llvm.org GIT mirror llvm / 0392724 lib / Target / AMDGPU / SILoadStoreOptimizer.cpp
0392724

Tree @0392724 (Download .tar.gz)

SILoadStoreOptimizer.cpp @0392724raw · 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
//===- SILoadStoreOptimizer.cpp -------------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This pass tries to fuse DS instructions with close by immediate offsets.
// This will fuse operations such as
//  ds_read_b32 v0, v2 offset:16
//  ds_read_b32 v1, v2 offset:32
// ==>
//   ds_read2_b32 v[0:1], v2, offset0:4 offset1:8
//
// The same is done for certain SMEM and VMEM opcodes, e.g.:
//  s_buffer_load_dword s4, s[0:3], 4
//  s_buffer_load_dword s5, s[0:3], 8
// ==>
//  s_buffer_load_dwordx2 s[4:5], s[0:3], 4
//
// This pass also tries to promote constant offset to the immediate by
// adjusting the base. It tries to use a base from the nearby instructions that
// allows it to have a 13bit constant offset and then promotes the 13bit offset
// to the immediate.
// E.g.
//  s_movk_i32 s0, 0x1800
//  v_add_co_u32_e32 v0, vcc, s0, v2
//  v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
//
//  s_movk_i32 s0, 0x1000
//  v_add_co_u32_e32 v5, vcc, s0, v2
//  v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
//  global_load_dwordx2 v[5:6], v[5:6], off
//  global_load_dwordx2 v[0:1], v[0:1], off
// =>
//  s_movk_i32 s0, 0x1000
//  v_add_co_u32_e32 v5, vcc, s0, v2
//  v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
//  global_load_dwordx2 v[5:6], v[5:6], off
//  global_load_dwordx2 v[0:1], v[5:6], off offset:2048
//
// Future improvements:
//
// - This currently relies on the scheduler to place loads and stores next to
//   each other, and then only merges adjacent pairs of instructions. It would
//   be good to be more flexible with interleaved instructions, and possibly run
//   before scheduling. It currently missing stores of constants because loading
//   the constant into the data register is placed between the stores, although
//   this is arguably a scheduling problem.
//
// - Live interval recomputing seems inefficient. This currently only matches
//   one pair, and recomputes live intervals and moves on to the next pair. It
//   would be better to compute a list of all merges that need to occur.
//
// - With a list of instructions to process, we can also merge more. If a
//   cluster of loads have offsets that are too large to fit in the 8-bit
//   offsets, but are close enough to fit in the 8 bits, we can add to the base
//   pointer and use the new reduced offsets.
//
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIInstrInfo.h"
#include "SIRegisterInfo.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/Pass.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <cstdlib>
#include <iterator>
#include <utility>

using namespace llvm;

#define DEBUG_TYPE "si-load-store-opt"

namespace {
enum InstClassEnum {
  UNKNOWN,
  DS_READ,
  DS_WRITE,
  S_BUFFER_LOAD_IMM,
  BUFFER_LOAD_OFFEN = AMDGPU::BUFFER_LOAD_DWORD_OFFEN,
  BUFFER_LOAD_OFFSET = AMDGPU::BUFFER_LOAD_DWORD_OFFSET,
  BUFFER_STORE_OFFEN = AMDGPU::BUFFER_STORE_DWORD_OFFEN,
  BUFFER_STORE_OFFSET = AMDGPU::BUFFER_STORE_DWORD_OFFSET,
  BUFFER_LOAD_OFFEN_exact = AMDGPU::BUFFER_LOAD_DWORD_OFFEN_exact,
  BUFFER_LOAD_OFFSET_exact = AMDGPU::BUFFER_LOAD_DWORD_OFFSET_exact,
  BUFFER_STORE_OFFEN_exact = AMDGPU::BUFFER_STORE_DWORD_OFFEN_exact,
  BUFFER_STORE_OFFSET_exact = AMDGPU::BUFFER_STORE_DWORD_OFFSET_exact,
};

enum RegisterEnum {
  SBASE = 0x1,
  SRSRC = 0x2,
  SOFFSET = 0x4,
  VADDR = 0x8,
  ADDR = 0x10,
};

class SILoadStoreOptimizer : public MachineFunctionPass {
  struct CombineInfo {
    MachineBasicBlock::iterator I;
    MachineBasicBlock::iterator Paired;
    unsigned EltSize;
    unsigned Offset0;
    unsigned Offset1;
    unsigned Width0;
    unsigned Width1;
    unsigned BaseOff;
    InstClassEnum InstClass;
    bool GLC0;
    bool GLC1;
    bool SLC0;
    bool SLC1;
    bool DLC0;
    bool DLC1;
    bool UseST64;
    SmallVector<MachineInstr *, 8> InstsToMove;
  };

  struct BaseRegisters {
    unsigned LoReg = 0;
    unsigned HiReg = 0;

    unsigned LoSubReg = 0;
    unsigned HiSubReg = 0;
  };

  struct MemAddress {
    BaseRegisters Base;
    int64_t Offset = 0;
  };

  using MemInfoMap = DenseMap<MachineInstr *, MemAddress>;

private:
  const GCNSubtarget *STM = nullptr;
  const SIInstrInfo *TII = nullptr;
  const SIRegisterInfo *TRI = nullptr;
  MachineRegisterInfo *MRI = nullptr;
  AliasAnalysis *AA = nullptr;
  bool OptimizeAgain;

  static bool offsetsCanBeCombined(CombineInfo &CI);
  static bool widthsFit(const GCNSubtarget &STM, const CombineInfo &CI);
  static unsigned getNewOpcode(const CombineInfo &CI);
  static std::pair<unsigned, unsigned> getSubRegIdxs(const CombineInfo &CI);
  const TargetRegisterClass *getTargetRegisterClass(const CombineInfo &CI);
  unsigned getOpcodeWidth(const MachineInstr &MI);
  InstClassEnum getInstClass(unsigned Opc);
  unsigned getRegs(unsigned Opc);

  bool findMatchingInst(CombineInfo &CI);

  unsigned read2Opcode(unsigned EltSize) const;
  unsigned read2ST64Opcode(unsigned EltSize) const;
  MachineBasicBlock::iterator mergeRead2Pair(CombineInfo &CI);

  unsigned write2Opcode(unsigned EltSize) const;
  unsigned write2ST64Opcode(unsigned EltSize) const;
  MachineBasicBlock::iterator mergeWrite2Pair(CombineInfo &CI);
  MachineBasicBlock::iterator mergeSBufferLoadImmPair(CombineInfo &CI);
  MachineBasicBlock::iterator mergeBufferLoadPair(CombineInfo &CI);
  MachineBasicBlock::iterator mergeBufferStorePair(CombineInfo &CI);

  void updateBaseAndOffset(MachineInstr &I, unsigned NewBase,
                           int32_t NewOffset);
  unsigned computeBase(MachineInstr &MI, const MemAddress &Addr);
  MachineOperand createRegOrImm(int32_t Val, MachineInstr &MI);
  Optional<int32_t> extractConstOffset(const MachineOperand &Op);
  void processBaseWithConstOffset(const MachineOperand &Base, MemAddress &Addr);
  /// Promotes constant offset to the immediate by adjusting the base. It
  /// tries to use a base from the nearby instructions that allows it to have
  /// a 13bit constant offset which gets promoted to the immediate.
  bool promoteConstantOffsetToImm(MachineInstr &CI,
                                  MemInfoMap &Visited,
                                  SmallPtrSet<MachineInstr *, 4> &Promoted);

public:
  static char ID;

  SILoadStoreOptimizer() : MachineFunctionPass(ID) {
    initializeSILoadStoreOptimizerPass(*PassRegistry::getPassRegistry());
  }

  bool optimizeBlock(MachineBasicBlock &MBB);

  bool runOnMachineFunction(MachineFunction &MF) override;

  StringRef getPassName() const override { return "SI Load Store Optimizer"; }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<AAResultsWrapperPass>();

    MachineFunctionPass::getAnalysisUsage(AU);
  }
};

} // end anonymous namespace.

