Training courses

Kernel and Embedded Linux

Bootlin training courses

Embedded Linux, kernel,
Yocto Project, Buildroot, real-time,
graphics, boot time, debugging...

Bootlin logo

Elixir Cross Referencer

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

#include "ARMTargetTransformInfo.h"
#include "ARMSubtarget.h"
#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/CodeGen/CostTable.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsARM.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/Type.h"
#include "llvm/MC/SubtargetFeature.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/MachineValueType.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/Utils/LoopUtils.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <utility>

using namespace llvm;

#define DEBUG_TYPE "armtti"

static cl::opt<bool> EnableMaskedLoadStores(
  "enable-arm-maskedldst", cl::Hidden, cl::init(true),
  cl::desc("Enable the generation of masked loads and stores"));

static cl::opt<bool> DisableLowOverheadLoops(
  "disable-arm-loloops", cl::Hidden, cl::init(false),
  cl::desc("Disable the generation of low-overhead loops"));

extern cl::opt<TailPredication::Mode> EnableTailPredication;

extern cl::opt<bool> EnableMaskedGatherScatters;

bool ARMTTIImpl::areInlineCompatible(const Function *Caller,
                                     const Function *Callee) const {
  const TargetMachine &TM = getTLI()->getTargetMachine();
  const FeatureBitset &CallerBits =
      TM.getSubtargetImpl(*Caller)->getFeatureBits();
  const FeatureBitset &CalleeBits =
      TM.getSubtargetImpl(*Callee)->getFeatureBits();

  // To inline a callee, all features not in the allowed list must match exactly.
  bool MatchExact = (CallerBits & ~InlineFeaturesAllowed) ==
                    (CalleeBits & ~InlineFeaturesAllowed);
  // For features in the allowed list, the callee's features must be a subset of
  // the callers'.
  bool MatchSubset = ((CallerBits & CalleeBits) & InlineFeaturesAllowed) ==
                     (CalleeBits & InlineFeaturesAllowed);
  return MatchExact && MatchSubset;
}

bool ARMTTIImpl::shouldFavorBackedgeIndex(const Loop *L) const {
  if (L->getHeader()->getParent()->hasOptSize())
    return false;
  if (ST->hasMVEIntegerOps())
    return false;
  return ST->isMClass() && ST->isThumb2() && L->getNumBlocks() == 1;
}

bool ARMTTIImpl::shouldFavorPostInc() const {
  if (ST->hasMVEIntegerOps())
    return true;
  return false;
}

int ARMTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
                              TTI::TargetCostKind CostKind) {
  assert(Ty->isIntegerTy());

 unsigned Bits = Ty->getPrimitiveSizeInBits();
 if (Bits == 0 || Imm.getActiveBits() >= 64)
   return 4;

  int64_t SImmVal = Imm.getSExtValue();
  uint64_t ZImmVal = Imm.getZExtValue();
  if (!ST->isThumb()) {
    if ((SImmVal >= 0 && SImmVal < 65536) ||
        (ARM_AM::getSOImmVal(ZImmVal) != -1) ||
        (ARM_AM::getSOImmVal(~ZImmVal) != -1))
      return 1;
    return ST->hasV6T2Ops() ? 2 : 3;
  }
  if (ST->isThumb2()) {
    if ((SImmVal >= 0 && SImmVal < 65536) ||
        (ARM_AM::getT2SOImmVal(ZImmVal) != -1) ||
        (ARM_AM::getT2SOImmVal(~ZImmVal) != -1))
      return 1;
    return ST->hasV6T2Ops() ? 2 : 3;
  }
  // Thumb1, any i8 imm cost 1.
  if (Bits == 8 || (SImmVal >= 0 && SImmVal < 256))
    return 1;
  if ((~SImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal))
    return 2;
  // Load from constantpool.
  return 3;
}

// Constants smaller than 256 fit in the immediate field of
// Thumb1 instructions so we return a zero cost and 1 otherwise.
int ARMTTIImpl::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx,
                                      const APInt &Imm, Type *Ty) {
  if (Imm.isNonNegative() && Imm.getLimitedValue() < 256)
    return 0;

  return 1;
}

int ARMTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm,
                                  Type *Ty, TTI::TargetCostKind CostKind) {
  // Division by a constant can be turned into multiplication, but only if we
  // know it's constant. So it's not so much that the immediate is cheap (it's
  // not), but that the alternative is worse.
  // FIXME: this is probably unneeded with GlobalISel.
  if ((Opcode == Instruction::SDiv || Opcode == Instruction::UDiv ||
       Opcode == Instruction::SRem || Opcode == Instruction::URem) &&
      Idx == 1)
    return 0;

  if (Opcode == Instruction::And) {
    // UXTB/UXTH
    if (Imm == 255 || Imm == 65535)
      return 0;
    // Conversion to BIC is free, and means we can use ~Imm instead.
    return std::min(getIntImmCost(Imm, Ty, CostKind),
                    getIntImmCost(~Imm, Ty, CostKind));
  }

  if (Opcode == Instruction::Add)
    // Conversion to SUB is free, and means we can use -Imm instead.
    return std::min(getIntImmCost(Imm, Ty, CostKind),
                    getIntImmCost(-Imm, Ty, CostKind));

  if (Opcode == Instruction::ICmp && Imm.isNegative() &&
      Ty->getIntegerBitWidth() == 32) {
    int64_t NegImm = -Imm.getSExtValue();
    if (ST->isThumb2() && NegImm < 1<<12)
      // icmp X, #-C -> cmn X, #C
      return 0;
    if (ST->isThumb() && NegImm < 1<<8)
      // icmp X, #-C -> adds X, #C
      return 0;
  }

  // xor a, -1 can always be folded to MVN
  if (Opcode == Instruction::Xor && Imm.isAllOnesValue())
    return 0;

  return getIntImmCost(Imm, Ty, CostKind);
}

int ARMTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
                                 TTI::TargetCostKind CostKind,
                                 const Instruction *I) {
  int ISD = TLI->InstructionOpcodeToISD(Opcode);
  assert(ISD && "Invalid opcode");

  // TODO: Allow non-throughput costs that aren't binary.
  auto AdjustCost = [&CostKind](int Cost) {
    if (CostKind != TTI::TCK_RecipThroughput)
      return Cost == 0 ? 0 : 1;
    return Cost;
  };

  EVT SrcTy = TLI->getValueType(DL, Src);
  EVT DstTy = TLI->getValueType(DL, Dst);

  if (!SrcTy.isSimple() || !DstTy.isSimple())
    return AdjustCost(BaseT::getCastInstrCost(Opcode, Dst, Src, CostKind, I));

  // The extend of a load is free
  if (I && isa<LoadInst>(I->getOperand(0))) {
    static const TypeConversionCostTblEntry LoadConversionTbl[] = {
        {ISD::SIGN_EXTEND, MVT::i32, MVT::i16, 0},
        {ISD::ZERO_EXTEND, MVT::i32, MVT::i16, 0},
        {ISD::SIGN_EXTEND, MVT::i32, MVT::i8, 0},
        {ISD::ZERO_EXTEND, MVT::i32, MVT::i8, 0},
        {ISD::SIGN_EXTEND, MVT::i16, MVT::i8, 0},
        {ISD::ZERO_EXTEND, MVT::i16, MVT::i8, 0},
        {ISD::SIGN_EXTEND, MVT::i64, MVT::i32, 1},
        {ISD::ZERO_EXTEND, MVT::i64, MVT::i32, 1},
        {ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 1},
        {ISD::ZERO_EXTEND, MVT::i64, MVT::i16, 1},
        {ISD::SIGN_EXTEND, MVT::i64, MVT::i8, 1},
        {ISD::ZERO_EXTEND, MVT::i64, MVT::i8, 1},
    };
    if (const auto *Entry = ConvertCostTableLookup(
            LoadConversionTbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
      return AdjustCost(Entry->Cost);

    static const TypeConversionCostTblEntry MVELoadConversionTbl[] = {
        {ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0},
        {ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0},
        {ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 0},
        {ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 0},
        {ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 0},
        {ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 0},
        // The following extend from a legal type to an illegal type, so need to
        // split the load. This introduced an extra load operation, but the
        // extend is still "free".
        {ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1},
        {ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1},
        {ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 3},
        {ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 3},
        {ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1},
        {ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1},
    };
    if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
      if (const auto *Entry =
              ConvertCostTableLookup(MVELoadConversionTbl, ISD,
                                     DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
        return AdjustCost(Entry->Cost * ST->getMVEVectorCostFactor());
    }

    static const TypeConversionCostTblEntry MVEFLoadConversionTbl[] = {
        // FPExtends are similar but also require the VCVT instructions.
        {ISD::FP_EXTEND, MVT::v4f32, MVT::v4f16, 1},
        {ISD::FP_EXTEND, MVT::v8f32, MVT::v8f16, 3},
    };
    if (SrcTy.isVector() && ST->hasMVEFloatOps()) {
      if (const auto *Entry =
              ConvertCostTableLookup(MVEFLoadConversionTbl, ISD,
                                     DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
        return AdjustCost(Entry->Cost * ST->getMVEVectorCostFactor());
    }
  }

