src/amd/compiler/tests/test_to_hw_instr.cpp

   1 /*
   2  * Copyright © 2020 Valve Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24 #include "helpers.h"
  25
  26 using namespace aco;
  27
  28 BEGIN_TEST(to_hw_instr.swap_subdword)
  29    PhysReg v0_lo{256};
  30    PhysReg v0_hi{256};
  31    PhysReg v0_b1{256};
  32    PhysReg v0_b3{256};
  33    PhysReg v1_lo{257};
  34    PhysReg v1_hi{257};
  35    PhysReg v1_b1{257};
  36    PhysReg v1_b3{257};
  37    PhysReg v2_lo{258};
  38    PhysReg v3_lo{259};
  39    v0_hi.reg_b += 2;
  40    v1_hi.reg_b += 2;
  41    v0_b1.reg_b += 1;
  42    v1_b1.reg_b += 1;
  43    v0_b3.reg_b += 3;
  44    v1_b3.reg_b += 3;
  45
  46    for (unsigned i = GFX6; i <= GFX7; i++) {
  47       if (!setup_cs(NULL, (chip_class)i))
  48          continue;
  49
  50       //~gfx[67]>>  p_unit_test 0
  51       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
  52       //~gfx[67]! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
  53       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
  54       bld.pseudo(aco_opcode::p_unit_test, Operand(0u));
  55       bld.pseudo(aco_opcode::p_parallelcopy,
  56                  Definition(v0_lo, v2b), Definition(v1_lo, v2b),
  57                  Operand(v1_lo, v2b), Operand(v0_lo, v2b));
  58
  59       //~gfx[67]! p_unit_test 1
  60       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:16]
  61       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xffff, %0:v[1][0:16]
  62       //~gfx[67]! v1: %0:v[1] = v_cvt_pk_u16_u32 %0:v[1][0:16], %0:v[0][0:16]
  63       //~gfx[67]! v1: %0:v[0] = v_mov_b32 %0:v[1]
  64       bld.pseudo(aco_opcode::p_unit_test, Operand(1u));
  65       bld.pseudo(aco_opcode::p_create_vector,
  66                  Definition(v0_lo, v1),
  67                  Operand(v1_lo, v2b), Operand(v0_lo, v2b));
  68
  69       //~gfx[67]! p_unit_test 2
  70       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:16]
  71       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xffff, %0:v[1][0:16]
  72       //~gfx[67]! v1: %0:v[1] = v_cvt_pk_u16_u32 %0:v[1][0:16], %0:v[0][0:16]
  73       //~gfx[67]! v1: %0:v[0] = v_mov_b32 %0:v[1]
  74       //~gfx[67]! v2b: %0:v[1][0:16] = v_mov_b32 %0:v[2][0:16]
  75       bld.pseudo(aco_opcode::p_unit_test, Operand(2u));
  76       bld.pseudo(aco_opcode::p_create_vector,
  77                  Definition(v0_lo, v6b), Operand(v1_lo, v2b),
  78                  Operand(v0_lo, v2b), Operand(v2_lo, v2b));
  79
  80       //~gfx[67]! p_unit_test 3
  81       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:16]
  82       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xffff, %0:v[1][0:16]
  83       //~gfx[67]! v1: %0:v[1] = v_cvt_pk_u16_u32 %0:v[1][0:16], %0:v[0][0:16]
  84       //~gfx[67]! v1: %0:v[0] = v_mov_b32 %0:v[1]
  85       //~gfx[67]! v1: %0:v[2] = v_and_b32 0xffff, %0:v[2][0:16]
  86       //~gfx[67]! v1: %0:v[3] = v_and_b32 0xffff, %0:v[3][0:16]
