gcc/testsuite/gcc.target/mips/loongson-simd.c

   1 /* Test cases for ST Microelectronics Loongson-2E/2F SIMD intrinsics.
   2    Copyright (C) 2008 Free Software Foundation, Inc.
   3    Contributed by CodeSourcery.
   4
   5 This file is part of GCC.
   6
   7 GCC is free software; you can redistribute it and/or modify
   8 it under the terms of the GNU General Public License as published by
   9 the Free Software Foundation; either version 3, or (at your option)
  10 any later version.
  11
  12 GCC is distributed in the hope that it will be useful,
  13 but WITHOUT ANY WARRANTY; without even the implied warranty of
  14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15 GNU General Public License for more details.
  16
  17 You should have received a copy of the GNU General Public License
  18 along with GCC; see the file COPYING3.  If not see
  19 <http://www.gnu.org/licenses/>.  */
  20
  21 /* { dg-do run } */
  22 /* loongson.h does not handle or check for MIPS16ness.  There doesn't
  23    seem any good reason for it to, given that the Loongson processors
  24    do not support MIPS16.  */
  25 /* { dg-options "isa=loongson -mhard-float -mno-mips16 -flax-vector-conversions" } */
  26
  27 #include "loongson.h"
  28 #include <stdio.h>
  29 #include <stdint.h>
  30 #include <assert.h>
  31 #include <limits.h>
  32
  33 typedef union { int32x2_t v; int32_t a[2]; } int32x2_encap_t;
  34 typedef union { int16x4_t v; int16_t a[4]; } int16x4_encap_t;
  35 typedef union { int8x8_t v; int8_t a[8]; } int8x8_encap_t;
  36 typedef union { uint32x2_t v; uint32_t a[2]; } uint32x2_encap_t;
  37 typedef union { uint16x4_t v; uint16_t a[4]; } uint16x4_encap_t;
  38 typedef union { uint8x8_t v; uint8_t a[8]; } uint8x8_encap_t;
  39
  40 #define UINT16x4_MAX USHRT_MAX
  41 #define UINT8x8_MAX UCHAR_MAX
  42 #define INT8x8_MAX SCHAR_MAX
  43 #define INT16x4_MAX SHRT_MAX
  44 #define INT32x2_MAX INT_MAX
  45
  46 static void test_packsswh (void)
  47 {
  48   int32x2_encap_t s, t;
  49   int16x4_encap_t r;
  50   s.a[0] = INT16x4_MAX - 2;
  51   s.a[1] = INT16x4_MAX - 1;
  52   t.a[0] = INT16x4_MAX;
  53   t.a[1] = INT16x4_MAX + 1;
  54   r.v = packsswh (s.v, t.v);
  55   assert (r.a[0] == INT16x4_MAX - 2);
  56   assert (r.a[1] == INT16x4_MAX - 1);
  57   assert (r.a[2] == INT16x4_MAX);
  58   assert (r.a[3] == INT16x4_MAX);
  59 }
  60
  61 static void test_packsshb (void)
  62 {
  63   int16x4_encap_t s, t;
  64   int8x8_encap_t r;
  65   s.a[0] = INT8x8_MAX - 6;
  66   s.a[1] = INT8x8_MAX - 5;
  67   s.a[2] = INT8x8_MAX - 4;
  68   s.a[3] = INT8x8_MAX - 3;
  69   t.a[0] = INT8x8_MAX - 2;
  70   t.a[1] = INT8x8_MAX - 1;
  71   t.a[2] = INT8x8_MAX;
  72   t.a[3] = INT8x8_MAX + 1;
  73   r.v = packsshb (s.v, t.v);
  74   assert (r.a[0] == INT8x8_MAX - 6);
  75   assert (r.a[1] == INT8x8_MAX - 5);
  76   assert (r.a[2] == INT8x8_MAX - 4);
  77   assert (r.a[3] == INT8x8_MAX - 3);
  78   assert (r.a[4] == INT8x8_MAX - 2);
  79   assert (r.a[5] == INT8x8_MAX - 1);
  80   assert (r.a[6] == INT8x8_MAX);
  81   assert (r.a[7] == INT8x8_MAX);
  82 }
  83
  84 static void test_packushb (void)
  85 {
  86   uint16x4_encap_t s, t;
  87   uint8x8_encap_t r;
  88   s.a[0] = UINT8x8_MAX - 6;
  89   s.a[1] = UINT8x8_MAX - 5;
  90   s.a[2] = UINT8x8_MAX - 4;
  91   s.a[3] = UINT8x8_MAX - 3;
  92   t.a[0] = UINT8x8_MAX - 2;
  93   t.a[1] = UINT8x8_MAX - 1;
  94   t.a[2] = UINT8x8_MAX;
  95   t.a[3] = UINT8x8_MAX + 1;
  96   r.v = packushb (s.v, t.v);
  97   assert (r.a[0] == UINT8x8_MAX - 6);
  98   assert (r.a[1] == UINT8x8_MAX - 5);
  99   assert (r.a[2] == UINT8x8_MAX - 4);
 100   assert (r.a[3] == UINT8x8_MAX - 3);
 101   assert (r.a[4] == UINT8x8_MAX - 2);
 102   assert (r.a[5] == UINT8x8_MAX - 1);
 103   assert (r.a[6] == UINT8x8_MAX);
 104   assert (r.a[7] == UINT8x8_MAX);
 105 }
 106
 107 static void test_paddw_u (void)
 108 {
 109   uint32x2_encap_t s, t;
 110   uint32x2_encap_t r;
 111   s.a[0] = 1;
 112   s.a[1] = 2;
 113   t.a[0] = 3;
 114   t.a[1] = 4;
 115   r.v = paddw_u (s.v, t.v);
 116   assert (r.a[0] == 4);
 117   assert (r.a[1] == 6);
 118 }
 119
 120 static void test_paddw_s (void)
 121 {
 122   int32x2_encap_t s, t;
 123   int32x2_encap_t r;
 124   s.a[0] = -2;
 125   s.a[1] = -1;
 126   t.a[0] = 3;
 127   t.a[1] = 4;
 128   r.v = paddw_s (s.v, t.v);
 129   assert (r.a[0] == 1);
 130   assert (r.a[1] == 3);
 131 }
 132
 133 static void test_paddh_u (void)
 134 {
 135   uint16x4_encap_t s, t;
 136   uint16x4_encap_t r;
 137   s.a[0] = 1;
 138   s.a[1] = 2;
 139   s.a[2] = 3;
 140   s.a[3] = 4;
 141   t.a[0] = 5;
 142   t.a[1] = 6;
 143   t.a[2] = 7;
 144   t.a[3] = 8;
 145   r.v = paddh_u (s.v, t.v);
 146   assert (r.a[0] == 6);
 147   assert (r.a[1] == 8);
 148   assert (r.a[2] == 10);
 149   assert (r.a[3] == 12);
 150 }
 151
 152 static void test_paddh_s (void)
 153 {
 154   int16x4_encap_t s, t;
 155   int16x4_encap_t r;
 156   s.a[0] = -10;
 157   s.a[1] = -20;
 158   s.a[2] = -30;
 159   s.a[3] = -40;
 160   t.a[0] = 1;
 161   t.a[1] = 2;
 162   t.a[2] = 3;
 163   t.a[3] = 4;
 164   r.v = paddh_s (s.v, t.v);
 165   assert (r.a[0] == -9);
 166   assert (r.a[1] == -18);
 167   assert (r.a[2] == -27);
 168   assert (r.a[3] == -36);
 169 }
 170
 171 static void test_paddb_u (void)
 172 {
 173   uint8x8_encap_t s, t;
 174   uint8x8_encap_t r;
 175   s.a[0] = 1;
 176   s.a[1] = 2;
 177   s.a[2] = 3;
 178   s.a[3] = 4;
 179   s.a[4] = 5;
 180   s.a[5] = 6;
 181   s.a[6] = 7;
 182   s.a[7] = 8;
 183   t.a[0] = 9;
 184   t.a[1] = 10;
 185   t.a[2] = 11;
 186   t.a[3] = 12;
 187   t.a[4] = 13;
 188   t.a[5] = 14;
 189   t.a[6] = 15;
 190   t.a[7] = 16;
 191   r.v = paddb_u (s.v, t.v);
 192   assert (r.a[0] == 10);
 193   assert (r.a[1] == 12);
 194   assert (r.a[2] == 14);
 195   assert (r.a[3] == 16);
 196   assert (r.a[4] == 18);
 197   assert (r.a[5] == 20);
 198   assert (r.a[6] == 22);
