 {
16571   return __builtin_aarch64_fmav4sf (__b,
16572                                     __aarch64_vdupq_lane_f32 (__c, __lane),
16573                                     __a);
16574 }
16575
16576 __extension__ extern __inline float64x2_t
16577 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16578 vfmaq_lane_f64 (float64x2_t __a, float64x2_t __b,
16579                 float64x1_t __c, const int __lane)
16580 {
16581   return __builtin_aarch64_fmav2df (__b, vdupq_n_f64 (__c[0]), __a);
16582 }
16583
16584 /* vfmaq_laneq  */
16585
16586 __extension__ extern __inline float32x4_t
16587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16588 vfmaq_laneq_f32 (float32x4_t __a, float32x4_t __b,
16589                  float32x4_t __c, const int __lane)
16590 {
16591   return __builtin_aarch64_fmav4sf (__b,
16592                                     __aarch64_vdupq_laneq_f32 (__c, __lane),
16593                                     __a);
16594 }
16595
16596 __extension__ extern __inline float64x2_t
16597 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16598 vfmaq_laneq_f64 (float64x2_t __a, float64x2_t __b,
16599                  float64x2_t __c, const int __lane)
16600 {
16601   return __builtin_aarch64_fmav2df (__b,
16602                                     __aarch64_vdupq_laneq_f64 (__c, __lane),
16603                                     __a);
16604 }
16605
16606 /* vfms  */
16607
16608 __extension__ extern __inline float64x1_t
16609 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16610 vfms_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
16611 {
16612   return (float64x1_t) {__builtin_fma (-__b[0], __c[0], __a[0])};
16613 }
16614
16615 __extension__ extern __inline float32x2_t
16616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16617 vfms_f32 (float32x2_t __a, float32x2_t __b, float32x2_t __c)
16618 {
16619   return __builtin_aarch64_fmav2sf (-__b, __c, __a);
16620 }
16621
16622 __extension__ extern __inline float32x4_t
16623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16624 vfmsq_f32 (float32x4_t __a, float32x4_t __b, float32x4_t __c)
16625 {
16626   return __builtin_aarch64_fmav4sf (-__b, __c, __a);
16627 }
16628
16629 __extension__ extern __inline float64x2_t
16630 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16631 vfmsq_f64 (float64x2_t __a, float64x2_t __b, float64x2_t __c)
16632 {
16633   return __builtin_aarch64_fmav2df (-__b, __c, __a);
16634 }
16635
16636 __extension__ extern __inline float32x2_t
16637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16638 vfms_n_f32 (float32x2_t __a, float32x2_t __b, float32_t __c)
16639 {
16640   return __builtin_aarch64_fmav2sf (-__b, vdup_n_f32 (__c), __a);
16641 }
16642
16643 __extension__ extern __inline float64x1_t
16644 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16645 vfms_n_f64 (float64x1_t __a, float64x1_t __b, float64_t __c)
16646 {
16647   return (float64x1_t) {-__b[0] * __c + __a[0]};
16648 }
16649
16650 __extension__ extern __inline float32x4_t
16651 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16652 vfmsq_n_f32 (float32x4_t __a, float32x4_t __b, float32_t __c)
16653 {
16654   return __builtin_aarch64_fmav4sf (-__b, vdupq_n_f32 (__c), __a);
16655 }
16656
16657 __extension__ extern __inline float64x2_t
16658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16659 vfmsq_n_f64 (float64x2_t __a, float64x2_t __b, float64_t __c)
16660 {
16661   return __builtin_aarch64_fmav2df (-__b, vdupq_n_f64 (__c), __a);
16662 }
16663
16664 /* vfms_lane  */
16665
16666 __extension__ extern __inline float32x2_t
16667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16668 vfms_lane_f32 (float32x2_t __a, float32x2_t __b,
16669                float32x2_t __c, const int __lane)
16670 {
16671   return __builtin_aarch64_fmav2sf (-__b,
16672                                     __aarch64_vdup_lane_f32 (__c, __lane),
16673                                     __a);
16674 }
16675
16676 __extension__ extern __inline float64x1_t
16677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16678 vfms_lane_f64 (float64x1_t __a, float64x1_t __b,
16679                float64x1_t __c, const int __lane)
16680 {
16681   return (float64x1_t) {__builtin_fma (-__b[0], __c[0], __a[0])};
16682 }
16683
16684 __extension__ extern __inline float64_t
16685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16686 vfmsd_lane_f64 (float64_t __a, float64_t __b,
16687                 float64x1_t __c, const int __lane)
16688 {
16689   return __builtin_fma (-__b, __c[0], __a);
16690 }
16691
16692 __extension__ extern __inline float32_t
16693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16694 vfmss_lane_f32 (float32_t __a, float32_t __b,
16695                 float32x2_t __c, const int __lane)
16696 {
16697   return __builtin_fmaf (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16698 }
16699
16700 /* vfms_laneq  */
16701
16702 __extension__ extern __inline float32x2_t
16703 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16704 vfms_laneq_f32 (float32x2_t __a, float32x2_t __b,
16705                 float32x4_t __c, const int __lane)
16706 {
16707   return __builtin_aarch64_fmav2sf (-__b,
16708                                     __aarch64_vdup_laneq_f32 (__c, __lane),
16709                                     __a);
16710 }
16711
16712 __extension__ extern __inline float64x1_t
16713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16714 vfms_laneq_f64 (float64x1_t __a, float64x1_t __b,
16715                 float64x2_t __c, const int __lane)
16716 {
16717   float64_t __c0 = __aarch64_vget_lane_any (__c, __lane);
16718   return (float64x1_t) {__builtin_fma (-__b[0], __c0, __a[0])};
16719 }
16720
16721 __extension__ extern __inline float64_t
16722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16723 vfmsd_laneq_f64 (float64_t __a, float64_t __b,
16724                  float64x2_t __c, const int __lane)
16725 {
16726   return __builtin_fma (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16727 }
16728
16729 __extension__ extern __inline float32_t
16730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16731 vfmss_laneq_f32 (float32_t __a, float32_t __b,
16732                  float32x4_t __c, const int __lane)
16733 {
16734   return __builtin_fmaf (-__b, __aarch64_vget_lane_any (__c, __lane), __a);
16735 }
16736
16737 /* vfmsq_lane  */
16738
16739 __extension__ extern __inline float32x4_t
16740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16741 vfmsq_lane_f32 (float32x4_t __a, float32x4_t __b,
16742                 float32x2_t __c, const int __lane)
16743 {
16744   return __builtin_aarch64_fmav4sf (-__b,
16745                                     __aarch64_vdupq_lane_f32 (__c, __lane),
16746                                     __a);
16747 }
16748
16749 __extension__ extern __inline float64x2_t
16750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16751 vfmsq_lane_f64 (float64x2_t __a, float64x2_t __b,
16752                 float64x1_t __c, const int __lane)
16753 {
16754   return __builtin_aarch64_fmav2df (-__b, vdupq_n_f64 (__c[0]), __a);
16755 }
16756
16757 /* vfmsq_laneq  */
16758
16759 __extension__ extern __inline float32x4_t
16760 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16761 vfmsq_laneq_f32 (float32x4_t __a, float32x4_t __b,
16762                  float32x4_t __c, const int __lane)
16763 {
16764   return __builtin_aarch64_fmav4sf (-__b,
16765                                     __aarch64_vdupq_laneq_f32 (__c, __lane),
16766                                     __a);
16767 }
16768
16769 __extension__ extern __inline float64x2_t
16770 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16771 vfmsq_laneq_f64 (float64x2_t __a, float64x2_t __b,
16772                  float64x2_t __c, const int __lane)
16773 {
16774   return __builtin_aarch64_fmav2df (-__b,
16775                                     __aarch64_vdupq_laneq_f64 (__c, __lane),
16776                                     __a);
16777 }
16778
16779 /* vld1 */
16780
16781 __extension__ extern __inline float16x4_t
16782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16783 vld1_f16 (const float16_t *__a)
16784 {
16785   return __builtin_aarch64_ld1v4hf (__a);
16786 }
16787
16788 __extension__ extern __inline float32x2_t
16789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16790 vld1_f32 (const float32_t *a)
16791 {
16792   return __builtin_aarch64_ld1v2sf ((const __builtin_aarch64_simd_sf *) a);
16793 }
16794
16795 __extension__ extern __inline float64x1_t
16796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16797 vld1_f64 (const float64_t *a)
16798 {
16799   return (float64x1_t) {*a};
16800 }
16801
16802 __extension__ extern __inline poly8x8_t
16803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16804 vld1_p8 (const poly8_t *a)
16805 {
16806   return (poly8x8_t)
16807     __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
16808 }
16809
16810 __extension__ extern __inline poly16x4_t
16811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16812 vld1_p16 (const poly16_t *a)
16813 {
16814   return (poly16x4_t)
16815     __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
16816 }
16817
16818 __extension__ extern __inline poly64x1_t
16819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16820 vld1_p64 (const poly64_t *a)
16821 {
16822   return (poly64x1_t) {*a};
16823 }
16824
16825 __extension__ extern __inline int8x8_t
16826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16827 vld1_s8 (const int8_t *a)
16828 {
16829   return __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
16830 }
16831
16832 __extension__ extern __inline int16x4_t
16833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16834 vld1_s16 (const int16_t *a)
16835 {
16836   return __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
16837 }
16838
16839 __extension__ extern __inline int32x2_t
16840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16841 vld1_s32 (const int32_t *a)
16842 {
16843   return __builtin_aarch64_ld1v2si ((const __builtin_aarch64_simd_si *) a);
16844 }
16845
16846 __extension__ extern __inline int64x1_t
16847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16848 vld1_s64 (const int64_t *a)
16849 {
16850   return (int64x1_t) {*a};
16851 }
16852
16853 __extension__ extern __inline uint8x8_t
16854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16855 vld1_u8 (const uint8_t *a)
16856 {
16857   return (uint8x8_t)
16858     __builtin_aarch64_ld1v8qi ((const __builtin_aarch64_simd_qi *) a);
16859 }
16860
16861 __extension__ extern __inline uint16x4_t
16862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16863 vld1_u16 (const uint16_t *a)
16864 {
16865   return (uint16x4_t)
16866     __builtin_aarch64_ld1v4hi ((const __builtin_aarch64_simd_hi *) a);
16867 }
16868
16869 __extension__ extern __inline uint32x2_t
16870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16871 vld1_u32 (const uint32_t *a)
16872 {
16873   return (uint32x2_t)
16874     __builtin_aarch64_ld1v2si ((const __builtin_aarch64_simd_si *) a);
16875 }
16876
16877 __extension__ extern __inline uint64x1_t
16878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16879 vld1_u64 (const uint64_t *a)
16880 {
16881   return (uint64x1_t) {*a};
16882 }
16883
16884 /* vld1q */
16885
16886 __extension__ extern __inline float16x8_t
16887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16888 vld1q_f16 (const float16_t *__a)
16889 {
16890   return __builtin_aarch64_ld1v8hf (__a);
16891 }
16892
16893 __extension__ extern __inline float32x4_t
16894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16895 vld1q_f32 (const float32_t *a)
16896 {
16897   return __builtin_aarch64_ld1v4sf ((const __builtin_aarch64_simd_sf *) a);
16898 }
16899
16900 __extension__ extern __inline float64x2_t
16901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16902 vld1q_f64 (const float64_t *a)
16903 {
16904   return __builtin_aarch64_ld1v2df ((const __builtin_aarch64_simd_df *) a);
16905 }
16906
16907 __extension__ extern __inline poly8x16_t
16908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16909 vld1q_p8 (const poly8_t *a)
16910 {
16911   return (poly8x16_t)
16912     __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
16913 }
16914
16915 __extension__ extern __inline poly16x8_t
16916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16917 vld1q_p16 (const poly16_t *a)
16918 {
16919   return (poly16x8_t)
16920     __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
16921 }
16922
16923 __extension__ extern __inline poly64x2_t
16924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16925 vld1q_p64 (const poly64_t *a)
16926 {
16927   return (poly64x2_t)
16928     __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
16929 }
16930
16931 __extension__ extern __inline int8x16_t
16932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16933 vld1q_s8 (const int8_t *a)
16934 {
16935   return __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
16936 }
16937
16938 __extension__ extern __inline int16x8_t
16939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16940 vld1q_s16 (const int16_t *a)
16941 {
16942   return __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
16943 }
16944
16945 __extension__ extern __inline int32x4_t
16946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16947 vld1q_s32 (const int32_t *a)
16948 {
16949   return __builtin_aarch64_ld1v4si ((const __builtin_aarch64_simd_si *) a);
16950 }
16951
16952 __extension__ extern __inline int64x2_t
16953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16954 vld1q_s64 (const int64_t *a)
16955 {
16956   return __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
16957 }
16958
16959 __extension__ extern __inline uint8x16_t
16960 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16961 vld1q_u8 (const uint8_t *a)
16962 {
16963   return (uint8x16_t)
16964     __builtin_aarch64_ld1v16qi ((const __builtin_aarch64_simd_qi *) a);
16965 }
16966
16967 __extension__ extern __inline uint16x8_t
16968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16969 vld1q_u16 (const uint16_t *a)
16970 {
16971   return (uint16x8_t)
16972     __builtin_aarch64_ld1v8hi ((const __builtin_aarch64_simd_hi *) a);
16973 }
16974
16975 __extension__ extern __inline uint32x4_t
16976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16977 vld1q_u32 (const uint32_t *a)
16978 {
16979   return (uint32x4_t)
16980     __builtin_aarch64_ld1v4si ((const __builtin_aarch64_simd_si *) a);
16981 }
16982
16983 __extension__ extern __inline uint64x2_t
16984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16985 vld1q_u64 (const uint64_t *a)
16986 {
16987   return (uint64x2_t)
16988     __builtin_aarch64_ld1v2di ((const __builtin_aarch64_simd_di *) a);
16989 }
16990
16991 /* vld1_dup  */
16992
16993 __extension__ extern __inline float16x4_t
16994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
16995 vld1_dup_f16 (const float16_t* __a)
16996 {
16997   return vdup_n_f16 (*__a);
16998 }
16999
17000 __extension__ extern __inline float32x2_t
17001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17002 vld1_dup_f32 (const float32_t* __a)
17003 {
17004   return vdup_n_f32 (*__a);
17005 }
17006
17007 __extension__ extern __inline float64x1_t
17008 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17009 vld1_dup_f64 (const float64_t* __a)
17010 {
17011   return vdup_n_f64 (*__a);
17012 }
17013
17014 __extension__ extern __inline poly8x8_t
17015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17016 vld1_dup_p8 (const poly8_t* __a)
17017 {
17018   return vdup_n_p8 (*__a);
17019 }
17020
17021 __extension__ extern __inline poly16x4_t
17022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17023 vld1_dup_p16 (const poly16_t* __a)
17024 {
17025   return vdup_n_p16 (*__a);
17026 }
17027
17028 __extension__ extern __inline poly64x1_t
17029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17030 vld1_dup_p64 (const poly64_t* __a)
17031 {
17032   return vdup_n_p64 (*__a);
17033 }
17034
17035 __extension__ extern __inline int8x8_t
17036 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17037 vld1_dup_s8 (const int8_t* __a)
17038 {
17039   return vdup_n_s8 (*__a);
17040 }
17041
17042 __extension__ extern __inline int16x4_t
17043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17044 vld1_dup_s16 (const int16_t* __a)
17045 {
17046   return vdup_n_s16 (*__a);
17047 }
17048
17049 __extension__ extern __inline int32x2_t
17050 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17051 vld1_dup_s32 (const int32_t* __a)
17052 {
17053   return vdup_n_s32 (*__a);
17054 }
17055
17056 __extension__ extern __inline int64x1_t
17057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17058 vld1_dup_s64 (const int64_t* __a)
17059 {
17060   return vdup_n_s64 (*__a);
17061 }
17062
17063 __extension__ extern __inline uint8x8_t
17064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17065 vld1_dup_u8 (const uint8_t* __a)
17066 {
17067   return vdup_n_u8 (*__a);
17068 }
17069
17070 __extension__ extern __inline uint16x4_t
17071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17072 vld1_dup_u16 (const uint16_t* __a)
17073 {
17074   return vdup_n_u16 (*__a);
17075 }
17076
17077 __extension__ extern __inline uint32x2_t
17078 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17079 vld1_dup_u32 (const uint32_t* __a)
17080 {
17081   return vdup_n_u32 (*__a);
17082 }
17083
17084 __extension__ extern __inline uint64x1_t
17085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17086 vld1_dup_u64 (const uint64_t* __a)
17087 {
17088   return vdup_n_u64 (*__a);
17089 }
17090
17091 /* vld1q_dup  */
17092
17093 __extension__ extern __inline float16x8_t
17094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17095 vld1q_dup_f16 (const float16_t* __a)
17096 {
17097   return vdupq_n_f16 (*__a);
17098 }
17099
17100 __extension__ extern __inline float32x4_t
17101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17102 vld1q_dup_f32 (const float32_t* __a)
17103 {
17104   return vdupq_n_f32 (*__a);
17105 }
17106
17107 __extension__ extern __inline float64x2_t
17108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17109 vld1q_dup_f64 (const float64_t* __a)
17110 {
17111   return vdupq_n_f64 (*__a);
17112 }
17113
17114 __extension__ extern __inline poly8x16_t
17115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17116 vld1q_dup_p8 (const poly8_t* __a)
17117 {
17118   return vdupq_n_p8 (*__a);
17119 }
17120
17121 __extension__ extern __inline poly16x8_t
17122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17123 vld1q_dup_p16 (const poly16_t* __a)
17124 {
17125   return vdupq_n_p16 (*__a);
17126 }
17127
17128 __extension__ extern __inline poly64x2_t
17129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17130 vld1q_dup_p64 (const poly64_t* __a)
17131 {
17132   return vdupq_n_p64 (*__a);
17133 }
17134
17135  __extension__ extern __inline int8x16_t
17136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17137 vld1q_dup_s8 (const int8_t* __a)
17138 {
17139   return vdupq_n_s8 (*__a);
17140 }
17141
17142 __extension__ extern __inline int16x8_t
17143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17144 vld1q_dup_s16 (const int16_t* __a)
17145 {
17146   return vdupq_n_s16 (*__a);
17147 }
17148
17149 __extension__ extern __inline int32x4_t
17150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17151 vld1q_dup_s32 (const int32_t* __a)
17152 {
17153   return vdupq_n_s32 (*__a);
17154 }
17155
17156 __extension__ extern __inline int64x2_t
17157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17158 vld1q_dup_s64 (const int64_t* __a)
17159 {
17160   return vdupq_n_s64 (*__a);
17161 }
17162
17163 __extension__ extern __inline uint8x16_t
17164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17165 vld1q_dup_u8 (const uint8_t* __a)
17166 {
17167   return vdupq_n_u8 (*__a);
17168 }
17169
17170 __extension__ extern __inline uint16x8_t
17171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17172 vld1q_dup_u16 (const uint16_t* __a)
17173 {
17174   return vdupq_n_u16 (*__a);
17175 }
17176
17177 __extension__ extern __inline uint32x4_t
17178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17179 vld1q_dup_u32 (const uint32_t* __a)
17180 {
17181   return vdupq_n_u32 (*__a);
17182 }
17183
17184 __extension__ extern __inline uint64x2_t
17185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17186 vld1q_dup_u64 (const uint64_t* __a)
17187 {
17188   return vdupq_n_u64 (*__a);
17189 }
17190
17191 /* vld1_lane  */
17192
17193 __extension__ extern __inline float16x4_t
17194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17195 vld1_lane_f16 (const float16_t *__src, float16x4_t __vec, const int __lane)
17196 {
17197   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17198 }
17199
17200 __extension__ extern __inline float32x2_t
17201 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17202 vld1_lane_f32 (const float32_t *__src, float32x2_t __vec, const int __lane)
17203 {
17204   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17205 }
17206
17207 __extension__ extern __inline float64x1_t
17208 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17209 vld1_lane_f64 (const float64_t *__src, float64x1_t __vec, const int __lane)
17210 {
17211   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17212 }
17213
17214 __extension__ extern __inline poly8x8_t
17215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17216 vld1_lane_p8 (const poly8_t *__src, poly8x8_t __vec, const int __lane)
17217 {
17218   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17219 }
17220
17221 __extension__ extern __inline poly16x4_t
17222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17223 vld1_lane_p16 (const poly16_t *__src, poly16x4_t __vec, const int __lane)
17224 {
17225   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17226 }
17227
17228 __extension__ extern __inline poly64x1_t
17229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17230 vld1_lane_p64 (const poly64_t *__src, poly64x1_t __vec, const int __lane)
17231 {
17232   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17233 }
17234
17235 __extension__ extern __inline int8x8_t
17236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17237 vld1_lane_s8 (const int8_t *__src, int8x8_t __vec, const int __lane)
17238 {
17239   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17240 }
17241
17242 __extension__ extern __inline int16x4_t
17243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17244 vld1_lane_s16 (const int16_t *__src, int16x4_t __vec, const int __lane)
17245 {
17246   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17247 }
17248
17249 __extension__ extern __inline int32x2_t
17250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17251 vld1_lane_s32 (const int32_t *__src, int32x2_t __vec, const int __lane)
17252 {
17253   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17254 }
17255
17256 __extension__ extern __inline int64x1_t
17257 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17258 vld1_lane_s64 (const int64_t *__src, int64x1_t __vec, const int __lane)
17259 {
17260   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17261 }
17262
17263 __extension__ extern __inline uint8x8_t
17264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17265 vld1_lane_u8 (const uint8_t *__src, uint8x8_t __vec, const int __lane)
17266 {
17267   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17268 }
17269
17270 __extension__ extern __inline uint16x4_t
17271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17272 vld1_lane_u16 (const uint16_t *__src, uint16x4_t __vec, const int __lane)
17273 {
17274   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17275 }
17276
17277 __extension__ extern __inline uint32x2_t
17278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17279 vld1_lane_u32 (const uint32_t *__src, uint32x2_t __vec, const int __lane)
17280 {
17281   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17282 }
17283
17284 __extension__ extern __inline uint64x1_t
17285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17286 vld1_lane_u64 (const uint64_t *__src, uint64x1_t __vec, const int __lane)
17287 {
17288   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17289 }
17290
17291 /* vld1q_lane  */
17292
17293 __extension__ extern __inline float16x8_t
17294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17295 vld1q_lane_f16 (const float16_t *__src, float16x8_t __vec, const int __lane)
17296 {
17297   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17298 }
17299
17300 __extension__ extern __inline float32x4_t
17301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17302 vld1q_lane_f32 (const float32_t *__src, float32x4_t __vec, const int __lane)
17303 {
17304   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17305 }
17306
17307 __extension__ extern __inline float64x2_t
17308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17309 vld1q_lane_f64 (const float64_t *__src, float64x2_t __vec, const int __lane)
17310 {
17311   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17312 }
17313
17314 __extension__ extern __inline poly8x16_t
17315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17316 vld1q_lane_p8 (const poly8_t *__src, poly8x16_t __vec, const int __lane)
17317 {
17318   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17319 }
17320
17321 __extension__ extern __inline poly16x8_t
17322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17323 vld1q_lane_p16 (const poly16_t *__src, poly16x8_t __vec, const int __lane)
17324 {
17325   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17326 }
17327
17328 __extension__ extern __inline poly64x2_t
17329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17330 vld1q_lane_p64 (const poly64_t *__src, poly64x2_t __vec, const int __lane)
17331 {
17332   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17333 }
17334
17335 __extension__ extern __inline int8x16_t
17336 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17337 vld1q_lane_s8 (const int8_t *__src, int8x16_t __vec, const int __lane)
17338 {
17339   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17340 }
17341
17342 __extension__ extern __inline int16x8_t
17343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17344 vld1q_lane_s16 (const int16_t *__src, int16x8_t __vec, const int __lane)
17345 {
17346   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17347 }
17348
17349 __extension__ extern __inline int32x4_t
17350 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17351 vld1q_lane_s32 (const int32_t *__src, int32x4_t __vec, const int __lane)
17352 {
17353   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17354 }
17355
17356 __extension__ extern __inline int64x2_t
17357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17358 vld1q_lane_s64 (const int64_t *__src, int64x2_t __vec, const int __lane)
17359 {
17360   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17361 }
17362
17363 __extension__ extern __inline uint8x16_t
17364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17365 vld1q_lane_u8 (const uint8_t *__src, uint8x16_t __vec, const int __lane)
17366 {
17367   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17368 }
17369
17370 __extension__ extern __inline uint16x8_t
17371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17372 vld1q_lane_u16 (const uint16_t *__src, uint16x8_t __vec, const int __lane)
17373 {
17374   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17375 }
17376
17377 __extension__ extern __inline uint32x4_t
17378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17379 vld1q_lane_u32 (const uint32_t *__src, uint32x4_t __vec, const int __lane)
17380 {
17381   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17382 }
17383
17384 __extension__ extern __inline uint64x2_t
17385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17386 vld1q_lane_u64 (const uint64_t *__src, uint64x2_t __vec, const int __lane)
17387 {
17388   return __aarch64_vset_lane_any (*__src, __vec, __lane);
17389 }
17390
17391 /* vldn */
17392
17393 __extension__ extern __inline int64x1x2_t
17394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17395 vld2_s64 (const int64_t * __a)
17396 {
17397   int64x1x2_t ret;
17398   __builtin_aarch64_simd_oi __o;
17399   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
17400   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17401   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17402   return ret;
17403 }
17404
17405 __extension__ extern __inline uint64x1x2_t
17406 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17407 vld2_u64 (const uint64_t * __a)
17408 {
17409   uint64x1x2_t ret;
17410   __builtin_aarch64_simd_oi __o;
17411   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
17412   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
17413   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
17414   return ret;
17415 }
17416
17417 __extension__ extern __inline float64x1x2_t
17418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17419 vld2_f64 (const float64_t * __a)
17420 {
17421   float64x1x2_t ret;
17422   __builtin_aarch64_simd_oi __o;
17423   __o = __builtin_aarch64_ld2df ((const __builtin_aarch64_simd_df *) __a);
17424   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
17425   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
17426   return ret;
17427 }
17428
17429 __extension__ extern __inline int8x8x2_t
17430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17431 vld2_s8 (const int8_t * __a)
17432 {
17433   int8x8x2_t ret;
17434   __builtin_aarch64_simd_oi __o;
17435   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17436   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17437   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17438   return ret;
17439 }
17440
17441 __extension__ extern __inline poly8x8x2_t
17442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17443 vld2_p8 (const poly8_t * __a)
17444 {
17445   poly8x8x2_t ret;
17446   __builtin_aarch64_simd_oi __o;
17447   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17448   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17449   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17450   return ret;
17451 }
17452
17453 __extension__ extern __inline poly64x1x2_t
17454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17455 vld2_p64 (const poly64_t * __a)
17456 {
17457   poly64x1x2_t ret;
17458   __builtin_aarch64_simd_oi __o;
17459   __o = __builtin_aarch64_ld2di ((const __builtin_aarch64_simd_di *) __a);
17460   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 0);
17461   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 1);
17462   return ret;
17463 }
17464
17465 __extension__ extern __inline int16x4x2_t
17466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17467 vld2_s16 (const int16_t * __a)
17468 {
17469   int16x4x2_t ret;
17470   __builtin_aarch64_simd_oi __o;
17471   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17472   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17473   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17474   return ret;
17475 }
17476
17477 __extension__ extern __inline poly16x4x2_t
17478 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17479 vld2_p16 (const poly16_t * __a)
17480 {
17481   poly16x4x2_t ret;
17482   __builtin_aarch64_simd_oi __o;
17483   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17484   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17485   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17486   return ret;
17487 }
17488
17489 __extension__ extern __inline int32x2x2_t
17490 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17491 vld2_s32 (const int32_t * __a)
17492 {
17493   int32x2x2_t ret;
17494   __builtin_aarch64_simd_oi __o;
17495   __o = __builtin_aarch64_ld2v2si ((const __builtin_aarch64_simd_si *) __a);
17496   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
17497   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
17498   return ret;
17499 }
17500
17501 __extension__ extern __inline uint8x8x2_t
17502 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17503 vld2_u8 (const uint8_t * __a)
17504 {
17505   uint8x8x2_t ret;
17506   __builtin_aarch64_simd_oi __o;
17507   __o = __builtin_aarch64_ld2v8qi ((const __builtin_aarch64_simd_qi *) __a);
17508   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
17509   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
17510   return ret;
17511 }
17512
17513 __extension__ extern __inline uint16x4x2_t
17514 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17515 vld2_u16 (const uint16_t * __a)
17516 {
17517   uint16x4x2_t ret;
17518   __builtin_aarch64_simd_oi __o;
17519   __o = __builtin_aarch64_ld2v4hi ((const __builtin_aarch64_simd_hi *) __a);
17520   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
17521   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
17522   return ret;
17523 }
17524
17525 __extension__ extern __inline uint32x2x2_t
17526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17527 vld2_u32 (const uint32_t * __a)
17528 {
17529   uint32x2x2_t ret;
17530   __builtin_aarch64_simd_oi __o;
17531   __o = __builtin_aarch64_ld2v2si ((const __builtin_aarch64_simd_si *) __a);
17532   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
17533   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
17534   return ret;
17535 }
17536
17537 __extension__ extern __inline float16x4x2_t
17538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17539 vld2_f16 (const float16_t * __a)
17540 {
17541   float16x4x2_t ret;
17542   __builtin_aarch64_simd_oi __o;
17543   __o = __builtin_aarch64_ld2v4hf (__a);
17544   ret.val[0] = __builtin_aarch64_get_dregoiv4hf (__o, 0);
17545   ret.val[1] = __builtin_aarch64_get_dregoiv4hf (__o, 1);
17546   return ret;
17547 }
17548
17549 __extension__ extern __inline float32x2x2_t
17550 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17551 vld2_f32 (const float32_t * __a)
17552 {
17553   float32x2x2_t ret;
17554   __builtin_aarch64_simd_oi __o;
17555   __o = __builtin_aarch64_ld2v2sf ((const __builtin_aarch64_simd_sf *) __a);
17556   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
17557   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
17558   return ret;
17559 }
17560
17561 __extension__ extern __inline int8x16x2_t
17562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17563 vld2q_s8 (const int8_t * __a)
17564 {
17565   int8x16x2_t ret;
17566   __builtin_aarch64_simd_oi __o;
17567   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17568   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17569   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17570   return ret;
17571 }
17572
17573 __extension__ extern __inline poly8x16x2_t
17574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17575 vld2q_p8 (const poly8_t * __a)
17576 {
17577   poly8x16x2_t ret;
17578   __builtin_aarch64_simd_oi __o;
17579   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17580   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17581   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17582   return ret;
17583 }
17584
17585 __extension__ extern __inline int16x8x2_t
17586 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17587 vld2q_s16 (const int16_t * __a)
17588 {
17589   int16x8x2_t ret;
17590   __builtin_aarch64_simd_oi __o;
17591   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17592   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17593   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17594   return ret;
17595 }
17596
17597 __extension__ extern __inline poly16x8x2_t
17598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17599 vld2q_p16 (const poly16_t * __a)
17600 {
17601   poly16x8x2_t ret;
17602   __builtin_aarch64_simd_oi __o;
17603   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17604   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17605   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17606   return ret;
17607 }
17608
17609 __extension__ extern __inline poly64x2x2_t
17610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17611 vld2q_p64 (const poly64_t * __a)
17612 {
17613   poly64x2x2_t ret;
17614   __builtin_aarch64_simd_oi __o;
17615   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
17616   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 0);
17617   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 1);
17618   return ret;
17619 }
17620
17621 __extension__ extern __inline int32x4x2_t
17622 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17623 vld2q_s32 (const int32_t * __a)
17624 {
17625   int32x4x2_t ret;
17626   __builtin_aarch64_simd_oi __o;
17627   __o = __builtin_aarch64_ld2v4si ((const __builtin_aarch64_simd_si *) __a);
17628   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
17629   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
17630   return ret;
17631 }
17632
17633 __extension__ extern __inline int64x2x2_t
17634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17635 vld2q_s64 (const int64_t * __a)
17636 {
17637   int64x2x2_t ret;
17638   __builtin_aarch64_simd_oi __o;
17639   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
17640   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17641   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17642   return ret;
17643 }
17644
17645 __extension__ extern __inline uint8x16x2_t
17646 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17647 vld2q_u8 (const uint8_t * __a)
17648 {
17649   uint8x16x2_t ret;
17650   __builtin_aarch64_simd_oi __o;
17651   __o = __builtin_aarch64_ld2v16qi ((const __builtin_aarch64_simd_qi *) __a);
17652   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
17653   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
17654   return ret;
17655 }
17656
17657 __extension__ extern __inline uint16x8x2_t
17658 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17659 vld2q_u16 (const uint16_t * __a)
17660 {
17661   uint16x8x2_t ret;
17662   __builtin_aarch64_simd_oi __o;
17663   __o = __builtin_aarch64_ld2v8hi ((const __builtin_aarch64_simd_hi *) __a);
17664   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
17665   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
17666   return ret;
17667 }
17668
17669 __extension__ extern __inline uint32x4x2_t
17670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17671 vld2q_u32 (const uint32_t * __a)
17672 {
17673   uint32x4x2_t ret;
17674   __builtin_aarch64_simd_oi __o;
17675   __o = __builtin_aarch64_ld2v4si ((const __builtin_aarch64_simd_si *) __a);
17676   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
17677   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
17678   return ret;
17679 }
17680
17681 __extension__ extern __inline uint64x2x2_t
17682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17683 vld2q_u64 (const uint64_t * __a)
17684 {
17685   uint64x2x2_t ret;
17686   __builtin_aarch64_simd_oi __o;
17687   __o = __builtin_aarch64_ld2v2di ((const __builtin_aarch64_simd_di *) __a);
17688   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
17689   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
17690   return ret;
17691 }
17692
17693 __extension__ extern __inline float16x8x2_t
17694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17695 vld2q_f16 (const float16_t * __a)
17696 {
17697   float16x8x2_t ret;
17698   __builtin_aarch64_simd_oi __o;
17699   __o = __builtin_aarch64_ld2v8hf (__a);
17700   ret.val[0] = __builtin_aarch64_get_qregoiv8hf (__o, 0);
17701   ret.val[1] = __builtin_aarch64_get_qregoiv8hf (__o, 1);
17702   return ret;
17703 }
17704
17705 __extension__ extern __inline float32x4x2_t
17706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17707 vld2q_f32 (const float32_t * __a)
17708 {
17709   float32x4x2_t ret;
17710   __builtin_aarch64_simd_oi __o;
17711   __o = __builtin_aarch64_ld2v4sf ((const __builtin_aarch64_simd_sf *) __a);
17712   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
17713   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
17714   return ret;
17715 }
17716
17717 __extension__ extern __inline float64x2x2_t
17718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17719 vld2q_f64 (const float64_t * __a)
17720 {
17721   float64x2x2_t ret;
17722   __builtin_aarch64_simd_oi __o;
17723   __o = __builtin_aarch64_ld2v2df ((const __builtin_aarch64_simd_df *) __a);
17724   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
17725   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
17726   return ret;
17727 }
17728
17729 __extension__ extern __inline int64x1x3_t
17730 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17731 vld3_s64 (const int64_t * __a)
17732 {
17733   int64x1x3_t ret;
17734   __builtin_aarch64_simd_ci __o;
17735   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
17736   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17737   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17738   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17739   return ret;
17740 }
17741
17742 __extension__ extern __inline uint64x1x3_t
17743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17744 vld3_u64 (const uint64_t * __a)
17745 {
17746   uint64x1x3_t ret;
17747   __builtin_aarch64_simd_ci __o;
17748   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
17749   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
17750   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
17751   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
17752   return ret;
17753 }
17754
17755 __extension__ extern __inline float64x1x3_t
17756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17757 vld3_f64 (const float64_t * __a)
17758 {
17759   float64x1x3_t ret;
17760   __builtin_aarch64_simd_ci __o;
17761   __o = __builtin_aarch64_ld3df ((const __builtin_aarch64_simd_df *) __a);
17762   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 0)};
17763   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 1)};
17764   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 2)};
17765   return ret;
17766 }
17767
17768 __extension__ extern __inline int8x8x3_t
17769 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17770 vld3_s8 (const int8_t * __a)
17771 {
17772   int8x8x3_t ret;
17773   __builtin_aarch64_simd_ci __o;
17774   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17775   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
17776   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
17777   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
17778   return ret;
17779 }
17780
17781 __extension__ extern __inline poly8x8x3_t
17782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17783 vld3_p8 (const poly8_t * __a)
17784 {
17785   poly8x8x3_t ret;
17786   __builtin_aarch64_simd_ci __o;
17787   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17788   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
17789   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
17790   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
17791   return ret;
17792 }
17793
17794 __extension__ extern __inline int16x4x3_t
17795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17796 vld3_s16 (const int16_t * __a)
17797 {
17798   int16x4x3_t ret;
17799   __builtin_aarch64_simd_ci __o;
17800   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17801   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
17802   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
17803   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
17804   return ret;
17805 }
17806
17807 __extension__ extern __inline poly16x4x3_t
17808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17809 vld3_p16 (const poly16_t * __a)
17810 {
17811   poly16x4x3_t ret;
17812   __builtin_aarch64_simd_ci __o;
17813   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17814   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
17815   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
17816   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
17817   return ret;
17818 }
17819
17820 __extension__ extern __inline int32x2x3_t
17821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17822 vld3_s32 (const int32_t * __a)
17823 {
17824   int32x2x3_t ret;
17825   __builtin_aarch64_simd_ci __o;
17826   __o = __builtin_aarch64_ld3v2si ((const __builtin_aarch64_simd_si *) __a);
17827   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
17828   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
17829   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
17830   return ret;
17831 }
17832
17833 __extension__ extern __inline uint8x8x3_t
17834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17835 vld3_u8 (const uint8_t * __a)
17836 {
17837   uint8x8x3_t ret;
17838   __builtin_aarch64_simd_ci __o;
17839   __o = __builtin_aarch64_ld3v8qi ((const __builtin_aarch64_simd_qi *) __a);
17840   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
17841   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
17842   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
17843   return ret;
17844 }
17845
17846 __extension__ extern __inline uint16x4x3_t
17847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17848 vld3_u16 (const uint16_t * __a)
17849 {
17850   uint16x4x3_t ret;
17851   __builtin_aarch64_simd_ci __o;
17852   __o = __builtin_aarch64_ld3v4hi ((const __builtin_aarch64_simd_hi *) __a);
17853   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
17854   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
17855   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
17856   return ret;
17857 }
17858
17859 __extension__ extern __inline uint32x2x3_t
17860 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17861 vld3_u32 (const uint32_t * __a)
17862 {
17863   uint32x2x3_t ret;
17864   __builtin_aarch64_simd_ci __o;
17865   __o = __builtin_aarch64_ld3v2si ((const __builtin_aarch64_simd_si *) __a);
17866   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
17867   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
17868   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
17869   return ret;
17870 }
17871
17872 __extension__ extern __inline float16x4x3_t
17873 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17874 vld3_f16 (const float16_t * __a)
17875 {
17876   float16x4x3_t ret;
17877   __builtin_aarch64_simd_ci __o;
17878   __o = __builtin_aarch64_ld3v4hf (__a);
17879   ret.val[0] = __builtin_aarch64_get_dregciv4hf (__o, 0);
17880   ret.val[1] = __builtin_aarch64_get_dregciv4hf (__o, 1);
17881   ret.val[2] = __builtin_aarch64_get_dregciv4hf (__o, 2);
17882   return ret;
17883 }
17884
17885 __extension__ extern __inline float32x2x3_t
17886 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17887 vld3_f32 (const float32_t * __a)
17888 {
17889   float32x2x3_t ret;
17890   __builtin_aarch64_simd_ci __o;
17891   __o = __builtin_aarch64_ld3v2sf ((const __builtin_aarch64_simd_sf *) __a);
17892   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 0);
17893   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 1);
17894   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 2);
17895   return ret;
17896 }
17897
17898 __extension__ extern __inline poly64x1x3_t
17899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17900 vld3_p64 (const poly64_t * __a)
17901 {
17902   poly64x1x3_t ret;
17903   __builtin_aarch64_simd_ci __o;
17904   __o = __builtin_aarch64_ld3di ((const __builtin_aarch64_simd_di *) __a);
17905   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 0);
17906   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 1);
17907   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 2);
17908   return ret;
17909 }
17910
17911 __extension__ extern __inline int8x16x3_t
17912 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17913 vld3q_s8 (const int8_t * __a)
17914 {
17915   int8x16x3_t ret;
17916   __builtin_aarch64_simd_ci __o;
17917   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17918   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
17919   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
17920   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
17921   return ret;
17922 }
17923
17924 __extension__ extern __inline poly8x16x3_t
17925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17926 vld3q_p8 (const poly8_t * __a)
17927 {
17928   poly8x16x3_t ret;
17929   __builtin_aarch64_simd_ci __o;
17930   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17931   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
17932   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
17933   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
17934   return ret;
17935 }
17936
17937 __extension__ extern __inline int16x8x3_t
17938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17939 vld3q_s16 (const int16_t * __a)
17940 {
17941   int16x8x3_t ret;
17942   __builtin_aarch64_simd_ci __o;
17943   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
17944   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
17945   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
17946   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
17947   return ret;
17948 }
17949
17950 __extension__ extern __inline poly16x8x3_t
17951 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17952 vld3q_p16 (const poly16_t * __a)
17953 {
17954   poly16x8x3_t ret;
17955   __builtin_aarch64_simd_ci __o;
17956   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
17957   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
17958   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
17959   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
17960   return ret;
17961 }
17962
17963 __extension__ extern __inline int32x4x3_t
17964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17965 vld3q_s32 (const int32_t * __a)
17966 {
17967   int32x4x3_t ret;
17968   __builtin_aarch64_simd_ci __o;
17969   __o = __builtin_aarch64_ld3v4si ((const __builtin_aarch64_simd_si *) __a);
17970   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
17971   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
17972   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
17973   return ret;
17974 }
17975
17976 __extension__ extern __inline int64x2x3_t
17977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17978 vld3q_s64 (const int64_t * __a)
17979 {
17980   int64x2x3_t ret;
17981   __builtin_aarch64_simd_ci __o;
17982   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
17983   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
17984   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
17985   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
17986   return ret;
17987 }
17988
17989 __extension__ extern __inline uint8x16x3_t
17990 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
17991 vld3q_u8 (const uint8_t * __a)
17992 {
17993   uint8x16x3_t ret;
17994   __builtin_aarch64_simd_ci __o;
17995   __o = __builtin_aarch64_ld3v16qi ((const __builtin_aarch64_simd_qi *) __a);
17996   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
17997   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
17998   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
17999   return ret;
18000 }
18001
18002 __extension__ extern __inline uint16x8x3_t
18003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18004 vld3q_u16 (const uint16_t * __a)
18005 {
18006   uint16x8x3_t ret;
18007   __builtin_aarch64_simd_ci __o;
18008   __o = __builtin_aarch64_ld3v8hi ((const __builtin_aarch64_simd_hi *) __a);
18009   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
18010   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
18011   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
18012   return ret;
18013 }
18014
18015 __extension__ extern __inline uint32x4x3_t
18016 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18017 vld3q_u32 (const uint32_t * __a)
18018 {
18019   uint32x4x3_t ret;
18020   __builtin_aarch64_simd_ci __o;
18021   __o = __builtin_aarch64_ld3v4si ((const __builtin_aarch64_simd_si *) __a);
18022   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
18023   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
18024   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
18025   return ret;
18026 }
18027
18028 __extension__ extern __inline uint64x2x3_t
18029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18030 vld3q_u64 (const uint64_t * __a)
18031 {
18032   uint64x2x3_t ret;
18033   __builtin_aarch64_simd_ci __o;
18034   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
18035   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
18036   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
18037   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
18038   return ret;
18039 }
18040
18041 __extension__ extern __inline float16x8x3_t
18042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18043 vld3q_f16 (const float16_t * __a)
18044 {
18045   float16x8x3_t ret;
18046   __builtin_aarch64_simd_ci __o;
18047   __o = __builtin_aarch64_ld3v8hf (__a);
18048   ret.val[0] = __builtin_aarch64_get_qregciv8hf (__o, 0);
18049   ret.val[1] = __builtin_aarch64_get_qregciv8hf (__o, 1);
18050   ret.val[2] = __builtin_aarch64_get_qregciv8hf (__o, 2);
18051   return ret;
18052 }
18053
18054 __extension__ extern __inline float32x4x3_t
18055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18056 vld3q_f32 (const float32_t * __a)
18057 {
18058   float32x4x3_t ret;
18059   __builtin_aarch64_simd_ci __o;
18060   __o = __builtin_aarch64_ld3v4sf ((const __builtin_aarch64_simd_sf *) __a);
18061   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 0);
18062   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 1);
18063   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 2);
18064   return ret;
18065 }
18066
18067 __extension__ extern __inline float64x2x3_t
18068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18069 vld3q_f64 (const float64_t * __a)
18070 {
18071   float64x2x3_t ret;
18072   __builtin_aarch64_simd_ci __o;
18073   __o = __builtin_aarch64_ld3v2df ((const __builtin_aarch64_simd_df *) __a);
18074   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 0);
18075   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 1);
18076   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 2);
18077   return ret;
18078 }
18079
18080 __extension__ extern __inline poly64x2x3_t
18081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18082 vld3q_p64 (const poly64_t * __a)
18083 {
18084   poly64x2x3_t ret;
18085   __builtin_aarch64_simd_ci __o;
18086   __o = __builtin_aarch64_ld3v2di ((const __builtin_aarch64_simd_di *) __a);
18087   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 0);
18088   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 1);
18089   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 2);
18090   return ret;
18091 }
18092
18093 __extension__ extern __inline int64x1x4_t
18094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18095 vld4_s64 (const int64_t * __a)
18096 {
18097   int64x1x4_t ret;
18098   __builtin_aarch64_simd_xi __o;
18099   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
18100   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
18101   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
18102   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
18103   ret.val[3] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
18104   return ret;
18105 }
18106
18107 __extension__ extern __inline uint64x1x4_t
18108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18109 vld4_u64 (const uint64_t * __a)
18110 {
18111   uint64x1x4_t ret;
18112   __builtin_aarch64_simd_xi __o;
18113   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
18114   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
18115   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
18116   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
18117   ret.val[3] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
18118   return ret;
18119 }
18120
18121 __extension__ extern __inline float64x1x4_t
18122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18123 vld4_f64 (const float64_t * __a)
18124 {
18125   float64x1x4_t ret;
18126   __builtin_aarch64_simd_xi __o;
18127   __o = __builtin_aarch64_ld4df ((const __builtin_aarch64_simd_df *) __a);
18128   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 0)};
18129   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 1)};
18130   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 2)};
18131   ret.val[3] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 3)};
18132   return ret;
18133 }
18134
18135 __extension__ extern __inline int8x8x4_t
18136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18137 vld4_s8 (const int8_t * __a)
18138 {
18139   int8x8x4_t ret;
18140   __builtin_aarch64_simd_xi __o;
18141   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
18142   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
18143   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
18144   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
18145   ret.val[3] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
18146   return ret;
18147 }
18148
18149 __extension__ extern __inline poly8x8x4_t
18150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18151 vld4_p8 (const poly8_t * __a)
18152 {
18153   poly8x8x4_t ret;
18154   __builtin_aarch64_simd_xi __o;
18155   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
18156   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
18157   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
18158   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
18159   ret.val[3] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
18160   return ret;
18161 }
18162
18163 __extension__ extern __inline int16x4x4_t
18164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18165 vld4_s16 (const int16_t * __a)
18166 {
18167   int16x4x4_t ret;
18168   __builtin_aarch64_simd_xi __o;
18169   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
18170   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
18171   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
18172   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
18173   ret.val[3] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
18174   return ret;
18175 }
18176
18177 __extension__ extern __inline poly16x4x4_t
18178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18179 vld4_p16 (const poly16_t * __a)
18180 {
18181   poly16x4x4_t ret;
18182   __builtin_aarch64_simd_xi __o;
18183   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
18184   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
18185   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
18186   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
18187   ret.val[3] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
18188   return ret;
18189 }
18190
18191 __extension__ extern __inline int32x2x4_t
18192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18193 vld4_s32 (const int32_t * __a)
18194 {
18195   int32x2x4_t ret;
18196   __builtin_aarch64_simd_xi __o;
18197   __o = __builtin_aarch64_ld4v2si ((const __builtin_aarch64_simd_si *) __a);
18198   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
18199   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
18200   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
18201   ret.val[3] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
18202   return ret;
18203 }
18204
18205 __extension__ extern __inline uint8x8x4_t
18206 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18207 vld4_u8 (const uint8_t * __a)
18208 {
18209   uint8x8x4_t ret;
18210   __builtin_aarch64_simd_xi __o;
18211   __o = __builtin_aarch64_ld4v8qi ((const __builtin_aarch64_simd_qi *) __a);
18212   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
18213   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
18214   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
18215   ret.val[3] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
18216   return ret;
18217 }
18218
18219 __extension__ extern __inline uint16x4x4_t
18220 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18221 vld4_u16 (const uint16_t * __a)
18222 {
18223   uint16x4x4_t ret;
18224   __builtin_aarch64_simd_xi __o;
18225   __o = __builtin_aarch64_ld4v4hi ((const __builtin_aarch64_simd_hi *) __a);
18226   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
18227   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
18228   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
18229   ret.val[3] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
18230   return ret;
18231 }
18232
18233 __extension__ extern __inline uint32x2x4_t
18234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18235 vld4_u32 (const uint32_t * __a)
18236 {
18237   uint32x2x4_t ret;
18238   __builtin_aarch64_simd_xi __o;
18239   __o = __builtin_aarch64_ld4v2si ((const __builtin_aarch64_simd_si *) __a);
18240   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
18241   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
18242   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
18243   ret.val[3] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
18244   return ret;
18245 }
18246
18247 __extension__ extern __inline float16x4x4_t
18248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18249 vld4_f16 (const float16_t * __a)
18250 {
18251   float16x4x4_t ret;
18252   __builtin_aarch64_simd_xi __o;
18253   __o = __builtin_aarch64_ld4v4hf (__a);
18254   ret.val[0] = __builtin_aarch64_get_dregxiv4hf (__o, 0);
18255   ret.val[1] = __builtin_aarch64_get_dregxiv4hf (__o, 1);
18256   ret.val[2] = __builtin_aarch64_get_dregxiv4hf (__o, 2);
18257   ret.val[3] = __builtin_aarch64_get_dregxiv4hf (__o, 3);
18258   return ret;
18259 }
18260
18261 __extension__ extern __inline float32x2x4_t
18262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18263 vld4_f32 (const float32_t * __a)
18264 {
18265   float32x2x4_t ret;
18266   __builtin_aarch64_simd_xi __o;
18267   __o = __builtin_aarch64_ld4v2sf ((const __builtin_aarch64_simd_sf *) __a);
18268   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 0);
18269   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 1);
18270   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 2);
18271   ret.val[3] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 3);
18272   return ret;
18273 }
18274
18275 __extension__ extern __inline poly64x1x4_t
18276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18277 vld4_p64 (const poly64_t * __a)
18278 {
18279   poly64x1x4_t  ret;
18280   __builtin_aarch64_simd_xi __o;
18281   __o = __builtin_aarch64_ld4di ((const __builtin_aarch64_simd_di *) __a);
18282   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 0);
18283   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 1);
18284   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 2);
18285   ret.val[3] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 3);
18286   return ret;
18287 }
18288
18289 __extension__ extern __inline int8x16x4_t
18290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18291 vld4q_s8 (const int8_t * __a)
18292 {
18293   int8x16x4_t ret;
18294   __builtin_aarch64_simd_xi __o;
18295   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
18296   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
18297   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
18298   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
18299   ret.val[3] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
18300   return ret;
18301 }
18302
18303 __extension__ extern __inline poly8x16x4_t
18304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18305 vld4q_p8 (const poly8_t * __a)
18306 {
18307   poly8x16x4_t ret;
18308   __builtin_aarch64_simd_xi __o;
18309   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
18310   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
18311   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
18312   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
18313   ret.val[3] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
18314   return ret;
18315 }
18316
18317 __extension__ extern __inline int16x8x4_t
18318 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18319 vld4q_s16 (const int16_t * __a)
18320 {
18321   int16x8x4_t ret;
18322   __builtin_aarch64_simd_xi __o;
18323   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
18324   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
18325   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
18326   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
18327   ret.val[3] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
18328   return ret;
18329 }
18330
18331 __extension__ extern __inline poly16x8x4_t
18332 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18333 vld4q_p16 (const poly16_t * __a)
18334 {
18335   poly16x8x4_t ret;
18336   __builtin_aarch64_simd_xi __o;
18337   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
18338   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
18339   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
18340   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
18341   ret.val[3] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
18342   return ret;
18343 }
18344
18345 __extension__ extern __inline int32x4x4_t
18346 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18347 vld4q_s32 (const int32_t * __a)
18348 {
18349   int32x4x4_t ret;
18350   __builtin_aarch64_simd_xi __o;
18351   __o = __builtin_aarch64_ld4v4si ((const __builtin_aarch64_simd_si *) __a);
18352   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
18353   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
18354   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
18355   ret.val[3] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
18356   return ret;
18357 }
18358
18359 __extension__ extern __inline int64x2x4_t
18360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18361 vld4q_s64 (const int64_t * __a)
18362 {
18363   int64x2x4_t ret;
18364   __builtin_aarch64_simd_xi __o;
18365   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
18366   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
18367   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
18368   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
18369   ret.val[3] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
18370   return ret;
18371 }
18372
18373 __extension__ extern __inline uint8x16x4_t
18374 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18375 vld4q_u8 (const uint8_t * __a)
18376 {
18377   uint8x16x4_t ret;
18378   __builtin_aarch64_simd_xi __o;
18379   __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
18380   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
18381   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
18382   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
18383   ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
18384   return ret;
18385 }
18386
18387 __extension__ extern __inline uint16x8x4_t
18388 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18389 vld4q_u16 (const uint16_t * __a)
18390 {
18391   uint16x8x4_t ret;
18392   __builtin_aarch64_simd_xi __o;
18393   __o = __builtin_aarch64_ld4v8hi ((const __builtin_aarch64_simd_hi *) __a);
18394   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
18395   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
18396   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