  87       //~gfx[67]! v1: %0:v[1] = v_cvt_pk_u16_u32 %0:v[2][0:16], %0:v[3][0:16]
  88       bld.pseudo(aco_opcode::p_unit_test, Operand(3u));
  89       bld.pseudo(aco_opcode::p_create_vector,
  90                  Definition(v0_lo, v2),
  91                  Operand(v1_lo, v2b), Operand(v0_lo, v2b),
  92                  Operand(v2_lo, v2b), Operand(v3_lo, v2b));
  93
  94       //~gfx[67]! p_unit_test 4
  95       //~gfx[67]! v1: %0:v[2] = v_and_b32 0xffff, %0:v[2][0:16]
  96       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xffff, %0:v[1][0:16]
  97       //~gfx[67]! v1: %0:v[1] = v_cvt_pk_u16_u32 %0:v[1][0:16], %0:v[2][0:16]
  98       //~gfx[67]! v1: %0:v[3] = v_and_b32 0xffff, %0:v[3][0:16]
  99       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:16]
 100       //~gfx[67]! v1: %0:v[0] = v_cvt_pk_u16_u32 %0:v[0][0:16], %0:v[3][0:16]
 101       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 102       //~gfx[67]! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 103       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 104       bld.pseudo(aco_opcode::p_unit_test, Operand(4u));
 105       bld.pseudo(aco_opcode::p_create_vector,
 106                  Definition(v0_lo, v2),
 107                  Operand(v1_lo, v2b), Operand(v2_lo, v2b),
 108                  Operand(v0_lo, v2b), Operand(v3_lo, v2b));
 109
 110       //~gfx[67]! p_unit_test 5
 111       //~gfx[67]! v2b: %0:v[1][0:16] = v_mov_b32 %0:v[0][0:16]
 112       //~gfx[67]! v2b: %0:v[0][0:16] = v_lshrrev_b32 16, %0:v[1][16:32]
 113       bld.pseudo(aco_opcode::p_unit_test, Operand(5u));
 114       bld.pseudo(aco_opcode::p_split_vector,
 115                  Definition(v1_lo, v2b), Definition(v0_lo, v2b),
 116                  Operand(v0_lo, v1));
 117
 118       //~gfx[67]! p_unit_test 6
 119       //~gfx[67]! v2b: %0:v[2][0:16] = v_mov_b32 %0:v[1][0:16]
 120       //~gfx[67]! v2b: %0:v[1][0:16] = v_mov_b32 %0:v[0][0:16]
 121       //~gfx[67]! v2b: %0:v[0][0:16] = v_lshrrev_b32 16, %0:v[1][16:32]
 122       bld.pseudo(aco_opcode::p_unit_test, Operand(6u));
 123       bld.pseudo(aco_opcode::p_split_vector,
 124                  Definition(v1_lo, v2b), Definition(v0_lo, v2b),
 125                  Definition(v2_lo, v2b), Operand(v0_lo, v6b));
 126
 127       //~gfx[67]! p_unit_test 7
 128       //~gfx[67]! v2b: %0:v[2][0:16] = v_mov_b32 %0:v[1][0:16]
 129       //~gfx[67]! v2b: %0:v[1][0:16] = v_mov_b32 %0:v[0][0:16]
 130       //~gfx[67]! v2b: %0:v[0][0:16] = v_lshrrev_b32 16, %0:v[1][16:32]
 131       //~gfx[67]! v2b: %0:v[3][0:16] = v_lshrrev_b32 16, %0:v[2][16:32]
 132       bld.pseudo(aco_opcode::p_unit_test, Operand(7u));
 133       bld.pseudo(aco_opcode::p_split_vector,
 134                  Definition(v1_lo, v2b), Definition(v0_lo, v2b),
 135                  Definition(v2_lo, v2b), Definition(v3_lo, v2b),
 136                  Operand(v0_lo, v2));
 137
 138       //~gfx[67]! p_unit_test 8
