llvm.org GIT mirror llvm / release_26 lib / CodeGen / SelectionDAG / FastISel.cpp
release_26

Tree @release_26 (Download .tar.gz)

FastISel.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
///===-- FastISel.cpp - Implementation of the FastISel class --------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains the implementation of the FastISel class.
//
// "Fast" instruction selection is designed to emit very poor code quickly.
// Also, it is not designed to be able to do much lowering, so most illegal
// types (e.g. i64 on 32-bit targets) and operations are not supported.  It is
// also not intended to be able to do much optimization, except in a few cases
// where doing optimizations reduces overall compile time.  For example, folding
// constants into immediate fields is often done, because it's cheap and it
// reduces the number of instructions later phases have to examine.
//
// "Fast" instruction selection is able to fail gracefully and transfer
// control to the SelectionDAG selector for operations that it doesn't
// support.  In many cases, this allows us to avoid duplicating a lot of
// the complicated lowering logic that SelectionDAG currently has.
//
// The intended use for "fast" instruction selection is "-O0" mode
// compilation, where the quality of the generated code is irrelevant when
// weighed against the speed at which the code can be generated.  Also,
// at -O0, the LLVM optimizers are not running, and this makes the
// compile time of codegen a much higher portion of the overall compile
// time.  Despite its limitations, "fast" instruction selection is able to
// handle enough code on its own to provide noticeable overall speedups
// in -O0 compiles.
//
// Basic operations are supported in a target-independent way, by reading
// the same instruction descriptions that the SelectionDAG selector reads,
// and identifying simple arithmetic operations that can be directly selected
// from simple operators.  More complicated operations currently require
// target-specific code.
//
//===----------------------------------------------------------------------===//

#include "llvm/Function.h"
#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/DwarfWriter.h"
#include "llvm/Analysis/DebugInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "SelectionDAGBuild.h"
using namespace llvm;

unsigned FastISel::getRegForValue(Value *V) {
  EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
  // Don't handle non-simple values in FastISel.
  if (!RealVT.isSimple())
    return 0;

  // Ignore illegal types. We must do this before looking up the value
  // in ValueMap because Arguments are given virtual registers regardless
  // of whether FastISel can handle them.
  MVT VT = RealVT.getSimpleVT();
  if (!TLI.isTypeLegal(VT)) {
    // Promote MVT::i1 to a legal type though, because it's common and easy.
    if (VT == MVT::i1)
      VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
    else
      return 0;
  }

  // Look up the value to see if we already have a register for it. We
  // cache values defined by Instructions across blocks, and other values
  // only locally. This is because Instructions already have the SSA
  // def-dominatess-use requirement enforced.
  if (ValueMap.count(V))
    return ValueMap[V];
  unsigned Reg = LocalValueMap[V];
  if (Reg != 0)
    return Reg;

  if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
    if (CI->getValue().getActiveBits() <= 64)
      Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
  } else if (isa<AllocaInst>(V)) {
    Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
  } else if (isa<ConstantPointerNull>(V)) {
    // Translate this as an integer zero so that it can be
    // local-CSE'd with actual integer zeros.
    Reg =
      getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
  } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
    Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);

    if (!Reg) {
      const APFloat &Flt = CF->getValueAPF();
      EVT IntVT = TLI.getPointerTy();

      uint64_t x[2];
      uint32_t IntBitWidth = IntVT.getSizeInBits();
      bool isExact;
      (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
                                APFloat::rmTowardZero, &isExact);
      if (isExact) {
        APInt IntVal(IntBitWidth, 2, x);

        unsigned IntegerReg =
          getRegForValue(ConstantInt::get(V->getContext(), IntVal));
        if (IntegerReg != 0)
          Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg);
      }
    }
  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
    if (!SelectOperator(CE, CE->getOpcode())) return 0;
    Reg = LocalValueMap[CE];
  } else if (isa<UndefValue>(V)) {
    Reg = createResultReg(TLI.getRegClassFor(VT));
    BuildMI(MBB, DL, TII.get(TargetInstrInfo::IMPLICIT_DEF), Reg);
  }
  