 199   assert (r.a[7] == 24);
 200 }
 201
 202 static void test_paddb_s (void)
 203 {
 204   int8x8_encap_t s, t;
 205   int8x8_encap_t r;
 206   s.a[0] = -10;
 207   s.a[1] = -20;
 208   s.a[2] = -30;
 209   s.a[3] = -40;
 210   s.a[4] = -50;
 211   s.a[5] = -60;
 212   s.a[6] = -70;
 213   s.a[7] = -80;
 214   t.a[0] = 1;
 215   t.a[1] = 2;
 216   t.a[2] = 3;
 217   t.a[3] = 4;
 218   t.a[4] = 5;
 219   t.a[5] = 6;
 220   t.a[6] = 7;
 221   t.a[7] = 8;
 222   r.v = paddb_s (s.v, t.v);
 223   assert (r.a[0] == -9);
 224   assert (r.a[1] == -18);
 225   assert (r.a[2] == -27);
 226   assert (r.a[3] == -36);
 227   assert (r.a[4] == -45);
 228   assert (r.a[5] == -54);
 229   assert (r.a[6] == -63);
 230   assert (r.a[7] == -72);
 231 }
 232
 233 static void test_paddd_u (void)
 234 {
 235   uint64_t d = 123456;
 236   uint64_t e = 789012;
 237   uint64_t r;
 238   r = paddd_u (d, e);
 239   assert (r == 912468);
 240 }
 241
 242 static void test_paddd_s (void)
 243 {
 244   int64_t d = 123456;
 245   int64_t e = -789012;
 246   int64_t r;
 247   r = paddd_s (d, e);
 248   assert (r == -665556);
 249 }
 250
 251 static void test_paddsh (void)
 252 {
 253   int16x4_encap_t s, t;
 254   int16x4_encap_t r;
 255   s.a[0] = -1;
 256   s.a[1] = 0;
 257   s.a[2] = 1;
 258   s.a[3] = 2;
 259   t.a[0] = INT16x4_MAX;
 260   t.a[1] = INT16x4_MAX;
 261   t.a[2] = INT16x4_MAX;
 262   t.a[3] = INT16x4_MAX;
 263   r.v = paddsh (s.v, t.v);
 264   assert (r.a[0] == INT16x4_MAX - 1);
 265   assert (r.a[1] == INT16x4_MAX);
 266   assert (r.a[2] == INT16x4_MAX);
 267   assert (r.a[3] == INT16x4_MAX);
 268 }
 269
 270 static void test_paddsb (void)
 271 {
 272   int8x8_encap_t s, t;
 273   int8x8_encap_t r;
 274   s.a[0] = -6;
 275   s.a[1] = -5;
 276   s.a[2] = -4;
 277   s.a[3] = -3;
 278   s.a[4] = -2;
 279   s.a[5] = -1;
 280   s.a[6] = 0;
 281   s.a[7] = 1;
 282   t.a[0] = INT8x8_MAX;
 283   t.a[1] = INT8x8_MAX;
 284   t.a[2] = INT8x8_MAX;
 285   t.a[3] = INT8x8_MAX;
 286   t.a[4] = INT8x8_MAX;
 287   t.a[5] = INT8x8_MAX;
 288   t.a[6] = INT8x8_MAX;
 289   t.a[7] = INT8x8_MAX;
 290   r.v = paddsb (s.v, t.v);
 291   assert (r.a[0] == INT8x8_MAX - 6);
 292   assert (r.a[1] == INT8x8_MAX - 5);
 293   assert (r.a[2] == INT8x8_MAX - 4);
 294   assert (r.a[3] == INT8x8_MAX - 3);
 295   assert (r.a[4] == INT8x8_MAX - 2);
 296   assert (r.a[5] == INT8x8_MAX - 1);
 297   assert (r.a[6] == INT8x8_MAX);
 298   assert (r.a[7] == INT8x8_MAX);
 299 }
 300
 301 static void test_paddush (void)
 302 {
 303   uint16x4_encap_t s, t;
 304   uint16x4_encap_t r;
 305   s.a[0] = 0;
 306   s.a[1] = 1;
 307   s.a[2] = 0;
 308   s.a[3] = 1;
 309   t.a[0] = UINT16x4_MAX;
 310   t.a[1] = UINT16x4_MAX;
 311   t.a[2] = UINT16x4_MAX;
 312   t.a[3] = UINT16x4_MAX;
 313   r.v = paddush (s.v, t.v);
 314   assert (r.a[0] == UINT16x4_MAX);
 315   assert (r.a[1] == UINT16x4_MAX);
 316   assert (r.a[2] == UINT16x4_MAX);
 317   assert (r.a[3] == UINT16x4_MAX);
 318 }
 319
 320 static void test_paddusb (void)
 321 {
 322   uint8x8_encap_t s, t;
 323   uint8x8_encap_t r;
 324   s.a[0] = 0;
 325   s.a[1] = 1;
 326   s.a[2] = 0;
 327   s.a[3] = 1;
 328   s.a[4] = 0;
 329   s.a[5] = 1;
 330   s.a[6] = 0;
 331   s.a[7] = 1;
 332   t.a[0] = UINT8x8_MAX;
 333   t.a[1] = UINT8x8_MAX;
 334   t.a[2] = UINT8x8_MAX;
 335   t.a[3] = UINT8x8_MAX;
 336   t.a[4] = UINT8x8_MAX;
 337   t.a[5] = UINT8x8_MAX;
 338   t.a[6] = UINT8x8_MAX;
 339   t.a[7] = UINT8x8_MAX;
 340   r.v = paddusb (s.v, t.v);
 341   assert (r.a[0] == UINT8x8_MAX);
 342   assert (r.a[1] == UINT8x8_MAX);
 343   assert (r.a[2] == UINT8x8_MAX);
 344   assert (r.a[3] == UINT8x8_MAX);
 345   assert (r.a[4] == UINT8x8_MAX);
 346   assert (r.a[5] == UINT8x8_MAX);
 347   assert (r.a[6] == UINT8x8_MAX);
 348   assert (r.a[7] == UINT8x8_MAX);
 349 }
 350
 351 static void test_pandn_ud (void)
 352 {
 353   uint64_t d1 = 0x0000ffff0000ffffull;
 354   uint64_t d2 = 0x0000ffff0000ffffull;
 355   uint64_t r;
 356   r = pandn_ud (d1, d2);
 357   assert (r == 0);
 358 }
 359
 360 static void test_pandn_sd (void)
 361 {
 362   int64_t d1 = (int64_t) 0x0000000000000000ull;
 363   int64_t d2 = (int64_t) 0xfffffffffffffffeull;
 364   int64_t r;
 365   r = pandn_sd (d1, d2);
 366   assert (r == -2);
 367 }
 368
 369 static void test_pandn_uw (void)
 370 {
 371   uint32x2_encap_t s, t;
 372   uint32x2_encap_t r;
 373   s.a[0] = 0xffffffff;
 374   s.a[1] = 0x00000000;
 375   t.a[0] = 0x00000000;
 376   t.a[1] = 0xffffffff;
 377   r.v = pandn_uw (s.v, t.v);
 378   assert (r.a[0] == 0x00000000);
 379   assert (r.a[1] == 0xffffffff);
 380 }
 381
 382 static void test_pandn_sw (void)
 383 {
 384   int32x2_encap_t s, t;
 385   int32x2_encap_t r;
 386   s.a[0] = 0xffffffff;
 387   s.a[1] = 0x00000000;
 388   t.a[0] = 0xffffffff;
 389   t.a[1] = 0xfffffffe;
 390   r.v = pandn_sw (s.v, t.v);
 391   assert (r.a[0] == 0);
 392   assert (r.a[1] == -2);
 393 }
 394
 395 static void test_pandn_uh (void)
 396 {
 397   uint16x4_encap_t s, t;
 398   uint16x4_encap_t r;
 399   s.a[0] = 0xffff;
 400   s.a[1] = 0x0000;
 401   s.a[2] = 0xffff;
 402   s.a[3] = 0x0000;
 403   t.a[0] = 0x0000;
 404   t.a[1] = 0xffff;
 405   t.a[2] = 0x0000;
 406   t.a[3] = 0xffff;
 407   r.v = pandn_uh (s.v, t.v);
 408   assert (r.a[0] == 0x0000);
 409   assert (r.a[1] == 0xffff);
 410   assert (r.a[2] == 0x0000);
 411   assert (r.a[3] == 0xffff);
 412 }
 413
 414 static void test_pandn_sh (void)
 415 {
 416   int16x4_encap_t s, t;
 417   int16x4_encap_t r;
 418   s.a[0] = 0xffff;
 419   s.a[1] = 0x0000;
 420   s.a[2] = 0xffff;
 421   s.a[3] = 0x0000;
 422   t.a[0] = 0xffff;
 423   t.a[1] = 0xfffe;
 424   t.a[2] = 0xffff;
 425   t.a[3] = 0xfffe;
 426   r.v = pandn_sh (s.v, t.v);
 427   assert (r.a[0] == 0);
 428   assert (r.a[1] == -2);
 429   assert (r.a[2] == 0);
 430   assert (r.a[3] == -2);
 431 }
 432
 433 static void test_pandn_ub (void)
 434 {