  // The truncate of a store is free. This is the mirror of extends above.
  if (I && I->hasOneUse() && isa<StoreInst>(*I->user_begin())) {
    static const TypeConversionCostTblEntry MVELoadConversionTbl[] = {
        {ISD::TRUNCATE, MVT::v4i32, MVT::v4i16, 0},
        {ISD::TRUNCATE, MVT::v4i32, MVT::v4i8, 0},
        {ISD::TRUNCATE, MVT::v8i16, MVT::v8i8, 0},
        {ISD::TRUNCATE, MVT::v8i32, MVT::v8i16, 1},
        {ISD::TRUNCATE, MVT::v16i32, MVT::v16i8, 3},
        {ISD::TRUNCATE, MVT::v16i16, MVT::v16i8, 1},
    };
    if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
      if (const auto *Entry =
              ConvertCostTableLookup(MVELoadConversionTbl, ISD, SrcTy.getSimpleVT(),
                                     DstTy.getSimpleVT()))
        return AdjustCost(Entry->Cost * ST->getMVEVectorCostFactor());
    }

    static const TypeConversionCostTblEntry MVEFLoadConversionTbl[] = {
        {ISD::FP_ROUND, MVT::v4f32, MVT::v4f16, 1},
        {ISD::FP_ROUND, MVT::v8f32, MVT::v8f16, 3},
    };
    if (SrcTy.isVector() && ST->hasMVEFloatOps()) {
      if (const auto *Entry =
              ConvertCostTableLookup(MVEFLoadConversionTbl, ISD, SrcTy.getSimpleVT(),
                                     DstTy.getSimpleVT()))
        return AdjustCost(Entry->Cost * ST->getMVEVectorCostFactor());
    }
  }

  // NEON vector operations that can extend their inputs.
  if ((ISD == ISD::SIGN_EXTEND || ISD == ISD::ZERO_EXTEND) &&
      I && I->hasOneUse() && ST->hasNEON() && SrcTy.isVector()) {
    static const TypeConversionCostTblEntry NEONDoubleWidthTbl[] = {
      // vaddl
      { ISD::ADD, MVT::v4i32, MVT::v4i16, 0 },
      { ISD::ADD, MVT::v8i16, MVT::v8i8,  0 },
      // vsubl
      { ISD::SUB, MVT::v4i32, MVT::v4i16, 0 },
      { ISD::SUB, MVT::v8i16, MVT::v8i8,  0 },
      // vmull
      { ISD::MUL, MVT::v4i32, MVT::v4i16, 0 },
      { ISD::MUL, MVT::v8i16, MVT::v8i8,  0 },
      // vshll
      { ISD::SHL, MVT::v4i32, MVT::v4i16, 0 },
      { ISD::SHL, MVT::v8i16, MVT::v8i8,  0 },
    };

    auto *User = cast<Instruction>(*I->user_begin());
    int UserISD = TLI->InstructionOpcodeToISD(User->getOpcode());
    if (auto *Entry = ConvertCostTableLookup(NEONDoubleWidthTbl, UserISD,
                                             DstTy.getSimpleVT(),
                                             SrcTy.getSimpleVT())) {
      return AdjustCost(Entry->Cost);
    }
  }

  // Single to/from double precision conversions.
  if (Src->isVectorTy() && ST->hasNEON() &&
      ((ISD == ISD::FP_ROUND && SrcTy.getScalarType() == MVT::f64 &&
        DstTy.getScalarType() == MVT::f32) ||
       (ISD == ISD::FP_EXTEND && SrcTy.getScalarType() == MVT::f32 &&
        DstTy.getScalarType() == MVT::f64))) {
    static const CostTblEntry NEONFltDblTbl[] = {
        // Vector fptrunc/fpext conversions.
        {ISD::FP_ROUND, MVT::v2f64, 2},
        {ISD::FP_EXTEND, MVT::v2f32, 2},
        {ISD::FP_EXTEND, MVT::v4f32, 4}};

    std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
    if (const auto *Entry = CostTableLookup(NEONFltDblTbl, ISD, LT.second))
      return AdjustCost(LT.first * Entry->Cost);
  }

  // Some arithmetic, load and store operations have specific instructions
  // to cast up/down their types automatically at no extra cost.
  // TODO: Get these tables to know at least what the related operations are.
  static const TypeConversionCostTblEntry NEONVectorConversionTbl[] = {
    { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
    { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
    { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
    { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
    { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64, 0 },
    { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i32, 1 },

    // The number of vmovl instructions for the extension.
    { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8,  1 },
    { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8,  1 },
    { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8,  2 },
    { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8,  2 },
    { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i8,  3 },
    { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i8,  3 },
    { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
    { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
    { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
    { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
    { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
    { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
    { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
    { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
    { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
    { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
    { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
    { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6 },

    // Operations that we legalize using splitting.
    { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i32, 6 },
    { ISD::TRUNCATE,    MVT::v8i8, MVT::v8i32, 3 },

    // Vector float <-> i32 conversions.
    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },

    { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i8, 3 },
    { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i8, 3 },
    { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i16, 2 },
    { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i16, 2 },
    { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i32, 1 },
    { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i32, 1 },
    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1, 3 },
    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1, 3 },
    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8, 3 },
    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8, 3 },
    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 4 },
    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 4 },
    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 2 },
    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 2 },
    { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 8 },
    { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 8 },
    { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 4 },
    { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 4 },

    { ISD::FP_TO_SINT,  MVT::v4i32, MVT::v4f32, 1 },
    { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f32, 1 },
    { ISD::FP_TO_SINT,  MVT::v4i8, MVT::v4f32, 3 },
    { ISD::FP_TO_UINT,  MVT::v4i8, MVT::v4f32, 3 },
    { ISD::FP_TO_SINT,  MVT::v4i16, MVT::v4f32, 2 },
    { ISD::FP_TO_UINT,  MVT::v4i16, MVT::v4f32, 2 },

    // Vector double <-> i32 conversions.
    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },

    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i8, 4 },
    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i8, 4 },
    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i16, 3 },
    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i16, 3 },
    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },

    { ISD::FP_TO_SINT,  MVT::v2i32, MVT::v2f64, 2 },
    { ISD::FP_TO_UINT,  MVT::v2i32, MVT::v2f64, 2 },
    { ISD::FP_TO_SINT,  MVT::v8i16, MVT::v8f32, 4 },
    { ISD::FP_TO_UINT,  MVT::v8i16, MVT::v8f32, 4 },
    { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v16f32, 8 },
    { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 8 }
  };

  if (SrcTy.isVector() && ST->hasNEON()) {
    if (const auto *Entry = ConvertCostTableLookup(NEONVectorConversionTbl, ISD,
                                                   DstTy.getSimpleVT(),
                                                   SrcTy.getSimpleVT()))
      return AdjustCost(Entry->Cost);
  }