INITIALIZE_PASS_BEGIN(SILoadStoreOptimizer, DEBUG_TYPE,
                      "SI Load Store Optimizer", false, false)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_END(SILoadStoreOptimizer, DEBUG_TYPE, "SI Load Store Optimizer",
                    false, false)

char SILoadStoreOptimizer::ID = 0;

char &llvm::SILoadStoreOptimizerID = SILoadStoreOptimizer::ID;

FunctionPass *llvm::createSILoadStoreOptimizerPass() {
  return new SILoadStoreOptimizer();
}

static void moveInstsAfter(MachineBasicBlock::iterator I,
                           ArrayRef<MachineInstr *> InstsToMove) {
  MachineBasicBlock *MBB = I->getParent();
  ++I;
  for (MachineInstr *MI : InstsToMove) {
    MI->removeFromParent();
    MBB->insert(I, MI);
  }
}

static void addDefsUsesToList(const MachineInstr &MI,
                              DenseSet<unsigned> &RegDefs,
                              DenseSet<unsigned> &PhysRegUses) {
  for (const MachineOperand &Op : MI.operands()) {
    if (Op.isReg()) {
      if (Op.isDef())
        RegDefs.insert(Op.getReg());
      else if (Op.readsReg() &&
               TargetRegisterInfo::isPhysicalRegister(Op.getReg()))
        PhysRegUses.insert(Op.getReg());
    }
  }
}

static bool memAccessesCanBeReordered(MachineBasicBlock::iterator A,
                                      MachineBasicBlock::iterator B,
                                      AliasAnalysis *AA) {
  // RAW or WAR - cannot reorder
  // WAW - cannot reorder
  // RAR - safe to reorder
  return !(A->mayStore() || B->mayStore()) || !A->mayAlias(AA, *B, true);
}

// Add MI and its defs to the lists if MI reads one of the defs that are
// already in the list. Returns true in that case.
static bool addToListsIfDependent(MachineInstr &MI, DenseSet<unsigned> &RegDefs,
                                  DenseSet<unsigned> &PhysRegUses,
                                  SmallVectorImpl<MachineInstr *> &Insts) {
  for (MachineOperand &Use : MI.operands()) {
    // If one of the defs is read, then there is a use of Def between I and the
    // instruction that I will potentially be merged with. We will need to move
    // this instruction after the merged instructions.
    //
    // Similarly, if there is a def which is read by an instruction that is to
    // be moved for merging, then we need to move the def-instruction as well.
    // This can only happen for physical registers such as M0; virtual
    // registers are in SSA form.
    if (Use.isReg() &&
        ((Use.readsReg() && RegDefs.count(Use.getReg())) ||
         (Use.isDef() && RegDefs.count(Use.getReg())) ||
         (Use.isDef() && TargetRegisterInfo::isPhysicalRegister(Use.getReg()) &&
          PhysRegUses.count(Use.getReg())))) {
      Insts.push_back(&MI);
      addDefsUsesToList(MI, RegDefs, PhysRegUses);
      return true;
    }
  }

  return false;
}

static bool canMoveInstsAcrossMemOp(MachineInstr &MemOp,
                                    ArrayRef<MachineInstr *> InstsToMove,
                                    AliasAnalysis *AA) {
  assert(MemOp.mayLoadOrStore());

  for (MachineInstr *InstToMove : InstsToMove) {
    if (!InstToMove->mayLoadOrStore())
      continue;
    if (!memAccessesCanBeReordered(MemOp, *InstToMove, AA))
      return false;
  }
  return true;
}

bool SILoadStoreOptimizer::offsetsCanBeCombined(CombineInfo &CI) {
  // XXX - Would the same offset be OK? Is there any reason this would happen or
  // be useful?
  if (CI.Offset0 == CI.Offset1)
    return false;

  // This won't be valid if the offset isn't aligned.
  if ((CI.Offset0 % CI.EltSize != 0) || (CI.Offset1 % CI.EltSize != 0))
    return false;

  unsigned EltOffset0 = CI.Offset0 / CI.EltSize;
  unsigned EltOffset1 = CI.Offset1 / CI.EltSize;
  CI.UseST64 = false;
  CI.BaseOff = 0;

  // Handle SMEM and VMEM instructions.
  if ((CI.InstClass != DS_READ) && (CI.InstClass != DS_WRITE)) {
    return (EltOffset0 + CI.Width0 == EltOffset1 ||
            EltOffset1 + CI.Width1 == EltOffset0) &&
           CI.GLC0 == CI.GLC1 && CI.DLC0 == CI.DLC1 &&
           (CI.InstClass == S_BUFFER_LOAD_IMM || CI.SLC0 == CI.SLC1);
  }

  // If the offset in elements doesn't fit in 8-bits, we might be able to use
  // the stride 64 versions.
  if ((EltOffset0 % 64 == 0) && (EltOffset1 % 64) == 0 &&
      isUInt<8>(EltOffset0 / 64) && isUInt<8>(EltOffset1 / 64)) {
    CI.Offset0 = EltOffset0 / 64;
    CI.Offset1 = EltOffset1 / 64;
    CI.UseST64 = true;
    return true;
  }

  // Check if the new offsets fit in the reduced 8-bit range.
  if (isUInt<8>(EltOffset0) && isUInt<8>(EltOffset1)) {
    CI.Offset0 = EltOffset0;
    CI.Offset1 = EltOffset1;
    return true;
  }

  // Try to shift base address to decrease offsets.
  unsigned OffsetDiff = std::abs((int)EltOffset1 - (int)EltOffset0);
  CI.BaseOff = std::min(CI.Offset0, CI.Offset1);

  if ((OffsetDiff % 64 == 0) && isUInt<8>(OffsetDiff / 64)) {
    CI.Offset0 = (EltOffset0 - CI.BaseOff / CI.EltSize) / 64;
    CI.Offset1 = (EltOffset1 - CI.BaseOff / CI.EltSize) / 64;
    CI.UseST64 = true;
    return true;
  }

  if (isUInt<8>(OffsetDiff)) {
    CI.Offset0 = EltOffset0 - CI.BaseOff / CI.EltSize;
    CI.Offset1 = EltOffset1 - CI.BaseOff / CI.EltSize;
    return true;
  }

  return false;
}

bool SILoadStoreOptimizer::widthsFit(const GCNSubtarget &STM,
                                     const CombineInfo &CI) {
  const unsigned Width = (CI.Width0 + CI.Width1);
  switch (CI.InstClass) {
  default:
    return (Width <= 4) && (STM.hasDwordx3LoadStores() || (Width != 3));
  case S_BUFFER_LOAD_IMM:
    switch (Width) {
    default:
      return false;
    case 2:
    case 4:
      return true;
    }
  }
}

unsigned SILoadStoreOptimizer::getOpcodeWidth(const MachineInstr &MI) {
  const unsigned Opc = MI.getOpcode();

  if (TII->isMUBUF(MI)) {
    return AMDGPU::getMUBUFDwords(Opc);
  }

  switch (Opc) {
  default:
    return 0;
  case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
    return 1;
  case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
    return 2;
  case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
    return 4;
  }
}

InstClassEnum SILoadStoreOptimizer::getInstClass(unsigned Opc) {
  if (TII->isMUBUF(Opc)) {
    const int baseOpcode = AMDGPU::getMUBUFBaseOpcode(Opc);