 435   uint8x8_encap_t s, t;
 436   uint8x8_encap_t r;
 437   s.a[0] = 0xff;
 438   s.a[1] = 0x00;
 439   s.a[2] = 0xff;
 440   s.a[3] = 0x00;
 441   s.a[4] = 0xff;
 442   s.a[5] = 0x00;
 443   s.a[6] = 0xff;
 444   s.a[7] = 0x00;
 445   t.a[0] = 0x00;
 446   t.a[1] = 0xff;
 447   t.a[2] = 0x00;
 448   t.a[3] = 0xff;
 449   t.a[4] = 0x00;
 450   t.a[5] = 0xff;
 451   t.a[6] = 0x00;
 452   t.a[7] = 0xff;
 453   r.v = pandn_ub (s.v, t.v);
 454   assert (r.a[0] == 0x00);
 455   assert (r.a[1] == 0xff);
 456   assert (r.a[2] == 0x00);
 457   assert (r.a[3] == 0xff);
 458   assert (r.a[4] == 0x00);
 459   assert (r.a[5] == 0xff);
 460   assert (r.a[6] == 0x00);
 461   assert (r.a[7] == 0xff);
 462 }
 463
 464 static void test_pandn_sb (void)
 465 {
 466   int8x8_encap_t s, t;
 467   int8x8_encap_t r;
 468   s.a[0] = 0xff;
 469   s.a[1] = 0x00;
 470   s.a[2] = 0xff;
 471   s.a[3] = 0x00;
 472   s.a[4] = 0xff;
 473   s.a[5] = 0x00;
 474   s.a[6] = 0xff;
 475   s.a[7] = 0x00;
 476   t.a[0] = 0xff;
 477   t.a[1] = 0xfe;
 478   t.a[2] = 0xff;
 479   t.a[3] = 0xfe;
 480   t.a[4] = 0xff;
 481   t.a[5] = 0xfe;
 482   t.a[6] = 0xff;
 483   t.a[7] = 0xfe;
 484   r.v = pandn_sb (s.v, t.v);
 485   assert (r.a[0] == 0);
 486   assert (r.a[1] == -2);
 487   assert (r.a[2] == 0);
 488   assert (r.a[3] == -2);
 489   assert (r.a[4] == 0);
 490   assert (r.a[5] == -2);
 491   assert (r.a[6] == 0);
 492   assert (r.a[7] == -2);
 493 }
 494
 495 static void test_pavgh (void)
 496 {
 497   uint16x4_encap_t s, t;
 498   uint16x4_encap_t r;
 499   s.a[0] = 1;
 500   s.a[1] = 2;
 501   s.a[2] = 3;
 502   s.a[3] = 4;
 503   t.a[0] = 5;
 504   t.a[1] = 6;
 505   t.a[2] = 7;
 506   t.a[3] = 8;
 507   r.v = pavgh (s.v, t.v);
 508   assert (r.a[0] == 3);
 509   assert (r.a[1] == 4);
 510   assert (r.a[2] == 5);
 511   assert (r.a[3] == 6);
 512 }
 513
 514 static void test_pavgb (void)
 515 {
 516   uint8x8_encap_t s, t;
 517   uint8x8_encap_t r;
 518   s.a[0] = 1;
 519   s.a[1] = 2;
 520   s.a[2] = 3;
 521   s.a[3] = 4;
 522   s.a[4] = 1;
 523   s.a[5] = 2;
 524   s.a[6] = 3;
 525   s.a[7] = 4;
 526   t.a[0] = 5;
 527   t.a[1] = 6;
 528   t.a[2] = 7;
 529   t.a[3] = 8;
 530   t.a[4] = 5;
 531   t.a[5] = 6;
 532   t.a[6] = 7;
 533   t.a[7] = 8;
 534   r.v = pavgb (s.v, t.v);
 535   assert (r.a[0] == 3);
 536   assert (r.a[1] == 4);
 537   assert (r.a[2] == 5);
 538   assert (r.a[3] == 6);
 539   assert (r.a[4] == 3);
 540   assert (r.a[5] == 4);
 541   assert (r.a[6] == 5);
 542   assert (r.a[7] == 6);
 543 }
 544
 545 static void test_pcmpeqw_u (void)
 546 {
 547   uint32x2_encap_t s, t;
 548   uint32x2_encap_t r;
 549   s.a[0] = 42;
 550   s.a[1] = 43;
 551   t.a[0] = 43;
 552   t.a[1] = 43;
 553   r.v = pcmpeqw_u (s.v, t.v);
 554   assert (r.a[0] == 0x00000000);
 555   assert (r.a[1] == 0xffffffff);
 556 }
 557
 558 static void test_pcmpeqh_u (void)
 559 {
 560   uint16x4_encap_t s, t;
 561   uint16x4_encap_t r;
 562   s.a[0] = 42;
 563   s.a[1] = 43;
 564   s.a[2] = 42;
 565   s.a[3] = 43;
 566   t.a[0] = 43;
 567   t.a[1] = 43;
 568   t.a[2] = 43;
 569   t.a[3] = 43;
 570   r.v = pcmpeqh_u (s.v, t.v);
 571   assert (r.a[0] == 0x0000);
 572   assert (r.a[1] == 0xffff);
 573   assert (r.a[2] == 0x0000);
 574   assert (r.a[3] == 0xffff);
 575 }
 576
 577 static void test_pcmpeqb_u (void)
 578 {
 579   uint8x8_encap_t s, t;
 580   uint8x8_encap_t r;
 581   s.a[0] = 42;
 582   s.a[1] = 43;
 583   s.a[2] = 42;
 584   s.a[3] = 43;
 585   s.a[4] = 42;
 586   s.a[5] = 43;
 587   s.a[6] = 42;
 588   s.a[7] = 43;
 589   t.a[0] = 43;
 590   t.a[1] = 43;
 591   t.a[2] = 43;
 592   t.a[3] = 43;
 593   t.a[4] = 43;
 594   t.a[5] = 43;
 595   t.a[6] = 43;
 596   t.a[7] = 43;
 597   r.v = pcmpeqb_u (s.v, t.v);
 598   assert (r.a[0] == 0x00);
 599   assert (r.a[1] == 0xff);
 600   assert (r.a[2] == 0x00);
 601   assert (r.a[3] == 0xff);
 602   assert (r.a[4] == 0x00);
 603   assert (r.a[5] == 0xff);
 604   assert (r.a[6] == 0x00);
 605   assert (r.a[7] == 0xff);
 606 }
 607
 608 static void test_pcmpeqw_s (void)
 609 {
 610   int32x2_encap_t s, t;
 611   int32x2_encap_t r;
 612   s.a[0] = -42;
 613   s.a[1] = -42;
 614   t.a[0] = 42;
 615   t.a[1] = -42;
 616   r.v = pcmpeqw_s (s.v, t.v);
 617   assert (r.a[0] == 0);
 618   assert (r.a[1] == -1);
 619 }
 620
 621 static void test_pcmpeqh_s (void)
 622 {
 623   int16x4_encap_t s, t;
 624   int16x4_encap_t r;
 625   s.a[0] = -42;
 626   s.a[1] = -42;
 627   s.a[2] = -42;
 628   s.a[3] = -42;
 629   t.a[0] = 42;
 630   t.a[1] = -42;
 631   t.a[2] = 42;
 632   t.a[3] = -42;
 633   r.v = pcmpeqh_s (s.v, t.v);
 634   assert (r.a[0] == 0);
 635   assert (r.a[1] == -1);
 636   assert (r.a[2] == 0);
 637   assert (r.a[3] == -1);
 638 }
 639
 640 static void test_pcmpeqb_s (void)
 641 {
 642   int8x8_encap_t s, t;
 643   int8x8_encap_t r;
 644   s.a[0] = -42;
 645   s.a[1] = -42;
 646   s.a[2] = -42;
 647   s.a[3] = -42;
 648   s.a[4] = -42;
 649   s.a[5] = -42;
 650   s.a[6] = -42;
 651   s.a[7] = -42;
 652   t.a[0] = 42;
 653   t.a[1] = -42;
 654   t.a[2] = 42;
 655   t.a[3] = -42;
 656   t.a[4] = 42;
 657   t.a[5] = -42;
 658   t.a[6] = 42;
 659   t.a[7] = -42;
 660   r.v = pcmpeqb_s (s.v, t.v);
 661   assert (r.a[0] == 0);
 662   assert (r.a[1] == -1);
 663   assert (r.a[2] == 0);
 664   assert (r.a[3] == -1);
 665   assert (r.a[4] == 0);
 666   assert (r.a[5] == -1);
 667   assert (r.a[6] == 0);
 668   assert (r.a[7] == -1);
 669 }
 670
 671 static void test_pcmpgtw_u (void)
 672 {
 673   uint32x2_encap_t s, t;
 674   uint32x2_encap_t r;
 675   s.a[0] = 42;
 676   s.a[1] = 43;
 677   t.a[0] = 43;
 678   t.a[1] = 42;
 679   r.v = pcmpgtw_u (s.v, t.v);
 680   assert (r.a[0] == 0x00000000);
 681   assert (r.a[1] == 0xffffffff);
 682 }
 683
 684 static void test_pcmpgth_u (void)
 685 {
 686   uint16x4_encap_t s, t;
 687   uint16x4_encap_t r;
 688   s.a[0] = 40;
 689   s.a[1] = 41;
 690   s.a[2] = 42;