  // Scalar float to integer conversions.
  static const TypeConversionCostTblEntry NEONFloatConversionTbl[] = {
    { ISD::FP_TO_SINT,  MVT::i1, MVT::f32, 2 },
    { ISD::FP_TO_UINT,  MVT::i1, MVT::f32, 2 },
    { ISD::FP_TO_SINT,  MVT::i1, MVT::f64, 2 },
    { ISD::FP_TO_UINT,  MVT::i1, MVT::f64, 2 },
    { ISD::FP_TO_SINT,  MVT::i8, MVT::f32, 2 },
    { ISD::FP_TO_UINT,  MVT::i8, MVT::f32, 2 },
    { ISD::FP_TO_SINT,  MVT::i8, MVT::f64, 2 },
    { ISD::FP_TO_UINT,  MVT::i8, MVT::f64, 2 },
    { ISD::FP_TO_SINT,  MVT::i16, MVT::f32, 2 },
    { ISD::FP_TO_UINT,  MVT::i16, MVT::f32, 2 },
    { ISD::FP_TO_SINT,  MVT::i16, MVT::f64, 2 },
    { ISD::FP_TO_UINT,  MVT::i16, MVT::f64, 2 },
    { ISD::FP_TO_SINT,  MVT::i32, MVT::f32, 2 },
    { ISD::FP_TO_UINT,  MVT::i32, MVT::f32, 2 },
    { ISD::FP_TO_SINT,  MVT::i32, MVT::f64, 2 },
    { ISD::FP_TO_UINT,  MVT::i32, MVT::f64, 2 },
    { ISD::FP_TO_SINT,  MVT::i64, MVT::f32, 10 },
    { ISD::FP_TO_UINT,  MVT::i64, MVT::f32, 10 },
    { ISD::FP_TO_SINT,  MVT::i64, MVT::f64, 10 },
    { ISD::FP_TO_UINT,  MVT::i64, MVT::f64, 10 }
  };
  if (SrcTy.isFloatingPoint() && ST->hasNEON()) {
    if (const auto *Entry = ConvertCostTableLookup(NEONFloatConversionTbl, ISD,
                                                   DstTy.getSimpleVT(),
                                                   SrcTy.getSimpleVT()))
      return AdjustCost(Entry->Cost);
  }

  // Scalar integer to float conversions.
  static const TypeConversionCostTblEntry NEONIntegerConversionTbl[] = {
    { ISD::SINT_TO_FP,  MVT::f32, MVT::i1, 2 },
    { ISD::UINT_TO_FP,  MVT::f32, MVT::i1, 2 },
    { ISD::SINT_TO_FP,  MVT::f64, MVT::i1, 2 },
    { ISD::UINT_TO_FP,  MVT::f64, MVT::i1, 2 },
    { ISD::SINT_TO_FP,  MVT::f32, MVT::i8, 2 },
    { ISD::UINT_TO_FP,  MVT::f32, MVT::i8, 2 },
    { ISD::SINT_TO_FP,  MVT::f64, MVT::i8, 2 },
    { ISD::UINT_TO_FP,  MVT::f64, MVT::i8, 2 },
    { ISD::SINT_TO_FP,  MVT::f32, MVT::i16, 2 },
    { ISD::UINT_TO_FP,  MVT::f32, MVT::i16, 2 },
    { ISD::SINT_TO_FP,  MVT::f64, MVT::i16, 2 },
    { ISD::UINT_TO_FP,  MVT::f64, MVT::i16, 2 },
    { ISD::SINT_TO_FP,  MVT::f32, MVT::i32, 2 },
    { ISD::UINT_TO_FP,  MVT::f32, MVT::i32, 2 },
    { ISD::SINT_TO_FP,  MVT::f64, MVT::i32, 2 },
    { ISD::UINT_TO_FP,  MVT::f64, MVT::i32, 2 },
    { ISD::SINT_TO_FP,  MVT::f32, MVT::i64, 10 },
    { ISD::UINT_TO_FP,  MVT::f32, MVT::i64, 10 },
    { ISD::SINT_TO_FP,  MVT::f64, MVT::i64, 10 },
    { ISD::UINT_TO_FP,  MVT::f64, MVT::i64, 10 }
  };

  if (SrcTy.isInteger() && ST->hasNEON()) {
    if (const auto *Entry = ConvertCostTableLookup(NEONIntegerConversionTbl,
                                                   ISD, DstTy.getSimpleVT(),
                                                   SrcTy.getSimpleVT()))
      return AdjustCost(Entry->Cost);
  }

  // MVE extend costs, taken from codegen tests. i8->i16 or i16->i32 is one
  // instruction, i8->i32 is two. i64 zexts are an VAND with a constant, sext
  // are linearised so take more.
  static const TypeConversionCostTblEntry MVEVectorConversionTbl[] = {
    { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
    { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
    { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
    { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
    { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i8, 10 },
    { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i8, 2 },
    { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
    { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
    { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i16, 10 },
    { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
    { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 8 },
    { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 2 },
  };

  if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
    if (const auto *Entry = ConvertCostTableLookup(MVEVectorConversionTbl,
                                                   ISD, DstTy.getSimpleVT(),
                                                   SrcTy.getSimpleVT()))
      return AdjustCost(Entry->Cost * ST->getMVEVectorCostFactor());
  }

  if (ISD == ISD::FP_ROUND || ISD == ISD::FP_EXTEND) {
    // As general rule, fp converts that were not matched above are scalarized
    // and cost 1 vcvt for each lane, so long as the instruction is available.
    // If not it will become a series of function calls.
    const int CallCost = getCallInstrCost(nullptr, Dst, {Src}, CostKind);
    int Lanes = 1;
    if (SrcTy.isFixedLengthVector())
      Lanes = SrcTy.getVectorNumElements();
    auto IsLegal = [this](EVT VT) {
      EVT EltVT = VT.getScalarType();
      return (EltVT == MVT::f32 && ST->hasVFP2Base()) ||
             (EltVT == MVT::f64 && ST->hasFP64()) ||
             (EltVT == MVT::f16 && ST->hasFullFP16());
    };

    if (IsLegal(SrcTy) && IsLegal(DstTy))
      return Lanes;
    else
      return Lanes * CallCost;
  }

  // Scalar integer conversion costs.
  static const TypeConversionCostTblEntry ARMIntegerConversionTbl[] = {
    // i16 -> i64 requires two dependent operations.
    { ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 },

    // Truncates on i64 are assumed to be free.
    { ISD::TRUNCATE,    MVT::i32, MVT::i64, 0 },
    { ISD::TRUNCATE,    MVT::i16, MVT::i64, 0 },
    { ISD::TRUNCATE,    MVT::i8,  MVT::i64, 0 },
    { ISD::TRUNCATE,    MVT::i1,  MVT::i64, 0 }
  };

  if (SrcTy.isInteger()) {
    if (const auto *Entry = ConvertCostTableLookup(ARMIntegerConversionTbl, ISD,
                                                   DstTy.getSimpleVT(),
                                                   SrcTy.getSimpleVT()))
      return AdjustCost(Entry->Cost);
  }

  int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
                     ? ST->getMVEVectorCostFactor()
                     : 1;
  return AdjustCost(
    BaseCost * BaseT::getCastInstrCost(Opcode, Dst, Src, CostKind, I));
}

int ARMTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
                                   unsigned Index) {
  // Penalize inserting into an D-subregister. We end up with a three times
  // lower estimated throughput on swift.
  if (ST->hasSlowLoadDSubregister() && Opcode == Instruction::InsertElement &&
      ValTy->isVectorTy() && ValTy->getScalarSizeInBits() <= 32)
    return 3;

  if (ST->hasNEON() && (Opcode == Instruction::InsertElement ||
                        Opcode == Instruction::ExtractElement)) {
    // Cross-class copies are expensive on many microarchitectures,
    // so assume they are expensive by default.
    if (cast<VectorType>(ValTy)->getElementType()->isIntegerTy())
      return 3;