  // If target-independent code couldn't handle the value, give target-specific
  // code a try.
  if (!Reg && isa<Constant>(V))
    Reg = TargetMaterializeConstant(cast<Constant>(V));
  
  // Don't cache constant materializations in the general ValueMap.
  // To do so would require tracking what uses they dominate.
  if (Reg != 0)
    LocalValueMap[V] = Reg;
  return Reg;
}

unsigned FastISel::lookUpRegForValue(Value *V) {
  // Look up the value to see if we already have a register for it. We
  // cache values defined by Instructions across blocks, and other values
  // only locally. This is because Instructions already have the SSA
  // def-dominatess-use requirement enforced.
  if (ValueMap.count(V))
    return ValueMap[V];
  return LocalValueMap[V];
}

/// UpdateValueMap - Update the value map to include the new mapping for this
/// instruction, or insert an extra copy to get the result in a previous
/// determined register.
/// NOTE: This is only necessary because we might select a block that uses
/// a value before we select the block that defines the value.  It might be
/// possible to fix this by selecting blocks in reverse postorder.
unsigned FastISel::UpdateValueMap(Value* I, unsigned Reg) {
  if (!isa<Instruction>(I)) {
    LocalValueMap[I] = Reg;
    return Reg;
  }
  
  unsigned &AssignedReg = ValueMap[I];
  if (AssignedReg == 0)
    AssignedReg = Reg;
  else if (Reg != AssignedReg) {
    const TargetRegisterClass *RegClass = MRI.getRegClass(Reg);
    TII.copyRegToReg(*MBB, MBB->end(), AssignedReg,
                     Reg, RegClass, RegClass);
  }
  return AssignedReg;
}

unsigned FastISel::getRegForGEPIndex(Value *Idx) {
  unsigned IdxN = getRegForValue(Idx);
  if (IdxN == 0)
    // Unhandled operand. Halt "fast" selection and bail.
    return 0;

  // If the index is smaller or larger than intptr_t, truncate or extend it.
  MVT PtrVT = TLI.getPointerTy();
  EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
  if (IdxVT.bitsLT(PtrVT))
    IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN);
  else if (IdxVT.bitsGT(PtrVT))
    IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, IdxN);
  return IdxN;
}

/// SelectBinaryOp - Select and emit code for a binary operator instruction,
/// which has an opcode which directly corresponds to the given ISD opcode.
///
bool FastISel::SelectBinaryOp(User *I, ISD::NodeType ISDOpcode) {
  EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
  if (VT == MVT::Other || !VT.isSimple())
    // Unhandled type. Halt "fast" selection and bail.
    return false;

  // We only handle legal types. For example, on x86-32 the instruction
  // selector contains all of the 64-bit instructions from x86-64,
  // under the assumption that i64 won't be used if the target doesn't
  // support it.
  if (!TLI.isTypeLegal(VT)) {
    // MVT::i1 is special. Allow AND, OR, or XOR because they
    // don't require additional zeroing, which makes them easy.
    if (VT == MVT::i1 &&
        (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
         ISDOpcode == ISD::XOR))
      VT = TLI.getTypeToTransformTo(I->getContext(), VT);
    else
      return false;
  }

  unsigned Op0 = getRegForValue(I->getOperand(0));
  if (Op0 == 0)
    // Unhandled operand. Halt "fast" selection and bail.
    return false;

  // Check if the second operand is a constant and handle it appropriately.
  if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
    unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
                                     ISDOpcode, Op0, CI->getZExtValue());
    if (ResultReg != 0) {
      // We successfully emitted code for the given LLVM Instruction.
      UpdateValueMap(I, ResultReg);
      return true;
    }
  }