18397   ret.val[3] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
18398   return ret;
18399 }
18400
18401 __extension__ extern __inline uint32x4x4_t
18402 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18403 vld4q_u32 (const uint32_t * __a)
18404 {
18405   uint32x4x4_t ret;
18406   __builtin_aarch64_simd_xi __o;
18407   __o = __builtin_aarch64_ld4v4si ((const __builtin_aarch64_simd_si *) __a);
18408   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
18409   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
18410   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
18411   ret.val[3] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
18412   return ret;
18413 }
18414
18415 __extension__ extern __inline uint64x2x4_t
18416 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18417 vld4q_u64 (const uint64_t * __a)
18418 {
18419   uint64x2x4_t ret;
18420   __builtin_aarch64_simd_xi __o;
18421   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
18422   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
18423   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
18424   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
18425   ret.val[3] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
18426   return ret;
18427 }
18428
18429 __extension__ extern __inline float16x8x4_t
18430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18431 vld4q_f16 (const float16_t * __a)
18432 {
18433   float16x8x4_t ret;
18434   __builtin_aarch64_simd_xi __o;
18435   __o = __builtin_aarch64_ld4v8hf (__a);
18436   ret.val[0] = __builtin_aarch64_get_qregxiv8hf (__o, 0);
18437   ret.val[1] = __builtin_aarch64_get_qregxiv8hf (__o, 1);
18438   ret.val[2] = __builtin_aarch64_get_qregxiv8hf (__o, 2);
18439   ret.val[3] = __builtin_aarch64_get_qregxiv8hf (__o, 3);
18440   return ret;
18441 }
18442
18443 __extension__ extern __inline float32x4x4_t
18444 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18445 vld4q_f32 (const float32_t * __a)
18446 {
18447   float32x4x4_t ret;
18448   __builtin_aarch64_simd_xi __o;
18449   __o = __builtin_aarch64_ld4v4sf ((const __builtin_aarch64_simd_sf *) __a);
18450   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 0);
18451   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 1);
18452   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 2);
18453   ret.val[3] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 3);
18454   return ret;
18455 }
18456
18457 __extension__ extern __inline float64x2x4_t
18458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18459 vld4q_f64 (const float64_t * __a)
18460 {
18461   float64x2x4_t ret;
18462   __builtin_aarch64_simd_xi __o;
18463   __o = __builtin_aarch64_ld4v2df ((const __builtin_aarch64_simd_df *) __a);
18464   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 0);
18465   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 1);
18466   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 2);
18467   ret.val[3] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 3);
18468   return ret;
18469 }
18470
18471 __extension__ extern __inline poly64x2x4_t
18472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18473 vld4q_p64 (const poly64_t * __a)
18474 {
18475   poly64x2x4_t  ret;
18476   __builtin_aarch64_simd_xi __o;
18477   __o = __builtin_aarch64_ld4v2di ((const __builtin_aarch64_simd_di *) __a);
18478   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 0);
18479   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 1);
18480   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 2);
18481   ret.val[3] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 3);
18482   return ret;
18483 }
18484
18485 /* vldn_dup */
18486
18487 __extension__ extern __inline int8x8x2_t
18488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18489 vld2_dup_s8 (const int8_t * __a)
18490 {
18491   int8x8x2_t ret;
18492   __builtin_aarch64_simd_oi __o;
18493   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18494   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18495   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18496   return ret;
18497 }
18498
18499 __extension__ extern __inline int16x4x2_t
18500 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18501 vld2_dup_s16 (const int16_t * __a)
18502 {
18503   int16x4x2_t ret;
18504   __builtin_aarch64_simd_oi __o;
18505   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18506   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18507   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18508   return ret;
18509 }
18510
18511 __extension__ extern __inline int32x2x2_t
18512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18513 vld2_dup_s32 (const int32_t * __a)
18514 {
18515   int32x2x2_t ret;
18516   __builtin_aarch64_simd_oi __o;
18517   __o = __builtin_aarch64_ld2rv2si ((const __builtin_aarch64_simd_si *) __a);
18518   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
18519   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
18520   return ret;
18521 }
18522
18523 __extension__ extern __inline float16x4x2_t
18524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18525 vld2_dup_f16 (const float16_t * __a)
18526 {
18527   float16x4x2_t ret;
18528   __builtin_aarch64_simd_oi __o;
18529   __o = __builtin_aarch64_ld2rv4hf ((const __builtin_aarch64_simd_hf *) __a);
18530   ret.val[0] = __builtin_aarch64_get_dregoiv4hf (__o, 0);
18531   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregoiv4hf (__o, 1);
18532   return ret;
18533 }
18534
18535 __extension__ extern __inline float32x2x2_t
18536 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18537 vld2_dup_f32 (const float32_t * __a)
18538 {
18539   float32x2x2_t ret;
18540   __builtin_aarch64_simd_oi __o;
18541   __o = __builtin_aarch64_ld2rv2sf ((const __builtin_aarch64_simd_sf *) __a);
18542   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 0);
18543   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregoiv2sf (__o, 1);
18544   return ret;
18545 }
18546
18547 __extension__ extern __inline float64x1x2_t
18548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18549 vld2_dup_f64 (const float64_t * __a)
18550 {
18551   float64x1x2_t ret;
18552   __builtin_aarch64_simd_oi __o;
18553   __o = __builtin_aarch64_ld2rdf ((const __builtin_aarch64_simd_df *) __a);
18554   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 0)};
18555   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregoidf (__o, 1)};
18556   return ret;
18557 }
18558
18559 __extension__ extern __inline uint8x8x2_t
18560 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18561 vld2_dup_u8 (const uint8_t * __a)
18562 {
18563   uint8x8x2_t ret;
18564   __builtin_aarch64_simd_oi __o;
18565   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18566   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18567   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18568   return ret;
18569 }
18570
18571 __extension__ extern __inline uint16x4x2_t
18572 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18573 vld2_dup_u16 (const uint16_t * __a)
18574 {
18575   uint16x4x2_t ret;
18576   __builtin_aarch64_simd_oi __o;
18577   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18578   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18579   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18580   return ret;
18581 }
18582
18583 __extension__ extern __inline uint32x2x2_t
18584 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18585 vld2_dup_u32 (const uint32_t * __a)
18586 {
18587   uint32x2x2_t ret;
18588   __builtin_aarch64_simd_oi __o;
18589   __o = __builtin_aarch64_ld2rv2si ((const __builtin_aarch64_simd_si *) __a);
18590   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 0);
18591   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregoiv2si (__o, 1);
18592   return ret;
18593 }
18594
18595 __extension__ extern __inline poly8x8x2_t
18596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18597 vld2_dup_p8 (const poly8_t * __a)
18598 {
18599   poly8x8x2_t ret;
18600   __builtin_aarch64_simd_oi __o;
18601   __o = __builtin_aarch64_ld2rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18602   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 0);
18603   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregoiv8qi (__o, 1);
18604   return ret;
18605 }
18606
18607 __extension__ extern __inline poly16x4x2_t
18608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18609 vld2_dup_p16 (const poly16_t * __a)
18610 {
18611   poly16x4x2_t ret;
18612   __builtin_aarch64_simd_oi __o;
18613   __o = __builtin_aarch64_ld2rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18614   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 0);
18615   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregoiv4hi (__o, 1);
18616   return ret;
18617 }
18618
18619 __extension__ extern __inline poly64x1x2_t
18620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18621 vld2_dup_p64 (const poly64_t * __a)
18622 {
18623   poly64x1x2_t ret;
18624   __builtin_aarch64_simd_oi __o;
18625   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18626   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 0);
18627   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregoidi_pss (__o, 1);
18628   return ret;
18629 }
18630
18631
18632 __extension__ extern __inline int64x1x2_t
18633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18634 vld2_dup_s64 (const int64_t * __a)
18635 {
18636   int64x1x2_t ret;
18637   __builtin_aarch64_simd_oi __o;
18638   __o = __builtin_aarch64_ld2rdi ((const __builtin_aarch64_simd_di *) __a);
18639   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
18640   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
18641   return ret;
18642 }
18643
18644 __extension__ extern __inline uint64x1x2_t
18645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18646 vld2_dup_u64 (const uint64_t * __a)
18647 {
18648   uint64x1x2_t ret;
18649   __builtin_aarch64_simd_oi __o;
18650   __o = __builtin_aarch64_ld2rdi ((const __builtin_aarch64_simd_di *) __a);
18651   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 0);
18652   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregoidi (__o, 1);
18653   return ret;
18654 }
18655
18656 __extension__ extern __inline int8x16x2_t
18657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18658 vld2q_dup_s8 (const int8_t * __a)
18659 {
18660   int8x16x2_t ret;
18661   __builtin_aarch64_simd_oi __o;
18662   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
18663   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18664   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18665   return ret;
18666 }
18667
18668 __extension__ extern __inline poly8x16x2_t
18669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18670 vld2q_dup_p8 (const poly8_t * __a)
18671 {
18672   poly8x16x2_t ret;
18673   __builtin_aarch64_simd_oi __o;
18674   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
18675   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18676   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18677   return ret;
18678 }
18679
18680 __extension__ extern __inline int16x8x2_t
18681 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18682 vld2q_dup_s16 (const int16_t * __a)
18683 {
18684   int16x8x2_t ret;
18685   __builtin_aarch64_simd_oi __o;
18686   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
18687   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18688   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18689   return ret;
18690 }
18691
18692 __extension__ extern __inline poly16x8x2_t
18693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18694 vld2q_dup_p16 (const poly16_t * __a)
18695 {
18696   poly16x8x2_t ret;
18697   __builtin_aarch64_simd_oi __o;
18698   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
18699   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18700   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18701   return ret;
18702 }
18703
18704 __extension__ extern __inline int32x4x2_t
18705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18706 vld2q_dup_s32 (const int32_t * __a)
18707 {
18708   int32x4x2_t ret;
18709   __builtin_aarch64_simd_oi __o;
18710   __o = __builtin_aarch64_ld2rv4si ((const __builtin_aarch64_simd_si *) __a);
18711   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
18712   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
18713   return ret;
18714 }
18715
18716 __extension__ extern __inline int64x2x2_t
18717 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18718 vld2q_dup_s64 (const int64_t * __a)
18719 {
18720   int64x2x2_t ret;
18721   __builtin_aarch64_simd_oi __o;
18722   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18723   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
18724   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
18725   return ret;
18726 }
18727
18728 __extension__ extern __inline uint8x16x2_t
18729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18730 vld2q_dup_u8 (const uint8_t * __a)
18731 {
18732   uint8x16x2_t ret;
18733   __builtin_aarch64_simd_oi __o;
18734   __o = __builtin_aarch64_ld2rv16qi ((const __builtin_aarch64_simd_qi *) __a);
18735   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 0);
18736   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregoiv16qi (__o, 1);
18737   return ret;
18738 }
18739
18740 __extension__ extern __inline uint16x8x2_t
18741 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18742 vld2q_dup_u16 (const uint16_t * __a)
18743 {
18744   uint16x8x2_t ret;
18745   __builtin_aarch64_simd_oi __o;
18746   __o = __builtin_aarch64_ld2rv8hi ((const __builtin_aarch64_simd_hi *) __a);
18747   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 0);
18748   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregoiv8hi (__o, 1);
18749   return ret;
18750 }
18751
18752 __extension__ extern __inline uint32x4x2_t
18753 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18754 vld2q_dup_u32 (const uint32_t * __a)
18755 {
18756   uint32x4x2_t ret;
18757   __builtin_aarch64_simd_oi __o;
18758   __o = __builtin_aarch64_ld2rv4si ((const __builtin_aarch64_simd_si *) __a);
18759   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 0);
18760   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregoiv4si (__o, 1);
18761   return ret;
18762 }
18763
18764 __extension__ extern __inline uint64x2x2_t
18765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18766 vld2q_dup_u64 (const uint64_t * __a)
18767 {
18768   uint64x2x2_t ret;
18769   __builtin_aarch64_simd_oi __o;
18770   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18771   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 0);
18772   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregoiv2di (__o, 1);
18773   return ret;
18774 }
18775
18776 __extension__ extern __inline float16x8x2_t
18777 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18778 vld2q_dup_f16 (const float16_t * __a)
18779 {
18780   float16x8x2_t ret;
18781   __builtin_aarch64_simd_oi __o;
18782   __o = __builtin_aarch64_ld2rv8hf ((const __builtin_aarch64_simd_hf *) __a);
18783   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregoiv8hf (__o, 0);
18784   ret.val[1] = __builtin_aarch64_get_qregoiv8hf (__o, 1);
18785   return ret;
18786 }
18787
18788 __extension__ extern __inline float32x4x2_t
18789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18790 vld2q_dup_f32 (const float32_t * __a)
18791 {
18792   float32x4x2_t ret;
18793   __builtin_aarch64_simd_oi __o;
18794   __o = __builtin_aarch64_ld2rv4sf ((const __builtin_aarch64_simd_sf *) __a);
18795   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 0);
18796   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregoiv4sf (__o, 1);
18797   return ret;
18798 }
18799
18800 __extension__ extern __inline float64x2x2_t
18801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18802 vld2q_dup_f64 (const float64_t * __a)
18803 {
18804   float64x2x2_t ret;
18805   __builtin_aarch64_simd_oi __o;
18806   __o = __builtin_aarch64_ld2rv2df ((const __builtin_aarch64_simd_df *) __a);
18807   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 0);
18808   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregoiv2df (__o, 1);
18809   return ret;
18810 }
18811
18812 __extension__ extern __inline poly64x2x2_t
18813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18814 vld2q_dup_p64 (const poly64_t * __a)
18815 {
18816   poly64x2x2_t ret;
18817   __builtin_aarch64_simd_oi __o;
18818   __o = __builtin_aarch64_ld2rv2di ((const __builtin_aarch64_simd_di *) __a);
18819   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 0);
18820   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregoiv2di_pss (__o, 1);
18821   return ret;
18822 }
18823
18824 __extension__ extern __inline int64x1x3_t
18825 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18826 vld3_dup_s64 (const int64_t * __a)
18827 {
18828   int64x1x3_t ret;
18829   __builtin_aarch64_simd_ci __o;
18830   __o = __builtin_aarch64_ld3rdi ((const __builtin_aarch64_simd_di *) __a);
18831   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
18832   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
18833   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
18834   return ret;
18835 }
18836
18837 __extension__ extern __inline uint64x1x3_t
18838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18839 vld3_dup_u64 (const uint64_t * __a)
18840 {
18841   uint64x1x3_t ret;
18842   __builtin_aarch64_simd_ci __o;
18843   __o = __builtin_aarch64_ld3rdi ((const __builtin_aarch64_simd_di *) __a);
18844   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 0);
18845   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 1);
18846   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregcidi (__o, 2);
18847   return ret;
18848 }
18849
18850 __extension__ extern __inline float64x1x3_t
18851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18852 vld3_dup_f64 (const float64_t * __a)
18853 {
18854   float64x1x3_t ret;
18855   __builtin_aarch64_simd_ci __o;
18856   __o = __builtin_aarch64_ld3rdf ((const __builtin_aarch64_simd_df *) __a);
18857   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 0)};
18858   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 1)};
18859   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregcidf (__o, 2)};
18860   return ret;
18861 }
18862
18863 __extension__ extern __inline int8x8x3_t
18864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18865 vld3_dup_s8 (const int8_t * __a)
18866 {
18867   int8x8x3_t ret;
18868   __builtin_aarch64_simd_ci __o;
18869   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18870   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18871   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18872   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18873   return ret;
18874 }
18875
18876 __extension__ extern __inline poly8x8x3_t
18877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18878 vld3_dup_p8 (const poly8_t * __a)
18879 {
18880   poly8x8x3_t ret;
18881   __builtin_aarch64_simd_ci __o;
18882   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18883   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18884   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18885   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18886   return ret;
18887 }
18888
18889 __extension__ extern __inline int16x4x3_t
18890 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18891 vld3_dup_s16 (const int16_t * __a)
18892 {
18893   int16x4x3_t ret;
18894   __builtin_aarch64_simd_ci __o;
18895   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18896   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18897   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18898   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18899   return ret;
18900 }
18901
18902 __extension__ extern __inline poly16x4x3_t
18903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18904 vld3_dup_p16 (const poly16_t * __a)
18905 {
18906   poly16x4x3_t ret;
18907   __builtin_aarch64_simd_ci __o;
18908   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18909   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18910   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18911   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18912   return ret;
18913 }
18914
18915 __extension__ extern __inline int32x2x3_t
18916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18917 vld3_dup_s32 (const int32_t * __a)
18918 {
18919   int32x2x3_t ret;
18920   __builtin_aarch64_simd_ci __o;
18921   __o = __builtin_aarch64_ld3rv2si ((const __builtin_aarch64_simd_si *) __a);
18922   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
18923   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
18924   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
18925   return ret;
18926 }
18927
18928 __extension__ extern __inline uint8x8x3_t
18929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18930 vld3_dup_u8 (const uint8_t * __a)
18931 {
18932   uint8x8x3_t ret;
18933   __builtin_aarch64_simd_ci __o;
18934   __o = __builtin_aarch64_ld3rv8qi ((const __builtin_aarch64_simd_qi *) __a);
18935   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 0);
18936   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 1);
18937   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregciv8qi (__o, 2);
18938   return ret;
18939 }
18940
18941 __extension__ extern __inline uint16x4x3_t
18942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18943 vld3_dup_u16 (const uint16_t * __a)
18944 {
18945   uint16x4x3_t ret;
18946   __builtin_aarch64_simd_ci __o;
18947   __o = __builtin_aarch64_ld3rv4hi ((const __builtin_aarch64_simd_hi *) __a);
18948   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 0);
18949   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 1);
18950   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregciv4hi (__o, 2);
18951   return ret;
18952 }
18953
18954 __extension__ extern __inline uint32x2x3_t
18955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18956 vld3_dup_u32 (const uint32_t * __a)
18957 {
18958   uint32x2x3_t ret;
18959   __builtin_aarch64_simd_ci __o;
18960   __o = __builtin_aarch64_ld3rv2si ((const __builtin_aarch64_simd_si *) __a);
18961   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 0);
18962   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 1);
18963   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregciv2si (__o, 2);
18964   return ret;
18965 }
18966
18967 __extension__ extern __inline float16x4x3_t
18968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18969 vld3_dup_f16 (const float16_t * __a)
18970 {
18971   float16x4x3_t ret;
18972   __builtin_aarch64_simd_ci __o;
18973   __o = __builtin_aarch64_ld3rv4hf ((const __builtin_aarch64_simd_hf *) __a);
18974   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 0);
18975   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 1);
18976   ret.val[2] = (float16x4_t) __builtin_aarch64_get_dregciv4hf (__o, 2);
18977   return ret;
18978 }
18979
18980 __extension__ extern __inline float32x2x3_t
18981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18982 vld3_dup_f32 (const float32_t * __a)
18983 {
18984   float32x2x3_t ret;
18985   __builtin_aarch64_simd_ci __o;
18986   __o = __builtin_aarch64_ld3rv2sf ((const __builtin_aarch64_simd_sf *) __a);
18987   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 0);
18988   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 1);
18989   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregciv2sf (__o, 2);
18990   return ret;
18991 }
18992
18993 __extension__ extern __inline poly64x1x3_t
18994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
18995 vld3_dup_p64 (const poly64_t * __a)
18996 {
18997   poly64x1x3_t ret;
18998   __builtin_aarch64_simd_ci __o;
18999   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19000   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 0);
19001   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 1);
19002   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregcidi_pss (__o, 2);
19003   return ret;
19004 }
19005
19006 __extension__ extern __inline int8x16x3_t
19007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19008 vld3q_dup_s8 (const int8_t * __a)
19009 {
19010   int8x16x3_t ret;
19011   __builtin_aarch64_simd_ci __o;
19012   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19013   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19014   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19015   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19016   return ret;
19017 }
19018
19019 __extension__ extern __inline poly8x16x3_t
19020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19021 vld3q_dup_p8 (const poly8_t * __a)
19022 {
19023   poly8x16x3_t ret;
19024   __builtin_aarch64_simd_ci __o;
19025   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19026   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19027   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19028   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19029   return ret;
19030 }
19031
19032 __extension__ extern __inline int16x8x3_t
19033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19034 vld3q_dup_s16 (const int16_t * __a)
19035 {
19036   int16x8x3_t ret;
19037   __builtin_aarch64_simd_ci __o;
19038   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19039   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
19040   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
19041   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
19042   return ret;
19043 }
19044
19045 __extension__ extern __inline poly16x8x3_t
19046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19047 vld3q_dup_p16 (const poly16_t * __a)
19048 {
19049   poly16x8x3_t ret;
19050   __builtin_aarch64_simd_ci __o;
19051   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19052   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
19053   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
19054   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
19055   return ret;
19056 }
19057
19058 __extension__ extern __inline int32x4x3_t
19059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19060 vld3q_dup_s32 (const int32_t * __a)
19061 {
19062   int32x4x3_t ret;
19063   __builtin_aarch64_simd_ci __o;
19064   __o = __builtin_aarch64_ld3rv4si ((const __builtin_aarch64_simd_si *) __a);
19065   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
19066   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
19067   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
19068   return ret;
19069 }
19070
19071 __extension__ extern __inline int64x2x3_t
19072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19073 vld3q_dup_s64 (const int64_t * __a)
19074 {
19075   int64x2x3_t ret;
19076   __builtin_aarch64_simd_ci __o;
19077   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19078   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
19079   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
19080   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
19081   return ret;
19082 }
19083
19084 __extension__ extern __inline uint8x16x3_t
19085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19086 vld3q_dup_u8 (const uint8_t * __a)
19087 {
19088   uint8x16x3_t ret;
19089   __builtin_aarch64_simd_ci __o;
19090   __o = __builtin_aarch64_ld3rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19091   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 0);
19092   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 1);
19093   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregciv16qi (__o, 2);
19094   return ret;
19095 }
19096
19097 __extension__ extern __inline uint16x8x3_t
19098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19099 vld3q_dup_u16 (const uint16_t * __a)
19100 {
19101   uint16x8x3_t ret;
19102   __builtin_aarch64_simd_ci __o;
19103   __o = __builtin_aarch64_ld3rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19104   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 0);
19105   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 1);
19106   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregciv8hi (__o, 2);
19107   return ret;
19108 }
19109
19110 __extension__ extern __inline uint32x4x3_t
19111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19112 vld3q_dup_u32 (const uint32_t * __a)
19113 {
19114   uint32x4x3_t ret;
19115   __builtin_aarch64_simd_ci __o;
19116   __o = __builtin_aarch64_ld3rv4si ((const __builtin_aarch64_simd_si *) __a);
19117   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 0);
19118   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 1);
19119   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregciv4si (__o, 2);
19120   return ret;
19121 }
19122
19123 __extension__ extern __inline uint64x2x3_t
19124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19125 vld3q_dup_u64 (const uint64_t * __a)
19126 {
19127   uint64x2x3_t ret;
19128   __builtin_aarch64_simd_ci __o;
19129   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19130   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 0);
19131   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 1);
19132   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregciv2di (__o, 2);
19133   return ret;
19134 }
19135
19136 __extension__ extern __inline float16x8x3_t
19137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19138 vld3q_dup_f16 (const float16_t * __a)
19139 {
19140   float16x8x3_t ret;
19141   __builtin_aarch64_simd_ci __o;
19142   __o = __builtin_aarch64_ld3rv8hf ((const __builtin_aarch64_simd_hf *) __a);
19143   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 0);
19144   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 1);
19145   ret.val[2] = (float16x8_t) __builtin_aarch64_get_qregciv8hf (__o, 2);
19146   return ret;
19147 }
19148
19149 __extension__ extern __inline float32x4x3_t
19150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19151 vld3q_dup_f32 (const float32_t * __a)
19152 {
19153   float32x4x3_t ret;
19154   __builtin_aarch64_simd_ci __o;
19155   __o = __builtin_aarch64_ld3rv4sf ((const __builtin_aarch64_simd_sf *) __a);
19156   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 0);
19157   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 1);
19158   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregciv4sf (__o, 2);
19159   return ret;
19160 }
19161
19162 __extension__ extern __inline float64x2x3_t
19163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19164 vld3q_dup_f64 (const float64_t * __a)
19165 {
19166   float64x2x3_t ret;
19167   __builtin_aarch64_simd_ci __o;
19168   __o = __builtin_aarch64_ld3rv2df ((const __builtin_aarch64_simd_df *) __a);
19169   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 0);
19170   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 1);
19171   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregciv2df (__o, 2);
19172   return ret;
19173 }
19174
19175 __extension__ extern __inline poly64x2x3_t
19176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19177 vld3q_dup_p64 (const poly64_t * __a)
19178 {
19179   poly64x2x3_t ret;
19180   __builtin_aarch64_simd_ci __o;
19181   __o = __builtin_aarch64_ld3rv2di ((const __builtin_aarch64_simd_di *) __a);
19182   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 0);
19183   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 1);
19184   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregciv2di_pss (__o, 2);
19185   return ret;
19186 }
19187
19188 __extension__ extern __inline int64x1x4_t
19189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19190 vld4_dup_s64 (const int64_t * __a)
19191 {
19192   int64x1x4_t ret;
19193   __builtin_aarch64_simd_xi __o;
19194   __o = __builtin_aarch64_ld4rdi ((const __builtin_aarch64_simd_di *) __a);
19195   ret.val[0] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
19196   ret.val[1] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
19197   ret.val[2] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
19198   ret.val[3] = (int64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
19199   return ret;
19200 }
19201
19202 __extension__ extern __inline uint64x1x4_t
19203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19204 vld4_dup_u64 (const uint64_t * __a)
19205 {
19206   uint64x1x4_t ret;
19207   __builtin_aarch64_simd_xi __o;
19208   __o = __builtin_aarch64_ld4rdi ((const __builtin_aarch64_simd_di *) __a);
19209   ret.val[0] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 0);
19210   ret.val[1] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 1);
19211   ret.val[2] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 2);
19212   ret.val[3] = (uint64x1_t) __builtin_aarch64_get_dregxidi (__o, 3);
19213   return ret;
19214 }
19215
19216 __extension__ extern __inline float64x1x4_t
19217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19218 vld4_dup_f64 (const float64_t * __a)
19219 {
19220   float64x1x4_t ret;
19221   __builtin_aarch64_simd_xi __o;
19222   __o = __builtin_aarch64_ld4rdf ((const __builtin_aarch64_simd_df *) __a);
19223   ret.val[0] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 0)};
19224   ret.val[1] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 1)};
19225   ret.val[2] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 2)};
19226   ret.val[3] = (float64x1_t) {__builtin_aarch64_get_dregxidf (__o, 3)};
19227   return ret;
19228 }
19229
19230 __extension__ extern __inline int8x8x4_t
19231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19232 vld4_dup_s8 (const int8_t * __a)
19233 {
19234   int8x8x4_t ret;
19235   __builtin_aarch64_simd_xi __o;
19236   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19237   ret.val[0] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19238   ret.val[1] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19239   ret.val[2] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19240   ret.val[3] = (int8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19241   return ret;
19242 }
19243
19244 __extension__ extern __inline poly8x8x4_t
19245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19246 vld4_dup_p8 (const poly8_t * __a)
19247 {
19248   poly8x8x4_t ret;
19249   __builtin_aarch64_simd_xi __o;
19250   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19251   ret.val[0] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19252   ret.val[1] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19253   ret.val[2] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19254   ret.val[3] = (poly8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19255   return ret;
19256 }
19257
19258 __extension__ extern __inline int16x4x4_t
19259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19260 vld4_dup_s16 (const int16_t * __a)
19261 {
19262   int16x4x4_t ret;
19263   __builtin_aarch64_simd_xi __o;
19264   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19265   ret.val[0] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19266   ret.val[1] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19267   ret.val[2] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19268   ret.val[3] = (int16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19269   return ret;
19270 }
19271
19272 __extension__ extern __inline poly16x4x4_t
19273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19274 vld4_dup_p16 (const poly16_t * __a)
19275 {
19276   poly16x4x4_t ret;
19277   __builtin_aarch64_simd_xi __o;
19278   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19279   ret.val[0] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19280   ret.val[1] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19281   ret.val[2] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19282   ret.val[3] = (poly16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19283   return ret;
19284 }
19285
19286 __extension__ extern __inline int32x2x4_t
19287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19288 vld4_dup_s32 (const int32_t * __a)
19289 {
19290   int32x2x4_t ret;
19291   __builtin_aarch64_simd_xi __o;
19292   __o = __builtin_aarch64_ld4rv2si ((const __builtin_aarch64_simd_si *) __a);
19293   ret.val[0] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
19294   ret.val[1] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
19295   ret.val[2] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
19296   ret.val[3] = (int32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
19297   return ret;
19298 }
19299
19300 __extension__ extern __inline uint8x8x4_t
19301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19302 vld4_dup_u8 (const uint8_t * __a)
19303 {
19304   uint8x8x4_t ret;
19305   __builtin_aarch64_simd_xi __o;
19306   __o = __builtin_aarch64_ld4rv8qi ((const __builtin_aarch64_simd_qi *) __a);
19307   ret.val[0] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 0);
19308   ret.val[1] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 1);
19309   ret.val[2] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 2);
19310   ret.val[3] = (uint8x8_t) __builtin_aarch64_get_dregxiv8qi (__o, 3);
19311   return ret;
19312 }
19313
19314 __extension__ extern __inline uint16x4x4_t
19315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19316 vld4_dup_u16 (const uint16_t * __a)
19317 {
19318   uint16x4x4_t ret;
19319   __builtin_aarch64_simd_xi __o;
19320   __o = __builtin_aarch64_ld4rv4hi ((const __builtin_aarch64_simd_hi *) __a);
19321   ret.val[0] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 0);
19322   ret.val[1] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 1);
19323   ret.val[2] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 2);
19324   ret.val[3] = (uint16x4_t) __builtin_aarch64_get_dregxiv4hi (__o, 3);
19325   return ret;
19326 }
19327
19328 __extension__ extern __inline uint32x2x4_t
19329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19330 vld4_dup_u32 (const uint32_t * __a)
19331 {
19332   uint32x2x4_t ret;
19333   __builtin_aarch64_simd_xi __o;
19334   __o = __builtin_aarch64_ld4rv2si ((const __builtin_aarch64_simd_si *) __a);
19335   ret.val[0] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 0);
19336   ret.val[1] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 1);
19337   ret.val[2] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 2);
19338   ret.val[3] = (uint32x2_t) __builtin_aarch64_get_dregxiv2si (__o, 3);
19339   return ret;
19340 }
19341
19342 __extension__ extern __inline float16x4x4_t
19343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19344 vld4_dup_f16 (const float16_t * __a)
19345 {
19346   float16x4x4_t ret;
19347   __builtin_aarch64_simd_xi __o;
19348   __o = __builtin_aarch64_ld4rv4hf ((const __builtin_aarch64_simd_hf *) __a);
19349   ret.val[0] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 0);
19350   ret.val[1] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 1);
19351   ret.val[2] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 2);
19352   ret.val[3] = (float16x4_t) __builtin_aarch64_get_dregxiv4hf (__o, 3);
19353   return ret;
19354 }
19355
19356 __extension__ extern __inline float32x2x4_t
19357 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19358 vld4_dup_f32 (const float32_t * __a)
19359 {
19360   float32x2x4_t ret;
19361   __builtin_aarch64_simd_xi __o;
19362   __o = __builtin_aarch64_ld4rv2sf ((const __builtin_aarch64_simd_sf *) __a);
19363   ret.val[0] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 0);
19364   ret.val[1] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 1);
19365   ret.val[2] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 2);
19366   ret.val[3] = (float32x2_t) __builtin_aarch64_get_dregxiv2sf (__o, 3);
19367   return ret;
19368 }
19369
19370 __extension__ extern __inline poly64x1x4_t
19371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19372 vld4_dup_p64 (const poly64_t * __a)
19373 {
19374   poly64x1x4_t ret;
19375   __builtin_aarch64_simd_xi __o;
19376   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19377   ret.val[0] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 0);
19378   ret.val[1] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 1);
19379   ret.val[2] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 2);
19380   ret.val[3] = (poly64x1_t) __builtin_aarch64_get_dregxidi_pss (__o, 3);
19381   return ret;
19382 }
19383
19384 __extension__ extern __inline int8x16x4_t
19385 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19386 vld4q_dup_s8 (const int8_t * __a)
19387 {
19388   int8x16x4_t ret;
19389   __builtin_aarch64_simd_xi __o;
19390   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19391   ret.val[0] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19392   ret.val[1] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19393   ret.val[2] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19394   ret.val[3] = (int8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19395   return ret;
19396 }
19397
19398 __extension__ extern __inline poly8x16x4_t
19399 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19400 vld4q_dup_p8 (const poly8_t * __a)
19401 {
19402   poly8x16x4_t ret;
19403   __builtin_aarch64_simd_xi __o;
19404   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19405   ret.val[0] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19406   ret.val[1] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19407   ret.val[2] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19408   ret.val[3] = (poly8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19409   return ret;
19410 }
19411
19412 __extension__ extern __inline int16x8x4_t
19413 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19414 vld4q_dup_s16 (const int16_t * __a)
19415 {
19416   int16x8x4_t ret;
19417   __builtin_aarch64_simd_xi __o;
19418   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19419   ret.val[0] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19420   ret.val[1] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19421   ret.val[2] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19422   ret.val[3] = (int16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19423   return ret;
19424 }
19425
19426 __extension__ extern __inline poly16x8x4_t
19427 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19428 vld4q_dup_p16 (const poly16_t * __a)
19429 {
19430   poly16x8x4_t ret;
19431   __builtin_aarch64_simd_xi __o;
19432   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19433   ret.val[0] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19434   ret.val[1] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19435   ret.val[2] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19436   ret.val[3] = (poly16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19437   return ret;
19438 }
19439
19440 __extension__ extern __inline int32x4x4_t
19441 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19442 vld4q_dup_s32 (const int32_t * __a)
19443 {
19444   int32x4x4_t ret;
19445   __builtin_aarch64_simd_xi __o;
19446   __o = __builtin_aarch64_ld4rv4si ((const __builtin_aarch64_simd_si *) __a);
19447   ret.val[0] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
19448   ret.val[1] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
19449   ret.val[2] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
19450   ret.val[3] = (int32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
19451   return ret;
19452 }
19453
19454 __extension__ extern __inline int64x2x4_t
19455 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19456 vld4q_dup_s64 (const int64_t * __a)
19457 {
19458   int64x2x4_t ret;
19459   __builtin_aarch64_simd_xi __o;
19460   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19461   ret.val[0] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
19462   ret.val[1] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
19463   ret.val[2] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
19464   ret.val[3] = (int64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
19465   return ret;
19466 }
19467
19468 __extension__ extern __inline uint8x16x4_t
19469 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19470 vld4q_dup_u8 (const uint8_t * __a)
19471 {
19472   uint8x16x4_t ret;
19473   __builtin_aarch64_simd_xi __o;
19474   __o = __builtin_aarch64_ld4rv16qi ((const __builtin_aarch64_simd_qi *) __a);
19475   ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
19476   ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
19477   ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
19478   ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
19479   return ret;
19480 }
19481
19482 __extension__ extern __inline uint16x8x4_t
19483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19484 vld4q_dup_u16 (const uint16_t * __a)
19485 {
19486   uint16x8x4_t ret;
19487   __builtin_aarch64_simd_xi __o;
19488   __o = __builtin_aarch64_ld4rv8hi ((const __builtin_aarch64_simd_hi *) __a);
19489   ret.val[0] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 0);
19490   ret.val[1] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 1);
19491   ret.val[2] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 2);
19492   ret.val[3] = (uint16x8_t) __builtin_aarch64_get_qregxiv8hi (__o, 3);
19493   return ret;
19494 }
19495
19496 __extension__ extern __inline uint32x4x4_t
19497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19498 vld4q_dup_u32 (const uint32_t * __a)
19499 {
19500   uint32x4x4_t ret;
19501   __builtin_aarch64_simd_xi __o;
19502   __o = __builtin_aarch64_ld4rv4si ((const __builtin_aarch64_simd_si *) __a);
19503   ret.val[0] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 0);
19504   ret.val[1] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 1);
19505   ret.val[2] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 2);
19506   ret.val[3] = (uint32x4_t) __builtin_aarch64_get_qregxiv4si (__o, 3);
19507   return ret;
19508 }
19509
19510 __extension__ extern __inline uint64x2x4_t
19511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19512 vld4q_dup_u64 (const uint64_t * __a)
19513 {
19514   uint64x2x4_t ret;
19515   __builtin_aarch64_simd_xi __o;
19516   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19517   ret.val[0] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 0);
19518   ret.val[1] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 1);
19519   ret.val[2] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 2);
19520   ret.val[3] = (uint64x2_t) __builtin_aarch64_get_qregxiv2di (__o, 3);
19521   return ret;
19522 }
19523
19524 __extension__ extern __inline float16x8x4_t
19525 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19526 vld4q_dup_f16 (const float16_t * __a)
19527 {
19528   float16x8x4_t ret;
19529   __builtin_aarch64_simd_xi __o;
19530   __o = __builtin_aarch64_ld4rv8hf ((const __builtin_aarch64_simd_hf *) __a);
19531   ret.val[0] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 0);
19532   ret.val[1] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 1);
19533   ret.val[2] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 2);
19534   ret.val[3] = (float16x8_t) __builtin_aarch64_get_qregxiv8hf (__o, 3);
19535   return ret;
19536 }
19537
19538 __extension__ extern __inline float32x4x4_t
19539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19540 vld4q_dup_f32 (const float32_t * __a)
19541 {
19542   float32x4x4_t ret;
19543   __builtin_aarch64_simd_xi __o;
19544   __o = __builtin_aarch64_ld4rv4sf ((const __builtin_aarch64_simd_sf *) __a);
19545   ret.val[0] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 0);
19546   ret.val[1] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 1);
19547   ret.val[2] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 2);
19548   ret.val[3] = (float32x4_t) __builtin_aarch64_get_qregxiv4sf (__o, 3);
19549   return ret;
19550 }
19551
19552 __extension__ extern __inline float64x2x4_t
19553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19554 vld4q_dup_f64 (const float64_t * __a)
19555 {
19556   float64x2x4_t ret;
19557   __builtin_aarch64_simd_xi __o;
19558   __o = __builtin_aarch64_ld4rv2df ((const __builtin_aarch64_simd_df *) __a);
19559   ret.val[0] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 0);
19560   ret.val[1] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 1);
19561   ret.val[2] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 2);
19562   ret.val[3] = (float64x2_t) __builtin_aarch64_get_qregxiv2df (__o, 3);
19563   return ret;
19564 }
19565
19566 __extension__ extern __inline poly64x2x4_t
19567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19568 vld4q_dup_p64 (const poly64_t * __a)
19569 {
19570   poly64x2x4_t ret;
19571   __builtin_aarch64_simd_xi __o;
19572   __o = __builtin_aarch64_ld4rv2di ((const __builtin_aarch64_simd_di *) __a);
19573   ret.val[0] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 0);
19574   ret.val[1] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 1);
19575   ret.val[2] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 2);
19576   ret.val[3] = (poly64x2_t) __builtin_aarch64_get_qregxiv2di_pss (__o, 3);
19577   return ret;
19578 }
19579
19580 /* vld2_lane */
19581
19582 #define __LD2_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
19583                          qmode, ptrmode, funcsuffix, signedtype)           \
19584 __extension__ extern __inline intype \
19585 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19586 vld2_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
19587 {                                                                          \
19588   __builtin_aarch64_simd_oi __o;                                           \
19589   largetype __temp;                                                        \
19590   __temp.val[0] =                                                          \
19591     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
19592   __temp.val[1] =                                                          \
19593     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
19594   __o = __builtin_aarch64_set_qregoi##qmode (__o,                          \
19595                                             (signedtype) __temp.val[0],    \
19596                                             0);                            \
19597   __o = __builtin_aarch64_set_qregoi##qmode (__o,                          \
19598                                             (signedtype) __temp.val[1],    \
19599                                             1);                            \
19600   __o = __builtin_aarch64_ld2_lane##mode (                                 \
19601           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
19602   __b.val[0] = (vectype) __builtin_aarch64_get_dregoidi (__o, 0);          \
19603   __b.val[1] = (vectype) __builtin_aarch64_get_dregoidi (__o, 1);          \
19604   return __b;                                                              \
19605 }
19606
19607 __LD2_LANE_FUNC (float16x4x2_t, float16x4_t, float16x8x2_t, float16_t, v4hf,
19608                  v8hf, hf, f16, float16x8_t)
19609 __LD2_LANE_FUNC (float32x2x2_t, float32x2_t, float32x4x2_t, float32_t, v2sf, v4sf,
19610                  sf, f32, float32x4_t)
19611 __LD2_LANE_FUNC (float64x1x2_t, float64x1_t, float64x2x2_t, float64_t, df, v2df,
19612                  df, f64, float64x2_t)
19613 __LD2_LANE_FUNC (poly8x8x2_t, poly8x8_t, poly8x16x2_t, poly8_t, v8qi, v16qi, qi, p8,
19614                  int8x16_t)
19615 __LD2_LANE_FUNC (poly16x4x2_t, poly16x4_t, poly16x8x2_t, poly16_t, v4hi, v8hi, hi,
19616                  p16, int16x8_t)
19617 __LD2_LANE_FUNC (poly64x1x2_t, poly64x1_t, poly64x2x2_t, poly64_t, di,
19618                  v2di_ssps, di, p64, poly64x2_t)
19619 __LD2_LANE_FUNC (int8x8x2_t, int8x8_t, int8x16x2_t, int8_t, v8qi, v16qi, qi, s8,
19620                  int8x16_t)
19621 __LD2_LANE_FUNC (int16x4x2_t, int16x4_t, int16x8x2_t, int16_t, v4hi, v8hi, hi, s16,
19622                  int16x8_t)
19623 __LD2_LANE_FUNC (int32x2x2_t, int32x2_t, int32x4x2_t, int32_t, v2si, v4si, si, s32,
19624                  int32x4_t)
19625 __LD2_LANE_FUNC (int64x1x2_t, int64x1_t, int64x2x2_t, int64_t, di, v2di, di, s64,
19626                  int64x2_t)
19627 __LD2_LANE_FUNC (uint8x8x2_t, uint8x8_t, uint8x16x2_t, uint8_t, v8qi, v16qi, qi, u8,
19628                  int8x16_t)
19629 __LD2_LANE_FUNC (uint16x4x2_t, uint16x4_t, uint16x8x2_t, uint16_t, v4hi, v8hi, hi,
19630                  u16, int16x8_t)
19631 __LD2_LANE_FUNC (uint32x2x2_t, uint32x2_t, uint32x4x2_t, uint32_t, v2si, v4si, si,
19632                  u32, int32x4_t)
19633 __LD2_LANE_FUNC (uint64x1x2_t, uint64x1_t, uint64x2x2_t, uint64_t, di, v2di, di,
19634                  u64, int64x2_t)
19635
19636 #undef __LD2_LANE_FUNC
19637
19638 /* vld2q_lane */
19639
19640 #define __LD2_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
19641 __extension__ extern __inline intype \
19642 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19643 vld2q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
19644 {                                                                          \
19645   __builtin_aarch64_simd_oi __o;                                           \
19646   intype ret;                                                              \
19647   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) __b.val[0], 0); \
19648   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) __b.val[1], 1); \
19649   __o = __builtin_aarch64_ld2_lane##mode (                                 \
19650         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
19651   ret.val[0] = (vtype) __builtin_aarch64_get_qregoiv4si (__o, 0);          \
19652   ret.val[1] = (vtype) __builtin_aarch64_get_qregoiv4si (__o, 1);          \
19653   return ret;                                                              \
19654 }
19655
19656 __LD2_LANE_FUNC (float16x8x2_t, float16x8_t, float16_t, v8hf, hf, f16)
19657 __LD2_LANE_FUNC (float32x4x2_t, float32x4_t, float32_t, v4sf, sf, f32)
19658 __LD2_LANE_FUNC (float64x2x2_t, float64x2_t, float64_t, v2df, df, f64)
19659 __LD2_LANE_FUNC (poly8x16x2_t, poly8x16_t, poly8_t, v16qi, qi, p8)
19660 __LD2_LANE_FUNC (poly16x8x2_t, poly16x8_t, poly16_t, v8hi, hi, p16)
19661 __LD2_LANE_FUNC (poly64x2x2_t, poly64x2_t, poly64_t, v2di, di, p64)
19662 __LD2_LANE_FUNC (int8x16x2_t, int8x16_t, int8_t, v16qi, qi, s8)
19663 __LD2_LANE_FUNC (int16x8x2_t, int16x8_t, int16_t, v8hi, hi, s16)
19664 __LD2_LANE_FUNC (int32x4x2_t, int32x4_t, int32_t, v4si, si, s32)
19665 __LD2_LANE_FUNC (int64x2x2_t, int64x2_t, int64_t, v2di, di, s64)
19666 __LD2_LANE_FUNC (uint8x16x2_t, uint8x16_t, uint8_t, v16qi, qi, u8)
19667 __LD2_LANE_FUNC (uint16x8x2_t, uint16x8_t, uint16_t, v8hi, hi, u16)
19668 __LD2_LANE_FUNC (uint32x4x2_t, uint32x4_t, uint32_t, v4si, si, u32)
19669 __LD2_LANE_FUNC (uint64x2x2_t, uint64x2_t, uint64_t, v2di, di, u64)
19670
19671 #undef __LD2_LANE_FUNC
19672
19673 /* vld3_lane */
19674
19675 #define __LD3_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
19676                          qmode, ptrmode, funcsuffix, signedtype)           \
19677 __extension__ extern __inline intype \
19678 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19679 vld3_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
19680 {                                                                          \
19681   __builtin_aarch64_simd_ci __o;                                           \
19682   largetype __temp;                                                        \
19683   __temp.val[0] =                                                          \
19684     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
19685   __temp.val[1] =                                                          \
19686     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
19687   __temp.val[2] =                                                          \
19688     vcombine_##funcsuffix (__b.val[2], vcreate_##funcsuffix (0));          \
19689   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
19690                                             (signedtype) __temp.val[0],    \
19691                                             0);                            \
19692   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
19693                                             (signedtype) __temp.val[1],    \
19694                                             1);                            \
19695   __o = __builtin_aarch64_set_qregci##qmode (__o,                          \
19696                                             (signedtype) __temp.val[2],    \
19697                                             2);                            \
19698   __o = __builtin_aarch64_ld3_lane##mode (                                 \
19699           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
19700   __b.val[0] = (vectype) __builtin_aarch64_get_dregcidi (__o, 0);          \
19701   __b.val[1] = (vectype) __builtin_aarch64_get_dregcidi (__o, 1);          \
19702   __b.val[2] = (vectype) __builtin_aarch64_get_dregcidi (__o, 2);          \
19703   return __b;                                                              \
19704 }
19705
19706 __LD3_LANE_FUNC (float16x4x3_t, float16x4_t, float16x8x3_t, float16_t, v4hf,
19707                  v8hf, hf, f16, float16x8_t)
19708 __LD3_LANE_FUNC (float32x2x3_t, float32x2_t, float32x4x3_t, float32_t, v2sf, v4sf,
19709                  sf, f32, float32x4_t)
19710 __LD3_LANE_FUNC (float64x1x3_t, float64x1_t, float64x2x3_t, float64_t, df, v2df,
19711                  df, f64, float64x2_t)
19712 __LD3_LANE_FUNC (poly8x8x3_t, poly8x8_t, poly8x16x3_t, poly8_t, v8qi, v16qi, qi, p8,
19713                  int8x16_t)
19714 __LD3_LANE_FUNC (poly16x4x3_t, poly16x4_t, poly16x8x3_t, poly16_t, v4hi, v8hi, hi,
19715                  p16, int16x8_t)
19716 __LD3_LANE_FUNC (poly64x1x3_t, poly64x1_t, poly64x2x3_t, poly64_t, di,
19717                  v2di_ssps, di, p64, poly64x2_t)
19718 __LD3_LANE_FUNC (int8x8x3_t, int8x8_t, int8x16x3_t, int8_t, v8qi, v16qi, qi, s8,
19719                  int8x16_t)
19720 __LD3_LANE_FUNC (int16x4x3_t, int16x4_t, int16x8x3_t, int16_t, v4hi, v8hi, hi, s16,
19721                  int16x8_t)
19722 __LD3_LANE_FUNC (int32x2x3_t, int32x2_t, int32x4x3_t, int32_t, v2si, v4si, si, s32,
19723                  int32x4_t)
19724 __LD3_LANE_FUNC (int64x1x3_t, int64x1_t, int64x2x3_t, int64_t, di, v2di, di, s64,
19725                  int64x2_t)
19726 __LD3_LANE_FUNC (uint8x8x3_t, uint8x8_t, uint8x16x3_t, uint8_t, v8qi, v16qi, qi, u8,
19727                  int8x16_t)
19728 __LD3_LANE_FUNC (uint16x4x3_t, uint16x4_t, uint16x8x3_t, uint16_t, v4hi, v8hi, hi,
19729                  u16, int16x8_t)
19730 __LD3_LANE_FUNC (uint32x2x3_t, uint32x2_t, uint32x4x3_t, uint32_t, v2si, v4si, si,
19731                  u32, int32x4_t)
19732 __LD3_LANE_FUNC (uint64x1x3_t, uint64x1_t, uint64x2x3_t, uint64_t, di, v2di, di,
19733                  u64, int64x2_t)
19734
19735 #undef __LD3_LANE_FUNC
19736
19737 /* vld3q_lane */
19738
19739 #define __LD3_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
19740 __extension__ extern __inline intype \
19741 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19742 vld3q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
19743 {                                                                          \
19744   __builtin_aarch64_simd_ci __o;                                           \
19745   intype ret;                                                              \
19746   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[0], 0); \
19747   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[1], 1); \
19748   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) __b.val[2], 2); \
19749   __o = __builtin_aarch64_ld3_lane##mode (                                 \
19750         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
19751   ret.val[0] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 0);          \
19752   ret.val[1] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 1);          \
19753   ret.val[2] = (vtype) __builtin_aarch64_get_qregciv4si (__o, 2);          \
19754   return ret;                                                              \
19755 }
19756
19757 __LD3_LANE_FUNC (float16x8x3_t, float16x8_t, float16_t, v8hf, hf, f16)
19758 __LD3_LANE_FUNC (float32x4x3_t, float32x4_t, float32_t, v4sf, sf, f32)
19759 __LD3_LANE_FUNC (float64x2x3_t, float64x2_t, float64_t, v2df, df, f64)
19760 __LD3_LANE_FUNC (poly8x16x3_t, poly8x16_t, poly8_t, v16qi, qi, p8)
19761 __LD3_LANE_FUNC (poly16x8x3_t, poly16x8_t, poly16_t, v8hi, hi, p16)
19762 __LD3_LANE_FUNC (poly64x2x3_t, poly64x2_t, poly64_t, v2di, di, p64)
19763 __LD3_LANE_FUNC (int8x16x3_t, int8x16_t, int8_t, v16qi, qi, s8)
19764 __LD3_LANE_FUNC (int16x8x3_t, int16x8_t, int16_t, v8hi, hi, s16)
19765 __LD3_LANE_FUNC (int32x4x3_t, int32x4_t, int32_t, v4si, si, s32)
19766 __LD3_LANE_FUNC (int64x2x3_t, int64x2_t, int64_t, v2di, di, s64)
19767 __LD3_LANE_FUNC (uint8x16x3_t, uint8x16_t, uint8_t, v16qi, qi, u8)
19768 __LD3_LANE_FUNC (uint16x8x3_t, uint16x8_t, uint16_t, v8hi, hi, u16)
19769 __LD3_LANE_FUNC (uint32x4x3_t, uint32x4_t, uint32_t, v4si, si, u32)
19770 __LD3_LANE_FUNC (uint64x2x3_t, uint64x2_t, uint64_t, v2di, di, u64)
19771
19772 #undef __LD3_LANE_FUNC
19773
19774 /* vld4_lane */
19775
19776 #define __LD4_LANE_FUNC(intype, vectype, largetype, ptrtype, mode,         \
19777                          qmode, ptrmode, funcsuffix, signedtype)           \
19778 __extension__ extern __inline intype \
19779 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19780 vld4_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c)  \
19781 {                                                                          \
19782   __builtin_aarch64_simd_xi __o;                                           \
19783   largetype __temp;                                                        \
19784   __temp.val[0] =                                                          \
19785     vcombine_##funcsuffix (__b.val[0], vcreate_##funcsuffix (0));          \
19786   __temp.val[1] =                                                          \
19787     vcombine_##funcsuffix (__b.val[1], vcreate_##funcsuffix (0));          \
19788   __temp.val[2] =                                                          \
19789     vcombine_##funcsuffix (__b.val[2], vcreate_##funcsuffix (0));          \
19790   __temp.val[3] =                                                          \
19791     vcombine_##funcsuffix (__b.val[3], vcreate_##funcsuffix (0));          \
19792   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19793                                             (signedtype) __temp.val[0],    \
19794                                             0);                            \
19795   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19796                                             (signedtype) __temp.val[1],    \
19797                                             1);                            \
19798   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19799                                             (signedtype) __temp.val[2],    \
19800                                             2);                            \
19801   __o = __builtin_aarch64_set_qregxi##qmode (__o,                          \
19802                                             (signedtype) __temp.val[3],    \
19803                                             3);                            \
19804   __o = __builtin_aarch64_ld4_lane##mode (                                 \
19805           (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);           \
19806   __b.val[0] = (vectype) __builtin_aarch64_get_dregxidi (__o, 0);          \
19807   __b.val[1] = (vectype) __builtin_aarch64_get_dregxidi (__o, 1);          \
19808   __b.val[2] = (vectype) __builtin_aarch64_get_dregxidi (__o, 2);          \
19809   __b.val[3] = (vectype) __builtin_aarch64_get_dregxidi (__o, 3);          \
19810   return __b;                                                              \
19811 }
19812
19813 /* vld4q_lane */
19814
19815 __LD4_LANE_FUNC (float16x4x4_t, float16x4_t, float16x8x4_t, float16_t, v4hf,
19816                  v8hf, hf, f16, float16x8_t)
19817 __LD4_LANE_FUNC (float32x2x4_t, float32x2_t, float32x4x4_t, float32_t, v2sf, v4sf,
19818                  sf, f32, float32x4_t)
19819 __LD4_LANE_FUNC (float64x1x4_t, float64x1_t, float64x2x4_t, float64_t, df, v2df,
19820                  df, f64, float64x2_t)
19821 __LD4_LANE_FUNC (poly8x8x4_t, poly8x8_t, poly8x16x4_t, poly8_t, v8qi, v16qi, qi, p8,
19822                  int8x16_t)
19823 __LD4_LANE_FUNC (poly16x4x4_t, poly16x4_t, poly16x8x4_t, poly16_t, v4hi, v8hi, hi,
19824                  p16, int16x8_t)
19825 __LD4_LANE_FUNC (poly64x1x4_t, poly64x1_t, poly64x2x4_t, poly64_t, di,
19826                  v2di_ssps, di, p64, poly64x2_t)
19827 __LD4_LANE_FUNC (int8x8x4_t, int8x8_t, int8x16x4_t, int8_t, v8qi, v16qi, qi, s8,
19828                  int8x16_t)
19829 __LD4_LANE_FUNC (int16x4x4_t, int16x4_t, int16x8x4_t, int16_t, v4hi, v8hi, hi, s16,
19830                  int16x8_t)
19831 __LD4_LANE_FUNC (int32x2x4_t, int32x2_t, int32x4x4_t, int32_t, v2si, v4si, si, s32,
19832                  int32x4_t)
19833 __LD4_LANE_FUNC (int64x1x4_t, int64x1_t, int64x2x4_t, int64_t, di, v2di, di, s64,
19834                  int64x2_t)
19835 __LD4_LANE_FUNC (uint8x8x4_t, uint8x8_t, uint8x16x4_t, uint8_t, v8qi, v16qi, qi, u8,
19836                  int8x16_t)
19837 __LD4_LANE_FUNC (uint16x4x4_t, uint16x4_t, uint16x8x4_t, uint16_t, v4hi, v8hi, hi,
19838                  u16, int16x8_t)
19839 __LD4_LANE_FUNC (uint32x2x4_t, uint32x2_t, uint32x4x4_t, uint32_t, v2si, v4si, si,
19840                  u32, int32x4_t)
19841 __LD4_LANE_FUNC (uint64x1x4_t, uint64x1_t, uint64x2x4_t, uint64_t, di, v2di, di,
19842                  u64, int64x2_t)
19843
19844 #undef __LD4_LANE_FUNC
19845
19846 /* vld4q_lane */
19847
19848 #define __LD4_LANE_FUNC(intype, vtype, ptrtype, mode, ptrmode, funcsuffix) \
19849 __extension__ extern __inline intype \
19850 __attribute__ ((__always_inline__, __gnu_inline__,__artificial__)) \
19851 vld4q_lane_##funcsuffix (const ptrtype * __ptr, intype __b, const int __c) \
19852 {                                                                          \
19853   __builtin_aarch64_simd_xi __o;                                           \
19854   intype ret;                                                              \
19855   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[0], 0); \
19856   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[1], 1); \
19857   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[2], 2); \
19858   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) __b.val[3], 3); \
19859   __o = __builtin_aarch64_ld4_lane##mode (                                 \
19860         (__builtin_aarch64_simd_##ptrmode *) __ptr, __o, __c);             \
19861   ret.val[0] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 0);          \
19862   ret.val[1] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 1);          \
19863   ret.val[2] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 2);          \
19864   ret.val[3] = (vtype) __builtin_aarch64_get_qregxiv4si (__o, 3);          \
19865   return ret;                                                              \
