llvm.org GIT mirror llvm / release_26 lib / Target / ARM / ARMLoadStoreOptimizer.cpp
release_26

Tree @release_26 (Download .tar.gz)

ARMLoadStoreOptimizer.cpp @release_26raw · 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
//===-- ARMLoadStoreOptimizer.cpp - ARM load / store opt. pass ----*- C++ -*-=//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains a pass that performs load / store related peephole
// optimizations. This pass should be run after register allocation.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "arm-ldst-opt"
#include "ARM.h"
#include "ARMAddressingModes.h"
#include "ARMBaseInstrInfo.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMRegisterInfo.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;

STATISTIC(NumLDMGened , "Number of ldm instructions generated");
STATISTIC(NumSTMGened , "Number of stm instructions generated");
STATISTIC(NumFLDMGened, "Number of fldm instructions generated");
STATISTIC(NumFSTMGened, "Number of fstm instructions generated");
STATISTIC(NumLdStMoved, "Number of load / store instructions moved");
STATISTIC(NumLDRDFormed,"Number of ldrd created before allocation");
STATISTIC(NumSTRDFormed,"Number of strd created before allocation");
STATISTIC(NumLDRD2LDM,  "Number of ldrd instructions turned back into ldm");
STATISTIC(NumSTRD2STM,  "Number of strd instructions turned back into stm");
STATISTIC(NumLDRD2LDR,  "Number of ldrd instructions turned back into ldr's");
STATISTIC(NumSTRD2STR,  "Number of strd instructions turned back into str's");

/// ARMAllocLoadStoreOpt - Post- register allocation pass the combine
/// load / store instructions to form ldm / stm instructions.

namespace {
  struct VISIBILITY_HIDDEN ARMLoadStoreOpt : public MachineFunctionPass {
    static char ID;
    ARMLoadStoreOpt() : MachineFunctionPass(&ID) {}

    const TargetInstrInfo *TII;
    const TargetRegisterInfo *TRI;
    ARMFunctionInfo *AFI;
    RegScavenger *RS;
    bool isThumb2;

    virtual bool runOnMachineFunction(MachineFunction &Fn);

    virtual const char *getPassName() const {
      return "ARM load / store optimization pass";
    }

  private:
    struct MemOpQueueEntry {
      int Offset;
      unsigned Position;
      MachineBasicBlock::iterator MBBI;
      bool Merged;
      MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
        : Offset(o), Position(p), MBBI(i), Merged(false) {};
    };
    typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
    typedef MemOpQueue::iterator MemOpQueueIter;

    bool MergeOps(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
                  int Offset, unsigned Base, bool BaseKill, int Opcode,
                  ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
                  DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
    void MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
                      int Opcode, unsigned Size,
                      ARMCC::CondCodes Pred, unsigned PredReg,
                      unsigned Scratch, MemOpQueue &MemOps,
                      SmallVector<MachineBasicBlock::iterator, 4> &Merges);

    void AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps);
    bool FixInvalidRegPairOp(MachineBasicBlock &MBB,
                             MachineBasicBlock::iterator &MBBI);
    bool MergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
                                  MachineBasicBlock::iterator MBBI,
                                  const TargetInstrInfo *TII,
                                  bool &Advance,
                                  MachineBasicBlock::iterator &I);
    bool MergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
                                   MachineBasicBlock::iterator MBBI,
                                   bool &Advance,
                                   MachineBasicBlock::iterator &I);
    bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
    bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
  };
  char ARMLoadStoreOpt::ID = 0;
}

static int getLoadStoreMultipleOpcode(int Opcode) {
  switch (Opcode) {
  case ARM::LDR:
    NumLDMGened++;
    return ARM::LDM;
  case ARM::STR:
    NumSTMGened++;
    return ARM::STM;
  case ARM::t2LDRi8:
  case ARM::t2LDRi12:
    NumLDMGened++;
    return ARM::t2LDM;
  case ARM::t2STRi8:
  case ARM::t2STRi12:
    NumSTMGened++;
    return ARM::t2STM;
  case ARM::FLDS:
    NumFLDMGened++;
    return ARM::FLDMS;
  case ARM::FSTS:
    NumFSTMGened++;
    return ARM::FSTMS;
  case ARM::FLDD:
    NumFLDMGened++;
    return ARM::FLDMD;
  case ARM::FSTD:
    NumFSTMGened++;
    return ARM::FSTMD;
  default: llvm_unreachable("Unhandled opcode!");
  }
  return 0;
}

static bool isT2i32Load(unsigned Opc) {
  return Opc == ARM::t2LDRi12 || Opc == ARM::t2LDRi8;
}

static bool isi32Load(unsigned Opc) {
  return Opc == ARM::LDR || isT2i32Load(Opc);
}

static bool isT2i32Store(unsigned Opc) {
  return Opc == ARM::t2STRi12 || Opc == ARM::t2STRi8;
}

static bool isi32Store(unsigned Opc) {
  return Opc == ARM::STR || isT2i32Store(Opc);
}

/// MergeOps - Create and insert a LDM or STM with Base as base register and
/// registers in Regs as the register operands that would be loaded / stored.
/// It returns true if the transformation is done.
bool
ARMLoadStoreOpt::MergeOps(MachineBasicBlock &MBB,
                          MachineBasicBlock::iterator MBBI,
                          int Offset, unsigned Base, bool BaseKill,
                          int Opcode, ARMCC::CondCodes Pred,
                          unsigned PredReg, unsigned Scratch, DebugLoc dl,
                          SmallVector<std::pair<unsigned, bool>, 8> &Regs) {
  // Only a single register to load / store. Don't bother.
  unsigned NumRegs = Regs.size();
  if (NumRegs <= 1)
    return false;

  ARM_AM::AMSubMode Mode = ARM_AM::ia;
  bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
  if (isAM4 && Offset == 4) {
    if (isThumb2)
      // Thumb2 does not support ldmib / stmib.
      return false;
    Mode = ARM_AM::ib;
  } else if (isAM4 && Offset == -4 * (int)NumRegs + 4) {
    if (isThumb2)
      // Thumb2 does not support ldmda / stmda.
      return false;
    Mode = ARM_AM::da;
  } else if (isAM4 && Offset == -4 * (int)NumRegs) {
    Mode = ARM_AM::db;
  } else if (Offset != 0) {
    // If starting offset isn't zero, insert a MI to materialize a new base.
    // But only do so if it is cost effective, i.e. merging more than two
    // loads / stores.
    if (NumRegs <= 2)
      return false;

    unsigned NewBase;
    if (isi32Load(Opcode))
      // If it is a load, then just use one of the destination register to
      // use as the new base.
      NewBase = Regs[NumRegs-1].first;
    else {
      // Use the scratch register to use as a new base.
      NewBase = Scratch;
      if (NewBase == 0)
        return false;
    }
    int BaseOpc = !isThumb2
      ? ARM::ADDri
      : ((Base == ARM::SP) ? ARM::t2ADDrSPi : ARM::t2ADDri);
    if (Offset < 0) {
      BaseOpc = !isThumb2
        ? ARM::SUBri
        : ((Base == ARM::SP) ? ARM::t2SUBrSPi : ARM::t2SUBri);
      Offset = - Offset;
    }
    int ImmedOffset = isThumb2
      ? ARM_AM::getT2SOImmVal(Offset) : ARM_AM::getSOImmVal(Offset);
    if (ImmedOffset == -1)
      // FIXME: Try t2ADDri12 or t2SUBri12?
      return false;  // Probably not worth it then.

