llvm.org GIT mirror llvm / release_90 lib / Target / AMDGPU / AMDGPUInstructionSelector.cpp
release_90

Tree @release_90 (Download .tar.gz)

AMDGPUInstructionSelector.cpp @release_90raw · history · blame

   1
   2
   3
   4
   5
   6
   7
   8
   9
  10
  11
  12
  13
  14
  15
  16
  17
  18
  19
  20
  21
  22
  23
  24
  25
  26
  27
  28
  29
  30
  31
  32
  33
  34
  35
  36
  37
  38
  39
  40
  41
  42
  43
  44
  45
  46
  47
  48
  49
  50
  51
  52
  53
  54
  55
  56
  57
  58
  59
  60
  61
  62
  63
  64
  65
  66
  67
  68
  69
  70
  71
  72
  73
  74
  75
  76
  77
  78
  79
  80
  81
  82
  83
  84
  85
  86
  87
  88
  89
  90
  91
  92
  93
  94
  95
  96
  97
  98
  99
 100
 101
 102
 103
 104
 105
 106
 107
 108
 109
 110
 111
 112
 113
 114
 115
 116
 117
 118
 119
 120
 121
 122
 123
 124
 125
 126
 127
 128
 129
 130
 131
 132
 133
 134
 135
 136
 137
 138
 139
 140
 141
 142
 143
 144
 145
 146
 147
 148
 149
 150
 151
 152
 153
 154
 155
 156
 157
 158
 159
 160
 161
 162
 163
 164
 165
 166
 167
 168
 169
 170
 171
 172
 173
 174
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
//===- AMDGPUInstructionSelector.cpp ----------------------------*- C++ -*-==//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements the targeting of the InstructionSelector class for
/// AMDGPU.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//

#include "AMDGPUInstructionSelector.h"
#include "AMDGPUInstrInfo.h"
#include "AMDGPURegisterBankInfo.h"
#include "AMDGPURegisterInfo.h"
#include "AMDGPUSubtarget.h"
#include "AMDGPUTargetMachine.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"

#define DEBUG_TYPE "amdgpu-isel"

using namespace llvm;
using namespace MIPatternMatch;

#define GET_GLOBALISEL_IMPL
#define AMDGPUSubtarget GCNSubtarget
#include "AMDGPUGenGlobalISel.inc"
#undef GET_GLOBALISEL_IMPL
#undef AMDGPUSubtarget

AMDGPUInstructionSelector::AMDGPUInstructionSelector(
    const GCNSubtarget &STI, const AMDGPURegisterBankInfo &RBI,
    const AMDGPUTargetMachine &TM)
    : InstructionSelector(), TII(*STI.getInstrInfo()),
      TRI(*STI.getRegisterInfo()), RBI(RBI), TM(TM),
      STI(STI),
      EnableLateStructurizeCFG(AMDGPUTargetMachine::EnableLateStructurizeCFG),
#define GET_GLOBALISEL_PREDICATES_INIT
#include "AMDGPUGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
#include "AMDGPUGenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_INIT
{
}

const char *AMDGPUInstructionSelector::getName() { return DEBUG_TYPE; }

static bool isSCC(Register Reg, const MachineRegisterInfo &MRI) {
  if (TargetRegisterInfo::isPhysicalRegister(Reg))
    return Reg == AMDGPU::SCC;

  auto &RegClassOrBank = MRI.getRegClassOrRegBank(Reg);
  const TargetRegisterClass *RC =
      RegClassOrBank.dyn_cast<const TargetRegisterClass*>();
  if (RC) {
    // FIXME: This is ambiguous for wave32. This could be SCC or VCC, but the
    // context of the register bank has been lost.
    if (RC->getID() != AMDGPU::SReg_32_XM0RegClassID)
      return false;
    const LLT Ty = MRI.getType(Reg);
    return Ty.isValid() && Ty.getSizeInBits() == 1;
  }

  const RegisterBank *RB = RegClassOrBank.get<const RegisterBank *>();
  return RB->getID() == AMDGPU::SCCRegBankID;
}

bool AMDGPUInstructionSelector::isVCC(Register Reg,
                                      const MachineRegisterInfo &MRI) const {
  if (TargetRegisterInfo::isPhysicalRegister(Reg))
    return Reg == TRI.getVCC();

  auto &RegClassOrBank = MRI.getRegClassOrRegBank(Reg);
  const TargetRegisterClass *RC =
      RegClassOrBank.dyn_cast<const TargetRegisterClass*>();
  if (RC) {
    const LLT Ty = MRI.getType(Reg);
    return RC->hasSuperClassEq(TRI.getBoolRC()) &&
           Ty.isValid() && Ty.getSizeInBits() == 1;
  }

  const RegisterBank *RB = RegClassOrBank.get<const RegisterBank *>();
  return RB->getID() == AMDGPU::VCCRegBankID;
}

bool AMDGPUInstructionSelector::selectCOPY(MachineInstr &I) const {
  const DebugLoc &DL = I.getDebugLoc();
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  I.setDesc(TII.get(TargetOpcode::COPY));

  const MachineOperand &Src = I.getOperand(1);
  MachineOperand &Dst = I.getOperand(0);
  Register DstReg = Dst.getReg();
  Register SrcReg = Src.getReg();

  if (isVCC(DstReg, MRI)) {
    if (SrcReg == AMDGPU::SCC) {
      const TargetRegisterClass *RC
        = TRI.getConstrainedRegClassForOperand(Dst, MRI);
      if (!RC)
        return true;
      return RBI.constrainGenericRegister(DstReg, *RC, MRI);
    }

    if (!isVCC(SrcReg, MRI)) {
      // TODO: Should probably leave the copy and let copyPhysReg expand it.
      if (!RBI.constrainGenericRegister(DstReg, *TRI.getBoolRC(), MRI))
        return false;

      BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_CMP_NE_U32_e64), DstReg)
        .addImm(0)
        .addReg(SrcReg);

      if (!MRI.getRegClassOrNull(SrcReg))
        MRI.setRegClass(SrcReg, TRI.getConstrainedRegClassForOperand(Src, MRI));
      I.eraseFromParent();
      return true;
    }

    const TargetRegisterClass *RC =
      TRI.getConstrainedRegClassForOperand(Dst, MRI);
    if (RC && !RBI.constrainGenericRegister(DstReg, *RC, MRI))
      return false;

    // Don't constrain the source register to a class so the def instruction
    // handles it (unless it's undef).
    //
    // FIXME: This is a hack. When selecting the def, we neeed to know
    // specifically know that the result is VCCRegBank, and not just an SGPR
    // with size 1. An SReg_32 with size 1 is ambiguous with wave32.
    if (Src.isUndef()) {
      const TargetRegisterClass *SrcRC =
        TRI.getConstrainedRegClassForOperand(Src, MRI);
      if (SrcRC && !RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI))
        return false;
    }

    return true;
  }

  for (const MachineOperand &MO : I.operands()) {
    if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
      continue;

    const TargetRegisterClass *RC =
            TRI.getConstrainedRegClassForOperand(MO, MRI);
    if (!RC)
      continue;
    RBI.constrainGenericRegister(MO.getReg(), *RC, MRI);
  }
  return true;
}

bool AMDGPUInstructionSelector::selectPHI(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();

  const Register DefReg = I.getOperand(0).getReg();
  const LLT DefTy = MRI.getType(DefReg);

  // TODO: Verify this doesn't have insane operands (i.e. VGPR to SGPR copy)

  const RegClassOrRegBank &RegClassOrBank =
    MRI.getRegClassOrRegBank(DefReg);

  const TargetRegisterClass *DefRC
    = RegClassOrBank.dyn_cast<const TargetRegisterClass *>();
  if (!DefRC) {
    if (!DefTy.isValid()) {
      LLVM_DEBUG(dbgs() << "PHI operand has no type, not a gvreg?\n");
      return false;
    }

    const RegisterBank &RB = *RegClassOrBank.get<const RegisterBank *>();
    if (RB.getID() == AMDGPU::SCCRegBankID) {
      LLVM_DEBUG(dbgs() << "illegal scc phi\n");
      return false;
    }

    DefRC = TRI.getRegClassForTypeOnBank(DefTy, RB, MRI);
    if (!DefRC) {
      LLVM_DEBUG(dbgs() << "PHI operand has unexpected size/bank\n");
      return false;
    }
  }