19866 }
19867
19868 __LD4_LANE_FUNC (float16x8x4_t, float16x8_t, float16_t, v8hf, hf, f16)
19869 __LD4_LANE_FUNC (float32x4x4_t, float32x4_t, float32_t, v4sf, sf, f32)
19870 __LD4_LANE_FUNC (float64x2x4_t, float64x2_t, float64_t, v2df, df, f64)
19871 __LD4_LANE_FUNC (poly8x16x4_t, poly8x16_t, poly8_t, v16qi, qi, p8)
19872 __LD4_LANE_FUNC (poly16x8x4_t, poly16x8_t, poly16_t, v8hi, hi, p16)
19873 __LD4_LANE_FUNC (poly64x2x4_t, poly64x2_t, poly64_t, v2di, di, p64)
19874 __LD4_LANE_FUNC (int8x16x4_t, int8x16_t, int8_t, v16qi, qi, s8)
19875 __LD4_LANE_FUNC (int16x8x4_t, int16x8_t, int16_t, v8hi, hi, s16)
19876 __LD4_LANE_FUNC (int32x4x4_t, int32x4_t, int32_t, v4si, si, s32)
19877 __LD4_LANE_FUNC (int64x2x4_t, int64x2_t, int64_t, v2di, di, s64)
19878 __LD4_LANE_FUNC (uint8x16x4_t, uint8x16_t, uint8_t, v16qi, qi, u8)
19879 __LD4_LANE_FUNC (uint16x8x4_t, uint16x8_t, uint16_t, v8hi, hi, u16)
19880 __LD4_LANE_FUNC (uint32x4x4_t, uint32x4_t, uint32_t, v4si, si, u32)
19881 __LD4_LANE_FUNC (uint64x2x4_t, uint64x2_t, uint64_t, v2di, di, u64)
19882
19883 #undef __LD4_LANE_FUNC
19884
19885 /* vmax */
19886
19887 __extension__ extern __inline float32x2_t
19888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19889 vmax_f32 (float32x2_t __a, float32x2_t __b)
19890 {
19891   return __builtin_aarch64_smax_nanv2sf (__a, __b);
19892 }
19893
19894 __extension__ extern __inline float64x1_t
19895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19896 vmax_f64 (float64x1_t __a, float64x1_t __b)
19897 {
19898     return (float64x1_t)
19899       { __builtin_aarch64_smax_nandf (vget_lane_f64 (__a, 0),
19900                                       vget_lane_f64 (__b, 0)) };
19901 }
19902
19903 __extension__ extern __inline int8x8_t
19904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19905 vmax_s8 (int8x8_t __a, int8x8_t __b)
19906 {
19907   return __builtin_aarch64_smaxv8qi (__a, __b);
19908 }
19909
19910 __extension__ extern __inline int16x4_t
19911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19912 vmax_s16 (int16x4_t __a, int16x4_t __b)
19913 {
19914   return __builtin_aarch64_smaxv4hi (__a, __b);
19915 }
19916
19917 __extension__ extern __inline int32x2_t
19918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19919 vmax_s32 (int32x2_t __a, int32x2_t __b)
19920 {
19921   return __builtin_aarch64_smaxv2si (__a, __b);
19922 }
19923
19924 __extension__ extern __inline uint8x8_t
19925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19926 vmax_u8 (uint8x8_t __a, uint8x8_t __b)
19927 {
19928   return (uint8x8_t) __builtin_aarch64_umaxv8qi ((int8x8_t) __a,
19929                                                  (int8x8_t) __b);
19930 }
19931
19932 __extension__ extern __inline uint16x4_t
19933 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19934 vmax_u16 (uint16x4_t __a, uint16x4_t __b)
19935 {
19936   return (uint16x4_t) __builtin_aarch64_umaxv4hi ((int16x4_t) __a,
19937                                                   (int16x4_t) __b);
19938 }
19939
19940 __extension__ extern __inline uint32x2_t
19941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19942 vmax_u32 (uint32x2_t __a, uint32x2_t __b)
19943 {
19944   return (uint32x2_t) __builtin_aarch64_umaxv2si ((int32x2_t) __a,
19945                                                   (int32x2_t) __b);
19946 }
19947
19948 __extension__ extern __inline float32x4_t
19949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19950 vmaxq_f32 (float32x4_t __a, float32x4_t __b)
19951 {
19952   return __builtin_aarch64_smax_nanv4sf (__a, __b);
19953 }
19954
19955 __extension__ extern __inline float64x2_t
19956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19957 vmaxq_f64 (float64x2_t __a, float64x2_t __b)
19958 {
19959   return __builtin_aarch64_smax_nanv2df (__a, __b);
19960 }
19961
19962 __extension__ extern __inline int8x16_t
19963 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19964 vmaxq_s8 (int8x16_t __a, int8x16_t __b)
19965 {
19966   return __builtin_aarch64_smaxv16qi (__a, __b);
19967 }
19968
19969 __extension__ extern __inline int16x8_t
19970 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19971 vmaxq_s16 (int16x8_t __a, int16x8_t __b)
19972 {
19973   return __builtin_aarch64_smaxv8hi (__a, __b);
19974 }
19975
19976 __extension__ extern __inline int32x4_t
19977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19978 vmaxq_s32 (int32x4_t __a, int32x4_t __b)
19979 {
19980   return __builtin_aarch64_smaxv4si (__a, __b);
19981 }
19982
19983 __extension__ extern __inline uint8x16_t
19984 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19985 vmaxq_u8 (uint8x16_t __a, uint8x16_t __b)
19986 {
19987   return (uint8x16_t) __builtin_aarch64_umaxv16qi ((int8x16_t) __a,
19988                                                    (int8x16_t) __b);
19989 }
19990
19991 __extension__ extern __inline uint16x8_t
19992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
19993 vmaxq_u16 (uint16x8_t __a, uint16x8_t __b)
19994 {
19995   return (uint16x8_t) __builtin_aarch64_umaxv8hi ((int16x8_t) __a,
19996                                                   (int16x8_t) __b);
19997 }
19998
19999 __extension__ extern __inline uint32x4_t
20000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20001 vmaxq_u32 (uint32x4_t __a, uint32x4_t __b)
20002 {
20003   return (uint32x4_t) __builtin_aarch64_umaxv4si ((int32x4_t) __a,
20004                                                   (int32x4_t) __b);
20005 }
20006 /* vmulx */
20007
20008 __extension__ extern __inline float32x2_t
20009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20010 vmulx_f32 (float32x2_t __a, float32x2_t __b)
20011 {
20012   return __builtin_aarch64_fmulxv2sf (__a, __b);
20013 }
20014
20015 __extension__ extern __inline float32x4_t
20016 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20017 vmulxq_f32 (float32x4_t __a, float32x4_t __b)
20018 {
20019   return __builtin_aarch64_fmulxv4sf (__a, __b);
20020 }
20021
20022 __extension__ extern __inline float64x1_t
20023 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20024 vmulx_f64 (float64x1_t __a, float64x1_t __b)
20025 {
20026   return (float64x1_t) {__builtin_aarch64_fmulxdf (__a[0], __b[0])};
20027 }
20028
20029 __extension__ extern __inline float64x2_t
20030 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20031 vmulxq_f64 (float64x2_t __a, float64x2_t __b)
20032 {
20033   return __builtin_aarch64_fmulxv2df (__a, __b);
20034 }
20035
20036 __extension__ extern __inline float32_t
20037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20038 vmulxs_f32 (float32_t __a, float32_t __b)
20039 {
20040   return __builtin_aarch64_fmulxsf (__a, __b);
20041 }
20042
20043 __extension__ extern __inline float64_t
20044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20045 vmulxd_f64 (float64_t __a, float64_t __b)
20046 {
20047   return __builtin_aarch64_fmulxdf (__a, __b);
20048 }
20049
20050 __extension__ extern __inline float32x2_t
20051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20052 vmulx_lane_f32 (float32x2_t __a, float32x2_t __v, const int __lane)
20053 {
20054   return vmulx_f32 (__a, __aarch64_vdup_lane_f32 (__v, __lane));
20055 }
20056
20057 __extension__ extern __inline float64x1_t
20058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20059 vmulx_lane_f64 (float64x1_t __a, float64x1_t __v, const int __lane)
20060 {
20061   return vmulx_f64 (__a, __aarch64_vdup_lane_f64 (__v, __lane));
20062 }
20063
20064 __extension__ extern __inline float32x4_t
20065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20066 vmulxq_lane_f32 (float32x4_t __a, float32x2_t __v, const int __lane)
20067 {
20068   return vmulxq_f32 (__a, __aarch64_vdupq_lane_f32 (__v, __lane));
20069 }
20070
20071 __extension__ extern __inline float64x2_t
20072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20073 vmulxq_lane_f64 (float64x2_t __a, float64x1_t __v, const int __lane)
20074 {
20075   return vmulxq_f64 (__a, __aarch64_vdupq_lane_f64 (__v, __lane));
20076 }
20077
20078 __extension__ extern __inline float32x2_t
20079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20080 vmulx_laneq_f32 (float32x2_t __a, float32x4_t __v, const int __lane)
20081 {
20082   return vmulx_f32 (__a, __aarch64_vdup_laneq_f32 (__v, __lane));
20083 }
20084
20085 __extension__ extern __inline float64x1_t
20086 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20087 vmulx_laneq_f64 (float64x1_t __a, float64x2_t __v, const int __lane)
20088 {
20089   return vmulx_f64 (__a, __aarch64_vdup_laneq_f64 (__v, __lane));
20090 }
20091
20092 __extension__ extern __inline float32x4_t
20093 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20094 vmulxq_laneq_f32 (float32x4_t __a, float32x4_t __v, const int __lane)
20095 {
20096   return vmulxq_f32 (__a, __aarch64_vdupq_laneq_f32 (__v, __lane));
20097 }
20098
20099 __extension__ extern __inline float64x2_t
20100 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20101 vmulxq_laneq_f64 (float64x2_t __a, float64x2_t __v, const int __lane)
20102 {
20103   return vmulxq_f64 (__a, __aarch64_vdupq_laneq_f64 (__v, __lane));
20104 }
20105
20106 __extension__ extern __inline float32_t
20107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20108 vmulxs_lane_f32 (float32_t __a, float32x2_t __v, const int __lane)
20109 {
20110   return vmulxs_f32 (__a, __aarch64_vget_lane_any (__v, __lane));
20111 }
20112
20113 __extension__ extern __inline float32_t
20114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20115 vmulxs_laneq_f32 (float32_t __a, float32x4_t __v, const int __lane)
20116 {
20117   return vmulxs_f32 (__a, __aarch64_vget_lane_any (__v, __lane));
20118 }
20119
20120 __extension__ extern __inline float64_t
20121 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20122 vmulxd_lane_f64 (float64_t __a, float64x1_t __v, const int __lane)
20123 {
20124   return vmulxd_f64 (__a, __aarch64_vget_lane_any (__v, __lane));
20125 }
20126
20127 __extension__ extern __inline float64_t
20128 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20129 vmulxd_laneq_f64 (float64_t __a, float64x2_t __v, const int __lane)
20130 {
20131   return vmulxd_f64 (__a, __aarch64_vget_lane_any (__v, __lane));
20132 }
20133
20134 /* vpmax  */
20135
20136 __extension__ extern __inline int8x8_t
20137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20138 vpmax_s8 (int8x8_t a, int8x8_t b)
20139 {
20140   return __builtin_aarch64_smaxpv8qi (a, b);
20141 }
20142
20143 __extension__ extern __inline int16x4_t
20144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20145 vpmax_s16 (int16x4_t a, int16x4_t b)
20146 {
20147   return __builtin_aarch64_smaxpv4hi (a, b);
20148 }
20149
20150 __extension__ extern __inline int32x2_t
20151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20152 vpmax_s32 (int32x2_t a, int32x2_t b)
20153 {
20154   return __builtin_aarch64_smaxpv2si (a, b);
20155 }
20156
20157 __extension__ extern __inline uint8x8_t
20158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20159 vpmax_u8 (uint8x8_t a, uint8x8_t b)
20160 {
20161   return (uint8x8_t) __builtin_aarch64_umaxpv8qi ((int8x8_t) a,
20162                                                   (int8x8_t) b);
20163 }
20164
20165 __extension__ extern __inline uint16x4_t
20166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20167 vpmax_u16 (uint16x4_t a, uint16x4_t b)
20168 {
20169   return (uint16x4_t) __builtin_aarch64_umaxpv4hi ((int16x4_t) a,
20170                                                    (int16x4_t) b);
20171 }
20172
20173 __extension__ extern __inline uint32x2_t
20174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20175 vpmax_u32 (uint32x2_t a, uint32x2_t b)
20176 {
20177   return (uint32x2_t) __builtin_aarch64_umaxpv2si ((int32x2_t) a,
20178                                                    (int32x2_t) b);
20179 }
20180
20181 __extension__ extern __inline int8x16_t
20182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20183 vpmaxq_s8 (int8x16_t a, int8x16_t b)
20184 {
20185   return __builtin_aarch64_smaxpv16qi (a, b);
20186 }
20187
20188 __extension__ extern __inline int16x8_t
20189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20190 vpmaxq_s16 (int16x8_t a, int16x8_t b)
20191 {
20192   return __builtin_aarch64_smaxpv8hi (a, b);
20193 }
20194
20195 __extension__ extern __inline int32x4_t
20196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20197 vpmaxq_s32 (int32x4_t a, int32x4_t b)
20198 {
20199   return __builtin_aarch64_smaxpv4si (a, b);
20200 }
20201
20202 __extension__ extern __inline uint8x16_t
20203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20204 vpmaxq_u8 (uint8x16_t a, uint8x16_t b)
20205 {
20206   return (uint8x16_t) __builtin_aarch64_umaxpv16qi ((int8x16_t) a,
20207                                                     (int8x16_t) b);
20208 }
20209
20210 __extension__ extern __inline uint16x8_t
20211 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20212 vpmaxq_u16 (uint16x8_t a, uint16x8_t b)
20213 {
20214   return (uint16x8_t) __builtin_aarch64_umaxpv8hi ((int16x8_t) a,
20215                                                    (int16x8_t) b);
20216 }
20217
20218 __extension__ extern __inline uint32x4_t
20219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20220 vpmaxq_u32 (uint32x4_t a, uint32x4_t b)
20221 {
20222   return (uint32x4_t) __builtin_aarch64_umaxpv4si ((int32x4_t) a,
20223                                                    (int32x4_t) b);
20224 }
20225
20226 __extension__ extern __inline float32x2_t
20227 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20228 vpmax_f32 (float32x2_t a, float32x2_t b)
20229 {
20230   return __builtin_aarch64_smax_nanpv2sf (a, b);
20231 }
20232
20233 __extension__ extern __inline float32x4_t
20234 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20235 vpmaxq_f32 (float32x4_t a, float32x4_t b)
20236 {
20237   return __builtin_aarch64_smax_nanpv4sf (a, b);
20238 }
20239
20240 __extension__ extern __inline float64x2_t
20241 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20242 vpmaxq_f64 (float64x2_t a, float64x2_t b)
20243 {
20244   return __builtin_aarch64_smax_nanpv2df (a, b);
20245 }
20246
20247 __extension__ extern __inline float64_t
20248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20249 vpmaxqd_f64 (float64x2_t a)
20250 {
20251   return __builtin_aarch64_reduc_smax_nan_scal_v2df (a);
20252 }
20253
20254 __extension__ extern __inline float32_t
20255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20256 vpmaxs_f32 (float32x2_t a)
20257 {
20258   return __builtin_aarch64_reduc_smax_nan_scal_v2sf (a);
20259 }
20260
20261 /* vpmaxnm  */
20262
20263 __extension__ extern __inline float32x2_t
20264 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20265 vpmaxnm_f32 (float32x2_t a, float32x2_t b)
20266 {
20267   return __builtin_aarch64_smaxpv2sf (a, b);
20268 }
20269
20270 __extension__ extern __inline float32x4_t
20271 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20272 vpmaxnmq_f32 (float32x4_t a, float32x4_t b)
20273 {
20274   return __builtin_aarch64_smaxpv4sf (a, b);
20275 }
20276
20277 __extension__ extern __inline float64x2_t
20278 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20279 vpmaxnmq_f64 (float64x2_t a, float64x2_t b)
20280 {
20281   return __builtin_aarch64_smaxpv2df (a, b);
20282 }
20283
20284 __extension__ extern __inline float64_t
20285 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20286 vpmaxnmqd_f64 (float64x2_t a)
20287 {
20288   return __builtin_aarch64_reduc_smax_scal_v2df (a);
20289 }
20290
20291 __extension__ extern __inline float32_t
20292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20293 vpmaxnms_f32 (float32x2_t a)
20294 {
20295   return __builtin_aarch64_reduc_smax_scal_v2sf (a);
20296 }
20297
20298 /* vpmin  */
20299
20300 __extension__ extern __inline int8x8_t
20301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20302 vpmin_s8 (int8x8_t a, int8x8_t b)
20303 {
20304   return __builtin_aarch64_sminpv8qi (a, b);
20305 }
20306
20307 __extension__ extern __inline int16x4_t
20308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20309 vpmin_s16 (int16x4_t a, int16x4_t b)
20310 {
20311   return __builtin_aarch64_sminpv4hi (a, b);
20312 }
20313
20314 __extension__ extern __inline int32x2_t
20315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20316 vpmin_s32 (int32x2_t a, int32x2_t b)
20317 {
20318   return __builtin_aarch64_sminpv2si (a, b);
20319 }
20320
20321 __extension__ extern __inline uint8x8_t
20322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20323 vpmin_u8 (uint8x8_t a, uint8x8_t b)
20324 {
20325   return (uint8x8_t) __builtin_aarch64_uminpv8qi ((int8x8_t) a,
20326                                                   (int8x8_t) b);
20327 }
20328
20329 __extension__ extern __inline uint16x4_t
20330 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20331 vpmin_u16 (uint16x4_t a, uint16x4_t b)
20332 {
20333   return (uint16x4_t) __builtin_aarch64_uminpv4hi ((int16x4_t) a,
20334                                                    (int16x4_t) b);
20335 }
20336
20337 __extension__ extern __inline uint32x2_t
20338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20339 vpmin_u32 (uint32x2_t a, uint32x2_t b)
20340 {
20341   return (uint32x2_t) __builtin_aarch64_uminpv2si ((int32x2_t) a,
20342                                                    (int32x2_t) b);
20343 }
20344
20345 __extension__ extern __inline int8x16_t
20346 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20347 vpminq_s8 (int8x16_t a, int8x16_t b)
20348 {
20349   return __builtin_aarch64_sminpv16qi (a, b);
20350 }
20351
20352 __extension__ extern __inline int16x8_t
20353 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20354 vpminq_s16 (int16x8_t a, int16x8_t b)
20355 {
20356   return __builtin_aarch64_sminpv8hi (a, b);
20357 }
20358
20359 __extension__ extern __inline int32x4_t
20360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20361 vpminq_s32 (int32x4_t a, int32x4_t b)
20362 {
20363   return __builtin_aarch64_sminpv4si (a, b);
20364 }
20365
20366 __extension__ extern __inline uint8x16_t
20367 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20368 vpminq_u8 (uint8x16_t a, uint8x16_t b)
20369 {
20370   return (uint8x16_t) __builtin_aarch64_uminpv16qi ((int8x16_t) a,
20371                                                     (int8x16_t) b);
20372 }
20373
20374 __extension__ extern __inline uint16x8_t
20375 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20376 vpminq_u16 (uint16x8_t a, uint16x8_t b)
20377 {
20378   return (uint16x8_t) __builtin_aarch64_uminpv8hi ((int16x8_t) a,
20379                                                    (int16x8_t) b);
20380 }
20381
20382 __extension__ extern __inline uint32x4_t
20383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20384 vpminq_u32 (uint32x4_t a, uint32x4_t b)
20385 {
20386   return (uint32x4_t) __builtin_aarch64_uminpv4si ((int32x4_t) a,
20387                                                    (int32x4_t) b);
20388 }
20389
20390 __extension__ extern __inline float32x2_t
20391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20392 vpmin_f32 (float32x2_t a, float32x2_t b)
20393 {
20394   return __builtin_aarch64_smin_nanpv2sf (a, b);
20395 }
20396
20397 __extension__ extern __inline float32x4_t
20398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20399 vpminq_f32 (float32x4_t a, float32x4_t b)
20400 {
20401   return __builtin_aarch64_smin_nanpv4sf (a, b);
20402 }
20403
20404 __extension__ extern __inline float64x2_t
20405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20406 vpminq_f64 (float64x2_t a, float64x2_t b)
20407 {
20408   return __builtin_aarch64_smin_nanpv2df (a, b);
20409 }
20410
20411 __extension__ extern __inline float64_t
20412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20413 vpminqd_f64 (float64x2_t a)
20414 {
20415   return __builtin_aarch64_reduc_smin_nan_scal_v2df (a);
20416 }
20417
20418 __extension__ extern __inline float32_t
20419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20420 vpmins_f32 (float32x2_t a)
20421 {
20422   return __builtin_aarch64_reduc_smin_nan_scal_v2sf (a);
20423 }
20424
20425 /* vpminnm  */
20426
20427 __extension__ extern __inline float32x2_t
20428 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20429 vpminnm_f32 (float32x2_t a, float32x2_t b)
20430 {
20431   return __builtin_aarch64_sminpv2sf (a, b);
20432 }
20433
20434 __extension__ extern __inline float32x4_t
20435 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20436 vpminnmq_f32 (float32x4_t a, float32x4_t b)
20437 {
20438   return __builtin_aarch64_sminpv4sf (a, b);
20439 }
20440
20441 __extension__ extern __inline float64x2_t
20442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20443 vpminnmq_f64 (float64x2_t a, float64x2_t b)
20444 {
20445   return __builtin_aarch64_sminpv2df (a, b);
20446 }
20447
20448 __extension__ extern __inline float64_t
20449 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20450 vpminnmqd_f64 (float64x2_t a)
20451 {
20452   return __builtin_aarch64_reduc_smin_scal_v2df (a);
20453 }
20454
20455 __extension__ extern __inline float32_t
20456 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20457 vpminnms_f32 (float32x2_t a)
20458 {
20459   return __builtin_aarch64_reduc_smin_scal_v2sf (a);
20460 }
20461
20462 /* vmaxnm  */
20463
20464 __extension__ extern __inline float32x2_t
20465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20466 vmaxnm_f32 (float32x2_t __a, float32x2_t __b)
20467 {
20468   return __builtin_aarch64_fmaxv2sf (__a, __b);
20469 }
20470
20471 __extension__ extern __inline float64x1_t
20472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20473 vmaxnm_f64 (float64x1_t __a, float64x1_t __b)
20474 {
20475   return (float64x1_t)
20476     { __builtin_aarch64_fmaxdf (vget_lane_f64 (__a, 0),
20477                                 vget_lane_f64 (__b, 0)) };
20478 }
20479
20480 __extension__ extern __inline float32x4_t
20481 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20482 vmaxnmq_f32 (float32x4_t __a, float32x4_t __b)
20483 {
20484   return __builtin_aarch64_fmaxv4sf (__a, __b);
20485 }
20486
20487 __extension__ extern __inline float64x2_t
20488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20489 vmaxnmq_f64 (float64x2_t __a, float64x2_t __b)
20490 {
20491   return __builtin_aarch64_fmaxv2df (__a, __b);
20492 }
20493
20494 /* vmaxv  */
20495
20496 __extension__ extern __inline float32_t
20497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20498 vmaxv_f32 (float32x2_t __a)
20499 {
20500   return __builtin_aarch64_reduc_smax_nan_scal_v2sf (__a);
20501 }
20502
20503 __extension__ extern __inline int8_t
20504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20505 vmaxv_s8 (int8x8_t __a)
20506 {
20507   return __builtin_aarch64_reduc_smax_scal_v8qi (__a);
20508 }
20509
20510 __extension__ extern __inline int16_t
20511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20512 vmaxv_s16 (int16x4_t __a)
20513 {
20514   return __builtin_aarch64_reduc_smax_scal_v4hi (__a);
20515 }
20516
20517 __extension__ extern __inline int32_t
20518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20519 vmaxv_s32 (int32x2_t __a)
20520 {
20521   return __builtin_aarch64_reduc_smax_scal_v2si (__a);
20522 }
20523
20524 __extension__ extern __inline uint8_t
20525 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20526 vmaxv_u8 (uint8x8_t __a)
20527 {
20528   return __builtin_aarch64_reduc_umax_scal_v8qi_uu (__a);
20529 }
20530
20531 __extension__ extern __inline uint16_t
20532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20533 vmaxv_u16 (uint16x4_t __a)
20534 {
20535   return __builtin_aarch64_reduc_umax_scal_v4hi_uu (__a);
20536 }
20537
20538 __extension__ extern __inline uint32_t
20539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20540 vmaxv_u32 (uint32x2_t __a)
20541 {
20542   return __builtin_aarch64_reduc_umax_scal_v2si_uu (__a);
20543 }
20544
20545 __extension__ extern __inline float32_t
20546 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20547 vmaxvq_f32 (float32x4_t __a)
20548 {
20549   return __builtin_aarch64_reduc_smax_nan_scal_v4sf (__a);
20550 }
20551
20552 __extension__ extern __inline float64_t
20553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20554 vmaxvq_f64 (float64x2_t __a)
20555 {
20556   return __builtin_aarch64_reduc_smax_nan_scal_v2df (__a);
20557 }
20558
20559 __extension__ extern __inline int8_t
20560 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20561 vmaxvq_s8 (int8x16_t __a)
20562 {
20563   return __builtin_aarch64_reduc_smax_scal_v16qi (__a);
20564 }
20565
20566 __extension__ extern __inline int16_t
20567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20568 vmaxvq_s16 (int16x8_t __a)
20569 {
20570   return __builtin_aarch64_reduc_smax_scal_v8hi (__a);
20571 }
20572
20573 __extension__ extern __inline int32_t
20574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20575 vmaxvq_s32 (int32x4_t __a)
20576 {
20577   return __builtin_aarch64_reduc_smax_scal_v4si (__a);
20578 }
20579
20580 __extension__ extern __inline uint8_t
20581 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20582 vmaxvq_u8 (uint8x16_t __a)
20583 {
20584   return __builtin_aarch64_reduc_umax_scal_v16qi_uu (__a);
20585 }
20586
20587 __extension__ extern __inline uint16_t
20588 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20589 vmaxvq_u16 (uint16x8_t __a)
20590 {
20591   return __builtin_aarch64_reduc_umax_scal_v8hi_uu (__a);
20592 }
20593
20594 __extension__ extern __inline uint32_t
20595 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20596 vmaxvq_u32 (uint32x4_t __a)
20597 {
20598   return __builtin_aarch64_reduc_umax_scal_v4si_uu (__a);
20599 }
20600
20601 /* vmaxnmv  */
20602
20603 __extension__ extern __inline float32_t
20604 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20605 vmaxnmv_f32 (float32x2_t __a)
20606 {
20607   return __builtin_aarch64_reduc_smax_scal_v2sf (__a);
20608 }
20609
20610 __extension__ extern __inline float32_t
20611 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20612 vmaxnmvq_f32 (float32x4_t __a)
20613 {
20614   return __builtin_aarch64_reduc_smax_scal_v4sf (__a);
20615 }
20616
20617 __extension__ extern __inline float64_t
20618 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20619 vmaxnmvq_f64 (float64x2_t __a)
20620 {
20621   return __builtin_aarch64_reduc_smax_scal_v2df (__a);
20622 }
20623
20624 /* vmin  */
20625
20626 __extension__ extern __inline float32x2_t
20627 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20628 vmin_f32 (float32x2_t __a, float32x2_t __b)
20629 {
20630   return __builtin_aarch64_smin_nanv2sf (__a, __b);
20631 }
20632
20633 __extension__ extern __inline float64x1_t
20634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20635 vmin_f64 (float64x1_t __a, float64x1_t __b)
20636 {
20637     return (float64x1_t)
20638           { __builtin_aarch64_smin_nandf (vget_lane_f64 (__a, 0),
20639                                           vget_lane_f64 (__b, 0)) };
20640 }
20641
20642 __extension__ extern __inline int8x8_t
20643 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20644 vmin_s8 (int8x8_t __a, int8x8_t __b)
20645 {
20646   return __builtin_aarch64_sminv8qi (__a, __b);
20647 }
20648
20649 __extension__ extern __inline int16x4_t
20650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20651 vmin_s16 (int16x4_t __a, int16x4_t __b)
20652 {
20653   return __builtin_aarch64_sminv4hi (__a, __b);
20654 }
20655
20656 __extension__ extern __inline int32x2_t
20657 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20658 vmin_s32 (int32x2_t __a, int32x2_t __b)
20659 {
20660   return __builtin_aarch64_sminv2si (__a, __b);
20661 }
20662
20663 __extension__ extern __inline uint8x8_t
20664 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20665 vmin_u8 (uint8x8_t __a, uint8x8_t __b)
20666 {
20667   return (uint8x8_t) __builtin_aarch64_uminv8qi ((int8x8_t) __a,
20668                                                  (int8x8_t) __b);
20669 }
20670
20671 __extension__ extern __inline uint16x4_t
20672 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20673 vmin_u16 (uint16x4_t __a, uint16x4_t __b)
20674 {
20675   return (uint16x4_t) __builtin_aarch64_uminv4hi ((int16x4_t) __a,
20676                                                   (int16x4_t) __b);
20677 }
20678
20679 __extension__ extern __inline uint32x2_t
20680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20681 vmin_u32 (uint32x2_t __a, uint32x2_t __b)
20682 {
20683   return (uint32x2_t) __builtin_aarch64_uminv2si ((int32x2_t) __a,
20684                                                   (int32x2_t) __b);
20685 }
20686
20687 __extension__ extern __inline float32x4_t
20688 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20689 vminq_f32 (float32x4_t __a, float32x4_t __b)
20690 {
20691   return __builtin_aarch64_smin_nanv4sf (__a, __b);
20692 }
20693
20694 __extension__ extern __inline float64x2_t
20695 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20696 vminq_f64 (float64x2_t __a, float64x2_t __b)
20697 {
20698   return __builtin_aarch64_smin_nanv2df (__a, __b);
20699 }
20700
20701 __extension__ extern __inline int8x16_t
20702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20703 vminq_s8 (int8x16_t __a, int8x16_t __b)
20704 {
20705   return __builtin_aarch64_sminv16qi (__a, __b);
20706 }
20707
20708 __extension__ extern __inline int16x8_t
20709 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20710 vminq_s16 (int16x8_t __a, int16x8_t __b)
20711 {
20712   return __builtin_aarch64_sminv8hi (__a, __b);
20713 }
20714
20715 __extension__ extern __inline int32x4_t
20716 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20717 vminq_s32 (int32x4_t __a, int32x4_t __b)
20718 {
20719   return __builtin_aarch64_sminv4si (__a, __b);
20720 }
20721
20722 __extension__ extern __inline uint8x16_t
20723 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20724 vminq_u8 (uint8x16_t __a, uint8x16_t __b)
20725 {
20726   return (uint8x16_t) __builtin_aarch64_uminv16qi ((int8x16_t) __a,
20727                                                    (int8x16_t) __b);
20728 }
20729
20730 __extension__ extern __inline uint16x8_t
20731 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20732 vminq_u16 (uint16x8_t __a, uint16x8_t __b)
20733 {
20734   return (uint16x8_t) __builtin_aarch64_uminv8hi ((int16x8_t) __a,
20735                                                   (int16x8_t) __b);
20736 }
20737
20738 __extension__ extern __inline uint32x4_t
20739 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20740 vminq_u32 (uint32x4_t __a, uint32x4_t __b)
20741 {
20742   return (uint32x4_t) __builtin_aarch64_uminv4si ((int32x4_t) __a,
20743                                                   (int32x4_t) __b);
20744 }
20745
20746 /* vminnm  */
20747
20748 __extension__ extern __inline float32x2_t
20749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20750 vminnm_f32 (float32x2_t __a, float32x2_t __b)
20751 {
20752   return __builtin_aarch64_fminv2sf (__a, __b);
20753 }
20754
20755 __extension__ extern __inline float64x1_t
20756 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20757 vminnm_f64 (float64x1_t __a, float64x1_t __b)
20758 {
20759   return (float64x1_t)
20760     { __builtin_aarch64_fmindf (vget_lane_f64 (__a, 0),
20761                                 vget_lane_f64 (__b, 0)) };
20762 }
20763
20764 __extension__ extern __inline float32x4_t
20765 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20766 vminnmq_f32 (float32x4_t __a, float32x4_t __b)
20767 {
20768   return __builtin_aarch64_fminv4sf (__a, __b);
20769 }
20770
20771 __extension__ extern __inline float64x2_t
20772 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20773 vminnmq_f64 (float64x2_t __a, float64x2_t __b)
20774 {
20775   return __builtin_aarch64_fminv2df (__a, __b);
20776 }
20777
20778 /* vminv  */
20779
20780 __extension__ extern __inline float32_t
20781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20782 vminv_f32 (float32x2_t __a)
20783 {
20784   return __builtin_aarch64_reduc_smin_nan_scal_v2sf (__a);
20785 }
20786
20787 __extension__ extern __inline int8_t
20788 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20789 vminv_s8 (int8x8_t __a)
20790 {
20791   return __builtin_aarch64_reduc_smin_scal_v8qi (__a);
20792 }
20793
20794 __extension__ extern __inline int16_t
20795 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20796 vminv_s16 (int16x4_t __a)
20797 {
20798   return __builtin_aarch64_reduc_smin_scal_v4hi (__a);
20799 }
20800
20801 __extension__ extern __inline int32_t
20802 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20803 vminv_s32 (int32x2_t __a)
20804 {
20805   return __builtin_aarch64_reduc_smin_scal_v2si (__a);
20806 }
20807
20808 __extension__ extern __inline uint8_t
20809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20810 vminv_u8 (uint8x8_t __a)
20811 {
20812   return __builtin_aarch64_reduc_umin_scal_v8qi_uu (__a);
20813 }
20814
20815 __extension__ extern __inline uint16_t
20816 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20817 vminv_u16 (uint16x4_t __a)
20818 {
20819   return __builtin_aarch64_reduc_umin_scal_v4hi_uu (__a);
20820 }
20821
20822 __extension__ extern __inline uint32_t
20823 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20824 vminv_u32 (uint32x2_t __a)
20825 {
20826   return __builtin_aarch64_reduc_umin_scal_v2si_uu (__a);
20827 }
20828
20829 __extension__ extern __inline float32_t
20830 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20831 vminvq_f32 (float32x4_t __a)
20832 {
20833   return __builtin_aarch64_reduc_smin_nan_scal_v4sf (__a);
20834 }
20835
20836 __extension__ extern __inline float64_t
20837 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20838 vminvq_f64 (float64x2_t __a)
20839 {
20840   return __builtin_aarch64_reduc_smin_nan_scal_v2df (__a);
20841 }
20842
20843 __extension__ extern __inline int8_t
20844 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20845 vminvq_s8 (int8x16_t __a)
20846 {
20847   return __builtin_aarch64_reduc_smin_scal_v16qi (__a);
20848 }
20849
20850 __extension__ extern __inline int16_t
20851 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20852 vminvq_s16 (int16x8_t __a)
20853 {
20854   return __builtin_aarch64_reduc_smin_scal_v8hi (__a);
20855 }
20856
20857 __extension__ extern __inline int32_t
20858 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20859 vminvq_s32 (int32x4_t __a)
20860 {
20861   return __builtin_aarch64_reduc_smin_scal_v4si (__a);
20862 }
20863
20864 __extension__ extern __inline uint8_t
20865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20866 vminvq_u8 (uint8x16_t __a)
20867 {
20868   return __builtin_aarch64_reduc_umin_scal_v16qi_uu (__a);
20869 }
20870
20871 __extension__ extern __inline uint16_t
20872 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20873 vminvq_u16 (uint16x8_t __a)
20874 {
20875   return __builtin_aarch64_reduc_umin_scal_v8hi_uu (__a);
20876 }
20877
20878 __extension__ extern __inline uint32_t
20879 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20880 vminvq_u32 (uint32x4_t __a)
20881 {
20882   return __builtin_aarch64_reduc_umin_scal_v4si_uu (__a);
20883 }
20884
20885 /* vminnmv  */
20886
20887 __extension__ extern __inline float32_t
20888 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20889 vminnmv_f32 (float32x2_t __a)
20890 {
20891   return __builtin_aarch64_reduc_smin_scal_v2sf (__a);
20892 }
20893
20894 __extension__ extern __inline float32_t
20895 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20896 vminnmvq_f32 (float32x4_t __a)
20897 {
20898   return __builtin_aarch64_reduc_smin_scal_v4sf (__a);
20899 }
20900
20901 __extension__ extern __inline float64_t
20902 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20903 vminnmvq_f64 (float64x2_t __a)
20904 {
20905   return __builtin_aarch64_reduc_smin_scal_v2df (__a);
20906 }
20907
20908 /* vmla */
20909
20910 __extension__ extern __inline float32x2_t
20911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20912 vmla_f32 (float32x2_t a, float32x2_t b, float32x2_t c)
20913 {
20914   return a + b * c;
20915 }
20916
20917 __extension__ extern __inline float64x1_t
20918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20919 vmla_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
20920 {
20921   return __a + __b * __c;
20922 }
20923
20924 __extension__ extern __inline float32x4_t
20925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20926 vmlaq_f32 (float32x4_t a, float32x4_t b, float32x4_t c)
20927 {
20928   return a + b * c;
20929 }
20930
20931 __extension__ extern __inline float64x2_t
20932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20933 vmlaq_f64 (float64x2_t a, float64x2_t b, float64x2_t c)
20934 {
20935   return a + b * c;
20936 }
20937
20938 /* vmla_lane  */
20939
20940 __extension__ extern __inline float32x2_t
20941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20942 vmla_lane_f32 (float32x2_t __a, float32x2_t __b,
20943                float32x2_t __c, const int __lane)
20944 {
20945   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20946 }
20947
20948 __extension__ extern __inline int16x4_t
20949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20950 vmla_lane_s16 (int16x4_t __a, int16x4_t __b,
20951                 int16x4_t __c, const int __lane)
20952 {
20953   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20954 }
20955
20956 __extension__ extern __inline int32x2_t
20957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20958 vmla_lane_s32 (int32x2_t __a, int32x2_t __b,
20959                 int32x2_t __c, const int __lane)
20960 {
20961   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20962 }
20963
20964 __extension__ extern __inline uint16x4_t
20965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20966 vmla_lane_u16 (uint16x4_t __a, uint16x4_t __b,
20967                 uint16x4_t __c, const int __lane)
20968 {
20969   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20970 }
20971
20972 __extension__ extern __inline uint32x2_t
20973 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20974 vmla_lane_u32 (uint32x2_t __a, uint32x2_t __b,
20975                uint32x2_t __c, const int __lane)
20976 {
20977   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20978 }
20979
20980 /* vmla_laneq  */
20981
20982 __extension__ extern __inline float32x2_t
20983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20984 vmla_laneq_f32 (float32x2_t __a, float32x2_t __b,
20985                 float32x4_t __c, const int __lane)
20986 {
20987   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20988 }
20989
20990 __extension__ extern __inline int16x4_t
20991 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
20992 vmla_laneq_s16 (int16x4_t __a, int16x4_t __b,
20993                 int16x8_t __c, const int __lane)
20994 {
20995   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
20996 }
20997
20998 __extension__ extern __inline int32x2_t
20999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21000 vmla_laneq_s32 (int32x2_t __a, int32x2_t __b,
21001                 int32x4_t __c, const int __lane)
21002 {
21003   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21004 }
21005
21006 __extension__ extern __inline uint16x4_t
21007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21008 vmla_laneq_u16 (uint16x4_t __a, uint16x4_t __b,
21009                 uint16x8_t __c, const int __lane)
21010 {
21011   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21012 }
21013
21014 __extension__ extern __inline uint32x2_t
21015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21016 vmla_laneq_u32 (uint32x2_t __a, uint32x2_t __b,
21017                 uint32x4_t __c, const int __lane)
21018 {
21019   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21020 }
21021
21022 /* vmlaq_lane  */
21023
21024 __extension__ extern __inline float32x4_t
21025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21026 vmlaq_lane_f32 (float32x4_t __a, float32x4_t __b,
21027                 float32x2_t __c, const int __lane)
21028 {
21029   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21030 }
21031
21032 __extension__ extern __inline int16x8_t
21033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21034 vmlaq_lane_s16 (int16x8_t __a, int16x8_t __b,
21035                 int16x4_t __c, const int __lane)
21036 {
21037   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21038 }
21039
21040 __extension__ extern __inline int32x4_t
21041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21042 vmlaq_lane_s32 (int32x4_t __a, int32x4_t __b,
21043                 int32x2_t __c, const int __lane)
21044 {
21045   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21046 }
21047
21048 __extension__ extern __inline uint16x8_t
21049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21050 vmlaq_lane_u16 (uint16x8_t __a, uint16x8_t __b,
21051                 uint16x4_t __c, const int __lane)
21052 {
21053   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21054 }
21055
21056 __extension__ extern __inline uint32x4_t
21057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21058 vmlaq_lane_u32 (uint32x4_t __a, uint32x4_t __b,
21059                 uint32x2_t __c, const int __lane)
21060 {
21061   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21062 }
21063
21064   /* vmlaq_laneq  */
21065
21066 __extension__ extern __inline float32x4_t
21067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21068 vmlaq_laneq_f32 (float32x4_t __a, float32x4_t __b,
21069                  float32x4_t __c, const int __lane)
21070 {
21071   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21072 }
21073
21074 __extension__ extern __inline int16x8_t
21075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21076 vmlaq_laneq_s16 (int16x8_t __a, int16x8_t __b,
21077                 int16x8_t __c, const int __lane)
21078 {
21079   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21080 }
21081
21082 __extension__ extern __inline int32x4_t
21083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21084 vmlaq_laneq_s32 (int32x4_t __a, int32x4_t __b,
21085                 int32x4_t __c, const int __lane)
21086 {
21087   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21088 }
21089
21090 __extension__ extern __inline uint16x8_t
21091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21092 vmlaq_laneq_u16 (uint16x8_t __a, uint16x8_t __b,
21093                 uint16x8_t __c, const int __lane)
21094 {
21095   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21096 }
21097
21098 __extension__ extern __inline uint32x4_t
21099 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21100 vmlaq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
21101                 uint32x4_t __c, const int __lane)
21102 {
21103   return (__a + (__b * __aarch64_vget_lane_any (__c, __lane)));
21104 }
21105
21106 /* vmls  */
21107
21108 __extension__ extern __inline float32x2_t
21109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21110 vmls_f32 (float32x2_t a, float32x2_t b, float32x2_t c)
21111 {
21112   return a - b * c;
21113 }
21114
21115 __extension__ extern __inline float64x1_t
21116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21117 vmls_f64 (float64x1_t __a, float64x1_t __b, float64x1_t __c)
21118 {
21119   return __a - __b * __c;
21120 }
21121
21122 __extension__ extern __inline float32x4_t
21123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21124 vmlsq_f32 (float32x4_t a, float32x4_t b, float32x4_t c)
21125 {
21126   return a - b * c;
21127 }
21128
21129 __extension__ extern __inline float64x2_t
21130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21131 vmlsq_f64 (float64x2_t a, float64x2_t b, float64x2_t c)
21132 {
21133   return a - b * c;
21134 }
21135
21136 /* vmls_lane  */
21137
21138 __extension__ extern __inline float32x2_t
21139 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21140 vmls_lane_f32 (float32x2_t __a, float32x2_t __b,
21141                float32x2_t __c, const int __lane)
21142 {
21143   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21144 }
21145
21146 __extension__ extern __inline int16x4_t
21147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21148 vmls_lane_s16 (int16x4_t __a, int16x4_t __b,
21149                 int16x4_t __c, const int __lane)
21150 {
21151   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21152 }
21153
21154 __extension__ extern __inline int32x2_t
21155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21156 vmls_lane_s32 (int32x2_t __a, int32x2_t __b,
21157                 int32x2_t __c, const int __lane)
21158 {
21159   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21160 }
21161
21162 __extension__ extern __inline uint16x4_t
21163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21164 vmls_lane_u16 (uint16x4_t __a, uint16x4_t __b,
21165                 uint16x4_t __c, const int __lane)
21166 {
21167   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21168 }
21169
21170 __extension__ extern __inline uint32x2_t
21171 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21172 vmls_lane_u32 (uint32x2_t __a, uint32x2_t __b,
21173                uint32x2_t __c, const int __lane)
21174 {
21175   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21176 }
21177
21178 /* vmls_laneq  */
21179
21180 __extension__ extern __inline float32x2_t
21181 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21182 vmls_laneq_f32 (float32x2_t __a, float32x2_t __b,
21183                float32x4_t __c, const int __lane)
21184 {
21185   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21186 }
21187
21188 __extension__ extern __inline int16x4_t
21189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21190 vmls_laneq_s16 (int16x4_t __a, int16x4_t __b,
21191                 int16x8_t __c, const int __lane)
21192 {
21193   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21194 }
21195
21196 __extension__ extern __inline int32x2_t
21197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21198 vmls_laneq_s32 (int32x2_t __a, int32x2_t __b,
21199                 int32x4_t __c, const int __lane)
21200 {
21201   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21202 }
21203
21204 __extension__ extern __inline uint16x4_t
21205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21206 vmls_laneq_u16 (uint16x4_t __a, uint16x4_t __b,
21207                 uint16x8_t __c, const int __lane)
21208 {
21209   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21210 }
21211
21212 __extension__ extern __inline uint32x2_t
21213 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21214 vmls_laneq_u32 (uint32x2_t __a, uint32x2_t __b,
21215                 uint32x4_t __c, const int __lane)
21216 {
21217   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21218 }
21219
21220 /* vmlsq_lane  */
21221
21222 __extension__ extern __inline float32x4_t
21223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21224 vmlsq_lane_f32 (float32x4_t __a, float32x4_t __b,
21225                 float32x2_t __c, const int __lane)
21226 {
21227   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21228 }
21229
21230 __extension__ extern __inline int16x8_t
21231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21232 vmlsq_lane_s16 (int16x8_t __a, int16x8_t __b,
21233                 int16x4_t __c, const int __lane)
21234 {
21235   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21236 }
21237
21238 __extension__ extern __inline int32x4_t
21239 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21240 vmlsq_lane_s32 (int32x4_t __a, int32x4_t __b,
21241                 int32x2_t __c, const int __lane)
21242 {
21243   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21244 }
21245
21246 __extension__ extern __inline uint16x8_t
21247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21248 vmlsq_lane_u16 (uint16x8_t __a, uint16x8_t __b,
21249                 uint16x4_t __c, const int __lane)
21250 {
21251   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21252 }
21253
21254 __extension__ extern __inline uint32x4_t
21255 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21256 vmlsq_lane_u32 (uint32x4_t __a, uint32x4_t __b,
21257                 uint32x2_t __c, const int __lane)
21258 {
21259   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21260 }
21261
21262   /* vmlsq_laneq  */
21263
21264 __extension__ extern __inline float32x4_t
21265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21266 vmlsq_laneq_f32 (float32x4_t __a, float32x4_t __b,
21267                 float32x4_t __c, const int __lane)
21268 {
21269   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21270 }
21271
21272 __extension__ extern __inline int16x8_t
21273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21274 vmlsq_laneq_s16 (int16x8_t __a, int16x8_t __b,
21275                 int16x8_t __c, const int __lane)
21276 {
21277   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21278 }
21279
21280 __extension__ extern __inline int32x4_t
21281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21282 vmlsq_laneq_s32 (int32x4_t __a, int32x4_t __b,
21283                 int32x4_t __c, const int __lane)
21284 {
21285   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21286 }
21287 __extension__ extern __inline uint16x8_t
21288 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21289 vmlsq_laneq_u16 (uint16x8_t __a, uint16x8_t __b,
21290                 uint16x8_t __c, const int __lane)
21291 {
21292   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21293 }
21294
21295 __extension__ extern __inline uint32x4_t
21296 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21297 vmlsq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
21298                 uint32x4_t __c, const int __lane)
21299 {
21300   return (__a - (__b * __aarch64_vget_lane_any (__c, __lane)));
21301 }
21302
21303 /* vmov_n_  */
21304
21305 __extension__ extern __inline float16x4_t
21306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21307 vmov_n_f16 (float16_t __a)
21308 {
21309   return vdup_n_f16 (__a);
21310 }
21311
21312 __extension__ extern __inline float32x2_t
21313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21314 vmov_n_f32 (float32_t __a)
21315 {
21316   return vdup_n_f32 (__a);
21317 }
21318
21319 __extension__ extern __inline float64x1_t
21320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21321 vmov_n_f64 (float64_t __a)
21322 {
21323   return (float64x1_t) {__a};
21324 }
21325
21326 __extension__ extern __inline poly8x8_t
21327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21328 vmov_n_p8 (poly8_t __a)
21329 {
21330   return vdup_n_p8 (__a);
21331 }
21332
21333 __extension__ extern __inline poly16x4_t
21334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21335 vmov_n_p16 (poly16_t __a)
21336 {
21337   return vdup_n_p16 (__a);
21338 }
21339
21340 __extension__ extern __inline poly64x1_t
21341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21342 vmov_n_p64 (poly64_t __a)
21343 {
21344   return vdup_n_p64 (__a);
21345 }
21346
21347 __extension__ extern __inline int8x8_t
21348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21349 vmov_n_s8 (int8_t __a)
21350 {
21351   return vdup_n_s8 (__a);
21352 }
21353
21354 __extension__ extern __inline int16x4_t
21355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21356 vmov_n_s16 (int16_t __a)
21357 {
21358   return vdup_n_s16 (__a);
21359 }
21360
21361 __extension__ extern __inline int32x2_t
21362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21363 vmov_n_s32 (int32_t __a)
21364 {
21365   return vdup_n_s32 (__a);
21366 }
21367
21368 __extension__ extern __inline int64x1_t
21369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21370 vmov_n_s64 (int64_t __a)
21371 {
21372   return (int64x1_t) {__a};
21373 }
21374
21375 __extension__ extern __inline uint8x8_t
21376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21377 vmov_n_u8 (uint8_t __a)
21378 {
21379   return vdup_n_u8 (__a);
21380 }
21381
21382 __extension__ extern __inline uint16x4_t
21383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21384 vmov_n_u16 (uint16_t __a)
21385 {
21386     return vdup_n_u16 (__a);
21387 }
21388
21389 __extension__ extern __inline uint32x2_t
21390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21391 vmov_n_u32 (uint32_t __a)
21392 {
21393    return vdup_n_u32 (__a);
21394 }
21395
21396 __extension__ extern __inline uint64x1_t
21397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21398 vmov_n_u64 (uint64_t __a)
21399 {
21400   return (uint64x1_t) {__a};
21401 }
21402
21403 __extension__ extern __inline float16x8_t
21404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21405 vmovq_n_f16 (float16_t __a)
21406 {
21407   return vdupq_n_f16 (__a);
21408 }
21409
21410 __extension__ extern __inline float32x4_t
21411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21412 vmovq_n_f32 (float32_t __a)
21413 {
21414   return vdupq_n_f32 (__a);
21415 }
21416
21417 __extension__ extern __inline float64x2_t
21418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21419 vmovq_n_f64 (float64_t __a)
21420 {
21421   return vdupq_n_f64 (__a);
21422 }
21423
21424 __extension__ extern __inline poly8x16_t
21425 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21426 vmovq_n_p8 (poly8_t __a)
21427 {
21428   return vdupq_n_p8 (__a);
21429 }
21430
21431 __extension__ extern __inline poly16x8_t
21432 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21433 vmovq_n_p16 (poly16_t __a)
21434 {
21435   return vdupq_n_p16 (__a);
21436 }
21437
21438 __extension__ extern __inline poly64x2_t
21439 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21440 vmovq_n_p64 (poly64_t __a)
21441 {
21442   return vdupq_n_p64 (__a);
21443 }
21444
21445 __extension__ extern __inline int8x16_t
21446 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21447 vmovq_n_s8 (int8_t __a)
21448 {
21449   return vdupq_n_s8 (__a);
21450 }
21451
21452 __extension__ extern __inline int16x8_t
21453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21454 vmovq_n_s16 (int16_t __a)
21455 {
21456   return vdupq_n_s16 (__a);
21457 }
21458
21459 __extension__ extern __inline int32x4_t
21460 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21461 vmovq_n_s32 (int32_t __a)
21462 {
21463   return vdupq_n_s32 (__a);
21464 }
21465
21466 __extension__ extern __inline int64x2_t
21467 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21468 vmovq_n_s64 (int64_t __a)
21469 {
21470   return vdupq_n_s64 (__a);
21471 }
21472
21473 __extension__ extern __inline uint8x16_t
21474 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21475 vmovq_n_u8 (uint8_t __a)
21476 {
21477   return vdupq_n_u8 (__a);
21478 }
21479
21480 __extension__ extern __inline uint16x8_t
21481 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21482 vmovq_n_u16 (uint16_t __a)
21483 {
21484   return vdupq_n_u16 (__a);
21485 }
21486
21487 __extension__ extern __inline uint32x4_t
21488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21489 vmovq_n_u32 (uint32_t __a)
21490 {
21491   return vdupq_n_u32 (__a);
21492 }
21493
21494 __extension__ extern __inline uint64x2_t
21495 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21496 vmovq_n_u64 (uint64_t __a)
21497 {
21498   return vdupq_n_u64 (__a);
21499 }
21500
21501 /* vmul_lane  */
21502
21503 __extension__ extern __inline float32x2_t
21504 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21505 vmul_lane_f32 (float32x2_t __a, float32x2_t __b, const int __lane)
21506 {
21507   return __a * __aarch64_vget_lane_any (__b, __lane);
21508 }
21509
21510 __extension__ extern __inline float64x1_t
21511 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21512 vmul_lane_f64 (float64x1_t __a, float64x1_t __b, const int __lane)
21513 {
21514   return __a * __b;
21515 }
21516
21517 __extension__ extern __inline int16x4_t
21518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21519 vmul_lane_s16 (int16x4_t __a, int16x4_t __b, const int __lane)
21520 {
21521   return __a * __aarch64_vget_lane_any (__b, __lane);
21522 }
21523
21524 __extension__ extern __inline int32x2_t
21525 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21526 vmul_lane_s32 (int32x2_t __a, int32x2_t __b, const int __lane)
21527 {
21528   return __a * __aarch64_vget_lane_any (__b, __lane);
21529 }
21530
21531 __extension__ extern __inline uint16x4_t
21532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21533 vmul_lane_u16 (uint16x4_t __a, uint16x4_t __b, const int __lane)
21534 {
21535   return __a * __aarch64_vget_lane_any (__b, __lane);
21536 }
21537
21538 __extension__ extern __inline uint32x2_t
21539 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21540 vmul_lane_u32 (uint32x2_t __a, uint32x2_t __b, const int __lane)
21541 {
21542   return __a * __aarch64_vget_lane_any (__b, __lane);
21543 }
21544
21545 /* vmuld_lane  */
21546
21547 __extension__ extern __inline float64_t
21548 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21549 vmuld_lane_f64 (float64_t __a, float64x1_t __b, const int __lane)
21550 {
21551   return __a * __aarch64_vget_lane_any (__b, __lane);
21552 }
21553
21554 __extension__ extern __inline float64_t
21555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21556 vmuld_laneq_f64 (float64_t __a, float64x2_t __b, const int __lane)
21557 {
21558   return __a * __aarch64_vget_lane_any (__b, __lane);
21559 }
21560
21561 /* vmuls_lane  */
21562
21563 __extension__ extern __inline float32_t
21564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21565 vmuls_lane_f32 (float32_t __a, float32x2_t __b, const int __lane)
21566 {
21567   return __a * __aarch64_vget_lane_any (__b, __lane);
21568 }
21569
21570 __extension__ extern __inline float32_t
21571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21572 vmuls_laneq_f32 (float32_t __a, float32x4_t __b, const int __lane)
21573 {
21574   return __a * __aarch64_vget_lane_any (__b, __lane);
21575 }
21576
21577 /* vmul_laneq  */
21578
21579 __extension__ extern __inline float32x2_t
21580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21581 vmul_laneq_f32 (float32x2_t __a, float32x4_t __b, const int __lane)
21582 {
21583   return __a * __aarch64_vget_lane_any (__b, __lane);
21584 }
21585
21586 __extension__ extern __inline float64x1_t
21587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21588 vmul_laneq_f64 (float64x1_t __a, float64x2_t __b, const int __lane)
21589 {
21590   return __a * __aarch64_vget_lane_any (__b, __lane);
21591 }
21592
21593 __extension__ extern __inline int16x4_t
21594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21595 vmul_laneq_s16 (int16x4_t __a, int16x8_t __b, const int __lane)
21596 {
21597   return __a * __aarch64_vget_lane_any (__b, __lane);
21598 }
21599
21600 __extension__ extern __inline int32x2_t
21601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21602 vmul_laneq_s32 (int32x2_t __a, int32x4_t __b, const int __lane)
21603 {
21604   return __a * __aarch64_vget_lane_any (__b, __lane);
21605 }
21606
21607 __extension__ extern __inline uint16x4_t
21608 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21609 vmul_laneq_u16 (uint16x4_t __a, uint16x8_t __b, const int __lane)
21610 {
21611   return __a * __aarch64_vget_lane_any (__b, __lane);
21612 }
21613
21614 __extension__ extern __inline uint32x2_t
21615 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21616 vmul_laneq_u32 (uint32x2_t __a, uint32x4_t __b, const int __lane)
21617 {
21618   return __a * __aarch64_vget_lane_any (__b, __lane);
21619 }
21620
21621 /* vmul_n  */
21622
21623 __extension__ extern __inline float64x1_t
21624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21625 vmul_n_f64  (float64x1_t __a, float64_t __b)
21626 {
21627   return (float64x1_t) { vget_lane_f64 (__a, 0) * __b };
21628 }
21629
21630 /* vmulq_lane  */
21631
21632 __extension__ extern __inline float32x4_t
21633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21634 vmulq_lane_f32 (float32x4_t __a, float32x2_t __b, const int __lane)
21635 {
21636   return __a * __aarch64_vget_lane_any (__b, __lane);
21637 }
21638
21639 __extension__ extern __inline float64x2_t
21640 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21641 vmulq_lane_f64 (float64x2_t __a, float64x1_t __b, const int __lane)
21642 {
21643   __AARCH64_LANE_CHECK (__a, __lane);
21644   return __a * __b[0];
21645 }
21646
21647 __extension__ extern __inline int16x8_t
21648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21649 vmulq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __lane)
21650 {
21651   return __a * __aarch64_vget_lane_any (__b, __lane);
21652 }
21653
21654 __extension__ extern __inline int32x4_t
21655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21656 vmulq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __lane)
21657 {
21658   return __a * __aarch64_vget_lane_any (__b, __lane);
21659 }
21660
21661 __extension__ extern __inline uint16x8_t
21662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21663 vmulq_lane_u16 (uint16x8_t __a, uint16x4_t __b, const int __lane)
21664 {
21665   return __a * __aarch64_vget_lane_any (__b, __lane);
21666 }
21667
21668 __extension__ extern __inline uint32x4_t
21669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21670 vmulq_lane_u32 (uint32x4_t __a, uint32x2_t __b, const int __lane)
21671 {
21672   return __a * __aarch64_vget_lane_any (__b, __lane);
21673 }
21674
21675 /* vmulq_laneq  */
21676
21677 __extension__ extern __inline float32x4_t
21678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21679 vmulq_laneq_f32 (float32x4_t __a, float32x4_t __b, const int __lane)
21680 {
21681   return __a * __aarch64_vget_lane_any (__b, __lane);
21682 }
21683
21684 __extension__ extern __inline float64x2_t
21685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21686 vmulq_laneq_f64 (float64x2_t __a, float64x2_t __b, const int __lane)
21687 {
21688   return __a * __aarch64_vget_lane_any (__b, __lane);
21689 }
21690
21691 __extension__ extern __inline int16x8_t
21692 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21693 vmulq_laneq_s16 (int16x8_t __a, int16x8_t __b, const int __lane)
21694 {
21695   return __a * __aarch64_vget_lane_any (__b, __lane);
21696 }
21697
21698 __extension__ extern __inline int32x4_t
21699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21700 vmulq_laneq_s32 (int32x4_t __a, int32x4_t __b, const int __lane)
21701 {
21702   return __a * __aarch64_vget_lane_any (__b, __lane);
21703 }
21704
21705 __extension__ extern __inline uint16x8_t
21706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21707 vmulq_laneq_u16 (uint16x8_t __a, uint16x8_t __b, const int __lane)
21708 {
21709   return __a * __aarch64_vget_lane_any (__b, __lane);
21710 }
21711
21712 __extension__ extern __inline uint32x4_t
21713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21714 vmulq_laneq_u32 (uint32x4_t __a, uint32x4_t __b, const int __lane)
21715 {
21716   return __a * __aarch64_vget_lane_any (__b, __lane);
21717 }
21718
21719 /* vmul_n.  */
21720
21721 __extension__ extern __inline float32x2_t
21722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21723 vmul_n_f32 (float32x2_t __a, float32_t __b)
21724 {
21725   return __a * __b;
21726 }
21727
21728 __extension__ extern __inline float32x4_t
21729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21730 vmulq_n_f32 (float32x4_t __a, float32_t __b)
21731 {
21732   return __a * __b;
21733 }
21734
21735 __extension__ extern __inline float64x2_t
21736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21737 vmulq_n_f64 (float64x2_t __a, float64_t __b)
21738 {
21739   return __a * __b;
21740 }
21741
21742 __extension__ extern __inline int16x4_t
21743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21744 vmul_n_s16 (int16x4_t __a, int16_t __b)
21745 {
21746   return __a * __b;
21747 }
21748
21749 __extension__ extern __inline int16x8_t
21750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21751 vmulq_n_s16 (int16x8_t __a, int16_t __b)
21752 {
21753   return __a * __b;
21754 }
21755
21756 __extension__ extern __inline int32x2_t
21757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21758 vmul_n_s32 (int32x2_t __a, int32_t __b)
21759 {
21760   return __a * __b;
21761 }
21762
21763 __extension__ extern __inline int32x4_t
21764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21765 vmulq_n_s32 (int32x4_t __a, int32_t __b)
21766 {
21767   return __a * __b;
21768 }
21769
21770 __extension__ extern __inline uint16x4_t
21771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21772 vmul_n_u16 (uint16x4_t __a, uint16_t __b)
21773 {
21774   return __a * __b;
21775 }
21776
21777 __extension__ extern __inline uint16x8_t
21778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21779 vmulq_n_u16 (uint16x8_t __a, uint16_t __b)
21780 {
21781   return __a * __b;
21782 }
21783
21784 __extension__ extern __inline uint32x2_t
21785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21786 vmul_n_u32 (uint32x2_t __a, uint32_t __b)
21787 {
21788   return __a * __b;
21789 }
21790
21791 __extension__ extern __inline uint32x4_t
21792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21793 vmulq_n_u32 (uint32x4_t __a, uint32_t __b)
21794 {
21795   return __a * __b;
21796 }
21797
21798 /* vmvn  */
21799
21800 __extension__ extern __inline poly8x8_t
21801 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21802 vmvn_p8 (poly8x8_t __a)
21803 {
21804   return (poly8x8_t) ~((int8x8_t) __a);
21805 }
21806
21807 __extension__ extern __inline int8x8_t
21808 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21809 vmvn_s8 (int8x8_t __a)
21810 {
21811   return ~__a;
21812 }
21813
21814 __extension__ extern __inline int16x4_t
21815 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21816 vmvn_s16 (int16x4_t __a)
21817 {
21818   return ~__a;
21819 }
21820
21821 __extension__ extern __inline int32x2_t
21822 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21823 vmvn_s32 (int32x2_t __a)
21824 {
21825   return ~__a;
21826 }
21827
21828 __extension__ extern __inline uint8x8_t
21829 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21830 vmvn_u8 (uint8x8_t __a)
21831 {
21832   return ~__a;
21833 }
21834
21835 __extension__ extern __inline uint16x4_t
21836 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21837 vmvn_u16 (uint16x4_t __a)
21838 {
21839   return ~__a;
21840 }
21841
21842 __extension__ extern __inline uint32x2_t
21843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21844 vmvn_u32 (uint32x2_t __a)
21845 {
21846   return ~__a;
21847 }
21848
21849 __extension__ extern __inline poly8x16_t
21850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21851 vmvnq_p8 (poly8x16_t __a)
21852 {
21853   return (poly8x16_t) ~((int8x16_t) __a);
21854 }
21855
21856 __extension__ extern __inline int8x16_t
21857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21858 vmvnq_s8 (int8x16_t __a)
21859 {
21860   return ~__a;
21861 }
21862
21863 __extension__ extern __inline int16x8_t
21864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21865 vmvnq_s16 (int16x8_t __a)
21866 {
21867   return ~__a;
21868 }
21869
21870 __extension__ extern __inline int32x4_t
21871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21872 vmvnq_s32 (int32x4_t __a)
21873 {
21874   return ~__a;
21875 }
21876
21877 __extension__ extern __inline uint8x16_t
21878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21879 vmvnq_u8 (uint8x16_t __a)
21880 {
21881   return ~__a;
21882 }
21883
21884 __extension__ extern __inline uint16x8_t
21885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21886 vmvnq_u16 (uint16x8_t __a)
21887 {
21888   return ~__a;
21889 }
21890