    BuildMI(MBB, MBBI, dl, TII->get(BaseOpc), NewBase)
      .addReg(Base, getKillRegState(BaseKill)).addImm(Offset)
      .addImm(Pred).addReg(PredReg).addReg(0);
    Base = NewBase;
    BaseKill = true;  // New base is always killed right its use.
  }

  bool isDPR = Opcode == ARM::FLDD || Opcode == ARM::FSTD;
  bool isDef = isi32Load(Opcode) || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
  Opcode = getLoadStoreMultipleOpcode(Opcode);
  MachineInstrBuilder MIB = (isAM4)
    ? BuildMI(MBB, MBBI, dl, TII->get(Opcode))
        .addReg(Base, getKillRegState(BaseKill))
        .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred).addReg(PredReg)
    : BuildMI(MBB, MBBI, dl, TII->get(Opcode))
        .addReg(Base, getKillRegState(BaseKill))
        .addImm(ARM_AM::getAM5Opc(Mode, false, isDPR ? NumRegs<<1 : NumRegs))
        .addImm(Pred).addReg(PredReg);
  for (unsigned i = 0; i != NumRegs; ++i)
    MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
                     | getKillRegState(Regs[i].second));

  return true;
}

/// MergeLDR_STR - Merge a number of load / store instructions into one or more
/// load / store multiple instructions.
void
ARMLoadStoreOpt::MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex,
                          unsigned Base, int Opcode, unsigned Size,
                          ARMCC::CondCodes Pred, unsigned PredReg,
                          unsigned Scratch, MemOpQueue &MemOps,
                          SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
  bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
  int Offset = MemOps[SIndex].Offset;
  int SOffset = Offset;
  unsigned Pos = MemOps[SIndex].Position;
  MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
  DebugLoc dl = Loc->getDebugLoc();
  unsigned PReg = Loc->getOperand(0).getReg();
  unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
  bool isKill = Loc->getOperand(0).isKill();

  SmallVector<std::pair<unsigned,bool>, 8> Regs;
  Regs.push_back(std::make_pair(PReg, isKill));
  for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
    int NewOffset = MemOps[i].Offset;
    unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
    unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
    isKill = MemOps[i].MBBI->getOperand(0).isKill();
    // AM4 - register numbers in ascending order.
    // AM5 - consecutive register numbers in ascending order.
    if (NewOffset == Offset + (int)Size &&
        ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
      Offset += Size;
      Regs.push_back(std::make_pair(Reg, isKill));
      PRegNum = RegNum;
    } else {
      // Can't merge this in. Try merge the earlier ones first.
      if (MergeOps(MBB, ++Loc, SOffset, Base, false, Opcode, Pred, PredReg,
                   Scratch, dl, Regs)) {
        Merges.push_back(prior(Loc));
        for (unsigned j = SIndex; j < i; ++j) {
          MBB.erase(MemOps[j].MBBI);
          MemOps[j].Merged = true;
        }
      }
      MergeLDR_STR(MBB, i, Base, Opcode, Size, Pred, PredReg, Scratch,
                   MemOps, Merges);
      return;
    }

    if (MemOps[i].Position > Pos) {
      Pos = MemOps[i].Position;
      Loc = MemOps[i].MBBI;
    }
  }

  bool BaseKill = Loc->findRegisterUseOperandIdx(Base, true) != -1;
  if (MergeOps(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Pred, PredReg,
               Scratch, dl, Regs)) {
    Merges.push_back(prior(Loc));
    for (unsigned i = SIndex, e = MemOps.size(); i != e; ++i) {
      MBB.erase(MemOps[i].MBBI);
      MemOps[i].Merged = true;
    }
  }

  return;
}

static inline bool isMatchingDecrement(MachineInstr *MI, unsigned Base,
                                       unsigned Bytes, unsigned Limit,
                                       ARMCC::CondCodes Pred, unsigned PredReg){
  unsigned MyPredReg = 0;
  if (!MI)
    return false;
  if (MI->getOpcode() != ARM::t2SUBri &&
      MI->getOpcode() != ARM::t2SUBrSPi &&
      MI->getOpcode() != ARM::t2SUBrSPi12 &&
      MI->getOpcode() != ARM::tSUBspi &&
      MI->getOpcode() != ARM::SUBri)
    return false;

  // Make sure the offset fits in 8 bits.
  if (Bytes <= 0 || (Limit && Bytes >= Limit))
    return false;

  unsigned Scale = (MI->getOpcode() == ARM::tSUBspi) ? 4 : 1; // FIXME
  return (MI->getOperand(0).getReg() == Base &&
          MI->getOperand(1).getReg() == Base &&
          (MI->getOperand(2).getImm()*Scale) == Bytes &&
          llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
          MyPredReg == PredReg);
}

static inline bool isMatchingIncrement(MachineInstr *MI, unsigned Base,
                                       unsigned Bytes, unsigned Limit,
                                       ARMCC::CondCodes Pred, unsigned PredReg){
  unsigned MyPredReg = 0;
  if (!MI)
    return false;
  if (MI->getOpcode() != ARM::t2ADDri &&
      MI->getOpcode() != ARM::t2ADDrSPi &&
      MI->getOpcode() != ARM::t2ADDrSPi12 &&
      MI->getOpcode() != ARM::tADDspi &&
      MI->getOpcode() != ARM::ADDri)
    return false;

  if (Bytes <= 0 || (Limit && Bytes >= Limit))
    // Make sure the offset fits in 8 bits.
    return false;

  unsigned Scale = (MI->getOpcode() == ARM::tADDspi) ? 4 : 1; // FIXME
  return (MI->getOperand(0).getReg() == Base &&
          MI->getOperand(1).getReg() == Base &&
          (MI->getOperand(2).getImm()*Scale) == Bytes &&
          llvm::getInstrPredicate(MI, MyPredReg) == Pred &&
          MyPredReg == PredReg);
}

static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
  switch (MI->getOpcode()) {
  default: return 0;
  case ARM::LDR:
  case ARM::STR:
  case ARM::t2LDRi8:
  case ARM::t2LDRi12:
  case ARM::t2STRi8:
  case ARM::t2STRi12:
  case ARM::FLDS:
  case ARM::FSTS:
    return 4;
  case ARM::FLDD:
  case ARM::FSTD:
    return 8;
  case ARM::LDM:
  case ARM::STM:
  case ARM::t2LDM:
  case ARM::t2STM:
    return (MI->getNumOperands() - 4) * 4;
  case ARM::FLDMS:
  case ARM::FSTMS:
  case ARM::FLDMD:
  case ARM::FSTMD:
    return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
  }
}

/// MergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
/// register into the LDM/STM/FLDM{D|S}/FSTM{D|S} op when possible:
///
/// stmia rn, <ra, rb, rc>
/// rn := rn + 4 * 3;
/// =>
/// stmia rn!, <ra, rb, rc>
///
/// rn := rn - 4 * 3;
/// ldmia rn, <ra, rb, rc>
/// =>
/// ldmdb rn!, <ra, rb, rc>
bool ARMLoadStoreOpt::MergeBaseUpdateLSMultiple(MachineBasicBlock &MBB,
                                               MachineBasicBlock::iterator MBBI,
                                               bool &Advance,
                                               MachineBasicBlock::iterator &I) {
  MachineInstr *MI = MBBI;
  unsigned Base = MI->getOperand(0).getReg();
  unsigned Bytes = getLSMultipleTransferSize(MI);
  unsigned PredReg = 0;
  ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
  int Opcode = MI->getOpcode();
  bool isAM4 = Opcode == ARM::LDM || Opcode == ARM::t2LDM ||
    Opcode == ARM::STM || Opcode == ARM::t2STM;

  if (isAM4) {
    if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
      return false;

    // Can't use the updating AM4 sub-mode if the base register is also a dest
    // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
    for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
      if (MI->getOperand(i).getReg() == Base)
        return false;
    }

    ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
    if (MBBI != MBB.begin()) {
      MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
      if (Mode == ARM_AM::ia &&
          isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
        MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
        MBB.erase(PrevMBBI);
        return true;
      } else if (Mode == ARM_AM::ib &&
                 isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
        MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
        MBB.erase(PrevMBBI);
        return true;
      }
    }

    if (MBBI != MBB.end()) {
      MachineBasicBlock::iterator NextMBBI = next(MBBI);
      if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
          isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
        MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
        if (NextMBBI == I) {
          Advance = true;
          ++I;
        }
        MBB.erase(NextMBBI);
        return true;
      } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
                 isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
        MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
        if (NextMBBI == I) {
          Advance = true;
          ++I;
        }
        MBB.erase(NextMBBI);
        return true;
      }
    }
  } else {
    // FLDM{D|S}, FSTM{D|S} addressing mode 5 ops.
    if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
      return false;

    ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
    unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
    if (MBBI != MBB.begin()) {
      MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
      if (Mode == ARM_AM::ia &&
          isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
        MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
        MBB.erase(PrevMBBI);
        return true;
      }
    }

    if (MBBI != MBB.end()) {
      MachineBasicBlock::iterator NextMBBI = next(MBBI);
      if (Mode == ARM_AM::ia &&
          isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
        MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
        if (NextMBBI == I) {
          Advance = true;
          ++I;
        }
        MBB.erase(NextMBBI);
      }
      return true;
    }
  }

  return false;
}

static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
  switch (Opc) {
  case ARM::LDR: return ARM::LDR_PRE;
  case ARM::STR: return ARM::STR_PRE;
  case ARM::FLDS: return ARM::FLDMS;
  case ARM::FLDD: return ARM::FLDMD;
  case ARM::FSTS: return ARM::FSTMS;
  case ARM::FSTD: return ARM::FSTMD;
  case ARM::t2LDRi8:
  case ARM::t2LDRi12:
    return ARM::t2LDR_PRE;
  case ARM::t2STRi8:
  case ARM::t2STRi12:
    return ARM::t2STR_PRE;
  default: llvm_unreachable("Unhandled opcode!");
  }
  return 0;
}

static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
  switch (Opc) {
  case ARM::LDR: return ARM::LDR_POST;
  case ARM::STR: return ARM::STR_POST;
  case ARM::FLDS: return ARM::FLDMS;
  case ARM::FLDD: return ARM::FLDMD;
  case ARM::FSTS: return ARM::FSTMS;
  case ARM::FSTD: return ARM::FSTMD;
  case ARM::t2LDRi8:
  case ARM::t2LDRi12:
    return ARM::t2LDR_POST;
  case ARM::t2STRi8:
  case ARM::t2STRi12:
    return ARM::t2STR_POST;
  default: llvm_unreachable("Unhandled opcode!");
  }
  return 0;
}

/// MergeBaseUpdateLoadStore - Fold proceeding/trailing inc/dec of base
/// register into the LDR/STR/FLD{D|S}/FST{D|S} op when possible:
bool ARMLoadStoreOpt::MergeBaseUpdateLoadStore(MachineBasicBlock &MBB,
                                               MachineBasicBlock::iterator MBBI,
                                               const TargetInstrInfo *TII,
                                               bool &Advance,
                                               MachineBasicBlock::iterator &I) {
  MachineInstr *MI = MBBI;
  unsigned Base = MI->getOperand(1).getReg();
  bool BaseKill = MI->getOperand(1).isKill();
  unsigned Bytes = getLSMultipleTransferSize(MI);
  int Opcode = MI->getOpcode();
  DebugLoc dl = MI->getDebugLoc();
  bool isAM5 = Opcode == ARM::FLDD || Opcode == ARM::FLDS ||
    Opcode == ARM::FSTD || Opcode == ARM::FSTS;
  bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
  if (isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0)
    return false;
  else if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
    return false;
  else if (isT2i32Load(Opcode) || isT2i32Store(Opcode))
    if (MI->getOperand(2).getImm() != 0)
      return false;

  bool isLd = isi32Load(Opcode) || Opcode == ARM::FLDS || Opcode == ARM::FLDD;
  // Can't do the merge if the destination register is the same as the would-be
  // writeback register.
  if (isLd && MI->getOperand(0).getReg() == Base)
    return false;