 691   s.a[3] = 43;
 692   t.a[0] = 40;
 693   t.a[1] = 41;
 694   t.a[2] = 43;
 695   t.a[3] = 42;
 696   r.v = pcmpgth_u (s.v, t.v);
 697   assert (r.a[0] == 0x0000);
 698   assert (r.a[1] == 0x0000);
 699   assert (r.a[2] == 0x0000);
 700   assert (r.a[3] == 0xffff);
 701 }
 702
 703 static void test_pcmpgtb_u (void)
 704 {
 705   uint8x8_encap_t s, t;
 706   uint8x8_encap_t r;
 707   s.a[0] = 40;
 708   s.a[1] = 41;
 709   s.a[2] = 42;
 710   s.a[3] = 43;
 711   s.a[4] = 44;
 712   s.a[5] = 45;
 713   s.a[6] = 46;
 714   s.a[7] = 47;
 715   t.a[0] = 48;
 716   t.a[1] = 47;
 717   t.a[2] = 46;
 718   t.a[3] = 45;
 719   t.a[4] = 44;
 720   t.a[5] = 43;
 721   t.a[6] = 42;
 722   t.a[7] = 41;
 723   r.v = pcmpgtb_u (s.v, t.v);
 724   assert (r.a[0] == 0x00);
 725   assert (r.a[1] == 0x00);
 726   assert (r.a[2] == 0x00);
 727   assert (r.a[3] == 0x00);
 728   assert (r.a[4] == 0x00);
 729   assert (r.a[5] == 0xff);
 730   assert (r.a[6] == 0xff);
 731   assert (r.a[7] == 0xff);
 732 }
 733
 734 static void test_pcmpgtw_s (void)
 735 {
 736   int32x2_encap_t s, t;
 737   int32x2_encap_t r;
 738   s.a[0] = 42;
 739   s.a[1] = -42;
 740   t.a[0] = -42;
 741   t.a[1] = -42;
 742   r.v = pcmpgtw_s (s.v, t.v);
 743   assert (r.a[0] == -1);
 744   assert (r.a[1] == 0);
 745 }
 746
 747 static void test_pcmpgth_s (void)
 748 {
 749   int16x4_encap_t s, t;
 750   int16x4_encap_t r;
 751   s.a[0] = -42;
 752   s.a[1] = -42;
 753   s.a[2] = -42;
 754   s.a[3] = -42;
 755   t.a[0] = 42;
 756   t.a[1] = 43;
 757   t.a[2] = 44;
 758   t.a[3] = -43;
 759   r.v = pcmpgth_s (s.v, t.v);
 760   assert (r.a[0] == 0);
 761   assert (r.a[1] == 0);
 762   assert (r.a[2] == 0);
 763   assert (r.a[3] == -1);
 764 }
 765
 766 static void test_pcmpgtb_s (void)
 767 {
 768   int8x8_encap_t s, t;
 769   int8x8_encap_t r;
 770   s.a[0] = -42;
 771   s.a[1] = -42;
 772   s.a[2] = -42;
 773   s.a[3] = -42;
 774   s.a[4] = 42;
 775   s.a[5] = 42;
 776   s.a[6] = 42;
 777   s.a[7] = 42;
 778   t.a[0] = -45;
 779   t.a[1] = -44;
 780   t.a[2] = -43;
 781   t.a[3] = -42;
 782   t.a[4] = 42;
 783   t.a[5] = 43;
 784   t.a[6] = 41;
 785   t.a[7] = 40;
 786   r.v = pcmpgtb_s (s.v, t.v);
 787   assert (r.a[0] == -1);
 788   assert (r.a[1] == -1);
 789   assert (r.a[2] == -1);
 790   assert (r.a[3] == 0);
 791   assert (r.a[4] == 0);
 792   assert (r.a[5] == 0);
 793   assert (r.a[6] == -1);
 794   assert (r.a[7] == -1);
 795 }
 796
 797 static void test_pextrh_u (void)
 798 {
 799   uint16x4_encap_t s;
 800   uint16x4_encap_t r;
 801   s.a[0] = 40;
 802   s.a[1] = 41;
 803   s.a[2] = 42;
 804   s.a[3] = 43;
 805   r.v = pextrh_u (s.v, 1);
 806   assert (r.a[0] == 41);
 807   assert (r.a[1] == 0);
 808   assert (r.a[2] == 0);
 809   assert (r.a[3] == 0);
 810 }
 811
 812 static void test_pextrh_s (void)
 813 {
 814   int16x4_encap_t s;
 815   int16x4_encap_t r;
 816   s.a[0] = -40;
 817   s.a[1] = -41;
 818   s.a[2] = -42;
 819   s.a[3] = -43;
 820   r.v = pextrh_s (s.v, 2);
 821   assert (r.a[0] == -42);
 822   assert (r.a[1] == 0);
 823   assert (r.a[2] == 0);
 824   assert (r.a[3] == 0);
 825 }
 826
 827 static void test_pinsrh_0123_u (void)
 828 {
 829   uint16x4_encap_t s, t;
 830   uint16x4_encap_t r;
 831   s.a[0] = 42;
 832   s.a[1] = 0;
 833   s.a[2] = 0;
 834   s.a[3] = 0;
 835   t.a[0] = 0;
 836   t.a[1] = 0;
 837   t.a[2] = 0;
 838   t.a[3] = 0;
 839   r.v = pinsrh_0_u (t.v, s.v);
 840   r.v = pinsrh_1_u (r.v, s.v);
 841   r.v = pinsrh_2_u (r.v, s.v);
 842   r.v = pinsrh_3_u (r.v, s.v);
 843   assert (r.a[0] == 42);
 844   assert (r.a[1] == 42);
 845   assert (r.a[2] == 42);
 846   assert (r.a[3] == 42);
 847 }
 848
 849 static void test_pinsrh_0123_s (void)
 850 {
 851   int16x4_encap_t s, t;
 852   int16x4_encap_t r;
 853   s.a[0] = -42;
 854   s.a[1] = 0;
 855   s.a[2] = 0;
 856   s.a[3] = 0;
 857   t.a[0] = 0;
 858   t.a[1] = 0;
 859   t.a[2] = 0;
 860   t.a[3] = 0;
 861   r.v = pinsrh_0_s (t.v, s.v);
 862   r.v = pinsrh_1_s (r.v, s.v);
 863   r.v = pinsrh_2_s (r.v, s.v);
 864   r.v = pinsrh_3_s (r.v, s.v);
 865   assert (r.a[0] == -42);
 866   assert (r.a[1] == -42);
 867   assert (r.a[2] == -42);
 868   assert (r.a[3] == -42);
 869 }
 870
 871 static void test_pmaddhw (void)
 872 {
 873   int16x4_encap_t s, t;
 874   int32x2_encap_t r;
 875   s.a[0] = -5;
 876   s.a[1] = -4;
 877   s.a[2] = -3;
 878   s.a[3] = -2;
 879   t.a[0] = 10;
 880   t.a[1] = 11;
 881   t.a[2] = 12;
 882   t.a[3] = 13;
 883   r.v = pmaddhw (s.v, t.v);
 884   assert (r.a[0] == (-5*10 + -4*11));
 885   assert (r.a[1] == (-3*12 + -2*13));
 886 }
 887
 888 static void test_pmaxsh (void)
 889 {
 890   int16x4_encap_t s, t;
 891   int16x4_encap_t r;
 892   s.a[0] = -20;
 893   s.a[1] = 40;
 894   s.a[2] = -10;
 895   s.a[3] = 50;
 896   t.a[0] = 20;
 897   t.a[1] = -40;
 898   t.a[2] = 10;
 899   t.a[3] = -50;
 900   r.v = pmaxsh (s.v, t.v);
 901   assert (r.a[0] == 20);
 902   assert (r.a[1] == 40);
 903   assert (r.a[2] == 10);
 904   assert (r.a[3] == 50);
 905 }
 906
 907 static void test_pmaxub (void)
 908 {
 909   uint8x8_encap_t s, t;
 910   uint8x8_encap_t r;
 911   s.a[0] = 10;
 912   s.a[1] = 20;
 913   s.a[2] = 30;
 914   s.a[3] = 40;
 915   s.a[4] = 50;
 916   s.a[5] = 60;
 917   s.a[6] = 70;
 918   s.a[7] = 80;
 919   t.a[0] = 80;
 920   t.a[1] = 70;
 921   t.a[2] = 60;
 922   t.a[3] = 50;
 923   t.a[4] = 40;
 924   t.a[5] = 30;
 925   t.a[6] = 20;
 926   t.a[7] = 10;
 927   r.v = pmaxub (s.v, t.v);
 928   assert (r.a[0] == 80);
 929   assert (r.a[1] == 70);
 930   assert (r.a[2] == 60);
 931   assert (r.a[3] == 50);
 932   assert (r.a[4] == 50);
 933   assert (r.a[5] == 60);
 934   assert (r.a[6] == 70);
 935   assert (r.a[7] == 80);
 936 }
 937
 938 static void test_pminsh (void)
 939 {
 940   int16x4_encap_t s, t;
 941   int16x4_encap_t r;
 942   s.a[0] = -20;
 943   s.a[1] = 40;
 944   s.a[2] = -10;
 945   s.a[3] = 50;