  I.setDesc(TII.get(TargetOpcode::PHI));
  return RBI.constrainGenericRegister(DefReg, *DefRC, MRI);
}

MachineOperand
AMDGPUInstructionSelector::getSubOperand64(MachineOperand &MO,
                                           const TargetRegisterClass &SubRC,
                                           unsigned SubIdx) const {

  MachineInstr *MI = MO.getParent();
  MachineBasicBlock *BB = MO.getParent()->getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  Register DstReg = MRI.createVirtualRegister(&SubRC);

  if (MO.isReg()) {
    unsigned ComposedSubIdx = TRI.composeSubRegIndices(MO.getSubReg(), SubIdx);
    unsigned Reg = MO.getReg();
    BuildMI(*BB, MI, MI->getDebugLoc(), TII.get(AMDGPU::COPY), DstReg)
            .addReg(Reg, 0, ComposedSubIdx);

    return MachineOperand::CreateReg(DstReg, MO.isDef(), MO.isImplicit(),
                                     MO.isKill(), MO.isDead(), MO.isUndef(),
                                     MO.isEarlyClobber(), 0, MO.isDebug(),
                                     MO.isInternalRead());
  }

  assert(MO.isImm());

  APInt Imm(64, MO.getImm());

  switch (SubIdx) {
  default:
    llvm_unreachable("do not know to split immediate with this sub index.");
  case AMDGPU::sub0:
    return MachineOperand::CreateImm(Imm.getLoBits(32).getSExtValue());
  case AMDGPU::sub1:
    return MachineOperand::CreateImm(Imm.getHiBits(32).getSExtValue());
  }
}

static int64_t getConstant(const MachineInstr *MI) {
  return MI->getOperand(1).getCImm()->getSExtValue();
}

static unsigned getLogicalBitOpcode(unsigned Opc, bool Is64) {
  switch (Opc) {
  case AMDGPU::G_AND:
    return Is64 ? AMDGPU::S_AND_B64 : AMDGPU::S_AND_B32;
  case AMDGPU::G_OR:
    return Is64 ? AMDGPU::S_OR_B64 : AMDGPU::S_OR_B32;
  case AMDGPU::G_XOR:
    return Is64 ? AMDGPU::S_XOR_B64 : AMDGPU::S_XOR_B32;
  default:
    llvm_unreachable("not a bit op");
  }
}

bool AMDGPUInstructionSelector::selectG_AND_OR_XOR(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  MachineOperand &Dst = I.getOperand(0);
  MachineOperand &Src0 = I.getOperand(1);
  MachineOperand &Src1 = I.getOperand(2);
  Register DstReg = Dst.getReg();
  unsigned Size = RBI.getSizeInBits(DstReg, MRI, TRI);

  const RegisterBank *DstRB = RBI.getRegBank(DstReg, MRI, TRI);
  if (DstRB->getID() == AMDGPU::VCCRegBankID) {
    const TargetRegisterClass *RC = TRI.getBoolRC();
    unsigned InstOpc = getLogicalBitOpcode(I.getOpcode(),
                                           RC == &AMDGPU::SReg_64RegClass);
    I.setDesc(TII.get(InstOpc));

    // FIXME: Hack to avoid turning the register bank into a register class.
    // The selector for G_ICMP relies on seeing the register bank for the result
    // is VCC. In wave32 if we constrain the registers to SReg_32 here, it will
    // be ambiguous whether it's a scalar or vector bool.
    if (Src0.isUndef() && !MRI.getRegClassOrNull(Src0.getReg()))
      MRI.setRegClass(Src0.getReg(), RC);
    if (Src1.isUndef() && !MRI.getRegClassOrNull(Src1.getReg()))
      MRI.setRegClass(Src1.getReg(), RC);

    return RBI.constrainGenericRegister(DstReg, *RC, MRI);
  }

  // TODO: Should this allow an SCC bank result, and produce a copy from SCC for
  // the result?
  if (DstRB->getID() == AMDGPU::SGPRRegBankID) {
    unsigned InstOpc = getLogicalBitOpcode(I.getOpcode(), Size > 32);
    I.setDesc(TII.get(InstOpc));

    const TargetRegisterClass *RC
      = TRI.getConstrainedRegClassForOperand(Dst, MRI);
    if (!RC)
      return false;
    return RBI.constrainGenericRegister(DstReg, *RC, MRI) &&
           RBI.constrainGenericRegister(Src0.getReg(), *RC, MRI) &&
           RBI.constrainGenericRegister(Src1.getReg(), *RC, MRI);
  }

  return false;
}

bool AMDGPUInstructionSelector::selectG_ADD_SUB(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  Register DstReg = I.getOperand(0).getReg();
  const DebugLoc &DL = I.getDebugLoc();
  unsigned Size = RBI.getSizeInBits(DstReg, MRI, TRI);
  const RegisterBank *DstRB = RBI.getRegBank(DstReg, MRI, TRI);
  const bool IsSALU = DstRB->getID() == AMDGPU::SGPRRegBankID;
  const bool Sub = I.getOpcode() == TargetOpcode::G_SUB;

  if (Size == 32) {
    if (IsSALU) {
      const unsigned Opc = Sub ? AMDGPU::S_SUB_U32 : AMDGPU::S_ADD_U32;
      MachineInstr *Add =
        BuildMI(*BB, &I, DL, TII.get(Opc), DstReg)
        .add(I.getOperand(1))
        .add(I.getOperand(2));
      I.eraseFromParent();
      return constrainSelectedInstRegOperands(*Add, TII, TRI, RBI);
    }

    if (STI.hasAddNoCarry()) {
      const unsigned Opc = Sub ? AMDGPU::V_SUB_U32_e64 : AMDGPU::V_ADD_U32_e64;
      I.setDesc(TII.get(Opc));
      I.addOperand(*MF, MachineOperand::CreateImm(0));
      I.addOperand(*MF, MachineOperand::CreateReg(AMDGPU::EXEC, false, true));
      return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
    }

    const unsigned Opc = Sub ? AMDGPU::V_SUB_I32_e64 : AMDGPU::V_ADD_I32_e64;

    Register UnusedCarry = MRI.createVirtualRegister(TRI.getWaveMaskRegClass());
    MachineInstr *Add
      = BuildMI(*BB, &I, DL, TII.get(Opc), DstReg)
      .addDef(UnusedCarry, RegState::Dead)
      .add(I.getOperand(1))
      .add(I.getOperand(2))
      .addImm(0);
    I.eraseFromParent();
    return constrainSelectedInstRegOperands(*Add, TII, TRI, RBI);
  }

  assert(!Sub && "illegal sub should not reach here");

  const TargetRegisterClass &RC
    = IsSALU ? AMDGPU::SReg_64_XEXECRegClass : AMDGPU::VReg_64RegClass;
  const TargetRegisterClass &HalfRC
    = IsSALU ? AMDGPU::SReg_32RegClass : AMDGPU::VGPR_32RegClass;

  MachineOperand Lo1(getSubOperand64(I.getOperand(1), HalfRC, AMDGPU::sub0));
  MachineOperand Lo2(getSubOperand64(I.getOperand(2), HalfRC, AMDGPU::sub0));
  MachineOperand Hi1(getSubOperand64(I.getOperand(1), HalfRC, AMDGPU::sub1));
  MachineOperand Hi2(getSubOperand64(I.getOperand(2), HalfRC, AMDGPU::sub1));

  Register DstLo = MRI.createVirtualRegister(&HalfRC);
  Register DstHi = MRI.createVirtualRegister(&HalfRC);

  if (IsSALU) {
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_ADD_U32), DstLo)
      .add(Lo1)
      .add(Lo2);
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_ADDC_U32), DstHi)
      .add(Hi1)
      .add(Hi2);
  } else {
    const TargetRegisterClass *CarryRC = TRI.getWaveMaskRegClass();
    Register CarryReg = MRI.createVirtualRegister(CarryRC);
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_ADD_I32_e64), DstLo)
      .addDef(CarryReg)
      .add(Lo1)
      .add(Lo2)
      .addImm(0);
    MachineInstr *Addc = BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_ADDC_U32_e64), DstHi)
      .addDef(MRI.createVirtualRegister(CarryRC), RegState::Dead)
      .add(Hi1)
      .add(Hi2)
      .addReg(CarryReg, RegState::Kill)
      .addImm(0);

    if (!constrainSelectedInstRegOperands(*Addc, TII, TRI, RBI))
      return false;
  }

  BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), DstReg)
    .addReg(DstLo)
    .addImm(AMDGPU::sub0)
    .addReg(DstHi)
    .addImm(AMDGPU::sub1);


  if (!RBI.constrainGenericRegister(DstReg, RC, MRI))
    return false;

  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_EXTRACT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  assert(I.getOperand(2).getImm() % 32 == 0);
  unsigned SubReg = TRI.getSubRegFromChannel(I.getOperand(2).getImm() / 32);
  const DebugLoc &DL = I.getDebugLoc();
  MachineInstr *Copy = BuildMI(*BB, &I, DL, TII.get(TargetOpcode::COPY),
                               I.getOperand(0).getReg())
                               .addReg(I.getOperand(1).getReg(), 0, SubReg);

  for (const MachineOperand &MO : Copy->operands()) {
    const TargetRegisterClass *RC =
            TRI.getConstrainedRegClassForOperand(MO, MRI);
    if (!RC)
      continue;
    RBI.constrainGenericRegister(MO.getReg(), *RC, MRI);
  }
  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_MERGE_VALUES(MachineInstr &MI) const {
  MachineBasicBlock *BB = MI.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  Register DstReg = MI.getOperand(0).getReg();
  LLT DstTy = MRI.getType(DstReg);
  LLT SrcTy = MRI.getType(MI.getOperand(1).getReg());

  const unsigned SrcSize = SrcTy.getSizeInBits();
  if (SrcSize < 32)
    return false;

  const DebugLoc &DL = MI.getDebugLoc();
  const RegisterBank *DstBank = RBI.getRegBank(DstReg, MRI, TRI);
  const unsigned DstSize = DstTy.getSizeInBits();
  const TargetRegisterClass *DstRC =
    TRI.getRegClassForSizeOnBank(DstSize, *DstBank, MRI);
  if (!DstRC)
    return false;

  ArrayRef<int16_t> SubRegs = TRI.getRegSplitParts(DstRC, SrcSize / 8);
  MachineInstrBuilder MIB =
    BuildMI(*BB, &MI, DL, TII.get(TargetOpcode::REG_SEQUENCE), DstReg);
  for (int I = 0, E = MI.getNumOperands() - 1; I != E; ++I) {
    MachineOperand &Src = MI.getOperand(I + 1);
    MIB.addReg(Src.getReg(), getUndefRegState(Src.isUndef()));
    MIB.addImm(SubRegs[I]);

    const TargetRegisterClass *SrcRC
      = TRI.getConstrainedRegClassForOperand(Src, MRI);
    if (SrcRC && !RBI.constrainGenericRegister(Src.getReg(), *SrcRC, MRI))
      return false;
  }

  if (!RBI.constrainGenericRegister(DstReg, *DstRC, MRI))
    return false;

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_UNMERGE_VALUES(MachineInstr &MI) const {
  MachineBasicBlock *BB = MI.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  const int NumDst = MI.getNumOperands() - 1;

  MachineOperand &Src = MI.getOperand(NumDst);

  Register SrcReg = Src.getReg();
  Register DstReg0 = MI.getOperand(0).getReg();
  LLT DstTy = MRI.getType(DstReg0);
  LLT SrcTy = MRI.getType(SrcReg);

  const unsigned DstSize = DstTy.getSizeInBits();
  const unsigned SrcSize = SrcTy.getSizeInBits();
  const DebugLoc &DL = MI.getDebugLoc();
  const RegisterBank *SrcBank = RBI.getRegBank(SrcReg, MRI, TRI);

  const TargetRegisterClass *SrcRC =
    TRI.getRegClassForSizeOnBank(SrcSize, *SrcBank, MRI);
  if (!SrcRC || !RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI))
    return false;

  const unsigned SrcFlags = getUndefRegState(Src.isUndef());

  // Note we could have mixed SGPR and VGPR destination banks for an SGPR
  // source, and this relies on the fact that the same subregister indices are
  // used for both.
  ArrayRef<int16_t> SubRegs = TRI.getRegSplitParts(SrcRC, DstSize / 8);
  for (int I = 0, E = NumDst; I != E; ++I) {
    MachineOperand &Dst = MI.getOperand(I);
    BuildMI(*BB, &MI, DL, TII.get(TargetOpcode::COPY), Dst.getReg())
      .addReg(SrcReg, SrcFlags, SubRegs[I]);

    const TargetRegisterClass *DstRC =
      TRI.getConstrainedRegClassForOperand(Dst, MRI);
    if (DstRC && !RBI.constrainGenericRegister(Dst.getReg(), *DstRC, MRI))
      return false;
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_GEP(MachineInstr &I) const {
  return selectG_ADD_SUB(I);
}

bool AMDGPUInstructionSelector::selectG_IMPLICIT_DEF(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  const MachineOperand &MO = I.getOperand(0);

  // FIXME: Interface for getConstrainedRegClassForOperand needs work. The
  // regbank check here is to know why getConstrainedRegClassForOperand failed.
  const TargetRegisterClass *RC = TRI.getConstrainedRegClassForOperand(MO, MRI);
  if ((!RC && !MRI.getRegBankOrNull(MO.getReg())) ||
      (RC && RBI.constrainGenericRegister(MO.getReg(), *RC, MRI))) {
    I.setDesc(TII.get(TargetOpcode::IMPLICIT_DEF));
    return true;
  }

  return false;
}

bool AMDGPUInstructionSelector::selectG_INSERT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  unsigned SubReg = TRI.getSubRegFromChannel(I.getOperand(3).getImm() / 32);
  DebugLoc DL = I.getDebugLoc();
  MachineInstr *Ins = BuildMI(*BB, &I, DL, TII.get(TargetOpcode::INSERT_SUBREG))
                               .addDef(I.getOperand(0).getReg())
                               .addReg(I.getOperand(1).getReg())
                               .addReg(I.getOperand(2).getReg())
                               .addImm(SubReg);

  for (const MachineOperand &MO : Ins->operands()) {
    if (!MO.isReg())
      continue;
    if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
      continue;

    const TargetRegisterClass *RC =
            TRI.getConstrainedRegClassForOperand(MO, MRI);
    if (!RC)
      continue;
    RBI.constrainGenericRegister(MO.getReg(), *RC, MRI);
  }
  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_INTRINSIC(
  MachineInstr &I, CodeGenCoverage &CoverageInfo) const {
  unsigned IntrinsicID =  I.getOperand(I.getNumExplicitDefs()).getIntrinsicID();
  switch (IntrinsicID) {
  case Intrinsic::maxnum:
  case Intrinsic::minnum:
  case Intrinsic::amdgcn_cvt_pkrtz:
    return selectImpl(I, CoverageInfo);
  case Intrinsic::amdgcn_if_break: {
    MachineBasicBlock *BB = I.getParent();
    MachineFunction *MF = BB->getParent();
    MachineRegisterInfo &MRI = MF->getRegInfo();

    // FIXME: Manually selecting to avoid dealiing with the SReg_1 trick
    // SelectionDAG uses for wave32 vs wave64.
    BuildMI(*BB, &I, I.getDebugLoc(), TII.get(AMDGPU::SI_IF_BREAK))
      .add(I.getOperand(0))
      .add(I.getOperand(2))
      .add(I.getOperand(3));

    Register DstReg = I.getOperand(0).getReg();
    Register Src0Reg = I.getOperand(2).getReg();
    Register Src1Reg = I.getOperand(3).getReg();