  unsigned PredReg = 0;
  ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
  bool DoMerge = false;
  ARM_AM::AddrOpc AddSub = ARM_AM::add;
  unsigned NewOpc = 0;
  // AM2 - 12 bits, thumb2 - 8 bits.
  unsigned Limit = isAM5 ? 0 : (isAM2 ? 0x1000 : 0x100);
  if (MBBI != MBB.begin()) {
    MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
    if (isMatchingDecrement(PrevMBBI, Base, Bytes, Limit, Pred, PredReg)) {
      DoMerge = true;
      AddSub = ARM_AM::sub;
      NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
    } else if (!isAM5 &&
               isMatchingIncrement(PrevMBBI, Base, Bytes, Limit,Pred,PredReg)) {
      DoMerge = true;
      NewOpc = getPreIndexedLoadStoreOpcode(Opcode);
    }
    if (DoMerge)
      MBB.erase(PrevMBBI);
  }

  if (!DoMerge && MBBI != MBB.end()) {
    MachineBasicBlock::iterator NextMBBI = next(MBBI);
    if (!isAM5 &&
        isMatchingDecrement(NextMBBI, Base, Bytes, Limit, Pred, PredReg)) {
      DoMerge = true;
      AddSub = ARM_AM::sub;
      NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
    } else if (isMatchingIncrement(NextMBBI, Base, Bytes, Limit,Pred,PredReg)) {
      DoMerge = true;
      NewOpc = getPostIndexedLoadStoreOpcode(Opcode);
    }
    if (DoMerge) {
      if (NextMBBI == I) {
        Advance = true;
        ++I;
      }
      MBB.erase(NextMBBI);
    }
  }

  if (!DoMerge)
    return false;

  bool isDPR = NewOpc == ARM::FLDMD || NewOpc == ARM::FSTMD;
  unsigned Offset = 0;
  if (isAM5)
    Offset = ARM_AM::getAM5Opc((AddSub == ARM_AM::sub)
                               ? ARM_AM::db
                               : ARM_AM::ia, true, (isDPR ? 2 : 1));
  else if (isAM2)
    Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
  else
    Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
  if (isLd) {
    if (isAM5)
      // FLDMS, FLDMD
      BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
        .addReg(Base, getKillRegState(BaseKill))
        .addImm(Offset).addImm(Pred).addReg(PredReg)
        .addReg(MI->getOperand(0).getReg(), RegState::Define);
    else if (isAM2)
      // LDR_PRE, LDR_POST,
      BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
        .addReg(Base, RegState::Define)
        .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
    else
      // t2LDR_PRE, t2LDR_POST
      BuildMI(MBB, MBBI, dl, TII->get(NewOpc), MI->getOperand(0).getReg())
        .addReg(Base, RegState::Define)
        .addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
  } else {
    MachineOperand &MO = MI->getOperand(0);
    if (isAM5)
      // FSTMS, FSTMD
      BuildMI(MBB, MBBI, dl, TII->get(NewOpc)).addReg(Base).addImm(Offset)
        .addImm(Pred).addReg(PredReg)
        .addReg(MO.getReg(), getKillRegState(MO.isKill()));
    else if (isAM2)
      // STR_PRE, STR_POST
      BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
        .addReg(MO.getReg(), getKillRegState(MO.isKill()))
        .addReg(Base).addReg(0).addImm(Offset).addImm(Pred).addReg(PredReg);
    else
      // t2STR_PRE, t2STR_POST
      BuildMI(MBB, MBBI, dl, TII->get(NewOpc), Base)
        .addReg(MO.getReg(), getKillRegState(MO.isKill()))
        .addReg(Base).addImm(Offset).addImm(Pred).addReg(PredReg);
  }
  MBB.erase(MBBI);

  return true;
}

/// isMemoryOp - Returns true if instruction is a memory operations (that this
/// pass is capable of operating on).
static bool isMemoryOp(const MachineInstr *MI) {
  int Opcode = MI->getOpcode();
  switch (Opcode) {
  default: break;
  case ARM::LDR:
  case ARM::STR:
    return MI->getOperand(1).isReg() && MI->getOperand(2).getReg() == 0;
  case ARM::FLDS:
  case ARM::FSTS:
    return MI->getOperand(1).isReg();
  case ARM::FLDD:
  case ARM::FSTD:
    return MI->getOperand(1).isReg();
  case ARM::t2LDRi8:
  case ARM::t2LDRi12:
  case ARM::t2STRi8:
  case ARM::t2STRi12:
    return true;
  }
  return false;
}

/// AdvanceRS - Advance register scavenger to just before the earliest memory
/// op that is being merged.
void ARMLoadStoreOpt::AdvanceRS(MachineBasicBlock &MBB, MemOpQueue &MemOps) {
  MachineBasicBlock::iterator Loc = MemOps[0].MBBI;
  unsigned Position = MemOps[0].Position;
  for (unsigned i = 1, e = MemOps.size(); i != e; ++i) {
    if (MemOps[i].Position < Position) {
      Position = MemOps[i].Position;
      Loc = MemOps[i].MBBI;
    }
  }

  if (Loc != MBB.begin())
    RS->forward(prior(Loc));
}

static int getMemoryOpOffset(const MachineInstr *MI) {
  int Opcode = MI->getOpcode();
  bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
  bool isAM3 = Opcode == ARM::LDRD || Opcode == ARM::STRD;
  unsigned NumOperands = MI->getDesc().getNumOperands();
  unsigned OffField = MI->getOperand(NumOperands-3).getImm();

  if (Opcode == ARM::t2LDRi12 || Opcode == ARM::t2LDRi8 ||
      Opcode == ARM::t2STRi12 || Opcode == ARM::t2STRi8 ||
      Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8)
    return OffField;

  int Offset = isAM2
    ? ARM_AM::getAM2Offset(OffField)
    : (isAM3 ? ARM_AM::getAM3Offset(OffField)
             : ARM_AM::getAM5Offset(OffField) * 4);
  if (isAM2) {
    if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
      Offset = -Offset;
  } else if (isAM3) {
    if (ARM_AM::getAM3Op(OffField) == ARM_AM::sub)
      Offset = -Offset;
  } else {
    if (ARM_AM::getAM5Op(OffField) == ARM_AM::sub)
      Offset = -Offset;
  }
  return Offset;
}