    // Even if it's not a cross class copy, this likely leads to mixing
    // of NEON and VFP code and should be therefore penalized.
    if (ValTy->isVectorTy() &&
        ValTy->getScalarSizeInBits() <= 32)
      return std::max(BaseT::getVectorInstrCost(Opcode, ValTy, Index), 2U);
  }

  if (ST->hasMVEIntegerOps() && (Opcode == Instruction::InsertElement ||
                                 Opcode == Instruction::ExtractElement)) {
    // We say MVE moves costs at least the MVEVectorCostFactor, even though
    // they are scalar instructions. This helps prevent mixing scalar and
    // vector, to prevent vectorising where we end up just scalarising the
    // result anyway.
    return std::max(BaseT::getVectorInstrCost(Opcode, ValTy, Index),
                    ST->getMVEVectorCostFactor()) *
           cast<FixedVectorType>(ValTy)->getNumElements() / 2;
  }

  return BaseT::getVectorInstrCost(Opcode, ValTy, Index);
}

int ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
                                   TTI::TargetCostKind CostKind,
                                   const Instruction *I) {
  // TODO: Handle other cost kinds.
  if (CostKind != TTI::TCK_RecipThroughput)
    return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, CostKind, I);

  int ISD = TLI->InstructionOpcodeToISD(Opcode);
  // On NEON a vector select gets lowered to vbsl.
  if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
    // Lowering of some vector selects is currently far from perfect.
    static const TypeConversionCostTblEntry NEONVectorSelectTbl[] = {
      { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
      { ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
      { ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
    };

    EVT SelCondTy = TLI->getValueType(DL, CondTy);
    EVT SelValTy = TLI->getValueType(DL, ValTy);
    if (SelCondTy.isSimple() && SelValTy.isSimple()) {
      if (const auto *Entry = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
                                                     SelCondTy.getSimpleVT(),
                                                     SelValTy.getSimpleVT()))
        return Entry->Cost;
    }

    std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
    return LT.first;
  }

  int BaseCost = ST->hasMVEIntegerOps() && ValTy->isVectorTy()
                     ? ST->getMVEVectorCostFactor()
                     : 1;
  return BaseCost * BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, CostKind,
                                              I);
}

int ARMTTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
                                          const SCEV *Ptr) {
  // Address computations in vectorized code with non-consecutive addresses will
  // likely result in more instructions compared to scalar code where the
  // computation can more often be merged into the index mode. The resulting
  // extra micro-ops can significantly decrease throughput.
  unsigned NumVectorInstToHideOverhead = 10;
  int MaxMergeDistance = 64;

  if (ST->hasNEON()) {
    if (Ty->isVectorTy() && SE &&
        !BaseT::isConstantStridedAccessLessThan(SE, Ptr, MaxMergeDistance + 1))
      return NumVectorInstToHideOverhead;

    // In many cases the address computation is not merged into the instruction
    // addressing mode.
    return 1;
  }
  return BaseT::getAddressComputationCost(Ty, SE, Ptr);
}

bool ARMTTIImpl::isProfitableLSRChainElement(Instruction *I) {
  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
    // If a VCTP is part of a chain, it's already profitable and shouldn't be
    // optimized, else LSR may block tail-predication.
    switch (II->getIntrinsicID()) {
    case Intrinsic::arm_mve_vctp8:
    case Intrinsic::arm_mve_vctp16:
    case Intrinsic::arm_mve_vctp32:
    case Intrinsic::arm_mve_vctp64:
      return true;
    default:
      break;
    }
  }
  return false;
}

bool ARMTTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
  if (!EnableMaskedLoadStores || !ST->hasMVEIntegerOps())
    return false;

  if (auto *VecTy = dyn_cast<FixedVectorType>(DataTy)) {
    // Don't support v2i1 yet.
    if (VecTy->getNumElements() == 2)
      return false;

    // We don't support extending fp types.
     unsigned VecWidth = DataTy->getPrimitiveSizeInBits();
    if (VecWidth != 128 && VecTy->getElementType()->isFloatingPointTy())
      return false;
  }

  unsigned EltWidth = DataTy->getScalarSizeInBits();
  return (EltWidth == 32 && Alignment >= 4) ||
         (EltWidth == 16 && Alignment >= 2) || (EltWidth == 8);
}

bool ARMTTIImpl::isLegalMaskedGather(Type *Ty, Align Alignment) {
  if (!EnableMaskedGatherScatters || !ST->hasMVEIntegerOps())
    return false;

  // This method is called in 2 places:
  //  - from the vectorizer with a scalar type, in which case we need to get
  //  this as good as we can with the limited info we have (and rely on the cost
  //  model for the rest).
  //  - from the masked intrinsic lowering pass with the actual vector type.
  // For MVE, we have a custom lowering pass that will already have custom
  // legalised any gathers that we can to MVE intrinsics, and want to expand all
  // the rest. The pass runs before the masked intrinsic lowering pass, so if we
  // are here, we know we want to expand.
  if (isa<VectorType>(Ty))
    return false;

  unsigned EltWidth = Ty->getScalarSizeInBits();
  return ((EltWidth == 32 && Alignment >= 4) ||
          (EltWidth == 16 && Alignment >= 2) || EltWidth == 8);
}

int ARMTTIImpl::getMemcpyCost(const Instruction *I) {
  const MemCpyInst *MI = dyn_cast<MemCpyInst>(I);
  assert(MI && "MemcpyInst expected");
  ConstantInt *C = dyn_cast<ConstantInt>(MI->getLength());

  // To model the cost of a library call, we assume 1 for the call, and
  // 3 for the argument setup.
  const unsigned LibCallCost = 4;

  // If 'size' is not a constant, a library call will be generated.
  if (!C)
    return LibCallCost;

  const unsigned Size = C->getValue().getZExtValue();
  const Align DstAlign = *MI->getDestAlign();
  const Align SrcAlign = *MI->getSourceAlign();
  const Function *F = I->getParent()->getParent();
  const unsigned Limit = TLI->getMaxStoresPerMemmove(F->hasMinSize());
  std::vector<EVT> MemOps;

  // MemOps will be poplulated with a list of data types that needs to be
  // loaded and stored. That's why we multiply the number of elements by 2 to
  // get the cost for this memcpy.
  if (getTLI()->findOptimalMemOpLowering(
          MemOps, Limit,
          MemOp::Copy(Size, /*DstAlignCanChange*/ false, DstAlign, SrcAlign,
                      /*IsVolatile*/ true),
          MI->getDestAddressSpace(), MI->getSourceAddressSpace(),
          F->getAttributes()))
    return MemOps.size() * 2;

  // If we can't find an optimal memop lowering, return the default cost
  return LibCallCost;
}

int ARMTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, VectorType *Tp,
                               int Index, VectorType *SubTp) {
  if (ST->hasNEON()) {
    if (Kind == TTI::SK_Broadcast) {
      static const CostTblEntry NEONDupTbl[] = {
          // VDUP handles these cases.
          {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v4i16, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v8i8, 1},

          {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v16i8, 1}};

      std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);

      if (const auto *Entry =
              CostTableLookup(NEONDupTbl, ISD::VECTOR_SHUFFLE, LT.second))
        return LT.first * Entry->Cost;
    }
    if (Kind == TTI::SK_Reverse) {
      static const CostTblEntry NEONShuffleTbl[] = {
          // Reverse shuffle cost one instruction if we are shuffling within a
          // double word (vrev) or two if we shuffle a quad word (vrev, vext).
          {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v4i16, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v8i8, 1},

          {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
          {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
          {ISD::VECTOR_SHUFFLE, MVT::v8i16, 2},
          {ISD::VECTOR_SHUFFLE, MVT::v16i8, 2}};

      std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);

      if (const auto *Entry =
              CostTableLookup(NEONShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second))
        return LT.first * Entry->Cost;
    }
    if (Kind == TTI::SK_Select) {
      static const CostTblEntry NEONSelShuffleTbl[] = {
          // Select shuffle cost table for ARM. Cost is the number of
          // instructions
          // required to create the shuffled vector.