    // If we couldn't identify the opcode, bail out.
    if (baseOpcode == -1) {
      return UNKNOWN;
    }

    switch (baseOpcode) {
    default:
      return UNKNOWN;
    case AMDGPU::BUFFER_LOAD_DWORD_OFFEN:
      return BUFFER_LOAD_OFFEN;
    case AMDGPU::BUFFER_LOAD_DWORD_OFFSET:
      return BUFFER_LOAD_OFFSET;
    case AMDGPU::BUFFER_STORE_DWORD_OFFEN:
      return BUFFER_STORE_OFFEN;
    case AMDGPU::BUFFER_STORE_DWORD_OFFSET:
      return BUFFER_STORE_OFFSET;
    case AMDGPU::BUFFER_LOAD_DWORD_OFFEN_exact:
      return BUFFER_LOAD_OFFEN_exact;
    case AMDGPU::BUFFER_LOAD_DWORD_OFFSET_exact:
      return BUFFER_LOAD_OFFSET_exact;
    case AMDGPU::BUFFER_STORE_DWORD_OFFEN_exact:
      return BUFFER_STORE_OFFEN_exact;
    case AMDGPU::BUFFER_STORE_DWORD_OFFSET_exact:
      return BUFFER_STORE_OFFSET_exact;
    }
  }

  switch (Opc) {
  default:
    return UNKNOWN;
  case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
  case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
  case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
    return S_BUFFER_LOAD_IMM;
  case AMDGPU::DS_READ_B32:
  case AMDGPU::DS_READ_B64:
  case AMDGPU::DS_READ_B32_gfx9:
  case AMDGPU::DS_READ_B64_gfx9:
    return DS_READ;
  case AMDGPU::DS_WRITE_B32:
  case AMDGPU::DS_WRITE_B64:
  case AMDGPU::DS_WRITE_B32_gfx9:
  case AMDGPU::DS_WRITE_B64_gfx9:
    return DS_WRITE;
  }
}

unsigned SILoadStoreOptimizer::getRegs(unsigned Opc) {
  if (TII->isMUBUF(Opc)) {
    unsigned result = 0;

    if (AMDGPU::getMUBUFHasVAddr(Opc)) {
      result |= VADDR;
    }

    if (AMDGPU::getMUBUFHasSrsrc(Opc)) {
      result |= SRSRC;
    }

    if (AMDGPU::getMUBUFHasSoffset(Opc)) {
      result |= SOFFSET;
    }

    return result;
  }

  switch (Opc) {
  default:
    return 0;
  case AMDGPU::S_BUFFER_LOAD_DWORD_IMM:
  case AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM:
  case AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM:
    return SBASE;
  case AMDGPU::DS_READ_B32:
  case AMDGPU::DS_READ_B64:
  case AMDGPU::DS_READ_B32_gfx9:
  case AMDGPU::DS_READ_B64_gfx9:
  case AMDGPU::DS_WRITE_B32:
  case AMDGPU::DS_WRITE_B64:
  case AMDGPU::DS_WRITE_B32_gfx9:
  case AMDGPU::DS_WRITE_B64_gfx9:
    return ADDR;
  }
}

bool SILoadStoreOptimizer::findMatchingInst(CombineInfo &CI) {
  MachineBasicBlock *MBB = CI.I->getParent();
  MachineBasicBlock::iterator E = MBB->end();
  MachineBasicBlock::iterator MBBI = CI.I;

  const unsigned Opc = CI.I->getOpcode();
  const InstClassEnum InstClass = getInstClass(Opc);

  if (InstClass == UNKNOWN) {
    return false;
  }

  const unsigned Regs = getRegs(Opc);

  unsigned AddrOpName[5] = {0};
  int AddrIdx[5];
  const MachineOperand *AddrReg[5];
  unsigned NumAddresses = 0;

  if (Regs & ADDR) {
    AddrOpName[NumAddresses++] = AMDGPU::OpName::addr;
  }

  if (Regs & SBASE) {
    AddrOpName[NumAddresses++] = AMDGPU::OpName::sbase;
  }

  if (Regs & SRSRC) {
    AddrOpName[NumAddresses++] = AMDGPU::OpName::srsrc;
  }

  if (Regs & SOFFSET) {
    AddrOpName[NumAddresses++] = AMDGPU::OpName::soffset;
  }

  if (Regs & VADDR) {
    AddrOpName[NumAddresses++] = AMDGPU::OpName::vaddr;
  }

  for (unsigned i = 0; i < NumAddresses; i++) {
    AddrIdx[i] = AMDGPU::getNamedOperandIdx(CI.I->getOpcode(), AddrOpName[i]);
    AddrReg[i] = &CI.I->getOperand(AddrIdx[i]);

    // We only ever merge operations with the same base address register, so
    // don't bother scanning forward if there are no other uses.
    if (AddrReg[i]->isReg() &&
        (TargetRegisterInfo::isPhysicalRegister(AddrReg[i]->getReg()) ||
         MRI->hasOneNonDBGUse(AddrReg[i]->getReg())))
      return false;
  }

  ++MBBI;

  DenseSet<unsigned> RegDefsToMove;
  DenseSet<unsigned> PhysRegUsesToMove;
  addDefsUsesToList(*CI.I, RegDefsToMove, PhysRegUsesToMove);

  for (; MBBI != E; ++MBBI) {
    const bool IsDS = (InstClass == DS_READ) || (InstClass == DS_WRITE);

    if ((getInstClass(MBBI->getOpcode()) != InstClass) ||
        (IsDS && (MBBI->getOpcode() != Opc))) {
      // This is not a matching DS instruction, but we can keep looking as
      // long as one of these conditions are met:
      // 1. It is safe to move I down past MBBI.
      // 2. It is safe to move MBBI down past the instruction that I will
      //    be merged into.

      if (MBBI->hasUnmodeledSideEffects()) {
        // We can't re-order this instruction with respect to other memory
        // operations, so we fail both conditions mentioned above.
        return false;
      }

      if (MBBI->mayLoadOrStore() &&
          (!memAccessesCanBeReordered(*CI.I, *MBBI, AA) ||
           !canMoveInstsAcrossMemOp(*MBBI, CI.InstsToMove, AA))) {
        // We fail condition #1, but we may still be able to satisfy condition
        // #2.  Add this instruction to the move list and then we will check
        // if condition #2 holds once we have selected the matching instruction.
        CI.InstsToMove.push_back(&*MBBI);
        addDefsUsesToList(*MBBI, RegDefsToMove, PhysRegUsesToMove);
        continue;
      }

      // When we match I with another DS instruction we will be moving I down
      // to the location of the matched instruction any uses of I will need to
      // be moved down as well.
      addToListsIfDependent(*MBBI, RegDefsToMove, PhysRegUsesToMove,
                            CI.InstsToMove);
      continue;
    }

    // Don't merge volatiles.
    if (MBBI->hasOrderedMemoryRef())
      return false;

    // Handle a case like
    //   DS_WRITE_B32 addr, v, idx0
    //   w = DS_READ_B32 addr, idx0
    //   DS_WRITE_B32 addr, f(w), idx1
    // where the DS_READ_B32 ends up in InstsToMove and therefore prevents
    // merging of the two writes.
    if (addToListsIfDependent(*MBBI, RegDefsToMove, PhysRegUsesToMove,
                              CI.InstsToMove))
      continue;

    bool Match = true;
    for (unsigned i = 0; i < NumAddresses; i++) {
      const MachineOperand &AddrRegNext = MBBI->getOperand(AddrIdx[i]);

      if (AddrReg[i]->isImm() || AddrRegNext.isImm()) {
        if (AddrReg[i]->isImm() != AddrRegNext.isImm() ||
            AddrReg[i]->getImm() != AddrRegNext.getImm()) {
          Match = false;
          break;
        }
        continue;
      }