 946   t.a[0] = 20;
 947   t.a[1] = -40;
 948   t.a[2] = 10;
 949   t.a[3] = -50;
 950   r.v = pminsh (s.v, t.v);
 951   assert (r.a[0] == -20);
 952   assert (r.a[1] == -40);
 953   assert (r.a[2] == -10);
 954   assert (r.a[3] == -50);
 955 }
 956
 957 static void test_pminub (void)
 958 {
 959   uint8x8_encap_t s, t;
 960   uint8x8_encap_t r;
 961   s.a[0] = 10;
 962   s.a[1] = 20;
 963   s.a[2] = 30;
 964   s.a[3] = 40;
 965   s.a[4] = 50;
 966   s.a[5] = 60;
 967   s.a[6] = 70;
 968   s.a[7] = 80;
 969   t.a[0] = 80;
 970   t.a[1] = 70;
 971   t.a[2] = 60;
 972   t.a[3] = 50;
 973   t.a[4] = 40;
 974   t.a[5] = 30;
 975   t.a[6] = 20;
 976   t.a[7] = 10;
 977   r.v = pminub (s.v, t.v);
 978   assert (r.a[0] == 10);
 979   assert (r.a[1] == 20);
 980   assert (r.a[2] == 30);
 981   assert (r.a[3] == 40);
 982   assert (r.a[4] == 40);
 983   assert (r.a[5] == 30);
 984   assert (r.a[6] == 20);
 985   assert (r.a[7] == 10);
 986 }
 987
 988 static void test_pmovmskb_u (void)
 989 {
 990   uint8x8_encap_t s;
 991   uint8x8_encap_t r;
 992   s.a[0] = 0xf0;
 993   s.a[1] = 0x40;
 994   s.a[2] = 0xf0;
 995   s.a[3] = 0x40;
 996   s.a[4] = 0xf0;
 997   s.a[5] = 0x40;
 998   s.a[6] = 0xf0;
 999   s.a[7] = 0x40;
1000   r.v = pmovmskb_u (s.v);
1001   assert (r.a[0] == 0x55);
1002   assert (r.a[1] == 0);
1003   assert (r.a[2] == 0);
1004   assert (r.a[3] == 0);
1005   assert (r.a[4] == 0);
1006   assert (r.a[5] == 0);
1007   assert (r.a[6] == 0);
1008   assert (r.a[7] == 0);
1009 }
1010
1011 static void test_pmovmskb_s (void)
1012 {
1013   int8x8_encap_t s;
1014   int8x8_encap_t r;
1015   s.a[0] = -1;
1016   s.a[1] = 1;
1017   s.a[2] = -1;
1018   s.a[3] = 1;
1019   s.a[4] = -1;
1020   s.a[5] = 1;
1021   s.a[6] = -1;
1022   s.a[7] = 1;
1023   r.v = pmovmskb_s (s.v);
1024   assert (r.a[0] == 0x55);
1025   assert (r.a[1] == 0);
1026   assert (r.a[2] == 0);
1027   assert (r.a[3] == 0);
1028   assert (r.a[4] == 0);
1029   assert (r.a[5] == 0);
1030   assert (r.a[6] == 0);
1031   assert (r.a[7] == 0);
1032 }
1033
1034 static void test_pmulhuh (void)
1035 {
1036   uint16x4_encap_t s, t;
1037   uint16x4_encap_t r;
1038   s.a[0] = 0xff00;
1039   s.a[1] = 0xff00;
1040   s.a[2] = 0xff00;
1041   s.a[3] = 0xff00;
1042   t.a[0] = 16;
1043   t.a[1] = 16;
1044   t.a[2] = 16;
1045   t.a[3] = 16;
1046   r.v = pmulhuh (s.v, t.v);
1047   assert (r.a[0] == 0x000f);
1048   assert (r.a[1] == 0x000f);
1049   assert (r.a[2] == 0x000f);
1050   assert (r.a[3] == 0x000f);
1051 }
1052
1053 static void test_pmulhh (void)
1054 {
1055   int16x4_encap_t s, t;
1056   int16x4_encap_t r;
1057   s.a[0] = 0x0ff0;
1058   s.a[1] = 0x0ff0;
1059   s.a[2] = 0x0ff0;
1060   s.a[3] = 0x0ff0;
1061   t.a[0] = -16*16;
1062   t.a[1] = -16*16;
1063   t.a[2] = -16*16;
1064   t.a[3] = -16*16;
1065   r.v = pmulhh (s.v, t.v);
1066   assert (r.a[0] == -16);
1067   assert (r.a[1] == -16);
1068   assert (r.a[2] == -16);
1069   assert (r.a[3] == -16);
1070 }
1071
1072 static void test_pmullh (void)
1073 {
1074   int16x4_encap_t s, t;
1075   int16x4_encap_t r;
1076   s.a[0] = 0x0ff0;
1077   s.a[1] = 0x0ff0;
1078   s.a[2] = 0x0ff0;
1079   s.a[3] = 0x0ff0;
1080   t.a[0] = -16*16;
1081   t.a[1] = -16*16;
1082   t.a[2] = -16*16;
1083   t.a[3] = -16*16;
1084   r.v = pmullh (s.v, t.v);
1085   assert (r.a[0] == 4096);
1086   assert (r.a[1] == 4096);
1087   assert (r.a[2] == 4096);
1088   assert (r.a[3] == 4096);
1089 }
1090
1091 static void test_pmuluw (void)
1092 {
1093   uint32x2_encap_t s, t;
1094   uint64_t r;
1095   s.a[0] = 0xdeadbeef;
1096   s.a[1] = 0;
1097   t.a[0] = 0x0f00baaa;
1098   t.a[1] = 0;
1099   r = pmuluw (s.v, t.v);
1100   assert (r == 0xd0cd08e1d1a70b6ull);
1101 }
1102
1103 static void test_pasubub (void)
1104 {
1105   uint8x8_encap_t s, t;
1106   uint8x8_encap_t r;
1107   s.a[0] = 10;
1108   s.a[1] = 20;
1109   s.a[2] = 30;
1110   s.a[3] = 40;
1111   s.a[4] = 50;
1112   s.a[5] = 60;
1113   s.a[6] = 70;
1114   s.a[7] = 80;
1115   t.a[0] = 80;
1116   t.a[1] = 70;
1117   t.a[2] = 60;
1118   t.a[3] = 50;
1119   t.a[4] = 40;
1120   t.a[5] = 30;
1121   t.a[6] = 20;
1122   t.a[7] = 10;
1123   r.v = pasubub (s.v, t.v);
1124   assert (r.a[0] == 70);
1125   assert (r.a[1] == 50);
1126   assert (r.a[2] == 30);
1127   assert (r.a[3] == 10);
1128   assert (r.a[4] == 10);
1129   assert (r.a[5] == 30);
1130   assert (r.a[6] == 50);
1131   assert (r.a[7] == 70);
1132 }
1133
1134 static void test_biadd (void)
1135 {
1136   uint8x8_encap_t s;
1137   uint16x4_encap_t r;
1138   s.a[0] = 10;
1139   s.a[1] = 20;
1140   s.a[2] = 30;
1141   s.a[3] = 40;
1142   s.a[4] = 50;
1143   s.a[5] = 60;
1144   s.a[6] = 70;
1145   s.a[7] = 80;
1146   r.v = biadd (s.v);
1147   assert (r.a[0] == 360);
1148   assert (r.a[1] == 0);
1149   assert (r.a[2] == 0);
1150   assert (r.a[3] == 0);
1151 }
1152
1153 static void test_psadbh (void)
1154 {
1155   uint8x8_encap_t s, t;
1156   uint16x4_encap_t r;
1157   s.a[0] = 10;
1158   s.a[1] = 20;
1159   s.a[2] = 30;
1160   s.a[3] = 40;
1161   s.a[4] = 50;
1162   s.a[5] = 60;
1163   s.a[6] = 70;
1164   s.a[7] = 80;
1165   t.a[0] = 80;
1166   t.a[1] = 70;
1167   t.a[2] = 60;
1168   t.a[3] = 50;
1169   t.a[4] = 40;
1170   t.a[5] = 30;
1171   t.a[6] = 20;
1172   t.a[7] = 10;
1173   r.v = psadbh (s.v, t.v);
1174   assert (r.a[0] == 0x0140);
1175   assert (r.a[1] == 0);
1176   assert (r.a[2] == 0);
1177   assert (r.a[3] == 0);
1178 }
1179
1180 static void test_pshufh_u (void)
1181 {
1182   uint16x4_encap_t s;
1183   uint16x4_encap_t r;
1184   s.a[0] = 1;
1185   s.a[1] = 2;
1186   s.a[2] = 3;
1187   s.a[3] = 4;
1188   r.a[0] = 0;
1189   r.a[1] = 0;
1190   r.a[2] = 0;
1191   r.a[3] = 0;
1192   r.v = pshufh_u (r.v, s.v, 0xe5);
1193   assert (r.a[0] == 2);
1194   assert (r.a[1] == 2);
1195   assert (r.a[2] == 3);
1196   assert (r.a[3] == 4);
1197 }
1198
1199 static void test_pshufh_s (void)
1200 {
1201   int16x4_encap_t s;
1202   int16x4_encap_t r;