          {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},

          {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
          {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
          {ISD::VECTOR_SHUFFLE, MVT::v4i16, 2},

          {ISD::VECTOR_SHUFFLE, MVT::v8i16, 16},

          {ISD::VECTOR_SHUFFLE, MVT::v16i8, 32}};

      std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
      if (const auto *Entry = CostTableLookup(NEONSelShuffleTbl,
                                              ISD::VECTOR_SHUFFLE, LT.second))
        return LT.first * Entry->Cost;
    }
  }
  if (ST->hasMVEIntegerOps()) {
    if (Kind == TTI::SK_Broadcast) {
      static const CostTblEntry MVEDupTbl[] = {
          // VDUP handles these cases.
          {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v16i8, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
          {ISD::VECTOR_SHUFFLE, MVT::v8f16, 1}};

      std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);

      if (const auto *Entry = CostTableLookup(MVEDupTbl, ISD::VECTOR_SHUFFLE,
                                              LT.second))
        return LT.first * Entry->Cost * ST->getMVEVectorCostFactor();
    }
  }
  int BaseCost = ST->hasMVEIntegerOps() && Tp->isVectorTy()
                     ? ST->getMVEVectorCostFactor()
                     : 1;
  return BaseCost * BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
}

int ARMTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                       TTI::TargetCostKind CostKind,
                                       TTI::OperandValueKind Op1Info,
                                       TTI::OperandValueKind Op2Info,
                                       TTI::OperandValueProperties Opd1PropInfo,
                                       TTI::OperandValueProperties Opd2PropInfo,
                                       ArrayRef<const Value *> Args,
                                       const Instruction *CxtI) {
  // TODO: Handle more cost kinds.
  if (CostKind != TTI::TCK_RecipThroughput)
    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                         Op2Info, Opd1PropInfo,
                                         Opd2PropInfo, Args, CxtI);

  int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
  std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);

  if (ST->hasNEON()) {
    const unsigned FunctionCallDivCost = 20;
    const unsigned ReciprocalDivCost = 10;
    static const CostTblEntry CostTbl[] = {
      // Division.
      // These costs are somewhat random. Choose a cost of 20 to indicate that
      // vectorizing devision (added function call) is going to be very expensive.
      // Double registers types.
      { ISD::SDIV, MVT::v1i64, 1 * FunctionCallDivCost},
      { ISD::UDIV, MVT::v1i64, 1 * FunctionCallDivCost},
      { ISD::SREM, MVT::v1i64, 1 * FunctionCallDivCost},
      { ISD::UREM, MVT::v1i64, 1 * FunctionCallDivCost},
      { ISD::SDIV, MVT::v2i32, 2 * FunctionCallDivCost},
      { ISD::UDIV, MVT::v2i32, 2 * FunctionCallDivCost},
      { ISD::SREM, MVT::v2i32, 2 * FunctionCallDivCost},
      { ISD::UREM, MVT::v2i32, 2 * FunctionCallDivCost},
      { ISD::SDIV, MVT::v4i16,     ReciprocalDivCost},
      { ISD::UDIV, MVT::v4i16,     ReciprocalDivCost},
      { ISD::SREM, MVT::v4i16, 4 * FunctionCallDivCost},
      { ISD::UREM, MVT::v4i16, 4 * FunctionCallDivCost},
      { ISD::SDIV, MVT::v8i8,      ReciprocalDivCost},
      { ISD::UDIV, MVT::v8i8,      ReciprocalDivCost},
      { ISD::SREM, MVT::v8i8,  8 * FunctionCallDivCost},
      { ISD::UREM, MVT::v8i8,  8 * FunctionCallDivCost},
      // Quad register types.
      { ISD::SDIV, MVT::v2i64, 2 * FunctionCallDivCost},
      { ISD::UDIV, MVT::v2i64, 2 * FunctionCallDivCost},
      { ISD::SREM, MVT::v2i64, 2 * FunctionCallDivCost},
      { ISD::UREM, MVT::v2i64, 2 * FunctionCallDivCost},
      { ISD::SDIV, MVT::v4i32, 4 * FunctionCallDivCost},
      { ISD::UDIV, MVT::v4i32, 4 * FunctionCallDivCost},
      { ISD::SREM, MVT::v4i32, 4 * FunctionCallDivCost},
      { ISD::UREM, MVT::v4i32, 4 * FunctionCallDivCost},
      { ISD::SDIV, MVT::v8i16, 8 * FunctionCallDivCost},
      { ISD::UDIV, MVT::v8i16, 8 * FunctionCallDivCost},
      { ISD::SREM, MVT::v8i16, 8 * FunctionCallDivCost},
      { ISD::UREM, MVT::v8i16, 8 * FunctionCallDivCost},
      { ISD::SDIV, MVT::v16i8, 16 * FunctionCallDivCost},
      { ISD::UDIV, MVT::v16i8, 16 * FunctionCallDivCost},
      { ISD::SREM, MVT::v16i8, 16 * FunctionCallDivCost},
      { ISD::UREM, MVT::v16i8, 16 * FunctionCallDivCost},
      // Multiplication.
    };

    if (const auto *Entry = CostTableLookup(CostTbl, ISDOpcode, LT.second))
      return LT.first * Entry->Cost;

    int Cost = BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                             Op2Info,
                                             Opd1PropInfo, Opd2PropInfo);

    // This is somewhat of a hack. The problem that we are facing is that SROA
    // creates a sequence of shift, and, or instructions to construct values.
    // These sequences are recognized by the ISel and have zero-cost. Not so for
    // the vectorized code. Because we have support for v2i64 but not i64 those
    // sequences look particularly beneficial to vectorize.
    // To work around this we increase the cost of v2i64 operations to make them
    // seem less beneficial.
    if (LT.second == MVT::v2i64 &&
        Op2Info == TargetTransformInfo::OK_UniformConstantValue)
      Cost += 4;

    return Cost;
  }

  // If this operation is a shift on arm/thumb2, it might well be folded into
  // the following instruction, hence having a cost of 0.
  auto LooksLikeAFreeShift = [&]() {
    if (ST->isThumb1Only() || Ty->isVectorTy())
      return false;

    if (!CxtI || !CxtI->hasOneUse() || !CxtI->isShift())
      return false;
    if (Op2Info != TargetTransformInfo::OK_UniformConstantValue)
      return false;

    // Folded into a ADC/ADD/AND/BIC/CMP/EOR/MVN/ORR/ORN/RSB/SBC/SUB
    switch (cast<Instruction>(CxtI->user_back())->getOpcode()) {
    case Instruction::Add:
    case Instruction::Sub:
    case Instruction::And:
    case Instruction::Xor:
    case Instruction::Or:
    case Instruction::ICmp:
      return true;
    default:
      return false;
    }
  };
  if (LooksLikeAFreeShift())
    return 0;

  int BaseCost = ST->hasMVEIntegerOps() && Ty->isVectorTy()
                     ? ST->getMVEVectorCostFactor()
                     : 1;

  // The rest of this mostly follows what is done in BaseT::getArithmeticInstrCost,
  // without treating floats as more expensive that scalars or increasing the
  // costs for custom operations. The results is also multiplied by the
  // MVEVectorCostFactor where appropriate.
  if (TLI->isOperationLegalOrCustomOrPromote(ISDOpcode, LT.second))
    return LT.first * BaseCost;

  // Else this is expand, assume that we need to scalarize this op.
  if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
    unsigned Num = VTy->getNumElements();
    unsigned Cost = getArithmeticInstrCost(Opcode, Ty->getScalarType(),
                                           CostKind);
    // Return the cost of multiple scalar invocation plus the cost of
    // inserting and extracting the values.
    return BaseT::getScalarizationOverhead(VTy, Args) + Num * Cost;
  }

  return BaseCost;
}

int ARMTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
                                MaybeAlign Alignment, unsigned AddressSpace,
                                TTI::TargetCostKind CostKind,
                                const Instruction *I) {
  // TODO: Handle other cost kinds.
  if (CostKind != TTI::TCK_RecipThroughput)
    return 1;

  // Type legalization can't handle structs
  if (TLI->getValueType(DL, Src, true) == MVT::Other)
    return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
                                  CostKind);

  if (ST->hasNEON() && Src->isVectorTy() &&
      (Alignment && *Alignment != Align(16)) &&
      cast<VectorType>(Src)->getElementType()->isDoubleTy()) {
    // Unaligned loads/stores are extremely inefficient.
    // We need 4 uops for vst.1/vld.1 vs 1uop for vldr/vstr.
    std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
    return LT.first * 4;
  }