21891 __extension__ extern __inline uint32x4_t
21892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21893 vmvnq_u32 (uint32x4_t __a)
21894 {
21895   return ~__a;
21896 }
21897
21898 /* vneg  */
21899
21900 __extension__ extern __inline float32x2_t
21901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21902 vneg_f32 (float32x2_t __a)
21903 {
21904   return -__a;
21905 }
21906
21907 __extension__ extern __inline float64x1_t
21908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21909 vneg_f64 (float64x1_t __a)
21910 {
21911   return -__a;
21912 }
21913
21914 __extension__ extern __inline int8x8_t
21915 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21916 vneg_s8 (int8x8_t __a)
21917 {
21918   return -__a;
21919 }
21920
21921 __extension__ extern __inline int16x4_t
21922 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21923 vneg_s16 (int16x4_t __a)
21924 {
21925   return -__a;
21926 }
21927
21928 __extension__ extern __inline int32x2_t
21929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21930 vneg_s32 (int32x2_t __a)
21931 {
21932   return -__a;
21933 }
21934
21935 __extension__ extern __inline int64x1_t
21936 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21937 vneg_s64 (int64x1_t __a)
21938 {
21939   return -__a;
21940 }
21941
21942 __extension__ extern __inline float32x4_t
21943 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21944 vnegq_f32 (float32x4_t __a)
21945 {
21946   return -__a;
21947 }
21948
21949 __extension__ extern __inline float64x2_t
21950 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21951 vnegq_f64 (float64x2_t __a)
21952 {
21953   return -__a;
21954 }
21955
21956 __extension__ extern __inline int8x16_t
21957 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21958 vnegq_s8 (int8x16_t __a)
21959 {
21960   return -__a;
21961 }
21962
21963 __extension__ extern __inline int16x8_t
21964 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21965 vnegq_s16 (int16x8_t __a)
21966 {
21967   return -__a;
21968 }
21969
21970 __extension__ extern __inline int32x4_t
21971 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21972 vnegq_s32 (int32x4_t __a)
21973 {
21974   return -__a;
21975 }
21976
21977 __extension__ extern __inline int64x2_t
21978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21979 vnegq_s64 (int64x2_t __a)
21980 {
21981   return -__a;
21982 }
21983
21984 /* vpadd  */
21985
21986 __extension__ extern __inline float32x2_t
21987 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21988 vpadd_f32 (float32x2_t __a, float32x2_t __b)
21989 {
21990   return __builtin_aarch64_faddpv2sf (__a, __b);
21991 }
21992
21993 __extension__ extern __inline float32x4_t
21994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
21995 vpaddq_f32 (float32x4_t __a, float32x4_t __b)
21996 {
21997   return __builtin_aarch64_faddpv4sf (__a, __b);
21998 }
21999
22000 __extension__ extern __inline float64x2_t
22001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22002 vpaddq_f64 (float64x2_t __a, float64x2_t __b)
22003 {
22004   return __builtin_aarch64_faddpv2df (__a, __b);
22005 }
22006
22007 __extension__ extern __inline int8x8_t
22008 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22009 vpadd_s8 (int8x8_t __a, int8x8_t __b)
22010 {
22011   return __builtin_aarch64_addpv8qi (__a, __b);
22012 }
22013
22014 __extension__ extern __inline int16x4_t
22015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22016 vpadd_s16 (int16x4_t __a, int16x4_t __b)
22017 {
22018   return __builtin_aarch64_addpv4hi (__a, __b);
22019 }
22020
22021 __extension__ extern __inline int32x2_t
22022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22023 vpadd_s32 (int32x2_t __a, int32x2_t __b)
22024 {
22025   return __builtin_aarch64_addpv2si (__a, __b);
22026 }
22027
22028 __extension__ extern __inline uint8x8_t
22029 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22030 vpadd_u8 (uint8x8_t __a, uint8x8_t __b)
22031 {
22032   return (uint8x8_t) __builtin_aarch64_addpv8qi ((int8x8_t) __a,
22033                                                  (int8x8_t) __b);
22034 }
22035
22036 __extension__ extern __inline uint16x4_t
22037 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22038 vpadd_u16 (uint16x4_t __a, uint16x4_t __b)
22039 {
22040   return (uint16x4_t) __builtin_aarch64_addpv4hi ((int16x4_t) __a,
22041                                                   (int16x4_t) __b);
22042 }
22043
22044 __extension__ extern __inline uint32x2_t
22045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22046 vpadd_u32 (uint32x2_t __a, uint32x2_t __b)
22047 {
22048   return (uint32x2_t) __builtin_aarch64_addpv2si ((int32x2_t) __a,
22049                                                   (int32x2_t) __b);
22050 }
22051
22052 __extension__ extern __inline float32_t
22053 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22054 vpadds_f32 (float32x2_t __a)
22055 {
22056   return __builtin_aarch64_reduc_plus_scal_v2sf (__a);
22057 }
22058
22059 __extension__ extern __inline float64_t
22060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22061 vpaddd_f64 (float64x2_t __a)
22062 {
22063   return __builtin_aarch64_reduc_plus_scal_v2df (__a);
22064 }
22065
22066 __extension__ extern __inline int64_t
22067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22068 vpaddd_s64 (int64x2_t __a)
22069 {
22070   return __builtin_aarch64_addpdi (__a);
22071 }
22072
22073 __extension__ extern __inline uint64_t
22074 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22075 vpaddd_u64 (uint64x2_t __a)
22076 {
22077   return __builtin_aarch64_addpdi ((int64x2_t) __a);
22078 }
22079
22080 /* vqabs */
22081
22082 __extension__ extern __inline int64x2_t
22083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22084 vqabsq_s64 (int64x2_t __a)
22085 {
22086   return (int64x2_t) __builtin_aarch64_sqabsv2di (__a);
22087 }
22088
22089 __extension__ extern __inline int8_t
22090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22091 vqabsb_s8 (int8_t __a)
22092 {
22093   return (int8_t) __builtin_aarch64_sqabsqi (__a);
22094 }
22095
22096 __extension__ extern __inline int16_t
22097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22098 vqabsh_s16 (int16_t __a)
22099 {
22100   return (int16_t) __builtin_aarch64_sqabshi (__a);
22101 }
22102
22103 __extension__ extern __inline int32_t
22104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22105 vqabss_s32 (int32_t __a)
22106 {
22107   return (int32_t) __builtin_aarch64_sqabssi (__a);
22108 }
22109
22110 __extension__ extern __inline int64_t
22111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22112 vqabsd_s64 (int64_t __a)
22113 {
22114   return __builtin_aarch64_sqabsdi (__a);
22115 }
22116
22117 /* vqadd */
22118
22119 __extension__ extern __inline int8_t
22120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22121 vqaddb_s8 (int8_t __a, int8_t __b)
22122 {
22123   return (int8_t) __builtin_aarch64_sqaddqi (__a, __b);
22124 }
22125
22126 __extension__ extern __inline int16_t
22127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22128 vqaddh_s16 (int16_t __a, int16_t __b)
22129 {
22130   return (int16_t) __builtin_aarch64_sqaddhi (__a, __b);
22131 }
22132
22133 __extension__ extern __inline int32_t
22134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22135 vqadds_s32 (int32_t __a, int32_t __b)
22136 {
22137   return (int32_t) __builtin_aarch64_sqaddsi (__a, __b);
22138 }
22139
22140 __extension__ extern __inline int64_t
22141 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22142 vqaddd_s64 (int64_t __a, int64_t __b)
22143 {
22144   return __builtin_aarch64_sqadddi (__a, __b);
22145 }
22146
22147 __extension__ extern __inline uint8_t
22148 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22149 vqaddb_u8 (uint8_t __a, uint8_t __b)
22150 {
22151   return (uint8_t) __builtin_aarch64_uqaddqi_uuu (__a, __b);
22152 }
22153
22154 __extension__ extern __inline uint16_t
22155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22156 vqaddh_u16 (uint16_t __a, uint16_t __b)
22157 {
22158   return (uint16_t) __builtin_aarch64_uqaddhi_uuu (__a, __b);
22159 }
22160
22161 __extension__ extern __inline uint32_t
22162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22163 vqadds_u32 (uint32_t __a, uint32_t __b)
22164 {
22165   return (uint32_t) __builtin_aarch64_uqaddsi_uuu (__a, __b);
22166 }
22167
22168 __extension__ extern __inline uint64_t
22169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22170 vqaddd_u64 (uint64_t __a, uint64_t __b)
22171 {
22172   return __builtin_aarch64_uqadddi_uuu (__a, __b);
22173 }
22174
22175 /* vqdmlal */
22176
22177 __extension__ extern __inline int32x4_t
22178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22179 vqdmlal_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c)
22180 {
22181   return __builtin_aarch64_sqdmlalv4hi (__a, __b, __c);
22182 }
22183
22184 __extension__ extern __inline int32x4_t
22185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22186 vqdmlal_high_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c)
22187 {
22188   return __builtin_aarch64_sqdmlal2v8hi (__a, __b, __c);
22189 }
22190
22191 __extension__ extern __inline int32x4_t
22192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22193 vqdmlal_high_lane_s16 (int32x4_t __a, int16x8_t __b, int16x4_t __c,
22194                        int const __d)
22195 {
22196   return __builtin_aarch64_sqdmlal2_lanev8hi (__a, __b, __c, __d);
22197 }
22198
22199 __extension__ extern __inline int32x4_t
22200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22201 vqdmlal_high_laneq_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c,
22202                         int const __d)
22203 {
22204   return __builtin_aarch64_sqdmlal2_laneqv8hi (__a, __b, __c, __d);
22205 }
22206
22207 __extension__ extern __inline int32x4_t
22208 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22209 vqdmlal_high_n_s16 (int32x4_t __a, int16x8_t __b, int16_t __c)
22210 {
22211   return __builtin_aarch64_sqdmlal2_nv8hi (__a, __b, __c);
22212 }
22213
22214 __extension__ extern __inline int32x4_t
22215 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22216 vqdmlal_lane_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c, int const __d)
22217 {
22218   return __builtin_aarch64_sqdmlal_lanev4hi (__a, __b, __c, __d);
22219 }
22220
22221 __extension__ extern __inline int32x4_t
22222 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22223 vqdmlal_laneq_s16 (int32x4_t __a, int16x4_t __b, int16x8_t __c, int const __d)
22224 {
22225   return __builtin_aarch64_sqdmlal_laneqv4hi (__a, __b, __c, __d);
22226 }
22227
22228 __extension__ extern __inline int32x4_t
22229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22230 vqdmlal_n_s16 (int32x4_t __a, int16x4_t __b, int16_t __c)
22231 {
22232   return __builtin_aarch64_sqdmlal_nv4hi (__a, __b, __c);
22233 }
22234
22235 __extension__ extern __inline int64x2_t
22236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22237 vqdmlal_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c)
22238 {
22239   return __builtin_aarch64_sqdmlalv2si (__a, __b, __c);
22240 }
22241
22242 __extension__ extern __inline int64x2_t
22243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22244 vqdmlal_high_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c)
22245 {
22246   return __builtin_aarch64_sqdmlal2v4si (__a, __b, __c);
22247 }
22248
22249 __extension__ extern __inline int64x2_t
22250 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22251 vqdmlal_high_lane_s32 (int64x2_t __a, int32x4_t __b, int32x2_t __c,
22252                        int const __d)
22253 {
22254   return __builtin_aarch64_sqdmlal2_lanev4si (__a, __b, __c, __d);
22255 }
22256
22257 __extension__ extern __inline int64x2_t
22258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22259 vqdmlal_high_laneq_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c,
22260                         int const __d)
22261 {
22262   return __builtin_aarch64_sqdmlal2_laneqv4si (__a, __b, __c, __d);
22263 }
22264
22265 __extension__ extern __inline int64x2_t
22266 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22267 vqdmlal_high_n_s32 (int64x2_t __a, int32x4_t __b, int32_t __c)
22268 {
22269   return __builtin_aarch64_sqdmlal2_nv4si (__a, __b, __c);
22270 }
22271
22272 __extension__ extern __inline int64x2_t
22273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22274 vqdmlal_lane_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c, int const __d)
22275 {
22276   return __builtin_aarch64_sqdmlal_lanev2si (__a, __b, __c, __d);
22277 }
22278
22279 __extension__ extern __inline int64x2_t
22280 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22281 vqdmlal_laneq_s32 (int64x2_t __a, int32x2_t __b, int32x4_t __c, int const __d)
22282 {
22283   return __builtin_aarch64_sqdmlal_laneqv2si (__a, __b, __c, __d);
22284 }
22285
22286 __extension__ extern __inline int64x2_t
22287 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22288 vqdmlal_n_s32 (int64x2_t __a, int32x2_t __b, int32_t __c)
22289 {
22290   return __builtin_aarch64_sqdmlal_nv2si (__a, __b, __c);
22291 }
22292
22293 __extension__ extern __inline int32_t
22294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22295 vqdmlalh_s16 (int32_t __a, int16_t __b, int16_t __c)
22296 {
22297   return __builtin_aarch64_sqdmlalhi (__a, __b, __c);
22298 }
22299
22300 __extension__ extern __inline int32_t
22301 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22302 vqdmlalh_lane_s16 (int32_t __a, int16_t __b, int16x4_t __c, const int __d)
22303 {
22304   return __builtin_aarch64_sqdmlal_lanehi (__a, __b, __c, __d);
22305 }
22306
22307 __extension__ extern __inline int32_t
22308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22309 vqdmlalh_laneq_s16 (int32_t __a, int16_t __b, int16x8_t __c, const int __d)
22310 {
22311   return __builtin_aarch64_sqdmlal_laneqhi (__a, __b, __c, __d);
22312 }
22313
22314 __extension__ extern __inline int64_t
22315 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22316 vqdmlals_s32 (int64_t __a, int32_t __b, int32_t __c)
22317 {
22318   return __builtin_aarch64_sqdmlalsi (__a, __b, __c);
22319 }
22320
22321 __extension__ extern __inline int64_t
22322 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22323 vqdmlals_lane_s32 (int64_t __a, int32_t __b, int32x2_t __c, const int __d)
22324 {
22325   return __builtin_aarch64_sqdmlal_lanesi (__a, __b, __c, __d);
22326 }
22327
22328 __extension__ extern __inline int64_t
22329 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22330 vqdmlals_laneq_s32 (int64_t __a, int32_t __b, int32x4_t __c, const int __d)
22331 {
22332   return __builtin_aarch64_sqdmlal_laneqsi (__a, __b, __c, __d);
22333 }
22334
22335 /* vqdmlsl */
22336
22337 __extension__ extern __inline int32x4_t
22338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22339 vqdmlsl_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c)
22340 {
22341   return __builtin_aarch64_sqdmlslv4hi (__a, __b, __c);
22342 }
22343
22344 __extension__ extern __inline int32x4_t
22345 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22346 vqdmlsl_high_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c)
22347 {
22348   return __builtin_aarch64_sqdmlsl2v8hi (__a, __b, __c);
22349 }
22350
22351 __extension__ extern __inline int32x4_t
22352 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22353 vqdmlsl_high_lane_s16 (int32x4_t __a, int16x8_t __b, int16x4_t __c,
22354                        int const __d)
22355 {
22356   return __builtin_aarch64_sqdmlsl2_lanev8hi (__a, __b, __c, __d);
22357 }
22358
22359 __extension__ extern __inline int32x4_t
22360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22361 vqdmlsl_high_laneq_s16 (int32x4_t __a, int16x8_t __b, int16x8_t __c,
22362                         int const __d)
22363 {
22364   return __builtin_aarch64_sqdmlsl2_laneqv8hi (__a, __b, __c, __d);
22365 }
22366
22367 __extension__ extern __inline int32x4_t
22368 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22369 vqdmlsl_high_n_s16 (int32x4_t __a, int16x8_t __b, int16_t __c)
22370 {
22371   return __builtin_aarch64_sqdmlsl2_nv8hi (__a, __b, __c);
22372 }
22373
22374 __extension__ extern __inline int32x4_t
22375 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22376 vqdmlsl_lane_s16 (int32x4_t __a, int16x4_t __b, int16x4_t __c, int const __d)
22377 {
22378   return __builtin_aarch64_sqdmlsl_lanev4hi (__a, __b, __c, __d);
22379 }
22380
22381 __extension__ extern __inline int32x4_t
22382 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22383 vqdmlsl_laneq_s16 (int32x4_t __a, int16x4_t __b, int16x8_t __c, int const __d)
22384 {
22385   return __builtin_aarch64_sqdmlsl_laneqv4hi (__a, __b, __c, __d);
22386 }
22387
22388 __extension__ extern __inline int32x4_t
22389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22390 vqdmlsl_n_s16 (int32x4_t __a, int16x4_t __b, int16_t __c)
22391 {
22392   return __builtin_aarch64_sqdmlsl_nv4hi (__a, __b, __c);
22393 }
22394
22395 __extension__ extern __inline int64x2_t
22396 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22397 vqdmlsl_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c)
22398 {
22399   return __builtin_aarch64_sqdmlslv2si (__a, __b, __c);
22400 }
22401
22402 __extension__ extern __inline int64x2_t
22403 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22404 vqdmlsl_high_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c)
22405 {
22406   return __builtin_aarch64_sqdmlsl2v4si (__a, __b, __c);
22407 }
22408
22409 __extension__ extern __inline int64x2_t
22410 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22411 vqdmlsl_high_lane_s32 (int64x2_t __a, int32x4_t __b, int32x2_t __c,
22412                        int const __d)
22413 {
22414   return __builtin_aarch64_sqdmlsl2_lanev4si (__a, __b, __c, __d);
22415 }
22416
22417 __extension__ extern __inline int64x2_t
22418 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22419 vqdmlsl_high_laneq_s32 (int64x2_t __a, int32x4_t __b, int32x4_t __c,
22420                         int const __d)
22421 {
22422   return __builtin_aarch64_sqdmlsl2_laneqv4si (__a, __b, __c, __d);
22423 }
22424
22425 __extension__ extern __inline int64x2_t
22426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22427 vqdmlsl_high_n_s32 (int64x2_t __a, int32x4_t __b, int32_t __c)
22428 {
22429   return __builtin_aarch64_sqdmlsl2_nv4si (__a, __b, __c);
22430 }
22431
22432 __extension__ extern __inline int64x2_t
22433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22434 vqdmlsl_lane_s32 (int64x2_t __a, int32x2_t __b, int32x2_t __c, int const __d)
22435 {
22436   return __builtin_aarch64_sqdmlsl_lanev2si (__a, __b, __c, __d);
22437 }
22438
22439 __extension__ extern __inline int64x2_t
22440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22441 vqdmlsl_laneq_s32 (int64x2_t __a, int32x2_t __b, int32x4_t __c, int const __d)
22442 {
22443   return __builtin_aarch64_sqdmlsl_laneqv2si (__a, __b, __c, __d);
22444 }
22445
22446 __extension__ extern __inline int64x2_t
22447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22448 vqdmlsl_n_s32 (int64x2_t __a, int32x2_t __b, int32_t __c)
22449 {
22450   return __builtin_aarch64_sqdmlsl_nv2si (__a, __b, __c);
22451 }
22452
22453 __extension__ extern __inline int32_t
22454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22455 vqdmlslh_s16 (int32_t __a, int16_t __b, int16_t __c)
22456 {
22457   return __builtin_aarch64_sqdmlslhi (__a, __b, __c);
22458 }
22459
22460 __extension__ extern __inline int32_t
22461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22462 vqdmlslh_lane_s16 (int32_t __a, int16_t __b, int16x4_t __c, const int __d)
22463 {
22464   return __builtin_aarch64_sqdmlsl_lanehi (__a, __b, __c, __d);
22465 }
22466
22467 __extension__ extern __inline int32_t
22468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22469 vqdmlslh_laneq_s16 (int32_t __a, int16_t __b, int16x8_t __c, const int __d)
22470 {
22471   return __builtin_aarch64_sqdmlsl_laneqhi (__a, __b, __c, __d);
22472 }
22473
22474 __extension__ extern __inline int64_t
22475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22476 vqdmlsls_s32 (int64_t __a, int32_t __b, int32_t __c)
22477 {
22478   return __builtin_aarch64_sqdmlslsi (__a, __b, __c);
22479 }
22480
22481 __extension__ extern __inline int64_t
22482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22483 vqdmlsls_lane_s32 (int64_t __a, int32_t __b, int32x2_t __c, const int __d)
22484 {
22485   return __builtin_aarch64_sqdmlsl_lanesi (__a, __b, __c, __d);
22486 }
22487
22488 __extension__ extern __inline int64_t
22489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22490 vqdmlsls_laneq_s32 (int64_t __a, int32_t __b, int32x4_t __c, const int __d)
22491 {
22492   return __builtin_aarch64_sqdmlsl_laneqsi (__a, __b, __c, __d);
22493 }
22494
22495 /* vqdmulh */
22496
22497 __extension__ extern __inline int16x4_t
22498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22499 vqdmulh_lane_s16 (int16x4_t __a, int16x4_t __b, const int __c)
22500 {
22501   return __builtin_aarch64_sqdmulh_lanev4hi (__a, __b, __c);
22502 }
22503
22504 __extension__ extern __inline int32x2_t
22505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22506 vqdmulh_lane_s32 (int32x2_t __a, int32x2_t __b, const int __c)
22507 {
22508   return __builtin_aarch64_sqdmulh_lanev2si (__a, __b, __c);
22509 }
22510
22511 __extension__ extern __inline int16x8_t
22512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22513 vqdmulhq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __c)
22514 {
22515   return __builtin_aarch64_sqdmulh_lanev8hi (__a, __b, __c);
22516 }
22517
22518 __extension__ extern __inline int32x4_t
22519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22520 vqdmulhq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __c)
22521 {
22522   return __builtin_aarch64_sqdmulh_lanev4si (__a, __b, __c);
22523 }
22524
22525 __extension__ extern __inline int16_t
22526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22527 vqdmulhh_s16 (int16_t __a, int16_t __b)
22528 {
22529   return (int16_t) __builtin_aarch64_sqdmulhhi (__a, __b);
22530 }
22531
22532 __extension__ extern __inline int16_t
22533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22534 vqdmulhh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
22535 {
22536   return __builtin_aarch64_sqdmulh_lanehi (__a, __b, __c);
22537 }
22538
22539 __extension__ extern __inline int16_t
22540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22541 vqdmulhh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
22542 {
22543   return __builtin_aarch64_sqdmulh_laneqhi (__a, __b, __c);
22544 }
22545
22546 __extension__ extern __inline int32_t
22547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22548 vqdmulhs_s32 (int32_t __a, int32_t __b)
22549 {
22550   return (int32_t) __builtin_aarch64_sqdmulhsi (__a, __b);
22551 }
22552
22553 __extension__ extern __inline int32_t
22554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22555 vqdmulhs_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
22556 {
22557   return __builtin_aarch64_sqdmulh_lanesi (__a, __b, __c);
22558 }
22559
22560 __extension__ extern __inline int32_t
22561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22562 vqdmulhs_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
22563 {
22564   return __builtin_aarch64_sqdmulh_laneqsi (__a, __b, __c);
22565 }
22566
22567 /* vqdmull */
22568
22569 __extension__ extern __inline int32x4_t
22570 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22571 vqdmull_s16 (int16x4_t __a, int16x4_t __b)
22572 {
22573   return __builtin_aarch64_sqdmullv4hi (__a, __b);
22574 }
22575
22576 __extension__ extern __inline int32x4_t
22577 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22578 vqdmull_high_s16 (int16x8_t __a, int16x8_t __b)
22579 {
22580   return __builtin_aarch64_sqdmull2v8hi (__a, __b);
22581 }
22582
22583 __extension__ extern __inline int32x4_t
22584 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22585 vqdmull_high_lane_s16 (int16x8_t __a, int16x4_t __b, int const __c)
22586 {
22587   return __builtin_aarch64_sqdmull2_lanev8hi (__a, __b,__c);
22588 }
22589
22590 __extension__ extern __inline int32x4_t
22591 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22592 vqdmull_high_laneq_s16 (int16x8_t __a, int16x8_t __b, int const __c)
22593 {
22594   return __builtin_aarch64_sqdmull2_laneqv8hi (__a, __b,__c);
22595 }
22596
22597 __extension__ extern __inline int32x4_t
22598 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22599 vqdmull_high_n_s16 (int16x8_t __a, int16_t __b)
22600 {
22601   return __builtin_aarch64_sqdmull2_nv8hi (__a, __b);
22602 }
22603
22604 __extension__ extern __inline int32x4_t
22605 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22606 vqdmull_lane_s16 (int16x4_t __a, int16x4_t __b, int const __c)
22607 {
22608   return __builtin_aarch64_sqdmull_lanev4hi (__a, __b, __c);
22609 }
22610
22611 __extension__ extern __inline int32x4_t
22612 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22613 vqdmull_laneq_s16 (int16x4_t __a, int16x8_t __b, int const __c)
22614 {
22615   return __builtin_aarch64_sqdmull_laneqv4hi (__a, __b, __c);
22616 }
22617
22618 __extension__ extern __inline int32x4_t
22619 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22620 vqdmull_n_s16 (int16x4_t __a, int16_t __b)
22621 {
22622   return __builtin_aarch64_sqdmull_nv4hi (__a, __b);
22623 }
22624
22625 __extension__ extern __inline int64x2_t
22626 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22627 vqdmull_s32 (int32x2_t __a, int32x2_t __b)
22628 {
22629   return __builtin_aarch64_sqdmullv2si (__a, __b);
22630 }
22631
22632 __extension__ extern __inline int64x2_t
22633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22634 vqdmull_high_s32 (int32x4_t __a, int32x4_t __b)
22635 {
22636   return __builtin_aarch64_sqdmull2v4si (__a, __b);
22637 }
22638
22639 __extension__ extern __inline int64x2_t
22640 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22641 vqdmull_high_lane_s32 (int32x4_t __a, int32x2_t __b, int const __c)
22642 {
22643   return __builtin_aarch64_sqdmull2_lanev4si (__a, __b, __c);
22644 }
22645
22646 __extension__ extern __inline int64x2_t
22647 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22648 vqdmull_high_laneq_s32 (int32x4_t __a, int32x4_t __b, int const __c)
22649 {
22650   return __builtin_aarch64_sqdmull2_laneqv4si (__a, __b, __c);
22651 }
22652
22653 __extension__ extern __inline int64x2_t
22654 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22655 vqdmull_high_n_s32 (int32x4_t __a, int32_t __b)
22656 {
22657   return __builtin_aarch64_sqdmull2_nv4si (__a, __b);
22658 }
22659
22660 __extension__ extern __inline int64x2_t
22661 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22662 vqdmull_lane_s32 (int32x2_t __a, int32x2_t __b, int const __c)
22663 {
22664   return __builtin_aarch64_sqdmull_lanev2si (__a, __b, __c);
22665 }
22666
22667 __extension__ extern __inline int64x2_t
22668 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22669 vqdmull_laneq_s32 (int32x2_t __a, int32x4_t __b, int const __c)
22670 {
22671   return __builtin_aarch64_sqdmull_laneqv2si (__a, __b, __c);
22672 }
22673
22674 __extension__ extern __inline int64x2_t
22675 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22676 vqdmull_n_s32 (int32x2_t __a, int32_t __b)
22677 {
22678   return __builtin_aarch64_sqdmull_nv2si (__a, __b);
22679 }
22680
22681 __extension__ extern __inline int32_t
22682 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22683 vqdmullh_s16 (int16_t __a, int16_t __b)
22684 {
22685   return (int32_t) __builtin_aarch64_sqdmullhi (__a, __b);
22686 }
22687
22688 __extension__ extern __inline int32_t
22689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22690 vqdmullh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
22691 {
22692   return __builtin_aarch64_sqdmull_lanehi (__a, __b, __c);
22693 }
22694
22695 __extension__ extern __inline int32_t
22696 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22697 vqdmullh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
22698 {
22699   return __builtin_aarch64_sqdmull_laneqhi (__a, __b, __c);
22700 }
22701
22702 __extension__ extern __inline int64_t
22703 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22704 vqdmulls_s32 (int32_t __a, int32_t __b)
22705 {
22706   return __builtin_aarch64_sqdmullsi (__a, __b);
22707 }
22708
22709 __extension__ extern __inline int64_t
22710 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22711 vqdmulls_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
22712 {
22713   return __builtin_aarch64_sqdmull_lanesi (__a, __b, __c);
22714 }
22715
22716 __extension__ extern __inline int64_t
22717 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22718 vqdmulls_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
22719 {
22720   return __builtin_aarch64_sqdmull_laneqsi (__a, __b, __c);
22721 }
22722
22723 /* vqmovn */
22724
22725 __extension__ extern __inline int8x8_t
22726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22727 vqmovn_s16 (int16x8_t __a)
22728 {
22729   return (int8x8_t) __builtin_aarch64_sqmovnv8hi (__a);
22730 }
22731
22732 __extension__ extern __inline int16x4_t
22733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22734 vqmovn_s32 (int32x4_t __a)
22735 {
22736   return (int16x4_t) __builtin_aarch64_sqmovnv4si (__a);
22737 }
22738
22739 __extension__ extern __inline int32x2_t
22740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22741 vqmovn_s64 (int64x2_t __a)
22742 {
22743   return (int32x2_t) __builtin_aarch64_sqmovnv2di (__a);
22744 }
22745
22746 __extension__ extern __inline uint8x8_t
22747 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22748 vqmovn_u16 (uint16x8_t __a)
22749 {
22750   return (uint8x8_t) __builtin_aarch64_uqmovnv8hi ((int16x8_t) __a);
22751 }
22752
22753 __extension__ extern __inline uint16x4_t
22754 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22755 vqmovn_u32 (uint32x4_t __a)
22756 {
22757   return (uint16x4_t) __builtin_aarch64_uqmovnv4si ((int32x4_t) __a);
22758 }
22759
22760 __extension__ extern __inline uint32x2_t
22761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22762 vqmovn_u64 (uint64x2_t __a)
22763 {
22764   return (uint32x2_t) __builtin_aarch64_uqmovnv2di ((int64x2_t) __a);
22765 }
22766
22767 __extension__ extern __inline int8_t
22768 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22769 vqmovnh_s16 (int16_t __a)
22770 {
22771   return (int8_t) __builtin_aarch64_sqmovnhi (__a);
22772 }
22773
22774 __extension__ extern __inline int16_t
22775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22776 vqmovns_s32 (int32_t __a)
22777 {
22778   return (int16_t) __builtin_aarch64_sqmovnsi (__a);
22779 }
22780
22781 __extension__ extern __inline int32_t
22782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22783 vqmovnd_s64 (int64_t __a)
22784 {
22785   return (int32_t) __builtin_aarch64_sqmovndi (__a);
22786 }
22787
22788 __extension__ extern __inline uint8_t
22789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22790 vqmovnh_u16 (uint16_t __a)
22791 {
22792   return (uint8_t) __builtin_aarch64_uqmovnhi (__a);
22793 }
22794
22795 __extension__ extern __inline uint16_t
22796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22797 vqmovns_u32 (uint32_t __a)
22798 {
22799   return (uint16_t) __builtin_aarch64_uqmovnsi (__a);
22800 }
22801
22802 __extension__ extern __inline uint32_t
22803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22804 vqmovnd_u64 (uint64_t __a)
22805 {
22806   return (uint32_t) __builtin_aarch64_uqmovndi (__a);
22807 }
22808
22809 /* vqmovun */
22810
22811 __extension__ extern __inline uint8x8_t
22812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22813 vqmovun_s16 (int16x8_t __a)
22814 {
22815   return (uint8x8_t) __builtin_aarch64_sqmovunv8hi (__a);
22816 }
22817
22818 __extension__ extern __inline uint16x4_t
22819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22820 vqmovun_s32 (int32x4_t __a)
22821 {
22822   return (uint16x4_t) __builtin_aarch64_sqmovunv4si (__a);
22823 }
22824
22825 __extension__ extern __inline uint32x2_t
22826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22827 vqmovun_s64 (int64x2_t __a)
22828 {
22829   return (uint32x2_t) __builtin_aarch64_sqmovunv2di (__a);
22830 }
22831
22832 __extension__ extern __inline int8_t
22833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22834 vqmovunh_s16 (int16_t __a)
22835 {
22836   return (int8_t) __builtin_aarch64_sqmovunhi (__a);
22837 }
22838
22839 __extension__ extern __inline int16_t
22840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22841 vqmovuns_s32 (int32_t __a)
22842 {
22843   return (int16_t) __builtin_aarch64_sqmovunsi (__a);
22844 }
22845
22846 __extension__ extern __inline int32_t
22847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22848 vqmovund_s64 (int64_t __a)
22849 {
22850   return (int32_t) __builtin_aarch64_sqmovundi (__a);
22851 }
22852
22853 /* vqneg */
22854
22855 __extension__ extern __inline int64x2_t
22856 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22857 vqnegq_s64 (int64x2_t __a)
22858 {
22859   return (int64x2_t) __builtin_aarch64_sqnegv2di (__a);
22860 }
22861
22862 __extension__ extern __inline int8_t
22863 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22864 vqnegb_s8 (int8_t __a)
22865 {
22866   return (int8_t) __builtin_aarch64_sqnegqi (__a);
22867 }
22868
22869 __extension__ extern __inline int16_t
22870 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22871 vqnegh_s16 (int16_t __a)
22872 {
22873   return (int16_t) __builtin_aarch64_sqneghi (__a);
22874 }
22875
22876 __extension__ extern __inline int32_t
22877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22878 vqnegs_s32 (int32_t __a)
22879 {
22880   return (int32_t) __builtin_aarch64_sqnegsi (__a);
22881 }
22882
22883 __extension__ extern __inline int64_t
22884 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22885 vqnegd_s64 (int64_t __a)
22886 {
22887   return __builtin_aarch64_sqnegdi (__a);
22888 }
22889
22890 /* vqrdmulh */
22891
22892 __extension__ extern __inline int16x4_t
22893 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22894 vqrdmulh_lane_s16 (int16x4_t __a, int16x4_t __b, const int __c)
22895 {
22896   return  __builtin_aarch64_sqrdmulh_lanev4hi (__a, __b, __c);
22897 }
22898
22899 __extension__ extern __inline int32x2_t
22900 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22901 vqrdmulh_lane_s32 (int32x2_t __a, int32x2_t __b, const int __c)
22902 {
22903   return __builtin_aarch64_sqrdmulh_lanev2si (__a, __b, __c);
22904 }
22905
22906 __extension__ extern __inline int16x8_t
22907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22908 vqrdmulhq_lane_s16 (int16x8_t __a, int16x4_t __b, const int __c)
22909 {
22910   return __builtin_aarch64_sqrdmulh_lanev8hi (__a, __b, __c);
22911 }
22912
22913 __extension__ extern __inline int32x4_t
22914 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22915 vqrdmulhq_lane_s32 (int32x4_t __a, int32x2_t __b, const int __c)
22916 {
22917   return __builtin_aarch64_sqrdmulh_lanev4si (__a, __b, __c);
22918 }
22919
22920 __extension__ extern __inline int16_t
22921 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22922 vqrdmulhh_s16 (int16_t __a, int16_t __b)
22923 {
22924   return (int16_t) __builtin_aarch64_sqrdmulhhi (__a, __b);
22925 }
22926
22927 __extension__ extern __inline int16_t
22928 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22929 vqrdmulhh_lane_s16 (int16_t __a, int16x4_t __b, const int __c)
22930 {
22931   return __builtin_aarch64_sqrdmulh_lanehi (__a, __b, __c);
22932 }
22933
22934 __extension__ extern __inline int16_t
22935 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22936 vqrdmulhh_laneq_s16 (int16_t __a, int16x8_t __b, const int __c)
22937 {
22938   return __builtin_aarch64_sqrdmulh_laneqhi (__a, __b, __c);
22939 }
22940
22941 __extension__ extern __inline int32_t
22942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22943 vqrdmulhs_s32 (int32_t __a, int32_t __b)
22944 {
22945   return (int32_t) __builtin_aarch64_sqrdmulhsi (__a, __b);
22946 }
22947
22948 __extension__ extern __inline int32_t
22949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22950 vqrdmulhs_lane_s32 (int32_t __a, int32x2_t __b, const int __c)
22951 {
22952   return __builtin_aarch64_sqrdmulh_lanesi (__a, __b, __c);
22953 }
22954
22955 __extension__ extern __inline int32_t
22956 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22957 vqrdmulhs_laneq_s32 (int32_t __a, int32x4_t __b, const int __c)
22958 {
22959   return __builtin_aarch64_sqrdmulh_laneqsi (__a, __b, __c);
22960 }
22961
22962 /* vqrshl */
22963
22964 __extension__ extern __inline int8x8_t
22965 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22966 vqrshl_s8 (int8x8_t __a, int8x8_t __b)
22967 {
22968   return __builtin_aarch64_sqrshlv8qi (__a, __b);
22969 }
22970
22971 __extension__ extern __inline int16x4_t
22972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22973 vqrshl_s16 (int16x4_t __a, int16x4_t __b)
22974 {
22975   return __builtin_aarch64_sqrshlv4hi (__a, __b);
22976 }
22977
22978 __extension__ extern __inline int32x2_t
22979 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22980 vqrshl_s32 (int32x2_t __a, int32x2_t __b)
22981 {
22982   return __builtin_aarch64_sqrshlv2si (__a, __b);
22983 }
22984
22985 __extension__ extern __inline int64x1_t
22986 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22987 vqrshl_s64 (int64x1_t __a, int64x1_t __b)
22988 {
22989   return (int64x1_t) {__builtin_aarch64_sqrshldi (__a[0], __b[0])};
22990 }
22991
22992 __extension__ extern __inline uint8x8_t
22993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
22994 vqrshl_u8 (uint8x8_t __a, int8x8_t __b)
22995 {
22996   return __builtin_aarch64_uqrshlv8qi_uus ( __a, __b);
22997 }
22998
22999 __extension__ extern __inline uint16x4_t
23000 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23001 vqrshl_u16 (uint16x4_t __a, int16x4_t __b)
23002 {
23003   return __builtin_aarch64_uqrshlv4hi_uus ( __a, __b);
23004 }
23005
23006 __extension__ extern __inline uint32x2_t
23007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23008 vqrshl_u32 (uint32x2_t __a, int32x2_t __b)
23009 {
23010   return __builtin_aarch64_uqrshlv2si_uus ( __a, __b);
23011 }
23012
23013 __extension__ extern __inline uint64x1_t
23014 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23015 vqrshl_u64 (uint64x1_t __a, int64x1_t __b)
23016 {
23017   return (uint64x1_t) {__builtin_aarch64_uqrshldi_uus (__a[0], __b[0])};
23018 }
23019
23020 __extension__ extern __inline int8x16_t
23021 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23022 vqrshlq_s8 (int8x16_t __a, int8x16_t __b)
23023 {
23024   return __builtin_aarch64_sqrshlv16qi (__a, __b);
23025 }
23026
23027 __extension__ extern __inline int16x8_t
23028 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23029 vqrshlq_s16 (int16x8_t __a, int16x8_t __b)
23030 {
23031   return __builtin_aarch64_sqrshlv8hi (__a, __b);
23032 }
23033
23034 __extension__ extern __inline int32x4_t
23035 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23036 vqrshlq_s32 (int32x4_t __a, int32x4_t __b)
23037 {
23038   return __builtin_aarch64_sqrshlv4si (__a, __b);
23039 }
23040
23041 __extension__ extern __inline int64x2_t
23042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23043 vqrshlq_s64 (int64x2_t __a, int64x2_t __b)
23044 {
23045   return __builtin_aarch64_sqrshlv2di (__a, __b);
23046 }
23047
23048 __extension__ extern __inline uint8x16_t
23049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23050 vqrshlq_u8 (uint8x16_t __a, int8x16_t __b)
23051 {
23052   return __builtin_aarch64_uqrshlv16qi_uus ( __a, __b);
23053 }
23054
23055 __extension__ extern __inline uint16x8_t
23056 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23057 vqrshlq_u16 (uint16x8_t __a, int16x8_t __b)
23058 {
23059   return __builtin_aarch64_uqrshlv8hi_uus ( __a, __b);
23060 }
23061
23062 __extension__ extern __inline uint32x4_t
23063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23064 vqrshlq_u32 (uint32x4_t __a, int32x4_t __b)
23065 {
23066   return __builtin_aarch64_uqrshlv4si_uus ( __a, __b);
23067 }
23068
23069 __extension__ extern __inline uint64x2_t
23070 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23071 vqrshlq_u64 (uint64x2_t __a, int64x2_t __b)
23072 {
23073   return __builtin_aarch64_uqrshlv2di_uus ( __a, __b);
23074 }
23075
23076 __extension__ extern __inline int8_t
23077 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23078 vqrshlb_s8 (int8_t __a, int8_t __b)
23079 {
23080   return __builtin_aarch64_sqrshlqi (__a, __b);
23081 }
23082
23083 __extension__ extern __inline int16_t
23084 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23085 vqrshlh_s16 (int16_t __a, int16_t __b)
23086 {
23087   return __builtin_aarch64_sqrshlhi (__a, __b);
23088 }
23089
23090 __extension__ extern __inline int32_t
23091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23092 vqrshls_s32 (int32_t __a, int32_t __b)
23093 {
23094   return __builtin_aarch64_sqrshlsi (__a, __b);
23095 }
23096
23097 __extension__ extern __inline int64_t
23098 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23099 vqrshld_s64 (int64_t __a, int64_t __b)
23100 {
23101   return __builtin_aarch64_sqrshldi (__a, __b);
23102 }
23103
23104 __extension__ extern __inline uint8_t
23105 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23106 vqrshlb_u8 (uint8_t __a, uint8_t __b)
23107 {
23108   return __builtin_aarch64_uqrshlqi_uus (__a, __b);
23109 }
23110
23111 __extension__ extern __inline uint16_t
23112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23113 vqrshlh_u16 (uint16_t __a, uint16_t __b)
23114 {
23115   return __builtin_aarch64_uqrshlhi_uus (__a, __b);
23116 }
23117
23118 __extension__ extern __inline uint32_t
23119 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23120 vqrshls_u32 (uint32_t __a, uint32_t __b)
23121 {
23122   return __builtin_aarch64_uqrshlsi_uus (__a, __b);
23123 }
23124
23125 __extension__ extern __inline uint64_t
23126 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23127 vqrshld_u64 (uint64_t __a, uint64_t __b)
23128 {
23129   return __builtin_aarch64_uqrshldi_uus (__a, __b);
23130 }
23131
23132 /* vqrshrn */
23133
23134 __extension__ extern __inline int8x8_t
23135 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23136 vqrshrn_n_s16 (int16x8_t __a, const int __b)
23137 {
23138   return (int8x8_t) __builtin_aarch64_sqrshrn_nv8hi (__a, __b);
23139 }
23140
23141 __extension__ extern __inline int16x4_t
23142 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23143 vqrshrn_n_s32 (int32x4_t __a, const int __b)
23144 {
23145   return (int16x4_t) __builtin_aarch64_sqrshrn_nv4si (__a, __b);
23146 }
23147
23148 __extension__ extern __inline int32x2_t
23149 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23150 vqrshrn_n_s64 (int64x2_t __a, const int __b)
23151 {
23152   return (int32x2_t) __builtin_aarch64_sqrshrn_nv2di (__a, __b);
23153 }
23154
23155 __extension__ extern __inline uint8x8_t
23156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23157 vqrshrn_n_u16 (uint16x8_t __a, const int __b)
23158 {
23159   return __builtin_aarch64_uqrshrn_nv8hi_uus ( __a, __b);
23160 }
23161
23162 __extension__ extern __inline uint16x4_t
23163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23164 vqrshrn_n_u32 (uint32x4_t __a, const int __b)
23165 {
23166   return __builtin_aarch64_uqrshrn_nv4si_uus ( __a, __b);
23167 }
23168
23169 __extension__ extern __inline uint32x2_t
23170 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23171 vqrshrn_n_u64 (uint64x2_t __a, const int __b)
23172 {
23173   return __builtin_aarch64_uqrshrn_nv2di_uus ( __a, __b);
23174 }
23175
23176 __extension__ extern __inline int8_t
23177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23178 vqrshrnh_n_s16 (int16_t __a, const int __b)
23179 {
23180   return (int8_t) __builtin_aarch64_sqrshrn_nhi (__a, __b);
23181 }
23182
23183 __extension__ extern __inline int16_t
23184 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23185 vqrshrns_n_s32 (int32_t __a, const int __b)
23186 {
23187   return (int16_t) __builtin_aarch64_sqrshrn_nsi (__a, __b);
23188 }
23189
23190 __extension__ extern __inline int32_t
23191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23192 vqrshrnd_n_s64 (int64_t __a, const int __b)
23193 {
23194   return (int32_t) __builtin_aarch64_sqrshrn_ndi (__a, __b);
23195 }
23196
23197 __extension__ extern __inline uint8_t
23198 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23199 vqrshrnh_n_u16 (uint16_t __a, const int __b)
23200 {
23201   return __builtin_aarch64_uqrshrn_nhi_uus (__a, __b);
23202 }
23203
23204 __extension__ extern __inline uint16_t
23205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23206 vqrshrns_n_u32 (uint32_t __a, const int __b)
23207 {
23208   return __builtin_aarch64_uqrshrn_nsi_uus (__a, __b);
23209 }
23210
23211 __extension__ extern __inline uint32_t
23212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23213 vqrshrnd_n_u64 (uint64_t __a, const int __b)
23214 {
23215   return __builtin_aarch64_uqrshrn_ndi_uus (__a, __b);
23216 }
23217
23218 /* vqrshrun */
23219
23220 __extension__ extern __inline uint8x8_t
23221 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23222 vqrshrun_n_s16 (int16x8_t __a, const int __b)
23223 {
23224   return (uint8x8_t) __builtin_aarch64_sqrshrun_nv8hi (__a, __b);
23225 }
23226
23227 __extension__ extern __inline uint16x4_t
23228 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23229 vqrshrun_n_s32 (int32x4_t __a, const int __b)
23230 {
23231   return (uint16x4_t) __builtin_aarch64_sqrshrun_nv4si (__a, __b);
23232 }
23233
23234 __extension__ extern __inline uint32x2_t
23235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23236 vqrshrun_n_s64 (int64x2_t __a, const int __b)
23237 {
23238   return (uint32x2_t) __builtin_aarch64_sqrshrun_nv2di (__a, __b);
23239 }
23240
23241 __extension__ extern __inline int8_t
23242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23243 vqrshrunh_n_s16 (int16_t __a, const int __b)
23244 {
23245   return (int8_t) __builtin_aarch64_sqrshrun_nhi (__a, __b);
23246 }
23247
23248 __extension__ extern __inline int16_t
23249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23250 vqrshruns_n_s32 (int32_t __a, const int __b)
23251 {
23252   return (int16_t) __builtin_aarch64_sqrshrun_nsi (__a, __b);
23253 }
23254
23255 __extension__ extern __inline int32_t
23256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23257 vqrshrund_n_s64 (int64_t __a, const int __b)
23258 {
23259   return (int32_t) __builtin_aarch64_sqrshrun_ndi (__a, __b);
23260 }
23261
23262 /* vqshl */
23263
23264 __extension__ extern __inline int8x8_t
23265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23266 vqshl_s8 (int8x8_t __a, int8x8_t __b)
23267 {
23268   return __builtin_aarch64_sqshlv8qi (__a, __b);
23269 }
23270
23271 __extension__ extern __inline int16x4_t
23272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23273 vqshl_s16 (int16x4_t __a, int16x4_t __b)
23274 {
23275   return __builtin_aarch64_sqshlv4hi (__a, __b);
23276 }
23277
23278 __extension__ extern __inline int32x2_t
23279 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23280 vqshl_s32 (int32x2_t __a, int32x2_t __b)
23281 {
23282   return __builtin_aarch64_sqshlv2si (__a, __b);
23283 }
23284
23285 __extension__ extern __inline int64x1_t
23286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23287 vqshl_s64 (int64x1_t __a, int64x1_t __b)
23288 {
23289   return (int64x1_t) {__builtin_aarch64_sqshldi (__a[0], __b[0])};
23290 }
23291
23292 __extension__ extern __inline uint8x8_t
23293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23294 vqshl_u8 (uint8x8_t __a, int8x8_t __b)
23295 {
23296   return __builtin_aarch64_uqshlv8qi_uus ( __a, __b);
23297 }
23298
23299 __extension__ extern __inline uint16x4_t
23300 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23301 vqshl_u16 (uint16x4_t __a, int16x4_t __b)
23302 {
23303   return __builtin_aarch64_uqshlv4hi_uus ( __a, __b);
23304 }
23305
23306 __extension__ extern __inline uint32x2_t
23307 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23308 vqshl_u32 (uint32x2_t __a, int32x2_t __b)
23309 {
23310   return __builtin_aarch64_uqshlv2si_uus ( __a, __b);
23311 }
23312
23313 __extension__ extern __inline uint64x1_t
23314 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23315 vqshl_u64 (uint64x1_t __a, int64x1_t __b)
23316 {
23317   return (uint64x1_t) {__builtin_aarch64_uqshldi_uus (__a[0], __b[0])};
23318 }
23319
23320 __extension__ extern __inline int8x16_t
23321 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23322 vqshlq_s8 (int8x16_t __a, int8x16_t __b)
23323 {
23324   return __builtin_aarch64_sqshlv16qi (__a, __b);
23325 }
23326
23327 __extension__ extern __inline int16x8_t
23328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23329 vqshlq_s16 (int16x8_t __a, int16x8_t __b)
23330 {
23331   return __builtin_aarch64_sqshlv8hi (__a, __b);
23332 }
23333
23334 __extension__ extern __inline int32x4_t
23335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23336 vqshlq_s32 (int32x4_t __a, int32x4_t __b)
23337 {
23338   return __builtin_aarch64_sqshlv4si (__a, __b);
23339 }
23340
23341 __extension__ extern __inline int64x2_t
23342 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23343 vqshlq_s64 (int64x2_t __a, int64x2_t __b)
23344 {
23345   return __builtin_aarch64_sqshlv2di (__a, __b);
23346 }
23347
23348 __extension__ extern __inline uint8x16_t
23349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23350 vqshlq_u8 (uint8x16_t __a, int8x16_t __b)
23351 {
23352   return __builtin_aarch64_uqshlv16qi_uus ( __a, __b);
23353 }
23354
23355 __extension__ extern __inline uint16x8_t
23356 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23357 vqshlq_u16 (uint16x8_t __a, int16x8_t __b)
23358 {
23359   return __builtin_aarch64_uqshlv8hi_uus ( __a, __b);
23360 }
23361
23362 __extension__ extern __inline uint32x4_t
23363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23364 vqshlq_u32 (uint32x4_t __a, int32x4_t __b)
23365 {
23366   return __builtin_aarch64_uqshlv4si_uus ( __a, __b);
23367 }
23368
23369 __extension__ extern __inline uint64x2_t
23370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23371 vqshlq_u64 (uint64x2_t __a, int64x2_t __b)
23372 {
23373   return __builtin_aarch64_uqshlv2di_uus ( __a, __b);
23374 }
23375
23376 __extension__ extern __inline int8_t
23377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23378 vqshlb_s8 (int8_t __a, int8_t __b)
23379 {
23380   return __builtin_aarch64_sqshlqi (__a, __b);
23381 }
23382
23383 __extension__ extern __inline int16_t
23384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23385 vqshlh_s16 (int16_t __a, int16_t __b)
23386 {
23387   return __builtin_aarch64_sqshlhi (__a, __b);
23388 }
23389
23390 __extension__ extern __inline int32_t
23391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23392 vqshls_s32 (int32_t __a, int32_t __b)
23393 {
23394   return __builtin_aarch64_sqshlsi (__a, __b);
23395 }
23396
23397 __extension__ extern __inline int64_t
23398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23399 vqshld_s64 (int64_t __a, int64_t __b)
23400 {
23401   return __builtin_aarch64_sqshldi (__a, __b);
23402 }
23403
23404 __extension__ extern __inline uint8_t
23405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23406 vqshlb_u8 (uint8_t __a, uint8_t __b)
23407 {
23408   return __builtin_aarch64_uqshlqi_uus (__a, __b);
23409 }
23410
23411 __extension__ extern __inline uint16_t
23412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23413 vqshlh_u16 (uint16_t __a, uint16_t __b)
23414 {
23415   return __builtin_aarch64_uqshlhi_uus (__a, __b);
23416 }
23417
23418 __extension__ extern __inline uint32_t
23419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23420 vqshls_u32 (uint32_t __a, uint32_t __b)
23421 {
23422   return __builtin_aarch64_uqshlsi_uus (__a, __b);
23423 }
23424
23425 __extension__ extern __inline uint64_t
23426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23427 vqshld_u64 (uint64_t __a, uint64_t __b)
23428 {
23429   return __builtin_aarch64_uqshldi_uus (__a, __b);
23430 }
23431
23432 __extension__ extern __inline int8x8_t
23433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23434 vqshl_n_s8 (int8x8_t __a, const int __b)
23435 {
23436   return (int8x8_t) __builtin_aarch64_sqshl_nv8qi (__a, __b);
23437 }
23438
23439 __extension__ extern __inline int16x4_t
23440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23441 vqshl_n_s16 (int16x4_t __a, const int __b)
23442 {
23443   return (int16x4_t) __builtin_aarch64_sqshl_nv4hi (__a, __b);
23444 }
23445
23446 __extension__ extern __inline int32x2_t
23447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23448 vqshl_n_s32 (int32x2_t __a, const int __b)
23449 {
23450   return (int32x2_t) __builtin_aarch64_sqshl_nv2si (__a, __b);
23451 }
23452
23453 __extension__ extern __inline int64x1_t
23454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23455 vqshl_n_s64 (int64x1_t __a, const int __b)
23456 {
23457   return (int64x1_t) {__builtin_aarch64_sqshl_ndi (__a[0], __b)};
23458 }
23459
23460 __extension__ extern __inline uint8x8_t
23461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23462 vqshl_n_u8 (uint8x8_t __a, const int __b)
23463 {
23464   return __builtin_aarch64_uqshl_nv8qi_uus (__a, __b);
23465 }
23466
23467 __extension__ extern __inline uint16x4_t
23468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23469 vqshl_n_u16 (uint16x4_t __a, const int __b)
23470 {
23471   return __builtin_aarch64_uqshl_nv4hi_uus (__a, __b);
23472 }
23473
23474 __extension__ extern __inline uint32x2_t
23475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23476 vqshl_n_u32 (uint32x2_t __a, const int __b)
23477 {
23478   return __builtin_aarch64_uqshl_nv2si_uus (__a, __b);
23479 }
23480
23481 __extension__ extern __inline uint64x1_t
23482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23483 vqshl_n_u64 (uint64x1_t __a, const int __b)
23484 {
23485   return (uint64x1_t) {__builtin_aarch64_uqshl_ndi_uus (__a[0], __b)};
23486 }
23487
23488 __extension__ extern __inline int8x16_t
23489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23490 vqshlq_n_s8 (int8x16_t __a, const int __b)
23491 {
23492   return (int8x16_t) __builtin_aarch64_sqshl_nv16qi (__a, __b);
23493 }
23494
23495 __extension__ extern __inline int16x8_t
23496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23497 vqshlq_n_s16 (int16x8_t __a, const int __b)
23498 {
23499   return (int16x8_t) __builtin_aarch64_sqshl_nv8hi (__a, __b);
23500 }
23501
23502 __extension__ extern __inline int32x4_t
23503 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23504 vqshlq_n_s32 (int32x4_t __a, const int __b)
23505 {
23506   return (int32x4_t) __builtin_aarch64_sqshl_nv4si (__a, __b);
23507 }
23508
23509 __extension__ extern __inline int64x2_t
23510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23511 vqshlq_n_s64 (int64x2_t __a, const int __b)
23512 {
23513   return (int64x2_t) __builtin_aarch64_sqshl_nv2di (__a, __b);
23514 }
23515
23516 __extension__ extern __inline uint8x16_t
23517 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23518 vqshlq_n_u8 (uint8x16_t __a, const int __b)
23519 {
23520   return __builtin_aarch64_uqshl_nv16qi_uus (__a, __b);
23521 }
23522
23523 __extension__ extern __inline uint16x8_t
23524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23525 vqshlq_n_u16 (uint16x8_t __a, const int __b)
23526 {
23527   return __builtin_aarch64_uqshl_nv8hi_uus (__a, __b);
23528 }
23529
23530 __extension__ extern __inline uint32x4_t
23531 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23532 vqshlq_n_u32 (uint32x4_t __a, const int __b)
23533 {
23534   return __builtin_aarch64_uqshl_nv4si_uus (__a, __b);
23535 }
23536
23537 __extension__ extern __inline uint64x2_t
23538 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23539 vqshlq_n_u64 (uint64x2_t __a, const int __b)
23540 {
23541   return __builtin_aarch64_uqshl_nv2di_uus (__a, __b);
23542 }
23543
23544 __extension__ extern __inline int8_t
23545 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23546 vqshlb_n_s8 (int8_t __a, const int __b)
23547 {
23548   return (int8_t) __builtin_aarch64_sqshl_nqi (__a, __b);
23549 }
23550
23551 __extension__ extern __inline int16_t
23552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23553 vqshlh_n_s16 (int16_t __a, const int __b)
23554 {
23555   return (int16_t) __builtin_aarch64_sqshl_nhi (__a, __b);
23556 }
23557
23558 __extension__ extern __inline int32_t
23559 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23560 vqshls_n_s32 (int32_t __a, const int __b)
23561 {
23562   return (int32_t) __builtin_aarch64_sqshl_nsi (__a, __b);
23563 }
23564
23565 __extension__ extern __inline int64_t
23566 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23567 vqshld_n_s64 (int64_t __a, const int __b)
23568 {
23569   return __builtin_aarch64_sqshl_ndi (__a, __b);
23570 }
23571
23572 __extension__ extern __inline uint8_t
23573 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23574 vqshlb_n_u8 (uint8_t __a, const int __b)
23575 {
23576   return __builtin_aarch64_uqshl_nqi_uus (__a, __b);
23577 }
23578
23579 __extension__ extern __inline uint16_t
23580 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23581 vqshlh_n_u16 (uint16_t __a, const int __b)
23582 {
23583   return __builtin_aarch64_uqshl_nhi_uus (__a, __b);
23584 }
23585
23586 __extension__ extern __inline uint32_t
23587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23588 vqshls_n_u32 (uint32_t __a, const int __b)
23589 {
23590   return __builtin_aarch64_uqshl_nsi_uus (__a, __b);
23591 }
23592
23593 __extension__ extern __inline uint64_t
23594 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23595 vqshld_n_u64 (uint64_t __a, const int __b)
23596 {
23597   return __builtin_aarch64_uqshl_ndi_uus (__a, __b);
23598 }
23599
23600 /* vqshlu */
23601
23602 __extension__ extern __inline uint8x8_t
23603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23604 vqshlu_n_s8 (int8x8_t __a, const int __b)
23605 {
23606   return __builtin_aarch64_sqshlu_nv8qi_uss (__a, __b);
23607 }
23608
23609 __extension__ extern __inline uint16x4_t
23610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23611 vqshlu_n_s16 (int16x4_t __a, const int __b)
23612 {
23613   return __builtin_aarch64_sqshlu_nv4hi_uss (__a, __b);
23614 }
23615
23616 __extension__ extern __inline uint32x2_t
23617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23618 vqshlu_n_s32 (int32x2_t __a, const int __b)
23619 {
23620   return __builtin_aarch64_sqshlu_nv2si_uss (__a, __b);
23621 }
23622
23623 __extension__ extern __inline uint64x1_t
23624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23625 vqshlu_n_s64 (int64x1_t __a, const int __b)
23626 {
23627   return (uint64x1_t) {__builtin_aarch64_sqshlu_ndi_uss (__a[0], __b)};
23628 }
23629
23630 __extension__ extern __inline uint8x16_t
23631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23632 vqshluq_n_s8 (int8x16_t __a, const int __b)
23633 {
23634   return __builtin_aarch64_sqshlu_nv16qi_uss (__a, __b);
23635 }
23636
23637 __extension__ extern __inline uint16x8_t
23638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23639 vqshluq_n_s16 (int16x8_t __a, const int __b)
23640 {
23641   return __builtin_aarch64_sqshlu_nv8hi_uss (__a, __b);
23642 }
23643
23644 __extension__ extern __inline uint32x4_t
23645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23646 vqshluq_n_s32 (int32x4_t __a, const int __b)
23647 {
23648   return __builtin_aarch64_sqshlu_nv4si_uss (__a, __b);
23649 }
23650
23651 __extension__ extern __inline uint64x2_t
23652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23653 vqshluq_n_s64 (int64x2_t __a, const int __b)
23654 {
23655   return __builtin_aarch64_sqshlu_nv2di_uss (__a, __b);
23656 }
23657
23658 __extension__ extern __inline int8_t
23659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23660 vqshlub_n_s8 (int8_t __a, const int __b)
23661 {
23662   return (int8_t) __builtin_aarch64_sqshlu_nqi_uss (__a, __b);
23663 }
23664
23665 __extension__ extern __inline int16_t
23666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23667 vqshluh_n_s16 (int16_t __a, const int __b)
23668 {
23669   return (int16_t) __builtin_aarch64_sqshlu_nhi_uss (__a, __b);
23670 }
23671
23672 __extension__ extern __inline int32_t
23673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23674 vqshlus_n_s32 (int32_t __a, const int __b)
23675 {
23676   return (int32_t) __builtin_aarch64_sqshlu_nsi_uss (__a, __b);
23677 }
23678
23679 __extension__ extern __inline uint64_t
23680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23681 vqshlud_n_s64 (int64_t __a, const int __b)
23682 {
23683   return __builtin_aarch64_sqshlu_ndi_uss (__a, __b);
23684 }
23685
23686 /* vqshrn */
23687
23688 __extension__ extern __inline int8x8_t
23689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23690 vqshrn_n_s16 (int16x8_t __a, const int __b)
23691 {
23692   return (int8x8_t) __builtin_aarch64_sqshrn_nv8hi (__a, __b);
23693 }
23694
23695 __extension__ extern __inline int16x4_t
23696 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23697 vqshrn_n_s32 (int32x4_t __a, const int __b)
23698 {
23699   return (int16x4_t) __builtin_aarch64_sqshrn_nv4si (__a, __b);
23700 }
23701
23702 __extension__ extern __inline int32x2_t
23703 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23704 vqshrn_n_s64 (int64x2_t __a, const int __b)
23705 {
23706   return (int32x2_t) __builtin_aarch64_sqshrn_nv2di (__a, __b);
23707 }
23708
23709 __extension__ extern __inline uint8x8_t
23710 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23711 vqshrn_n_u16 (uint16x8_t __a, const int __b)
23712 {
23713   return __builtin_aarch64_uqshrn_nv8hi_uus ( __a, __b);
23714 }
23715
23716 __extension__ extern __inline uint16x4_t
23717 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23718 vqshrn_n_u32 (uint32x4_t __a, const int __b)
23719 {
23720   return __builtin_aarch64_uqshrn_nv4si_uus ( __a, __b);
23721 }
23722
23723 __extension__ extern __inline uint32x2_t
23724 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23725 vqshrn_n_u64 (uint64x2_t __a, const int __b)
23726 {
23727   return __builtin_aarch64_uqshrn_nv2di_uus ( __a, __b);
23728 }
23729
23730 __extension__ extern __inline int8_t
23731 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23732 vqshrnh_n_s16 (int16_t __a, const int __b)
23733 {
23734   return (int8_t) __builtin_aarch64_sqshrn_nhi (__a, __b);
23735 }
23736
23737 __extension__ extern __inline int16_t
23738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23739 vqshrns_n_s32 (int32_t __a, const int __b)
23740 {
23741   return (int16_t) __builtin_aarch64_sqshrn_nsi (__a, __b);
23742 }
23743
23744 __extension__ extern __inline int32_t
23745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23746 vqshrnd_n_s64 (int64_t __a, const int __b)
23747 {
23748   return (int32_t) __builtin_aarch64_sqshrn_ndi (__a, __b);
23749 }
23750
23751 __extension__ extern __inline uint8_t
23752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23753 vqshrnh_n_u16 (uint16_t __a, const int __b)
23754 {
23755   return __builtin_aarch64_uqshrn_nhi_uus (__a, __b);
23756 }
23757
23758 __extension__ extern __inline uint16_t
23759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23760 vqshrns_n_u32 (uint32_t __a, const int __b)
23761 {
23762   return __builtin_aarch64_uqshrn_nsi_uus (__a, __b);
23763 }
23764
23765 __extension__ extern __inline uint32_t
23766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23767 vqshrnd_n_u64 (uint64_t __a, const int __b)
23768 {
23769   return __builtin_aarch64_uqshrn_ndi_uus (__a, __b);
23770 }
23771
23772 /* vqshrun */
23773
23774 __extension__ extern __inline uint8x8_t
23775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23776 vqshrun_n_s16 (int16x8_t __a, const int __b)
23777 {
23778   return (uint8x8_t) __builtin_aarch64_sqshrun_nv8hi (__a, __b);
23779 }
23780
23781 __extension__ extern __inline uint16x4_t
23782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23783 vqshrun_n_s32 (int32x4_t __a, const int __b)
23784 {
23785   return (uint16x4_t) __builtin_aarch64_sqshrun_nv4si (__a, __b);
23786 }
23787
23788 __extension__ extern __inline uint32x2_t
23789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23790 vqshrun_n_s64 (int64x2_t __a, const int __b)
23791 {
23792   return (uint32x2_t) __builtin_aarch64_sqshrun_nv2di (__a, __b);
23793 }
23794
23795 __extension__ extern __inline int8_t
23796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23797 vqshrunh_n_s16 (int16_t __a, const int __b)
23798 {
23799   return (int8_t) __builtin_aarch64_sqshrun_nhi (__a, __b);
23800 }
23801
23802 __extension__ extern __inline int16_t
23803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23804 vqshruns_n_s32 (int32_t __a, const int __b)
23805 {
23806   return (int16_t) __builtin_aarch64_sqshrun_nsi (__a, __b);
23807 }
23808
23809 __extension__ extern __inline int32_t
23810 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23811 vqshrund_n_s64 (int64_t __a, const int __b)
23812 {
23813   return (int32_t) __builtin_aarch64_sqshrun_ndi (__a, __b);
23814 }
23815
23816 /* vqsub */
23817
23818 __extension__ extern __inline int8_t
23819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23820 vqsubb_s8 (int8_t __a, int8_t __b)
23821 {
23822   return (int8_t) __builtin_aarch64_sqsubqi (__a, __b);
23823 }
23824
23825 __extension__ extern __inline int16_t
23826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23827 vqsubh_s16 (int16_t __a, int16_t __b)
23828 {
23829   return (int16_t) __builtin_aarch64_sqsubhi (__a, __b);
23830 }
23831
23832 __extension__ extern __inline int32_t
23833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23834 vqsubs_s32 (int32_t __a, int32_t __b)
23835 {
23836   return (int32_t) __builtin_aarch64_sqsubsi (__a, __b);
23837 }
23838
23839 __extension__ extern __inline int64_t
23840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23841 vqsubd_s64 (int64_t __a, int64_t __b)
23842 {
23843   return __builtin_aarch64_sqsubdi (__a, __b);
23844 }
23845
23846 __extension__ extern __inline uint8_t
23847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23848 vqsubb_u8 (uint8_t __a, uint8_t __b)
23849 {
23850   return (uint8_t) __builtin_aarch64_uqsubqi_uuu (__a, __b);
23851 }
23852
23853 __extension__ extern __inline uint16_t
23854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23855 vqsubh_u16 (uint16_t __a, uint16_t __b)
23856 {
23857   return (uint16_t) __builtin_aarch64_uqsubhi_uuu (__a, __b);
23858 }
23859
23860 __extension__ extern __inline uint32_t
23861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23862 vqsubs_u32 (uint32_t __a, uint32_t __b)
23863 {
23864   return (uint32_t) __builtin_aarch64_uqsubsi_uuu (__a, __b);
23865 }
23866
23867 __extension__ extern __inline uint64_t
23868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23869 vqsubd_u64 (uint64_t __a, uint64_t __b)
23870 {
23871   return __builtin_aarch64_uqsubdi_uuu (__a, __b);
23872 }
23873
23874 /* vqtbl2 */
23875
23876 __extension__ extern __inline int8x8_t
23877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23878 vqtbl2_s8 (int8x16x2_t tab, uint8x8_t idx)
23879 {
23880   __builtin_aarch64_simd_oi __o;
23881   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
23882   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