    I.eraseFromParent();

    for (Register Reg : { DstReg, Src0Reg, Src1Reg }) {
      if (!MRI.getRegClassOrNull(Reg))
        MRI.setRegClass(Reg, TRI.getWaveMaskRegClass());
    }

    return true;
  }
  default:
    return selectImpl(I, CoverageInfo);
  }
}

static int getV_CMPOpcode(CmpInst::Predicate P, unsigned Size) {
  if (Size != 32 && Size != 64)
    return -1;
  switch (P) {
  default:
    llvm_unreachable("Unknown condition code!");
  case CmpInst::ICMP_NE:
    return Size == 32 ? AMDGPU::V_CMP_NE_U32_e64 : AMDGPU::V_CMP_NE_U64_e64;
  case CmpInst::ICMP_EQ:
    return Size == 32 ? AMDGPU::V_CMP_EQ_U32_e64 : AMDGPU::V_CMP_EQ_U64_e64;
  case CmpInst::ICMP_SGT:
    return Size == 32 ? AMDGPU::V_CMP_GT_I32_e64 : AMDGPU::V_CMP_GT_I64_e64;
  case CmpInst::ICMP_SGE:
    return Size == 32 ? AMDGPU::V_CMP_GE_I32_e64 : AMDGPU::V_CMP_GE_I64_e64;
  case CmpInst::ICMP_SLT:
    return Size == 32 ? AMDGPU::V_CMP_LT_I32_e64 : AMDGPU::V_CMP_LT_I64_e64;
  case CmpInst::ICMP_SLE:
    return Size == 32 ? AMDGPU::V_CMP_LE_I32_e64 : AMDGPU::V_CMP_LE_I64_e64;
  case CmpInst::ICMP_UGT:
    return Size == 32 ? AMDGPU::V_CMP_GT_U32_e64 : AMDGPU::V_CMP_GT_U64_e64;
  case CmpInst::ICMP_UGE:
    return Size == 32 ? AMDGPU::V_CMP_GE_U32_e64 : AMDGPU::V_CMP_GE_U64_e64;
  case CmpInst::ICMP_ULT:
    return Size == 32 ? AMDGPU::V_CMP_LT_U32_e64 : AMDGPU::V_CMP_LT_U64_e64;
  case CmpInst::ICMP_ULE:
    return Size == 32 ? AMDGPU::V_CMP_LE_U32_e64 : AMDGPU::V_CMP_LE_U64_e64;
  }
}

int AMDGPUInstructionSelector::getS_CMPOpcode(CmpInst::Predicate P,
                                              unsigned Size) const {
  if (Size == 64) {
    if (!STI.hasScalarCompareEq64())
      return -1;

    switch (P) {
    case CmpInst::ICMP_NE:
      return AMDGPU::S_CMP_LG_U64;
    case CmpInst::ICMP_EQ:
      return AMDGPU::S_CMP_EQ_U64;
    default:
      return -1;
    }
  }

  if (Size != 32)
    return -1;

  switch (P) {
  case CmpInst::ICMP_NE:
    return AMDGPU::S_CMP_LG_U32;
  case CmpInst::ICMP_EQ:
    return AMDGPU::S_CMP_EQ_U32;
  case CmpInst::ICMP_SGT:
    return AMDGPU::S_CMP_GT_I32;
  case CmpInst::ICMP_SGE:
    return AMDGPU::S_CMP_GE_I32;
  case CmpInst::ICMP_SLT:
    return AMDGPU::S_CMP_LT_I32;
  case CmpInst::ICMP_SLE:
    return AMDGPU::S_CMP_LE_I32;
  case CmpInst::ICMP_UGT:
    return AMDGPU::S_CMP_GT_U32;
  case CmpInst::ICMP_UGE:
    return AMDGPU::S_CMP_GE_U32;
  case CmpInst::ICMP_ULT:
    return AMDGPU::S_CMP_LT_U32;
  case CmpInst::ICMP_ULE:
    return AMDGPU::S_CMP_LE_U32;
  default:
    llvm_unreachable("Unknown condition code!");
  }
}

bool AMDGPUInstructionSelector::selectG_ICMP(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  const DebugLoc &DL = I.getDebugLoc();

  unsigned SrcReg = I.getOperand(2).getReg();
  unsigned Size = RBI.getSizeInBits(SrcReg, MRI, TRI);

  auto Pred = (CmpInst::Predicate)I.getOperand(1).getPredicate();

  unsigned CCReg = I.getOperand(0).getReg();
  if (isSCC(CCReg, MRI)) {
    int Opcode = getS_CMPOpcode(Pred, Size);
    if (Opcode == -1)
      return false;
    MachineInstr *ICmp = BuildMI(*BB, &I, DL, TII.get(Opcode))
            .add(I.getOperand(2))
            .add(I.getOperand(3));
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), CCReg)
      .addReg(AMDGPU::SCC);
    bool Ret =
        constrainSelectedInstRegOperands(*ICmp, TII, TRI, RBI) &&
        RBI.constrainGenericRegister(CCReg, AMDGPU::SReg_32RegClass, MRI);
    I.eraseFromParent();
    return Ret;
  }

  int Opcode = getV_CMPOpcode(Pred, Size);
  if (Opcode == -1)
    return false;

  MachineInstr *ICmp = BuildMI(*BB, &I, DL, TII.get(Opcode),
            I.getOperand(0).getReg())
            .add(I.getOperand(2))
            .add(I.getOperand(3));
  RBI.constrainGenericRegister(ICmp->getOperand(0).getReg(),
                               *TRI.getBoolRC(), MRI);
  bool Ret = constrainSelectedInstRegOperands(*ICmp, TII, TRI, RBI);
  I.eraseFromParent();
  return Ret;
}

static MachineInstr *
buildEXP(const TargetInstrInfo &TII, MachineInstr *Insert, unsigned Tgt,
         unsigned Reg0, unsigned Reg1, unsigned Reg2, unsigned Reg3,
         unsigned VM, bool Compr, unsigned Enabled, bool Done) {
  const DebugLoc &DL = Insert->getDebugLoc();
  MachineBasicBlock &BB = *Insert->getParent();
  unsigned Opcode = Done ? AMDGPU::EXP_DONE : AMDGPU::EXP;
  return BuildMI(BB, Insert, DL, TII.get(Opcode))
          .addImm(Tgt)
          .addReg(Reg0)
          .addReg(Reg1)
          .addReg(Reg2)
          .addReg(Reg3)
          .addImm(VM)
          .addImm(Compr)
          .addImm(Enabled);
}

bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
  MachineInstr &I, CodeGenCoverage &CoverageInfo) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();

  unsigned IntrinsicID = I.getOperand(0).getIntrinsicID();
  switch (IntrinsicID) {
  case Intrinsic::amdgcn_exp: {
    int64_t Tgt = getConstant(MRI.getVRegDef(I.getOperand(1).getReg()));
    int64_t Enabled = getConstant(MRI.getVRegDef(I.getOperand(2).getReg()));
    int64_t Done = getConstant(MRI.getVRegDef(I.getOperand(7).getReg()));
    int64_t VM = getConstant(MRI.getVRegDef(I.getOperand(8).getReg()));

    MachineInstr *Exp = buildEXP(TII, &I, Tgt, I.getOperand(3).getReg(),
                                 I.getOperand(4).getReg(),
                                 I.getOperand(5).getReg(),
                                 I.getOperand(6).getReg(),
                                 VM, false, Enabled, Done);

    I.eraseFromParent();
    return constrainSelectedInstRegOperands(*Exp, TII, TRI, RBI);
  }
  case Intrinsic::amdgcn_exp_compr: {
    const DebugLoc &DL = I.getDebugLoc();
    int64_t Tgt = getConstant(MRI.getVRegDef(I.getOperand(1).getReg()));
    int64_t Enabled = getConstant(MRI.getVRegDef(I.getOperand(2).getReg()));
    unsigned Reg0 = I.getOperand(3).getReg();
    unsigned Reg1 = I.getOperand(4).getReg();
    unsigned Undef = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
    int64_t Done = getConstant(MRI.getVRegDef(I.getOperand(5).getReg()));
    int64_t VM = getConstant(MRI.getVRegDef(I.getOperand(6).getReg()));

    BuildMI(*BB, &I, DL, TII.get(AMDGPU::IMPLICIT_DEF), Undef);
    MachineInstr *Exp = buildEXP(TII, &I, Tgt, Reg0, Reg1, Undef, Undef, VM,
                                 true,  Enabled, Done);

    I.eraseFromParent();
    return constrainSelectedInstRegOperands(*Exp, TII, TRI, RBI);
  }
  case Intrinsic::amdgcn_end_cf: {
    // FIXME: Manually selecting to avoid dealiing with the SReg_1 trick
    // SelectionDAG uses for wave32 vs wave64.
    BuildMI(*BB, &I, I.getDebugLoc(),
            TII.get(AMDGPU::SI_END_CF))
      .add(I.getOperand(1));