 139       //~gfx[67]! v2b: %0:v[2][0:16] = v_lshrrev_b32 16, %0:v[0][16:32]
 140       //~gfx[67]! v2b: %0:v[3][0:16] = v_lshrrev_b32 16, %0:v[1][16:32]
 141       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 142       //~gfx[67]! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 143       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 144       bld.pseudo(aco_opcode::p_unit_test, Operand(8u));
 145       bld.pseudo(aco_opcode::p_split_vector,
 146                  Definition(v1_lo, v2b), Definition(v2_lo, v2b),
 147                  Definition(v0_lo, v2b), Definition(v3_lo, v2b),
 148                  Operand(v0_lo, v2));
 149
 150       //~gfx[67]! p_unit_test 9
 151       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 152       //~gfx[67]! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 153       //~gfx[67]! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 154       bld.pseudo(aco_opcode::p_unit_test, Operand(9u));
 155       bld.pseudo(aco_opcode::p_parallelcopy,
 156                  Definition(v0_lo, v1b), Definition(v1_lo, v1b),
 157                  Operand(v1_lo, v1b), Operand(v0_lo, v1b));
 158
 159       //~gfx[67]! p_unit_test 10
 160       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xff, %0:v[1][0:8]
 161       //~gfx[67]! v1: %0:v[0] = v_lshlrev_b32 8, %0:v[0][0:8]
 162       //~gfx[67]! v1: %0:v[1] = v_or_b32 %0:v[1][0:8], %0:v[0][0:8]
 163       //~gfx[67]! v1: %0:v[0] = v_lshrrev_b32 8, %0:v[0][0:8]
 164       //~gfx[67]! v2b: %0:v[0][0:16] = v_mov_b32 %0:v[1][0:16]
 165       bld.pseudo(aco_opcode::p_unit_test, Operand(10u));
 166       bld.pseudo(aco_opcode::p_create_vector,
 167                  Definition(v0_lo, v2b),
 168                  Operand(v1_lo, v1b), Operand(v0_lo, v1b));
 169
 170       //~gfx[67]! p_unit_test 11
 171       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xff, %0:v[1][0:8]
 172       //~gfx[67]! v1: %0:v[0] = v_lshlrev_b32 8, %0:v[0][0:8]
 173       //~gfx[67]! v1: %0:v[1] = v_or_b32 %0:v[1][0:8], %0:v[0][0:8]
 174       //~gfx[67]! v1: %0:v[0] = v_lshrrev_b32 8, %0:v[0][0:8]
 175       //~gfx[67]! v2b: %0:v[0][0:16] = v_mov_b32 %0:v[1][0:16]
 176       //~gfx[67]! v1: %0:v[2] = v_and_b32 0xffff, %0:v[2][0:8]
 177       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:16]
 178       //~gfx[67]! v1: %0:v[0] = v_cvt_pk_u16_u32 %0:v[0][0:16], %0:v[2][0:8]
 179       bld.pseudo(aco_opcode::p_unit_test, Operand(11u));
 180       bld.pseudo(aco_opcode::p_create_vector,
 181                  Definition(v0_lo, v3b), Operand(v1_lo, v1b),
 182                  Operand(v0_lo, v1b), Operand(v2_lo, v1b));
 183
 184       //~gfx[67]! p_unit_test 12
 185       //~gfx[67]! v1: %0:v[1] = v_and_b32 0xff, %0:v[1][0:8]
 186       //~gfx[67]! v1: %0:v[0] = v_lshlrev_b32 8, %0:v[0][0:8]
 187       //~gfx[67]! v1: %0:v[1] = v_or_b32 %0:v[1][0:8], %0:v[0][0:8]
 188       //~gfx[67]! v1: %0:v[0] = v_lshrrev_b32 8, %0:v[0][0:8]
 189       //~gfx[67]! v2b: %0:v[0][0:16] = v_mov_b32 %0:v[1][0:16]