23883   return __builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
23884 }
23885
23886 __extension__ extern __inline uint8x8_t
23887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23888 vqtbl2_u8 (uint8x16x2_t tab, uint8x8_t idx)
23889 {
23890   __builtin_aarch64_simd_oi __o;
23891   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23892   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23893   return (uint8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
23894 }
23895
23896 __extension__ extern __inline poly8x8_t
23897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23898 vqtbl2_p8 (poly8x16x2_t tab, uint8x8_t idx)
23899 {
23900   __builtin_aarch64_simd_oi __o;
23901   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23902   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23903   return (poly8x8_t)__builtin_aarch64_tbl3v8qi (__o, (int8x8_t)idx);
23904 }
23905
23906 __extension__ extern __inline int8x16_t
23907 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23908 vqtbl2q_s8 (int8x16x2_t tab, uint8x16_t idx)
23909 {
23910   __builtin_aarch64_simd_oi __o;
23911   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23912   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23913   return __builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
23914 }
23915
23916 __extension__ extern __inline uint8x16_t
23917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23918 vqtbl2q_u8 (uint8x16x2_t tab, uint8x16_t idx)
23919 {
23920   __builtin_aarch64_simd_oi __o;
23921   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23922   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23923   return (uint8x16_t)__builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
23924 }
23925
23926 __extension__ extern __inline poly8x16_t
23927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23928 vqtbl2q_p8 (poly8x16x2_t tab, uint8x16_t idx)
23929 {
23930   __builtin_aarch64_simd_oi __o;
23931   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
23932   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
23933   return (poly8x16_t)__builtin_aarch64_tbl3v16qi (__o, (int8x16_t)idx);
23934 }
23935
23936 /* vqtbl3 */
23937
23938 __extension__ extern __inline int8x8_t
23939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23940 vqtbl3_s8 (int8x16x3_t tab, uint8x8_t idx)
23941 {
23942   __builtin_aarch64_simd_ci __o;
23943   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
23944   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
23945   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
23946   return __builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
23947 }
23948
23949 __extension__ extern __inline uint8x8_t
23950 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23951 vqtbl3_u8 (uint8x16x3_t tab, uint8x8_t idx)
23952 {
23953   __builtin_aarch64_simd_ci __o;
23954   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
23955   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
23956   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
23957   return (uint8x8_t)__builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
23958 }
23959
23960 __extension__ extern __inline poly8x8_t
23961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23962 vqtbl3_p8 (poly8x16x3_t tab, uint8x8_t idx)
23963 {
23964   __builtin_aarch64_simd_ci __o;
23965   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
23966   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
23967   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
23968   return (poly8x8_t)__builtin_aarch64_qtbl3v8qi (__o, (int8x8_t)idx);
23969 }
23970
23971 __extension__ extern __inline int8x16_t
23972 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23973 vqtbl3q_s8 (int8x16x3_t tab, uint8x16_t idx)
23974 {
23975   __builtin_aarch64_simd_ci __o;
23976   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
23977   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
23978   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
23979   return __builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
23980 }
23981
23982 __extension__ extern __inline uint8x16_t
23983 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23984 vqtbl3q_u8 (uint8x16x3_t tab, uint8x16_t idx)
23985 {
23986   __builtin_aarch64_simd_ci __o;
23987   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
23988   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
23989   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
23990   return (uint8x16_t)__builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
23991 }
23992
23993 __extension__ extern __inline poly8x16_t
23994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
23995 vqtbl3q_p8 (poly8x16x3_t tab, uint8x16_t idx)
23996 {
23997   __builtin_aarch64_simd_ci __o;
23998   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
23999   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24000   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24001   return (poly8x16_t)__builtin_aarch64_qtbl3v16qi (__o, (int8x16_t)idx);
24002 }
24003
24004 /* vqtbl4 */
24005
24006 __extension__ extern __inline int8x8_t
24007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24008 vqtbl4_s8 (int8x16x4_t tab, uint8x8_t idx)
24009 {
24010   __builtin_aarch64_simd_xi __o;
24011   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24012   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24013   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24014   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24015   return __builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24016 }
24017
24018 __extension__ extern __inline uint8x8_t
24019 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24020 vqtbl4_u8 (uint8x16x4_t tab, uint8x8_t idx)
24021 {
24022   __builtin_aarch64_simd_xi __o;
24023   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24024   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24025   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24026   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24027   return (uint8x8_t)__builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24028 }
24029
24030 __extension__ extern __inline poly8x8_t
24031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24032 vqtbl4_p8 (poly8x16x4_t tab, uint8x8_t idx)
24033 {
24034   __builtin_aarch64_simd_xi __o;
24035   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24036   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24037   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24038   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24039   return (poly8x8_t)__builtin_aarch64_qtbl4v8qi (__o, (int8x8_t)idx);
24040 }
24041
24042 __extension__ extern __inline int8x16_t
24043 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24044 vqtbl4q_s8 (int8x16x4_t tab, uint8x16_t idx)
24045 {
24046   __builtin_aarch64_simd_xi __o;
24047   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24048   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24049   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24050   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24051   return __builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
24052 }
24053
24054 __extension__ extern __inline uint8x16_t
24055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24056 vqtbl4q_u8 (uint8x16x4_t tab, uint8x16_t idx)
24057 {
24058   __builtin_aarch64_simd_xi __o;
24059   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24060   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24061   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24062   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24063   return (uint8x16_t)__builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
24064 }
24065
24066 __extension__ extern __inline poly8x16_t
24067 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24068 vqtbl4q_p8 (poly8x16x4_t tab, uint8x16_t idx)
24069 {
24070   __builtin_aarch64_simd_xi __o;
24071   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24072   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24073   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24074   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24075   return (poly8x16_t)__builtin_aarch64_qtbl4v16qi (__o, (int8x16_t)idx);
24076 }
24077
24078
24079 /* vqtbx2 */
24080 __extension__ extern __inline int8x8_t
24081 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24082 vqtbx2_s8 (int8x8_t r, int8x16x2_t tab, uint8x8_t idx)
24083 {
24084   __builtin_aarch64_simd_oi __o;
24085   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
24086   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
24087   return __builtin_aarch64_tbx4v8qi (r, __o, (int8x8_t)idx);
24088 }
24089
24090 __extension__ extern __inline uint8x8_t
24091 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24092 vqtbx2_u8 (uint8x8_t r, uint8x16x2_t tab, uint8x8_t idx)
24093 {
24094   __builtin_aarch64_simd_oi __o;
24095   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24096   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24097   return (uint8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)r, __o,
24098                                                 (int8x8_t)idx);
24099 }
24100
24101 __extension__ extern __inline poly8x8_t
24102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24103 vqtbx2_p8 (poly8x8_t r, poly8x16x2_t tab, uint8x8_t idx)
24104 {
24105   __builtin_aarch64_simd_oi __o;
24106   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24107   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24108   return (poly8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)r, __o,
24109                                                 (int8x8_t)idx);
24110 }
24111
24112 __extension__ extern __inline int8x16_t
24113 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24114 vqtbx2q_s8 (int8x16_t r, int8x16x2_t tab, uint8x16_t idx)
24115 {
24116   __builtin_aarch64_simd_oi __o;
24117   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[0], 0);
24118   __o = __builtin_aarch64_set_qregoiv16qi (__o, tab.val[1], 1);
24119   return __builtin_aarch64_tbx4v16qi (r, __o, (int8x16_t)idx);
24120 }
24121
24122 __extension__ extern __inline uint8x16_t
24123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24124 vqtbx2q_u8 (uint8x16_t r, uint8x16x2_t tab, uint8x16_t idx)
24125 {
24126   __builtin_aarch64_simd_oi __o;
24127   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24128   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24129   return (uint8x16_t)__builtin_aarch64_tbx4v16qi ((int8x16_t)r, __o,
24130                                                   (int8x16_t)idx);
24131 }
24132
24133 __extension__ extern __inline poly8x16_t
24134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24135 vqtbx2q_p8 (poly8x16_t r, poly8x16x2_t tab, uint8x16_t idx)
24136 {
24137   __builtin_aarch64_simd_oi __o;
24138   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[0], 0);
24139   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t)tab.val[1], 1);
24140   return (poly8x16_t)__builtin_aarch64_tbx4v16qi ((int8x16_t)r, __o,
24141                                                   (int8x16_t)idx);
24142 }
24143
24144 /* vqtbx3 */
24145 __extension__ extern __inline int8x8_t
24146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24147 vqtbx3_s8 (int8x8_t r, int8x16x3_t tab, uint8x8_t idx)
24148 {
24149   __builtin_aarch64_simd_ci __o;
24150   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[0], 0);
24151   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[1], 1);
24152   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[2], 2);
24153   return __builtin_aarch64_qtbx3v8qi (r, __o, (int8x8_t)idx);
24154 }
24155
24156 __extension__ extern __inline uint8x8_t
24157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24158 vqtbx3_u8 (uint8x8_t r, uint8x16x3_t tab, uint8x8_t idx)
24159 {
24160   __builtin_aarch64_simd_ci __o;
24161   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24162   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24163   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24164   return (uint8x8_t)__builtin_aarch64_qtbx3v8qi ((int8x8_t)r, __o,
24165                                                  (int8x8_t)idx);
24166 }
24167
24168 __extension__ extern __inline poly8x8_t
24169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24170 vqtbx3_p8 (poly8x8_t r, poly8x16x3_t tab, uint8x8_t idx)
24171 {
24172   __builtin_aarch64_simd_ci __o;
24173   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24174   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24175   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24176   return (poly8x8_t)__builtin_aarch64_qtbx3v8qi ((int8x8_t)r, __o,
24177                                                  (int8x8_t)idx);
24178 }
24179
24180 __extension__ extern __inline int8x16_t
24181 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24182 vqtbx3q_s8 (int8x16_t r, int8x16x3_t tab, uint8x16_t idx)
24183 {
24184   __builtin_aarch64_simd_ci __o;
24185   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[0], 0);
24186   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[1], 1);
24187   __o = __builtin_aarch64_set_qregciv16qi (__o, tab.val[2], 2);
24188   return __builtin_aarch64_qtbx3v16qi (r, __o, (int8x16_t)idx);
24189 }
24190
24191 __extension__ extern __inline uint8x16_t
24192 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24193 vqtbx3q_u8 (uint8x16_t r, uint8x16x3_t tab, uint8x16_t idx)
24194 {
24195   __builtin_aarch64_simd_ci __o;
24196   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24197   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24198   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24199   return (uint8x16_t)__builtin_aarch64_qtbx3v16qi ((int8x16_t)r, __o,
24200                                                    (int8x16_t)idx);
24201 }
24202
24203 __extension__ extern __inline poly8x16_t
24204 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24205 vqtbx3q_p8 (poly8x16_t r, poly8x16x3_t tab, uint8x16_t idx)
24206 {
24207   __builtin_aarch64_simd_ci __o;
24208   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[0], 0);
24209   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[1], 1);
24210   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t)tab.val[2], 2);
24211   return (poly8x16_t)__builtin_aarch64_qtbx3v16qi ((int8x16_t)r, __o,
24212                                                    (int8x16_t)idx);
24213 }
24214
24215 /* vqtbx4 */
24216
24217 __extension__ extern __inline int8x8_t
24218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24219 vqtbx4_s8 (int8x8_t r, int8x16x4_t tab, uint8x8_t idx)
24220 {
24221   __builtin_aarch64_simd_xi __o;
24222   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[0], 0);
24223   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[1], 1);
24224   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[2], 2);
24225   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[3], 3);
24226   return __builtin_aarch64_qtbx4v8qi (r, __o, (int8x8_t)idx);
24227 }
24228
24229 __extension__ extern __inline uint8x8_t
24230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24231 vqtbx4_u8 (uint8x8_t r, uint8x16x4_t tab, uint8x8_t idx)
24232 {
24233   __builtin_aarch64_simd_xi __o;
24234   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24235   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24236   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24237   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24238   return (uint8x8_t)__builtin_aarch64_qtbx4v8qi ((int8x8_t)r, __o,
24239                                                  (int8x8_t)idx);
24240 }
24241
24242 __extension__ extern __inline poly8x8_t
24243 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24244 vqtbx4_p8 (poly8x8_t r, poly8x16x4_t tab, uint8x8_t idx)
24245 {
24246   __builtin_aarch64_simd_xi __o;
24247   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24248   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24249   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24250   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24251   return (poly8x8_t)__builtin_aarch64_qtbx4v8qi ((int8x8_t)r, __o,
24252                                                  (int8x8_t)idx);
24253 }
24254
24255 __extension__ extern __inline int8x16_t
24256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24257 vqtbx4q_s8 (int8x16_t r, int8x16x4_t tab, uint8x16_t idx)
24258 {
24259   __builtin_aarch64_simd_xi __o;
24260   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[0], 0);
24261   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[1], 1);
24262   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[2], 2);
24263   __o = __builtin_aarch64_set_qregxiv16qi (__o, tab.val[3], 3);
24264   return __builtin_aarch64_qtbx4v16qi (r, __o, (int8x16_t)idx);
24265 }
24266
24267 __extension__ extern __inline uint8x16_t
24268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24269 vqtbx4q_u8 (uint8x16_t r, uint8x16x4_t tab, uint8x16_t idx)
24270 {
24271   __builtin_aarch64_simd_xi __o;
24272   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24273   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24274   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24275   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24276   return (uint8x16_t)__builtin_aarch64_qtbx4v16qi ((int8x16_t)r, __o,
24277                                                    (int8x16_t)idx);
24278 }
24279
24280 __extension__ extern __inline poly8x16_t
24281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24282 vqtbx4q_p8 (poly8x16_t r, poly8x16x4_t tab, uint8x16_t idx)
24283 {
24284   __builtin_aarch64_simd_xi __o;
24285   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[0], 0);
24286   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[1], 1);
24287   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[2], 2);
24288   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t)tab.val[3], 3);
24289   return (poly8x16_t)__builtin_aarch64_qtbx4v16qi ((int8x16_t)r, __o,
24290                                                    (int8x16_t)idx);
24291 }
24292
24293 /* vrbit  */
24294
24295 __extension__ extern __inline poly8x8_t
24296 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24297 vrbit_p8 (poly8x8_t __a)
24298 {
24299   return (poly8x8_t) __builtin_aarch64_rbitv8qi ((int8x8_t) __a);
24300 }
24301
24302 __extension__ extern __inline int8x8_t
24303 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24304 vrbit_s8 (int8x8_t __a)
24305 {
24306   return __builtin_aarch64_rbitv8qi (__a);
24307 }
24308
24309 __extension__ extern __inline uint8x8_t
24310 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24311 vrbit_u8 (uint8x8_t __a)
24312 {
24313   return (uint8x8_t) __builtin_aarch64_rbitv8qi ((int8x8_t) __a);
24314 }
24315
24316 __extension__ extern __inline poly8x16_t
24317 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24318 vrbitq_p8 (poly8x16_t __a)
24319 {
24320   return (poly8x16_t) __builtin_aarch64_rbitv16qi ((int8x16_t)__a);
24321 }
24322
24323 __extension__ extern __inline int8x16_t
24324 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24325 vrbitq_s8 (int8x16_t __a)
24326 {
24327   return __builtin_aarch64_rbitv16qi (__a);
24328 }
24329
24330 __extension__ extern __inline uint8x16_t
24331 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24332 vrbitq_u8 (uint8x16_t __a)
24333 {
24334   return (uint8x16_t) __builtin_aarch64_rbitv16qi ((int8x16_t) __a);
24335 }
24336
24337 /* vrecpe  */
24338
24339 __extension__ extern __inline uint32x2_t
24340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24341 vrecpe_u32 (uint32x2_t __a)
24342 {
24343   return (uint32x2_t) __builtin_aarch64_urecpev2si ((int32x2_t) __a);
24344 }
24345
24346 __extension__ extern __inline uint32x4_t
24347 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24348 vrecpeq_u32 (uint32x4_t __a)
24349 {
24350   return (uint32x4_t) __builtin_aarch64_urecpev4si ((int32x4_t) __a);
24351 }
24352
24353 __extension__ extern __inline float32_t
24354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24355 vrecpes_f32 (float32_t __a)
24356 {
24357   return __builtin_aarch64_frecpesf (__a);
24358 }
24359
24360 __extension__ extern __inline float64_t
24361 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24362 vrecped_f64 (float64_t __a)
24363 {
24364   return __builtin_aarch64_frecpedf (__a);
24365 }
24366
24367 __extension__ extern __inline float32x2_t
24368 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24369 vrecpe_f32 (float32x2_t __a)
24370 {
24371   return __builtin_aarch64_frecpev2sf (__a);
24372 }
24373
24374 __extension__ extern __inline float64x1_t
24375 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24376 vrecpe_f64 (float64x1_t __a)
24377 {
24378   return (float64x1_t) { vrecped_f64 (vget_lane_f64 (__a, 0)) };
24379 }
24380
24381 __extension__ extern __inline float32x4_t
24382 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24383 vrecpeq_f32 (float32x4_t __a)
24384 {
24385   return __builtin_aarch64_frecpev4sf (__a);
24386 }
24387
24388 __extension__ extern __inline float64x2_t
24389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24390 vrecpeq_f64 (float64x2_t __a)
24391 {
24392   return __builtin_aarch64_frecpev2df (__a);
24393 }
24394
24395 /* vrecps  */
24396
24397 __extension__ extern __inline float32_t
24398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24399 vrecpss_f32 (float32_t __a, float32_t __b)
24400 {
24401   return __builtin_aarch64_frecpssf (__a, __b);
24402 }
24403
24404 __extension__ extern __inline float64_t
24405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24406 vrecpsd_f64 (float64_t __a, float64_t __b)
24407 {
24408   return __builtin_aarch64_frecpsdf (__a, __b);
24409 }
24410
24411 __extension__ extern __inline float32x2_t
24412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24413 vrecps_f32 (float32x2_t __a, float32x2_t __b)
24414 {
24415   return __builtin_aarch64_frecpsv2sf (__a, __b);
24416 }
24417
24418 __extension__ extern __inline float64x1_t
24419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24420 vrecps_f64 (float64x1_t __a, float64x1_t __b)
24421 {
24422   return (float64x1_t) { vrecpsd_f64 (vget_lane_f64 (__a, 0),
24423                                       vget_lane_f64 (__b, 0)) };
24424 }
24425
24426 __extension__ extern __inline float32x4_t
24427 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24428 vrecpsq_f32 (float32x4_t __a, float32x4_t __b)
24429 {
24430   return __builtin_aarch64_frecpsv4sf (__a, __b);
24431 }
24432
24433 __extension__ extern __inline float64x2_t
24434 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24435 vrecpsq_f64 (float64x2_t __a, float64x2_t __b)
24436 {
24437   return __builtin_aarch64_frecpsv2df (__a, __b);
24438 }
24439
24440 /* vrecpx  */
24441
24442 __extension__ extern __inline float32_t
24443 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24444 vrecpxs_f32 (float32_t __a)
24445 {
24446   return __builtin_aarch64_frecpxsf (__a);
24447 }
24448
24449 __extension__ extern __inline float64_t
24450 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24451 vrecpxd_f64 (float64_t __a)
24452 {
24453   return __builtin_aarch64_frecpxdf (__a);
24454 }
24455
24456
24457 /* vrev  */
24458
24459 __extension__ extern __inline poly8x8_t
24460 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24461 vrev16_p8 (poly8x8_t a)
24462 {
24463   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24464 }
24465
24466 __extension__ extern __inline int8x8_t
24467 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24468 vrev16_s8 (int8x8_t a)
24469 {
24470   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24471 }
24472
24473 __extension__ extern __inline uint8x8_t
24474 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24475 vrev16_u8 (uint8x8_t a)
24476 {
24477   return __builtin_shuffle (a, (uint8x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24478 }
24479
24480 __extension__ extern __inline poly8x16_t
24481 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24482 vrev16q_p8 (poly8x16_t a)
24483 {
24484   return __builtin_shuffle (a,
24485       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
24486 }
24487
24488 __extension__ extern __inline int8x16_t
24489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24490 vrev16q_s8 (int8x16_t a)
24491 {
24492   return __builtin_shuffle (a,
24493       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
24494 }
24495
24496 __extension__ extern __inline uint8x16_t
24497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24498 vrev16q_u8 (uint8x16_t a)
24499 {
24500   return __builtin_shuffle (a,
24501       (uint8x16_t) { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 });
24502 }
24503
24504 __extension__ extern __inline poly8x8_t
24505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24506 vrev32_p8 (poly8x8_t a)
24507 {
24508   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24509 }
24510
24511 __extension__ extern __inline poly16x4_t
24512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24513 vrev32_p16 (poly16x4_t a)
24514 {
24515   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
24516 }
24517
24518 __extension__ extern __inline int8x8_t
24519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24520 vrev32_s8 (int8x8_t a)
24521 {
24522   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24523 }
24524
24525 __extension__ extern __inline int16x4_t
24526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24527 vrev32_s16 (int16x4_t a)
24528 {
24529   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
24530 }
24531
24532 __extension__ extern __inline uint8x8_t
24533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24534 vrev32_u8 (uint8x8_t a)
24535 {
24536   return __builtin_shuffle (a, (uint8x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24537 }
24538
24539 __extension__ extern __inline uint16x4_t
24540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24541 vrev32_u16 (uint16x4_t a)
24542 {
24543   return __builtin_shuffle (a, (uint16x4_t) { 1, 0, 3, 2 });
24544 }
24545
24546 __extension__ extern __inline poly8x16_t
24547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24548 vrev32q_p8 (poly8x16_t a)
24549 {
24550   return __builtin_shuffle (a,
24551       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
24552 }
24553
24554 __extension__ extern __inline poly16x8_t
24555 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24556 vrev32q_p16 (poly16x8_t a)
24557 {
24558   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24559 }
24560
24561 __extension__ extern __inline int8x16_t
24562 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24563 vrev32q_s8 (int8x16_t a)
24564 {
24565   return __builtin_shuffle (a,
24566       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
24567 }
24568
24569 __extension__ extern __inline int16x8_t
24570 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24571 vrev32q_s16 (int16x8_t a)
24572 {
24573   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24574 }
24575
24576 __extension__ extern __inline uint8x16_t
24577 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24578 vrev32q_u8 (uint8x16_t a)
24579 {
24580   return __builtin_shuffle (a,
24581       (uint8x16_t) { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 });
24582 }
24583
24584 __extension__ extern __inline uint16x8_t
24585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24586 vrev32q_u16 (uint16x8_t a)
24587 {
24588   return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
24589 }
24590
24591 __extension__ extern __inline float16x4_t
24592 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24593 vrev64_f16 (float16x4_t __a)
24594 {
24595   return __builtin_shuffle (__a, (uint16x4_t) { 3, 2, 1, 0 });
24596 }
24597
24598 __extension__ extern __inline float32x2_t
24599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24600 vrev64_f32 (float32x2_t a)
24601 {
24602   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
24603 }
24604
24605 __extension__ extern __inline poly8x8_t
24606 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24607 vrev64_p8 (poly8x8_t a)
24608 {
24609   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
24610 }
24611
24612 __extension__ extern __inline poly16x4_t
24613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24614 vrev64_p16 (poly16x4_t a)
24615 {
24616   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
24617 }
24618
24619 __extension__ extern __inline int8x8_t
24620 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24621 vrev64_s8 (int8x8_t a)
24622 {
24623   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
24624 }
24625
24626 __extension__ extern __inline int16x4_t
24627 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24628 vrev64_s16 (int16x4_t a)
24629 {
24630   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
24631 }
24632
24633 __extension__ extern __inline int32x2_t
24634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24635 vrev64_s32 (int32x2_t a)
24636 {
24637   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
24638 }
24639
24640 __extension__ extern __inline uint8x8_t
24641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24642 vrev64_u8 (uint8x8_t a)
24643 {
24644   return __builtin_shuffle (a, (uint8x8_t) { 7, 6, 5, 4, 3, 2, 1, 0 });
24645 }
24646
24647 __extension__ extern __inline uint16x4_t
24648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24649 vrev64_u16 (uint16x4_t a)
24650 {
24651   return __builtin_shuffle (a, (uint16x4_t) { 3, 2, 1, 0 });
24652 }
24653
24654 __extension__ extern __inline uint32x2_t
24655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24656 vrev64_u32 (uint32x2_t a)
24657 {
24658   return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
24659 }
24660
24661 __extension__ extern __inline float16x8_t
24662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24663 vrev64q_f16 (float16x8_t __a)
24664 {
24665   return __builtin_shuffle (__a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24666 }
24667
24668 __extension__ extern __inline float32x4_t
24669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24670 vrev64q_f32 (float32x4_t a)
24671 {
24672   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
24673 }
24674
24675 __extension__ extern __inline poly8x16_t
24676 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24677 vrev64q_p8 (poly8x16_t a)
24678 {
24679   return __builtin_shuffle (a,
24680       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
24681 }
24682
24683 __extension__ extern __inline poly16x8_t
24684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24685 vrev64q_p16 (poly16x8_t a)
24686 {
24687   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24688 }
24689
24690 __extension__ extern __inline int8x16_t
24691 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24692 vrev64q_s8 (int8x16_t a)
24693 {
24694   return __builtin_shuffle (a,
24695       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
24696 }
24697
24698 __extension__ extern __inline int16x8_t
24699 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24700 vrev64q_s16 (int16x8_t a)
24701 {
24702   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24703 }
24704
24705 __extension__ extern __inline int32x4_t
24706 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24707 vrev64q_s32 (int32x4_t a)
24708 {
24709   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
24710 }
24711
24712 __extension__ extern __inline uint8x16_t
24713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24714 vrev64q_u8 (uint8x16_t a)
24715 {
24716   return __builtin_shuffle (a,
24717       (uint8x16_t) { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 });
24718 }
24719
24720 __extension__ extern __inline uint16x8_t
24721 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24722 vrev64q_u16 (uint16x8_t a)
24723 {
24724   return __builtin_shuffle (a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
24725 }
24726
24727 __extension__ extern __inline uint32x4_t
24728 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24729 vrev64q_u32 (uint32x4_t a)
24730 {
24731   return __builtin_shuffle (a, (uint32x4_t) { 1, 0, 3, 2 });
24732 }
24733
24734 /* vrnd  */
24735
24736 __extension__ extern __inline float32x2_t
24737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24738 vrnd_f32 (float32x2_t __a)
24739 {
24740   return __builtin_aarch64_btruncv2sf (__a);
24741 }
24742
24743 __extension__ extern __inline float64x1_t
24744 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24745 vrnd_f64 (float64x1_t __a)
24746 {
24747   return vset_lane_f64 (__builtin_trunc (vget_lane_f64 (__a, 0)), __a, 0);
24748 }
24749
24750 __extension__ extern __inline float32x4_t
24751 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24752 vrndq_f32 (float32x4_t __a)
24753 {
24754   return __builtin_aarch64_btruncv4sf (__a);
24755 }
24756
24757 __extension__ extern __inline float64x2_t
24758 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24759 vrndq_f64 (float64x2_t __a)
24760 {
24761   return __builtin_aarch64_btruncv2df (__a);
24762 }
24763
24764 /* vrnda  */
24765
24766 __extension__ extern __inline float32x2_t
24767 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24768 vrnda_f32 (float32x2_t __a)
24769 {
24770   return __builtin_aarch64_roundv2sf (__a);
24771 }
24772
24773 __extension__ extern __inline float64x1_t
24774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24775 vrnda_f64 (float64x1_t __a)
24776 {
24777   return vset_lane_f64 (__builtin_round (vget_lane_f64 (__a, 0)), __a, 0);
24778 }
24779
24780 __extension__ extern __inline float32x4_t
24781 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24782 vrndaq_f32 (float32x4_t __a)
24783 {
24784   return __builtin_aarch64_roundv4sf (__a);
24785 }
24786
24787 __extension__ extern __inline float64x2_t
24788 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24789 vrndaq_f64 (float64x2_t __a)
24790 {
24791   return __builtin_aarch64_roundv2df (__a);
24792 }
24793
24794 /* vrndi  */
24795
24796 __extension__ extern __inline float32x2_t
24797 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24798 vrndi_f32 (float32x2_t __a)
24799 {
24800   return __builtin_aarch64_nearbyintv2sf (__a);
24801 }
24802
24803 __extension__ extern __inline float64x1_t
24804 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24805 vrndi_f64 (float64x1_t __a)
24806 {
24807   return vset_lane_f64 (__builtin_nearbyint (vget_lane_f64 (__a, 0)), __a, 0);
24808 }
24809
24810 __extension__ extern __inline float32x4_t
24811 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24812 vrndiq_f32 (float32x4_t __a)
24813 {
24814   return __builtin_aarch64_nearbyintv4sf (__a);
24815 }
24816
24817 __extension__ extern __inline float64x2_t
24818 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24819 vrndiq_f64 (float64x2_t __a)
24820 {
24821   return __builtin_aarch64_nearbyintv2df (__a);
24822 }
24823
24824 /* vrndm  */
24825
24826 __extension__ extern __inline float32x2_t
24827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24828 vrndm_f32 (float32x2_t __a)
24829 {
24830   return __builtin_aarch64_floorv2sf (__a);
24831 }
24832
24833 __extension__ extern __inline float64x1_t
24834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24835 vrndm_f64 (float64x1_t __a)
24836 {
24837   return vset_lane_f64 (__builtin_floor (vget_lane_f64 (__a, 0)), __a, 0);
24838 }
24839
24840 __extension__ extern __inline float32x4_t
24841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24842 vrndmq_f32 (float32x4_t __a)
24843 {
24844   return __builtin_aarch64_floorv4sf (__a);
24845 }
24846
24847 __extension__ extern __inline float64x2_t
24848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24849 vrndmq_f64 (float64x2_t __a)
24850 {
24851   return __builtin_aarch64_floorv2df (__a);
24852 }
24853
24854 /* vrndn  */
24855
24856 __extension__ extern __inline float32x2_t
24857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24858 vrndn_f32 (float32x2_t __a)
24859 {
24860   return __builtin_aarch64_frintnv2sf (__a);
24861 }
24862
24863 __extension__ extern __inline float64x1_t
24864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24865 vrndn_f64 (float64x1_t __a)
24866 {
24867   return (float64x1_t) {__builtin_aarch64_frintndf (__a[0])};
24868 }
24869
24870 __extension__ extern __inline float32x4_t
24871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24872 vrndnq_f32 (float32x4_t __a)
24873 {
24874   return __builtin_aarch64_frintnv4sf (__a);
24875 }
24876
24877 __extension__ extern __inline float64x2_t
24878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24879 vrndnq_f64 (float64x2_t __a)
24880 {
24881   return __builtin_aarch64_frintnv2df (__a);
24882 }
24883
24884 /* vrndp  */
24885
24886 __extension__ extern __inline float32x2_t
24887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24888 vrndp_f32 (float32x2_t __a)
24889 {
24890   return __builtin_aarch64_ceilv2sf (__a);
24891 }
24892
24893 __extension__ extern __inline float64x1_t
24894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24895 vrndp_f64 (float64x1_t __a)
24896 {
24897   return vset_lane_f64 (__builtin_ceil (vget_lane_f64 (__a, 0)), __a, 0);
24898 }
24899
24900 __extension__ extern __inline float32x4_t
24901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24902 vrndpq_f32 (float32x4_t __a)
24903 {
24904   return __builtin_aarch64_ceilv4sf (__a);
24905 }
24906
24907 __extension__ extern __inline float64x2_t
24908 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24909 vrndpq_f64 (float64x2_t __a)
24910 {
24911   return __builtin_aarch64_ceilv2df (__a);
24912 }
24913
24914 /* vrndx  */
24915
24916 __extension__ extern __inline float32x2_t
24917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24918 vrndx_f32 (float32x2_t __a)
24919 {
24920   return __builtin_aarch64_rintv2sf (__a);
24921 }
24922
24923 __extension__ extern __inline float64x1_t
24924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24925 vrndx_f64 (float64x1_t __a)
24926 {
24927   return vset_lane_f64 (__builtin_rint (vget_lane_f64 (__a, 0)), __a, 0);
24928 }
24929
24930 __extension__ extern __inline float32x4_t
24931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24932 vrndxq_f32 (float32x4_t __a)
24933 {
24934   return __builtin_aarch64_rintv4sf (__a);
24935 }
24936
24937 __extension__ extern __inline float64x2_t
24938 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24939 vrndxq_f64 (float64x2_t __a)
24940 {
24941   return __builtin_aarch64_rintv2df (__a);
24942 }
24943
24944 /* vrshl */
24945
24946 __extension__ extern __inline int8x8_t
24947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24948 vrshl_s8 (int8x8_t __a, int8x8_t __b)
24949 {
24950   return (int8x8_t) __builtin_aarch64_srshlv8qi (__a, __b);
24951 }
24952
24953 __extension__ extern __inline int16x4_t
24954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24955 vrshl_s16 (int16x4_t __a, int16x4_t __b)
24956 {
24957   return (int16x4_t) __builtin_aarch64_srshlv4hi (__a, __b);
24958 }
24959
24960 __extension__ extern __inline int32x2_t
24961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24962 vrshl_s32 (int32x2_t __a, int32x2_t __b)
24963 {
24964   return (int32x2_t) __builtin_aarch64_srshlv2si (__a, __b);
24965 }
24966
24967 __extension__ extern __inline int64x1_t
24968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24969 vrshl_s64 (int64x1_t __a, int64x1_t __b)
24970 {
24971   return (int64x1_t) {__builtin_aarch64_srshldi (__a[0], __b[0])};
24972 }
24973
24974 __extension__ extern __inline uint8x8_t
24975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24976 vrshl_u8 (uint8x8_t __a, int8x8_t __b)
24977 {
24978   return __builtin_aarch64_urshlv8qi_uus (__a, __b);
24979 }
24980
24981 __extension__ extern __inline uint16x4_t
24982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24983 vrshl_u16 (uint16x4_t __a, int16x4_t __b)
24984 {
24985   return __builtin_aarch64_urshlv4hi_uus (__a, __b);
24986 }
24987
24988 __extension__ extern __inline uint32x2_t
24989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24990 vrshl_u32 (uint32x2_t __a, int32x2_t __b)
24991 {
24992   return __builtin_aarch64_urshlv2si_uus (__a, __b);
24993 }
24994
24995 __extension__ extern __inline uint64x1_t
24996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
24997 vrshl_u64 (uint64x1_t __a, int64x1_t __b)
24998 {
24999   return (uint64x1_t) {__builtin_aarch64_urshldi_uus (__a[0], __b[0])};
25000 }
25001
25002 __extension__ extern __inline int8x16_t
25003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25004 vrshlq_s8 (int8x16_t __a, int8x16_t __b)
25005 {
25006   return (int8x16_t) __builtin_aarch64_srshlv16qi (__a, __b);
25007 }
25008
25009 __extension__ extern __inline int16x8_t
25010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25011 vrshlq_s16 (int16x8_t __a, int16x8_t __b)
25012 {
25013   return (int16x8_t) __builtin_aarch64_srshlv8hi (__a, __b);
25014 }
25015
25016 __extension__ extern __inline int32x4_t
25017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25018 vrshlq_s32 (int32x4_t __a, int32x4_t __b)
25019 {
25020   return (int32x4_t) __builtin_aarch64_srshlv4si (__a, __b);
25021 }
25022
25023 __extension__ extern __inline int64x2_t
25024 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25025 vrshlq_s64 (int64x2_t __a, int64x2_t __b)
25026 {
25027   return (int64x2_t) __builtin_aarch64_srshlv2di (__a, __b);
25028 }
25029
25030 __extension__ extern __inline uint8x16_t
25031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25032 vrshlq_u8 (uint8x16_t __a, int8x16_t __b)
25033 {
25034   return __builtin_aarch64_urshlv16qi_uus (__a, __b);
25035 }
25036
25037 __extension__ extern __inline uint16x8_t
25038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25039 vrshlq_u16 (uint16x8_t __a, int16x8_t __b)
25040 {
25041   return __builtin_aarch64_urshlv8hi_uus (__a, __b);
25042 }
25043
25044 __extension__ extern __inline uint32x4_t
25045 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25046 vrshlq_u32 (uint32x4_t __a, int32x4_t __b)
25047 {
25048   return __builtin_aarch64_urshlv4si_uus (__a, __b);
25049 }
25050
25051 __extension__ extern __inline uint64x2_t
25052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25053 vrshlq_u64 (uint64x2_t __a, int64x2_t __b)
25054 {
25055   return __builtin_aarch64_urshlv2di_uus (__a, __b);
25056 }
25057
25058 __extension__ extern __inline int64_t
25059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25060 vrshld_s64 (int64_t __a, int64_t __b)
25061 {
25062   return __builtin_aarch64_srshldi (__a, __b);
25063 }
25064
25065 __extension__ extern __inline uint64_t
25066 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25067 vrshld_u64 (uint64_t __a, int64_t __b)
25068 {
25069   return __builtin_aarch64_urshldi_uus (__a, __b);
25070 }
25071
25072 /* vrshr */
25073
25074 __extension__ extern __inline int8x8_t
25075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25076 vrshr_n_s8 (int8x8_t __a, const int __b)
25077 {
25078   return (int8x8_t) __builtin_aarch64_srshr_nv8qi (__a, __b);
25079 }
25080
25081 __extension__ extern __inline int16x4_t
25082 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25083 vrshr_n_s16 (int16x4_t __a, const int __b)
25084 {
25085   return (int16x4_t) __builtin_aarch64_srshr_nv4hi (__a, __b);
25086 }
25087
25088 __extension__ extern __inline int32x2_t
25089 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25090 vrshr_n_s32 (int32x2_t __a, const int __b)
25091 {
25092   return (int32x2_t) __builtin_aarch64_srshr_nv2si (__a, __b);
25093 }
25094
25095 __extension__ extern __inline int64x1_t
25096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25097 vrshr_n_s64 (int64x1_t __a, const int __b)
25098 {
25099   return (int64x1_t) {__builtin_aarch64_srshr_ndi (__a[0], __b)};
25100 }
25101
25102 __extension__ extern __inline uint8x8_t
25103 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25104 vrshr_n_u8 (uint8x8_t __a, const int __b)
25105 {
25106   return __builtin_aarch64_urshr_nv8qi_uus (__a, __b);
25107 }
25108
25109 __extension__ extern __inline uint16x4_t
25110 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25111 vrshr_n_u16 (uint16x4_t __a, const int __b)
25112 {
25113   return __builtin_aarch64_urshr_nv4hi_uus (__a, __b);
25114 }
25115
25116 __extension__ extern __inline uint32x2_t
25117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25118 vrshr_n_u32 (uint32x2_t __a, const int __b)
25119 {
25120   return __builtin_aarch64_urshr_nv2si_uus (__a, __b);
25121 }
25122
25123 __extension__ extern __inline uint64x1_t
25124 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25125 vrshr_n_u64 (uint64x1_t __a, const int __b)
25126 {
25127   return (uint64x1_t) {__builtin_aarch64_urshr_ndi_uus (__a[0], __b)};
25128 }
25129
25130 __extension__ extern __inline int8x16_t
25131 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25132 vrshrq_n_s8 (int8x16_t __a, const int __b)
25133 {
25134   return (int8x16_t) __builtin_aarch64_srshr_nv16qi (__a, __b);
25135 }
25136
25137 __extension__ extern __inline int16x8_t
25138 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25139 vrshrq_n_s16 (int16x8_t __a, const int __b)
25140 {
25141   return (int16x8_t) __builtin_aarch64_srshr_nv8hi (__a, __b);
25142 }
25143
25144 __extension__ extern __inline int32x4_t
25145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25146 vrshrq_n_s32 (int32x4_t __a, const int __b)
25147 {
25148   return (int32x4_t) __builtin_aarch64_srshr_nv4si (__a, __b);
25149 }
25150
25151 __extension__ extern __inline int64x2_t
25152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25153 vrshrq_n_s64 (int64x2_t __a, const int __b)
25154 {
25155   return (int64x2_t) __builtin_aarch64_srshr_nv2di (__a, __b);
25156 }
25157
25158 __extension__ extern __inline uint8x16_t
25159 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25160 vrshrq_n_u8 (uint8x16_t __a, const int __b)
25161 {
25162   return __builtin_aarch64_urshr_nv16qi_uus (__a, __b);
25163 }
25164
25165 __extension__ extern __inline uint16x8_t
25166 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25167 vrshrq_n_u16 (uint16x8_t __a, const int __b)
25168 {
25169   return __builtin_aarch64_urshr_nv8hi_uus (__a, __b);
25170 }
25171
25172 __extension__ extern __inline uint32x4_t
25173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25174 vrshrq_n_u32 (uint32x4_t __a, const int __b)
25175 {
25176   return __builtin_aarch64_urshr_nv4si_uus (__a, __b);
25177 }
25178
25179 __extension__ extern __inline uint64x2_t
25180 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25181 vrshrq_n_u64 (uint64x2_t __a, const int __b)
25182 {
25183   return __builtin_aarch64_urshr_nv2di_uus (__a, __b);
25184 }
25185
25186 __extension__ extern __inline int64_t
25187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25188 vrshrd_n_s64 (int64_t __a, const int __b)
25189 {
25190   return __builtin_aarch64_srshr_ndi (__a, __b);
25191 }
25192
25193 __extension__ extern __inline uint64_t
25194 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25195 vrshrd_n_u64 (uint64_t __a, const int __b)
25196 {
25197   return __builtin_aarch64_urshr_ndi_uus (__a, __b);
25198 }
25199
25200 /* vrsqrte.  */
25201
25202 __extension__ extern __inline float32_t
25203 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25204 vrsqrtes_f32 (float32_t __a)
25205 {
25206   return __builtin_aarch64_rsqrtesf (__a);
25207 }
25208
25209 __extension__ extern __inline float64_t
25210 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25211 vrsqrted_f64 (float64_t __a)
25212 {
25213   return __builtin_aarch64_rsqrtedf (__a);
25214 }
25215
25216 __extension__ extern __inline float32x2_t
25217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25218 vrsqrte_f32 (float32x2_t __a)
25219 {
25220   return __builtin_aarch64_rsqrtev2sf (__a);
25221 }
25222
25223 __extension__ extern __inline float64x1_t
25224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25225 vrsqrte_f64 (float64x1_t __a)
25226 {
25227   return (float64x1_t) {vrsqrted_f64 (vget_lane_f64 (__a, 0))};
25228 }
25229
25230 __extension__ extern __inline float32x4_t
25231 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25232 vrsqrteq_f32 (float32x4_t __a)
25233 {
25234   return __builtin_aarch64_rsqrtev4sf (__a);
25235 }
25236
25237 __extension__ extern __inline float64x2_t
25238 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25239 vrsqrteq_f64 (float64x2_t __a)
25240 {
25241   return __builtin_aarch64_rsqrtev2df (__a);
25242 }
25243
25244 /* vrsqrts.  */
25245
25246 __extension__ extern __inline float32_t
25247 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25248 vrsqrtss_f32 (float32_t __a, float32_t __b)
25249 {
25250   return __builtin_aarch64_rsqrtssf (__a, __b);
25251 }
25252
25253 __extension__ extern __inline float64_t
25254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25255 vrsqrtsd_f64 (float64_t __a, float64_t __b)
25256 {
25257   return __builtin_aarch64_rsqrtsdf (__a, __b);
25258 }
25259
25260 __extension__ extern __inline float32x2_t
25261 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25262 vrsqrts_f32 (float32x2_t __a, float32x2_t __b)
25263 {
25264   return __builtin_aarch64_rsqrtsv2sf (__a, __b);
25265 }
25266
25267 __extension__ extern __inline float64x1_t
25268 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25269 vrsqrts_f64 (float64x1_t __a, float64x1_t __b)
25270 {
25271   return (float64x1_t) {vrsqrtsd_f64 (vget_lane_f64 (__a, 0),
25272                                       vget_lane_f64 (__b, 0))};
25273 }
25274
25275 __extension__ extern __inline float32x4_t
25276 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25277 vrsqrtsq_f32 (float32x4_t __a, float32x4_t __b)
25278 {
25279   return __builtin_aarch64_rsqrtsv4sf (__a, __b);
25280 }
25281
25282 __extension__ extern __inline float64x2_t
25283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25284 vrsqrtsq_f64 (float64x2_t __a, float64x2_t __b)
25285 {
25286   return __builtin_aarch64_rsqrtsv2df (__a, __b);
25287 }
25288
25289 /* vrsra */
25290
25291 __extension__ extern __inline int8x8_t
25292 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25293 vrsra_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
25294 {
25295   return (int8x8_t) __builtin_aarch64_srsra_nv8qi (__a, __b, __c);
25296 }
25297
25298 __extension__ extern __inline int16x4_t
25299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25300 vrsra_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
25301 {
25302   return (int16x4_t) __builtin_aarch64_srsra_nv4hi (__a, __b, __c);
25303 }
25304
25305 __extension__ extern __inline int32x2_t
25306 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25307 vrsra_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
25308 {
25309   return (int32x2_t) __builtin_aarch64_srsra_nv2si (__a, __b, __c);
25310 }
25311
25312 __extension__ extern __inline int64x1_t
25313 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25314 vrsra_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
25315 {
25316   return (int64x1_t) {__builtin_aarch64_srsra_ndi (__a[0], __b[0], __c)};
25317 }
25318
25319 __extension__ extern __inline uint8x8_t
25320 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25321 vrsra_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
25322 {
25323   return __builtin_aarch64_ursra_nv8qi_uuus (__a, __b, __c);
25324 }
25325
25326 __extension__ extern __inline uint16x4_t
25327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25328 vrsra_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
25329 {
25330   return __builtin_aarch64_ursra_nv4hi_uuus (__a, __b, __c);
25331 }
25332
25333 __extension__ extern __inline uint32x2_t
25334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25335 vrsra_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
25336 {
25337   return __builtin_aarch64_ursra_nv2si_uuus (__a, __b, __c);
25338 }
25339
25340 __extension__ extern __inline uint64x1_t
25341 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25342 vrsra_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
25343 {
25344   return (uint64x1_t) {__builtin_aarch64_ursra_ndi_uuus (__a[0], __b[0], __c)};
25345 }
25346
25347 __extension__ extern __inline int8x16_t
25348 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25349 vrsraq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
25350 {
25351   return (int8x16_t) __builtin_aarch64_srsra_nv16qi (__a, __b, __c);
25352 }
25353
25354 __extension__ extern __inline int16x8_t
25355 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25356 vrsraq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
25357 {
25358   return (int16x8_t) __builtin_aarch64_srsra_nv8hi (__a, __b, __c);
25359 }
25360
25361 __extension__ extern __inline int32x4_t
25362 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25363 vrsraq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
25364 {
25365   return (int32x4_t) __builtin_aarch64_srsra_nv4si (__a, __b, __c);
25366 }
25367
25368 __extension__ extern __inline int64x2_t
25369 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25370 vrsraq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
25371 {
25372   return (int64x2_t) __builtin_aarch64_srsra_nv2di (__a, __b, __c);
25373 }
25374
25375 __extension__ extern __inline uint8x16_t
25376 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25377 vrsraq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
25378 {
25379   return __builtin_aarch64_ursra_nv16qi_uuus (__a, __b, __c);
25380 }
25381
25382 __extension__ extern __inline uint16x8_t
25383 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25384 vrsraq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
25385 {
25386   return __builtin_aarch64_ursra_nv8hi_uuus (__a, __b, __c);
25387 }
25388
25389 __extension__ extern __inline uint32x4_t
25390 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25391 vrsraq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
25392 {
25393   return __builtin_aarch64_ursra_nv4si_uuus (__a, __b, __c);
25394 }
25395
25396 __extension__ extern __inline uint64x2_t
25397 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25398 vrsraq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
25399 {
25400   return __builtin_aarch64_ursra_nv2di_uuus (__a, __b, __c);
25401 }
25402
25403 __extension__ extern __inline int64_t
25404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25405 vrsrad_n_s64 (int64_t __a, int64_t __b, const int __c)
25406 {
25407   return __builtin_aarch64_srsra_ndi (__a, __b, __c);
25408 }
25409
25410 __extension__ extern __inline uint64_t
25411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25412 vrsrad_n_u64 (uint64_t __a, uint64_t __b, const int __c)
25413 {
25414   return __builtin_aarch64_ursra_ndi_uuus (__a, __b, __c);
25415 }
25416
25417 #pragma GCC push_options
25418 #pragma GCC target ("+nothing+crypto")
25419
25420 /* vsha1  */
25421
25422 __extension__ extern __inline uint32x4_t
25423 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25424 vsha1cq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
25425 {
25426   return __builtin_aarch64_crypto_sha1cv4si_uuuu (hash_abcd, hash_e, wk);
25427 }
25428
25429 __extension__ extern __inline uint32x4_t
25430 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25431 vsha1mq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
25432 {
25433   return __builtin_aarch64_crypto_sha1mv4si_uuuu (hash_abcd, hash_e, wk);
25434 }
25435
25436 __extension__ extern __inline uint32x4_t
25437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25438 vsha1pq_u32 (uint32x4_t hash_abcd, uint32_t hash_e, uint32x4_t wk)
25439 {
25440   return __builtin_aarch64_crypto_sha1pv4si_uuuu (hash_abcd, hash_e, wk);
25441 }
25442
25443 __extension__ extern __inline uint32_t
25444 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25445 vsha1h_u32 (uint32_t hash_e)
25446 {
25447   return __builtin_aarch64_crypto_sha1hsi_uu (hash_e);
25448 }
25449
25450 __extension__ extern __inline uint32x4_t
25451 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25452 vsha1su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7, uint32x4_t w8_11)
25453 {
25454   return __builtin_aarch64_crypto_sha1su0v4si_uuuu (w0_3, w4_7, w8_11);
25455 }
25456
25457 __extension__ extern __inline uint32x4_t
25458 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25459 vsha1su1q_u32 (uint32x4_t tw0_3, uint32x4_t w12_15)
25460 {
25461   return __builtin_aarch64_crypto_sha1su1v4si_uuu (tw0_3, w12_15);
25462 }
25463
25464 __extension__ extern __inline uint32x4_t
25465 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25466 vsha256hq_u32 (uint32x4_t hash_abcd, uint32x4_t hash_efgh, uint32x4_t wk)
25467 {
25468   return __builtin_aarch64_crypto_sha256hv4si_uuuu (hash_abcd, hash_efgh, wk);
25469 }
25470
25471 __extension__ extern __inline uint32x4_t
25472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25473 vsha256h2q_u32 (uint32x4_t hash_efgh, uint32x4_t hash_abcd, uint32x4_t wk)
25474 {
25475   return __builtin_aarch64_crypto_sha256h2v4si_uuuu (hash_efgh, hash_abcd, wk);
25476 }
25477
25478 __extension__ extern __inline uint32x4_t
25479 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25480 vsha256su0q_u32 (uint32x4_t w0_3, uint32x4_t w4_7)
25481 {
25482   return __builtin_aarch64_crypto_sha256su0v4si_uuu (w0_3, w4_7);
25483 }
25484
25485 __extension__ extern __inline uint32x4_t
25486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25487 vsha256su1q_u32 (uint32x4_t tw0_3, uint32x4_t w8_11, uint32x4_t w12_15)
25488 {
25489   return __builtin_aarch64_crypto_sha256su1v4si_uuuu (tw0_3, w8_11, w12_15);
25490 }
25491
25492 __extension__ extern __inline poly128_t
25493 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25494 vmull_p64 (poly64_t a, poly64_t b)
25495 {
25496   return
25497     __builtin_aarch64_crypto_pmulldi_ppp (a, b);
25498 }
25499
25500 __extension__ extern __inline poly128_t
25501 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25502 vmull_high_p64 (poly64x2_t a, poly64x2_t b)
25503 {
25504   return __builtin_aarch64_crypto_pmullv2di_ppp (a, b);
25505 }
25506
25507 #pragma GCC pop_options
25508
25509 /* vshl */
25510
25511 __extension__ extern __inline int8x8_t
25512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25513 vshl_n_s8 (int8x8_t __a, const int __b)
25514 {
25515   return (int8x8_t) __builtin_aarch64_ashlv8qi (__a, __b);
25516 }
25517
25518 __extension__ extern __inline int16x4_t
25519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25520 vshl_n_s16 (int16x4_t __a, const int __b)
25521 {
25522   return (int16x4_t) __builtin_aarch64_ashlv4hi (__a, __b);
25523 }
25524
25525 __extension__ extern __inline int32x2_t
25526 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25527 vshl_n_s32 (int32x2_t __a, const int __b)
25528 {
25529   return (int32x2_t) __builtin_aarch64_ashlv2si (__a, __b);
25530 }
25531
25532 __extension__ extern __inline int64x1_t
25533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25534 vshl_n_s64 (int64x1_t __a, const int __b)
25535 {
25536   return (int64x1_t) {__builtin_aarch64_ashldi (__a[0], __b)};
25537 }
25538
25539 __extension__ extern __inline uint8x8_t
25540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25541 vshl_n_u8 (uint8x8_t __a, const int __b)
25542 {
25543   return (uint8x8_t) __builtin_aarch64_ashlv8qi ((int8x8_t) __a, __b);
25544 }
25545
25546 __extension__ extern __inline uint16x4_t
25547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25548 vshl_n_u16 (uint16x4_t __a, const int __b)
25549 {
25550   return (uint16x4_t) __builtin_aarch64_ashlv4hi ((int16x4_t) __a, __b);
25551 }
25552
25553 __extension__ extern __inline uint32x2_t
25554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25555 vshl_n_u32 (uint32x2_t __a, const int __b)
25556 {
25557   return (uint32x2_t) __builtin_aarch64_ashlv2si ((int32x2_t) __a, __b);
25558 }
25559
25560 __extension__ extern __inline uint64x1_t
25561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25562 vshl_n_u64 (uint64x1_t __a, const int __b)
25563 {
25564   return (uint64x1_t) {__builtin_aarch64_ashldi ((int64_t) __a[0], __b)};
25565 }
25566
25567 __extension__ extern __inline int8x16_t
25568 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25569 vshlq_n_s8 (int8x16_t __a, const int __b)
25570 {
25571   return (int8x16_t) __builtin_aarch64_ashlv16qi (__a, __b);
25572 }
25573
25574 __extension__ extern __inline int16x8_t
25575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25576 vshlq_n_s16 (int16x8_t __a, const int __b)
25577 {
25578   return (int16x8_t) __builtin_aarch64_ashlv8hi (__a, __b);
25579 }
25580
25581 __extension__ extern __inline int32x4_t
25582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25583 vshlq_n_s32 (int32x4_t __a, const int __b)
25584 {
25585   return (int32x4_t) __builtin_aarch64_ashlv4si (__a, __b);
25586 }
25587
25588 __extension__ extern __inline int64x2_t
25589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25590 vshlq_n_s64 (int64x2_t __a, const int __b)
25591 {
25592   return (int64x2_t) __builtin_aarch64_ashlv2di (__a, __b);
25593 }
25594
25595 __extension__ extern __inline uint8x16_t
25596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25597 vshlq_n_u8 (uint8x16_t __a, const int __b)
25598 {
25599   return (uint8x16_t) __builtin_aarch64_ashlv16qi ((int8x16_t) __a, __b);
25600 }
25601
25602 __extension__ extern __inline uint16x8_t
25603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25604 vshlq_n_u16 (uint16x8_t __a, const int __b)
25605 {
25606   return (uint16x8_t) __builtin_aarch64_ashlv8hi ((int16x8_t) __a, __b);
25607 }
25608
25609 __extension__ extern __inline uint32x4_t
25610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25611 vshlq_n_u32 (uint32x4_t __a, const int __b)
25612 {
25613   return (uint32x4_t) __builtin_aarch64_ashlv4si ((int32x4_t) __a, __b);
25614 }
25615
25616 __extension__ extern __inline uint64x2_t
25617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25618 vshlq_n_u64 (uint64x2_t __a, const int __b)
25619 {
25620   return (uint64x2_t) __builtin_aarch64_ashlv2di ((int64x2_t) __a, __b);
25621 }
25622
25623 __extension__ extern __inline int64_t
25624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25625 vshld_n_s64 (int64_t __a, const int __b)
25626 {
25627   return __builtin_aarch64_ashldi (__a, __b);
25628 }
25629
25630 __extension__ extern __inline uint64_t
25631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25632 vshld_n_u64 (uint64_t __a, const int __b)
25633 {
25634   return (uint64_t) __builtin_aarch64_ashldi (__a, __b);
25635 }
25636
25637 __extension__ extern __inline int8x8_t
25638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25639 vshl_s8 (int8x8_t __a, int8x8_t __b)
25640 {
25641   return __builtin_aarch64_sshlv8qi (__a, __b);
25642 }
25643
25644 __extension__ extern __inline int16x4_t
25645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25646 vshl_s16 (int16x4_t __a, int16x4_t __b)
25647 {
25648   return __builtin_aarch64_sshlv4hi (__a, __b);
25649 }
25650
25651 __extension__ extern __inline int32x2_t
25652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25653 vshl_s32 (int32x2_t __a, int32x2_t __b)
25654 {
25655   return __builtin_aarch64_sshlv2si (__a, __b);
25656 }
25657
25658 __extension__ extern __inline int64x1_t
25659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25660 vshl_s64 (int64x1_t __a, int64x1_t __b)
25661 {
25662   return (int64x1_t) {__builtin_aarch64_sshldi (__a[0], __b[0])};
25663 }
25664
25665 __extension__ extern __inline uint8x8_t
25666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25667 vshl_u8 (uint8x8_t __a, int8x8_t __b)
25668 {
25669   return __builtin_aarch64_ushlv8qi_uus (__a, __b);
25670 }
25671
25672 __extension__ extern __inline uint16x4_t
25673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25674 vshl_u16 (uint16x4_t __a, int16x4_t __b)
25675 {
25676   return __builtin_aarch64_ushlv4hi_uus (__a, __b);
25677 }
25678
25679 __extension__ extern __inline uint32x2_t
25680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25681 vshl_u32 (uint32x2_t __a, int32x2_t __b)
25682 {
25683   return __builtin_aarch64_ushlv2si_uus (__a, __b);
25684 }
25685
25686 __extension__ extern __inline uint64x1_t
25687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25688 vshl_u64 (uint64x1_t __a, int64x1_t __b)
25689 {
25690   return (uint64x1_t) {__builtin_aarch64_ushldi_uus (__a[0], __b[0])};
25691 }
25692
25693 __extension__ extern __inline int8x16_t
25694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25695 vshlq_s8 (int8x16_t __a, int8x16_t __b)
25696 {
25697   return __builtin_aarch64_sshlv16qi (__a, __b);
25698 }
25699
25700 __extension__ extern __inline int16x8_t
25701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25702 vshlq_s16 (int16x8_t __a, int16x8_t __b)
25703 {
25704   return __builtin_aarch64_sshlv8hi (__a, __b);
25705 }
25706
25707 __extension__ extern __inline int32x4_t
25708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25709 vshlq_s32 (int32x4_t __a, int32x4_t __b)
25710 {
25711   return __builtin_aarch64_sshlv4si (__a, __b);
25712 }
25713
25714 __extension__ extern __inline int64x2_t
25715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25716 vshlq_s64 (int64x2_t __a, int64x2_t __b)
25717 {
25718   return __builtin_aarch64_sshlv2di (__a, __b);
25719 }
25720
25721 __extension__ extern __inline uint8x16_t
25722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25723 vshlq_u8 (uint8x16_t __a, int8x16_t __b)
25724 {
25725   return __builtin_aarch64_ushlv16qi_uus (__a, __b);
25726 }
25727
25728 __extension__ extern __inline uint16x8_t
25729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25730 vshlq_u16 (uint16x8_t __a, int16x8_t __b)
25731 {
25732   return __builtin_aarch64_ushlv8hi_uus (__a, __b);
25733 }
25734
25735 __extension__ extern __inline uint32x4_t
25736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25737 vshlq_u32 (uint32x4_t __a, int32x4_t __b)
25738 {
25739   return __builtin_aarch64_ushlv4si_uus (__a, __b);
25740 }
25741
25742 __extension__ extern __inline uint64x2_t
25743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25744 vshlq_u64 (uint64x2_t __a, int64x2_t __b)
25745 {
25746   return __builtin_aarch64_ushlv2di_uus (__a, __b);
25747 }
25748
25749 __extension__ extern __inline int64_t
25750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25751 vshld_s64 (int64_t __a, int64_t __b)
25752 {
25753   return __builtin_aarch64_sshldi (__a, __b);
25754 }
25755
25756 __extension__ extern __inline uint64_t
25757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25758 vshld_u64 (uint64_t __a, uint64_t __b)
25759 {
25760   return __builtin_aarch64_ushldi_uus (__a, __b);
25761 }
25762
25763 __extension__ extern __inline int16x8_t
25764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25765 vshll_high_n_s8 (int8x16_t __a, const int __b)
25766 {
25767   return __builtin_aarch64_sshll2_nv16qi (__a, __b);
25768 }
25769
25770 __extension__ extern __inline int32x4_t
25771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25772 vshll_high_n_s16 (int16x8_t __a, const int __b)
25773 {
25774   return __builtin_aarch64_sshll2_nv8hi (__a, __b);
25775 }
25776
25777 __extension__ extern __inline int64x2_t
25778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25779 vshll_high_n_s32 (int32x4_t __a, const int __b)