  // MVE can optimize a fpext(load(4xhalf)) using an extending integer load.
  // Same for stores.
  if (ST->hasMVEFloatOps() && isa<FixedVectorType>(Src) && I &&
      ((Opcode == Instruction::Load && I->hasOneUse() &&
        isa<FPExtInst>(*I->user_begin())) ||
       (Opcode == Instruction::Store && isa<FPTruncInst>(I->getOperand(0))))) {
    FixedVectorType *SrcVTy = cast<FixedVectorType>(Src);
    Type *DstTy =
        Opcode == Instruction::Load
            ? (*I->user_begin())->getType()
            : cast<Instruction>(I->getOperand(0))->getOperand(0)->getType();
    if (SrcVTy->getNumElements() == 4 && SrcVTy->getScalarType()->isHalfTy() &&
        DstTy->getScalarType()->isFloatTy())
      return ST->getMVEVectorCostFactor();
  }

  int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
                     ? ST->getMVEVectorCostFactor()
                     : 1;
  return BaseCost * BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
                                           CostKind, I);
}

int ARMTTIImpl::getInterleavedMemoryOpCost(
    unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
    Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
    bool UseMaskForCond, bool UseMaskForGaps) {
  assert(Factor >= 2 && "Invalid interleave factor");
  assert(isa<VectorType>(VecTy) && "Expect a vector type");

  // vldN/vstN doesn't support vector types of i64/f64 element.
  bool EltIs64Bits = DL.getTypeSizeInBits(VecTy->getScalarType()) == 64;

  if (Factor <= TLI->getMaxSupportedInterleaveFactor() && !EltIs64Bits &&
      !UseMaskForCond && !UseMaskForGaps) {
    unsigned NumElts = cast<FixedVectorType>(VecTy)->getNumElements();
    auto *SubVecTy =
        FixedVectorType::get(VecTy->getScalarType(), NumElts / Factor);

    // vldN/vstN only support legal vector types of size 64 or 128 in bits.
    // Accesses having vector types that are a multiple of 128 bits can be
    // matched to more than one vldN/vstN instruction.
    int BaseCost = ST->hasMVEIntegerOps() ? ST->getMVEVectorCostFactor() : 1;
    if (NumElts % Factor == 0 &&
        TLI->isLegalInterleavedAccessType(Factor, SubVecTy, DL))
      return Factor * BaseCost * TLI->getNumInterleavedAccesses(SubVecTy, DL);

    // Some smaller than legal interleaved patterns are cheap as we can make
    // use of the vmovn or vrev patterns to interleave a standard load. This is
    // true for v4i8, v8i8 and v4i16 at least (but not for v4f16 as it is
    // promoted differently). The cost of 2 here is then a load and vrev or
    // vmovn.
    if (ST->hasMVEIntegerOps() && Factor == 2 && NumElts / Factor > 2 &&
        VecTy->isIntOrIntVectorTy() && DL.getTypeSizeInBits(SubVecTy) <= 64)
      return 2 * BaseCost;
  }

  return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                           Alignment, AddressSpace, CostKind,
                                           UseMaskForCond, UseMaskForGaps);
}

unsigned ARMTTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
                                            const Value *Ptr, bool VariableMask,
                                            Align Alignment,
                                            TTI::TargetCostKind CostKind,
                                            const Instruction *I) {
  using namespace PatternMatch;
  if (!ST->hasMVEIntegerOps() || !EnableMaskedGatherScatters)
    return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
                                         Alignment, CostKind, I);

  assert(DataTy->isVectorTy() && "Can't do gather/scatters on scalar!");
  auto *VTy = cast<FixedVectorType>(DataTy);

  // TODO: Splitting, once we do that.

  unsigned NumElems = VTy->getNumElements();
  unsigned EltSize = VTy->getScalarSizeInBits();
  std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, DataTy);

  // For now, it is assumed that for the MVE gather instructions the loads are
  // all effectively serialised. This means the cost is the scalar cost
  // multiplied by the number of elements being loaded. This is possibly very
  // conservative, but even so we still end up vectorising loops because the
  // cost per iteration for many loops is lower than for scalar loops.
  unsigned VectorCost = NumElems * LT.first;
  // The scalarization cost should be a lot higher. We use the number of vector
  // elements plus the scalarization overhead.
  unsigned ScalarCost =
      NumElems * LT.first + BaseT::getScalarizationOverhead(VTy, {});

  if (Alignment < EltSize / 8)
    return ScalarCost;

  unsigned ExtSize = EltSize;
  // Check whether there's a single user that asks for an extended type
  if (I != nullptr) {
    // Dependent of the caller of this function, a gather instruction will
    // either have opcode Instruction::Load or be a call to the masked_gather
    // intrinsic
    if ((I->getOpcode() == Instruction::Load ||
         match(I, m_Intrinsic<Intrinsic::masked_gather>())) &&
        I->hasOneUse()) {
      const User *Us = *I->users().begin();
      if (isa<ZExtInst>(Us) || isa<SExtInst>(Us)) {
        // only allow valid type combinations
        unsigned TypeSize =
            cast<Instruction>(Us)->getType()->getScalarSizeInBits();
        if (((TypeSize == 32 && (EltSize == 8 || EltSize == 16)) ||
             (TypeSize == 16 && EltSize == 8)) &&
            TypeSize * NumElems == 128) {
          ExtSize = TypeSize;
        }
      }
    }
    // Check whether the input data needs to be truncated
    TruncInst *T;
    if ((I->getOpcode() == Instruction::Store ||
         match(I, m_Intrinsic<Intrinsic::masked_scatter>())) &&
        (T = dyn_cast<TruncInst>(I->getOperand(0)))) {
      // Only allow valid type combinations
      unsigned TypeSize = T->getOperand(0)->getType()->getScalarSizeInBits();
      if (((EltSize == 16 && TypeSize == 32) ||
           (EltSize == 8 && (TypeSize == 32 || TypeSize == 16))) &&
          TypeSize * NumElems == 128)
        ExtSize = TypeSize;
    }
  }

  if (ExtSize * NumElems != 128 || NumElems < 4)
    return ScalarCost;

  // Any (aligned) i32 gather will not need to be scalarised.
  if (ExtSize == 32)
    return VectorCost;
  // For smaller types, we need to ensure that the gep's inputs are correctly
  // extended from a small enough value. Other sizes (including i64) are
  // scalarized for now.
  if (ExtSize != 8 && ExtSize != 16)
    return ScalarCost;

  if (const auto *BC = dyn_cast<BitCastInst>(Ptr))
    Ptr = BC->getOperand(0);
  if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
    if (GEP->getNumOperands() != 2)
      return ScalarCost;
    unsigned Scale = DL.getTypeAllocSize(GEP->getResultElementType());
    // Scale needs to be correct (which is only relevant for i16s).
    if (Scale != 1 && Scale * 8 != ExtSize)
      return ScalarCost;
    // And we need to zext (not sext) the indexes from a small enough type.
    if (const auto *ZExt = dyn_cast<ZExtInst>(GEP->getOperand(1))) {
      if (ZExt->getOperand(0)->getType()->getScalarSizeInBits() <= ExtSize)
        return VectorCost;
    }
    return ScalarCost;
  }
  return ScalarCost;
}

bool ARMTTIImpl::isLoweredToCall(const Function *F) {
  if (!F->isIntrinsic())
    BaseT::isLoweredToCall(F);