      // Check same base pointer. Be careful of subregisters, which can occur
      // with vectors of pointers.
      if (AddrReg[i]->getReg() != AddrRegNext.getReg() ||
          AddrReg[i]->getSubReg() != AddrRegNext.getSubReg()) {
        Match = false;
        break;
      }
    }

    if (Match) {
      int OffsetIdx =
          AMDGPU::getNamedOperandIdx(CI.I->getOpcode(), AMDGPU::OpName::offset);
      CI.Offset0 = CI.I->getOperand(OffsetIdx).getImm();
      CI.Width0 = getOpcodeWidth(*CI.I);
      CI.Offset1 = MBBI->getOperand(OffsetIdx).getImm();
      CI.Width1 = getOpcodeWidth(*MBBI);
      CI.Paired = MBBI;

      if ((CI.InstClass == DS_READ) || (CI.InstClass == DS_WRITE)) {
        CI.Offset0 &= 0xffff;
        CI.Offset1 &= 0xffff;
      } else {
        CI.GLC0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::glc)->getImm();
        CI.GLC1 = TII->getNamedOperand(*MBBI, AMDGPU::OpName::glc)->getImm();
        if (CI.InstClass != S_BUFFER_LOAD_IMM) {
          CI.SLC0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::slc)->getImm();
          CI.SLC1 = TII->getNamedOperand(*MBBI, AMDGPU::OpName::slc)->getImm();
        }
        CI.DLC0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::dlc)->getImm();
        CI.DLC1 = TII->getNamedOperand(*MBBI, AMDGPU::OpName::dlc)->getImm();
      }

      // Check both offsets fit in the reduced range.
      // We also need to go through the list of instructions that we plan to
      // move and make sure they are all safe to move down past the merged
      // instruction.
      if (widthsFit(*STM, CI) && offsetsCanBeCombined(CI))
        if (canMoveInstsAcrossMemOp(*MBBI, CI.InstsToMove, AA))
          return true;
    }

    // We've found a load/store that we couldn't merge for some reason.
    // We could potentially keep looking, but we'd need to make sure that
    // it was safe to move I and also all the instruction in InstsToMove
    // down past this instruction.
    // check if we can move I across MBBI and if we can move all I's users
    if (!memAccessesCanBeReordered(*CI.I, *MBBI, AA) ||
        !canMoveInstsAcrossMemOp(*MBBI, CI.InstsToMove, AA))
      break;
  }
  return false;
}

unsigned SILoadStoreOptimizer::read2Opcode(unsigned EltSize) const {
  if (STM->ldsRequiresM0Init())
    return (EltSize == 4) ? AMDGPU::DS_READ2_B32 : AMDGPU::DS_READ2_B64;
  return (EltSize == 4) ? AMDGPU::DS_READ2_B32_gfx9 : AMDGPU::DS_READ2_B64_gfx9;
}

unsigned SILoadStoreOptimizer::read2ST64Opcode(unsigned EltSize) const {
  if (STM->ldsRequiresM0Init())
    return (EltSize == 4) ? AMDGPU::DS_READ2ST64_B32 : AMDGPU::DS_READ2ST64_B64;

  return (EltSize == 4) ? AMDGPU::DS_READ2ST64_B32_gfx9
                        : AMDGPU::DS_READ2ST64_B64_gfx9;
}

MachineBasicBlock::iterator
SILoadStoreOptimizer::mergeRead2Pair(CombineInfo &CI) {
  MachineBasicBlock *MBB = CI.I->getParent();

  // Be careful, since the addresses could be subregisters themselves in weird
  // cases, like vectors of pointers.
  const auto *AddrReg = TII->getNamedOperand(*CI.I, AMDGPU::OpName::addr);

  const auto *Dest0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::vdst);
  const auto *Dest1 = TII->getNamedOperand(*CI.Paired, AMDGPU::OpName::vdst);

  unsigned NewOffset0 = CI.Offset0;
  unsigned NewOffset1 = CI.Offset1;
  unsigned Opc =
      CI.UseST64 ? read2ST64Opcode(CI.EltSize) : read2Opcode(CI.EltSize);

  unsigned SubRegIdx0 = (CI.EltSize == 4) ? AMDGPU::sub0 : AMDGPU::sub0_sub1;
  unsigned SubRegIdx1 = (CI.EltSize == 4) ? AMDGPU::sub1 : AMDGPU::sub2_sub3;

  if (NewOffset0 > NewOffset1) {
    // Canonicalize the merged instruction so the smaller offset comes first.
    std::swap(NewOffset0, NewOffset1);
    std::swap(SubRegIdx0, SubRegIdx1);
  }

  assert((isUInt<8>(NewOffset0) && isUInt<8>(NewOffset1)) &&
         (NewOffset0 != NewOffset1) && "Computed offset doesn't fit");

  const MCInstrDesc &Read2Desc = TII->get(Opc);

  const TargetRegisterClass *SuperRC =
      (CI.EltSize == 4) ? &AMDGPU::VReg_64RegClass : &AMDGPU::VReg_128RegClass;
  unsigned DestReg = MRI->createVirtualRegister(SuperRC);

  DebugLoc DL = CI.I->getDebugLoc();

  unsigned BaseReg = AddrReg->getReg();
  unsigned BaseSubReg = AddrReg->getSubReg();
  unsigned BaseRegFlags = 0;
  if (CI.BaseOff) {
    unsigned ImmReg = MRI->createVirtualRegister(&AMDGPU::SGPR_32RegClass);
    BuildMI(*MBB, CI.Paired, DL, TII->get(AMDGPU::S_MOV_B32), ImmReg)
        .addImm(CI.BaseOff);

    BaseReg = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
    BaseRegFlags = RegState::Kill;

    TII->getAddNoCarry(*MBB, CI.Paired, DL, BaseReg)
        .addReg(ImmReg)
        .addReg(AddrReg->getReg(), 0, BaseSubReg)
        .addImm(0); // clamp bit
    BaseSubReg = 0;
  }

  MachineInstrBuilder Read2 =
      BuildMI(*MBB, CI.Paired, DL, Read2Desc, DestReg)
          .addReg(BaseReg, BaseRegFlags, BaseSubReg) // addr
          .addImm(NewOffset0)                        // offset0
          .addImm(NewOffset1)                        // offset1
          .addImm(0)                                 // gds
          .cloneMergedMemRefs({&*CI.I, &*CI.Paired});

  (void)Read2;

  const MCInstrDesc &CopyDesc = TII->get(TargetOpcode::COPY);

  // Copy to the old destination registers.
  BuildMI(*MBB, CI.Paired, DL, CopyDesc)
      .add(*Dest0) // Copy to same destination including flags and sub reg.
      .addReg(DestReg, 0, SubRegIdx0);
  MachineInstr *Copy1 = BuildMI(*MBB, CI.Paired, DL, CopyDesc)
                            .add(*Dest1)
                            .addReg(DestReg, RegState::Kill, SubRegIdx1);

  moveInstsAfter(Copy1, CI.InstsToMove);

  MachineBasicBlock::iterator Next = std::next(CI.I);
  CI.I->eraseFromParent();
  CI.Paired->eraseFromParent();

  LLVM_DEBUG(dbgs() << "Inserted read2: " << *Read2 << '\n');
  return Next;
}

unsigned SILoadStoreOptimizer::write2Opcode(unsigned EltSize) const {
  if (STM->ldsRequiresM0Init())
    return (EltSize == 4) ? AMDGPU::DS_WRITE2_B32 : AMDGPU::DS_WRITE2_B64;
  return (EltSize == 4) ? AMDGPU::DS_WRITE2_B32_gfx9
                        : AMDGPU::DS_WRITE2_B64_gfx9;
}

unsigned SILoadStoreOptimizer::write2ST64Opcode(unsigned EltSize) const {
  if (STM->ldsRequiresM0Init())
    return (EltSize == 4) ? AMDGPU::DS_WRITE2ST64_B32
                          : AMDGPU::DS_WRITE2ST64_B64;

  return (EltSize == 4) ? AMDGPU::DS_WRITE2ST64_B32_gfx9
                        : AMDGPU::DS_WRITE2ST64_B64_gfx9;
}