1203   s.a[0] = -1;
1204   s.a[1] = 2;
1205   s.a[2] = -3;
1206   s.a[3] = 4;
1207   r.a[0] = 0;
1208   r.a[1] = 0;
1209   r.a[2] = 0;
1210   r.a[3] = 0;
1211   r.v = pshufh_s (r.v, s.v, 0xe5);
1212   assert (r.a[0] == 2);
1213   assert (r.a[1] == 2);
1214   assert (r.a[2] == -3);
1215   assert (r.a[3] == 4);
1216 }
1217
1218 static void test_psllh_u (void)
1219 {
1220   uint16x4_encap_t s;
1221   uint16x4_encap_t r;
1222   s.a[0] = 0xffff;
1223   s.a[1] = 0xffff;
1224   s.a[2] = 0xffff;
1225   s.a[3] = 0xffff;
1226   r.v = psllh_u (s.v, 1);
1227   assert (r.a[0] == 0xfffe);
1228   assert (r.a[1] == 0xfffe);
1229   assert (r.a[2] == 0xfffe);
1230   assert (r.a[3] == 0xfffe);
1231 }
1232
1233 static void test_psllw_u (void)
1234 {
1235   uint32x2_encap_t s;
1236   uint32x2_encap_t r;
1237   s.a[0] = 0xffffffff;
1238   s.a[1] = 0xffffffff;
1239   r.v = psllw_u (s.v, 2);
1240   assert (r.a[0] == 0xfffffffc);
1241   assert (r.a[1] == 0xfffffffc);
1242 }
1243
1244 static void test_psllh_s (void)
1245 {
1246   int16x4_encap_t s;
1247   int16x4_encap_t r;
1248   s.a[0] = -1;
1249   s.a[1] = -1;
1250   s.a[2] = -1;
1251   s.a[3] = -1;
1252   r.v = psllh_s (s.v, 1);
1253   assert (r.a[0] == -2);
1254   assert (r.a[1] == -2);
1255   assert (r.a[2] == -2);
1256   assert (r.a[3] == -2);
1257 }
1258
1259 static void test_psllw_s (void)
1260 {
1261   int32x2_encap_t s;
1262   int32x2_encap_t r;
1263   s.a[0] = -1;
1264   s.a[1] = -1;
1265   r.v = psllw_s (s.v, 2);
1266   assert (r.a[0] == -4);
1267   assert (r.a[1] == -4);
1268 }
1269
1270 static void test_psrah_u (void)
1271 {
1272   uint16x4_encap_t s;
1273   uint16x4_encap_t r;
1274   s.a[0] = 0xffef;
1275   s.a[1] = 0xffef;
1276   s.a[2] = 0xffef;
1277   s.a[3] = 0xffef;
1278   r.v = psrah_u (s.v, 1);
1279   assert (r.a[0] == 0xfff7);
1280   assert (r.a[1] == 0xfff7);
1281   assert (r.a[2] == 0xfff7);
1282   assert (r.a[3] == 0xfff7);
1283 }
1284
1285 static void test_psraw_u (void)
1286 {
1287   uint32x2_encap_t s;
1288   uint32x2_encap_t r;
1289   s.a[0] = 0xffffffef;
1290   s.a[1] = 0xffffffef;
1291   r.v = psraw_u (s.v, 1);
1292   assert (r.a[0] == 0xfffffff7);
1293   assert (r.a[1] == 0xfffffff7);
1294 }
1295
1296 static void test_psrah_s (void)
1297 {
1298   int16x4_encap_t s;
1299   int16x4_encap_t r;
1300   s.a[0] = -2;
1301   s.a[1] = -2;
1302   s.a[2] = -2;
1303   s.a[3] = -2;
1304   r.v = psrah_s (s.v, 1);
1305   assert (r.a[0] == -1);
1306   assert (r.a[1] == -1);
1307   assert (r.a[2] == -1);
1308   assert (r.a[3] == -1);
1309 }
1310
1311 static void test_psraw_s (void)
1312 {
1313   int32x2_encap_t s;
1314   int32x2_encap_t r;
1315   s.a[0] = -2;
1316   s.a[1] = -2;
1317   r.v = psraw_s (s.v, 1);
1318   assert (r.a[0] == -1);
1319   assert (r.a[1] == -1);
1320 }
1321
1322 static void test_psrlh_u (void)
1323 {
1324   uint16x4_encap_t s;
1325   uint16x4_encap_t r;
1326   s.a[0] = 0xffef;
1327   s.a[1] = 0xffef;
1328   s.a[2] = 0xffef;
1329   s.a[3] = 0xffef;
1330   r.v = psrlh_u (s.v, 1);
1331   assert (r.a[0] == 0x7ff7);
1332   assert (r.a[1] == 0x7ff7);
1333   assert (r.a[2] == 0x7ff7);
1334   assert (r.a[3] == 0x7ff7);
1335 }
1336
1337 static void test_psrlw_u (void)
1338 {
1339   uint32x2_encap_t s;
1340   uint32x2_encap_t r;
1341   s.a[0] = 0xffffffef;
1342   s.a[1] = 0xffffffef;
1343   r.v = psrlw_u (s.v, 1);
1344   assert (r.a[0] == 0x7ffffff7);
1345   assert (r.a[1] == 0x7ffffff7);
1346 }
1347
1348 static void test_psrlh_s (void)
1349 {
1350   int16x4_encap_t s;
1351   int16x4_encap_t r;
1352   s.a[0] = -1;
1353   s.a[1] = -1;
1354   s.a[2] = -1;
1355   s.a[3] = -1;
1356   r.v = psrlh_s (s.v, 1);
1357   assert (r.a[0] == INT16x4_MAX);
1358   assert (r.a[1] == INT16x4_MAX);
1359   assert (r.a[2] == INT16x4_MAX);
1360   assert (r.a[3] == INT16x4_MAX);
1361 }
1362
1363 static void test_psrlw_s (void)
1364 {
1365   int32x2_encap_t s;
1366   int32x2_encap_t r;
1367   s.a[0] = -1;
1368   s.a[1] = -1;
1369   r.v = psrlw_s (s.v, 1);
1370   assert (r.a[0] == INT32x2_MAX);
1371   assert (r.a[1] == INT32x2_MAX);
1372 }
1373
1374 static void test_psubw_u (void)
1375 {
1376   uint32x2_encap_t s, t;
1377   uint32x2_encap_t r;
1378   s.a[0] = 3;
1379   s.a[1] = 4;
1380   t.a[0] = 2;
1381   t.a[1] = 1;
1382   r.v = psubw_u (s.v, t.v);
1383   assert (r.a[0] == 1);
1384   assert (r.a[1] == 3);
1385 }
1386
1387 static void test_psubw_s (void)
1388 {
1389   int32x2_encap_t s, t;
1390   int32x2_encap_t r;
1391   s.a[0] = -2;
1392   s.a[1] = -1;
1393   t.a[0] = 3;
1394   t.a[1] = -4;
1395   r.v = psubw_s (s.v, t.v);
1396   assert (r.a[0] == -5);
1397   assert (r.a[1] == 3);
1398 }
1399
1400 static void test_psubh_u (void)
1401 {
1402   uint16x4_encap_t s, t;
1403   uint16x4_encap_t r;
1404   s.a[0] = 5;
1405   s.a[1] = 6;
1406   s.a[2] = 7;
1407   s.a[3] = 8;
1408   t.a[0] = 1;
1409   t.a[1] = 2;
1410   t.a[2] = 3;
1411   t.a[3] = 4;
1412   r.v = psubh_u (s.v, t.v);
1413   assert (r.a[0] == 4);
1414   assert (r.a[1] == 4);
1415   assert (r.a[2] == 4);
1416   assert (r.a[3] == 4);
1417 }
1418
1419 static void test_psubh_s (void)
1420 {
1421   int16x4_encap_t s, t;
1422   int16x4_encap_t r;
1423   s.a[0] = -10;
1424   s.a[1] = -20;
1425   s.a[2] = -30;
1426   s.a[3] = -40;
1427   t.a[0] = 1;
1428   t.a[1] = 2;
1429   t.a[2] = 3;
1430   t.a[3] = 4;
1431   r.v = psubh_s (s.v, t.v);
1432   assert (r.a[0] == -11);
1433   assert (r.a[1] == -22);
1434   assert (r.a[2] == -33);
1435   assert (r.a[3] == -44);
1436 }
1437
1438 static void test_psubb_u (void)
1439 {
1440   uint8x8_encap_t s, t;
1441   uint8x8_encap_t r;
1442   s.a[0] = 10;
1443   s.a[1] = 11;
1444   s.a[2] = 12;
1445   s.a[3] = 13;
1446   s.a[4] = 14;
1447   s.a[5] = 15;
1448   s.a[6] = 16;
1449   s.a[7] = 17;
1450   t.a[0] = 1;
1451   t.a[1] = 2;
1452   t.a[2] = 3;
1453   t.a[3] = 4;
1454   t.a[4] = 5;
1455   t.a[5] = 6;
1456   t.a[6] = 7;
1457   t.a[7] = 8;
1458   r.v = psubb_u (s.v, t.v);