    Register Reg = I.getOperand(1).getReg();
    I.eraseFromParent();

    if (!MRI.getRegClassOrNull(Reg))
      MRI.setRegClass(Reg, TRI.getWaveMaskRegClass());
    return true;
  }
  default:
    return selectImpl(I, CoverageInfo);
  }
}

bool AMDGPUInstructionSelector::selectG_SELECT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  const DebugLoc &DL = I.getDebugLoc();

  unsigned DstReg = I.getOperand(0).getReg();
  unsigned Size = RBI.getSizeInBits(DstReg, MRI, TRI);
  assert(Size <= 32 || Size == 64);
  const MachineOperand &CCOp = I.getOperand(1);
  unsigned CCReg = CCOp.getReg();
  if (isSCC(CCReg, MRI)) {
    unsigned SelectOpcode = Size == 64 ? AMDGPU::S_CSELECT_B64 :
                                         AMDGPU::S_CSELECT_B32;
    MachineInstr *CopySCC = BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), AMDGPU::SCC)
            .addReg(CCReg);

    // The generic constrainSelectedInstRegOperands doesn't work for the scc register
    // bank, because it does not cover the register class that we used to represent
    // for it.  So we need to manually set the register class here.
    if (!MRI.getRegClassOrNull(CCReg))
        MRI.setRegClass(CCReg, TRI.getConstrainedRegClassForOperand(CCOp, MRI));
    MachineInstr *Select = BuildMI(*BB, &I, DL, TII.get(SelectOpcode), DstReg)
            .add(I.getOperand(2))
            .add(I.getOperand(3));

    bool Ret = constrainSelectedInstRegOperands(*Select, TII, TRI, RBI) |
               constrainSelectedInstRegOperands(*CopySCC, TII, TRI, RBI);
    I.eraseFromParent();
    return Ret;
  }

  // Wide VGPR select should have been split in RegBankSelect.
  if (Size > 32)
    return false;

  MachineInstr *Select =
      BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
              .addImm(0)
              .add(I.getOperand(3))
              .addImm(0)
              .add(I.getOperand(2))
              .add(I.getOperand(1));

  bool Ret = constrainSelectedInstRegOperands(*Select, TII, TRI, RBI);
  I.eraseFromParent();
  return Ret;
}

bool AMDGPUInstructionSelector::selectG_STORE(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  DebugLoc DL = I.getDebugLoc();
  unsigned PtrSize = RBI.getSizeInBits(I.getOperand(1).getReg(), MRI, TRI);
  if (PtrSize != 64) {
    LLVM_DEBUG(dbgs() << "Unhandled address space\n");
    return false;
  }

  unsigned StoreSize = RBI.getSizeInBits(I.getOperand(0).getReg(), MRI, TRI);
  unsigned Opcode;

  // FIXME: Remove this when integers > s32 naturally selected.
  switch (StoreSize) {
  default:
    return false;
  case 32:
    Opcode = AMDGPU::FLAT_STORE_DWORD;
    break;
  case 64:
    Opcode = AMDGPU::FLAT_STORE_DWORDX2;
    break;
  case 96:
    Opcode = AMDGPU::FLAT_STORE_DWORDX3;
    break;
  case 128:
    Opcode = AMDGPU::FLAT_STORE_DWORDX4;
    break;
  }

  MachineInstr *Flat = BuildMI(*BB, &I, DL, TII.get(Opcode))
          .add(I.getOperand(1))
          .add(I.getOperand(0))
          .addImm(0)  // offset
          .addImm(0)  // glc
          .addImm(0)  // slc
          .addImm(0); // dlc


  // Now that we selected an opcode, we need to constrain the register
  // operands to use appropriate classes.
  bool Ret = constrainSelectedInstRegOperands(*Flat, TII, TRI, RBI);

  I.eraseFromParent();
  return Ret;
}

static int sizeToSubRegIndex(unsigned Size) {
  switch (Size) {
  case 32:
    return AMDGPU::sub0;
  case 64:
    return AMDGPU::sub0_sub1;
  case 96:
    return AMDGPU::sub0_sub1_sub2;
  case 128:
    return AMDGPU::sub0_sub1_sub2_sub3;
  case 256:
    return AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
  default:
    if (Size < 32)
      return AMDGPU::sub0;
    if (Size > 256)
      return -1;
    return sizeToSubRegIndex(PowerOf2Ceil(Size));
  }
}

bool AMDGPUInstructionSelector::selectG_TRUNC(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();

  unsigned DstReg = I.getOperand(0).getReg();
  unsigned SrcReg = I.getOperand(1).getReg();
  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);
  if (!DstTy.isScalar())
    return false;

  const RegisterBank *DstRB = RBI.getRegBank(DstReg, MRI, TRI);
  const RegisterBank *SrcRB = RBI.getRegBank(SrcReg, MRI, TRI);
  if (SrcRB != DstRB)
    return false;

  unsigned DstSize = DstTy.getSizeInBits();
  unsigned SrcSize = SrcTy.getSizeInBits();

  const TargetRegisterClass *SrcRC
    = TRI.getRegClassForSizeOnBank(SrcSize, *SrcRB, MRI);
  const TargetRegisterClass *DstRC
    = TRI.getRegClassForSizeOnBank(DstSize, *DstRB, MRI);

  if (SrcSize > 32) {
    int SubRegIdx = sizeToSubRegIndex(DstSize);
    if (SubRegIdx == -1)
      return false;

    // Deal with weird cases where the class only partially supports the subreg
    // index.
    SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubRegIdx);
    if (!SrcRC)
      return false;

    I.getOperand(1).setSubReg(SubRegIdx);
  }

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain G_TRUNC\n");
    return false;
  }

  I.setDesc(TII.get(TargetOpcode::COPY));
  return true;
}

/// \returns true if a bitmask for \p Size bits will be an inline immediate.
static bool shouldUseAndMask(unsigned Size, unsigned &Mask) {
  Mask = maskTrailingOnes<unsigned>(Size);
  int SignedMask = static_cast<int>(Mask);
  return SignedMask >= -16 && SignedMask <= 64;
}

bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
  bool Signed = I.getOpcode() == AMDGPU::G_SEXT;
  const DebugLoc &DL = I.getDebugLoc();
  MachineBasicBlock &MBB = *I.getParent();
  MachineFunction &MF = *MBB.getParent();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  const unsigned DstReg = I.getOperand(0).getReg();
  const unsigned SrcReg = I.getOperand(1).getReg();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);
  const LLT S1 = LLT::scalar(1);
  const unsigned SrcSize = SrcTy.getSizeInBits();
  const unsigned DstSize = DstTy.getSizeInBits();
  if (!DstTy.isScalar())
    return false;

  const RegisterBank *SrcBank = RBI.getRegBank(SrcReg, MRI, TRI);

  if (SrcBank->getID() == AMDGPU::SCCRegBankID) {
    if (SrcTy != S1 || DstSize > 64) // Invalid
      return false;

    unsigned Opcode =
        DstSize > 32 ? AMDGPU::S_CSELECT_B64 : AMDGPU::S_CSELECT_B32;
    const TargetRegisterClass *DstRC =
        DstSize > 32 ? &AMDGPU::SReg_64RegClass : &AMDGPU::SReg_32RegClass;

    // FIXME: Create an extra copy to avoid incorrectly constraining the result
    // of the scc producer.
    unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
    BuildMI(MBB, I, DL, TII.get(AMDGPU::COPY), TmpReg)
      .addReg(SrcReg);
    BuildMI(MBB, I, DL, TII.get(AMDGPU::COPY), AMDGPU::SCC)
      .addReg(TmpReg);

    // The instruction operands are backwards from what you would expect.
    BuildMI(MBB, I, DL, TII.get(Opcode), DstReg)
      .addImm(0)
      .addImm(Signed ? -1 : 1);
    return RBI.constrainGenericRegister(DstReg, *DstRC, MRI);
  }

  if (SrcBank->getID() == AMDGPU::VCCRegBankID && DstSize <= 32) {
    if (SrcTy != S1) // Invalid
      return false;

    MachineInstr *ExtI =
      BuildMI(MBB, I, DL, TII.get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
      .addImm(0)               // src0_modifiers
      .addImm(0)               // src0
      .addImm(0)               // src1_modifiers
      .addImm(Signed ? -1 : 1) // src1
      .addUse(SrcReg);
    return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
  }

  if (I.getOpcode() == AMDGPU::G_ANYEXT)
    return selectCOPY(I);

  if (SrcBank->getID() == AMDGPU::VGPRRegBankID && DstSize <= 32) {
    // 64-bit should have been split up in RegBankSelect