static void InsertLDR_STR(MachineBasicBlock &MBB,
                          MachineBasicBlock::iterator &MBBI,
                          int OffImm, bool isDef,
                          DebugLoc dl, unsigned NewOpc,
                          unsigned Reg, bool RegDeadKill,
                          unsigned BaseReg, bool BaseKill,
                          unsigned OffReg, bool OffKill,
                          ARMCC::CondCodes Pred, unsigned PredReg,
                          const TargetInstrInfo *TII) {
  unsigned Offset;
  if (OffImm < 0)
    Offset = ARM_AM::getAM2Opc(ARM_AM::sub, -OffImm, ARM_AM::no_shift);
  else
    Offset = ARM_AM::getAM2Opc(ARM_AM::add, OffImm, ARM_AM::no_shift);
  if (isDef)
    BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
      .addReg(Reg, getDefRegState(true) | getDeadRegState(RegDeadKill))
      .addReg(BaseReg, getKillRegState(BaseKill))
      .addReg(OffReg,  getKillRegState(OffKill))
      .addImm(Offset)
      .addImm(Pred).addReg(PredReg);
  else
    BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
      .addReg(Reg, getKillRegState(RegDeadKill))
      .addReg(BaseReg, getKillRegState(BaseKill))
      .addReg(OffReg,  getKillRegState(OffKill))
      .addImm(Offset)
      .addImm(Pred).addReg(PredReg);
}

bool ARMLoadStoreOpt::FixInvalidRegPairOp(MachineBasicBlock &MBB,
                                          MachineBasicBlock::iterator &MBBI) {
  MachineInstr *MI = &*MBBI;
  unsigned Opcode = MI->getOpcode();
  if (Opcode == ARM::LDRD || Opcode == ARM::STRD) {
    unsigned EvenReg = MI->getOperand(0).getReg();
    unsigned OddReg  = MI->getOperand(1).getReg();
    unsigned EvenRegNum = TRI->getDwarfRegNum(EvenReg, false);
    unsigned OddRegNum  = TRI->getDwarfRegNum(OddReg, false);
    if ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum)
      return false;

    bool isLd = Opcode == ARM::LDRD;
    bool EvenDeadKill = isLd ?
      MI->getOperand(0).isDead() : MI->getOperand(0).isKill();
    bool OddDeadKill  = isLd ?
      MI->getOperand(1).isDead() : MI->getOperand(1).isKill();
    const MachineOperand &BaseOp = MI->getOperand(2);
    unsigned BaseReg = BaseOp.getReg();
    bool BaseKill = BaseOp.isKill();
    const MachineOperand &OffOp = MI->getOperand(3);
    unsigned OffReg = OffOp.getReg();
    bool OffKill = OffOp.isKill();
    int OffImm = getMemoryOpOffset(MI);
    unsigned PredReg = 0;
    ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);

    if (OddRegNum > EvenRegNum && OffReg == 0 && OffImm == 0) {
      // Ascending register numbers and no offset. It's safe to change it to a
      // ldm or stm.
      unsigned NewOpc = (Opcode == ARM::LDRD) ? ARM::LDM : ARM::STM;
      if (isLd) {
        BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
          .addReg(BaseReg, getKillRegState(BaseKill))
          .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
          .addImm(Pred).addReg(PredReg)
          .addReg(EvenReg, getDefRegState(isLd) | getDeadRegState(EvenDeadKill))
          .addReg(OddReg, getDefRegState(isLd) | getDeadRegState(OddDeadKill));
        ++NumLDRD2LDM;
      } else {
        BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(NewOpc))
          .addReg(BaseReg, getKillRegState(BaseKill))
          .addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
          .addImm(Pred).addReg(PredReg)
          .addReg(EvenReg, getKillRegState(EvenDeadKill))
          .addReg(OddReg, getKillRegState(OddDeadKill));
        ++NumSTRD2STM;
      }
    } else {
      // Split into two instructions.
      unsigned NewOpc = (Opcode == ARM::LDRD) ? ARM::LDR : ARM::STR;
      DebugLoc dl = MBBI->getDebugLoc();
      // If this is a load and base register is killed, it may have been
      // re-defed by the load, make sure the first load does not clobber it.
      if (isLd &&
          (BaseKill || OffKill) &&
          (TRI->regsOverlap(EvenReg, BaseReg) ||
           (OffReg && TRI->regsOverlap(EvenReg, OffReg)))) {
        assert(!TRI->regsOverlap(OddReg, BaseReg) &&
               (!OffReg || !TRI->regsOverlap(OddReg, OffReg)));
        InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc, OddReg, OddDeadKill,
                      BaseReg, false, OffReg, false, Pred, PredReg, TII);
        InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc, EvenReg, EvenDeadKill,
                      BaseReg, BaseKill, OffReg, OffKill, Pred, PredReg, TII);
      } else {
        InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
                      EvenReg, EvenDeadKill, BaseReg, false, OffReg, false,
                      Pred, PredReg, TII);
        InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
                      OddReg, OddDeadKill, BaseReg, BaseKill, OffReg, OffKill,
                      Pred, PredReg, TII);
      }
      if (isLd)
        ++NumLDRD2LDR;
      else
        ++NumSTRD2STR;
    }

    MBBI = prior(MBBI);
    MBB.erase(MI);
  }
  return false;
}

/// LoadStoreMultipleOpti - An optimization pass to turn multiple LDR / STR
/// ops of the same base and incrementing offset into LDM / STM ops.
bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
  unsigned NumMerges = 0;
  unsigned NumMemOps = 0;
  MemOpQueue MemOps;
  unsigned CurrBase = 0;
  int CurrOpc = -1;
  unsigned CurrSize = 0;
  ARMCC::CondCodes CurrPred = ARMCC::AL;
  unsigned CurrPredReg = 0;
  unsigned Position = 0;
  SmallVector<MachineBasicBlock::iterator,4> Merges;

  RS->enterBasicBlock(&MBB);
  MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
  while (MBBI != E) {
    if (FixInvalidRegPairOp(MBB, MBBI))
      continue;

    bool Advance  = false;
    bool TryMerge = false;
    bool Clobber  = false;

    bool isMemOp = isMemoryOp(MBBI);
    if (isMemOp) {
      int Opcode = MBBI->getOpcode();
      unsigned Size = getLSMultipleTransferSize(MBBI);
      unsigned Base = MBBI->getOperand(1).getReg();
      unsigned PredReg = 0;
      ARMCC::CondCodes Pred = llvm::getInstrPredicate(MBBI, PredReg);
      int Offset = getMemoryOpOffset(MBBI);
      // Watch out for:
      // r4 := ldr [r5]
      // r5 := ldr [r5, #4]
      // r6 := ldr [r5, #8]
      //
      // The second ldr has effectively broken the chain even though it
      // looks like the later ldr(s) use the same base register. Try to
      // merge the ldr's so far, including this one. But don't try to
      // combine the following ldr(s).
      Clobber = (isi32Load(Opcode) && Base == MBBI->getOperand(0).getReg());
      if (CurrBase == 0 && !Clobber) {
        // Start of a new chain.
        CurrBase = Base;
        CurrOpc  = Opcode;
        CurrSize = Size;
        CurrPred = Pred;
        CurrPredReg = PredReg;
        MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
        NumMemOps++;
        Advance = true;
      } else {
        if (Clobber) {
          TryMerge = true;
          Advance = true;
        }