1459   assert (r.a[0] == 9);
1460   assert (r.a[1] == 9);
1461   assert (r.a[2] == 9);
1462   assert (r.a[3] == 9);
1463   assert (r.a[4] == 9);
1464   assert (r.a[5] == 9);
1465   assert (r.a[6] == 9);
1466   assert (r.a[7] == 9);
1467 }
1468
1469 static void test_psubb_s (void)
1470 {
1471   int8x8_encap_t s, t;
1472   int8x8_encap_t r;
1473   s.a[0] = -10;
1474   s.a[1] = -20;
1475   s.a[2] = -30;
1476   s.a[3] = -40;
1477   s.a[4] = -50;
1478   s.a[5] = -60;
1479   s.a[6] = -70;
1480   s.a[7] = -80;
1481   t.a[0] = 1;
1482   t.a[1] = 2;
1483   t.a[2] = 3;
1484   t.a[3] = 4;
1485   t.a[4] = 5;
1486   t.a[5] = 6;
1487   t.a[6] = 7;
1488   t.a[7] = 8;
1489   r.v = psubb_s (s.v, t.v);
1490   assert (r.a[0] == -11);
1491   assert (r.a[1] == -22);
1492   assert (r.a[2] == -33);
1493   assert (r.a[3] == -44);
1494   assert (r.a[4] == -55);
1495   assert (r.a[5] == -66);
1496   assert (r.a[6] == -77);
1497   assert (r.a[7] == -88);
1498 }
1499
1500 static void test_psubd_u (void)
1501 {
1502   uint64_t d = 789012;
1503   uint64_t e = 123456;
1504   uint64_t r;
1505   r = psubd_u (d, e);
1506   assert (r == 665556);
1507 }
1508
1509 static void test_psubd_s (void)
1510 {
1511   int64_t d = 123456;
1512   int64_t e = -789012;
1513   int64_t r;
1514   r = psubd_s (d, e);
1515   assert (r == 912468);
1516 }
1517
1518 static void test_psubsh (void)
1519 {
1520   int16x4_encap_t s, t;
1521   int16x4_encap_t r;
1522   s.a[0] = -1;
1523   s.a[1] = 0;
1524   s.a[2] = 1;
1525   s.a[3] = 2;
1526   t.a[0] = -INT16x4_MAX;
1527   t.a[1] = -INT16x4_MAX;
1528   t.a[2] = -INT16x4_MAX;
1529   t.a[3] = -INT16x4_MAX;
1530   r.v = psubsh (s.v, t.v);
1531   assert (r.a[0] == INT16x4_MAX - 1);
1532   assert (r.a[1] == INT16x4_MAX);
1533   assert (r.a[2] == INT16x4_MAX);
1534   assert (r.a[3] == INT16x4_MAX);
1535 }
1536
1537 static void test_psubsb (void)
1538 {
1539   int8x8_encap_t s, t;
1540   int8x8_encap_t r;
1541   s.a[0] = -6;
1542   s.a[1] = -5;
1543   s.a[2] = -4;
1544   s.a[3] = -3;
1545   s.a[4] = -2;
1546   s.a[5] = -1;
1547   s.a[6] = 0;
1548   s.a[7] = 1;
1549   t.a[0] = -INT8x8_MAX;
1550   t.a[1] = -INT8x8_MAX;
1551   t.a[2] = -INT8x8_MAX;
1552   t.a[3] = -INT8x8_MAX;
1553   t.a[4] = -INT8x8_MAX;
1554   t.a[5] = -INT8x8_MAX;
1555   t.a[6] = -INT8x8_MAX;
1556   t.a[7] = -INT8x8_MAX;
1557   r.v = psubsb (s.v, t.v);
1558   assert (r.a[0] == INT8x8_MAX - 6);
1559   assert (r.a[1] == INT8x8_MAX - 5);
1560   assert (r.a[2] == INT8x8_MAX - 4);
1561   assert (r.a[3] == INT8x8_MAX - 3);
1562   assert (r.a[4] == INT8x8_MAX - 2);
1563   assert (r.a[5] == INT8x8_MAX - 1);
1564   assert (r.a[6] == INT8x8_MAX);
1565   assert (r.a[7] == INT8x8_MAX);
1566 }
1567
1568 static void test_psubush (void)
1569 {
1570   uint16x4_encap_t s, t;
1571   uint16x4_encap_t r;
1572   s.a[0] = 0;
1573   s.a[1] = 1;
1574   s.a[2] = 2;
1575   s.a[3] = 3;
1576   t.a[0] = 1;
1577   t.a[1] = 1;
1578   t.a[2] = 3;
1579   t.a[3] = 3;
1580   r.v = psubush (s.v, t.v);
1581   assert (r.a[0] == 0);
1582   assert (r.a[1] == 0);
1583   assert (r.a[2] == 0);
1584   assert (r.a[3] == 0);
1585 }
1586
1587 static void test_psubusb (void)
1588 {
1589   uint8x8_encap_t s, t;
1590   uint8x8_encap_t r;
1591   s.a[0] = 0;
1592   s.a[1] = 1;
1593   s.a[2] = 2;
1594   s.a[3] = 3;
1595   s.a[4] = 4;
1596   s.a[5] = 5;
1597   s.a[6] = 6;
1598   s.a[7] = 7;
1599   t.a[0] = 1;
1600   t.a[1] = 1;
1601   t.a[2] = 3;
1602   t.a[3] = 3;
1603   t.a[4] = 5;
1604   t.a[5] = 5;
1605   t.a[6] = 7;
1606   t.a[7] = 7;
1607   r.v = psubusb (s.v, t.v);
1608   assert (r.a[0] == 0);
1609   assert (r.a[1] == 0);
1610   assert (r.a[2] == 0);
1611   assert (r.a[3] == 0);
1612   assert (r.a[4] == 0);
1613   assert (r.a[5] == 0);
1614   assert (r.a[6] == 0);
1615   assert (r.a[7] == 0);
1616 }
1617
1618 static void test_punpckhbh_s (void)
1619 {
1620   int8x8_encap_t s, t;
1621   int8x8_encap_t r;
1622   s.a[0] = -1;
1623   s.a[1] = -3;
1624   s.a[2] = -5;
1625   s.a[3] = -7;
1626   s.a[4] = -9;
1627   s.a[5] = -11;
1628   s.a[6] = -13;
1629   s.a[7] = -15;
1630   t.a[0] = 2;
1631   t.a[1] = 4;
1632   t.a[2] = 6;
1633   t.a[3] = 8;
1634   t.a[4] = 10;
1635   t.a[5] = 12;
1636   t.a[6] = 14;
1637   t.a[7] = 16;
1638   r.v = punpckhbh_s (s.v, t.v);
1639   assert (r.a[0] == -9);
1640   assert (r.a[1] == 10);
1641   assert (r.a[2] == -11);
1642   assert (r.a[3] == 12);
1643   assert (r.a[4] == -13);
1644   assert (r.a[5] == 14);
1645   assert (r.a[6] == -15);
1646   assert (r.a[7] == 16);
1647 }
1648
1649 static void test_punpckhbh_u (void)
1650 {
1651   uint8x8_encap_t s, t;
1652   uint8x8_encap_t r;
1653   s.a[0] = 1;
1654   s.a[1] = 3;
1655   s.a[2] = 5;
1656   s.a[3] = 7;
1657   s.a[4] = 9;
1658   s.a[5] = 11;
1659   s.a[6] = 13;
1660   s.a[7] = 15;
1661   t.a[0] = 2;
1662   t.a[1] = 4;
1663   t.a[2] = 6;
1664   t.a[3] = 8;
1665   t.a[4] = 10;
1666   t.a[5] = 12;
1667   t.a[6] = 14;
1668   t.a[7] = 16;
1669   r.v = punpckhbh_u (s.v, t.v);
1670   assert (r.a[0] == 9);
1671   assert (r.a[1] == 10);
1672   assert (r.a[2] == 11);
1673   assert (r.a[3] == 12);
1674   assert (r.a[4] == 13);
1675   assert (r.a[5] == 14);
1676   assert (r.a[6] == 15);
1677   assert (r.a[7] == 16);
1678 }
1679
1680 static void test_punpckhhw_s (void)
1681 {
1682   int16x4_encap_t s, t;
1683   int16x4_encap_t r;
1684   s.a[0] = -1;
1685   s.a[1] = 3;
1686   s.a[2] = -5;
1687   s.a[3] = 7;
1688   t.a[0] = -2;
1689   t.a[1] = 4;
1690   t.a[2] = -6;
1691   t.a[3] = 8;
1692   r.v = punpckhhw_s (s.v, t.v);
1693   assert (r.a[0] == -5);
1694   assert (r.a[1] == -6);
1695   assert (r.a[2] == 7);
1696   assert (r.a[3] == 8);
1697 }
1698
1699 static void test_punpckhhw_u (void)
1700 {
1701   uint16x4_encap_t s, t;
1702   uint16x4_encap_t r;
1703   s.a[0] = 1;
1704   s.a[1] = 3;
1705   s.a[2] = 5;
1706   s.a[3] = 7;
1707   t.a[0] = 2;
1708   t.a[1] = 4;
1709   t.a[2] = 6;
1710   t.a[3] = 8;
1711   r.v = punpckhhw_u (s.v, t.v);