    // Try to use an and with a mask if it will save code size.
    unsigned Mask;
    if (!Signed && shouldUseAndMask(SrcSize, Mask)) {
      MachineInstr *ExtI =
      BuildMI(MBB, I, DL, TII.get(AMDGPU::V_AND_B32_e32), DstReg)
        .addImm(Mask)
        .addReg(SrcReg);
      return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
    }

    const unsigned BFE = Signed ? AMDGPU::V_BFE_I32 : AMDGPU::V_BFE_U32;
    MachineInstr *ExtI =
      BuildMI(MBB, I, DL, TII.get(BFE), DstReg)
      .addReg(SrcReg)
      .addImm(0) // Offset
      .addImm(SrcSize); // Width
    return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
  }

  if (SrcBank->getID() == AMDGPU::SGPRRegBankID && DstSize <= 64) {
    if (!RBI.constrainGenericRegister(SrcReg, AMDGPU::SReg_32RegClass, MRI))
      return false;

    if (Signed && DstSize == 32 && (SrcSize == 8 || SrcSize == 16)) {
      const unsigned SextOpc = SrcSize == 8 ?
        AMDGPU::S_SEXT_I32_I8 : AMDGPU::S_SEXT_I32_I16;
      BuildMI(MBB, I, DL, TII.get(SextOpc), DstReg)
        .addReg(SrcReg);
      return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_32RegClass, MRI);
    }

    const unsigned BFE64 = Signed ? AMDGPU::S_BFE_I64 : AMDGPU::S_BFE_U64;
    const unsigned BFE32 = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32;

    // Scalar BFE is encoded as S1[5:0] = offset, S1[22:16]= width.
    if (DstSize > 32 && SrcSize <= 32) {
      // We need a 64-bit register source, but the high bits don't matter.
      unsigned ExtReg
        = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
      unsigned UndefReg
        = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
      BuildMI(MBB, I, DL, TII.get(AMDGPU::IMPLICIT_DEF), UndefReg);
      BuildMI(MBB, I, DL, TII.get(AMDGPU::REG_SEQUENCE), ExtReg)
        .addReg(SrcReg)
        .addImm(AMDGPU::sub0)
        .addReg(UndefReg)
        .addImm(AMDGPU::sub1);

      BuildMI(MBB, I, DL, TII.get(BFE64), DstReg)
        .addReg(ExtReg)
        .addImm(SrcSize << 16);

      return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_64RegClass, MRI);
    }

    unsigned Mask;
    if (!Signed && shouldUseAndMask(SrcSize, Mask)) {
      BuildMI(MBB, I, DL, TII.get(AMDGPU::S_AND_B32), DstReg)
        .addReg(SrcReg)
        .addImm(Mask);
    } else {
      BuildMI(MBB, I, DL, TII.get(BFE32), DstReg)
        .addReg(SrcReg)
        .addImm(SrcSize << 16);
    }

    return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_32RegClass, MRI);
  }

  return false;
}

bool AMDGPUInstructionSelector::selectG_CONSTANT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  MachineOperand &ImmOp = I.getOperand(1);

  // The AMDGPU backend only supports Imm operands and not CImm or FPImm.
  if (ImmOp.isFPImm()) {
    const APInt &Imm = ImmOp.getFPImm()->getValueAPF().bitcastToAPInt();
    ImmOp.ChangeToImmediate(Imm.getZExtValue());
  } else if (ImmOp.isCImm()) {
    ImmOp.ChangeToImmediate(ImmOp.getCImm()->getZExtValue());
  }

  unsigned DstReg = I.getOperand(0).getReg();
  unsigned Size;
  bool IsSgpr;
  const RegisterBank *RB = MRI.getRegBankOrNull(I.getOperand(0).getReg());
  if (RB) {
    IsSgpr = RB->getID() == AMDGPU::SGPRRegBankID;
    Size = MRI.getType(DstReg).getSizeInBits();
  } else {
    const TargetRegisterClass *RC = TRI.getRegClassForReg(MRI, DstReg);
    IsSgpr = TRI.isSGPRClass(RC);
    Size = TRI.getRegSizeInBits(*RC);
  }

  if (Size != 32 && Size != 64)
    return false;

  unsigned Opcode = IsSgpr ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
  if (Size == 32) {
    I.setDesc(TII.get(Opcode));
    I.addImplicitDefUseOperands(*MF);
    return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
  }

  DebugLoc DL = I.getDebugLoc();
  const TargetRegisterClass *RC = IsSgpr ? &AMDGPU::SReg_32_XM0RegClass :
                                           &AMDGPU::VGPR_32RegClass;
  unsigned LoReg = MRI.createVirtualRegister(RC);
  unsigned HiReg = MRI.createVirtualRegister(RC);
  const APInt &Imm = APInt(Size, I.getOperand(1).getImm());

  BuildMI(*BB, &I, DL, TII.get(Opcode), LoReg)
          .addImm(Imm.trunc(32).getZExtValue());

  BuildMI(*BB, &I, DL, TII.get(Opcode), HiReg)
          .addImm(Imm.ashr(32).getZExtValue());

  const MachineInstr *RS =
      BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), DstReg)
              .addReg(LoReg)
              .addImm(AMDGPU::sub0)
              .addReg(HiReg)
              .addImm(AMDGPU::sub1);

  // We can't call constrainSelectedInstRegOperands here, because it doesn't
  // work for target independent opcodes
  I.eraseFromParent();
  const TargetRegisterClass *DstRC =
      TRI.getConstrainedRegClassForOperand(RS->getOperand(0), MRI);
  if (!DstRC)
    return true;
  return RBI.constrainGenericRegister(DstReg, *DstRC, MRI);
}

static bool isConstant(const MachineInstr &MI) {
  return MI.getOpcode() == TargetOpcode::G_CONSTANT;
}

void AMDGPUInstructionSelector::getAddrModeInfo(const MachineInstr &Load,
    const MachineRegisterInfo &MRI, SmallVectorImpl<GEPInfo> &AddrInfo) const {

  const MachineInstr *PtrMI = MRI.getUniqueVRegDef(Load.getOperand(1).getReg());

  assert(PtrMI);

  if (PtrMI->getOpcode() != TargetOpcode::G_GEP)
    return;

  GEPInfo GEPInfo(*PtrMI);

  for (unsigned i = 1, e = 3; i < e; ++i) {
    const MachineOperand &GEPOp = PtrMI->getOperand(i);
    const MachineInstr *OpDef = MRI.getUniqueVRegDef(GEPOp.getReg());
    assert(OpDef);
    if (isConstant(*OpDef)) {
      // FIXME: Is it possible to have multiple Imm parts?  Maybe if we
      // are lacking other optimizations.
      assert(GEPInfo.Imm == 0);
      GEPInfo.Imm = OpDef->getOperand(1).getCImm()->getSExtValue();
      continue;
    }
    const RegisterBank *OpBank = RBI.getRegBank(GEPOp.getReg(), MRI, TRI);
    if (OpBank->getID() == AMDGPU::SGPRRegBankID)
      GEPInfo.SgprParts.push_back(GEPOp.getReg());
    else
      GEPInfo.VgprParts.push_back(GEPOp.getReg());
  }

  AddrInfo.push_back(GEPInfo);
  getAddrModeInfo(*PtrMI, MRI, AddrInfo);
}

bool AMDGPUInstructionSelector::isInstrUniform(const MachineInstr &MI) const {
  if (!MI.hasOneMemOperand())
    return false;

  const MachineMemOperand *MMO = *MI.memoperands_begin();
  const Value *Ptr = MMO->getValue();