MachineBasicBlock::iterator
SILoadStoreOptimizer::mergeWrite2Pair(CombineInfo &CI) {
  MachineBasicBlock *MBB = CI.I->getParent();

  // Be sure to use .addOperand(), and not .addReg() with these. We want to be
  // sure we preserve the subregister index and any register flags set on them.
  const MachineOperand *AddrReg =
      TII->getNamedOperand(*CI.I, AMDGPU::OpName::addr);
  const MachineOperand *Data0 =
      TII->getNamedOperand(*CI.I, AMDGPU::OpName::data0);
  const MachineOperand *Data1 =
      TII->getNamedOperand(*CI.Paired, AMDGPU::OpName::data0);

  unsigned NewOffset0 = CI.Offset0;
  unsigned NewOffset1 = CI.Offset1;
  unsigned Opc =
      CI.UseST64 ? write2ST64Opcode(CI.EltSize) : write2Opcode(CI.EltSize);

  if (NewOffset0 > NewOffset1) {
    // Canonicalize the merged instruction so the smaller offset comes first.
    std::swap(NewOffset0, NewOffset1);
    std::swap(Data0, Data1);
  }

  assert((isUInt<8>(NewOffset0) && isUInt<8>(NewOffset1)) &&
         (NewOffset0 != NewOffset1) && "Computed offset doesn't fit");

  const MCInstrDesc &Write2Desc = TII->get(Opc);
  DebugLoc DL = CI.I->getDebugLoc();

  unsigned BaseReg = AddrReg->getReg();
  unsigned BaseSubReg = AddrReg->getSubReg();
  unsigned BaseRegFlags = 0;
  if (CI.BaseOff) {
    unsigned ImmReg = MRI->createVirtualRegister(&AMDGPU::SGPR_32RegClass);
    BuildMI(*MBB, CI.Paired, DL, TII->get(AMDGPU::S_MOV_B32), ImmReg)
        .addImm(CI.BaseOff);

    BaseReg = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
    BaseRegFlags = RegState::Kill;

    TII->getAddNoCarry(*MBB, CI.Paired, DL, BaseReg)
        .addReg(ImmReg)
        .addReg(AddrReg->getReg(), 0, BaseSubReg)
        .addImm(0); // clamp bit
    BaseSubReg = 0;
  }

  MachineInstrBuilder Write2 =
      BuildMI(*MBB, CI.Paired, DL, Write2Desc)
          .addReg(BaseReg, BaseRegFlags, BaseSubReg) // addr
          .add(*Data0)                               // data0
          .add(*Data1)                               // data1
          .addImm(NewOffset0)                        // offset0
          .addImm(NewOffset1)                        // offset1
          .addImm(0)                                 // gds
          .cloneMergedMemRefs({&*CI.I, &*CI.Paired});

  moveInstsAfter(Write2, CI.InstsToMove);

  MachineBasicBlock::iterator Next = std::next(CI.I);
  CI.I->eraseFromParent();
  CI.Paired->eraseFromParent();

  LLVM_DEBUG(dbgs() << "Inserted write2 inst: " << *Write2 << '\n');
  return Next;
}

MachineBasicBlock::iterator
SILoadStoreOptimizer::mergeSBufferLoadImmPair(CombineInfo &CI) {
  MachineBasicBlock *MBB = CI.I->getParent();
  DebugLoc DL = CI.I->getDebugLoc();
  const unsigned Opcode = getNewOpcode(CI);

  const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI);

  unsigned DestReg = MRI->createVirtualRegister(SuperRC);
  unsigned MergedOffset = std::min(CI.Offset0, CI.Offset1);

  BuildMI(*MBB, CI.Paired, DL, TII->get(Opcode), DestReg)
      .add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::sbase))
      .addImm(MergedOffset) // offset
      .addImm(CI.GLC0)      // glc
      .addImm(CI.DLC0)      // dlc
      .cloneMergedMemRefs({&*CI.I, &*CI.Paired});

  std::pair<unsigned, unsigned> SubRegIdx = getSubRegIdxs(CI);
  const unsigned SubRegIdx0 = std::get<0>(SubRegIdx);
  const unsigned SubRegIdx1 = std::get<1>(SubRegIdx);

  // Copy to the old destination registers.
  const MCInstrDesc &CopyDesc = TII->get(TargetOpcode::COPY);
  const auto *Dest0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::sdst);
  const auto *Dest1 = TII->getNamedOperand(*CI.Paired, AMDGPU::OpName::sdst);

  BuildMI(*MBB, CI.Paired, DL, CopyDesc)
      .add(*Dest0) // Copy to same destination including flags and sub reg.
      .addReg(DestReg, 0, SubRegIdx0);
  MachineInstr *Copy1 = BuildMI(*MBB, CI.Paired, DL, CopyDesc)
                            .add(*Dest1)
                            .addReg(DestReg, RegState::Kill, SubRegIdx1);

  moveInstsAfter(Copy1, CI.InstsToMove);

  MachineBasicBlock::iterator Next = std::next(CI.I);
  CI.I->eraseFromParent();
  CI.Paired->eraseFromParent();
  return Next;
}

MachineBasicBlock::iterator
SILoadStoreOptimizer::mergeBufferLoadPair(CombineInfo &CI) {
  MachineBasicBlock *MBB = CI.I->getParent();
  DebugLoc DL = CI.I->getDebugLoc();

  const unsigned Opcode = getNewOpcode(CI);

  const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI);

  // Copy to the new source register.
  unsigned DestReg = MRI->createVirtualRegister(SuperRC);
  unsigned MergedOffset = std::min(CI.Offset0, CI.Offset1);

  auto MIB = BuildMI(*MBB, CI.Paired, DL, TII->get(Opcode), DestReg);

  const unsigned Regs = getRegs(Opcode);

  if (Regs & VADDR)
    MIB.add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::vaddr));

  MIB.add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::srsrc))
      .add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::soffset))
      .addImm(MergedOffset) // offset
      .addImm(CI.GLC0)      // glc
      .addImm(CI.SLC0)      // slc
      .addImm(0)            // tfe
      .addImm(CI.DLC0)      // dlc
      .cloneMergedMemRefs({&*CI.I, &*CI.Paired});

  std::pair<unsigned, unsigned> SubRegIdx = getSubRegIdxs(CI);
  const unsigned SubRegIdx0 = std::get<0>(SubRegIdx);
  const unsigned SubRegIdx1 = std::get<1>(SubRegIdx);

  // Copy to the old destination registers.
  const MCInstrDesc &CopyDesc = TII->get(TargetOpcode::COPY);
  const auto *Dest0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::vdata);
  const auto *Dest1 = TII->getNamedOperand(*CI.Paired, AMDGPU::OpName::vdata);

  BuildMI(*MBB, CI.Paired, DL, CopyDesc)
      .add(*Dest0) // Copy to same destination including flags and sub reg.
      .addReg(DestReg, 0, SubRegIdx0);
  MachineInstr *Copy1 = BuildMI(*MBB, CI.Paired, DL, CopyDesc)
                            .add(*Dest1)
                            .addReg(DestReg, RegState::Kill, SubRegIdx1);

  moveInstsAfter(Copy1, CI.InstsToMove);