        if (CurrOpc == Opcode && CurrBase == Base && CurrPred == Pred) {
          // No need to match PredReg.
          // Continue adding to the queue.
          if (Offset > MemOps.back().Offset) {
            MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
            NumMemOps++;
            Advance = true;
          } else {
            for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
                 I != E; ++I) {
              if (Offset < I->Offset) {
                MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
                NumMemOps++;
                Advance = true;
                break;
              } else if (Offset == I->Offset) {
                // Collision! This can't be merged!
                break;
              }
            }
          }
        }
      }
    }

    if (Advance) {
      ++Position;
      ++MBBI;
    } else
      TryMerge = true;

    if (TryMerge) {
      if (NumMemOps > 1) {
        // Try to find a free register to use as a new base in case it's needed.
        // First advance to the instruction just before the start of the chain.
        AdvanceRS(MBB, MemOps);
        // Find a scratch register.
        unsigned Scratch = RS->FindUnusedReg(&ARM::GPRRegClass);
        // Process the load / store instructions.
        RS->forward(prior(MBBI));

        // Merge ops.
        Merges.clear();
        MergeLDR_STR(MBB, 0, CurrBase, CurrOpc, CurrSize,
                     CurrPred, CurrPredReg, Scratch, MemOps, Merges);

        // Try folding preceeding/trailing base inc/dec into the generated
        // LDM/STM ops.
        for (unsigned i = 0, e = Merges.size(); i < e; ++i)
          if (MergeBaseUpdateLSMultiple(MBB, Merges[i], Advance, MBBI))
            ++NumMerges;
        NumMerges += Merges.size();

        // Try folding preceeding/trailing base inc/dec into those load/store
        // that were not merged to form LDM/STM ops.
        for (unsigned i = 0; i != NumMemOps; ++i)
          if (!MemOps[i].Merged)
            if (MergeBaseUpdateLoadStore(MBB, MemOps[i].MBBI, TII,Advance,MBBI))
              ++NumMerges;

        // RS may be pointing to an instruction that's deleted.
        RS->skipTo(prior(MBBI));
      } else if (NumMemOps == 1) {
        // Try folding preceeding/trailing base inc/dec into the single
        // load/store.
        if (MergeBaseUpdateLoadStore(MBB, MemOps[0].MBBI, TII, Advance, MBBI)) {
          ++NumMerges;
          RS->forward(prior(MBBI));
        }
      }

      CurrBase = 0;
      CurrOpc = -1;
      CurrSize = 0;
      CurrPred = ARMCC::AL;
      CurrPredReg = 0;
      if (NumMemOps) {
        MemOps.clear();
        NumMemOps = 0;
      }

      // If iterator hasn't been advanced and this is not a memory op, skip it.
      // It can't start a new chain anyway.
      if (!Advance && !isMemOp && MBBI != E) {
        ++Position;
        ++MBBI;
      }
    }
  }
  return NumMerges > 0;
}

namespace {
  struct OffsetCompare {
    bool operator()(const MachineInstr *LHS, const MachineInstr *RHS) const {
      int LOffset = getMemoryOpOffset(LHS);
      int ROffset = getMemoryOpOffset(RHS);
      assert(LHS == RHS || LOffset != ROffset);
      return LOffset > ROffset;
    }
  };
}

/// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
/// (bx lr) into the preceeding stack restore so it directly restore the value
/// of LR into pc.
///   ldmfd sp!, {r7, lr}
///   bx lr
/// =>
///   ldmfd sp!, {r7, pc}
bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
  if (MBB.empty()) return false;

  MachineBasicBlock::iterator MBBI = prior(MBB.end());
  if (MBBI != MBB.begin() &&
      (MBBI->getOpcode() == ARM::BX_RET || MBBI->getOpcode() == ARM::tBX_RET)) {
    MachineInstr *PrevMI = prior(MBBI);
    if (PrevMI->getOpcode() == ARM::LDM || PrevMI->getOpcode() == ARM::t2LDM) {
      MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
      if (MO.getReg() != ARM::LR)
        return false;
      unsigned NewOpc = isThumb2 ? ARM::t2LDM_RET : ARM::LDM_RET;
      PrevMI->setDesc(TII->get(NewOpc));
      MO.setReg(ARM::PC);
      MBB.erase(MBBI);
      return true;
    }
  }
  return false;
}

bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
  const TargetMachine &TM = Fn.getTarget();
  AFI = Fn.getInfo<ARMFunctionInfo>();
  TII = TM.getInstrInfo();
  TRI = TM.getRegisterInfo();
  RS = new RegScavenger();
  isThumb2 = AFI->isThumb2Function();

  bool Modified = false;
  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
       ++MFI) {
    MachineBasicBlock &MBB = *MFI;
    Modified |= LoadStoreMultipleOpti(MBB);
    Modified |= MergeReturnIntoLDM(MBB);
  }

  delete RS;
  return Modified;
}


/// ARMPreAllocLoadStoreOpt - Pre- register allocation pass that move
/// load / stores from consecutive locations close to make it more
/// likely they will be combined later.

namespace {
  struct VISIBILITY_HIDDEN ARMPreAllocLoadStoreOpt : public MachineFunctionPass{
    static char ID;
    ARMPreAllocLoadStoreOpt() : MachineFunctionPass(&ID) {}

    const TargetData *TD;
    const TargetInstrInfo *TII;
    const TargetRegisterInfo *TRI;
    const ARMSubtarget *STI;
    MachineRegisterInfo *MRI;

    virtual bool runOnMachineFunction(MachineFunction &Fn);

    virtual const char *getPassName() const {
      return "ARM pre- register allocation load / store optimization pass";
    }

  private:
    bool CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl,
                          unsigned &NewOpc, unsigned &EvenReg,
                          unsigned &OddReg, unsigned &BaseReg,
                          unsigned &OffReg, unsigned &Offset,
                          unsigned &PredReg, ARMCC::CondCodes &Pred);
    bool RescheduleOps(MachineBasicBlock *MBB,
                       SmallVector<MachineInstr*, 4> &Ops,
                       unsigned Base, bool isLd,
                       DenseMap<MachineInstr*, unsigned> &MI2LocMap);
    bool RescheduleLoadStoreInstrs(MachineBasicBlock *MBB);
  };
  char ARMPreAllocLoadStoreOpt::ID = 0;
}

bool ARMPreAllocLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
  TD  = Fn.getTarget().getTargetData();
  TII = Fn.getTarget().getInstrInfo();
  TRI = Fn.getTarget().getRegisterInfo();
  STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
  MRI = &Fn.getRegInfo();