  // UndefValue means this is a load of a kernel input.  These are uniform.
  // Sometimes LDS instructions have constant pointers.
  // If Ptr is null, then that means this mem operand contains a
  // PseudoSourceValue like GOT.
  if (!Ptr || isa<UndefValue>(Ptr) || isa<Argument>(Ptr) ||
      isa<Constant>(Ptr) || isa<GlobalValue>(Ptr))
    return true;

  if (MMO->getAddrSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT)
    return true;

  const Instruction *I = dyn_cast<Instruction>(Ptr);
  return I && I->getMetadata("amdgpu.uniform");
}

bool AMDGPUInstructionSelector::hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const {
  for (const GEPInfo &GEPInfo : AddrInfo) {
    if (!GEPInfo.VgprParts.empty())
      return true;
  }
  return false;
}

bool AMDGPUInstructionSelector::selectG_LOAD(MachineInstr &I) const {
  // TODO: Can/should we insert m0 initialization here for DS instructions and
  // call the normal selector?
  return false;
}

bool AMDGPUInstructionSelector::selectG_BRCOND(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  MachineOperand &CondOp = I.getOperand(0);
  Register CondReg = CondOp.getReg();
  const DebugLoc &DL = I.getDebugLoc();

  unsigned BrOpcode;
  Register CondPhysReg;
  const TargetRegisterClass *ConstrainRC;

  // In SelectionDAG, we inspect the IR block for uniformity metadata to decide
  // whether the branch is uniform when selecting the instruction. In
  // GlobalISel, we should push that decision into RegBankSelect. Assume for now
  // RegBankSelect knows what it's doing if the branch condition is scc, even
  // though it currently does not.
  if (isSCC(CondReg, MRI)) {
    CondPhysReg = AMDGPU::SCC;
    BrOpcode = AMDGPU::S_CBRANCH_SCC1;
    ConstrainRC = &AMDGPU::SReg_32_XM0RegClass;
  } else if (isVCC(CondReg, MRI)) {
    // FIXME: Do we have to insert an and with exec here, like in SelectionDAG?
    // We sort of know that a VCC producer based on the register bank, that ands
    // inactive lanes with 0. What if there was a logical operation with vcc
    // producers in different blocks/with different exec masks?
    // FIXME: Should scc->vcc copies and with exec?
    CondPhysReg = TRI.getVCC();
    BrOpcode = AMDGPU::S_CBRANCH_VCCNZ;
    ConstrainRC = TRI.getBoolRC();
  } else
    return false;

  if (!MRI.getRegClassOrNull(CondReg))
    MRI.setRegClass(CondReg, ConstrainRC);

  BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), CondPhysReg)
    .addReg(CondReg);
  BuildMI(*BB, &I, DL, TII.get(BrOpcode))
    .addMBB(I.getOperand(1).getMBB());

  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_FRAME_INDEX(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();

  Register DstReg = I.getOperand(0).getReg();
  const RegisterBank *DstRB = RBI.getRegBank(DstReg, MRI, TRI);
  const bool IsVGPR = DstRB->getID() == AMDGPU::VGPRRegBankID;
  I.setDesc(TII.get(IsVGPR ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32));
  if (IsVGPR)
    I.addOperand(*MF, MachineOperand::CreateReg(AMDGPU::EXEC, false, true));

  return RBI.constrainGenericRegister(
    DstReg, IsVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::SReg_32RegClass, MRI);
}

bool AMDGPUInstructionSelector::select(MachineInstr &I,
                                       CodeGenCoverage &CoverageInfo) const {
  if (I.isPHI())
    return selectPHI(I);

  if (!isPreISelGenericOpcode(I.getOpcode())) {
    if (I.isCopy())
      return selectCOPY(I);
    return true;
  }

  switch (I.getOpcode()) {
  case TargetOpcode::G_AND:
  case TargetOpcode::G_OR:
  case TargetOpcode::G_XOR:
    if (selectG_AND_OR_XOR(I))
      return true;
    return selectImpl(I, CoverageInfo);
  case TargetOpcode::G_ADD:
  case TargetOpcode::G_SUB:
    if (selectG_ADD_SUB(I))
      return true;
    LLVM_FALLTHROUGH;
  default:
    return selectImpl(I, CoverageInfo);
  case TargetOpcode::G_INTTOPTR:
  case TargetOpcode::G_BITCAST:
    return selectCOPY(I);
  case TargetOpcode::G_CONSTANT:
  case TargetOpcode::G_FCONSTANT:
    return selectG_CONSTANT(I);
  case TargetOpcode::G_EXTRACT:
    return selectG_EXTRACT(I);
  case TargetOpcode::G_MERGE_VALUES:
  case TargetOpcode::G_BUILD_VECTOR:
  case TargetOpcode::G_CONCAT_VECTORS:
    return selectG_MERGE_VALUES(I);
  case TargetOpcode::G_UNMERGE_VALUES:
    return selectG_UNMERGE_VALUES(I);
  case TargetOpcode::G_GEP:
    return selectG_GEP(I);
  case TargetOpcode::G_IMPLICIT_DEF:
    return selectG_IMPLICIT_DEF(I);
  case TargetOpcode::G_INSERT:
    return selectG_INSERT(I);
  case TargetOpcode::G_INTRINSIC:
    return selectG_INTRINSIC(I, CoverageInfo);
  case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
    return selectG_INTRINSIC_W_SIDE_EFFECTS(I, CoverageInfo);
  case TargetOpcode::G_ICMP:
    if (selectG_ICMP(I))
      return true;
    return selectImpl(I, CoverageInfo);
  case TargetOpcode::G_LOAD:
    return selectImpl(I, CoverageInfo);
  case TargetOpcode::G_SELECT:
    return selectG_SELECT(I);
  case TargetOpcode::G_STORE:
    if (selectImpl(I, CoverageInfo))
      return true;
    return selectG_STORE(I);
  case TargetOpcode::G_TRUNC:
    return selectG_TRUNC(I);
  case TargetOpcode::G_SEXT:
  case TargetOpcode::G_ZEXT:
  case TargetOpcode::G_ANYEXT:
    if (selectG_SZA_EXT(I)) {
      I.eraseFromParent();
      return true;
    }

    return false;
  case TargetOpcode::G_BRCOND:
    return selectG_BRCOND(I);
  case TargetOpcode::G_FRAME_INDEX:
    return selectG_FRAME_INDEX(I);
  case TargetOpcode::G_FENCE:
    // FIXME: Tablegen importer doesn't handle the imm operands correctly, and
    // is checking for G_CONSTANT
    I.setDesc(TII.get(AMDGPU::ATOMIC_FENCE));
    return true;
  }
  return false;
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVCSRC(MachineOperand &Root) const {
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.add(Root); }
  }};

}

std::pair<Register, unsigned>
AMDGPUInstructionSelector::selectVOP3ModsImpl(
  Register Src, const MachineRegisterInfo &MRI) const {
  unsigned Mods = 0;
  MachineInstr *MI = MRI.getVRegDef(Src);

  if (MI && MI->getOpcode() == AMDGPU::G_FNEG) {
    Src = MI->getOperand(1).getReg();
    Mods |= SISrcMods::NEG;
    MI = MRI.getVRegDef(Src);
  }

  if (MI && MI->getOpcode() == AMDGPU::G_FABS) {
    Src = MI->getOperand(1).getReg();
    Mods |= SISrcMods::ABS;
  }

  return std::make_pair(Src, Mods);
}

///
/// This will select either an SGPR or VGPR operand and will save us from
/// having to write an extra tablegen pattern.
InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVSRC0(MachineOperand &Root) const {
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.add(Root); }
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3Mods0(MachineOperand &Root) const {
  MachineRegisterInfo &MRI
    = Root.getParent()->getParent()->getParent()->getRegInfo();

  Register Src;
  unsigned Mods;
  std::tie(Src, Mods) = selectVOP3ModsImpl(Root.getReg(), MRI);

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Src); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Mods); }, // src0_mods
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); },    // clamp
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }     // omod
  }};
}
InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3OMods(MachineOperand &Root) const {
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.add(Root); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }, // clamp
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // omod
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3Mods(MachineOperand &Root) const {
  MachineRegisterInfo &MRI
    = Root.getParent()->getParent()->getParent()->getRegInfo();

  Register Src;
  unsigned Mods;
  std::tie(Src, Mods) = selectVOP3ModsImpl(Root.getReg(), MRI);

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Src); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Mods); }  // src_mods
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectSmrdImm(MachineOperand &Root) const {
  MachineRegisterInfo &MRI =
      Root.getParent()->getParent()->getParent()->getRegInfo();