  MachineBasicBlock::iterator Next = std::next(CI.I);
  CI.I->eraseFromParent();
  CI.Paired->eraseFromParent();
  return Next;
}

unsigned SILoadStoreOptimizer::getNewOpcode(const CombineInfo &CI) {
  const unsigned Width = CI.Width0 + CI.Width1;

  switch (CI.InstClass) {
  default:
    return AMDGPU::getMUBUFOpcode(CI.InstClass, Width);
  case UNKNOWN:
    llvm_unreachable("Unknown instruction class");
  case S_BUFFER_LOAD_IMM:
    switch (Width) {
    default:
      return 0;
    case 2:
      return AMDGPU::S_BUFFER_LOAD_DWORDX2_IMM;
    case 4:
      return AMDGPU::S_BUFFER_LOAD_DWORDX4_IMM;
    }
  }
}

std::pair<unsigned, unsigned>
SILoadStoreOptimizer::getSubRegIdxs(const CombineInfo &CI) {
  if (CI.Offset0 > CI.Offset1) {
    switch (CI.Width0) {
    default:
      return std::make_pair(0, 0);
    case 1:
      switch (CI.Width1) {
      default:
        return std::make_pair(0, 0);
      case 1:
        return std::make_pair(AMDGPU::sub1, AMDGPU::sub0);
      case 2:
        return std::make_pair(AMDGPU::sub2, AMDGPU::sub0_sub1);
      case 3:
        return std::make_pair(AMDGPU::sub3, AMDGPU::sub0_sub1_sub2);
      }
    case 2:
      switch (CI.Width1) {
      default:
        return std::make_pair(0, 0);
      case 1:
        return std::make_pair(AMDGPU::sub1_sub2, AMDGPU::sub0);
      case 2:
        return std::make_pair(AMDGPU::sub2_sub3, AMDGPU::sub0_sub1);
      }
    case 3:
      switch (CI.Width1) {
      default:
        return std::make_pair(0, 0);
      case 1:
        return std::make_pair(AMDGPU::sub1_sub2_sub3, AMDGPU::sub0);
      }
    }
  } else {
    switch (CI.Width0) {
    default:
      return std::make_pair(0, 0);
    case 1:
      switch (CI.Width1) {
      default:
        return std::make_pair(0, 0);
      case 1:
        return std::make_pair(AMDGPU::sub0, AMDGPU::sub1);
      case 2:
        return std::make_pair(AMDGPU::sub0, AMDGPU::sub1_sub2);
      case 3:
        return std::make_pair(AMDGPU::sub0, AMDGPU::sub1_sub2_sub3);
      }
    case 2:
      switch (CI.Width1) {
      default:
        return std::make_pair(0, 0);
      case 1:
        return std::make_pair(AMDGPU::sub0_sub1, AMDGPU::sub2);
      case 2:
        return std::make_pair(AMDGPU::sub0_sub1, AMDGPU::sub2_sub3);
      }
    case 3:
      switch (CI.Width1) {
      default:
        return std::make_pair(0, 0);
      case 1:
        return std::make_pair(AMDGPU::sub0_sub1_sub2, AMDGPU::sub3);
      }
    }
  }
}

const TargetRegisterClass *
SILoadStoreOptimizer::getTargetRegisterClass(const CombineInfo &CI) {
  if (CI.InstClass == S_BUFFER_LOAD_IMM) {
    switch (CI.Width0 + CI.Width1) {
    default:
      return nullptr;
    case 2:
      return &AMDGPU::SReg_64_XEXECRegClass;
    case 4:
      return &AMDGPU::SReg_128RegClass;
    case 8:
      return &AMDGPU::SReg_256RegClass;
    case 16:
      return &AMDGPU::SReg_512RegClass;
    }
  } else {
    switch (CI.Width0 + CI.Width1) {
    default:
      return nullptr;
    case 2:
      return &AMDGPU::VReg_64RegClass;
    case 3:
      return &AMDGPU::VReg_96RegClass;
    case 4:
      return &AMDGPU::VReg_128RegClass;
    }
  }
}

MachineBasicBlock::iterator
SILoadStoreOptimizer::mergeBufferStorePair(CombineInfo &CI) {
  MachineBasicBlock *MBB = CI.I->getParent();
  DebugLoc DL = CI.I->getDebugLoc();

  const unsigned Opcode = getNewOpcode(CI);

  std::pair<unsigned, unsigned> SubRegIdx = getSubRegIdxs(CI);
  const unsigned SubRegIdx0 = std::get<0>(SubRegIdx);
  const unsigned SubRegIdx1 = std::get<1>(SubRegIdx);

  // Copy to the new source register.
  const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI);
  unsigned SrcReg = MRI->createVirtualRegister(SuperRC);

  const auto *Src0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::vdata);
  const auto *Src1 = TII->getNamedOperand(*CI.Paired, AMDGPU::OpName::vdata);

  BuildMI(*MBB, CI.Paired, DL, TII->get(AMDGPU::REG_SEQUENCE), SrcReg)
      .add(*Src0)
      .addImm(SubRegIdx0)
      .add(*Src1)
      .addImm(SubRegIdx1);

  auto MIB = BuildMI(*MBB, CI.Paired, DL, TII->get(Opcode))
                 .addReg(SrcReg, RegState::Kill);

  const unsigned Regs = getRegs(Opcode);

  if (Regs & VADDR)
    MIB.add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::vaddr));

  MIB.add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::srsrc))
      .add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::soffset))
      .addImm(std::min(CI.Offset0, CI.Offset1)) // offset
      .addImm(CI.GLC0)      // glc
      .addImm(CI.SLC0)      // slc
      .addImm(0)            // tfe
      .addImm(CI.DLC0)      // dlc
      .cloneMergedMemRefs({&*CI.I, &*CI.Paired});

  moveInstsAfter(MIB, CI.InstsToMove);

  MachineBasicBlock::iterator Next = std::next(CI.I);
  CI.I->eraseFromParent();
  CI.Paired->eraseFromParent();
  return Next;
}

MachineOperand
SILoadStoreOptimizer::createRegOrImm(int32_t Val, MachineInstr &MI) {
  APInt V(32, Val, true);
  if (TII->isInlineConstant(V))
    return MachineOperand::CreateImm(Val);

  unsigned Reg = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);
  MachineInstr *Mov =
  BuildMI(*MI.getParent(), MI.getIterator(), MI.getDebugLoc(),
          TII->get(AMDGPU::S_MOV_B32), Reg)
    .addImm(Val);
  (void)Mov;
  LLVM_DEBUG(dbgs() << "    "; Mov->dump());
  return MachineOperand::CreateReg(Reg, false);
}

// Compute base address using Addr and return the final register.
unsigned SILoadStoreOptimizer::computeBase(MachineInstr &MI,
                                           const MemAddress &Addr) {
  MachineBasicBlock *MBB = MI.getParent();
  MachineBasicBlock::iterator MBBI = MI.getIterator();
  DebugLoc DL = MI.getDebugLoc();

  assert((TRI->getRegSizeInBits(Addr.Base.LoReg, *MRI) == 32 ||
          Addr.Base.LoSubReg) &&
         "Expected 32-bit Base-Register-Low!!");

  assert((TRI->getRegSizeInBits(Addr.Base.HiReg, *MRI) == 32 ||
          Addr.Base.HiSubReg) &&
         "Expected 32-bit Base-Register-Hi!!");

  LLVM_DEBUG(dbgs() << "  Re-Computed Anchor-Base:\n");
  MachineOperand OffsetLo = createRegOrImm(static_cast<int32_t>(Addr.Offset), MI);
  MachineOperand OffsetHi =
    createRegOrImm(static_cast<int32_t>(Addr.Offset >> 32), MI);

  const auto *CarryRC = TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
  unsigned CarryReg = MRI->createVirtualRegister(CarryRC);
  unsigned DeadCarryReg = MRI->createVirtualRegister(CarryRC);

  unsigned DestSub0 = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
  unsigned DestSub1 = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
  MachineInstr *LoHalf =
    BuildMI(*MBB, MBBI, DL, TII->get(AMDGPU::V_ADD_I32_e64), DestSub0)
      .addReg(CarryReg, RegState::Define)
      .addReg(Addr.Base.LoReg, 0, Addr.Base.LoSubReg)
      .add(OffsetLo)
      .addImm(0); // clamp bit
  (void)LoHalf;
  LLVM_DEBUG(dbgs() << "    "; LoHalf->dump(););