  // Assume all Arm-specific intrinsics map to an instruction.
  if (F->getName().startswith("llvm.arm"))
    return false;

  switch (F->getIntrinsicID()) {
  default: break;
  case Intrinsic::powi:
  case Intrinsic::sin:
  case Intrinsic::cos:
  case Intrinsic::pow:
  case Intrinsic::log:
  case Intrinsic::log10:
  case Intrinsic::log2:
  case Intrinsic::exp:
  case Intrinsic::exp2:
    return true;
  case Intrinsic::sqrt:
  case Intrinsic::fabs:
  case Intrinsic::copysign:
  case Intrinsic::floor:
  case Intrinsic::ceil:
  case Intrinsic::trunc:
  case Intrinsic::rint:
  case Intrinsic::nearbyint:
  case Intrinsic::round:
  case Intrinsic::canonicalize:
  case Intrinsic::lround:
  case Intrinsic::llround:
  case Intrinsic::lrint:
  case Intrinsic::llrint:
    if (F->getReturnType()->isDoubleTy() && !ST->hasFP64())
      return true;
    if (F->getReturnType()->isHalfTy() && !ST->hasFullFP16())
      return true;
    // Some operations can be handled by vector instructions and assume
    // unsupported vectors will be expanded into supported scalar ones.
    // TODO Handle scalar operations properly.
    return !ST->hasFPARMv8Base() && !ST->hasVFP2Base();
  case Intrinsic::masked_store:
  case Intrinsic::masked_load:
  case Intrinsic::masked_gather:
  case Intrinsic::masked_scatter:
    return !ST->hasMVEIntegerOps();
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::uadd_with_overflow:
  case Intrinsic::ssub_with_overflow:
  case Intrinsic::usub_with_overflow:
  case Intrinsic::sadd_sat:
  case Intrinsic::uadd_sat:
  case Intrinsic::ssub_sat:
  case Intrinsic::usub_sat:
    return false;
  }

  return BaseT::isLoweredToCall(F);
}

bool ARMTTIImpl::isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
                                          AssumptionCache &AC,
                                          TargetLibraryInfo *LibInfo,
                                          HardwareLoopInfo &HWLoopInfo) {
  // Low-overhead branches are only supported in the 'low-overhead branch'
  // extension of v8.1-m.
  if (!ST->hasLOB() || DisableLowOverheadLoops) {
    LLVM_DEBUG(dbgs() << "ARMHWLoops: Disabled\n");
    return false;
  }

  if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
    LLVM_DEBUG(dbgs() << "ARMHWLoops: No BETC\n");
    return false;
  }

  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
    LLVM_DEBUG(dbgs() << "ARMHWLoops: Uncomputable BETC\n");
    return false;
  }

  const SCEV *TripCountSCEV =
    SE.getAddExpr(BackedgeTakenCount,
                  SE.getOne(BackedgeTakenCount->getType()));

  // We need to store the trip count in LR, a 32-bit register.
  if (SE.getUnsignedRangeMax(TripCountSCEV).getBitWidth() > 32) {
    LLVM_DEBUG(dbgs() << "ARMHWLoops: Trip count does not fit into 32bits\n");
    return false;
  }

  // Making a call will trash LR and clear LO_BRANCH_INFO, so there's little
  // point in generating a hardware loop if that's going to happen.
  auto MaybeCall = [this](Instruction &I) {
    const ARMTargetLowering *TLI = getTLI();
    unsigned ISD = TLI->InstructionOpcodeToISD(I.getOpcode());
    EVT VT = TLI->getValueType(DL, I.getType(), true);
    if (TLI->getOperationAction(ISD, VT) == TargetLowering::LibCall)
      return true;

    // Check if an intrinsic will be lowered to a call and assume that any
    // other CallInst will generate a bl.
    if (auto *Call = dyn_cast<CallInst>(&I)) {
      if (isa<IntrinsicInst>(Call)) {
        if (const Function *F = Call->getCalledFunction())
          return isLoweredToCall(F);
      }
      return true;
    }

    // FPv5 provides conversions between integer, double-precision,
    // single-precision, and half-precision formats.
    switch (I.getOpcode()) {
    default:
      break;
    case Instruction::FPToSI:
    case Instruction::FPToUI:
    case Instruction::SIToFP:
    case Instruction::UIToFP:
    case Instruction::FPTrunc:
    case Instruction::FPExt:
      return !ST->hasFPARMv8Base();
    }

    // FIXME: Unfortunately the approach of checking the Operation Action does
    // not catch all cases of Legalization that use library calls. Our
    // Legalization step categorizes some transformations into library calls as
    // Custom, Expand or even Legal when doing type legalization. So for now
    // we have to special case for instance the SDIV of 64bit integers and the
    // use of floating point emulation.
    if (VT.isInteger() && VT.getSizeInBits() >= 64) {
      switch (ISD) {
      default:
        break;
      case ISD::SDIV:
      case ISD::UDIV:
      case ISD::SREM:
      case ISD::UREM:
      case ISD::SDIVREM:
      case ISD::UDIVREM:
        return true;
      }
    }

    // Assume all other non-float operations are supported.
    if (!VT.isFloatingPoint())
      return false;

    // We'll need a library call to handle most floats when using soft.
    if (TLI->useSoftFloat()) {
      switch (I.getOpcode()) {
      default:
        return true;
      case Instruction::Alloca:
      case Instruction::Load:
      case Instruction::Store:
      case Instruction::Select:
      case Instruction::PHI:
        return false;
      }
    }

    // We'll need a libcall to perform double precision operations on a single
    // precision only FPU.
    if (I.getType()->isDoubleTy() && !ST->hasFP64())
      return true;

    // Likewise for half precision arithmetic.
    if (I.getType()->isHalfTy() && !ST->hasFullFP16())
      return true;

    return false;
  };

  auto IsHardwareLoopIntrinsic = [](Instruction &I) {
    if (auto *Call = dyn_cast<IntrinsicInst>(&I)) {
      switch (Call->getIntrinsicID()) {
      default:
        break;
      case Intrinsic::set_loop_iterations:
      case Intrinsic::test_set_loop_iterations:
      case Intrinsic::loop_decrement:
      case Intrinsic::loop_decrement_reg:
        return true;
      }
    }
    return false;
  };

  // Scan the instructions to see if there's any that we know will turn into a
  // call or if this loop is already a low-overhead loop.
  auto ScanLoop = [&](Loop *L) {
    for (auto *BB : L->getBlocks()) {
      for (auto &I : *BB) {
        if (MaybeCall(I) || IsHardwareLoopIntrinsic(I)) {
          LLVM_DEBUG(dbgs() << "ARMHWLoops: Bad instruction: " << I << "\n");
          return false;
        }
      }
    }
    return true;
  };

  // Visit inner loops.
  for (auto Inner : *L)
    if (!ScanLoop(Inner))
      return false;

  if (!ScanLoop(L))
    return false;

  // TODO: Check whether the trip count calculation is expensive. If L is the
  // inner loop but we know it has a low trip count, calculating that trip
  // count (in the parent loop) may be detrimental.

  LLVMContext &C = L->getHeader()->getContext();
  HWLoopInfo.CounterInReg = true;
  HWLoopInfo.IsNestingLegal = false;
  HWLoopInfo.PerformEntryTest = true;
  HWLoopInfo.CountType = Type::getInt32Ty(C);
  HWLoopInfo.LoopDecrement = ConstantInt::get(HWLoopInfo.CountType, 1);
  return true;
}

static bool canTailPredicateInstruction(Instruction &I, int &ICmpCount) {
  // We don't allow icmp's, and because we only look at single block loops,
  // we simply count the icmps, i.e. there should only be 1 for the backedge.
  if (isa<ICmpInst>(&I) && ++ICmpCount > 1)
    return false;

  if (isa<FCmpInst>(&I))
    return false;

  // We could allow extending/narrowing FP loads/stores, but codegen is
  // too inefficient so reject this for now.
  if (isa<FPExtInst>(&I) || isa<FPTruncInst>(&I))
    return false;

  // Extends have to be extending-loads
  if (isa<SExtInst>(&I) || isa<ZExtInst>(&I) )
    if (!I.getOperand(0)->hasOneUse() || !isa<LoadInst>(I.getOperand(0)))
      return false;

  // Truncs have to be narrowing-stores
  if (isa<TruncInst>(&I) )
    if (!I.hasOneUse() || !isa<StoreInst>(*I.user_begin()))
      return false;

  return true;
}

// To set up a tail-predicated loop, we need to know the total number of
// elements processed by that loop. Thus, we need to determine the element
// size and:
// 1) it should be uniform for all operations in the vector loop, so we
//    e.g. don't want any widening/narrowing operations.
// 2) it should be smaller than i64s because we don't have vector operations
//    that work on i64s.
// 3) we don't want elements to be reversed or shuffled, to make sure the
//    tail-predication masks/predicates the right lanes.
//
static bool canTailPredicateLoop(Loop *L, LoopInfo *LI, ScalarEvolution &SE,
                                 const DataLayout &DL,
                                 const LoopAccessInfo *LAI) {
  LLVM_DEBUG(dbgs() << "Tail-predication: checking allowed instructions\n");