  SmallVector<GEPInfo, 4> AddrInfo;
  getAddrModeInfo(*Root.getParent(), MRI, AddrInfo);

  if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
    return None;

  const GEPInfo &GEPInfo = AddrInfo[0];

  if (!AMDGPU::isLegalSMRDImmOffset(STI, GEPInfo.Imm))
    return None;

  unsigned PtrReg = GEPInfo.SgprParts[0];
  int64_t EncodedImm = AMDGPU::getSMRDEncodedOffset(STI, GEPInfo.Imm);
  return {{
    [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
    [=](MachineInstrBuilder &MIB) { MIB.addImm(EncodedImm); }
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectSmrdImm32(MachineOperand &Root) const {
  MachineRegisterInfo &MRI =
      Root.getParent()->getParent()->getParent()->getRegInfo();

  SmallVector<GEPInfo, 4> AddrInfo;
  getAddrModeInfo(*Root.getParent(), MRI, AddrInfo);

  if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
    return None;

  const GEPInfo &GEPInfo = AddrInfo[0];
  unsigned PtrReg = GEPInfo.SgprParts[0];
  int64_t EncodedImm = AMDGPU::getSMRDEncodedOffset(STI, GEPInfo.Imm);
  if (!isUInt<32>(EncodedImm))
    return None;

  return {{
    [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
    [=](MachineInstrBuilder &MIB) { MIB.addImm(EncodedImm); }
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectSmrdSgpr(MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();
  MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();

  SmallVector<GEPInfo, 4> AddrInfo;
  getAddrModeInfo(*MI, MRI, AddrInfo);

  // FIXME: We should shrink the GEP if the offset is known to be <= 32-bits,
  // then we can select all ptr + 32-bit offsets not just immediate offsets.
  if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
    return None;

  const GEPInfo &GEPInfo = AddrInfo[0];
  if (!GEPInfo.Imm || !isUInt<32>(GEPInfo.Imm))
    return None;

  // If we make it this far we have a load with an 32-bit immediate offset.
  // It is OK to select this using a sgpr offset, because we have already
  // failed trying to select this load into one of the _IMM variants since
  // the _IMM Patterns are considered before the _SGPR patterns.
  unsigned PtrReg = GEPInfo.SgprParts[0];
  unsigned OffsetReg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
  BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::S_MOV_B32), OffsetReg)
          .addImm(GEPInfo.Imm);
  return {{
    [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
    [=](MachineInstrBuilder &MIB) { MIB.addReg(OffsetReg); }
  }};
}

template <bool Signed>
InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectFlatOffsetImpl(MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();
  MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();

  InstructionSelector::ComplexRendererFns Default = {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Root.getReg()); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); },  // offset
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // slc
    }};

  if (!STI.hasFlatInstOffsets())
    return Default;

  const MachineInstr *OpDef = MRI.getVRegDef(Root.getReg());
  if (!OpDef || OpDef->getOpcode() != AMDGPU::G_GEP)
    return Default;

  Optional<int64_t> Offset =
    getConstantVRegVal(OpDef->getOperand(2).getReg(), MRI);
  if (!Offset.hasValue())
    return Default;

  unsigned AddrSpace = (*MI->memoperands_begin())->getAddrSpace();
  if (!TII.isLegalFLATOffset(Offset.getValue(), AddrSpace, Signed))
    return Default;

  Register BasePtr = OpDef->getOperand(1).getReg();

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(BasePtr); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Offset.getValue()); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // slc
    }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectFlatOffset(MachineOperand &Root) const {
  return selectFlatOffsetImpl<false>(Root);
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectFlatOffsetSigned(MachineOperand &Root) const {
  return selectFlatOffsetImpl<true>(Root);
}

// FIXME: Implement
static bool signBitIsZero(const MachineOperand &Op,
                          const MachineRegisterInfo &MRI) {
  return false;
}

static bool isStackPtrRelative(const MachinePointerInfo &PtrInfo) {
  auto PSV = PtrInfo.V.dyn_cast<const PseudoSourceValue *>();
  return PSV && PSV->isStack();
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectMUBUFScratchOffen(MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();
  MachineFunction *MF = MBB->getParent();
  MachineRegisterInfo &MRI = MF->getRegInfo();
  const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();

  int64_t Offset = 0;
  if (mi_match(Root.getReg(), MRI, m_ICst(Offset))) {
    Register HighBits = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);

    // TODO: Should this be inside the render function? The iterator seems to
    // move.
    BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::V_MOV_B32_e32),
            HighBits)
      .addImm(Offset & ~4095);

    return {{[=](MachineInstrBuilder &MIB) { // rsrc
               MIB.addReg(Info->getScratchRSrcReg());
             },
             [=](MachineInstrBuilder &MIB) { // vaddr
               MIB.addReg(HighBits);
             },
             [=](MachineInstrBuilder &MIB) { // soffset
               const MachineMemOperand *MMO = *MI->memoperands_begin();
               const MachinePointerInfo &PtrInfo = MMO->getPointerInfo();

               Register SOffsetReg = isStackPtrRelative(PtrInfo)
                                         ? Info->getStackPtrOffsetReg()
                                         : Info->getScratchWaveOffsetReg();
               MIB.addReg(SOffsetReg);
             },
             [=](MachineInstrBuilder &MIB) { // offset
               MIB.addImm(Offset & 4095);
             }}};
  }

  assert(Offset == 0);

  // Try to fold a frame index directly into the MUBUF vaddr field, and any
  // offsets.
  Optional<int> FI;
  Register VAddr = Root.getReg();
  if (const MachineInstr *RootDef = MRI.getVRegDef(Root.getReg())) {
    if (isBaseWithConstantOffset(Root, MRI)) {
      const MachineOperand &LHS = RootDef->getOperand(1);
      const MachineOperand &RHS = RootDef->getOperand(2);
      const MachineInstr *LHSDef = MRI.getVRegDef(LHS.getReg());
      const MachineInstr *RHSDef = MRI.getVRegDef(RHS.getReg());
      if (LHSDef && RHSDef) {
        int64_t PossibleOffset =
            RHSDef->getOperand(1).getCImm()->getSExtValue();
        if (SIInstrInfo::isLegalMUBUFImmOffset(PossibleOffset) &&
            (!STI.privateMemoryResourceIsRangeChecked() ||
             signBitIsZero(LHS, MRI))) {
          if (LHSDef->getOpcode() == AMDGPU::G_FRAME_INDEX)
            FI = LHSDef->getOperand(1).getIndex();
          else
            VAddr = LHS.getReg();
          Offset = PossibleOffset;
        }
      }
    } else if (RootDef->getOpcode() == AMDGPU::G_FRAME_INDEX) {
      FI = RootDef->getOperand(1).getIndex();
    }
  }

  // If we don't know this private access is a local stack object, it needs to
  // be relative to the entry point's scratch wave offset register.
  // TODO: Should split large offsets that don't fit like above.
  // TODO: Don't use scratch wave offset just because the offset didn't fit.
  Register SOffset = FI.hasValue() ? Info->getStackPtrOffsetReg()
                                   : Info->getScratchWaveOffsetReg();

  return {{[=](MachineInstrBuilder &MIB) { // rsrc
             MIB.addReg(Info->getScratchRSrcReg());
           },
           [=](MachineInstrBuilder &MIB) { // vaddr
             if (FI.hasValue())
               MIB.addFrameIndex(FI.getValue());
             else
               MIB.addReg(VAddr);
           },
           [=](MachineInstrBuilder &MIB) { // soffset
             MIB.addReg(SOffset);
           },
           [=](MachineInstrBuilder &MIB) { // offset
             MIB.addImm(Offset);
           }}};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectMUBUFScratchOffset(
    MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();
  MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();

  int64_t Offset = 0;
  if (!mi_match(Root.getReg(), MRI, m_ICst(Offset)) ||
      !SIInstrInfo::isLegalMUBUFImmOffset(Offset))
    return {};

  const MachineFunction *MF = MBB->getParent();
  const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
  const MachineMemOperand *MMO = *MI->memoperands_begin();
  const MachinePointerInfo &PtrInfo = MMO->getPointerInfo();

  Register SOffsetReg = isStackPtrRelative(PtrInfo)
                            ? Info->getStackPtrOffsetReg()
                            : Info->getScratchWaveOffsetReg();
  return {{
      [=](MachineInstrBuilder &MIB) {
        MIB.addReg(Info->getScratchRSrcReg());
      },                                                         // rsrc
      [=](MachineInstrBuilder &MIB) { MIB.addReg(SOffsetReg); }, // soffset
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Offset); }      // offset
  }};
}