  MachineInstr *HiHalf =
  BuildMI(*MBB, MBBI, DL, TII->get(AMDGPU::V_ADDC_U32_e64), DestSub1)
    .addReg(DeadCarryReg, RegState::Define | RegState::Dead)
    .addReg(Addr.Base.HiReg, 0, Addr.Base.HiSubReg)
    .add(OffsetHi)
    .addReg(CarryReg, RegState::Kill)
    .addImm(0); // clamp bit
  (void)HiHalf;
  LLVM_DEBUG(dbgs() << "    "; HiHalf->dump(););

  unsigned FullDestReg = MRI->createVirtualRegister(&AMDGPU::VReg_64RegClass);
  MachineInstr *FullBase =
    BuildMI(*MBB, MBBI, DL, TII->get(TargetOpcode::REG_SEQUENCE), FullDestReg)
      .addReg(DestSub0)
      .addImm(AMDGPU::sub0)
      .addReg(DestSub1)
      .addImm(AMDGPU::sub1);
  (void)FullBase;
  LLVM_DEBUG(dbgs() << "    "; FullBase->dump(); dbgs() << "\n";);

  return FullDestReg;
}

// Update base and offset with the NewBase and NewOffset in MI.
void SILoadStoreOptimizer::updateBaseAndOffset(MachineInstr &MI,
                                               unsigned NewBase,
                                               int32_t NewOffset) {
  TII->getNamedOperand(MI, AMDGPU::OpName::vaddr)->setReg(NewBase);
  TII->getNamedOperand(MI, AMDGPU::OpName::offset)->setImm(NewOffset);
}

Optional<int32_t>
SILoadStoreOptimizer::extractConstOffset(const MachineOperand &Op) {
  if (Op.isImm())
    return Op.getImm();

  if (!Op.isReg())
    return None;

  MachineInstr *Def = MRI->getUniqueVRegDef(Op.getReg());
  if (!Def || Def->getOpcode() != AMDGPU::S_MOV_B32 ||
      !Def->getOperand(1).isImm())
    return None;

  return Def->getOperand(1).getImm();
}

// Analyze Base and extracts:
//  - 32bit base registers, subregisters
//  - 64bit constant offset
// Expecting base computation as:
//   %OFFSET0:sgpr_32 = S_MOV_B32 8000
//   %LO:vgpr_32, %c:sreg_64_xexec =
//       V_ADD_I32_e64 %BASE_LO:vgpr_32, %103:sgpr_32,
//   %HI:vgpr_32, = V_ADDC_U32_e64 %BASE_HI:vgpr_32, 0, killed %c:sreg_64_xexec
//   %Base:vreg_64 =
//       REG_SEQUENCE %LO:vgpr_32, %subreg.sub0, %HI:vgpr_32, %subreg.sub1
void SILoadStoreOptimizer::processBaseWithConstOffset(const MachineOperand &Base,
                                                      MemAddress &Addr) {
  if (!Base.isReg())
    return;

  MachineInstr *Def = MRI->getUniqueVRegDef(Base.getReg());
  if (!Def || Def->getOpcode() != AMDGPU::REG_SEQUENCE
      || Def->getNumOperands() != 5)
    return;

  MachineOperand BaseLo = Def->getOperand(1);
  MachineOperand BaseHi = Def->getOperand(3);
  if (!BaseLo.isReg() || !BaseHi.isReg())
    return;

  MachineInstr *BaseLoDef = MRI->getUniqueVRegDef(BaseLo.getReg());
  MachineInstr *BaseHiDef = MRI->getUniqueVRegDef(BaseHi.getReg());

  if (!BaseLoDef || BaseLoDef->getOpcode() != AMDGPU::V_ADD_I32_e64 ||
      !BaseHiDef || BaseHiDef->getOpcode() != AMDGPU::V_ADDC_U32_e64)
    return;

  const auto *Src0 = TII->getNamedOperand(*BaseLoDef, AMDGPU::OpName::src0);
  const auto *Src1 = TII->getNamedOperand(*BaseLoDef, AMDGPU::OpName::src1);

  auto Offset0P = extractConstOffset(*Src0);
  if (Offset0P)
    BaseLo = *Src1;
  else {
    if (!(Offset0P = extractConstOffset(*Src1)))
      return;
    BaseLo = *Src0;
  }

  Src0 = TII->getNamedOperand(*BaseHiDef, AMDGPU::OpName::src0);
  Src1 = TII->getNamedOperand(*BaseHiDef, AMDGPU::OpName::src1);

  if (Src0->isImm())
    std::swap(Src0, Src1);

  if (!Src1->isImm())
    return;

  uint64_t Offset1 = Src1->getImm();
  BaseHi = *Src0;

  Addr.Base.LoReg = BaseLo.getReg();
  Addr.Base.HiReg = BaseHi.getReg();
  Addr.Base.LoSubReg = BaseLo.getSubReg();
  Addr.Base.HiSubReg = BaseHi.getSubReg();
  Addr.Offset = (*Offset0P & 0x00000000ffffffff) | (Offset1 << 32);
}

bool SILoadStoreOptimizer::promoteConstantOffsetToImm(
    MachineInstr &MI,
    MemInfoMap &Visited,
    SmallPtrSet<MachineInstr *, 4> &AnchorList) {

  // TODO: Support flat and scratch.
  if (AMDGPU::getGlobalSaddrOp(MI.getOpcode()) < 0 ||
      TII->getNamedOperand(MI, AMDGPU::OpName::vdata) != NULL)
    return false;

  // TODO: Support Store.
  if (!MI.mayLoad())
    return false;

  if (AnchorList.count(&MI))
    return false;

  LLVM_DEBUG(dbgs() << "\nTryToPromoteConstantOffsetToImmFor "; MI.dump());

  if (TII->getNamedOperand(MI, AMDGPU::OpName::offset)->getImm()) {
    LLVM_DEBUG(dbgs() << "  Const-offset is already promoted.\n";);
    return false;
  }

  // Step1: Find the base-registers and a 64bit constant offset.
  MachineOperand &Base = *TII->getNamedOperand(MI, AMDGPU::OpName::vaddr);
  MemAddress MAddr;
  if (Visited.find(&MI) == Visited.end()) {
    processBaseWithConstOffset(Base, MAddr);
    Visited[&MI] = MAddr;
  } else
    MAddr = Visited[&MI];

  if (MAddr.Offset == 0) {
    LLVM_DEBUG(dbgs() << "  Failed to extract constant-offset or there are no"
                         " constant offsets that can be promoted.\n";);
    return false;
  }

  LLVM_DEBUG(dbgs() << "  BASE: {" << MAddr.Base.HiReg << ", "
             << MAddr.Base.LoReg << "} Offset: " << MAddr.Offset << "\n\n";);

  // Step2: Traverse through MI's basic block and find an anchor(that has the
  // same base-registers) with the highest 13bit distance from MI's offset.
  // E.g. (64bit loads)
  // bb:
  //   addr1 = &a + 4096;   load1 = load(addr1,  0)
  //   addr2 = &a + 6144;   load2 = load(addr2,  0)
  //   addr3 = &a + 8192;   load3 = load(addr3,  0)
  //   addr4 = &a + 10240;  load4 = load(addr4,  0)
  //   addr5 = &a + 12288;  load5 = load(addr5,  0)
  //
  // Starting from the first load, the optimization will try to find a new base
  // from which (&a + 4096) has 13 bit distance. Both &a + 6144 and &a + 8192
  // has 13bit distance from &a + 4096. The heuristic considers &a + 8192
  // as the new-base(anchor) because of the maximum distance which can
  // accomodate more intermediate bases presumeably.
  //
  // Step3: move (&a + 8192) above load1. Compute and promote offsets from
  // (&a + 8192) for load1, load2, load4.
  //   addr = &a + 8192
  //   load1 = load(addr,       -4096)
  //   load2 = load(addr,       -2048)
  //   load3 = load(addr,       0)
  //   load4 = load(addr,       2048)
  //   addr5 = &a + 12288;  load5 = load(addr5,  0)
  //
  MachineInstr *AnchorInst = nullptr;
  MemAddress AnchorAddr;
  uint32_t MaxDist = std::numeric_limits<uint32_t>::min();
  SmallVector<std::pair<MachineInstr *, int64_t>, 4> InstsWCommonBase;