1712   assert (r.a[0] == 5);
1713   assert (r.a[1] == 6);
1714   assert (r.a[2] == 7);
1715   assert (r.a[3] == 8);
1716 }
1717
1718 static void test_punpckhwd_s (void)
1719 {
1720   int32x2_encap_t s, t;
1721   int32x2_encap_t r;
1722   s.a[0] = 1;
1723   s.a[1] = 3;
1724   t.a[0] = 2;
1725   t.a[1] = -4;
1726   r.v = punpckhwd_s (s.v, t.v);
1727   assert (r.a[0] == 3);
1728   assert (r.a[1] == -4);
1729 }
1730
1731 static void test_punpckhwd_u (void)
1732 {
1733   uint32x2_encap_t s, t;
1734   uint32x2_encap_t r;
1735   s.a[0] = 1;
1736   s.a[1] = 3;
1737   t.a[0] = 2;
1738   t.a[1] = 4;
1739   r.v = punpckhwd_u (s.v, t.v);
1740   assert (r.a[0] == 3);
1741   assert (r.a[1] == 4);
1742 }
1743
1744 static void test_punpcklbh_s (void)
1745 {
1746   int8x8_encap_t s, t;
1747   int8x8_encap_t r;
1748   s.a[0] = -1;
1749   s.a[1] = -3;
1750   s.a[2] = -5;
1751   s.a[3] = -7;
1752   s.a[4] = -9;
1753   s.a[5] = -11;
1754   s.a[6] = -13;
1755   s.a[7] = -15;
1756   t.a[0] = 2;
1757   t.a[1] = 4;
1758   t.a[2] = 6;
1759   t.a[3] = 8;
1760   t.a[4] = 10;
1761   t.a[5] = 12;
1762   t.a[6] = 14;
1763   t.a[7] = 16;
1764   r.v = punpcklbh_s (s.v, t.v);
1765   assert (r.a[0] == -1);
1766   assert (r.a[1] == 2);
1767   assert (r.a[2] == -3);
1768   assert (r.a[3] == 4);
1769   assert (r.a[4] == -5);
1770   assert (r.a[5] == 6);
1771   assert (r.a[6] == -7);
1772   assert (r.a[7] == 8);
1773 }
1774
1775 static void test_punpcklbh_u (void)
1776 {
1777   uint8x8_encap_t s, t;
1778   uint8x8_encap_t r;
1779   s.a[0] = 1;
1780   s.a[1] = 3;
1781   s.a[2] = 5;
1782   s.a[3] = 7;
1783   s.a[4] = 9;
1784   s.a[5] = 11;
1785   s.a[6] = 13;
1786   s.a[7] = 15;
1787   t.a[0] = 2;
1788   t.a[1] = 4;
1789   t.a[2] = 6;
1790   t.a[3] = 8;
1791   t.a[4] = 10;
1792   t.a[5] = 12;
1793   t.a[6] = 14;
1794   t.a[7] = 16;
1795   r.v = punpcklbh_u (s.v, t.v);
1796   assert (r.a[0] == 1);
1797   assert (r.a[1] == 2);
1798   assert (r.a[2] == 3);
1799   assert (r.a[3] == 4);
1800   assert (r.a[4] == 5);
1801   assert (r.a[5] == 6);
1802   assert (r.a[6] == 7);
1803   assert (r.a[7] == 8);
1804 }
1805
1806 static void test_punpcklhw_s (void)
1807 {
1808   int16x4_encap_t s, t;
1809   int16x4_encap_t r;
1810   s.a[0] = -1;
1811   s.a[1] = 3;
1812   s.a[2] = -5;
1813   s.a[3] = 7;
1814   t.a[0] = -2;
1815   t.a[1] = 4;
1816   t.a[2] = -6;
1817   t.a[3] = 8;
1818   r.v = punpcklhw_s (s.v, t.v);
1819   assert (r.a[0] == -1);
1820   assert (r.a[1] == -2);
1821   assert (r.a[2] == 3);
1822   assert (r.a[3] == 4);
1823 }
1824
1825 static void test_punpcklhw_u (void)
1826 {
1827   uint16x4_encap_t s, t;
1828   uint16x4_encap_t r;
1829   s.a[0] = 1;
1830   s.a[1] = 3;
1831   s.a[2] = 5;
1832   s.a[3] = 7;
1833   t.a[0] = 2;
1834   t.a[1] = 4;
1835   t.a[2] = 6;
1836   t.a[3] = 8;
1837   r.v = punpcklhw_u (s.v, t.v);
1838   assert (r.a[0] == 1);
1839   assert (r.a[1] == 2);
1840   assert (r.a[2] == 3);
1841   assert (r.a[3] == 4);
1842 }
1843
1844 static void test_punpcklwd_s (void)
1845 {
1846   int32x2_encap_t s, t;
1847   int32x2_encap_t r;
1848   s.a[0] = 1;
1849   s.a[1] = 3;
1850   t.a[0] = -2;
1851   t.a[1] = 4;
1852   r.v = punpcklwd_s (s.v, t.v);
1853   assert (r.a[0] == 1);
1854   assert (r.a[1] == -2);
1855 }
1856
1857 static void test_punpcklwd_u (void)
1858 {
1859   uint32x2_encap_t s, t;
1860   uint32x2_encap_t r;
1861   s.a[0] = 1;
1862   s.a[1] = 3;
1863   t.a[0] = 2;
1864   t.a[1] = 4;
1865   r.v = punpcklwd_u (s.v, t.v);
1866   assert (r.a[0] == 1);
1867   assert (r.a[1] == 2);
1868 }
1869
1870 int main (void)
1871 {
1872   test_packsswh ();
1873   test_packsshb ();
1874   test_packushb ();
1875   test_paddw_u ();
1876   test_paddw_s ();
1877   test_paddh_u ();
1878   test_paddh_s ();
1879   test_paddb_u ();
1880   test_paddb_s ();
1881   test_paddd_u ();
1882   test_paddd_s ();
1883   test_paddsh ();
1884   test_paddsb ();
1885   test_paddush ();
1886   test_paddusb ();
1887   test_pandn_ud ();
1888   test_pandn_sd ();
1889   test_pandn_uw ();
1890   test_pandn_sw ();
1891   test_pandn_uh ();
1892   test_pandn_sh ();
1893   test_pandn_ub ();
1894   test_pandn_sb ();
1895   test_pavgh ();
1896   test_pavgb ();
1897   test_pcmpeqw_u ();
1898   test_pcmpeqh_u ();
1899   test_pcmpeqb_u ();
1900   test_pcmpeqw_s ();
1901   test_pcmpeqh_s ();
1902   test_pcmpeqb_s ();
1903   test_pcmpgtw_u ();
1904   test_pcmpgth_u ();
1905   test_pcmpgtb_u ();
1906   test_pcmpgtw_s ();
1907   test_pcmpgth_s ();
1908   test_pcmpgtb_s ();
1909   test_pextrh_u ();
1910   test_pextrh_s ();
1911   test_pinsrh_0123_u ();
1912   test_pinsrh_0123_s ();
1913   test_pmaddhw ();
1914   test_pmaxsh ();
1915   test_pmaxub ();
1916   test_pminsh ();
1917   test_pminub ();
1918   test_pmovmskb_u ();
1919   test_pmovmskb_s ();
1920   test_pmulhuh ();
1921   test_pmulhh ();
1922   test_pmullh ();
1923   test_pmuluw ();
1924   test_pasubub ();
1925   test_biadd ();
1926   test_psadbh ();
1927   test_pshufh_u ();
1928   test_pshufh_s ();
1929   test_psllh_u ();
1930   test_psllw_u ();
1931   test_psllh_s ();
1932   test_psllw_s ();
1933   test_psrah_u ();
1934   test_psraw_u ();
1935   test_psrah_s ();
1936   test_psraw_s ();
1937   test_psrlh_u ();
1938   test_psrlw_u ();
1939   test_psrlh_s ();
1940   test_psrlw_s ();
1941   test_psubw_u ();
1942   test_psubw_s ();
1943   test_psubh_u ();
1944   test_psubh_s ();
1945   test_psubb_u ();
1946   test_psubb_s ();
1947   test_psubd_u ();
1948   test_psubd_s ();
1949   test_psubsh ();
1950   test_psubsb ();
1951   test_psubush ();
1952   test_psubusb ();
1953   test_punpckhbh_s ();
1954   test_punpckhbh_u ();
1955   test_punpckhhw_s ();
1956   test_punpckhhw_u ();
1957   test_punpckhwd_s ();
1958   test_punpckhwd_u ();
1959   test_punpcklbh_s ();
1960   test_punpcklbh_u ();
1961   test_punpcklhw_s ();
1962   test_punpcklhw_u ();
1963   test_punpcklwd_s ();
1964   test_punpcklwd_u ();
1965   return 0;
1966 }