  bool Modified = false;
  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
       ++MFI)
    Modified |= RescheduleLoadStoreInstrs(MFI);

  return Modified;
}

static bool IsSafeAndProfitableToMove(bool isLd, unsigned Base,
                                      MachineBasicBlock::iterator I,
                                      MachineBasicBlock::iterator E,
                                      SmallPtrSet<MachineInstr*, 4> &MemOps,
                                      SmallSet<unsigned, 4> &MemRegs,
                                      const TargetRegisterInfo *TRI) {
  // Are there stores / loads / calls between them?
  // FIXME: This is overly conservative. We should make use of alias information
  // some day.
  SmallSet<unsigned, 4> AddedRegPressure;
  while (++I != E) {
    if (MemOps.count(&*I))
      continue;
    const TargetInstrDesc &TID = I->getDesc();
    if (TID.isCall() || TID.isTerminator() || TID.hasUnmodeledSideEffects())
      return false;
    if (isLd && TID.mayStore())
      return false;
    if (!isLd) {
      if (TID.mayLoad())
        return false;
      // It's not safe to move the first 'str' down.
      // str r1, [r0]
      // strh r5, [r0]
      // str r4, [r0, #+4]
      if (TID.mayStore())
        return false;
    }
    for (unsigned j = 0, NumOps = I->getNumOperands(); j != NumOps; ++j) {
      MachineOperand &MO = I->getOperand(j);
      if (!MO.isReg())
        continue;
      unsigned Reg = MO.getReg();
      if (MO.isDef() && TRI->regsOverlap(Reg, Base))
        return false;
      if (Reg != Base && !MemRegs.count(Reg))
        AddedRegPressure.insert(Reg);
    }
  }

  // Estimate register pressure increase due to the transformation.
  if (MemRegs.size() <= 4)
    // Ok if we are moving small number of instructions.
    return true;
  return AddedRegPressure.size() <= MemRegs.size() * 2;
}

bool
ARMPreAllocLoadStoreOpt::CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1,
                                          DebugLoc &dl,
                                          unsigned &NewOpc, unsigned &EvenReg,
                                          unsigned &OddReg, unsigned &BaseReg,
                                          unsigned &OffReg, unsigned &Offset,
                                          unsigned &PredReg,
                                          ARMCC::CondCodes &Pred) {
  // FIXME: FLDS / FSTS -> FLDD / FSTD
  unsigned Opcode = Op0->getOpcode();
  if (Opcode == ARM::LDR)
    NewOpc = ARM::LDRD;
  else if (Opcode == ARM::STR)
    NewOpc = ARM::STRD;
  else
    return 0;

  // Must sure the base address satisfies i64 ld / st alignment requirement.
  if (!Op0->hasOneMemOperand() ||
      !Op0->memoperands_begin()->getValue() ||
      Op0->memoperands_begin()->isVolatile())
    return false;

  unsigned Align = Op0->memoperands_begin()->getAlignment();
  unsigned ReqAlign = STI->hasV6Ops()
    ? TD->getPrefTypeAlignment(
  Type::getInt64Ty(Op0->getParent()->getParent()->getFunction()->getContext())) 
  : 8; // Pre-v6 need 8-byte align
  if (Align < ReqAlign)
    return false;

  // Then make sure the immediate offset fits.
  int OffImm = getMemoryOpOffset(Op0);
  ARM_AM::AddrOpc AddSub = ARM_AM::add;
  if (OffImm < 0) {
    AddSub = ARM_AM::sub;
    OffImm = - OffImm;
  }
  if (OffImm >= 256) // 8 bits
    return false;
  Offset = ARM_AM::getAM3Opc(AddSub, OffImm);

  EvenReg = Op0->getOperand(0).getReg();
  OddReg  = Op1->getOperand(0).getReg();
  if (EvenReg == OddReg)
    return false;
  BaseReg = Op0->getOperand(1).getReg();
  OffReg = Op0->getOperand(2).getReg();
  Pred = llvm::getInstrPredicate(Op0, PredReg);
  dl = Op0->getDebugLoc();
  return true;
}

bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
                                 SmallVector<MachineInstr*, 4> &Ops,
                                 unsigned Base, bool isLd,
                                 DenseMap<MachineInstr*, unsigned> &MI2LocMap) {
  bool RetVal = false;

  // Sort by offset (in reverse order).
  std::sort(Ops.begin(), Ops.end(), OffsetCompare());

  // The loads / stores of the same base are in order. Scan them from first to
  // last and check for the followins:
  // 1. Any def of base.
  // 2. Any gaps.
  while (Ops.size() > 1) {
    unsigned FirstLoc = ~0U;
    unsigned LastLoc = 0;
    MachineInstr *FirstOp = 0;
    MachineInstr *LastOp = 0;
    int LastOffset = 0;
    unsigned LastOpcode = 0;
    unsigned LastBytes = 0;
    unsigned NumMove = 0;
    for (int i = Ops.size() - 1; i >= 0; --i) {
      MachineInstr *Op = Ops[i];
      unsigned Loc = MI2LocMap[Op];
      if (Loc <= FirstLoc) {
        FirstLoc = Loc;
        FirstOp = Op;
      }
      if (Loc >= LastLoc) {
        LastLoc = Loc;
        LastOp = Op;
      }

      unsigned Opcode = Op->getOpcode();
      if (LastOpcode && Opcode != LastOpcode)
        break;

      int Offset = getMemoryOpOffset(Op);
      unsigned Bytes = getLSMultipleTransferSize(Op);
      if (LastBytes) {
        if (Bytes != LastBytes || Offset != (LastOffset + (int)Bytes))
          break;
      }
      LastOffset = Offset;
      LastBytes = Bytes;
      LastOpcode = Opcode;
      if (++NumMove == 8) // FIXME: Tune
        break;
    }

    if (NumMove <= 1)
      Ops.pop_back();
    else {
      SmallPtrSet<MachineInstr*, 4> MemOps;
      SmallSet<unsigned, 4> MemRegs;
      for (int i = NumMove-1; i >= 0; --i) {
        MemOps.insert(Ops[i]);
        MemRegs.insert(Ops[i]->getOperand(0).getReg());
      }

      // Be conservative, if the instructions are too far apart, don't
      // move them. We want to limit the increase of register pressure.
      bool DoMove = (LastLoc - FirstLoc) <= NumMove*4; // FIXME: Tune this.
      if (DoMove)
        DoMove = IsSafeAndProfitableToMove(isLd, Base, FirstOp, LastOp,
                                           MemOps, MemRegs, TRI);
      if (!DoMove) {
        for (unsigned i = 0; i != NumMove; ++i)
          Ops.pop_back();
      } else {
        // This is the new location for the loads / stores.
        MachineBasicBlock::iterator InsertPos = isLd ? FirstOp : LastOp;
        while (InsertPos != MBB->end() && MemOps.count(InsertPos))
          ++InsertPos;