  // If there are live-out values, it is probably a reduction, which needs a
  // final reduction step after the loop. MVE has a VADDV instruction to reduce
  // integer vectors, but doesn't have an equivalent one for float vectors. A
  // live-out value that is not recognised as a reduction will result in the
  // tail-predicated loop to be reverted to a non-predicated loop and this is
  // very expensive, i.e. it has a significant performance impact. So, in this
  // case it's better not to tail-predicate the loop, which is what we check
  // here. Thus, we allow only 1 live-out value, which has to be an integer
  // reduction, which matches the loops supported by ARMLowOverheadLoops.
  // It is important to keep ARMLowOverheadLoops and canTailPredicateLoop in
  // sync with each other.
  SmallVector< Instruction *, 8 > LiveOuts;
  LiveOuts = llvm::findDefsUsedOutsideOfLoop(L);
  bool IntReductionsDisabled =
      EnableTailPredication == TailPredication::EnabledNoReductions ||
      EnableTailPredication == TailPredication::ForceEnabledNoReductions;

  for (auto *I : LiveOuts) {
    if (!I->getType()->isIntegerTy()) {
      LLVM_DEBUG(dbgs() << "Don't tail-predicate loop with non-integer "
                           "live-out value\n");
      return false;
    }
    if (I->getOpcode() != Instruction::Add) {
      LLVM_DEBUG(dbgs() << "Only add reductions supported\n");
      return false;
    }
    if (IntReductionsDisabled) {
      LLVM_DEBUG(dbgs() << "Integer add reductions not enabled\n");
      return false;
    }
  }

  // Next, check that all instructions can be tail-predicated.
  PredicatedScalarEvolution PSE = LAI->getPSE();
  SmallVector<Instruction *, 16> LoadStores;
  int ICmpCount = 0;
  int Stride = 0;

  for (BasicBlock *BB : L->blocks()) {
    for (Instruction &I : BB->instructionsWithoutDebug()) {
      if (isa<PHINode>(&I))
        continue;
      if (!canTailPredicateInstruction(I, ICmpCount)) {
        LLVM_DEBUG(dbgs() << "Instruction not allowed: "; I.dump());
        return false;
      }

      Type *T  = I.getType();
      if (T->isPointerTy())
        T = T->getPointerElementType();

      if (T->getScalarSizeInBits() > 32) {
        LLVM_DEBUG(dbgs() << "Unsupported Type: "; T->dump());
        return false;
      }

      if (isa<StoreInst>(I) || isa<LoadInst>(I)) {
        Value *Ptr = isa<LoadInst>(I) ? I.getOperand(0) : I.getOperand(1);
        int64_t NextStride = getPtrStride(PSE, Ptr, L);
        // TODO: for now only allow consecutive strides of 1. We could support
        // other strides as long as it is uniform, but let's keep it simple for
        // now.
        if (Stride == 0 && NextStride == 1) {
          Stride = NextStride;
          continue;
        }
        if (Stride != NextStride) {
          LLVM_DEBUG(dbgs() << "Different strides found, can't "
                               "tail-predicate\n.");
          return false;
        }
      }
    }
  }

  LLVM_DEBUG(dbgs() << "tail-predication: all instructions allowed!\n");
  return true;
}

bool ARMTTIImpl::preferPredicateOverEpilogue(Loop *L, LoopInfo *LI,
                                             ScalarEvolution &SE,
                                             AssumptionCache &AC,
                                             TargetLibraryInfo *TLI,
                                             DominatorTree *DT,
                                             const LoopAccessInfo *LAI) {
  if (!EnableTailPredication) {
    LLVM_DEBUG(dbgs() << "Tail-predication not enabled.\n");
    return false;
  }

  // Creating a predicated vector loop is the first step for generating a
  // tail-predicated hardware loop, for which we need the MVE masked
  // load/stores instructions:
  if (!ST->hasMVEIntegerOps())
    return false;

  // For now, restrict this to single block loops.
  if (L->getNumBlocks() > 1) {
    LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: not a single block "
                         "loop.\n");
    return false;
  }

  assert(L->empty() && "preferPredicateOverEpilogue: inner-loop expected");

  HardwareLoopInfo HWLoopInfo(L);
  if (!HWLoopInfo.canAnalyze(*LI)) {
    LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
                         "analyzable.\n");
    return false;
  }

  // This checks if we have the low-overhead branch architecture
  // extension, and if we will create a hardware-loop:
  if (!isHardwareLoopProfitable(L, SE, AC, TLI, HWLoopInfo)) {
    LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
                         "profitable.\n");
    return false;
  }

  if (!HWLoopInfo.isHardwareLoopCandidate(SE, *LI, *DT)) {
    LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
                         "a candidate.\n");
    return false;
  }

  return canTailPredicateLoop(L, LI, SE, DL, LAI);
}

bool ARMTTIImpl::emitGetActiveLaneMask() const {
  if (!ST->hasMVEIntegerOps() || !EnableTailPredication)
    return false;

  // Intrinsic @llvm.get.active.lane.mask is supported.
  // It is used in the MVETailPredication pass, which requires the number of
  // elements processed by this vector loop to setup the tail-predicated
  // loop.
  return true;
}
void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
                                         TTI::UnrollingPreferences &UP) {
  // Only currently enable these preferences for M-Class cores.
  if (!ST->isMClass())
    return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP);

  // Disable loop unrolling for Oz and Os.
  UP.OptSizeThreshold = 0;
  UP.PartialOptSizeThreshold = 0;
  if (L->getHeader()->getParent()->hasOptSize())
    return;

  // Only enable on Thumb-2 targets.
  if (!ST->isThumb2())
    return;

  SmallVector<BasicBlock*, 4> ExitingBlocks;
  L->getExitingBlocks(ExitingBlocks);
  LLVM_DEBUG(dbgs() << "Loop has:\n"
                    << "Blocks: " << L->getNumBlocks() << "\n"
                    << "Exit blocks: " << ExitingBlocks.size() << "\n");

  // Only allow another exit other than the latch. This acts as an early exit
  // as it mirrors the profitability calculation of the runtime unroller.
  if (ExitingBlocks.size() > 2)
    return;

  // Limit the CFG of the loop body for targets with a branch predictor.
  // Allowing 4 blocks permits if-then-else diamonds in the body.
  if (ST->hasBranchPredictor() && L->getNumBlocks() > 4)
    return;

  // Scan the loop: don't unroll loops with calls as this could prevent
  // inlining.
  unsigned Cost = 0;
  for (auto *BB : L->getBlocks()) {
    for (auto &I : *BB) {
      // Don't unroll vectorised loop. MVE does not benefit from it as much as
      // scalar code.
      if (I.getType()->isVectorTy())
        return;

      if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
        if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
          if (!isLoweredToCall(F))
            continue;
        }
        return;
      }

      SmallVector<const Value*, 4> Operands(I.value_op_begin(),
                                            I.value_op_end());
      Cost += getUserCost(&I, Operands, TargetTransformInfo::TCK_CodeSize);
    }
  }

  LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n");

  UP.Partial = true;
  UP.Runtime = true;
  UP.UpperBound = true;
  UP.UnrollRemainder = true;
  UP.DefaultUnrollRuntimeCount = 4;
  UP.UnrollAndJam = true;
  UP.UnrollAndJamInnerLoopThreshold = 60;

  // Force unrolling small loops can be very useful because of the branch
  // taken cost of the backedge.
  if (Cost < 12)
    UP.Force = true;
}

void ARMTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
                                       TTI::PeelingPreferences &PP) {
  BaseT::getPeelingPreferences(L, SE, PP);
}

bool ARMTTIImpl::useReductionIntrinsic(unsigned Opcode, Type *Ty,
                                       TTI::ReductionFlags Flags) const {
  return ST->hasMVEIntegerOps();
}