 190       //~gfx[67]! v1: %0:v[2] = v_and_b32 0xffff, %0:v[2][0:8]
 191       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:16]
 192       //~gfx[67]! v1: %0:v[0] = v_cvt_pk_u16_u32 %0:v[0][0:16], %0:v[2][0:8]
 193       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffffff, %0:v[0][0:8]
 194       //~gfx[67]! v1: %0:v[3] = v_lshlrev_b32 24, %0:v[3][0:8]
 195       //~gfx[67]! v1: %0:v[0] = v_or_b32 %0:v[0][0:8], %0:v[3][0:8]
 196       //~gfx[67]! v1: %0:v[3] = v_lshrrev_b32 24, %0:v[3][0:8]
 197       bld.pseudo(aco_opcode::p_unit_test, Operand(12u));
 198       bld.pseudo(aco_opcode::p_create_vector,
 199                  Definition(v0_lo, v1),
 200                  Operand(v1_lo, v1b), Operand(v0_lo, v1b),
 201                  Operand(v2_lo, v1b), Operand(v3_lo, v1b));
 202
 203       //~gfx[67]! p_unit_test 13
 204       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xff, %0:v[0][0:8]
 205       //~gfx[67]! v1: %0:v[0] = v_mul_u32_u24 0x101, %0:v[0][0:8]
 206       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffff, %0:v[0][0:8]
 207       //~gfx[67]! v1: %0:v[0] = v_cvt_pk_u16_u32 %0:v[0][0:16], %0:v[0][0:8]
 208       //~gfx[67]! v1: %0:v[0] = v_and_b32 0xffffff, %0:v[0][0:8]
 209       //~gfx[67]! s1: %0:m0 = s_mov_b32 0x1000001
 210       //~gfx[67]! v1: %0:v[0] = v_mul_lo_u32 %0:m0, %0:v[0][0:8]
 211       bld.pseudo(aco_opcode::p_unit_test, Operand(13u));
 212       Instruction* pseudo = bld.pseudo(aco_opcode::p_create_vector,
 213                                        Definition(v0_lo, v1),
 214                                        Operand(v0_lo, v1b), Operand(v0_lo, v1b),
 215                                        Operand(v0_lo, v1b), Operand(v0_lo, v1b));
 216       static_cast<Pseudo_instruction*>(pseudo)->scratch_sgpr = m0;
 217
 218       //~gfx[67]! p_unit_test 14
 219       //~gfx[67]! v1b: %0:v[1][0:8] = v_mov_b32 %0:v[0][0:8]
 220       //~gfx[67]! v1b: %0:v[0][0:8] = v_lshrrev_b32 8, %0:v[1][8:16]
 221       bld.pseudo(aco_opcode::p_unit_test, Operand(14u));
 222       bld.pseudo(aco_opcode::p_split_vector,
 223                  Definition(v1_lo, v1b), Definition(v0_lo, v1b),
 224                  Operand(v0_lo, v2b));
 225
 226       //~gfx[67]! p_unit_test 15
 227       //~gfx[67]! v1b: %0:v[1][0:8] = v_mov_b32 %0:v[0][0:8]
 228       //~gfx[67]! v1b: %0:v[0][0:8] = v_lshrrev_b32 8, %0:v[1][8:16]
 229       //~gfx[67]! v1b: %0:v[2][0:8] = v_lshrrev_b32 16, %0:v[1][16:24]
 230       //~gfx[67]! v1b: %0:v[3][0:8] = v_lshrrev_b32 24, %0:v[1][24:32]
 231       bld.pseudo(aco_opcode::p_unit_test, Operand(15u));
 232       bld.pseudo(aco_opcode::p_split_vector,
 233                  Definition(v1_lo, v1b), Definition(v0_lo, v1b),
 234                  Definition(v2_lo, v1b), Definition(v3_lo, v1b),
 235                  Operand(v0_lo, v1));
 236
 237       //~gfx[67]! s_endpgm
 238
 239       finish_to_hw_instr_test();
 240    }
 241