        // If we are moving a pair of loads / stores, see if it makes sense
        // to try to allocate a pair of registers that can form register pairs.
        MachineInstr *Op0 = Ops.back();
        MachineInstr *Op1 = Ops[Ops.size()-2];
        unsigned EvenReg = 0, OddReg = 0;
        unsigned BaseReg = 0, OffReg = 0, PredReg = 0;
        ARMCC::CondCodes Pred = ARMCC::AL;
        unsigned NewOpc = 0;
        unsigned Offset = 0;
        DebugLoc dl;
        if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
                                             EvenReg, OddReg, BaseReg, OffReg,
                                             Offset, PredReg, Pred)) {
          Ops.pop_back();
          Ops.pop_back();

          // Form the pair instruction.
          if (isLd) {
            BuildMI(*MBB, InsertPos, dl, TII->get(NewOpc))
              .addReg(EvenReg, RegState::Define)
              .addReg(OddReg, RegState::Define)
              .addReg(BaseReg).addReg(0).addImm(Offset)
              .addImm(Pred).addReg(PredReg);
            ++NumLDRDFormed;
          } else {
            BuildMI(*MBB, InsertPos, dl, TII->get(NewOpc))
              .addReg(EvenReg)
              .addReg(OddReg)
              .addReg(BaseReg).addReg(0).addImm(Offset)
              .addImm(Pred).addReg(PredReg);
            ++NumSTRDFormed;
          }
          MBB->erase(Op0);
          MBB->erase(Op1);

          // Add register allocation hints to form register pairs.
          MRI->setRegAllocationHint(EvenReg, ARMRI::RegPairEven, OddReg);
          MRI->setRegAllocationHint(OddReg,  ARMRI::RegPairOdd, EvenReg);
        } else {
          for (unsigned i = 0; i != NumMove; ++i) {
            MachineInstr *Op = Ops.back();
            Ops.pop_back();
            MBB->splice(InsertPos, MBB, Op);
          }
        }

        NumLdStMoved += NumMove;
        RetVal = true;
      }
    }
  }

  return RetVal;
}

bool
ARMPreAllocLoadStoreOpt::RescheduleLoadStoreInstrs(MachineBasicBlock *MBB) {
  bool RetVal = false;

  DenseMap<MachineInstr*, unsigned> MI2LocMap;
  DenseMap<unsigned, SmallVector<MachineInstr*, 4> > Base2LdsMap;
  DenseMap<unsigned, SmallVector<MachineInstr*, 4> > Base2StsMap;
  SmallVector<unsigned, 4> LdBases;
  SmallVector<unsigned, 4> StBases;

  unsigned Loc = 0;
  MachineBasicBlock::iterator MBBI = MBB->begin();
  MachineBasicBlock::iterator E = MBB->end();
  while (MBBI != E) {
    for (; MBBI != E; ++MBBI) {
      MachineInstr *MI = MBBI;
      const TargetInstrDesc &TID = MI->getDesc();
      if (TID.isCall() || TID.isTerminator()) {
        // Stop at barriers.
        ++MBBI;
        break;
      }

      MI2LocMap[MI] = Loc++;
      if (!isMemoryOp(MI))
        continue;
      unsigned PredReg = 0;
      if (llvm::getInstrPredicate(MI, PredReg) != ARMCC::AL)
        continue;

      int Opcode = MI->getOpcode();
      bool isLd = Opcode == ARM::LDR ||
        Opcode == ARM::FLDS || Opcode == ARM::FLDD;
      unsigned Base = MI->getOperand(1).getReg();
      int Offset = getMemoryOpOffset(MI);

      bool StopHere = false;
      if (isLd) {
        DenseMap<unsigned, SmallVector<MachineInstr*, 4> >::iterator BI =
          Base2LdsMap.find(Base);
        if (BI != Base2LdsMap.end()) {
          for (unsigned i = 0, e = BI->second.size(); i != e; ++i) {
            if (Offset == getMemoryOpOffset(BI->second[i])) {
              StopHere = true;
              break;
            }
          }
          if (!StopHere)
            BI->second.push_back(MI);
        } else {
          SmallVector<MachineInstr*, 4> MIs;
          MIs.push_back(MI);
          Base2LdsMap[Base] = MIs;
          LdBases.push_back(Base);
        }
      } else {
        DenseMap<unsigned, SmallVector<MachineInstr*, 4> >::iterator BI =
          Base2StsMap.find(Base);
        if (BI != Base2StsMap.end()) {
          for (unsigned i = 0, e = BI->second.size(); i != e; ++i) {
            if (Offset == getMemoryOpOffset(BI->second[i])) {
              StopHere = true;
              break;
            }
          }
          if (!StopHere)
            BI->second.push_back(MI);
        } else {
          SmallVector<MachineInstr*, 4> MIs;
          MIs.push_back(MI);
          Base2StsMap[Base] = MIs;
          StBases.push_back(Base);
        }
      }

      if (StopHere) {
        // Found a duplicate (a base+offset combination that's seen earlier).
        // Backtrack.
        --Loc;
        break;
      }
    }

    // Re-schedule loads.
    for (unsigned i = 0, e = LdBases.size(); i != e; ++i) {
      unsigned Base = LdBases[i];
      SmallVector<MachineInstr*, 4> &Lds = Base2LdsMap[Base];
      if (Lds.size() > 1)
        RetVal |= RescheduleOps(MBB, Lds, Base, true, MI2LocMap);
    }

    // Re-schedule stores.
    for (unsigned i = 0, e = StBases.size(); i != e; ++i) {
      unsigned Base = StBases[i];
      SmallVector<MachineInstr*, 4> &Sts = Base2StsMap[Base];
      if (Sts.size() > 1)
        RetVal |= RescheduleOps(MBB, Sts, Base, false, MI2LocMap);
    }

    if (MBBI != E) {
      Base2LdsMap.clear();
      Base2StsMap.clear();
      LdBases.clear();
      StBases.clear();
    }
  }

  return RetVal;
}


/// createARMLoadStoreOptimizationPass - returns an instance of the load / store
/// optimization pass.
FunctionPass *llvm::createARMLoadStoreOptimizationPass(bool PreAlloc) {
  if (PreAlloc)
    return new ARMPreAllocLoadStoreOpt();
  return new ARMLoadStoreOpt();
}