25780 {
25781   return __builtin_aarch64_sshll2_nv4si (__a, __b);
25782 }
25783
25784 __extension__ extern __inline uint16x8_t
25785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25786 vshll_high_n_u8 (uint8x16_t __a, const int __b)
25787 {
25788   return (uint16x8_t) __builtin_aarch64_ushll2_nv16qi ((int8x16_t) __a, __b);
25789 }
25790
25791 __extension__ extern __inline uint32x4_t
25792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25793 vshll_high_n_u16 (uint16x8_t __a, const int __b)
25794 {
25795   return (uint32x4_t) __builtin_aarch64_ushll2_nv8hi ((int16x8_t) __a, __b);
25796 }
25797
25798 __extension__ extern __inline uint64x2_t
25799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25800 vshll_high_n_u32 (uint32x4_t __a, const int __b)
25801 {
25802   return (uint64x2_t) __builtin_aarch64_ushll2_nv4si ((int32x4_t) __a, __b);
25803 }
25804
25805 __extension__ extern __inline int16x8_t
25806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25807 vshll_n_s8 (int8x8_t __a, const int __b)
25808 {
25809   return __builtin_aarch64_sshll_nv8qi (__a, __b);
25810 }
25811
25812 __extension__ extern __inline int32x4_t
25813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25814 vshll_n_s16 (int16x4_t __a, const int __b)
25815 {
25816   return __builtin_aarch64_sshll_nv4hi (__a, __b);
25817 }
25818
25819 __extension__ extern __inline int64x2_t
25820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25821 vshll_n_s32 (int32x2_t __a, const int __b)
25822 {
25823   return __builtin_aarch64_sshll_nv2si (__a, __b);
25824 }
25825
25826 __extension__ extern __inline uint16x8_t
25827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25828 vshll_n_u8 (uint8x8_t __a, const int __b)
25829 {
25830   return __builtin_aarch64_ushll_nv8qi_uus (__a, __b);
25831 }
25832
25833 __extension__ extern __inline uint32x4_t
25834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25835 vshll_n_u16 (uint16x4_t __a, const int __b)
25836 {
25837   return __builtin_aarch64_ushll_nv4hi_uus (__a, __b);
25838 }
25839
25840 __extension__ extern __inline uint64x2_t
25841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25842 vshll_n_u32 (uint32x2_t __a, const int __b)
25843 {
25844   return __builtin_aarch64_ushll_nv2si_uus (__a, __b);
25845 }
25846
25847 /* vshr */
25848
25849 __extension__ extern __inline int8x8_t
25850 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25851 vshr_n_s8 (int8x8_t __a, const int __b)
25852 {
25853   return (int8x8_t) __builtin_aarch64_ashrv8qi (__a, __b);
25854 }
25855
25856 __extension__ extern __inline int16x4_t
25857 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25858 vshr_n_s16 (int16x4_t __a, const int __b)
25859 {
25860   return (int16x4_t) __builtin_aarch64_ashrv4hi (__a, __b);
25861 }
25862
25863 __extension__ extern __inline int32x2_t
25864 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25865 vshr_n_s32 (int32x2_t __a, const int __b)
25866 {
25867   return (int32x2_t) __builtin_aarch64_ashrv2si (__a, __b);
25868 }
25869
25870 __extension__ extern __inline int64x1_t
25871 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25872 vshr_n_s64 (int64x1_t __a, const int __b)
25873 {
25874   return (int64x1_t) {__builtin_aarch64_ashr_simddi (__a[0], __b)};
25875 }
25876
25877 __extension__ extern __inline uint8x8_t
25878 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25879 vshr_n_u8 (uint8x8_t __a, const int __b)
25880 {
25881   return (uint8x8_t) __builtin_aarch64_lshrv8qi ((int8x8_t) __a, __b);
25882 }
25883
25884 __extension__ extern __inline uint16x4_t
25885 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25886 vshr_n_u16 (uint16x4_t __a, const int __b)
25887 {
25888   return (uint16x4_t) __builtin_aarch64_lshrv4hi ((int16x4_t) __a, __b);
25889 }
25890
25891 __extension__ extern __inline uint32x2_t
25892 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25893 vshr_n_u32 (uint32x2_t __a, const int __b)
25894 {
25895   return (uint32x2_t) __builtin_aarch64_lshrv2si ((int32x2_t) __a, __b);
25896 }
25897
25898 __extension__ extern __inline uint64x1_t
25899 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25900 vshr_n_u64 (uint64x1_t __a, const int __b)
25901 {
25902   return (uint64x1_t) {__builtin_aarch64_lshr_simddi_uus ( __a[0], __b)};
25903 }
25904
25905 __extension__ extern __inline int8x16_t
25906 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25907 vshrq_n_s8 (int8x16_t __a, const int __b)
25908 {
25909   return (int8x16_t) __builtin_aarch64_ashrv16qi (__a, __b);
25910 }
25911
25912 __extension__ extern __inline int16x8_t
25913 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25914 vshrq_n_s16 (int16x8_t __a, const int __b)
25915 {
25916   return (int16x8_t) __builtin_aarch64_ashrv8hi (__a, __b);
25917 }
25918
25919 __extension__ extern __inline int32x4_t
25920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25921 vshrq_n_s32 (int32x4_t __a, const int __b)
25922 {
25923   return (int32x4_t) __builtin_aarch64_ashrv4si (__a, __b);
25924 }
25925
25926 __extension__ extern __inline int64x2_t
25927 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25928 vshrq_n_s64 (int64x2_t __a, const int __b)
25929 {
25930   return (int64x2_t) __builtin_aarch64_ashrv2di (__a, __b);
25931 }
25932
25933 __extension__ extern __inline uint8x16_t
25934 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25935 vshrq_n_u8 (uint8x16_t __a, const int __b)
25936 {
25937   return (uint8x16_t) __builtin_aarch64_lshrv16qi ((int8x16_t) __a, __b);
25938 }
25939
25940 __extension__ extern __inline uint16x8_t
25941 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25942 vshrq_n_u16 (uint16x8_t __a, const int __b)
25943 {
25944   return (uint16x8_t) __builtin_aarch64_lshrv8hi ((int16x8_t) __a, __b);
25945 }
25946
25947 __extension__ extern __inline uint32x4_t
25948 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25949 vshrq_n_u32 (uint32x4_t __a, const int __b)
25950 {
25951   return (uint32x4_t) __builtin_aarch64_lshrv4si ((int32x4_t) __a, __b);
25952 }
25953
25954 __extension__ extern __inline uint64x2_t
25955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25956 vshrq_n_u64 (uint64x2_t __a, const int __b)
25957 {
25958   return (uint64x2_t) __builtin_aarch64_lshrv2di ((int64x2_t) __a, __b);
25959 }
25960
25961 __extension__ extern __inline int64_t
25962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25963 vshrd_n_s64 (int64_t __a, const int __b)
25964 {
25965   return __builtin_aarch64_ashr_simddi (__a, __b);
25966 }
25967
25968 __extension__ extern __inline uint64_t
25969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25970 vshrd_n_u64 (uint64_t __a, const int __b)
25971 {
25972   return __builtin_aarch64_lshr_simddi_uus (__a, __b);
25973 }
25974
25975 /* vsli */
25976
25977 __extension__ extern __inline int8x8_t
25978 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25979 vsli_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
25980 {
25981   return (int8x8_t) __builtin_aarch64_ssli_nv8qi (__a, __b, __c);
25982 }
25983
25984 __extension__ extern __inline int16x4_t
25985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25986 vsli_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
25987 {
25988   return (int16x4_t) __builtin_aarch64_ssli_nv4hi (__a, __b, __c);
25989 }
25990
25991 __extension__ extern __inline int32x2_t
25992 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
25993 vsli_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
25994 {
25995   return (int32x2_t) __builtin_aarch64_ssli_nv2si (__a, __b, __c);
25996 }
25997
25998 __extension__ extern __inline int64x1_t
25999 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26000 vsli_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26001 {
26002   return (int64x1_t) {__builtin_aarch64_ssli_ndi (__a[0], __b[0], __c)};
26003 }
26004
26005 __extension__ extern __inline uint8x8_t
26006 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26007 vsli_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26008 {
26009   return __builtin_aarch64_usli_nv8qi_uuus (__a, __b, __c);
26010 }
26011
26012 __extension__ extern __inline uint16x4_t
26013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26014 vsli_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26015 {
26016   return __builtin_aarch64_usli_nv4hi_uuus (__a, __b, __c);
26017 }
26018
26019 __extension__ extern __inline uint32x2_t
26020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26021 vsli_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26022 {
26023   return __builtin_aarch64_usli_nv2si_uuus (__a, __b, __c);
26024 }
26025
26026 __extension__ extern __inline uint64x1_t
26027 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26028 vsli_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26029 {
26030   return (uint64x1_t) {__builtin_aarch64_usli_ndi_uuus (__a[0], __b[0], __c)};
26031 }
26032
26033 __extension__ extern __inline poly64x1_t
26034 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26035 vsli_n_p64 (poly64x1_t __a, poly64x1_t __b, const int __c)
26036 {
26037   return (poly64x1_t) {__builtin_aarch64_ssli_ndi_ppps (__a[0], __b[0], __c)};
26038 }
26039
26040 __extension__ extern __inline int8x16_t
26041 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26042 vsliq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26043 {
26044   return (int8x16_t) __builtin_aarch64_ssli_nv16qi (__a, __b, __c);
26045 }
26046
26047 __extension__ extern __inline int16x8_t
26048 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26049 vsliq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26050 {
26051   return (int16x8_t) __builtin_aarch64_ssli_nv8hi (__a, __b, __c);
26052 }
26053
26054 __extension__ extern __inline int32x4_t
26055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26056 vsliq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26057 {
26058   return (int32x4_t) __builtin_aarch64_ssli_nv4si (__a, __b, __c);
26059 }
26060
26061 __extension__ extern __inline int64x2_t
26062 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26063 vsliq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26064 {
26065   return (int64x2_t) __builtin_aarch64_ssli_nv2di (__a, __b, __c);
26066 }
26067
26068 __extension__ extern __inline uint8x16_t
26069 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26070 vsliq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26071 {
26072   return __builtin_aarch64_usli_nv16qi_uuus (__a, __b, __c);
26073 }
26074
26075 __extension__ extern __inline uint16x8_t
26076 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26077 vsliq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26078 {
26079   return __builtin_aarch64_usli_nv8hi_uuus (__a, __b, __c);
26080 }
26081
26082 __extension__ extern __inline uint32x4_t
26083 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26084 vsliq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26085 {
26086   return __builtin_aarch64_usli_nv4si_uuus (__a, __b, __c);
26087 }
26088
26089 __extension__ extern __inline uint64x2_t
26090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26091 vsliq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26092 {
26093   return __builtin_aarch64_usli_nv2di_uuus (__a, __b, __c);
26094 }
26095
26096 __extension__ extern __inline poly64x2_t
26097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26098 vsliq_n_p64 (poly64x2_t __a, poly64x2_t __b, const int __c)
26099 {
26100   return __builtin_aarch64_ssli_nv2di_ppps (__a, __b, __c);
26101 }
26102
26103 __extension__ extern __inline int64_t
26104 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26105 vslid_n_s64 (int64_t __a, int64_t __b, const int __c)
26106 {
26107   return __builtin_aarch64_ssli_ndi (__a, __b, __c);
26108 }
26109
26110 __extension__ extern __inline uint64_t
26111 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26112 vslid_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26113 {
26114   return __builtin_aarch64_usli_ndi_uuus (__a, __b, __c);
26115 }
26116
26117 /* vsqadd */
26118
26119 __extension__ extern __inline uint8x8_t
26120 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26121 vsqadd_u8 (uint8x8_t __a, int8x8_t __b)
26122 {
26123   return __builtin_aarch64_usqaddv8qi_uus (__a, __b);
26124 }
26125
26126 __extension__ extern __inline uint16x4_t
26127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26128 vsqadd_u16 (uint16x4_t __a, int16x4_t __b)
26129 {
26130   return __builtin_aarch64_usqaddv4hi_uus (__a, __b);
26131 }
26132
26133 __extension__ extern __inline uint32x2_t
26134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26135 vsqadd_u32 (uint32x2_t __a, int32x2_t __b)
26136 {
26137   return __builtin_aarch64_usqaddv2si_uus (__a, __b);
26138 }
26139
26140 __extension__ extern __inline uint64x1_t
26141 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26142 vsqadd_u64 (uint64x1_t __a, int64x1_t __b)
26143 {
26144   return (uint64x1_t) {__builtin_aarch64_usqadddi_uus (__a[0], __b[0])};
26145 }
26146
26147 __extension__ extern __inline uint8x16_t
26148 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26149 vsqaddq_u8 (uint8x16_t __a, int8x16_t __b)
26150 {
26151   return __builtin_aarch64_usqaddv16qi_uus (__a, __b);
26152 }
26153
26154 __extension__ extern __inline uint16x8_t
26155 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26156 vsqaddq_u16 (uint16x8_t __a, int16x8_t __b)
26157 {
26158   return __builtin_aarch64_usqaddv8hi_uus (__a, __b);
26159 }
26160
26161 __extension__ extern __inline uint32x4_t
26162 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26163 vsqaddq_u32 (uint32x4_t __a, int32x4_t __b)
26164 {
26165   return __builtin_aarch64_usqaddv4si_uus (__a, __b);
26166 }
26167
26168 __extension__ extern __inline uint64x2_t
26169 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26170 vsqaddq_u64 (uint64x2_t __a, int64x2_t __b)
26171 {
26172   return __builtin_aarch64_usqaddv2di_uus (__a, __b);
26173 }
26174
26175 __extension__ extern __inline uint8_t
26176 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26177 vsqaddb_u8 (uint8_t __a, int8_t __b)
26178 {
26179   return __builtin_aarch64_usqaddqi_uus (__a, __b);
26180 }
26181
26182 __extension__ extern __inline uint16_t
26183 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26184 vsqaddh_u16 (uint16_t __a, int16_t __b)
26185 {
26186   return __builtin_aarch64_usqaddhi_uus (__a, __b);
26187 }
26188
26189 __extension__ extern __inline uint32_t
26190 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26191 vsqadds_u32 (uint32_t __a, int32_t __b)
26192 {
26193   return __builtin_aarch64_usqaddsi_uus (__a, __b);
26194 }
26195
26196 __extension__ extern __inline uint64_t
26197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26198 vsqaddd_u64 (uint64_t __a, int64_t __b)
26199 {
26200   return __builtin_aarch64_usqadddi_uus (__a, __b);
26201 }
26202
26203 /* vsqrt */
26204 __extension__ extern __inline float32x2_t
26205 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26206 vsqrt_f32 (float32x2_t a)
26207 {
26208   return __builtin_aarch64_sqrtv2sf (a);
26209 }
26210
26211 __extension__ extern __inline float32x4_t
26212 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26213 vsqrtq_f32 (float32x4_t a)
26214 {
26215   return __builtin_aarch64_sqrtv4sf (a);
26216 }
26217
26218 __extension__ extern __inline float64x1_t
26219 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26220 vsqrt_f64 (float64x1_t a)
26221 {
26222   return (float64x1_t) { __builtin_aarch64_sqrtdf (a[0]) };
26223 }
26224
26225 __extension__ extern __inline float64x2_t
26226 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26227 vsqrtq_f64 (float64x2_t a)
26228 {
26229   return __builtin_aarch64_sqrtv2df (a);
26230 }
26231
26232 /* vsra */
26233
26234 __extension__ extern __inline int8x8_t
26235 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26236 vsra_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26237 {
26238   return (int8x8_t) __builtin_aarch64_ssra_nv8qi (__a, __b, __c);
26239 }
26240
26241 __extension__ extern __inline int16x4_t
26242 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26243 vsra_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26244 {
26245   return (int16x4_t) __builtin_aarch64_ssra_nv4hi (__a, __b, __c);
26246 }
26247
26248 __extension__ extern __inline int32x2_t
26249 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26250 vsra_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26251 {
26252   return (int32x2_t) __builtin_aarch64_ssra_nv2si (__a, __b, __c);
26253 }
26254
26255 __extension__ extern __inline int64x1_t
26256 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26257 vsra_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26258 {
26259   return (int64x1_t) {__builtin_aarch64_ssra_ndi (__a[0], __b[0], __c)};
26260 }
26261
26262 __extension__ extern __inline uint8x8_t
26263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26264 vsra_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26265 {
26266   return __builtin_aarch64_usra_nv8qi_uuus (__a, __b, __c);
26267 }
26268
26269 __extension__ extern __inline uint16x4_t
26270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26271 vsra_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26272 {
26273   return __builtin_aarch64_usra_nv4hi_uuus (__a, __b, __c);
26274 }
26275
26276 __extension__ extern __inline uint32x2_t
26277 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26278 vsra_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26279 {
26280   return __builtin_aarch64_usra_nv2si_uuus (__a, __b, __c);
26281 }
26282
26283 __extension__ extern __inline uint64x1_t
26284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26285 vsra_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26286 {
26287   return (uint64x1_t) {__builtin_aarch64_usra_ndi_uuus (__a[0], __b[0], __c)};
26288 }
26289
26290 __extension__ extern __inline int8x16_t
26291 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26292 vsraq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26293 {
26294   return (int8x16_t) __builtin_aarch64_ssra_nv16qi (__a, __b, __c);
26295 }
26296
26297 __extension__ extern __inline int16x8_t
26298 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26299 vsraq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26300 {
26301   return (int16x8_t) __builtin_aarch64_ssra_nv8hi (__a, __b, __c);
26302 }
26303
26304 __extension__ extern __inline int32x4_t
26305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26306 vsraq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26307 {
26308   return (int32x4_t) __builtin_aarch64_ssra_nv4si (__a, __b, __c);
26309 }
26310
26311 __extension__ extern __inline int64x2_t
26312 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26313 vsraq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26314 {
26315   return (int64x2_t) __builtin_aarch64_ssra_nv2di (__a, __b, __c);
26316 }
26317
26318 __extension__ extern __inline uint8x16_t
26319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26320 vsraq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26321 {
26322   return __builtin_aarch64_usra_nv16qi_uuus (__a, __b, __c);
26323 }
26324
26325 __extension__ extern __inline uint16x8_t
26326 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26327 vsraq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26328 {
26329   return __builtin_aarch64_usra_nv8hi_uuus (__a, __b, __c);
26330 }
26331
26332 __extension__ extern __inline uint32x4_t
26333 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26334 vsraq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26335 {
26336   return __builtin_aarch64_usra_nv4si_uuus (__a, __b, __c);
26337 }
26338
26339 __extension__ extern __inline uint64x2_t
26340 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26341 vsraq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26342 {
26343   return __builtin_aarch64_usra_nv2di_uuus (__a, __b, __c);
26344 }
26345
26346 __extension__ extern __inline int64_t
26347 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26348 vsrad_n_s64 (int64_t __a, int64_t __b, const int __c)
26349 {
26350   return __builtin_aarch64_ssra_ndi (__a, __b, __c);
26351 }
26352
26353 __extension__ extern __inline uint64_t
26354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26355 vsrad_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26356 {
26357   return __builtin_aarch64_usra_ndi_uuus (__a, __b, __c);
26358 }
26359
26360 /* vsri */
26361
26362 __extension__ extern __inline int8x8_t
26363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26364 vsri_n_s8 (int8x8_t __a, int8x8_t __b, const int __c)
26365 {
26366   return (int8x8_t) __builtin_aarch64_ssri_nv8qi (__a, __b, __c);
26367 }
26368
26369 __extension__ extern __inline int16x4_t
26370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26371 vsri_n_s16 (int16x4_t __a, int16x4_t __b, const int __c)
26372 {
26373   return (int16x4_t) __builtin_aarch64_ssri_nv4hi (__a, __b, __c);
26374 }
26375
26376 __extension__ extern __inline int32x2_t
26377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26378 vsri_n_s32 (int32x2_t __a, int32x2_t __b, const int __c)
26379 {
26380   return (int32x2_t) __builtin_aarch64_ssri_nv2si (__a, __b, __c);
26381 }
26382
26383 __extension__ extern __inline int64x1_t
26384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26385 vsri_n_s64 (int64x1_t __a, int64x1_t __b, const int __c)
26386 {
26387   return (int64x1_t) {__builtin_aarch64_ssri_ndi (__a[0], __b[0], __c)};
26388 }
26389
26390 __extension__ extern __inline uint8x8_t
26391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26392 vsri_n_u8 (uint8x8_t __a, uint8x8_t __b, const int __c)
26393 {
26394   return __builtin_aarch64_usri_nv8qi_uuus (__a, __b, __c);
26395 }
26396
26397 __extension__ extern __inline uint16x4_t
26398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26399 vsri_n_u16 (uint16x4_t __a, uint16x4_t __b, const int __c)
26400 {
26401   return __builtin_aarch64_usri_nv4hi_uuus (__a, __b, __c);
26402 }
26403
26404 __extension__ extern __inline uint32x2_t
26405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26406 vsri_n_u32 (uint32x2_t __a, uint32x2_t __b, const int __c)
26407 {
26408   return __builtin_aarch64_usri_nv2si_uuus (__a, __b, __c);
26409 }
26410
26411 __extension__ extern __inline uint64x1_t
26412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26413 vsri_n_u64 (uint64x1_t __a, uint64x1_t __b, const int __c)
26414 {
26415   return (uint64x1_t) {__builtin_aarch64_usri_ndi_uuus (__a[0], __b[0], __c)};
26416 }
26417
26418 __extension__ extern __inline int8x16_t
26419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26420 vsriq_n_s8 (int8x16_t __a, int8x16_t __b, const int __c)
26421 {
26422   return (int8x16_t) __builtin_aarch64_ssri_nv16qi (__a, __b, __c);
26423 }
26424
26425 __extension__ extern __inline int16x8_t
26426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26427 vsriq_n_s16 (int16x8_t __a, int16x8_t __b, const int __c)
26428 {
26429   return (int16x8_t) __builtin_aarch64_ssri_nv8hi (__a, __b, __c);
26430 }
26431
26432 __extension__ extern __inline int32x4_t
26433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26434 vsriq_n_s32 (int32x4_t __a, int32x4_t __b, const int __c)
26435 {
26436   return (int32x4_t) __builtin_aarch64_ssri_nv4si (__a, __b, __c);
26437 }
26438
26439 __extension__ extern __inline int64x2_t
26440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26441 vsriq_n_s64 (int64x2_t __a, int64x2_t __b, const int __c)
26442 {
26443   return (int64x2_t) __builtin_aarch64_ssri_nv2di (__a, __b, __c);
26444 }
26445
26446 __extension__ extern __inline uint8x16_t
26447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26448 vsriq_n_u8 (uint8x16_t __a, uint8x16_t __b, const int __c)
26449 {
26450   return __builtin_aarch64_usri_nv16qi_uuus (__a, __b, __c);
26451 }
26452
26453 __extension__ extern __inline uint16x8_t
26454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26455 vsriq_n_u16 (uint16x8_t __a, uint16x8_t __b, const int __c)
26456 {
26457   return __builtin_aarch64_usri_nv8hi_uuus (__a, __b, __c);
26458 }
26459
26460 __extension__ extern __inline uint32x4_t
26461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26462 vsriq_n_u32 (uint32x4_t __a, uint32x4_t __b, const int __c)
26463 {
26464   return __builtin_aarch64_usri_nv4si_uuus (__a, __b, __c);
26465 }
26466
26467 __extension__ extern __inline uint64x2_t
26468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26469 vsriq_n_u64 (uint64x2_t __a, uint64x2_t __b, const int __c)
26470 {
26471   return __builtin_aarch64_usri_nv2di_uuus (__a, __b, __c);
26472 }
26473
26474 __extension__ extern __inline int64_t
26475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26476 vsrid_n_s64 (int64_t __a, int64_t __b, const int __c)
26477 {
26478   return __builtin_aarch64_ssri_ndi (__a, __b, __c);
26479 }
26480
26481 __extension__ extern __inline uint64_t
26482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26483 vsrid_n_u64 (uint64_t __a, uint64_t __b, const int __c)
26484 {
26485   return __builtin_aarch64_usri_ndi_uuus (__a, __b, __c);
26486 }
26487
26488 /* vst1 */
26489
26490 __extension__ extern __inline void
26491 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26492 vst1_f16 (float16_t *__a, float16x4_t __b)
26493 {
26494   __builtin_aarch64_st1v4hf (__a, __b);
26495 }
26496
26497 __extension__ extern __inline void
26498 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26499 vst1_f32 (float32_t *a, float32x2_t b)
26500 {
26501   __builtin_aarch64_st1v2sf ((__builtin_aarch64_simd_sf *) a, b);
26502 }
26503
26504 __extension__ extern __inline void
26505 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26506 vst1_f64 (float64_t *a, float64x1_t b)
26507 {
26508   *a = b[0];
26509 }
26510
26511 __extension__ extern __inline void
26512 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26513 vst1_p8 (poly8_t *a, poly8x8_t b)
26514 {
26515   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a,
26516                              (int8x8_t) b);
26517 }
26518
26519 __extension__ extern __inline void
26520 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26521 vst1_p16 (poly16_t *a, poly16x4_t b)
26522 {
26523   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a,
26524                              (int16x4_t) b);
26525 }
26526
26527 __extension__ extern __inline void
26528 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26529 vst1_p64 (poly64_t *a, poly64x1_t b)
26530 {
26531   *a = b[0];
26532 }
26533
26534 __extension__ extern __inline void
26535 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26536 vst1_s8 (int8_t *a, int8x8_t b)
26537 {
26538   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a, b);
26539 }
26540
26541 __extension__ extern __inline void
26542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26543 vst1_s16 (int16_t *a, int16x4_t b)
26544 {
26545   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a, b);
26546 }
26547
26548 __extension__ extern __inline void
26549 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26550 vst1_s32 (int32_t *a, int32x2_t b)
26551 {
26552   __builtin_aarch64_st1v2si ((__builtin_aarch64_simd_si *) a, b);
26553 }
26554
26555 __extension__ extern __inline void
26556 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26557 vst1_s64 (int64_t *a, int64x1_t b)
26558 {
26559   *a = b[0];
26560 }
26561
26562 __extension__ extern __inline void
26563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26564 vst1_u8 (uint8_t *a, uint8x8_t b)
26565 {
26566   __builtin_aarch64_st1v8qi ((__builtin_aarch64_simd_qi *) a,
26567                              (int8x8_t) b);
26568 }
26569
26570 __extension__ extern __inline void
26571 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26572 vst1_u16 (uint16_t *a, uint16x4_t b)
26573 {
26574   __builtin_aarch64_st1v4hi ((__builtin_aarch64_simd_hi *) a,
26575                              (int16x4_t) b);
26576 }
26577
26578 __extension__ extern __inline void
26579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26580 vst1_u32 (uint32_t *a, uint32x2_t b)
26581 {
26582   __builtin_aarch64_st1v2si ((__builtin_aarch64_simd_si *) a,
26583                              (int32x2_t) b);
26584 }
26585
26586 __extension__ extern __inline void
26587 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26588 vst1_u64 (uint64_t *a, uint64x1_t b)
26589 {
26590   *a = b[0];
26591 }
26592
26593 /* vst1q */
26594
26595 __extension__ extern __inline void
26596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26597 vst1q_f16 (float16_t *__a, float16x8_t __b)
26598 {
26599   __builtin_aarch64_st1v8hf (__a, __b);
26600 }
26601
26602 __extension__ extern __inline void
26603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26604 vst1q_f32 (float32_t *a, float32x4_t b)
26605 {
26606   __builtin_aarch64_st1v4sf ((__builtin_aarch64_simd_sf *) a, b);
26607 }
26608
26609 __extension__ extern __inline void
26610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26611 vst1q_f64 (float64_t *a, float64x2_t b)
26612 {
26613   __builtin_aarch64_st1v2df ((__builtin_aarch64_simd_df *) a, b);
26614 }
26615
26616 __extension__ extern __inline void
26617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26618 vst1q_p8 (poly8_t *a, poly8x16_t b)
26619 {
26620   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a,
26621                               (int8x16_t) b);
26622 }
26623
26624 __extension__ extern __inline void
26625 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26626 vst1q_p16 (poly16_t *a, poly16x8_t b)
26627 {
26628   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a,
26629                              (int16x8_t) b);
26630 }
26631
26632 __extension__ extern __inline void
26633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26634 vst1q_p64 (poly64_t *a, poly64x2_t b)
26635 {
26636   __builtin_aarch64_st1v2di_sp ((__builtin_aarch64_simd_di *) a,
26637                                 (poly64x2_t) b);
26638 }
26639
26640 __extension__ extern __inline void
26641 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26642 vst1q_s8 (int8_t *a, int8x16_t b)
26643 {
26644   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a, b);
26645 }
26646
26647 __extension__ extern __inline void
26648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26649 vst1q_s16 (int16_t *a, int16x8_t b)
26650 {
26651   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a, b);
26652 }
26653
26654 __extension__ extern __inline void
26655 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26656 vst1q_s32 (int32_t *a, int32x4_t b)
26657 {
26658   __builtin_aarch64_st1v4si ((__builtin_aarch64_simd_si *) a, b);
26659 }
26660
26661 __extension__ extern __inline void
26662 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26663 vst1q_s64 (int64_t *a, int64x2_t b)
26664 {
26665   __builtin_aarch64_st1v2di ((__builtin_aarch64_simd_di *) a, b);
26666 }
26667
26668 __extension__ extern __inline void
26669 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26670 vst1q_u8 (uint8_t *a, uint8x16_t b)
26671 {
26672   __builtin_aarch64_st1v16qi ((__builtin_aarch64_simd_qi *) a,
26673                               (int8x16_t) b);
26674 }
26675
26676 __extension__ extern __inline void
26677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26678 vst1q_u16 (uint16_t *a, uint16x8_t b)
26679 {
26680   __builtin_aarch64_st1v8hi ((__builtin_aarch64_simd_hi *) a,
26681                              (int16x8_t) b);
26682 }
26683
26684 __extension__ extern __inline void
26685 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26686 vst1q_u32 (uint32_t *a, uint32x4_t b)
26687 {
26688   __builtin_aarch64_st1v4si ((__builtin_aarch64_simd_si *) a,
26689                              (int32x4_t) b);
26690 }
26691
26692 __extension__ extern __inline void
26693 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26694 vst1q_u64 (uint64_t *a, uint64x2_t b)
26695 {
26696   __builtin_aarch64_st1v2di ((__builtin_aarch64_simd_di *) a,
26697                              (int64x2_t) b);
26698 }
26699
26700 /* vst1_lane */
26701
26702 __extension__ extern __inline void
26703 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26704 vst1_lane_f16 (float16_t *__a, float16x4_t __b, const int __lane)
26705 {
26706   *__a = __aarch64_vget_lane_any (__b, __lane);
26707 }
26708
26709 __extension__ extern __inline void
26710 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26711 vst1_lane_f32 (float32_t *__a, float32x2_t __b, const int __lane)
26712 {
26713   *__a = __aarch64_vget_lane_any (__b, __lane);
26714 }
26715
26716 __extension__ extern __inline void
26717 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26718 vst1_lane_f64 (float64_t *__a, float64x1_t __b, const int __lane)
26719 {
26720   *__a = __aarch64_vget_lane_any (__b, __lane);
26721 }
26722
26723 __extension__ extern __inline void
26724 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26725 vst1_lane_p8 (poly8_t *__a, poly8x8_t __b, const int __lane)
26726 {
26727   *__a = __aarch64_vget_lane_any (__b, __lane);
26728 }
26729
26730 __extension__ extern __inline void
26731 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26732 vst1_lane_p16 (poly16_t *__a, poly16x4_t __b, const int __lane)
26733 {
26734   *__a = __aarch64_vget_lane_any (__b, __lane);
26735 }
26736
26737 __extension__ extern __inline void
26738 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26739 vst1_lane_p64 (poly64_t *__a, poly64x1_t __b, const int __lane)
26740 {
26741   *__a = __aarch64_vget_lane_any (__b, __lane);
26742 }
26743
26744 __extension__ extern __inline void
26745 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26746 vst1_lane_s8 (int8_t *__a, int8x8_t __b, const int __lane)
26747 {
26748   *__a = __aarch64_vget_lane_any (__b, __lane);
26749 }
26750
26751 __extension__ extern __inline void
26752 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26753 vst1_lane_s16 (int16_t *__a, int16x4_t __b, const int __lane)
26754 {
26755   *__a = __aarch64_vget_lane_any (__b, __lane);
26756 }
26757
26758 __extension__ extern __inline void
26759 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26760 vst1_lane_s32 (int32_t *__a, int32x2_t __b, const int __lane)
26761 {
26762   *__a = __aarch64_vget_lane_any (__b, __lane);
26763 }
26764
26765 __extension__ extern __inline void
26766 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26767 vst1_lane_s64 (int64_t *__a, int64x1_t __b, const int __lane)
26768 {
26769   *__a = __aarch64_vget_lane_any (__b, __lane);
26770 }
26771
26772 __extension__ extern __inline void
26773 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26774 vst1_lane_u8 (uint8_t *__a, uint8x8_t __b, const int __lane)
26775 {
26776   *__a = __aarch64_vget_lane_any (__b, __lane);
26777 }
26778
26779 __extension__ extern __inline void
26780 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26781 vst1_lane_u16 (uint16_t *__a, uint16x4_t __b, const int __lane)
26782 {
26783   *__a = __aarch64_vget_lane_any (__b, __lane);
26784 }
26785
26786 __extension__ extern __inline void
26787 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26788 vst1_lane_u32 (uint32_t *__a, uint32x2_t __b, const int __lane)
26789 {
26790   *__a = __aarch64_vget_lane_any (__b, __lane);
26791 }
26792
26793 __extension__ extern __inline void
26794 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26795 vst1_lane_u64 (uint64_t *__a, uint64x1_t __b, const int __lane)
26796 {
26797   *__a = __aarch64_vget_lane_any (__b, __lane);
26798 }
26799
26800 /* vst1q_lane */
26801
26802 __extension__ extern __inline void
26803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26804 vst1q_lane_f16 (float16_t *__a, float16x8_t __b, const int __lane)
26805 {
26806   *__a = __aarch64_vget_lane_any (__b, __lane);
26807 }
26808
26809 __extension__ extern __inline void
26810 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26811 vst1q_lane_f32 (float32_t *__a, float32x4_t __b, const int __lane)
26812 {
26813   *__a = __aarch64_vget_lane_any (__b, __lane);
26814 }
26815
26816 __extension__ extern __inline void
26817 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26818 vst1q_lane_f64 (float64_t *__a, float64x2_t __b, const int __lane)
26819 {
26820   *__a = __aarch64_vget_lane_any (__b, __lane);
26821 }
26822
26823 __extension__ extern __inline void
26824 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26825 vst1q_lane_p8 (poly8_t *__a, poly8x16_t __b, const int __lane)
26826 {
26827   *__a = __aarch64_vget_lane_any (__b, __lane);
26828 }
26829
26830 __extension__ extern __inline void
26831 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26832 vst1q_lane_p16 (poly16_t *__a, poly16x8_t __b, const int __lane)
26833 {
26834   *__a = __aarch64_vget_lane_any (__b, __lane);
26835 }
26836
26837 __extension__ extern __inline void
26838 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26839 vst1q_lane_p64 (poly64_t *__a, poly64x2_t __b, const int __lane)
26840 {
26841   *__a = __aarch64_vget_lane_any (__b, __lane);
26842 }
26843
26844 __extension__ extern __inline void
26845 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26846 vst1q_lane_s8 (int8_t *__a, int8x16_t __b, const int __lane)
26847 {
26848   *__a = __aarch64_vget_lane_any (__b, __lane);
26849 }
26850
26851 __extension__ extern __inline void
26852 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26853 vst1q_lane_s16 (int16_t *__a, int16x8_t __b, const int __lane)
26854 {
26855   *__a = __aarch64_vget_lane_any (__b, __lane);
26856 }
26857
26858 __extension__ extern __inline void
26859 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26860 vst1q_lane_s32 (int32_t *__a, int32x4_t __b, const int __lane)
26861 {
26862   *__a = __aarch64_vget_lane_any (__b, __lane);
26863 }
26864
26865 __extension__ extern __inline void
26866 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26867 vst1q_lane_s64 (int64_t *__a, int64x2_t __b, const int __lane)
26868 {
26869   *__a = __aarch64_vget_lane_any (__b, __lane);
26870 }
26871
26872 __extension__ extern __inline void
26873 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26874 vst1q_lane_u8 (uint8_t *__a, uint8x16_t __b, const int __lane)
26875 {
26876   *__a = __aarch64_vget_lane_any (__b, __lane);
26877 }
26878
26879 __extension__ extern __inline void
26880 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26881 vst1q_lane_u16 (uint16_t *__a, uint16x8_t __b, const int __lane)
26882 {
26883   *__a = __aarch64_vget_lane_any (__b, __lane);
26884 }
26885
26886 __extension__ extern __inline void
26887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26888 vst1q_lane_u32 (uint32_t *__a, uint32x4_t __b, const int __lane)
26889 {
26890   *__a = __aarch64_vget_lane_any (__b, __lane);
26891 }
26892
26893 __extension__ extern __inline void
26894 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26895 vst1q_lane_u64 (uint64_t *__a, uint64x2_t __b, const int __lane)
26896 {
26897   *__a = __aarch64_vget_lane_any (__b, __lane);
26898 }
26899
26900 /* vstn */
26901
26902 __extension__ extern __inline void
26903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26904 vst2_s64 (int64_t * __a, int64x1x2_t val)
26905 {
26906   __builtin_aarch64_simd_oi __o;
26907   int64x2x2_t temp;
26908   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
26909   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
26910   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
26911   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
26912   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
26913 }
26914
26915 __extension__ extern __inline void
26916 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26917 vst2_u64 (uint64_t * __a, uint64x1x2_t val)
26918 {
26919   __builtin_aarch64_simd_oi __o;
26920   uint64x2x2_t temp;
26921   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
26922   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
26923   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[0], 0);
26924   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) temp.val[1], 1);
26925   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
26926 }
26927
26928 __extension__ extern __inline void
26929 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26930 vst2_f64 (float64_t * __a, float64x1x2_t val)
26931 {
26932   __builtin_aarch64_simd_oi __o;
26933   float64x2x2_t temp;
26934   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
26935   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
26936   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[0], 0);
26937   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) temp.val[1], 1);
26938   __builtin_aarch64_st2df ((__builtin_aarch64_simd_df *) __a, __o);
26939 }
26940
26941 __extension__ extern __inline void
26942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26943 vst2_s8 (int8_t * __a, int8x8x2_t val)
26944 {
26945   __builtin_aarch64_simd_oi __o;
26946   int8x16x2_t temp;
26947   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
26948   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
26949   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
26950   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
26951   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
26952 }
26953
26954 __extension__ extern __inline void
26955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26956 vst2_p8 (poly8_t * __a, poly8x8x2_t val)
26957 {
26958   __builtin_aarch64_simd_oi __o;
26959   poly8x16x2_t temp;
26960   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
26961   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
26962   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
26963   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
26964   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
26965 }
26966
26967 __extension__ extern __inline void
26968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26969 vst2_s16 (int16_t * __a, int16x4x2_t val)
26970 {
26971   __builtin_aarch64_simd_oi __o;
26972   int16x8x2_t temp;
26973   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
26974   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
26975   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
26976   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
26977   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
26978 }
26979
26980 __extension__ extern __inline void
26981 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26982 vst2_p16 (poly16_t * __a, poly16x4x2_t val)
26983 {
26984   __builtin_aarch64_simd_oi __o;
26985   poly16x8x2_t temp;
26986   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
26987   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
26988   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
26989   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
26990   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
26991 }
26992
26993 __extension__ extern __inline void
26994 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
26995 vst2_s32 (int32_t * __a, int32x2x2_t val)
26996 {
26997   __builtin_aarch64_simd_oi __o;
26998   int32x4x2_t temp;
26999   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27000   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27001   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
27002   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
27003   __builtin_aarch64_st2v2si ((__builtin_aarch64_simd_si *) __a, __o);
27004 }
27005
27006 __extension__ extern __inline void
27007 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27008 vst2_u8 (uint8_t * __a, uint8x8x2_t val)
27009 {
27010   __builtin_aarch64_simd_oi __o;
27011   uint8x16x2_t temp;
27012   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27013   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27014   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[0], 0);
27015   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) temp.val[1], 1);
27016   __builtin_aarch64_st2v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27017 }
27018
27019 __extension__ extern __inline void
27020 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27021 vst2_u16 (uint16_t * __a, uint16x4x2_t val)
27022 {
27023   __builtin_aarch64_simd_oi __o;
27024   uint16x8x2_t temp;
27025   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27026   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27027   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[0], 0);
27028   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) temp.val[1], 1);
27029   __builtin_aarch64_st2v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27030 }
27031
27032 __extension__ extern __inline void
27033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27034 vst2_u32 (uint32_t * __a, uint32x2x2_t val)
27035 {
27036   __builtin_aarch64_simd_oi __o;
27037   uint32x4x2_t temp;
27038   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
27039   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
27040   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[0], 0);
27041   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) temp.val[1], 1);
27042   __builtin_aarch64_st2v2si ((__builtin_aarch64_simd_si *) __a, __o);
27043 }
27044
27045 __extension__ extern __inline void
27046 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27047 vst2_f16 (float16_t * __a, float16x4x2_t val)
27048 {
27049   __builtin_aarch64_simd_oi __o;
27050   float16x8x2_t temp;
27051   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
27052   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
27053   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[0], 0);
27054   __o = __builtin_aarch64_set_qregoiv8hf (__o, temp.val[1], 1);
27055   __builtin_aarch64_st2v4hf (__a, __o);
27056 }
27057
27058 __extension__ extern __inline void
27059 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27060 vst2_f32 (float32_t * __a, float32x2x2_t val)
27061 {
27062   __builtin_aarch64_simd_oi __o;
27063   float32x4x2_t temp;
27064   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
27065   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
27066   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[0], 0);
27067   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) temp.val[1], 1);
27068   __builtin_aarch64_st2v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
27069 }
27070
27071 __extension__ extern __inline void
27072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27073 vst2_p64 (poly64_t * __a, poly64x1x2_t val)
27074 {
27075   __builtin_aarch64_simd_oi __o;
27076   poly64x2x2_t temp;
27077   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
27078   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
27079   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27080                                                (poly64x2_t) temp.val[0], 0);
27081   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27082                                                (poly64x2_t) temp.val[1], 1);
27083   __builtin_aarch64_st2di ((__builtin_aarch64_simd_di *) __a, __o);
27084 }
27085
27086 __extension__ extern __inline void
27087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27088 vst2q_s8 (int8_t * __a, int8x16x2_t val)
27089 {
27090   __builtin_aarch64_simd_oi __o;
27091   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
27092   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
27093   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27094 }
27095
27096 __extension__ extern __inline void
27097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27098 vst2q_p8 (poly8_t * __a, poly8x16x2_t val)
27099 {
27100   __builtin_aarch64_simd_oi __o;
27101   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
27102   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
27103   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27104 }
27105
27106 __extension__ extern __inline void
27107 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27108 vst2q_s16 (int16_t * __a, int16x8x2_t val)
27109 {
27110   __builtin_aarch64_simd_oi __o;
27111   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
27112   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
27113   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27114 }
27115
27116 __extension__ extern __inline void
27117 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27118 vst2q_p16 (poly16_t * __a, poly16x8x2_t val)
27119 {
27120   __builtin_aarch64_simd_oi __o;
27121   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
27122   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
27123   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27124 }
27125
27126 __extension__ extern __inline void
27127 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27128 vst2q_s32 (int32_t * __a, int32x4x2_t val)
27129 {
27130   __builtin_aarch64_simd_oi __o;
27131   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
27132   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
27133   __builtin_aarch64_st2v4si ((__builtin_aarch64_simd_si *) __a, __o);
27134 }
27135
27136 __extension__ extern __inline void
27137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27138 vst2q_s64 (int64_t * __a, int64x2x2_t val)
27139 {
27140   __builtin_aarch64_simd_oi __o;
27141   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
27142   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
27143   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
27144 }
27145
27146 __extension__ extern __inline void
27147 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27148 vst2q_u8 (uint8_t * __a, uint8x16x2_t val)
27149 {
27150   __builtin_aarch64_simd_oi __o;
27151   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[0], 0);
27152   __o = __builtin_aarch64_set_qregoiv16qi (__o, (int8x16_t) val.val[1], 1);
27153   __builtin_aarch64_st2v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27154 }
27155
27156 __extension__ extern __inline void
27157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27158 vst2q_u16 (uint16_t * __a, uint16x8x2_t val)
27159 {
27160   __builtin_aarch64_simd_oi __o;
27161   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[0], 0);
27162   __o = __builtin_aarch64_set_qregoiv8hi (__o, (int16x8_t) val.val[1], 1);
27163   __builtin_aarch64_st2v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27164 }
27165
27166 __extension__ extern __inline void
27167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27168 vst2q_u32 (uint32_t * __a, uint32x4x2_t val)
27169 {
27170   __builtin_aarch64_simd_oi __o;
27171   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[0], 0);
27172   __o = __builtin_aarch64_set_qregoiv4si (__o, (int32x4_t) val.val[1], 1);
27173   __builtin_aarch64_st2v4si ((__builtin_aarch64_simd_si *) __a, __o);
27174 }
27175
27176 __extension__ extern __inline void
27177 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27178 vst2q_u64 (uint64_t * __a, uint64x2x2_t val)
27179 {
27180   __builtin_aarch64_simd_oi __o;
27181   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[0], 0);
27182   __o = __builtin_aarch64_set_qregoiv2di (__o, (int64x2_t) val.val[1], 1);
27183   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
27184 }
27185
27186 __extension__ extern __inline void
27187 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27188 vst2q_f16 (float16_t * __a, float16x8x2_t val)
27189 {
27190   __builtin_aarch64_simd_oi __o;
27191   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[0], 0);
27192   __o = __builtin_aarch64_set_qregoiv8hf (__o, val.val[1], 1);
27193   __builtin_aarch64_st2v8hf (__a, __o);
27194 }
27195
27196 __extension__ extern __inline void
27197 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27198 vst2q_f32 (float32_t * __a, float32x4x2_t val)
27199 {
27200   __builtin_aarch64_simd_oi __o;
27201   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[0], 0);
27202   __o = __builtin_aarch64_set_qregoiv4sf (__o, (float32x4_t) val.val[1], 1);
27203   __builtin_aarch64_st2v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
27204 }
27205
27206 __extension__ extern __inline void
27207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27208 vst2q_f64 (float64_t * __a, float64x2x2_t val)
27209 {
27210   __builtin_aarch64_simd_oi __o;
27211   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[0], 0);
27212   __o = __builtin_aarch64_set_qregoiv2df (__o, (float64x2_t) val.val[1], 1);
27213   __builtin_aarch64_st2v2df ((__builtin_aarch64_simd_df *) __a, __o);
27214 }
27215
27216 __extension__ extern __inline void
27217 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27218 vst2q_p64 (poly64_t * __a, poly64x2x2_t val)
27219 {
27220   __builtin_aarch64_simd_oi __o;
27221   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27222                                                (poly64x2_t) val.val[0], 0);
27223   __o = __builtin_aarch64_set_qregoiv2di_ssps (__o,
27224                                                (poly64x2_t) val.val[1], 1);
27225   __builtin_aarch64_st2v2di ((__builtin_aarch64_simd_di *) __a, __o);
27226 }
27227
27228 __extension__ extern __inline void
27229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27230 vst3_s64 (int64_t * __a, int64x1x3_t val)
27231 {
27232   __builtin_aarch64_simd_ci __o;
27233   int64x2x3_t temp;
27234   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
27235   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
27236   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
27237   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
27238   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
27239   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
27240   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
27241 }
27242
27243 __extension__ extern __inline void
27244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27245 vst3_u64 (uint64_t * __a, uint64x1x3_t val)
27246 {
27247   __builtin_aarch64_simd_ci __o;
27248   uint64x2x3_t temp;
27249   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
27250   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
27251   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
27252   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[0], 0);
27253   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[1], 1);
27254   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) temp.val[2], 2);
27255   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
27256 }
27257
27258 __extension__ extern __inline void
27259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27260 vst3_f64 (float64_t * __a, float64x1x3_t val)
27261 {
27262   __builtin_aarch64_simd_ci __o;
27263   float64x2x3_t temp;
27264   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
27265   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
27266   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
27267   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[0], 0);
27268   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[1], 1);
27269   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) temp.val[2], 2);
27270   __builtin_aarch64_st3df ((__builtin_aarch64_simd_df *) __a, __o);
27271 }
27272
27273 __extension__ extern __inline void
27274 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27275 vst3_s8 (int8_t * __a, int8x8x3_t val)
27276 {
27277   __builtin_aarch64_simd_ci __o;
27278   int8x16x3_t temp;
27279   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
27280   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
27281   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
27282   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
27283   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
27284   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
27285   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27286 }
27287
27288 __extension__ extern __inline void
27289 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27290 vst3_p8 (poly8_t * __a, poly8x8x3_t val)
27291 {
27292   __builtin_aarch64_simd_ci __o;
27293   poly8x16x3_t temp;
27294   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
27295   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
27296   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
27297   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
27298   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
27299   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
27300   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27301 }
27302
27303 __extension__ extern __inline void
27304 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27305 vst3_s16 (int16_t * __a, int16x4x3_t val)
27306 {
27307   __builtin_aarch64_simd_ci __o;
27308   int16x8x3_t temp;
27309   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
27310   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
27311   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
27312   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
27313   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
27314   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
27315   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27316 }
27317
27318 __extension__ extern __inline void
27319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27320 vst3_p16 (poly16_t * __a, poly16x4x3_t val)
27321 {
27322   __builtin_aarch64_simd_ci __o;
27323   poly16x8x3_t temp;
27324   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
27325   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
27326   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
27327   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
27328   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
27329   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
27330   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27331 }
27332
27333 __extension__ extern __inline void
27334 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27335 vst3_s32 (int32_t * __a, int32x2x3_t val)
27336 {
27337   __builtin_aarch64_simd_ci __o;
27338   int32x4x3_t temp;
27339   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27340   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27341   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
27342   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
27343   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
27344   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
27345   __builtin_aarch64_st3v2si ((__builtin_aarch64_simd_si *) __a, __o);
27346 }
27347
27348 __extension__ extern __inline void
27349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27350 vst3_u8 (uint8_t * __a, uint8x8x3_t val)
27351 {
27352   __builtin_aarch64_simd_ci __o;
27353   uint8x16x3_t temp;
27354   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27355   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27356   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
27357   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[0], 0);
27358   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[1], 1);
27359   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) temp.val[2], 2);
27360   __builtin_aarch64_st3v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27361 }
27362
27363 __extension__ extern __inline void
27364 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27365 vst3_u16 (uint16_t * __a, uint16x4x3_t val)
27366 {
27367   __builtin_aarch64_simd_ci __o;
27368   uint16x8x3_t temp;
27369   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27370   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27371   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
27372   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[0], 0);
27373   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[1], 1);
27374   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) temp.val[2], 2);
27375   __builtin_aarch64_st3v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27376 }
27377
27378 __extension__ extern __inline void
27379 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27380 vst3_u32 (uint32_t * __a, uint32x2x3_t val)
27381 {
27382   __builtin_aarch64_simd_ci __o;
27383   uint32x4x3_t temp;
27384   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
27385   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
27386   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
27387   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[0], 0);
27388   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[1], 1);
27389   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) temp.val[2], 2);
27390   __builtin_aarch64_st3v2si ((__builtin_aarch64_simd_si *) __a, __o);
27391 }
27392
27393 __extension__ extern __inline void
27394 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27395 vst3_f16 (float16_t * __a, float16x4x3_t val)
27396 {
27397   __builtin_aarch64_simd_ci __o;
27398   float16x8x3_t temp;
27399   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
27400   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
27401   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
27402   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[0], 0);
27403   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[1], 1);
27404   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) temp.val[2], 2);
27405   __builtin_aarch64_st3v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
27406 }
27407
27408 __extension__ extern __inline void
27409 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27410 vst3_f32 (float32_t * __a, float32x2x3_t val)
27411 {
27412   __builtin_aarch64_simd_ci __o;
27413   float32x4x3_t temp;
27414   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
27415   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
27416   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
27417   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[0], 0);
27418   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[1], 1);
27419   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) temp.val[2], 2);
27420   __builtin_aarch64_st3v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
27421 }
27422
27423 __extension__ extern __inline void
27424 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27425 vst3_p64 (poly64_t * __a, poly64x1x3_t val)
27426 {
27427   __builtin_aarch64_simd_ci __o;
27428   poly64x2x3_t temp;
27429   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
27430   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
27431   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
27432   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27433                                                (poly64x2_t) temp.val[0], 0);
27434   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27435                                                (poly64x2_t) temp.val[1], 1);
27436   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27437                                                (poly64x2_t) temp.val[2], 2);
27438   __builtin_aarch64_st3di ((__builtin_aarch64_simd_di *) __a, __o);
27439 }
27440
27441 __extension__ extern __inline void
27442 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27443 vst3q_s8 (int8_t * __a, int8x16x3_t val)
27444 {
27445   __builtin_aarch64_simd_ci __o;
27446   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
27447   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
27448   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
27449   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27450 }
27451
27452 __extension__ extern __inline void
27453 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27454 vst3q_p8 (poly8_t * __a, poly8x16x3_t val)
27455 {
27456   __builtin_aarch64_simd_ci __o;
27457   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
27458   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
27459   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
27460   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27461 }
27462
27463 __extension__ extern __inline void
27464 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27465 vst3q_s16 (int16_t * __a, int16x8x3_t val)
27466 {
27467   __builtin_aarch64_simd_ci __o;
27468   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
27469   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
27470   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
27471   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27472 }
27473
27474 __extension__ extern __inline void
27475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27476 vst3q_p16 (poly16_t * __a, poly16x8x3_t val)
27477 {
27478   __builtin_aarch64_simd_ci __o;
27479   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
27480   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
27481   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
27482   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27483 }
27484
27485 __extension__ extern __inline void
27486 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27487 vst3q_s32 (int32_t * __a, int32x4x3_t val)
27488 {
27489   __builtin_aarch64_simd_ci __o;
27490   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
27491   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
27492   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
27493   __builtin_aarch64_st3v4si ((__builtin_aarch64_simd_si *) __a, __o);
27494 }
27495
27496 __extension__ extern __inline void
27497 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27498 vst3q_s64 (int64_t * __a, int64x2x3_t val)
27499 {
27500   __builtin_aarch64_simd_ci __o;
27501   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
27502   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
27503   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
27504   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
27505 }
27506
27507 __extension__ extern __inline void
27508 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27509 vst3q_u8 (uint8_t * __a, uint8x16x3_t val)
27510 {
27511   __builtin_aarch64_simd_ci __o;
27512   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
27513   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
27514   __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
27515   __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27516 }
27517
27518 __extension__ extern __inline void
27519 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27520 vst3q_u16 (uint16_t * __a, uint16x8x3_t val)
27521 {
27522   __builtin_aarch64_simd_ci __o;
27523   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
27524   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
27525   __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[2], 2);
27526   __builtin_aarch64_st3v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27527 }
27528
27529 __extension__ extern __inline void
27530 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27531 vst3q_u32 (uint32_t * __a, uint32x4x3_t val)
27532 {
27533   __builtin_aarch64_simd_ci __o;
27534   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[0], 0);
27535   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[1], 1);
27536   __o = __builtin_aarch64_set_qregciv4si (__o, (int32x4_t) val.val[2], 2);
27537   __builtin_aarch64_st3v4si ((__builtin_aarch64_simd_si *) __a, __o);
27538 }
27539
27540 __extension__ extern __inline void
27541 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27542 vst3q_u64 (uint64_t * __a, uint64x2x3_t val)
27543 {
27544   __builtin_aarch64_simd_ci __o;
27545   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[0], 0);