 242    for (unsigned i = GFX8; i <= GFX9; i++) {
 243       if (!setup_cs(NULL, (chip_class)i))
 244          continue;
 245
 246       //~gfx[89]>> p_unit_test 0
 247       //~gfx8! v2b: %0:v[0][16:32] = v_xor_b32 %0:v[0][16:32], %0:v[0][0:16] dst_preserve
 248       //~gfx8! v2b: %0:v[0][0:16] = v_xor_b32 %0:v[0][16:32], %0:v[0][0:16] dst_preserve
 249       //~gfx8! v2b: %0:v[0][16:32] = v_xor_b32 %0:v[0][16:32], %0:v[0][0:16] dst_preserve
 250       //~gfx9! v1: %0:v[0] = v_pk_add_u16 %0:v[0].yx, 0
 251       bld.pseudo(aco_opcode::p_unit_test, Operand(0u));
 252       bld.pseudo(aco_opcode::p_parallelcopy,
 253                  Definition(v0_lo, v2b), Definition(v0_hi, v2b),
 254                  Operand(v0_hi, v2b), Operand(v0_lo, v2b));
 255
 256       //~gfx[89]! p_unit_test 1
 257       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 258       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 259       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 260       //~gfx9! v1: %0:v[0],  v1: %0:v[1] = v_swap_b32 %0:v[1], %0:v[0]
 261       //~gfx[89]! v2b: %0:v[1][16:32] = v_mov_b32 %0:v[0][16:32] dst_preserve
 262       bld.pseudo(aco_opcode::p_unit_test, Operand(1u));
 263       bld.pseudo(aco_opcode::p_parallelcopy,
 264                  Definition(v0_lo, v1), Definition(v1_lo, v2b),
 265                  Operand(v1_lo, v1), Operand(v0_lo, v2b));
 266
 267       //~gfx[89]! p_unit_test 2
 268       //~gfx[89]! v2b: %0:v[0][16:32] = v_mov_b32 %0:v[1][16:32] dst_preserve
 269       //~gfx[89]! v2b: %0:v[1][16:32] = v_mov_b32 %0:v[0][0:16] dst_preserve
 270       //~gfx[89]! v2b: %0:v[1][0:16] = v_xor_b32 %0:v[1][0:16], %0:v[0][0:16] dst_preserve
 271       //~gfx[89]! v2b: %0:v[0][0:16] = v_xor_b32 %0:v[1][0:16], %0:v[0][0:16] dst_preserve
 272       //~gfx[89]! v2b: %0:v[1][0:16] = v_xor_b32 %0:v[1][0:16], %0:v[0][0:16] dst_preserve
 273       bld.pseudo(aco_opcode::p_unit_test, Operand(2u));
 274       bld.pseudo(aco_opcode::p_parallelcopy,
 275                  Definition(v0_lo, v1), Definition(v1_lo, v2b), Definition(v1_hi, v2b),
 276                  Operand(v1_lo, v1), Operand(v0_lo, v2b), Operand(v0_lo, v2b));
 277
 278       //~gfx[89]! p_unit_test 3
 279       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 280       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 281       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 282       //~gfx9! v1: %0:v[0],  v1: %0:v[1] = v_swap_b32 %0:v[1], %0:v[0]
 283       //~gfx[89]! v2b: %0:v[1][0:16] = v_mov_b32 %0:v[0][0:16] dst_preserve
 284       //~gfx[89]! v1b: %0:v[1][16:24] = v_mov_b32 %0:v[0][16:24] dst_preserve
 285       bld.pseudo(aco_opcode::p_unit_test, Operand(3u));
 286       bld.pseudo(aco_opcode::p_parallelcopy,
 287                  Definition(v0_lo, v1), Definition(v1_b3, v1b),
 288                  Operand(v1_lo, v1), Operand(v0_b3, v1b));
 289
 290       //~gfx[89]! p_unit_test 4