  MachineBasicBlock *MBB = MI.getParent();
  MachineBasicBlock::iterator E = MBB->end();
  MachineBasicBlock::iterator MBBI = MI.getIterator();
  ++MBBI;
  const SITargetLowering *TLI =
    static_cast<const SITargetLowering *>(STM->getTargetLowering());

  for ( ; MBBI != E; ++MBBI) {
    MachineInstr &MINext = *MBBI;
    // TODO: Support finding an anchor(with same base) from store addresses or
    // any other load addresses where the opcodes are different.
    if (MINext.getOpcode() != MI.getOpcode() ||
        TII->getNamedOperand(MINext, AMDGPU::OpName::offset)->getImm())
      continue;

    const MachineOperand &BaseNext =
      *TII->getNamedOperand(MINext, AMDGPU::OpName::vaddr);
    MemAddress MAddrNext;
    if (Visited.find(&MINext) == Visited.end()) {
      processBaseWithConstOffset(BaseNext, MAddrNext);
      Visited[&MINext] = MAddrNext;
    } else
      MAddrNext = Visited[&MINext];

    if (MAddrNext.Base.LoReg != MAddr.Base.LoReg ||
        MAddrNext.Base.HiReg != MAddr.Base.HiReg ||
        MAddrNext.Base.LoSubReg != MAddr.Base.LoSubReg ||
        MAddrNext.Base.HiSubReg != MAddr.Base.HiSubReg)
      continue;

    InstsWCommonBase.push_back(std::make_pair(&MINext, MAddrNext.Offset));

    int64_t Dist = MAddr.Offset - MAddrNext.Offset;
    TargetLoweringBase::AddrMode AM;
    AM.HasBaseReg = true;
    AM.BaseOffs = Dist;
    if (TLI->isLegalGlobalAddressingMode(AM) &&
        (uint32_t)std::abs(Dist) > MaxDist) {
      MaxDist = std::abs(Dist);

      AnchorAddr = MAddrNext;
      AnchorInst = &MINext;
    }
  }

  if (AnchorInst) {
    LLVM_DEBUG(dbgs() << "  Anchor-Inst(with max-distance from Offset): ";
               AnchorInst->dump());
    LLVM_DEBUG(dbgs() << "  Anchor-Offset from BASE: "
               <<  AnchorAddr.Offset << "\n\n");

    // Instead of moving up, just re-compute anchor-instruction's base address.
    unsigned Base = computeBase(MI, AnchorAddr);

    updateBaseAndOffset(MI, Base, MAddr.Offset - AnchorAddr.Offset);
    LLVM_DEBUG(dbgs() << "  After promotion: "; MI.dump(););

    for (auto P : InstsWCommonBase) {
      TargetLoweringBase::AddrMode AM;
      AM.HasBaseReg = true;
      AM.BaseOffs = P.second - AnchorAddr.Offset;

      if (TLI->isLegalGlobalAddressingMode(AM)) {
        LLVM_DEBUG(dbgs() << "  Promote Offset(" << P.second;
                   dbgs() << ")"; P.first->dump());
        updateBaseAndOffset(*P.first, Base, P.second - AnchorAddr.Offset);
        LLVM_DEBUG(dbgs() << "     After promotion: "; P.first->dump());
      }
    }
    AnchorList.insert(AnchorInst);
    return true;
  }

  return false;
}

// Scan through looking for adjacent LDS operations with constant offsets from
// the same base register. We rely on the scheduler to do the hard work of
// clustering nearby loads, and assume these are all adjacent.
bool SILoadStoreOptimizer::optimizeBlock(MachineBasicBlock &MBB) {
  bool Modified = false;

  // Contain the list
  MemInfoMap Visited;
  // Contains the list of instructions for which constant offsets are being
  // promoted to the IMM.
  SmallPtrSet<MachineInstr *, 4> AnchorList;

  for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E;) {
    MachineInstr &MI = *I;

    if (promoteConstantOffsetToImm(MI, Visited, AnchorList))
      Modified = true;

    // Don't combine if volatile.
    if (MI.hasOrderedMemoryRef()) {
      ++I;
      continue;
    }

    const unsigned Opc = MI.getOpcode();

    CombineInfo CI;
    CI.I = I;
    CI.InstClass = getInstClass(Opc);

    switch (CI.InstClass) {
    default:
      break;
    case DS_READ:
      CI.EltSize =
          (Opc == AMDGPU::DS_READ_B64 || Opc == AMDGPU::DS_READ_B64_gfx9) ? 8
                                                                          : 4;
      if (findMatchingInst(CI)) {
        Modified = true;
        I = mergeRead2Pair(CI);
      } else {
        ++I;
      }
      continue;
    case DS_WRITE:
      CI.EltSize =
          (Opc == AMDGPU::DS_WRITE_B64 || Opc == AMDGPU::DS_WRITE_B64_gfx9) ? 8
                                                                            : 4;
      if (findMatchingInst(CI)) {
        Modified = true;
        I = mergeWrite2Pair(CI);
      } else {
        ++I;
      }
      continue;
    case S_BUFFER_LOAD_IMM:
      CI.EltSize = AMDGPU::getSMRDEncodedOffset(*STM, 4);
      if (findMatchingInst(CI)) {
        Modified = true;
        I = mergeSBufferLoadImmPair(CI);
        OptimizeAgain |= (CI.Width0 + CI.Width1) < 16;
      } else {
        ++I;
      }
      continue;
    case BUFFER_LOAD_OFFEN:
    case BUFFER_LOAD_OFFSET:
    case BUFFER_LOAD_OFFEN_exact:
    case BUFFER_LOAD_OFFSET_exact:
      CI.EltSize = 4;
      if (findMatchingInst(CI)) {
        Modified = true;
        I = mergeBufferLoadPair(CI);
        OptimizeAgain |= (CI.Width0 + CI.Width1) < 4;
      } else {
        ++I;
      }
      continue;
    case BUFFER_STORE_OFFEN:
    case BUFFER_STORE_OFFSET:
    case BUFFER_STORE_OFFEN_exact:
    case BUFFER_STORE_OFFSET_exact:
      CI.EltSize = 4;
      if (findMatchingInst(CI)) {
        Modified = true;
        I = mergeBufferStorePair(CI);
        OptimizeAgain |= (CI.Width0 + CI.Width1) < 4;
      } else {
        ++I;
      }
      continue;
    }

    ++I;
  }

  return Modified;
}

bool SILoadStoreOptimizer::runOnMachineFunction(MachineFunction &MF) {
  if (skipFunction(MF.getFunction()))
    return false;

  STM = &MF.getSubtarget<GCNSubtarget>();
  if (!STM->loadStoreOptEnabled())
    return false;

  TII = STM->getInstrInfo();
  TRI = &TII->getRegisterInfo();

  MRI = &MF.getRegInfo();
  AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();

  assert(MRI->isSSA() && "Must be run on SSA");

  LLVM_DEBUG(dbgs() << "Running SILoadStoreOptimizer\n");

  bool Modified = false;

  for (MachineBasicBlock &MBB : MF) {
    do {
      OptimizeAgain = false;
      Modified |= optimizeBlock(MBB);
    } while (OptimizeAgain);
  }

  return Modified;
}