27546   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[1], 1);
27547   __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
27548   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
27549 }
27550
27551 __extension__ extern __inline void
27552 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27553 vst3q_f16 (float16_t * __a, float16x8x3_t val)
27554 {
27555   __builtin_aarch64_simd_ci __o;
27556   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[0], 0);
27557   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[1], 1);
27558   __o = __builtin_aarch64_set_qregciv8hf (__o, (float16x8_t) val.val[2], 2);
27559   __builtin_aarch64_st3v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
27560 }
27561
27562 __extension__ extern __inline void
27563 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27564 vst3q_f32 (float32_t * __a, float32x4x3_t val)
27565 {
27566   __builtin_aarch64_simd_ci __o;
27567   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[0], 0);
27568   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[1], 1);
27569   __o = __builtin_aarch64_set_qregciv4sf (__o, (float32x4_t) val.val[2], 2);
27570   __builtin_aarch64_st3v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
27571 }
27572
27573 __extension__ extern __inline void
27574 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27575 vst3q_f64 (float64_t * __a, float64x2x3_t val)
27576 {
27577   __builtin_aarch64_simd_ci __o;
27578   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[0], 0);
27579   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[1], 1);
27580   __o = __builtin_aarch64_set_qregciv2df (__o, (float64x2_t) val.val[2], 2);
27581   __builtin_aarch64_st3v2df ((__builtin_aarch64_simd_df *) __a, __o);
27582 }
27583
27584 __extension__ extern __inline void
27585 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27586 vst3q_p64 (poly64_t * __a, poly64x2x3_t val)
27587 {
27588   __builtin_aarch64_simd_ci __o;
27589   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27590                                                (poly64x2_t) val.val[0], 0);
27591   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27592                                                (poly64x2_t) val.val[1], 1);
27593   __o = __builtin_aarch64_set_qregciv2di_ssps (__o,
27594                                                (poly64x2_t) val.val[2], 2);
27595   __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
27596 }
27597
27598 __extension__ extern __inline void
27599 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27600 vst4_s64 (int64_t * __a, int64x1x4_t val)
27601 {
27602   __builtin_aarch64_simd_xi __o;
27603   int64x2x4_t temp;
27604   temp.val[0] = vcombine_s64 (val.val[0], vcreate_s64 (__AARCH64_INT64_C (0)));
27605   temp.val[1] = vcombine_s64 (val.val[1], vcreate_s64 (__AARCH64_INT64_C (0)));
27606   temp.val[2] = vcombine_s64 (val.val[2], vcreate_s64 (__AARCH64_INT64_C (0)));
27607   temp.val[3] = vcombine_s64 (val.val[3], vcreate_s64 (__AARCH64_INT64_C (0)));
27608   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[0], 0);
27609   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[1], 1);
27610   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[2], 2);
27611   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[3], 3);
27612   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
27613 }
27614
27615 __extension__ extern __inline void
27616 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27617 vst4_u64 (uint64_t * __a, uint64x1x4_t val)
27618 {
27619   __builtin_aarch64_simd_xi __o;
27620   uint64x2x4_t temp;
27621   temp.val[0] = vcombine_u64 (val.val[0], vcreate_u64 (__AARCH64_UINT64_C (0)));
27622   temp.val[1] = vcombine_u64 (val.val[1], vcreate_u64 (__AARCH64_UINT64_C (0)));
27623   temp.val[2] = vcombine_u64 (val.val[2], vcreate_u64 (__AARCH64_UINT64_C (0)));
27624   temp.val[3] = vcombine_u64 (val.val[3], vcreate_u64 (__AARCH64_UINT64_C (0)));
27625   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[0], 0);
27626   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[1], 1);
27627   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[2], 2);
27628   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) temp.val[3], 3);
27629   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
27630 }
27631
27632 __extension__ extern __inline void
27633 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27634 vst4_f64 (float64_t * __a, float64x1x4_t val)
27635 {
27636   __builtin_aarch64_simd_xi __o;
27637   float64x2x4_t temp;
27638   temp.val[0] = vcombine_f64 (val.val[0], vcreate_f64 (__AARCH64_UINT64_C (0)));
27639   temp.val[1] = vcombine_f64 (val.val[1], vcreate_f64 (__AARCH64_UINT64_C (0)));
27640   temp.val[2] = vcombine_f64 (val.val[2], vcreate_f64 (__AARCH64_UINT64_C (0)));
27641   temp.val[3] = vcombine_f64 (val.val[3], vcreate_f64 (__AARCH64_UINT64_C (0)));
27642   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[0], 0);
27643   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[1], 1);
27644   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[2], 2);
27645   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) temp.val[3], 3);
27646   __builtin_aarch64_st4df ((__builtin_aarch64_simd_df *) __a, __o);
27647 }
27648
27649 __extension__ extern __inline void
27650 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27651 vst4_s8 (int8_t * __a, int8x8x4_t val)
27652 {
27653   __builtin_aarch64_simd_xi __o;
27654   int8x16x4_t temp;
27655   temp.val[0] = vcombine_s8 (val.val[0], vcreate_s8 (__AARCH64_INT64_C (0)));
27656   temp.val[1] = vcombine_s8 (val.val[1], vcreate_s8 (__AARCH64_INT64_C (0)));
27657   temp.val[2] = vcombine_s8 (val.val[2], vcreate_s8 (__AARCH64_INT64_C (0)));
27658   temp.val[3] = vcombine_s8 (val.val[3], vcreate_s8 (__AARCH64_INT64_C (0)));
27659   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
27660   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
27661   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
27662   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
27663   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27664 }
27665
27666 __extension__ extern __inline void
27667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27668 vst4_p8 (poly8_t * __a, poly8x8x4_t val)
27669 {
27670   __builtin_aarch64_simd_xi __o;
27671   poly8x16x4_t temp;
27672   temp.val[0] = vcombine_p8 (val.val[0], vcreate_p8 (__AARCH64_UINT64_C (0)));
27673   temp.val[1] = vcombine_p8 (val.val[1], vcreate_p8 (__AARCH64_UINT64_C (0)));
27674   temp.val[2] = vcombine_p8 (val.val[2], vcreate_p8 (__AARCH64_UINT64_C (0)));
27675   temp.val[3] = vcombine_p8 (val.val[3], vcreate_p8 (__AARCH64_UINT64_C (0)));
27676   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[0], 0);
27677   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[1], 1);
27678   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[2], 2);
27679   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) temp.val[3], 3);
27680   __builtin_aarch64_st4v8qi ((__builtin_aarch64_simd_qi *) __a, __o);
27681 }
27682
27683 __extension__ extern __inline void
27684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27685 vst4_s16 (int16_t * __a, int16x4x4_t val)
27686 {
27687   __builtin_aarch64_simd_xi __o;
27688   int16x8x4_t temp;
27689   temp.val[0] = vcombine_s16 (val.val[0], vcreate_s16 (__AARCH64_INT64_C (0)));
27690   temp.val[1] = vcombine_s16 (val.val[1], vcreate_s16 (__AARCH64_INT64_C (0)));
27691   temp.val[2] = vcombine_s16 (val.val[2], vcreate_s16 (__AARCH64_INT64_C (0)));
27692   temp.val[3] = vcombine_s16 (val.val[3], vcreate_s16 (__AARCH64_INT64_C (0)));
27693   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
27694   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
27695   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
27696   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
27697   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27698 }
27699
27700 __extension__ extern __inline void
27701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27702 vst4_p16 (poly16_t * __a, poly16x4x4_t val)
27703 {
27704   __builtin_aarch64_simd_xi __o;
27705   poly16x8x4_t temp;
27706   temp.val[0] = vcombine_p16 (val.val[0], vcreate_p16 (__AARCH64_UINT64_C (0)));
27707   temp.val[1] = vcombine_p16 (val.val[1], vcreate_p16 (__AARCH64_UINT64_C (0)));
27708   temp.val[2] = vcombine_p16 (val.val[2], vcreate_p16 (__AARCH64_UINT64_C (0)));
27709   temp.val[3] = vcombine_p16 (val.val[3], vcreate_p16 (__AARCH64_UINT64_C (0)));
27710   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[0], 0);
27711   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[1], 1);
27712   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[2], 2);
27713   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) temp.val[3], 3);
27714   __builtin_aarch64_st4v4hi ((__builtin_aarch64_simd_hi *) __a, __o);
27715 }
27716
27717 __extension__ extern __inline void
27718 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27719 vst4_s32 (int32_t * __a, int32x2x4_t val)
27720 {
27721   __builtin_aarch64_simd_xi __o;
27722   int32x4x4_t temp;
27723   temp.val[0] = vcombine_s32 (val.val[0], vcreate_s32 (__AARCH64_INT64_C (0)));
27724   temp.val[1] = vcombine_s32 (val.val[1], vcreate_s32 (__AARCH64_INT64_C (0)));
27725   temp.val[2] = vcombine_s32 (val.val[2], vcreate_s32 (__AARCH64_INT64_C (0)));
27726   temp.val[3] = vcombine_s32 (val.val[3], vcreate_s32 (__AARCH64_INT64_C (0)));
27727   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[0], 0);
27728   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[1], 1);
27729   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[2], 2);
27730   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[3], 3);
27731   __builtin_aarch64_st4v2si ((__builtin_aarch64_simd_si *) __a, __o);
27732 }
27733
27734 __extension__ extern __inline void
27735 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27736 vst4_u8 (uint8_t * __a, uint8x8x4_t val)
27737 {
27738   __builtin_aarch64_simd_xi __o;
27739   uint8x16x4_t temp;
27740   temp.val[0] = vcombine_u8 (val.val[0], vcreate_u8 (__AARCH64_UINT64_C (0)));
27741   temp.val[1] = vcombine_u8 (val.val[1], vcreate_u8 (__AARCH64_UINT64_C (0)));
27742   temp.val[2] = vcombine_u8 (val.val[2], vcreate_u8 (__AARCH64_UINT64_C (0)));
27743   temp.val[3] = vcombine_u8 (val.val[3], vcreate_u8 (__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_u16 (uint16_t * __a, uint16x4x4_t val)
27754 {
27755   __builtin_aarch64_simd_xi __o;
27756   uint16x8x4_t temp;
27757   temp.val[0] = vcombine_u16 (val.val[0], vcreate_u16 (__AARCH64_UINT64_C (0)));
27758   temp.val[1] = vcombine_u16 (val.val[1], vcreate_u16 (__AARCH64_UINT64_C (0)));
27759   temp.val[2] = vcombine_u16 (val.val[2], vcreate_u16 (__AARCH64_UINT64_C (0)));
27760   temp.val[3] = vcombine_u16 (val.val[3], vcreate_u16 (__AARCH64_UINT64_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_u32 (uint32_t * __a, uint32x2x4_t val)
27771 {
27772   __builtin_aarch64_simd_xi __o;
27773   uint32x4x4_t temp;
27774   temp.val[0] = vcombine_u32 (val.val[0], vcreate_u32 (__AARCH64_UINT64_C (0)));
27775   temp.val[1] = vcombine_u32 (val.val[1], vcreate_u32 (__AARCH64_UINT64_C (0)));
27776   temp.val[2] = vcombine_u32 (val.val[2], vcreate_u32 (__AARCH64_UINT64_C (0)));
27777   temp.val[3] = vcombine_u32 (val.val[3], vcreate_u32 (__AARCH64_UINT64_C (0)));
27778   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[0], 0);
27779   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[1], 1);
27780   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[2], 2);
27781   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) temp.val[3], 3);
27782   __builtin_aarch64_st4v2si ((__builtin_aarch64_simd_si *) __a, __o);
27783 }
27784
27785 __extension__ extern __inline void
27786 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27787 vst4_f16 (float16_t * __a, float16x4x4_t val)
27788 {
27789   __builtin_aarch64_simd_xi __o;
27790   float16x8x4_t temp;
27791   temp.val[0] = vcombine_f16 (val.val[0], vcreate_f16 (__AARCH64_UINT64_C (0)));
27792   temp.val[1] = vcombine_f16 (val.val[1], vcreate_f16 (__AARCH64_UINT64_C (0)));
27793   temp.val[2] = vcombine_f16 (val.val[2], vcreate_f16 (__AARCH64_UINT64_C (0)));
27794   temp.val[3] = vcombine_f16 (val.val[3], vcreate_f16 (__AARCH64_UINT64_C (0)));
27795   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[0], 0);
27796   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[1], 1);
27797   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[2], 2);
27798   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) temp.val[3], 3);
27799   __builtin_aarch64_st4v4hf ((__builtin_aarch64_simd_hf *) __a, __o);
27800 }
27801
27802 __extension__ extern __inline void
27803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27804 vst4_f32 (float32_t * __a, float32x2x4_t val)
27805 {
27806   __builtin_aarch64_simd_xi __o;
27807   float32x4x4_t temp;
27808   temp.val[0] = vcombine_f32 (val.val[0], vcreate_f32 (__AARCH64_UINT64_C (0)));
27809   temp.val[1] = vcombine_f32 (val.val[1], vcreate_f32 (__AARCH64_UINT64_C (0)));
27810   temp.val[2] = vcombine_f32 (val.val[2], vcreate_f32 (__AARCH64_UINT64_C (0)));
27811   temp.val[3] = vcombine_f32 (val.val[3], vcreate_f32 (__AARCH64_UINT64_C (0)));
27812   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[0], 0);
27813   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[1], 1);
27814   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[2], 2);
27815   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) temp.val[3], 3);
27816   __builtin_aarch64_st4v2sf ((__builtin_aarch64_simd_sf *) __a, __o);
27817 }
27818
27819 __extension__ extern __inline void
27820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27821 vst4_p64 (poly64_t * __a, poly64x1x4_t val)
27822 {
27823   __builtin_aarch64_simd_xi __o;
27824   poly64x2x4_t temp;
27825   temp.val[0] = vcombine_p64 (val.val[0], vcreate_p64 (__AARCH64_UINT64_C (0)));
27826   temp.val[1] = vcombine_p64 (val.val[1], vcreate_p64 (__AARCH64_UINT64_C (0)));
27827   temp.val[2] = vcombine_p64 (val.val[2], vcreate_p64 (__AARCH64_UINT64_C (0)));
27828   temp.val[3] = vcombine_p64 (val.val[3], vcreate_p64 (__AARCH64_UINT64_C (0)));
27829   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27830                                                (poly64x2_t) temp.val[0], 0);
27831   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27832                                                (poly64x2_t) temp.val[1], 1);
27833   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27834                                                (poly64x2_t) temp.val[2], 2);
27835   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
27836                                                (poly64x2_t) temp.val[3], 3);
27837   __builtin_aarch64_st4di ((__builtin_aarch64_simd_di *) __a, __o);
27838 }
27839
27840 __extension__ extern __inline void
27841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27842 vst4q_s8 (int8_t * __a, int8x16x4_t val)
27843 {
27844   __builtin_aarch64_simd_xi __o;
27845   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
27846   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
27847   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
27848   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
27849   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27850 }
27851
27852 __extension__ extern __inline void
27853 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27854 vst4q_p8 (poly8_t * __a, poly8x16x4_t val)
27855 {
27856   __builtin_aarch64_simd_xi __o;
27857   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
27858   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
27859   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
27860   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
27861   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27862 }
27863
27864 __extension__ extern __inline void
27865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27866 vst4q_s16 (int16_t * __a, int16x8x4_t val)
27867 {
27868   __builtin_aarch64_simd_xi __o;
27869   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
27870   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
27871   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
27872   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
27873   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27874 }
27875
27876 __extension__ extern __inline void
27877 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27878 vst4q_p16 (poly16_t * __a, poly16x8x4_t val)
27879 {
27880   __builtin_aarch64_simd_xi __o;
27881   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
27882   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
27883   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
27884   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
27885   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27886 }
27887
27888 __extension__ extern __inline void
27889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27890 vst4q_s32 (int32_t * __a, int32x4x4_t val)
27891 {
27892   __builtin_aarch64_simd_xi __o;
27893   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[0], 0);
27894   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[1], 1);
27895   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[2], 2);
27896   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[3], 3);
27897   __builtin_aarch64_st4v4si ((__builtin_aarch64_simd_si *) __a, __o);
27898 }
27899
27900 __extension__ extern __inline void
27901 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27902 vst4q_s64 (int64_t * __a, int64x2x4_t val)
27903 {
27904   __builtin_aarch64_simd_xi __o;
27905   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[0], 0);
27906   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[1], 1);
27907   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[2], 2);
27908   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[3], 3);
27909   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
27910 }
27911
27912 __extension__ extern __inline void
27913 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27914 vst4q_u8 (uint8_t * __a, uint8x16x4_t val)
27915 {
27916   __builtin_aarch64_simd_xi __o;
27917   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[0], 0);
27918   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[1], 1);
27919   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[2], 2);
27920   __o = __builtin_aarch64_set_qregxiv16qi (__o, (int8x16_t) val.val[3], 3);
27921   __builtin_aarch64_st4v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
27922 }
27923
27924 __extension__ extern __inline void
27925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27926 vst4q_u16 (uint16_t * __a, uint16x8x4_t val)
27927 {
27928   __builtin_aarch64_simd_xi __o;
27929   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[0], 0);
27930   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[1], 1);
27931   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[2], 2);
27932   __o = __builtin_aarch64_set_qregxiv8hi (__o, (int16x8_t) val.val[3], 3);
27933   __builtin_aarch64_st4v8hi ((__builtin_aarch64_simd_hi *) __a, __o);
27934 }
27935
27936 __extension__ extern __inline void
27937 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27938 vst4q_u32 (uint32_t * __a, uint32x4x4_t val)
27939 {
27940   __builtin_aarch64_simd_xi __o;
27941   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[0], 0);
27942   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[1], 1);
27943   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[2], 2);
27944   __o = __builtin_aarch64_set_qregxiv4si (__o, (int32x4_t) val.val[3], 3);
27945   __builtin_aarch64_st4v4si ((__builtin_aarch64_simd_si *) __a, __o);
27946 }
27947
27948 __extension__ extern __inline void
27949 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27950 vst4q_u64 (uint64_t * __a, uint64x2x4_t val)
27951 {
27952   __builtin_aarch64_simd_xi __o;
27953   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[0], 0);
27954   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[1], 1);
27955   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[2], 2);
27956   __o = __builtin_aarch64_set_qregxiv2di (__o, (int64x2_t) val.val[3], 3);
27957   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
27958 }
27959
27960 __extension__ extern __inline void
27961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27962 vst4q_f16 (float16_t * __a, float16x8x4_t val)
27963 {
27964   __builtin_aarch64_simd_xi __o;
27965   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[0], 0);
27966   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[1], 1);
27967   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[2], 2);
27968   __o = __builtin_aarch64_set_qregxiv8hf (__o, (float16x8_t) val.val[3], 3);
27969   __builtin_aarch64_st4v8hf ((__builtin_aarch64_simd_hf *) __a, __o);
27970 }
27971
27972 __extension__ extern __inline void
27973 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27974 vst4q_f32 (float32_t * __a, float32x4x4_t val)
27975 {
27976   __builtin_aarch64_simd_xi __o;
27977   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[0], 0);
27978   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[1], 1);
27979   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[2], 2);
27980   __o = __builtin_aarch64_set_qregxiv4sf (__o, (float32x4_t) val.val[3], 3);
27981   __builtin_aarch64_st4v4sf ((__builtin_aarch64_simd_sf *) __a, __o);
27982 }
27983
27984 __extension__ extern __inline void
27985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27986 vst4q_f64 (float64_t * __a, float64x2x4_t val)
27987 {
27988   __builtin_aarch64_simd_xi __o;
27989   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[0], 0);
27990   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[1], 1);
27991   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[2], 2);
27992   __o = __builtin_aarch64_set_qregxiv2df (__o, (float64x2_t) val.val[3], 3);
27993   __builtin_aarch64_st4v2df ((__builtin_aarch64_simd_df *) __a, __o);
27994 }
27995
27996 __extension__ extern __inline void
27997 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
27998 vst4q_p64 (poly64_t * __a, poly64x2x4_t val)
27999 {
28000   __builtin_aarch64_simd_xi __o;
28001   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28002                                                (poly64x2_t) val.val[0], 0);
28003   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28004                                                (poly64x2_t) val.val[1], 1);
28005   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28006                                                (poly64x2_t) val.val[2], 2);
28007   __o = __builtin_aarch64_set_qregxiv2di_ssps (__o,
28008                                                (poly64x2_t) val.val[3], 3);
28009   __builtin_aarch64_st4v2di ((__builtin_aarch64_simd_di *) __a, __o);
28010 }
28011
28012 /* vsub */
28013
28014 __extension__ extern __inline int64_t
28015 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28016 vsubd_s64 (int64_t __a, int64_t __b)
28017 {
28018   return __a - __b;
28019 }
28020
28021 __extension__ extern __inline uint64_t
28022 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28023 vsubd_u64 (uint64_t __a, uint64_t __b)
28024 {
28025   return __a - __b;
28026 }
28027
28028 /* vtbx1  */
28029
28030 __extension__ extern __inline int8x8_t
28031 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28032 vtbx1_s8 (int8x8_t __r, int8x8_t __tab, int8x8_t __idx)
28033 {
28034   uint8x8_t __mask = vclt_u8 (vreinterpret_u8_s8 (__idx),
28035                               vmov_n_u8 (8));
28036   int8x8_t __tbl = vtbl1_s8 (__tab, __idx);
28037
28038   return vbsl_s8 (__mask, __tbl, __r);
28039 }
28040
28041 __extension__ extern __inline uint8x8_t
28042 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28043 vtbx1_u8 (uint8x8_t __r, uint8x8_t __tab, uint8x8_t __idx)
28044 {
28045   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (8));
28046   uint8x8_t __tbl = vtbl1_u8 (__tab, __idx);
28047
28048   return vbsl_u8 (__mask, __tbl, __r);
28049 }
28050
28051 __extension__ extern __inline poly8x8_t
28052 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28053 vtbx1_p8 (poly8x8_t __r, poly8x8_t __tab, uint8x8_t __idx)
28054 {
28055   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (8));
28056   poly8x8_t __tbl = vtbl1_p8 (__tab, __idx);
28057
28058   return vbsl_p8 (__mask, __tbl, __r);
28059 }
28060
28061 /* vtbx3  */
28062
28063 __extension__ extern __inline int8x8_t
28064 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28065 vtbx3_s8 (int8x8_t __r, int8x8x3_t __tab, int8x8_t __idx)
28066 {
28067   uint8x8_t __mask = vclt_u8 (vreinterpret_u8_s8 (__idx),
28068                               vmov_n_u8 (24));
28069   int8x8_t __tbl = vtbl3_s8 (__tab, __idx);
28070
28071   return vbsl_s8 (__mask, __tbl, __r);
28072 }
28073
28074 __extension__ extern __inline uint8x8_t
28075 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28076 vtbx3_u8 (uint8x8_t __r, uint8x8x3_t __tab, uint8x8_t __idx)
28077 {
28078   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (24));
28079   uint8x8_t __tbl = vtbl3_u8 (__tab, __idx);
28080
28081   return vbsl_u8 (__mask, __tbl, __r);
28082 }
28083
28084 __extension__ extern __inline poly8x8_t
28085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28086 vtbx3_p8 (poly8x8_t __r, poly8x8x3_t __tab, uint8x8_t __idx)
28087 {
28088   uint8x8_t __mask = vclt_u8 (__idx, vmov_n_u8 (24));
28089   poly8x8_t __tbl = vtbl3_p8 (__tab, __idx);
28090
28091   return vbsl_p8 (__mask, __tbl, __r);
28092 }
28093
28094 /* vtbx4  */
28095
28096 __extension__ extern __inline int8x8_t
28097 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28098 vtbx4_s8 (int8x8_t __r, int8x8x4_t __tab, int8x8_t __idx)
28099 {
28100   int8x8_t result;
28101   int8x16x2_t temp;
28102   __builtin_aarch64_simd_oi __o;
28103   temp.val[0] = vcombine_s8 (__tab.val[0], __tab.val[1]);
28104   temp.val[1] = vcombine_s8 (__tab.val[2], __tab.val[3]);
28105   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28106                                            (int8x16_t) temp.val[0], 0);
28107   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28108                                            (int8x16_t) temp.val[1], 1);
28109   result = __builtin_aarch64_tbx4v8qi (__r, __o, __idx);
28110   return result;
28111 }
28112
28113 __extension__ extern __inline uint8x8_t
28114 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28115 vtbx4_u8 (uint8x8_t __r, uint8x8x4_t __tab, uint8x8_t __idx)
28116 {
28117   uint8x8_t result;
28118   uint8x16x2_t temp;
28119   __builtin_aarch64_simd_oi __o;
28120   temp.val[0] = vcombine_u8 (__tab.val[0], __tab.val[1]);
28121   temp.val[1] = vcombine_u8 (__tab.val[2], __tab.val[3]);
28122   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28123                                            (int8x16_t) temp.val[0], 0);
28124   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28125                                            (int8x16_t) temp.val[1], 1);
28126   result = (uint8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)__r, __o,
28127                                                   (int8x8_t)__idx);
28128   return result;
28129 }
28130
28131 __extension__ extern __inline poly8x8_t
28132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28133 vtbx4_p8 (poly8x8_t __r, poly8x8x4_t __tab, uint8x8_t __idx)
28134 {
28135   poly8x8_t result;
28136   poly8x16x2_t temp;
28137   __builtin_aarch64_simd_oi __o;
28138   temp.val[0] = vcombine_p8 (__tab.val[0], __tab.val[1]);
28139   temp.val[1] = vcombine_p8 (__tab.val[2], __tab.val[3]);
28140   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28141                                            (int8x16_t) temp.val[0], 0);
28142   __o = __builtin_aarch64_set_qregoiv16qi (__o,
28143                                            (int8x16_t) temp.val[1], 1);
28144   result = (poly8x8_t)__builtin_aarch64_tbx4v8qi ((int8x8_t)__r, __o,
28145                                                   (int8x8_t)__idx);
28146   return result;
28147 }
28148
28149 /* vtrn */
28150
28151 __extension__ extern __inline float16x4_t
28152 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28153 vtrn1_f16 (float16x4_t __a, float16x4_t __b)
28154 {
28155 #ifdef __AARCH64EB__
28156   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28157 #else
28158   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28159 #endif
28160 }
28161
28162 __extension__ extern __inline float32x2_t
28163 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28164 vtrn1_f32 (float32x2_t __a, float32x2_t __b)
28165 {
28166 #ifdef __AARCH64EB__
28167   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
28168 #else
28169   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
28170 #endif
28171 }
28172
28173 __extension__ extern __inline poly8x8_t
28174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28175 vtrn1_p8 (poly8x8_t __a, poly8x8_t __b)
28176 {
28177 #ifdef __AARCH64EB__
28178   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28179 #else
28180   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28181 #endif
28182 }
28183
28184 __extension__ extern __inline poly16x4_t
28185 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28186 vtrn1_p16 (poly16x4_t __a, poly16x4_t __b)
28187 {
28188 #ifdef __AARCH64EB__
28189   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28190 #else
28191   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28192 #endif
28193 }
28194
28195 __extension__ extern __inline int8x8_t
28196 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28197 vtrn1_s8 (int8x8_t __a, int8x8_t __b)
28198 {
28199 #ifdef __AARCH64EB__
28200   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28201 #else
28202   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28203 #endif
28204 }
28205
28206 __extension__ extern __inline int16x4_t
28207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28208 vtrn1_s16 (int16x4_t __a, int16x4_t __b)
28209 {
28210 #ifdef __AARCH64EB__
28211   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28212 #else
28213   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28214 #endif
28215 }
28216
28217 __extension__ extern __inline int32x2_t
28218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28219 vtrn1_s32 (int32x2_t __a, int32x2_t __b)
28220 {
28221 #ifdef __AARCH64EB__
28222   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
28223 #else
28224   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
28225 #endif
28226 }
28227
28228 __extension__ extern __inline uint8x8_t
28229 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28230 vtrn1_u8 (uint8x8_t __a, uint8x8_t __b)
28231 {
28232 #ifdef __AARCH64EB__
28233   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28234 #else
28235   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28236 #endif
28237 }
28238
28239 __extension__ extern __inline uint16x4_t
28240 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28241 vtrn1_u16 (uint16x4_t __a, uint16x4_t __b)
28242 {
28243 #ifdef __AARCH64EB__
28244   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
28245 #else
28246   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
28247 #endif
28248 }
28249
28250 __extension__ extern __inline uint32x2_t
28251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28252 vtrn1_u32 (uint32x2_t __a, uint32x2_t __b)
28253 {
28254 #ifdef __AARCH64EB__
28255   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
28256 #else
28257   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
28258 #endif
28259 }
28260
28261 __extension__ extern __inline float16x8_t
28262 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28263 vtrn1q_f16 (float16x8_t __a, float16x8_t __b)
28264 {
28265 #ifdef __AARCH64EB__
28266   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28267 #else
28268   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28269 #endif
28270 }
28271
28272 __extension__ extern __inline float32x4_t
28273 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28274 vtrn1q_f32 (float32x4_t __a, float32x4_t __b)
28275 {
28276 #ifdef __AARCH64EB__
28277   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
28278 #else
28279   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
28280 #endif
28281 }
28282
28283 __extension__ extern __inline float64x2_t
28284 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28285 vtrn1q_f64 (float64x2_t __a, float64x2_t __b)
28286 {
28287 #ifdef __AARCH64EB__
28288   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
28289 #else
28290   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
28291 #endif
28292 }
28293
28294 __extension__ extern __inline poly8x16_t
28295 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28296 vtrn1q_p8 (poly8x16_t __a, poly8x16_t __b)
28297 {
28298 #ifdef __AARCH64EB__
28299   return __builtin_shuffle (__a, __b,
28300       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
28301 #else
28302   return __builtin_shuffle (__a, __b,
28303       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
28304 #endif
28305 }
28306
28307 __extension__ extern __inline poly16x8_t
28308 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28309 vtrn1q_p16 (poly16x8_t __a, poly16x8_t __b)
28310 {
28311 #ifdef __AARCH64EB__
28312   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28313 #else
28314   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28315 #endif
28316 }
28317
28318 __extension__ extern __inline int8x16_t
28319 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28320 vtrn1q_s8 (int8x16_t __a, int8x16_t __b)
28321 {
28322 #ifdef __AARCH64EB__
28323   return __builtin_shuffle (__a, __b,
28324       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
28325 #else
28326   return __builtin_shuffle (__a, __b,
28327       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
28328 #endif
28329 }
28330
28331 __extension__ extern __inline int16x8_t
28332 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28333 vtrn1q_s16 (int16x8_t __a, int16x8_t __b)
28334 {
28335 #ifdef __AARCH64EB__
28336   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28337 #else
28338   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28339 #endif
28340 }
28341
28342 __extension__ extern __inline int32x4_t
28343 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28344 vtrn1q_s32 (int32x4_t __a, int32x4_t __b)
28345 {
28346 #ifdef __AARCH64EB__
28347   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
28348 #else
28349   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
28350 #endif
28351 }
28352
28353 __extension__ extern __inline int64x2_t
28354 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28355 vtrn1q_s64 (int64x2_t __a, int64x2_t __b)
28356 {
28357 #ifdef __AARCH64EB__
28358   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
28359 #else
28360   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
28361 #endif
28362 }
28363
28364 __extension__ extern __inline uint8x16_t
28365 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28366 vtrn1q_u8 (uint8x16_t __a, uint8x16_t __b)
28367 {
28368 #ifdef __AARCH64EB__
28369   return __builtin_shuffle (__a, __b,
28370       (uint8x16_t) {17, 1, 19, 3, 21, 5, 23, 7, 25, 9, 27, 11, 29, 13, 31, 15});
28371 #else
28372   return __builtin_shuffle (__a, __b,
28373       (uint8x16_t) {0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30});
28374 #endif
28375 }
28376
28377 __extension__ extern __inline uint16x8_t
28378 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28379 vtrn1q_u16 (uint16x8_t __a, uint16x8_t __b)
28380 {
28381 #ifdef __AARCH64EB__
28382   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
28383 #else
28384   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
28385 #endif
28386 }
28387
28388 __extension__ extern __inline uint32x4_t
28389 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28390 vtrn1q_u32 (uint32x4_t __a, uint32x4_t __b)
28391 {
28392 #ifdef __AARCH64EB__
28393   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 1, 7, 3});
28394 #else
28395   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 2, 6});
28396 #endif
28397 }
28398
28399 __extension__ extern __inline uint64x2_t
28400 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28401 vtrn1q_u64 (uint64x2_t __a, uint64x2_t __b)
28402 {
28403 #ifdef __AARCH64EB__
28404   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
28405 #else
28406   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
28407 #endif
28408 }
28409
28410 __extension__ extern __inline float16x4_t
28411 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28412 vtrn2_f16 (float16x4_t __a, float16x4_t __b)
28413 {
28414 #ifdef __AARCH64EB__
28415   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28416 #else
28417   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28418 #endif
28419 }
28420
28421 __extension__ extern __inline float32x2_t
28422 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28423 vtrn2_f32 (float32x2_t __a, float32x2_t __b)
28424 {
28425 #ifdef __AARCH64EB__
28426   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
28427 #else
28428   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
28429 #endif
28430 }
28431
28432 __extension__ extern __inline poly8x8_t
28433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28434 vtrn2_p8 (poly8x8_t __a, poly8x8_t __b)
28435 {
28436 #ifdef __AARCH64EB__
28437   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28438 #else
28439   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28440 #endif
28441 }
28442
28443 __extension__ extern __inline poly16x4_t
28444 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28445 vtrn2_p16 (poly16x4_t __a, poly16x4_t __b)
28446 {
28447 #ifdef __AARCH64EB__
28448   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28449 #else
28450   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28451 #endif
28452 }
28453
28454 __extension__ extern __inline int8x8_t
28455 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28456 vtrn2_s8 (int8x8_t __a, int8x8_t __b)
28457 {
28458 #ifdef __AARCH64EB__
28459   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28460 #else
28461   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28462 #endif
28463 }
28464
28465 __extension__ extern __inline int16x4_t
28466 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28467 vtrn2_s16 (int16x4_t __a, int16x4_t __b)
28468 {
28469 #ifdef __AARCH64EB__
28470   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28471 #else
28472   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28473 #endif
28474 }
28475
28476 __extension__ extern __inline int32x2_t
28477 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28478 vtrn2_s32 (int32x2_t __a, int32x2_t __b)
28479 {
28480 #ifdef __AARCH64EB__
28481   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
28482 #else
28483   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
28484 #endif
28485 }
28486
28487 __extension__ extern __inline uint8x8_t
28488 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28489 vtrn2_u8 (uint8x8_t __a, uint8x8_t __b)
28490 {
28491 #ifdef __AARCH64EB__
28492   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28493 #else
28494   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28495 #endif
28496 }
28497
28498 __extension__ extern __inline uint16x4_t
28499 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28500 vtrn2_u16 (uint16x4_t __a, uint16x4_t __b)
28501 {
28502 #ifdef __AARCH64EB__
28503   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
28504 #else
28505   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
28506 #endif
28507 }
28508
28509 __extension__ extern __inline uint32x2_t
28510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28511 vtrn2_u32 (uint32x2_t __a, uint32x2_t __b)
28512 {
28513 #ifdef __AARCH64EB__
28514   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
28515 #else
28516   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
28517 #endif
28518 }
28519
28520 __extension__ extern __inline float16x8_t
28521 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28522 vtrn2q_f16 (float16x8_t __a, float16x8_t __b)
28523 {
28524 #ifdef __AARCH64EB__
28525   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28526 #else
28527   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28528 #endif
28529 }
28530
28531 __extension__ extern __inline float32x4_t
28532 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28533 vtrn2q_f32 (float32x4_t __a, float32x4_t __b)
28534 {
28535 #ifdef __AARCH64EB__
28536   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
28537 #else
28538   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
28539 #endif
28540 }
28541
28542 __extension__ extern __inline float64x2_t
28543 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28544 vtrn2q_f64 (float64x2_t __a, float64x2_t __b)
28545 {
28546 #ifdef __AARCH64EB__
28547   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
28548 #else
28549   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
28550 #endif
28551 }
28552
28553 __extension__ extern __inline poly8x16_t
28554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28555 vtrn2q_p8 (poly8x16_t __a, poly8x16_t __b)
28556 {
28557 #ifdef __AARCH64EB__
28558   return __builtin_shuffle (__a, __b,
28559       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
28560 #else
28561   return __builtin_shuffle (__a, __b,
28562       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
28563 #endif
28564 }
28565
28566 __extension__ extern __inline poly16x8_t
28567 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28568 vtrn2q_p16 (poly16x8_t __a, poly16x8_t __b)
28569 {
28570 #ifdef __AARCH64EB__
28571   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28572 #else
28573   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28574 #endif
28575 }
28576
28577 __extension__ extern __inline int8x16_t
28578 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28579 vtrn2q_s8 (int8x16_t __a, int8x16_t __b)
28580 {
28581 #ifdef __AARCH64EB__
28582   return __builtin_shuffle (__a, __b,
28583       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
28584 #else
28585   return __builtin_shuffle (__a, __b,
28586       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
28587 #endif
28588 }
28589
28590 __extension__ extern __inline int16x8_t
28591 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28592 vtrn2q_s16 (int16x8_t __a, int16x8_t __b)
28593 {
28594 #ifdef __AARCH64EB__
28595   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28596 #else
28597   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28598 #endif
28599 }
28600
28601 __extension__ extern __inline int32x4_t
28602 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28603 vtrn2q_s32 (int32x4_t __a, int32x4_t __b)
28604 {
28605 #ifdef __AARCH64EB__
28606   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
28607 #else
28608   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
28609 #endif
28610 }
28611
28612 __extension__ extern __inline int64x2_t
28613 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28614 vtrn2q_s64 (int64x2_t __a, int64x2_t __b)
28615 {
28616 #ifdef __AARCH64EB__
28617   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
28618 #else
28619   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
28620 #endif
28621 }
28622
28623 __extension__ extern __inline uint8x16_t
28624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28625 vtrn2q_u8 (uint8x16_t __a, uint8x16_t __b)
28626 {
28627 #ifdef __AARCH64EB__
28628   return __builtin_shuffle (__a, __b,
28629       (uint8x16_t) {16, 0, 18, 2, 20, 4, 22, 6, 24, 8, 26, 10, 28, 12, 30, 14});
28630 #else
28631   return __builtin_shuffle (__a, __b,
28632       (uint8x16_t) {1, 17, 3, 19, 5, 21, 7, 23, 9, 25, 11, 27, 13, 29, 15, 31});
28633 #endif
28634 }
28635
28636 __extension__ extern __inline uint16x8_t
28637 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28638 vtrn2q_u16 (uint16x8_t __a, uint16x8_t __b)
28639 {
28640 #ifdef __AARCH64EB__
28641   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
28642 #else
28643   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
28644 #endif
28645 }
28646
28647 __extension__ extern __inline uint32x4_t
28648 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28649 vtrn2q_u32 (uint32x4_t __a, uint32x4_t __b)
28650 {
28651 #ifdef __AARCH64EB__
28652   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 6, 2});
28653 #else
28654   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 5, 3, 7});
28655 #endif
28656 }
28657
28658 __extension__ extern __inline uint64x2_t
28659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28660 vtrn2q_u64 (uint64x2_t __a, uint64x2_t __b)
28661 {
28662 #ifdef __AARCH64EB__
28663   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
28664 #else
28665   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
28666 #endif
28667 }
28668
28669 __extension__ extern __inline float16x4x2_t
28670 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28671 vtrn_f16 (float16x4_t __a, float16x4_t __b)
28672 {
28673   return (float16x4x2_t) {vtrn1_f16 (__a, __b), vtrn2_f16 (__a, __b)};
28674 }
28675
28676 __extension__ extern __inline float32x2x2_t
28677 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28678 vtrn_f32 (float32x2_t a, float32x2_t b)
28679 {
28680   return (float32x2x2_t) {vtrn1_f32 (a, b), vtrn2_f32 (a, b)};
28681 }
28682
28683 __extension__ extern __inline poly8x8x2_t
28684 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28685 vtrn_p8 (poly8x8_t a, poly8x8_t b)
28686 {
28687   return (poly8x8x2_t) {vtrn1_p8 (a, b), vtrn2_p8 (a, b)};
28688 }
28689
28690 __extension__ extern __inline poly16x4x2_t
28691 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28692 vtrn_p16 (poly16x4_t a, poly16x4_t b)
28693 {
28694   return (poly16x4x2_t) {vtrn1_p16 (a, b), vtrn2_p16 (a, b)};
28695 }
28696
28697 __extension__ extern __inline int8x8x2_t
28698 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28699 vtrn_s8 (int8x8_t a, int8x8_t b)
28700 {
28701   return (int8x8x2_t) {vtrn1_s8 (a, b), vtrn2_s8 (a, b)};
28702 }
28703
28704 __extension__ extern __inline int16x4x2_t
28705 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28706 vtrn_s16 (int16x4_t a, int16x4_t b)
28707 {
28708   return (int16x4x2_t) {vtrn1_s16 (a, b), vtrn2_s16 (a, b)};
28709 }
28710
28711 __extension__ extern __inline int32x2x2_t
28712 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28713 vtrn_s32 (int32x2_t a, int32x2_t b)
28714 {
28715   return (int32x2x2_t) {vtrn1_s32 (a, b), vtrn2_s32 (a, b)};
28716 }
28717
28718 __extension__ extern __inline uint8x8x2_t
28719 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28720 vtrn_u8 (uint8x8_t a, uint8x8_t b)
28721 {
28722   return (uint8x8x2_t) {vtrn1_u8 (a, b), vtrn2_u8 (a, b)};
28723 }
28724
28725 __extension__ extern __inline uint16x4x2_t
28726 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28727 vtrn_u16 (uint16x4_t a, uint16x4_t b)
28728 {
28729   return (uint16x4x2_t) {vtrn1_u16 (a, b), vtrn2_u16 (a, b)};
28730 }
28731
28732 __extension__ extern __inline uint32x2x2_t
28733 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28734 vtrn_u32 (uint32x2_t a, uint32x2_t b)
28735 {
28736   return (uint32x2x2_t) {vtrn1_u32 (a, b), vtrn2_u32 (a, b)};
28737 }
28738
28739 __extension__ extern __inline float16x8x2_t
28740 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28741 vtrnq_f16 (float16x8_t __a, float16x8_t __b)
28742 {
28743   return (float16x8x2_t) {vtrn1q_f16 (__a, __b), vtrn2q_f16 (__a, __b)};
28744 }
28745
28746 __extension__ extern __inline float32x4x2_t
28747 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28748 vtrnq_f32 (float32x4_t a, float32x4_t b)
28749 {
28750   return (float32x4x2_t) {vtrn1q_f32 (a, b), vtrn2q_f32 (a, b)};
28751 }
28752
28753 __extension__ extern __inline poly8x16x2_t
28754 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28755 vtrnq_p8 (poly8x16_t a, poly8x16_t b)
28756 {
28757   return (poly8x16x2_t) {vtrn1q_p8 (a, b), vtrn2q_p8 (a, b)};
28758 }
28759
28760 __extension__ extern __inline poly16x8x2_t
28761 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28762 vtrnq_p16 (poly16x8_t a, poly16x8_t b)
28763 {
28764   return (poly16x8x2_t) {vtrn1q_p16 (a, b), vtrn2q_p16 (a, b)};
28765 }
28766
28767 __extension__ extern __inline int8x16x2_t
28768 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28769 vtrnq_s8 (int8x16_t a, int8x16_t b)
28770 {
28771   return (int8x16x2_t) {vtrn1q_s8 (a, b), vtrn2q_s8 (a, b)};
28772 }
28773
28774 __extension__ extern __inline int16x8x2_t
28775 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28776 vtrnq_s16 (int16x8_t a, int16x8_t b)
28777 {
28778   return (int16x8x2_t) {vtrn1q_s16 (a, b), vtrn2q_s16 (a, b)};
28779 }
28780
28781 __extension__ extern __inline int32x4x2_t
28782 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28783 vtrnq_s32 (int32x4_t a, int32x4_t b)
28784 {
28785   return (int32x4x2_t) {vtrn1q_s32 (a, b), vtrn2q_s32 (a, b)};
28786 }
28787
28788 __extension__ extern __inline uint8x16x2_t
28789 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28790 vtrnq_u8 (uint8x16_t a, uint8x16_t b)
28791 {
28792   return (uint8x16x2_t) {vtrn1q_u8 (a, b), vtrn2q_u8 (a, b)};
28793 }
28794
28795 __extension__ extern __inline uint16x8x2_t
28796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28797 vtrnq_u16 (uint16x8_t a, uint16x8_t b)
28798 {
28799   return (uint16x8x2_t) {vtrn1q_u16 (a, b), vtrn2q_u16 (a, b)};
28800 }
28801
28802 __extension__ extern __inline uint32x4x2_t
28803 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28804 vtrnq_u32 (uint32x4_t a, uint32x4_t b)
28805 {
28806   return (uint32x4x2_t) {vtrn1q_u32 (a, b), vtrn2q_u32 (a, b)};
28807 }
28808
28809 /* vtst */
28810
28811 __extension__ extern __inline uint8x8_t
28812 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28813 vtst_s8 (int8x8_t __a, int8x8_t __b)
28814 {
28815   return (uint8x8_t) ((__a & __b) != 0);
28816 }
28817
28818 __extension__ extern __inline uint16x4_t
28819 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28820 vtst_s16 (int16x4_t __a, int16x4_t __b)
28821 {
28822   return (uint16x4_t) ((__a & __b) != 0);
28823 }
28824
28825 __extension__ extern __inline uint32x2_t
28826 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28827 vtst_s32 (int32x2_t __a, int32x2_t __b)
28828 {
28829   return (uint32x2_t) ((__a & __b) != 0);
28830 }
28831
28832 __extension__ extern __inline uint64x1_t
28833 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28834 vtst_s64 (int64x1_t __a, int64x1_t __b)
28835 {
28836   return (uint64x1_t) ((__a & __b) != __AARCH64_INT64_C (0));
28837 }
28838
28839 __extension__ extern __inline uint8x8_t
28840 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28841 vtst_u8 (uint8x8_t __a, uint8x8_t __b)
28842 {
28843   return ((__a & __b) != 0);
28844 }
28845
28846 __extension__ extern __inline uint16x4_t
28847 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28848 vtst_u16 (uint16x4_t __a, uint16x4_t __b)
28849 {
28850   return ((__a & __b) != 0);
28851 }
28852
28853 __extension__ extern __inline uint32x2_t
28854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28855 vtst_u32 (uint32x2_t __a, uint32x2_t __b)
28856 {
28857   return ((__a & __b) != 0);
28858 }
28859
28860 __extension__ extern __inline uint64x1_t
28861 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28862 vtst_u64 (uint64x1_t __a, uint64x1_t __b)
28863 {
28864   return ((__a & __b) != __AARCH64_UINT64_C (0));
28865 }
28866
28867 __extension__ extern __inline uint8x16_t
28868 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28869 vtstq_s8 (int8x16_t __a, int8x16_t __b)
28870 {
28871   return (uint8x16_t) ((__a & __b) != 0);
28872 }
28873
28874 __extension__ extern __inline uint16x8_t
28875 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28876 vtstq_s16 (int16x8_t __a, int16x8_t __b)
28877 {
28878   return (uint16x8_t) ((__a & __b) != 0);
28879 }
28880
28881 __extension__ extern __inline uint32x4_t
28882 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28883 vtstq_s32 (int32x4_t __a, int32x4_t __b)
28884 {
28885   return (uint32x4_t) ((__a & __b) != 0);
28886 }
28887
28888 __extension__ extern __inline uint64x2_t
28889 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28890 vtstq_s64 (int64x2_t __a, int64x2_t __b)
28891 {
28892   return (uint64x2_t) ((__a & __b) != __AARCH64_INT64_C (0));
28893 }
28894
28895 __extension__ extern __inline uint8x16_t
28896 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28897 vtstq_u8 (uint8x16_t __a, uint8x16_t __b)
28898 {
28899   return ((__a & __b) != 0);
28900 }
28901
28902 __extension__ extern __inline uint16x8_t
28903 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28904 vtstq_u16 (uint16x8_t __a, uint16x8_t __b)
28905 {
28906   return ((__a & __b) != 0);
28907 }
28908
28909 __extension__ extern __inline uint32x4_t
28910 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28911 vtstq_u32 (uint32x4_t __a, uint32x4_t __b)
28912 {
28913   return ((__a & __b) != 0);
28914 }
28915
28916 __extension__ extern __inline uint64x2_t
28917 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28918 vtstq_u64 (uint64x2_t __a, uint64x2_t __b)
28919 {
28920   return ((__a & __b) != __AARCH64_UINT64_C (0));
28921 }
28922
28923 __extension__ extern __inline uint64_t
28924 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28925 vtstd_s64 (int64_t __a, int64_t __b)
28926 {
28927   return (__a & __b) ? -1ll : 0ll;
28928 }
28929
28930 __extension__ extern __inline uint64_t
28931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28932 vtstd_u64 (uint64_t __a, uint64_t __b)
28933 {
28934   return (__a & __b) ? -1ll : 0ll;
28935 }
28936
28937 /* vuqadd */
28938
28939 __extension__ extern __inline int8x8_t
28940 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28941 vuqadd_s8 (int8x8_t __a, uint8x8_t __b)
28942 {
28943   return __builtin_aarch64_suqaddv8qi_ssu (__a,  __b);
28944 }
28945
28946 __extension__ extern __inline int16x4_t
28947 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28948 vuqadd_s16 (int16x4_t __a, uint16x4_t __b)
28949 {
28950   return __builtin_aarch64_suqaddv4hi_ssu (__a,  __b);
28951 }
28952
28953 __extension__ extern __inline int32x2_t
28954 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28955 vuqadd_s32 (int32x2_t __a, uint32x2_t __b)
28956 {
28957   return __builtin_aarch64_suqaddv2si_ssu (__a,  __b);
28958 }
28959
28960 __extension__ extern __inline int64x1_t
28961 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28962 vuqadd_s64 (int64x1_t __a, uint64x1_t __b)
28963 {
28964   return (int64x1_t) {__builtin_aarch64_suqadddi_ssu (__a[0], __b[0])};
28965 }
28966
28967 __extension__ extern __inline int8x16_t
28968 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28969 vuqaddq_s8 (int8x16_t __a, uint8x16_t __b)
28970 {
28971   return __builtin_aarch64_suqaddv16qi_ssu (__a,  __b);
28972 }
28973
28974 __extension__ extern __inline int16x8_t
28975 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28976 vuqaddq_s16 (int16x8_t __a, uint16x8_t __b)
28977 {
28978   return __builtin_aarch64_suqaddv8hi_ssu (__a,  __b);
28979 }
28980
28981 __extension__ extern __inline int32x4_t
28982 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28983 vuqaddq_s32 (int32x4_t __a, uint32x4_t __b)
28984 {
28985   return __builtin_aarch64_suqaddv4si_ssu (__a,  __b);
28986 }
28987
28988 __extension__ extern __inline int64x2_t
28989 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28990 vuqaddq_s64 (int64x2_t __a, uint64x2_t __b)
28991 {
28992   return __builtin_aarch64_suqaddv2di_ssu (__a,  __b);
28993 }
28994
28995 __extension__ extern __inline int8_t
28996 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
28997 vuqaddb_s8 (int8_t __a, uint8_t __b)
28998 {
28999   return __builtin_aarch64_suqaddqi_ssu (__a,  __b);
29000 }
29001
29002 __extension__ extern __inline int16_t
29003 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29004 vuqaddh_s16 (int16_t __a, uint16_t __b)
29005 {
29006   return __builtin_aarch64_suqaddhi_ssu (__a,  __b);
29007 }
29008
29009 __extension__ extern __inline int32_t
29010 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29011 vuqadds_s32 (int32_t __a, uint32_t __b)
29012 {
29013   return __builtin_aarch64_suqaddsi_ssu (__a,  __b);
29014 }
29015
29016 __extension__ extern __inline int64_t
29017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29018 vuqaddd_s64 (int64_t __a, uint64_t __b)
29019 {
29020   return __builtin_aarch64_suqadddi_ssu (__a,  __b);
29021 }
29022
29023 #define __DEFINTERLEAVE(op, rettype, intype, funcsuffix, Q)             \
29024   __extension__ extern __inline rettype                                 \
29025   __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
29026   v ## op ## Q ## _ ## funcsuffix (intype a, intype b)                  \
29027   {                                                                     \
29028     return (rettype) {v ## op ## 1 ## Q ## _ ## funcsuffix (a, b),      \
29029                       v ## op ## 2 ## Q ## _ ## funcsuffix (a, b)};     \
29030   }
29031
29032 #define __INTERLEAVE_LIST(op)                                   \
29033   __DEFINTERLEAVE (op, float16x4x2_t, float16x4_t, f16,)        \
29034   __DEFINTERLEAVE (op, float32x2x2_t, float32x2_t, f32,)        \
29035   __DEFINTERLEAVE (op, poly8x8x2_t, poly8x8_t, p8,)             \
29036   __DEFINTERLEAVE (op, poly16x4x2_t, poly16x4_t, p16,)          \
29037   __DEFINTERLEAVE (op, int8x8x2_t, int8x8_t, s8,)               \
29038   __DEFINTERLEAVE (op, int16x4x2_t, int16x4_t, s16,)            \
29039   __DEFINTERLEAVE (op, int32x2x2_t, int32x2_t, s32,)            \
29040   __DEFINTERLEAVE (op, uint8x8x2_t, uint8x8_t, u8,)             \
29041   __DEFINTERLEAVE (op, uint16x4x2_t, uint16x4_t, u16,)          \
29042   __DEFINTERLEAVE (op, uint32x2x2_t, uint32x2_t, u32,)          \
29043   __DEFINTERLEAVE (op, float16x8x2_t, float16x8_t, f16, q)      \
29044   __DEFINTERLEAVE (op, float32x4x2_t, float32x4_t, f32, q)      \
29045   __DEFINTERLEAVE (op, poly8x16x2_t, poly8x16_t, p8, q)         \
29046   __DEFINTERLEAVE (op, poly16x8x2_t, poly16x8_t, p16, q)        \
29047   __DEFINTERLEAVE (op, int8x16x2_t, int8x16_t, s8, q)           \
29048   __DEFINTERLEAVE (op, int16x8x2_t, int16x8_t, s16, q)          \
29049   __DEFINTERLEAVE (op, int32x4x2_t, int32x4_t, s32, q)          \
29050   __DEFINTERLEAVE (op, uint8x16x2_t, uint8x16_t, u8, q)         \
29051   __DEFINTERLEAVE (op, uint16x8x2_t, uint16x8_t, u16, q)        \
29052   __DEFINTERLEAVE (op, uint32x4x2_t, uint32x4_t, u32, q)
29053
29054 /* vuzp */
29055
29056 __extension__ extern __inline float16x4_t
29057 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29058 vuzp1_f16 (float16x4_t __a, float16x4_t __b)
29059 {
29060 #ifdef __AARCH64EB__
29061   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29062 #else
29063   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29064 #endif
29065 }
29066
29067 __extension__ extern __inline float32x2_t
29068 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29069 vuzp1_f32 (float32x2_t __a, float32x2_t __b)
29070 {
29071 #ifdef __AARCH64EB__
29072   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29073 #else
29074   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29075 #endif
29076 }
29077
29078 __extension__ extern __inline poly8x8_t
29079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29080 vuzp1_p8 (poly8x8_t __a, poly8x8_t __b)
29081 {
29082 #ifdef __AARCH64EB__
29083   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29084 #else
29085   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29086 #endif
29087 }
29088
29089 __extension__ extern __inline poly16x4_t
29090 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29091 vuzp1_p16 (poly16x4_t __a, poly16x4_t __b)
29092 {
29093 #ifdef __AARCH64EB__
29094   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29095 #else
29096   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29097 #endif
29098 }
29099
29100 __extension__ extern __inline int8x8_t
29101 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29102 vuzp1_s8 (int8x8_t __a, int8x8_t __b)
29103 {
29104 #ifdef __AARCH64EB__
29105   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29106 #else
29107   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29108 #endif
29109 }
29110
29111 __extension__ extern __inline int16x4_t
29112 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29113 vuzp1_s16 (int16x4_t __a, int16x4_t __b)
29114 {
29115 #ifdef __AARCH64EB__
29116   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29117 #else
29118   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29119 #endif
29120 }
29121
29122 __extension__ extern __inline int32x2_t
29123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29124 vuzp1_s32 (int32x2_t __a, int32x2_t __b)
29125 {
29126 #ifdef __AARCH64EB__
29127   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29128 #else
29129   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29130 #endif
29131 }
29132
29133 __extension__ extern __inline uint8x8_t
29134 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29135 vuzp1_u8 (uint8x8_t __a, uint8x8_t __b)
29136 {
29137 #ifdef __AARCH64EB__
29138   return __builtin_shuffle (__a, __b, (uint8x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29139 #else
29140   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29141 #endif
29142 }
29143
29144 __extension__ extern __inline uint16x4_t
29145 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29146 vuzp1_u16 (uint16x4_t __a, uint16x4_t __b)
29147 {
29148 #ifdef __AARCH64EB__
29149   return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
29150 #else
29151   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
29152 #endif
29153 }
29154
29155 __extension__ extern __inline uint32x2_t
29156 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29157 vuzp1_u32 (uint32x2_t __a, uint32x2_t __b)
29158 {
29159 #ifdef __AARCH64EB__
29160   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29161 #else
29162   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29163 #endif
29164 }
29165
29166 __extension__ extern __inline float16x8_t
29167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29168 vuzp1q_f16 (float16x8_t __a, float16x8_t __b)
29169 {
29170 #ifdef __AARCH64EB__
29171   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29172 #else
29173   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29174 #endif
29175 }
29176
29177 __extension__ extern __inline float32x4_t
29178 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29179 vuzp1q_f32 (float32x4_t __a, float32x4_t __b)
29180 {
29181 #ifdef __AARCH64EB__
29182   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
29183 #else
29184   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
29185 #endif
29186 }
29187
29188 __extension__ extern __inline float64x2_t
29189 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29190 vuzp1q_f64 (float64x2_t __a, float64x2_t __b)
29191 {
29192 #ifdef __AARCH64EB__
29193   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29194 #else
29195   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29196 #endif
29197 }
29198
29199 __extension__ extern __inline poly8x16_t
29200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29201 vuzp1q_p8 (poly8x16_t __a, poly8x16_t __b)
29202 {
29203 #ifdef __AARCH64EB__
29204   return __builtin_shuffle (__a, __b, (uint8x16_t)
29205       {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
29206 #else
29207   return __builtin_shuffle (__a, __b, (uint8x16_t)
29208       {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
29209 #endif
29210 }
29211
29212 __extension__ extern __inline poly16x8_t
29213 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29214 vuzp1q_p16 (poly16x8_t __a, poly16x8_t __b)
29215 {
29216 #ifdef __AARCH64EB__
29217   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29218 #else
29219   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29220 #endif
29221 }
29222
29223 __extension__ extern __inline int8x16_t
29224 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29225 vuzp1q_s8 (int8x16_t __a, int8x16_t __b)
29226 {
29227 #ifdef __AARCH64EB__
29228   return __builtin_shuffle (__a, __b,
29229       (uint8x16_t) {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
29230 #else
29231   return __builtin_shuffle (__a, __b,
29232       (uint8x16_t) {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
29233 #endif
29234 }
29235
29236 __extension__ extern __inline int16x8_t
29237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29238 vuzp1q_s16 (int16x8_t __a, int16x8_t __b)
29239 {
29240 #ifdef __AARCH64EB__
29241   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29242 #else
29243   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29244 #endif
29245 }
29246
29247 __extension__ extern __inline int32x4_t
29248 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29249 vuzp1q_s32 (int32x4_t __a, int32x4_t __b)
29250 {
29251 #ifdef __AARCH64EB__
29252   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
29253 #else
29254   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
29255 #endif
29256 }
29257
29258 __extension__ extern __inline int64x2_t
29259 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29260 vuzp1q_s64 (int64x2_t __a, int64x2_t __b)
29261 {
29262 #ifdef __AARCH64EB__
29263   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29264 #else
29265   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29266 #endif
29267 }
29268
29269 __extension__ extern __inline uint8x16_t
29270 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29271 vuzp1q_u8 (uint8x16_t __a, uint8x16_t __b)
29272 {
29273 #ifdef __AARCH64EB__
29274   return __builtin_shuffle (__a, __b,
29275       (uint8x16_t) {17, 19, 21, 23, 25, 27, 29, 31, 1, 3, 5, 7, 9, 11, 13, 15});
29276 #else
29277   return __builtin_shuffle (__a, __b,
29278       (uint8x16_t) {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
29279 #endif
29280 }
29281
29282 __extension__ extern __inline uint16x8_t
29283 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29284 vuzp1q_u16 (uint16x8_t __a, uint16x8_t __b)
29285 {
29286 #ifdef __AARCH64EB__
29287   return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
29288 #else
29289   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
29290 #endif
29291 }
29292
29293 __extension__ extern __inline uint32x4_t
29294 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29295 vuzp1q_u32 (uint32x4_t __a, uint32x4_t __b)
29296 {
29297 #ifdef __AARCH64EB__
29298   return __builtin_shuffle (__a, __b, (uint32x4_t) {5, 7, 1, 3});
29299 #else
29300   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 2, 4, 6});
29301 #endif
29302 }
29303
29304 __extension__ extern __inline uint64x2_t
29305 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29306 vuzp1q_u64 (uint64x2_t __a, uint64x2_t __b)
29307 {
29308 #ifdef __AARCH64EB__
29309   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29310 #else
29311   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29312 #endif
29313 }
29314
29315 __extension__ extern __inline float16x4_t