 291       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 292       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 293       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 294       //~gfx9! v1: %0:v[0],  v1: %0:v[1] = v_swap_b32 %0:v[1], %0:v[0]
 295       //~gfx[89]! v1b: %0:v[1][8:16] = v_mov_b32 %0:v[0][8:16] dst_preserve
 296       //~gfx[89]! v2b: %0:v[1][16:32] = v_mov_b32 %0:v[0][16:32] dst_preserve
 297       bld.pseudo(aco_opcode::p_unit_test, Operand(4u));
 298       bld.pseudo(aco_opcode::p_parallelcopy,
 299                  Definition(v0_lo, v1), Definition(v1_lo, v1b),
 300                  Operand(v1_lo, v1), Operand(v0_lo, v1b));
 301
 302       //~gfx[89]! p_unit_test 5
 303       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[0], %0:v[1]
 304       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[0], %0:v[1]
 305       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[0], %0:v[1]
 306       //~gfx9! v1: %0:v[1],  v1: %0:v[0] = v_swap_b32 %0:v[0], %0:v[1]
 307       //~gfx[89]! v1b: %0:v[0][8:16] = v_mov_b32 %0:v[1][8:16] dst_preserve
 308       //~gfx[89]! v1b: %0:v[0][24:32] = v_mov_b32 %0:v[1][24:32] dst_preserve
 309       bld.pseudo(aco_opcode::p_unit_test, Operand(5u));
 310       bld.pseudo(aco_opcode::p_parallelcopy,
 311                  Definition(v0_lo, v1b), Definition(v0_hi, v1b), Definition(v1_lo, v1),
 312                  Operand(v1_lo, v1b), Operand(v1_hi, v1b), Operand(v0_lo, v1));
 313
 314       //~gfx[89]! p_unit_test 6
 315       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 316       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 317       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 318       //~gfx9! v1: %0:v[0],  v1: %0:v[1] = v_swap_b32 %0:v[1], %0:v[0]
 319       bld.pseudo(aco_opcode::p_unit_test, Operand(6u));
 320       bld.pseudo(aco_opcode::p_parallelcopy,
 321                  Definition(v0_lo, v2b), Definition(v0_hi, v2b), Definition(v1_lo, v1),
 322                  Operand(v1_lo, v2b), Operand(v1_hi, v2b), Operand(v0_lo, v1));
 323
 324       //~gfx[89]! p_unit_test 7
 325       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[0], %0:v[1]
 326       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[0], %0:v[1]
 327       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[0], %0:v[1]
 328       //~gfx8! v2b: %0:v[0][16:32] = v_xor_b32 %0:v[0][16:32], %0:v[0][0:16] dst_preserve
 329       //~gfx8! v2b: %0:v[0][0:16] = v_xor_b32 %0:v[0][16:32], %0:v[0][0:16] dst_preserve
 330       //~gfx8! v2b: %0:v[0][16:32] = v_xor_b32 %0:v[0][16:32], %0:v[0][0:16] dst_preserve
 331       //~gfx9! v1: %0:v[1],  v1: %0:v[0] = v_swap_b32 %0:v[0], %0:v[1]
 332       //~gfx9! v1: %0:v[0] = v_pk_add_u16 %0:v[0].yx, 0
 333       bld.pseudo(aco_opcode::p_unit_test, Operand(7u));
 334       bld.pseudo(aco_opcode::p_parallelcopy,
 335                  Definition(v0_lo, v2b), Definition(v0_hi, v2b), Definition(v1_lo, v1),