29316 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29317 vuzp2_f16 (float16x4_t __a, float16x4_t __b)
29318 {
29319 #ifdef __AARCH64EB__
29320   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29321 #else
29322   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29323 #endif
29324 }
29325
29326 __extension__ extern __inline float32x2_t
29327 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29328 vuzp2_f32 (float32x2_t __a, float32x2_t __b)
29329 {
29330 #ifdef __AARCH64EB__
29331   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29332 #else
29333   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29334 #endif
29335 }
29336
29337 __extension__ extern __inline poly8x8_t
29338 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29339 vuzp2_p8 (poly8x8_t __a, poly8x8_t __b)
29340 {
29341 #ifdef __AARCH64EB__
29342   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29343 #else
29344   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29345 #endif
29346 }
29347
29348 __extension__ extern __inline poly16x4_t
29349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29350 vuzp2_p16 (poly16x4_t __a, poly16x4_t __b)
29351 {
29352 #ifdef __AARCH64EB__
29353   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29354 #else
29355   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29356 #endif
29357 }
29358
29359 __extension__ extern __inline int8x8_t
29360 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29361 vuzp2_s8 (int8x8_t __a, int8x8_t __b)
29362 {
29363 #ifdef __AARCH64EB__
29364   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29365 #else
29366   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29367 #endif
29368 }
29369
29370 __extension__ extern __inline int16x4_t
29371 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29372 vuzp2_s16 (int16x4_t __a, int16x4_t __b)
29373 {
29374 #ifdef __AARCH64EB__
29375   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29376 #else
29377   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29378 #endif
29379 }
29380
29381 __extension__ extern __inline int32x2_t
29382 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29383 vuzp2_s32 (int32x2_t __a, int32x2_t __b)
29384 {
29385 #ifdef __AARCH64EB__
29386   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29387 #else
29388   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29389 #endif
29390 }
29391
29392 __extension__ extern __inline uint8x8_t
29393 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29394 vuzp2_u8 (uint8x8_t __a, uint8x8_t __b)
29395 {
29396 #ifdef __AARCH64EB__
29397   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29398 #else
29399   return __builtin_shuffle (__a, __b, (uint8x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29400 #endif
29401 }
29402
29403 __extension__ extern __inline uint16x4_t
29404 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29405 vuzp2_u16 (uint16x4_t __a, uint16x4_t __b)
29406 {
29407 #ifdef __AARCH64EB__
29408   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
29409 #else
29410   return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
29411 #endif
29412 }
29413
29414 __extension__ extern __inline uint32x2_t
29415 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29416 vuzp2_u32 (uint32x2_t __a, uint32x2_t __b)
29417 {
29418 #ifdef __AARCH64EB__
29419   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29420 #else
29421   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29422 #endif
29423 }
29424
29425 __extension__ extern __inline float16x8_t
29426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29427 vuzp2q_f16 (float16x8_t __a, float16x8_t __b)
29428 {
29429 #ifdef __AARCH64EB__
29430   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29431 #else
29432   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29433 #endif
29434 }
29435
29436 __extension__ extern __inline float32x4_t
29437 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29438 vuzp2q_f32 (float32x4_t __a, float32x4_t __b)
29439 {
29440 #ifdef __AARCH64EB__
29441   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
29442 #else
29443   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
29444 #endif
29445 }
29446
29447 __extension__ extern __inline float64x2_t
29448 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29449 vuzp2q_f64 (float64x2_t __a, float64x2_t __b)
29450 {
29451 #ifdef __AARCH64EB__
29452   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29453 #else
29454   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29455 #endif
29456 }
29457
29458 __extension__ extern __inline poly8x16_t
29459 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29460 vuzp2q_p8 (poly8x16_t __a, poly8x16_t __b)
29461 {
29462 #ifdef __AARCH64EB__
29463   return __builtin_shuffle (__a, __b,
29464       (uint8x16_t) {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
29465 #else
29466   return __builtin_shuffle (__a, __b,
29467       (uint8x16_t) {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
29468 #endif
29469 }
29470
29471 __extension__ extern __inline poly16x8_t
29472 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29473 vuzp2q_p16 (poly16x8_t __a, poly16x8_t __b)
29474 {
29475 #ifdef __AARCH64EB__
29476   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29477 #else
29478   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29479 #endif
29480 }
29481
29482 __extension__ extern __inline int8x16_t
29483 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29484 vuzp2q_s8 (int8x16_t __a, int8x16_t __b)
29485 {
29486 #ifdef __AARCH64EB__
29487   return __builtin_shuffle (__a, __b,
29488       (uint8x16_t) {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
29489 #else
29490   return __builtin_shuffle (__a, __b,
29491       (uint8x16_t) {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
29492 #endif
29493 }
29494
29495 __extension__ extern __inline int16x8_t
29496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29497 vuzp2q_s16 (int16x8_t __a, int16x8_t __b)
29498 {
29499 #ifdef __AARCH64EB__
29500   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29501 #else
29502   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29503 #endif
29504 }
29505
29506 __extension__ extern __inline int32x4_t
29507 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29508 vuzp2q_s32 (int32x4_t __a, int32x4_t __b)
29509 {
29510 #ifdef __AARCH64EB__
29511   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
29512 #else
29513   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
29514 #endif
29515 }
29516
29517 __extension__ extern __inline int64x2_t
29518 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29519 vuzp2q_s64 (int64x2_t __a, int64x2_t __b)
29520 {
29521 #ifdef __AARCH64EB__
29522   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29523 #else
29524   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29525 #endif
29526 }
29527
29528 __extension__ extern __inline uint8x16_t
29529 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29530 vuzp2q_u8 (uint8x16_t __a, uint8x16_t __b)
29531 {
29532 #ifdef __AARCH64EB__
29533   return __builtin_shuffle (__a, __b, (uint8x16_t)
29534       {16, 18, 20, 22, 24, 26, 28, 30, 0, 2, 4, 6, 8, 10, 12, 14});
29535 #else
29536   return __builtin_shuffle (__a, __b, (uint8x16_t)
29537       {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31});
29538 #endif
29539 }
29540
29541 __extension__ extern __inline uint16x8_t
29542 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29543 vuzp2q_u16 (uint16x8_t __a, uint16x8_t __b)
29544 {
29545 #ifdef __AARCH64EB__
29546   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
29547 #else
29548   return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
29549 #endif
29550 }
29551
29552 __extension__ extern __inline uint32x4_t
29553 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29554 vuzp2q_u32 (uint32x4_t __a, uint32x4_t __b)
29555 {
29556 #ifdef __AARCH64EB__
29557   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 6, 0, 2});
29558 #else
29559   return __builtin_shuffle (__a, __b, (uint32x4_t) {1, 3, 5, 7});
29560 #endif
29561 }
29562
29563 __extension__ extern __inline uint64x2_t
29564 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29565 vuzp2q_u64 (uint64x2_t __a, uint64x2_t __b)
29566 {
29567 #ifdef __AARCH64EB__
29568   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29569 #else
29570   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29571 #endif
29572 }
29573
29574 __INTERLEAVE_LIST (uzp)
29575
29576 /* vzip */
29577
29578 __extension__ extern __inline float16x4_t
29579 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29580 vzip1_f16 (float16x4_t __a, float16x4_t __b)
29581 {
29582 #ifdef __AARCH64EB__
29583   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29584 #else
29585   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29586 #endif
29587 }
29588
29589 __extension__ extern __inline float32x2_t
29590 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29591 vzip1_f32 (float32x2_t __a, float32x2_t __b)
29592 {
29593 #ifdef __AARCH64EB__
29594   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29595 #else
29596   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29597 #endif
29598 }
29599
29600 __extension__ extern __inline poly8x8_t
29601 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29602 vzip1_p8 (poly8x8_t __a, poly8x8_t __b)
29603 {
29604 #ifdef __AARCH64EB__
29605   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29606 #else
29607   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29608 #endif
29609 }
29610
29611 __extension__ extern __inline poly16x4_t
29612 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29613 vzip1_p16 (poly16x4_t __a, poly16x4_t __b)
29614 {
29615 #ifdef __AARCH64EB__
29616   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29617 #else
29618   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29619 #endif
29620 }
29621
29622 __extension__ extern __inline int8x8_t
29623 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29624 vzip1_s8 (int8x8_t __a, int8x8_t __b)
29625 {
29626 #ifdef __AARCH64EB__
29627   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29628 #else
29629   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29630 #endif
29631 }
29632
29633 __extension__ extern __inline int16x4_t
29634 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29635 vzip1_s16 (int16x4_t __a, int16x4_t __b)
29636 {
29637 #ifdef __AARCH64EB__
29638   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29639 #else
29640   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29641 #endif
29642 }
29643
29644 __extension__ extern __inline int32x2_t
29645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29646 vzip1_s32 (int32x2_t __a, int32x2_t __b)
29647 {
29648 #ifdef __AARCH64EB__
29649   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29650 #else
29651   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29652 #endif
29653 }
29654
29655 __extension__ extern __inline uint8x8_t
29656 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29657 vzip1_u8 (uint8x8_t __a, uint8x8_t __b)
29658 {
29659 #ifdef __AARCH64EB__
29660   return __builtin_shuffle (__a, __b, (uint8x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29661 #else
29662   return __builtin_shuffle (__a, __b, (uint8x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29663 #endif
29664 }
29665
29666 __extension__ extern __inline uint16x4_t
29667 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29668 vzip1_u16 (uint16x4_t __a, uint16x4_t __b)
29669 {
29670 #ifdef __AARCH64EB__
29671   return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
29672 #else
29673   return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
29674 #endif
29675 }
29676
29677 __extension__ extern __inline uint32x2_t
29678 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29679 vzip1_u32 (uint32x2_t __a, uint32x2_t __b)
29680 {
29681 #ifdef __AARCH64EB__
29682   return __builtin_shuffle (__a, __b, (uint32x2_t) {3, 1});
29683 #else
29684   return __builtin_shuffle (__a, __b, (uint32x2_t) {0, 2});
29685 #endif
29686 }
29687
29688 __extension__ extern __inline float16x8_t
29689 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29690 vzip1q_f16 (float16x8_t __a, float16x8_t __b)
29691 {
29692 #ifdef __AARCH64EB__
29693   return __builtin_shuffle (__a, __b,
29694                             (uint16x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
29695 #else
29696   return __builtin_shuffle (__a, __b,
29697                             (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29698 #endif
29699 }
29700
29701 __extension__ extern __inline float32x4_t
29702 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29703 vzip1q_f32 (float32x4_t __a, float32x4_t __b)
29704 {
29705 #ifdef __AARCH64EB__
29706   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
29707 #else
29708   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
29709 #endif
29710 }
29711
29712 __extension__ extern __inline float64x2_t
29713 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29714 vzip1q_f64 (float64x2_t __a, float64x2_t __b)
29715 {
29716 #ifdef __AARCH64EB__
29717   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29718 #else
29719   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29720 #endif
29721 }
29722
29723 __extension__ extern __inline poly8x16_t
29724 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29725 vzip1q_p8 (poly8x16_t __a, poly8x16_t __b)
29726 {
29727 #ifdef __AARCH64EB__
29728   return __builtin_shuffle (__a, __b, (uint8x16_t)
29729       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
29730 #else
29731   return __builtin_shuffle (__a, __b, (uint8x16_t)
29732       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
29733 #endif
29734 }
29735
29736 __extension__ extern __inline poly16x8_t
29737 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29738 vzip1q_p16 (poly16x8_t __a, poly16x8_t __b)
29739 {
29740 #ifdef __AARCH64EB__
29741   return __builtin_shuffle (__a, __b, (uint16x8_t)
29742       {12, 4, 13, 5, 14, 6, 15, 7});
29743 #else
29744   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29745 #endif
29746 }
29747
29748 __extension__ extern __inline int8x16_t
29749 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29750 vzip1q_s8 (int8x16_t __a, int8x16_t __b)
29751 {
29752 #ifdef __AARCH64EB__
29753   return __builtin_shuffle (__a, __b, (uint8x16_t)
29754       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
29755 #else
29756   return __builtin_shuffle (__a, __b, (uint8x16_t)
29757       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
29758 #endif
29759 }
29760
29761 __extension__ extern __inline int16x8_t
29762 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29763 vzip1q_s16 (int16x8_t __a, int16x8_t __b)
29764 {
29765 #ifdef __AARCH64EB__
29766   return __builtin_shuffle (__a, __b, (uint16x8_t)
29767       {12, 4, 13, 5, 14, 6, 15, 7});
29768 #else
29769   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29770 #endif
29771 }
29772
29773 __extension__ extern __inline int32x4_t
29774 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29775 vzip1q_s32 (int32x4_t __a, int32x4_t __b)
29776 {
29777 #ifdef __AARCH64EB__
29778   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
29779 #else
29780   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
29781 #endif
29782 }
29783
29784 __extension__ extern __inline int64x2_t
29785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29786 vzip1q_s64 (int64x2_t __a, int64x2_t __b)
29787 {
29788 #ifdef __AARCH64EB__
29789   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29790 #else
29791   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29792 #endif
29793 }
29794
29795 __extension__ extern __inline uint8x16_t
29796 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29797 vzip1q_u8 (uint8x16_t __a, uint8x16_t __b)
29798 {
29799 #ifdef __AARCH64EB__
29800   return __builtin_shuffle (__a, __b, (uint8x16_t)
29801       {24, 8, 25, 9, 26, 10, 27, 11, 28, 12, 29, 13, 30, 14, 31, 15});
29802 #else
29803   return __builtin_shuffle (__a, __b, (uint8x16_t)
29804       {0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23});
29805 #endif
29806 }
29807
29808 __extension__ extern __inline uint16x8_t
29809 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29810 vzip1q_u16 (uint16x8_t __a, uint16x8_t __b)
29811 {
29812 #ifdef __AARCH64EB__
29813   return __builtin_shuffle (__a, __b, (uint16x8_t)
29814       {12, 4, 13, 5, 14, 6, 15, 7});
29815 #else
29816   return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
29817 #endif
29818 }
29819
29820 __extension__ extern __inline uint32x4_t
29821 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29822 vzip1q_u32 (uint32x4_t __a, uint32x4_t __b)
29823 {
29824 #ifdef __AARCH64EB__
29825   return __builtin_shuffle (__a, __b, (uint32x4_t) {6, 2, 7, 3});
29826 #else
29827   return __builtin_shuffle (__a, __b, (uint32x4_t) {0, 4, 1, 5});
29828 #endif
29829 }
29830
29831 __extension__ extern __inline uint64x2_t
29832 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29833 vzip1q_u64 (uint64x2_t __a, uint64x2_t __b)
29834 {
29835 #ifdef __AARCH64EB__
29836   return __builtin_shuffle (__a, __b, (uint64x2_t) {3, 1});
29837 #else
29838   return __builtin_shuffle (__a, __b, (uint64x2_t) {0, 2});
29839 #endif
29840 }
29841
29842 __extension__ extern __inline float16x4_t
29843 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29844 vzip2_f16 (float16x4_t __a, float16x4_t __b)
29845 {
29846 #ifdef __AARCH64EB__
29847   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29848 #else
29849   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29850 #endif
29851 }
29852
29853 __extension__ extern __inline float32x2_t
29854 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29855 vzip2_f32 (float32x2_t __a, float32x2_t __b)
29856 {
29857 #ifdef __AARCH64EB__
29858   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29859 #else
29860   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29861 #endif
29862 }
29863
29864 __extension__ extern __inline poly8x8_t
29865 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29866 vzip2_p8 (poly8x8_t __a, poly8x8_t __b)
29867 {
29868 #ifdef __AARCH64EB__
29869   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29870 #else
29871   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29872 #endif
29873 }
29874
29875 __extension__ extern __inline poly16x4_t
29876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29877 vzip2_p16 (poly16x4_t __a, poly16x4_t __b)
29878 {
29879 #ifdef __AARCH64EB__
29880   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29881 #else
29882   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29883 #endif
29884 }
29885
29886 __extension__ extern __inline int8x8_t
29887 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29888 vzip2_s8 (int8x8_t __a, int8x8_t __b)
29889 {
29890 #ifdef __AARCH64EB__
29891   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29892 #else
29893   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29894 #endif
29895 }
29896
29897 __extension__ extern __inline int16x4_t
29898 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29899 vzip2_s16 (int16x4_t __a, int16x4_t __b)
29900 {
29901 #ifdef __AARCH64EB__
29902   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29903 #else
29904   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29905 #endif
29906 }
29907
29908 __extension__ extern __inline int32x2_t
29909 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29910 vzip2_s32 (int32x2_t __a, int32x2_t __b)
29911 {
29912 #ifdef __AARCH64EB__
29913   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29914 #else
29915   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29916 #endif
29917 }
29918
29919 __extension__ extern __inline uint8x8_t
29920 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29921 vzip2_u8 (uint8x8_t __a, uint8x8_t __b)
29922 {
29923 #ifdef __AARCH64EB__
29924   return __builtin_shuffle (__a, __b, (uint8x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29925 #else
29926   return __builtin_shuffle (__a, __b, (uint8x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29927 #endif
29928 }
29929
29930 __extension__ extern __inline uint16x4_t
29931 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29932 vzip2_u16 (uint16x4_t __a, uint16x4_t __b)
29933 {
29934 #ifdef __AARCH64EB__
29935   return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
29936 #else
29937   return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
29938 #endif
29939 }
29940
29941 __extension__ extern __inline uint32x2_t
29942 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29943 vzip2_u32 (uint32x2_t __a, uint32x2_t __b)
29944 {
29945 #ifdef __AARCH64EB__
29946   return __builtin_shuffle (__a, __b, (uint32x2_t) {2, 0});
29947 #else
29948   return __builtin_shuffle (__a, __b, (uint32x2_t) {1, 3});
29949 #endif
29950 }
29951
29952 __extension__ extern __inline float16x8_t
29953 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29954 vzip2q_f16 (float16x8_t __a, float16x8_t __b)
29955 {
29956 #ifdef __AARCH64EB__
29957   return __builtin_shuffle (__a, __b,
29958                             (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
29959 #else
29960   return __builtin_shuffle (__a, __b,
29961                             (uint16x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
29962 #endif
29963 }
29964
29965 __extension__ extern __inline float32x4_t
29966 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29967 vzip2q_f32 (float32x4_t __a, float32x4_t __b)
29968 {
29969 #ifdef __AARCH64EB__
29970   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
29971 #else
29972   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
29973 #endif
29974 }
29975
29976 __extension__ extern __inline float64x2_t
29977 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29978 vzip2q_f64 (float64x2_t __a, float64x2_t __b)
29979 {
29980 #ifdef __AARCH64EB__
29981   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
29982 #else
29983   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
29984 #endif
29985 }
29986
29987 __extension__ extern __inline poly8x16_t
29988 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
29989 vzip2q_p8 (poly8x16_t __a, poly8x16_t __b)
29990 {
29991 #ifdef __AARCH64EB__
29992   return __builtin_shuffle (__a, __b, (uint8x16_t)
29993       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
29994 #else
29995   return __builtin_shuffle (__a, __b, (uint8x16_t)
29996       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
29997 #endif
29998 }
29999
30000 __extension__ extern __inline poly16x8_t
30001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30002 vzip2q_p16 (poly16x8_t __a, poly16x8_t __b)
30003 {
30004 #ifdef __AARCH64EB__
30005   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30006 #else
30007   return __builtin_shuffle (__a, __b, (uint16x8_t)
30008       {4, 12, 5, 13, 6, 14, 7, 15});
30009 #endif
30010 }
30011
30012 __extension__ extern __inline int8x16_t
30013 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30014 vzip2q_s8 (int8x16_t __a, int8x16_t __b)
30015 {
30016 #ifdef __AARCH64EB__
30017   return __builtin_shuffle (__a, __b, (uint8x16_t)
30018       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
30019 #else
30020   return __builtin_shuffle (__a, __b, (uint8x16_t)
30021       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
30022 #endif
30023 }
30024
30025 __extension__ extern __inline int16x8_t
30026 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30027 vzip2q_s16 (int16x8_t __a, int16x8_t __b)
30028 {
30029 #ifdef __AARCH64EB__
30030   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30031 #else
30032   return __builtin_shuffle (__a, __b, (uint16x8_t)
30033       {4, 12, 5, 13, 6, 14, 7, 15});
30034 #endif
30035 }
30036
30037 __extension__ extern __inline int32x4_t
30038 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30039 vzip2q_s32 (int32x4_t __a, int32x4_t __b)
30040 {
30041 #ifdef __AARCH64EB__
30042   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
30043 #else
30044   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
30045 #endif
30046 }
30047
30048 __extension__ extern __inline int64x2_t
30049 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30050 vzip2q_s64 (int64x2_t __a, int64x2_t __b)
30051 {
30052 #ifdef __AARCH64EB__
30053   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30054 #else
30055   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30056 #endif
30057 }
30058
30059 __extension__ extern __inline uint8x16_t
30060 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30061 vzip2q_u8 (uint8x16_t __a, uint8x16_t __b)
30062 {
30063 #ifdef __AARCH64EB__
30064   return __builtin_shuffle (__a, __b, (uint8x16_t)
30065       {16, 0, 17, 1, 18, 2, 19, 3, 20, 4, 21, 5, 22, 6, 23, 7});
30066 #else
30067   return __builtin_shuffle (__a, __b, (uint8x16_t)
30068       {8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31});
30069 #endif
30070 }
30071
30072 __extension__ extern __inline uint16x8_t
30073 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30074 vzip2q_u16 (uint16x8_t __a, uint16x8_t __b)
30075 {
30076 #ifdef __AARCH64EB__
30077   return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
30078 #else
30079   return __builtin_shuffle (__a, __b, (uint16x8_t)
30080       {4, 12, 5, 13, 6, 14, 7, 15});
30081 #endif
30082 }
30083
30084 __extension__ extern __inline uint32x4_t
30085 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30086 vzip2q_u32 (uint32x4_t __a, uint32x4_t __b)
30087 {
30088 #ifdef __AARCH64EB__
30089   return __builtin_shuffle (__a, __b, (uint32x4_t) {4, 0, 5, 1});
30090 #else
30091   return __builtin_shuffle (__a, __b, (uint32x4_t) {2, 6, 3, 7});
30092 #endif
30093 }
30094
30095 __extension__ extern __inline uint64x2_t
30096 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30097 vzip2q_u64 (uint64x2_t __a, uint64x2_t __b)
30098 {
30099 #ifdef __AARCH64EB__
30100   return __builtin_shuffle (__a, __b, (uint64x2_t) {2, 0});
30101 #else
30102   return __builtin_shuffle (__a, __b, (uint64x2_t) {1, 3});
30103 #endif
30104 }
30105
30106 __INTERLEAVE_LIST (zip)
30107
30108 #undef __INTERLEAVE_LIST
30109 #undef __DEFINTERLEAVE
30110
30111 /* End of optimal implementations in approved order.  */
30112
30113 #pragma GCC pop_options
30114
30115 /* ARMv8.2-A FP16 intrinsics.  */
30116
30117 #include "arm_fp16.h"
30118
30119 #pragma GCC push_options
30120 #pragma GCC target ("arch=armv8.2-a+fp16")
30121
30122 /* ARMv8.2-A FP16 one operand vector intrinsics.  */
30123
30124 __extension__ extern __inline float16x4_t
30125 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30126 vabs_f16 (float16x4_t __a)
30127 {
30128   return __builtin_aarch64_absv4hf (__a);
30129 }
30130
30131 __extension__ extern __inline float16x8_t
30132 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30133 vabsq_f16 (float16x8_t __a)
30134 {
30135   return __builtin_aarch64_absv8hf (__a);
30136 }
30137
30138 __extension__ extern __inline uint16x4_t
30139 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30140 vceqz_f16 (float16x4_t __a)
30141 {
30142   return __builtin_aarch64_cmeqv4hf_uss (__a, vdup_n_f16 (0.0f));
30143 }
30144
30145 __extension__ extern __inline uint16x8_t
30146 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30147 vceqzq_f16 (float16x8_t __a)
30148 {
30149   return __builtin_aarch64_cmeqv8hf_uss (__a, vdupq_n_f16 (0.0f));
30150 }
30151
30152 __extension__ extern __inline uint16x4_t
30153 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30154 vcgez_f16 (float16x4_t __a)
30155 {
30156   return __builtin_aarch64_cmgev4hf_uss (__a, vdup_n_f16 (0.0f));
30157 }
30158
30159 __extension__ extern __inline uint16x8_t
30160 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30161 vcgezq_f16 (float16x8_t __a)
30162 {
30163   return __builtin_aarch64_cmgev8hf_uss (__a, vdupq_n_f16 (0.0f));
30164 }
30165
30166 __extension__ extern __inline uint16x4_t
30167 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30168 vcgtz_f16 (float16x4_t __a)
30169 {
30170   return __builtin_aarch64_cmgtv4hf_uss (__a, vdup_n_f16 (0.0f));
30171 }
30172
30173 __extension__ extern __inline uint16x8_t
30174 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30175 vcgtzq_f16 (float16x8_t __a)
30176 {
30177   return __builtin_aarch64_cmgtv8hf_uss (__a, vdupq_n_f16 (0.0f));
30178 }
30179
30180 __extension__ extern __inline uint16x4_t
30181 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30182 vclez_f16 (float16x4_t __a)
30183 {
30184   return __builtin_aarch64_cmlev4hf_uss (__a, vdup_n_f16 (0.0f));
30185 }
30186
30187 __extension__ extern __inline uint16x8_t
30188 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30189 vclezq_f16 (float16x8_t __a)
30190 {
30191   return __builtin_aarch64_cmlev8hf_uss (__a, vdupq_n_f16 (0.0f));
30192 }
30193
30194 __extension__ extern __inline uint16x4_t
30195 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30196 vcltz_f16 (float16x4_t __a)
30197 {
30198   return __builtin_aarch64_cmltv4hf_uss (__a, vdup_n_f16 (0.0f));
30199 }
30200
30201 __extension__ extern __inline uint16x8_t
30202 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30203 vcltzq_f16 (float16x8_t __a)
30204 {
30205   return __builtin_aarch64_cmltv8hf_uss (__a, vdupq_n_f16 (0.0f));
30206 }
30207
30208 __extension__ extern __inline float16x4_t
30209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30210 vcvt_f16_s16 (int16x4_t __a)
30211 {
30212   return __builtin_aarch64_floatv4hiv4hf (__a);
30213 }
30214
30215 __extension__ extern __inline float16x8_t
30216 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30217 vcvtq_f16_s16 (int16x8_t __a)
30218 {
30219   return __builtin_aarch64_floatv8hiv8hf (__a);
30220 }
30221
30222 __extension__ extern __inline float16x4_t
30223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30224 vcvt_f16_u16 (uint16x4_t __a)
30225 {
30226   return __builtin_aarch64_floatunsv4hiv4hf ((int16x4_t) __a);
30227 }
30228
30229 __extension__ extern __inline float16x8_t
30230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30231 vcvtq_f16_u16 (uint16x8_t __a)
30232 {
30233   return __builtin_aarch64_floatunsv8hiv8hf ((int16x8_t) __a);
30234 }
30235
30236 __extension__ extern __inline int16x4_t
30237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30238 vcvt_s16_f16 (float16x4_t __a)
30239 {
30240   return __builtin_aarch64_lbtruncv4hfv4hi (__a);
30241 }
30242
30243 __extension__ extern __inline int16x8_t
30244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30245 vcvtq_s16_f16 (float16x8_t __a)
30246 {
30247   return __builtin_aarch64_lbtruncv8hfv8hi (__a);
30248 }
30249
30250 __extension__ extern __inline uint16x4_t
30251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30252 vcvt_u16_f16 (float16x4_t __a)
30253 {
30254   return __builtin_aarch64_lbtruncuv4hfv4hi_us (__a);
30255 }
30256
30257 __extension__ extern __inline uint16x8_t
30258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30259 vcvtq_u16_f16 (float16x8_t __a)
30260 {
30261   return __builtin_aarch64_lbtruncuv8hfv8hi_us (__a);
30262 }
30263
30264 __extension__ extern __inline int16x4_t
30265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30266 vcvta_s16_f16 (float16x4_t __a)
30267 {
30268   return __builtin_aarch64_lroundv4hfv4hi (__a);
30269 }
30270
30271 __extension__ extern __inline int16x8_t
30272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30273 vcvtaq_s16_f16 (float16x8_t __a)
30274 {
30275   return __builtin_aarch64_lroundv8hfv8hi (__a);
30276 }
30277
30278 __extension__ extern __inline uint16x4_t
30279 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30280 vcvta_u16_f16 (float16x4_t __a)
30281 {
30282   return __builtin_aarch64_lrounduv4hfv4hi_us (__a);
30283 }
30284
30285 __extension__ extern __inline uint16x8_t
30286 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30287 vcvtaq_u16_f16 (float16x8_t __a)
30288 {
30289   return __builtin_aarch64_lrounduv8hfv8hi_us (__a);
30290 }
30291
30292 __extension__ extern __inline int16x4_t
30293 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30294 vcvtm_s16_f16 (float16x4_t __a)
30295 {
30296   return __builtin_aarch64_lfloorv4hfv4hi (__a);
30297 }
30298
30299 __extension__ extern __inline int16x8_t
30300 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30301 vcvtmq_s16_f16 (float16x8_t __a)
30302 {
30303   return __builtin_aarch64_lfloorv8hfv8hi (__a);
30304 }
30305
30306 __extension__ extern __inline uint16x4_t
30307 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30308 vcvtm_u16_f16 (float16x4_t __a)
30309 {
30310   return __builtin_aarch64_lflooruv4hfv4hi_us (__a);
30311 }
30312
30313 __extension__ extern __inline uint16x8_t
30314 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30315 vcvtmq_u16_f16 (float16x8_t __a)
30316 {
30317   return __builtin_aarch64_lflooruv8hfv8hi_us (__a);
30318 }
30319
30320 __extension__ extern __inline int16x4_t
30321 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30322 vcvtn_s16_f16 (float16x4_t __a)
30323 {
30324   return __builtin_aarch64_lfrintnv4hfv4hi (__a);
30325 }
30326
30327 __extension__ extern __inline int16x8_t
30328 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30329 vcvtnq_s16_f16 (float16x8_t __a)
30330 {
30331   return __builtin_aarch64_lfrintnv8hfv8hi (__a);
30332 }
30333
30334 __extension__ extern __inline uint16x4_t
30335 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30336 vcvtn_u16_f16 (float16x4_t __a)
30337 {
30338   return __builtin_aarch64_lfrintnuv4hfv4hi_us (__a);
30339 }
30340
30341 __extension__ extern __inline uint16x8_t
30342 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30343 vcvtnq_u16_f16 (float16x8_t __a)
30344 {
30345   return __builtin_aarch64_lfrintnuv8hfv8hi_us (__a);
30346 }
30347
30348 __extension__ extern __inline int16x4_t
30349 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30350 vcvtp_s16_f16 (float16x4_t __a)
30351 {
30352   return __builtin_aarch64_lceilv4hfv4hi (__a);
30353 }
30354
30355 __extension__ extern __inline int16x8_t
30356 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30357 vcvtpq_s16_f16 (float16x8_t __a)
30358 {
30359   return __builtin_aarch64_lceilv8hfv8hi (__a);
30360 }
30361
30362 __extension__ extern __inline uint16x4_t
30363 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30364 vcvtp_u16_f16 (float16x4_t __a)
30365 {
30366   return __builtin_aarch64_lceiluv4hfv4hi_us (__a);
30367 }
30368
30369 __extension__ extern __inline uint16x8_t
30370 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30371 vcvtpq_u16_f16 (float16x8_t __a)
30372 {
30373   return __builtin_aarch64_lceiluv8hfv8hi_us (__a);
30374 }
30375
30376 __extension__ extern __inline float16x4_t
30377 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30378 vneg_f16 (float16x4_t __a)
30379 {
30380   return -__a;
30381 }
30382
30383 __extension__ extern __inline float16x8_t
30384 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30385 vnegq_f16 (float16x8_t __a)
30386 {
30387   return -__a;
30388 }
30389
30390 __extension__ extern __inline float16x4_t
30391 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30392 vrecpe_f16 (float16x4_t __a)
30393 {
30394   return __builtin_aarch64_frecpev4hf (__a);
30395 }
30396
30397 __extension__ extern __inline float16x8_t
30398 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30399 vrecpeq_f16 (float16x8_t __a)
30400 {
30401   return __builtin_aarch64_frecpev8hf (__a);
30402 }
30403
30404 __extension__ extern __inline float16x4_t
30405 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30406 vrnd_f16 (float16x4_t __a)
30407 {
30408   return __builtin_aarch64_btruncv4hf (__a);
30409 }
30410
30411 __extension__ extern __inline float16x8_t
30412 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30413 vrndq_f16 (float16x8_t __a)
30414 {
30415   return __builtin_aarch64_btruncv8hf (__a);
30416 }
30417
30418 __extension__ extern __inline float16x4_t
30419 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30420 vrnda_f16 (float16x4_t __a)
30421 {
30422   return __builtin_aarch64_roundv4hf (__a);
30423 }
30424
30425 __extension__ extern __inline float16x8_t
30426 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30427 vrndaq_f16 (float16x8_t __a)
30428 {
30429   return __builtin_aarch64_roundv8hf (__a);
30430 }
30431
30432 __extension__ extern __inline float16x4_t
30433 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30434 vrndi_f16 (float16x4_t __a)
30435 {
30436   return __builtin_aarch64_nearbyintv4hf (__a);
30437 }
30438
30439 __extension__ extern __inline float16x8_t
30440 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30441 vrndiq_f16 (float16x8_t __a)
30442 {
30443   return __builtin_aarch64_nearbyintv8hf (__a);
30444 }
30445
30446 __extension__ extern __inline float16x4_t
30447 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30448 vrndm_f16 (float16x4_t __a)
30449 {
30450   return __builtin_aarch64_floorv4hf (__a);
30451 }
30452
30453 __extension__ extern __inline float16x8_t
30454 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30455 vrndmq_f16 (float16x8_t __a)
30456 {
30457   return __builtin_aarch64_floorv8hf (__a);
30458 }
30459
30460 __extension__ extern __inline float16x4_t
30461 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30462 vrndn_f16 (float16x4_t __a)
30463 {
30464   return __builtin_aarch64_frintnv4hf (__a);
30465 }
30466
30467 __extension__ extern __inline float16x8_t
30468 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30469 vrndnq_f16 (float16x8_t __a)
30470 {
30471   return __builtin_aarch64_frintnv8hf (__a);
30472 }
30473
30474 __extension__ extern __inline float16x4_t
30475 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30476 vrndp_f16 (float16x4_t __a)
30477 {
30478   return __builtin_aarch64_ceilv4hf (__a);
30479 }
30480
30481 __extension__ extern __inline float16x8_t
30482 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30483 vrndpq_f16 (float16x8_t __a)
30484 {
30485   return __builtin_aarch64_ceilv8hf (__a);
30486 }
30487
30488 __extension__ extern __inline float16x4_t
30489 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30490 vrndx_f16 (float16x4_t __a)
30491 {
30492   return __builtin_aarch64_rintv4hf (__a);
30493 }
30494
30495 __extension__ extern __inline float16x8_t
30496 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30497 vrndxq_f16 (float16x8_t __a)
30498 {
30499   return __builtin_aarch64_rintv8hf (__a);
30500 }
30501
30502 __extension__ extern __inline float16x4_t
30503 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30504 vrsqrte_f16 (float16x4_t a)
30505 {
30506   return __builtin_aarch64_rsqrtev4hf (a);
30507 }
30508
30509 __extension__ extern __inline float16x8_t
30510 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30511 vrsqrteq_f16 (float16x8_t a)
30512 {
30513   return __builtin_aarch64_rsqrtev8hf (a);
30514 }
30515
30516 __extension__ extern __inline float16x4_t
30517 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30518 vsqrt_f16 (float16x4_t a)
30519 {
30520   return __builtin_aarch64_sqrtv4hf (a);
30521 }
30522
30523 __extension__ extern __inline float16x8_t
30524 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30525 vsqrtq_f16 (float16x8_t a)
30526 {
30527   return __builtin_aarch64_sqrtv8hf (a);
30528 }
30529
30530 /* ARMv8.2-A FP16 two operands vector intrinsics.  */
30531
30532 __extension__ extern __inline float16x4_t
30533 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30534 vadd_f16 (float16x4_t __a, float16x4_t __b)
30535 {
30536   return __a + __b;
30537 }
30538
30539 __extension__ extern __inline float16x8_t
30540 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30541 vaddq_f16 (float16x8_t __a, float16x8_t __b)
30542 {
30543   return __a + __b;
30544 }
30545
30546 __extension__ extern __inline float16x4_t
30547 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30548 vabd_f16 (float16x4_t a, float16x4_t b)
30549 {
30550   return __builtin_aarch64_fabdv4hf (a, b);
30551 }
30552
30553 __extension__ extern __inline float16x8_t
30554 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30555 vabdq_f16 (float16x8_t a, float16x8_t b)
30556 {
30557   return __builtin_aarch64_fabdv8hf (a, b);
30558 }
30559
30560 __extension__ extern __inline uint16x4_t
30561 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30562 vcage_f16 (float16x4_t __a, float16x4_t __b)
30563 {
30564   return __builtin_aarch64_facgev4hf_uss (__a, __b);
30565 }
30566
30567 __extension__ extern __inline uint16x8_t
30568 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30569 vcageq_f16 (float16x8_t __a, float16x8_t __b)
30570 {
30571   return __builtin_aarch64_facgev8hf_uss (__a, __b);
30572 }
30573
30574 __extension__ extern __inline uint16x4_t
30575 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30576 vcagt_f16 (float16x4_t __a, float16x4_t __b)
30577 {
30578   return __builtin_aarch64_facgtv4hf_uss (__a, __b);
30579 }
30580
30581 __extension__ extern __inline uint16x8_t
30582 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30583 vcagtq_f16 (float16x8_t __a, float16x8_t __b)
30584 {
30585   return __builtin_aarch64_facgtv8hf_uss (__a, __b);
30586 }
30587
30588 __extension__ extern __inline uint16x4_t
30589 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30590 vcale_f16 (float16x4_t __a, float16x4_t __b)
30591 {
30592   return __builtin_aarch64_faclev4hf_uss (__a, __b);
30593 }
30594
30595 __extension__ extern __inline uint16x8_t
30596 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30597 vcaleq_f16 (float16x8_t __a, float16x8_t __b)
30598 {
30599   return __builtin_aarch64_faclev8hf_uss (__a, __b);
30600 }
30601
30602 __extension__ extern __inline uint16x4_t
30603 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30604 vcalt_f16 (float16x4_t __a, float16x4_t __b)
30605 {
30606   return __builtin_aarch64_facltv4hf_uss (__a, __b);
30607 }
30608
30609 __extension__ extern __inline uint16x8_t
30610 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30611 vcaltq_f16 (float16x8_t __a, float16x8_t __b)
30612 {
30613   return __builtin_aarch64_facltv8hf_uss (__a, __b);
30614 }
30615
30616 __extension__ extern __inline uint16x4_t
30617 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30618 vceq_f16 (float16x4_t __a, float16x4_t __b)
30619 {
30620   return __builtin_aarch64_cmeqv4hf_uss (__a, __b);
30621 }
30622
30623 __extension__ extern __inline uint16x8_t
30624 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30625 vceqq_f16 (float16x8_t __a, float16x8_t __b)
30626 {
30627   return __builtin_aarch64_cmeqv8hf_uss (__a, __b);
30628 }
30629
30630 __extension__ extern __inline uint16x4_t
30631 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30632 vcge_f16 (float16x4_t __a, float16x4_t __b)
30633 {
30634   return __builtin_aarch64_cmgev4hf_uss (__a, __b);
30635 }
30636
30637 __extension__ extern __inline uint16x8_t
30638 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30639 vcgeq_f16 (float16x8_t __a, float16x8_t __b)
30640 {
30641   return __builtin_aarch64_cmgev8hf_uss (__a, __b);
30642 }
30643
30644 __extension__ extern __inline uint16x4_t
30645 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30646 vcgt_f16 (float16x4_t __a, float16x4_t __b)
30647 {
30648   return __builtin_aarch64_cmgtv4hf_uss (__a, __b);
30649 }
30650
30651 __extension__ extern __inline uint16x8_t
30652 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30653 vcgtq_f16 (float16x8_t __a, float16x8_t __b)
30654 {
30655   return __builtin_aarch64_cmgtv8hf_uss (__a, __b);
30656 }
30657
30658 __extension__ extern __inline uint16x4_t
30659 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30660 vcle_f16 (float16x4_t __a, float16x4_t __b)
30661 {
30662   return __builtin_aarch64_cmlev4hf_uss (__a, __b);
30663 }
30664
30665 __extension__ extern __inline uint16x8_t
30666 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30667 vcleq_f16 (float16x8_t __a, float16x8_t __b)
30668 {
30669   return __builtin_aarch64_cmlev8hf_uss (__a, __b);
30670 }
30671
30672 __extension__ extern __inline uint16x4_t
30673 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30674 vclt_f16 (float16x4_t __a, float16x4_t __b)
30675 {
30676   return __builtin_aarch64_cmltv4hf_uss (__a, __b);
30677 }
30678
30679 __extension__ extern __inline uint16x8_t
30680 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30681 vcltq_f16 (float16x8_t __a, float16x8_t __b)
30682 {
30683   return __builtin_aarch64_cmltv8hf_uss (__a, __b);
30684 }
30685
30686 __extension__ extern __inline float16x4_t
30687 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30688 vcvt_n_f16_s16 (int16x4_t __a, const int __b)
30689 {
30690   return __builtin_aarch64_scvtfv4hi (__a, __b);
30691 }
30692
30693 __extension__ extern __inline float16x8_t
30694 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30695 vcvtq_n_f16_s16 (int16x8_t __a, const int __b)
30696 {
30697   return __builtin_aarch64_scvtfv8hi (__a, __b);
30698 }
30699
30700 __extension__ extern __inline float16x4_t
30701 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30702 vcvt_n_f16_u16 (uint16x4_t __a, const int __b)
30703 {
30704   return __builtin_aarch64_ucvtfv4hi_sus (__a, __b);
30705 }
30706
30707 __extension__ extern __inline float16x8_t
30708 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30709 vcvtq_n_f16_u16 (uint16x8_t __a, const int __b)
30710 {
30711   return __builtin_aarch64_ucvtfv8hi_sus (__a, __b);
30712 }
30713
30714 __extension__ extern __inline int16x4_t
30715 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30716 vcvt_n_s16_f16 (float16x4_t __a, const int __b)
30717 {
30718   return __builtin_aarch64_fcvtzsv4hf (__a, __b);
30719 }
30720
30721 __extension__ extern __inline int16x8_t
30722 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30723 vcvtq_n_s16_f16 (float16x8_t __a, const int __b)
30724 {
30725   return __builtin_aarch64_fcvtzsv8hf (__a, __b);
30726 }
30727
30728 __extension__ extern __inline uint16x4_t
30729 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30730 vcvt_n_u16_f16 (float16x4_t __a, const int __b)
30731 {
30732   return __builtin_aarch64_fcvtzuv4hf_uss (__a, __b);
30733 }
30734
30735 __extension__ extern __inline uint16x8_t
30736 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30737 vcvtq_n_u16_f16 (float16x8_t __a, const int __b)
30738 {
30739   return __builtin_aarch64_fcvtzuv8hf_uss (__a, __b);
30740 }
30741
30742 __extension__ extern __inline float16x4_t
30743 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30744 vdiv_f16 (float16x4_t __a, float16x4_t __b)
30745 {
30746   return __a / __b;
30747 }
30748
30749 __extension__ extern __inline float16x8_t
30750 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30751 vdivq_f16 (float16x8_t __a, float16x8_t __b)
30752 {
30753   return __a / __b;
30754 }
30755
30756 __extension__ extern __inline float16x4_t
30757 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30758 vmax_f16 (float16x4_t __a, float16x4_t __b)
30759 {
30760   return __builtin_aarch64_smax_nanv4hf (__a, __b);
30761 }
30762
30763 __extension__ extern __inline float16x8_t
30764 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30765 vmaxq_f16 (float16x8_t __a, float16x8_t __b)
30766 {
30767   return __builtin_aarch64_smax_nanv8hf (__a, __b);
30768 }
30769
30770 __extension__ extern __inline float16x4_t
30771 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30772 vmaxnm_f16 (float16x4_t __a, float16x4_t __b)
30773 {
30774   return __builtin_aarch64_fmaxv4hf (__a, __b);
30775 }
30776
30777 __extension__ extern __inline float16x8_t
30778 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30779 vmaxnmq_f16 (float16x8_t __a, float16x8_t __b)
30780 {
30781   return __builtin_aarch64_fmaxv8hf (__a, __b);
30782 }
30783
30784 __extension__ extern __inline float16x4_t
30785 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30786 vmin_f16 (float16x4_t __a, float16x4_t __b)
30787 {
30788   return __builtin_aarch64_smin_nanv4hf (__a, __b);
30789 }
30790
30791 __extension__ extern __inline float16x8_t
30792 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30793 vminq_f16 (float16x8_t __a, float16x8_t __b)
30794 {
30795   return __builtin_aarch64_smin_nanv8hf (__a, __b);
30796 }
30797
30798 __extension__ extern __inline float16x4_t
30799 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30800 vminnm_f16 (float16x4_t __a, float16x4_t __b)
30801 {
30802   return __builtin_aarch64_fminv4hf (__a, __b);
30803 }
30804
30805 __extension__ extern __inline float16x8_t
30806 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30807 vminnmq_f16 (float16x8_t __a, float16x8_t __b)
30808 {
30809   return __builtin_aarch64_fminv8hf (__a, __b);
30810 }
30811
30812 __extension__ extern __inline float16x4_t
30813 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30814 vmul_f16 (float16x4_t __a, float16x4_t __b)
30815 {
30816   return __a * __b;
30817 }
30818
30819 __extension__ extern __inline float16x8_t
30820 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30821 vmulq_f16 (float16x8_t __a, float16x8_t __b)
30822 {
30823   return __a * __b;
30824 }
30825
30826 __extension__ extern __inline float16x4_t
30827 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30828 vmulx_f16 (float16x4_t __a, float16x4_t __b)
30829 {
30830   return __builtin_aarch64_fmulxv4hf (__a, __b);
30831 }
30832
30833 __extension__ extern __inline float16x8_t
30834 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30835 vmulxq_f16 (float16x8_t __a, float16x8_t __b)
30836 {
30837   return __builtin_aarch64_fmulxv8hf (__a, __b);
30838 }
30839
30840 __extension__ extern __inline float16x4_t
30841 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30842 vpadd_f16 (float16x4_t a, float16x4_t b)
30843 {
30844   return __builtin_aarch64_faddpv4hf (a, b);
30845 }
30846
30847 __extension__ extern __inline float16x8_t
30848 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30849 vpaddq_f16 (float16x8_t a, float16x8_t b)
30850 {
30851   return __builtin_aarch64_faddpv8hf (a, b);
30852 }
30853
30854 __extension__ extern __inline float16x4_t
30855 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30856 vpmax_f16 (float16x4_t a, float16x4_t b)
30857 {
30858   return __builtin_aarch64_smax_nanpv4hf (a, b);
30859 }
30860
30861 __extension__ extern __inline float16x8_t
30862 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30863 vpmaxq_f16 (float16x8_t a, float16x8_t b)
30864 {
30865   return __builtin_aarch64_smax_nanpv8hf (a, b);
30866 }
30867
30868 __extension__ extern __inline float16x4_t
30869 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30870 vpmaxnm_f16 (float16x4_t a, float16x4_t b)
30871 {
30872   return __builtin_aarch64_smaxpv4hf (a, b);
30873 }
30874
30875 __extension__ extern __inline float16x8_t
30876 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30877 vpmaxnmq_f16 (float16x8_t a, float16x8_t b)
30878 {
30879   return __builtin_aarch64_smaxpv8hf (a, b);
30880 }
30881
30882 __extension__ extern __inline float16x4_t
30883 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30884 vpmin_f16 (float16x4_t a, float16x4_t b)
30885 {
30886   return __builtin_aarch64_smin_nanpv4hf (a, b);
30887 }
30888
30889 __extension__ extern __inline float16x8_t
30890 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30891 vpminq_f16 (float16x8_t a, float16x8_t b)
30892 {
30893   return __builtin_aarch64_smin_nanpv8hf (a, b);
30894 }
30895
30896 __extension__ extern __inline float16x4_t
30897 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30898 vpminnm_f16 (float16x4_t a, float16x4_t b)
30899 {
30900   return __builtin_aarch64_sminpv4hf (a, b);
30901 }
30902
30903 __extension__ extern __inline float16x8_t
30904 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30905 vpminnmq_f16 (float16x8_t a, float16x8_t b)
30906 {
30907   return __builtin_aarch64_sminpv8hf (a, b);
30908 }
30909
30910 __extension__ extern __inline float16x4_t
30911 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30912 vrecps_f16 (float16x4_t __a, float16x4_t __b)
30913 {
30914   return __builtin_aarch64_frecpsv4hf (__a, __b);
30915 }
30916
30917 __extension__ extern __inline float16x8_t
30918 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30919 vrecpsq_f16 (float16x8_t __a, float16x8_t __b)
30920 {
30921   return __builtin_aarch64_frecpsv8hf (__a, __b);
30922 }
30923
30924 __extension__ extern __inline float16x4_t
30925 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30926 vrsqrts_f16 (float16x4_t a, float16x4_t b)
30927 {
30928   return __builtin_aarch64_rsqrtsv4hf (a, b);
30929 }
30930
30931 __extension__ extern __inline float16x8_t
30932 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30933 vrsqrtsq_f16 (float16x8_t a, float16x8_t b)
30934 {
30935   return __builtin_aarch64_rsqrtsv8hf (a, b);
30936 }
30937
30938 __extension__ extern __inline float16x4_t
30939 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30940 vsub_f16 (float16x4_t __a, float16x4_t __b)
30941 {
30942   return __a - __b;
30943 }
30944
30945 __extension__ extern __inline float16x8_t
30946 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30947 vsubq_f16 (float16x8_t __a, float16x8_t __b)
30948 {
30949   return __a - __b;
30950 }
30951
30952 /* ARMv8.2-A FP16 three operands vector intrinsics.  */
30953
30954 __extension__ extern __inline float16x4_t
30955 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30956 vfma_f16 (float16x4_t __a, float16x4_t __b, float16x4_t __c)
30957 {
30958   return __builtin_aarch64_fmav4hf (__b, __c, __a);
30959 }
30960
30961 __extension__ extern __inline float16x8_t
30962 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30963 vfmaq_f16 (float16x8_t __a, float16x8_t __b, float16x8_t __c)
30964 {
30965   return __builtin_aarch64_fmav8hf (__b, __c, __a);
30966 }
30967
30968 __extension__ extern __inline float16x4_t
30969 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30970 vfms_f16 (float16x4_t __a, float16x4_t __b, float16x4_t __c)
30971 {
30972   return __builtin_aarch64_fnmav4hf (__b, __c, __a);
30973 }
30974
30975 __extension__ extern __inline float16x8_t
30976 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30977 vfmsq_f16 (float16x8_t __a, float16x8_t __b, float16x8_t __c)
30978 {
30979   return __builtin_aarch64_fnmav8hf (__b, __c, __a);
30980 }
30981
30982 /* ARMv8.2-A FP16 lane vector intrinsics.  */
30983
30984 __extension__ extern __inline float16_t
30985 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30986 vfmah_lane_f16 (float16_t __a, float16_t __b,
30987                 float16x4_t __c, const int __lane)
30988 {
30989   return vfmah_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
30990 }
30991
30992 __extension__ extern __inline float16_t
30993 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
30994 vfmah_laneq_f16 (float16_t __a, float16_t __b,
30995                  float16x8_t __c, const int __lane)
30996 {
30997   return vfmah_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
30998 }
30999
31000 __extension__ extern __inline float16x4_t
31001 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31002 vfma_lane_f16 (float16x4_t __a, float16x4_t __b,
31003                float16x4_t __c, const int __lane)
31004 {
31005   return vfma_f16 (__a, __b, __aarch64_vdup_lane_f16 (__c, __lane));
31006 }
31007
31008 __extension__ extern __inline float16x8_t
31009 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31010 vfmaq_lane_f16 (float16x8_t __a, float16x8_t __b,
31011                 float16x4_t __c, const int __lane)
31012 {
31013   return vfmaq_f16 (__a, __b, __aarch64_vdupq_lane_f16 (__c, __lane));
31014 }
31015
31016 __extension__ extern __inline float16x4_t
31017 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31018 vfma_laneq_f16 (float16x4_t __a, float16x4_t __b,
31019                 float16x8_t __c, const int __lane)
31020 {
31021   return vfma_f16 (__a, __b, __aarch64_vdup_laneq_f16 (__c, __lane));
31022 }
31023
31024 __extension__ extern __inline float16x8_t
31025 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31026 vfmaq_laneq_f16 (float16x8_t __a, float16x8_t __b,
31027                  float16x8_t __c, const int __lane)
31028 {
31029   return vfmaq_f16 (__a, __b, __aarch64_vdupq_laneq_f16 (__c, __lane));
31030 }
31031
31032 __extension__ extern __inline float16x4_t
31033 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31034 vfma_n_f16 (float16x4_t __a, float16x4_t __b, float16_t __c)
31035 {
31036   return vfma_f16 (__a, __b, vdup_n_f16 (__c));
31037 }
31038
31039 __extension__ extern __inline float16x8_t
31040 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31041 vfmaq_n_f16 (float16x8_t __a, float16x8_t __b, float16_t __c)
31042 {
31043   return vfmaq_f16 (__a, __b, vdupq_n_f16 (__c));
31044 }
31045
31046 __extension__ extern __inline float16_t
31047 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31048 vfmsh_lane_f16 (float16_t __a, float16_t __b,
31049                 float16x4_t __c, const int __lane)
31050 {
31051   return vfmsh_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
31052 }
31053
31054 __extension__ extern __inline float16_t
31055 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31056 vfmsh_laneq_f16 (float16_t __a, float16_t __b,
31057                  float16x8_t __c, const int __lane)
31058 {
31059   return vfmsh_f16 (__a, __b, __aarch64_vget_lane_any (__c, __lane));
31060 }
31061
31062 __extension__ extern __inline float16x4_t
31063 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31064 vfms_lane_f16 (float16x4_t __a, float16x4_t __b,
31065                float16x4_t __c, const int __lane)
31066 {
31067   return vfms_f16 (__a, __b, __aarch64_vdup_lane_f16 (__c, __lane));
31068 }
31069
31070 __extension__ extern __inline float16x8_t
31071 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31072 vfmsq_lane_f16 (float16x8_t __a, float16x8_t __b,
31073                 float16x4_t __c, const int __lane)
31074 {
31075   return vfmsq_f16 (__a, __b, __aarch64_vdupq_lane_f16 (__c, __lane));
31076 }
31077
31078 __extension__ extern __inline float16x4_t
31079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31080 vfms_laneq_f16 (float16x4_t __a, float16x4_t __b,
31081                 float16x8_t __c, const int __lane)
31082 {
31083   return vfms_f16 (__a, __b, __aarch64_vdup_laneq_f16 (__c, __lane));
31084 }
31085
31086 __extension__ extern __inline float16x8_t
31087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31088 vfmsq_laneq_f16 (float16x8_t __a, float16x8_t __b,
31089                  float16x8_t __c, const int __lane)
31090 {
31091   return vfmsq_f16 (__a, __b, __aarch64_vdupq_laneq_f16 (__c, __lane));
31092 }
31093
31094 __extension__ extern __inline float16x4_t
31095 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31096 vfms_n_f16 (float16x4_t __a, float16x4_t __b, float16_t __c)
31097 {
31098   return vfms_f16 (__a, __b, vdup_n_f16 (__c));
31099 }
31100
31101 __extension__ extern __inline float16x8_t
31102 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31103 vfmsq_n_f16 (float16x8_t __a, float16x8_t __b, float16_t __c)
31104 {
31105   return vfmsq_f16 (__a, __b, vdupq_n_f16 (__c));
31106 }
31107
31108 __extension__ extern __inline float16_t
31109 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31110 vmulh_lane_f16 (float16_t __a, float16x4_t __b, const int __lane)
31111 {
31112   return __a * __aarch64_vget_lane_any (__b, __lane);
31113 }
31114
31115 __extension__ extern __inline float16x4_t
31116 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31117 vmul_lane_f16 (float16x4_t __a, float16x4_t __b, const int __lane)
31118 {
31119   return vmul_f16 (__a, vdup_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31120 }
31121
31122 __extension__ extern __inline float16x8_t
31123 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31124 vmulq_lane_f16 (float16x8_t __a, float16x4_t __b, const int __lane)
31125 {
31126   return vmulq_f16 (__a, vdupq_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31127 }
31128
31129 __extension__ extern __inline float16_t
31130 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31131 vmulh_laneq_f16 (float16_t __a, float16x8_t __b, const int __lane)
31132 {
31133   return __a * __aarch64_vget_lane_any (__b, __lane);
31134 }
31135
31136 __extension__ extern __inline float16x4_t
31137 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31138 vmul_laneq_f16 (float16x4_t __a, float16x8_t __b, const int __lane)
31139 {
31140   return vmul_f16 (__a, vdup_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31141 }
31142
31143 __extension__ extern __inline float16x8_t
31144 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31145 vmulq_laneq_f16 (float16x8_t __a, float16x8_t __b, const int __lane)
31146 {
31147   return vmulq_f16 (__a, vdupq_n_f16 (__aarch64_vget_lane_any (__b, __lane)));
31148 }
31149
31150 __extension__ extern __inline float16x4_t
31151 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31152 vmul_n_f16 (float16x4_t __a, float16_t __b)
31153 {
31154   return vmul_lane_f16 (__a, vdup_n_f16 (__b), 0);
31155 }
31156
31157 __extension__ extern __inline float16x8_t
31158 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31159 vmulq_n_f16 (float16x8_t __a, float16_t __b)
31160 {
31161   return vmulq_laneq_f16 (__a, vdupq_n_f16 (__b), 0);
31162 }
31163
31164 __extension__ extern __inline float16_t
31165 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31166 vmulxh_lane_f16 (float16_t __a, float16x4_t __b, const int __lane)
31167 {
31168   return vmulxh_f16 (__a, __aarch64_vget_lane_any (__b, __lane));
31169 }
31170
31171 __extension__ extern __inline float16x4_t
31172 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31173 vmulx_lane_f16 (float16x4_t __a, float16x4_t __b, const int __lane)
31174 {
31175   return vmulx_f16 (__a, __aarch64_vdup_lane_f16 (__b, __lane));
31176 }
31177
31178 __extension__ extern __inline float16x8_t
31179 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31180 vmulxq_lane_f16 (float16x8_t __a, float16x4_t __b, const int __lane)
31181 {
31182   return vmulxq_f16 (__a, __aarch64_vdupq_lane_f16 (__b, __lane));
31183 }
31184
31185 __extension__ extern __inline float16_t
31186 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31187 vmulxh_laneq_f16 (float16_t __a, float16x8_t __b, const int __lane)
31188 {
31189   return vmulxh_f16 (__a, __aarch64_vget_lane_any (__b, __lane));
31190 }
31191
31192 __extension__ extern __inline float16x4_t
31193 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31194 vmulx_laneq_f16 (float16x4_t __a, float16x8_t __b, const int __lane)
31195 {
31196   return vmulx_f16 (__a, __aarch64_vdup_laneq_f16 (__b, __lane));
31197 }
31198
31199 __extension__ extern __inline float16x8_t
31200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31201 vmulxq_laneq_f16 (float16x8_t __a, float16x8_t __b, const int __lane)
31202 {
31203   return vmulxq_f16 (__a, __aarch64_vdupq_laneq_f16 (__b, __lane));
31204 }
31205
31206 __extension__ extern __inline float16x4_t
31207 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31208 vmulx_n_f16 (float16x4_t __a, float16_t __b)
31209 {
31210   return vmulx_f16 (__a, vdup_n_f16 (__b));
31211 }
31212
31213 __extension__ extern __inline float16x8_t
31214 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31215 vmulxq_n_f16 (float16x8_t __a, float16_t __b)
31216 {
31217   return vmulxq_f16 (__a, vdupq_n_f16 (__b));
31218 }
31219
31220 /* ARMv8.2-A FP16 reduction vector intrinsics.  */
31221
31222 __extension__ extern __inline float16_t
31223 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31224 vmaxv_f16 (float16x4_t __a)
31225 {
31226   return __builtin_aarch64_reduc_smax_nan_scal_v4hf (__a);
31227 }
31228
31229 __extension__ extern __inline float16_t
31230 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31231 vmaxvq_f16 (float16x8_t __a)
31232 {
31233   return __builtin_aarch64_reduc_smax_nan_scal_v8hf (__a);
31234 }
31235
31236 __extension__ extern __inline float16_t
31237 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31238 vminv_f16 (float16x4_t __a)
31239 {
31240   return __builtin_aarch64_reduc_smin_nan_scal_v4hf (__a);
31241 }
31242
31243 __extension__ extern __inline float16_t
31244 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31245 vminvq_f16 (float16x8_t __a)
31246 {
31247   return __builtin_aarch64_reduc_smin_nan_scal_v8hf (__a);
31248 }
31249
31250 __extension__ extern __inline float16_t
31251 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31252 vmaxnmv_f16 (float16x4_t __a)
31253 {
31254   return __builtin_aarch64_reduc_smax_scal_v4hf (__a);
31255 }
31256
31257 __extension__ extern __inline float16_t
31258 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31259 vmaxnmvq_f16 (float16x8_t __a)
31260 {
31261   return __builtin_aarch64_reduc_smax_scal_v8hf (__a);
31262 }
31263
31264 __extension__ extern __inline float16_t
31265 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31266 vminnmv_f16 (float16x4_t __a)
31267 {
31268   return __builtin_aarch64_reduc_smin_scal_v4hf (__a);
31269 }
31270
31271 __extension__ extern __inline float16_t
31272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
31273 vminnmvq_f16 (float16x8_t __a)
31274 {
31275   return __builtin_aarch64_reduc_smin_scal_v8hf (__a);
31276 }
31277
31278 #pragma GCC pop_options
31279
31280 #undef __aarch64_vget_lane_any
31281
31282 #undef __aarch64_vdup_lane_any
31283 #undef __aarch64_vdup_lane_f16
31284 #undef __aarch64_vdup_lane_f32
31285 #undef __aarch64_vdup_lane_f64
31286 #undef __aarch64_vdup_lane_p8
31287 #undef __aarch64_vdup_lane_p16
31288 #undef __aarch64_vdup_lane_s8
31289 #undef __aarch64_vdup_lane_s16
31290 #undef __aarch64_vdup_lane_s32
31291 #undef __aarch64_vdup_lane_s64
31292 #undef __aarch64_vdup_lane_u8
31293 #undef __aarch64_vdup_lane_u16
31294 #undef __aarch64_vdup_lane_u32
31295 #undef __aarch64_vdup_lane_u64
31296 #undef __aarch64_vdup_laneq_f16
31297 #undef __aarch64_vdup_laneq_f32
31298 #undef __aarch64_vdup_laneq_f64
31299 #undef __aarch64_vdup_laneq_p8
31300 #undef __aarch64_vdup_laneq_p16
31301 #undef __aarch64_vdup_laneq_s8
31302 #undef __aarch64_vdup_laneq_s16
31303 #undef __aarch64_vdup_laneq_s32
31304 #undef __aarch64_vdup_laneq_s64
31305 #undef __aarch64_vdup_laneq_u8
31306 #undef __aarch64_vdup_laneq_u16
31307 #undef __aarch64_vdup_laneq_u32
31308 #undef __aarch64_vdup_laneq_u64
31309 #undef __aarch64_vdupq_lane_f16
31310 #undef __aarch64_vdupq_lane_f32
31311 #undef __aarch64_vdupq_lane_f64
31312 #undef __aarch64_vdupq_lane_p8
31313 #undef __aarch64_vdupq_lane_p16
31314 #undef __aarch64_vdupq_lane_s8
31315 #undef __aarch64_vdupq_lane_s16
31316 #undef __aarch64_vdupq_lane_s32
31317 #undef __aarch64_vdupq_lane_s64
31318 #undef __aarch64_vdupq_lane_u8
31319 #undef __aarch64_vdupq_lane_u16
31320 #undef __aarch64_vdupq_lane_u32
31321 #undef __aarch64_vdupq_lane_u64
31322 #undef __aarch64_vdupq_laneq_f16
31323 #undef __aarch64_vdupq_laneq_f32
31324 #undef __aarch64_vdupq_laneq_f64
31325 #undef __aarch64_vdupq_laneq_p8
31326 #undef __aarch64_vdupq_laneq_p16
31327 #undef __aarch64_vdupq_laneq_s8
31328 #undef __aarch64_vdupq_laneq_s16
31329 #undef __aarch64_vdupq_laneq_s32
31330 #undef __aarch64_vdupq_laneq_s64
31331 #undef __aarch64_vdupq_laneq_u8
31332 #undef __aarch64_vdupq_laneq_u16
31333 #undef __aarch64_vdupq_laneq_u32
31334 #undef __aarch64_vdupq_laneq_u64
31335
31336 #endif