 336                  Operand(v1_hi, v2b), Operand(v1_lo, v2b), Operand(v0_lo, v1));
 337
 338       //~gfx[89]! p_unit_test 8
 339       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 340       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 341       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 342       //~gfx9! v1: %0:v[0],  v1: %0:v[1] = v_swap_b32 %0:v[1], %0:v[0]
 343       //~gfx[89]! v1b: %0:v[1][24:32] = v_xor_b32 %0:v[1][24:32], %0:v[0][24:32] dst_preserve
 344       //~gfx[89]! v1b: %0:v[0][24:32] = v_xor_b32 %0:v[1][24:32], %0:v[0][24:32] dst_preserve
 345       //~gfx[89]! v1b: %0:v[1][24:32] = v_xor_b32 %0:v[1][24:32], %0:v[0][24:32] dst_preserve
 346       bld.pseudo(aco_opcode::p_unit_test, Operand(8u));
 347       bld.pseudo(aco_opcode::p_parallelcopy,
 348                  Definition(v0_lo, v3b), Definition(v1_lo, v3b),
 349                  Operand(v1_lo, v3b), Operand(v0_lo, v3b));
 350
 351       //~gfx[89]! p_unit_test 9
 352       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 353       //~gfx8! v1: %0:v[0] = v_xor_b32 %0:v[1], %0:v[0]
 354       //~gfx8! v1: %0:v[1] = v_xor_b32 %0:v[1], %0:v[0]
 355       //~gfx9! v1: %0:v[0],  v1: %0:v[1] = v_swap_b32 %0:v[1], %0:v[0]
 356       //~gfx[89]! v1b: %0:v[1][24:32] = v_mov_b32 %0:v[0][24:32] dst_preserve
 357       bld.pseudo(aco_opcode::p_unit_test, Operand(9u));
 358       bld.pseudo(aco_opcode::p_parallelcopy,
 359                  Definition(v0_lo, v3b), Definition(v1_lo, v3b), Definition(v0_b3, v1b),
 360                  Operand(v1_lo, v3b), Operand(v0_lo, v3b), Operand(v1_b3, v1b));
 361
 362       //~gfx[89]! p_unit_test 10
 363       //~gfx[89]! v1b: %0:v[1][8:16] = v_xor_b32 %0:v[1][8:16], %0:v[0][8:16] dst_preserve
 364       //~gfx[89]! v1b: %0:v[0][8:16] = v_xor_b32 %0:v[1][8:16], %0:v[0][8:16] dst_preserve
 365       //~gfx[89]! v1b: %0:v[1][8:16] = v_xor_b32 %0:v[1][8:16], %0:v[0][8:16] dst_preserve
 366       //~gfx[89]! v1b: %0:v[1][16:24] = v_xor_b32 %0:v[1][16:24], %0:v[0][16:24] dst_preserve
 367       //~gfx[89]! v1b: %0:v[0][16:24] = v_xor_b32 %0:v[1][16:24], %0:v[0][16:24] dst_preserve
 368       //~gfx[89]! v1b: %0:v[1][16:24] = v_xor_b32 %0:v[1][16:24], %0:v[0][16:24] dst_preserve
 369       bld.pseudo(aco_opcode::p_unit_test, Operand(10u));
 370       bld.pseudo(aco_opcode::p_parallelcopy,
 371                  Definition(v0_b1, v2b), Definition(v1_b1, v2b),
 372                  Operand(v1_b1, v2b), Operand(v0_b1, v2b));
 373
 374       //~gfx[89]! p_unit_test 11
 375       //~gfx[89]! v2b: %0:v[1][0:16] = v_mov_b32 %0:v[0][16:32] dst_preserve
 376       //~gfx[89]! v1: %0:v[0] = v_mov_b32 42
 377       bld.pseudo(aco_opcode::p_unit_test, Operand(11u));
 378       bld.pseudo(aco_opcode::p_parallelcopy,
 379                  Definition(v0_lo, v1), Definition(v1_lo, v2b),
 380                  Operand(42u), Operand(v0_hi, v2b));
 381
 382       //~gfx[89]! s_endpgm
 383
 384       finish_to_hw_instr_test();
 385    }
 386 END_TEST