  // Check if the second operand is a constant float.
  if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
    unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
                                     ISDOpcode, Op0, CF);
    if (ResultReg != 0) {
      // We successfully emitted code for the given LLVM Instruction.
      UpdateValueMap(I, ResultReg);
      return true;
    }
  }

  unsigned Op1 = getRegForValue(I->getOperand(1));
  if (Op1 == 0)
    // Unhandled operand. Halt "fast" selection and bail.
    return false;

  // Now we have both operands in registers. Emit the instruction.
  unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
                                   ISDOpcode, Op0, Op1);
  if (ResultReg == 0)
    // Target-specific code wasn't able to find a machine opcode for
    // the given ISD opcode and type. Halt "fast" selection and bail.
    return false;

  // We successfully emitted code for the given LLVM Instruction.
  UpdateValueMap(I, ResultReg);
  return true;
}

bool FastISel::SelectGetElementPtr(User *I) {
  unsigned N = getRegForValue(I->getOperand(0));
  if (N == 0)
    // Unhandled operand. Halt "fast" selection and bail.
    return false;

  const Type *Ty = I->getOperand(0)->getType();
  MVT VT = TLI.getPointerTy();
  for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end();
       OI != E; ++OI) {
    Value *Idx = *OI;
    if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
      unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
      if (Field) {
        // N = N + Offset
        uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
        // FIXME: This can be optimized by combining the add with a
        // subsequent one.
        N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
        if (N == 0)
          // Unhandled operand. Halt "fast" selection and bail.
          return false;
      }
      Ty = StTy->getElementType(Field);
    } else {
      Ty = cast<SequentialType>(Ty)->getElementType();

      // If this is a constant subscript, handle it quickly.
      if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
        if (CI->getZExtValue() == 0) continue;
        uint64_t Offs = 
          TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
        N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
        if (N == 0)
          // Unhandled operand. Halt "fast" selection and bail.
          return false;
        continue;
      }
      
      // N = N + Idx * ElementSize;
      uint64_t ElementSize = TD.getTypeAllocSize(Ty);
      unsigned IdxN = getRegForGEPIndex(Idx);
      if (IdxN == 0)
        // Unhandled operand. Halt "fast" selection and bail.
        return false;

      if (ElementSize != 1) {
        IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT);
        if (IdxN == 0)
          // Unhandled operand. Halt "fast" selection and bail.
          return false;
      }
      N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
      if (N == 0)
        // Unhandled operand. Halt "fast" selection and bail.
        return false;
    }
  }

  // We successfully emitted code for the given LLVM Instruction.
  UpdateValueMap(I, N);
  return true;
}

bool FastISel::SelectCall(User *I) {
  Function *F = cast<CallInst>(I)->getCalledFunction();
  if (!F) return false;

  unsigned IID = F->getIntrinsicID();
  switch (IID) {
  default: break;
  case Intrinsic::dbg_stoppoint: {
    DbgStopPointInst *SPI = cast<DbgStopPointInst>(I);
    if (isValidDebugInfoIntrinsic(*SPI, CodeGenOpt::None))
      setCurDebugLoc(ExtractDebugLocation(*SPI, MF.getDebugLocInfo()));
    return true;
  }
  case Intrinsic::dbg_region_start: {
    DbgRegionStartInst *RSI = cast<DbgRegionStartInst>(I);
    if (isValidDebugInfoIntrinsic(*RSI, CodeGenOpt::None) && DW
        && DW->ShouldEmitDwarfDebug()) {
      unsigned ID = 
        DW->RecordRegionStart(cast<GlobalVariable>(RSI->getContext()));
      const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
      BuildMI(MBB, DL, II).addImm(ID);
    }
    return true;
  }
  case Intrinsic::dbg_region_end: {
    DbgRegionEndInst *REI = cast<DbgRegionEndInst>(I);
    if (isValidDebugInfoIntrinsic(*REI, CodeGenOpt::None) && DW
        && DW->ShouldEmitDwarfDebug()) {
     unsigned ID = 0;
     DISubprogram Subprogram(cast<GlobalVariable>(REI->getContext()));
     if (isInlinedFnEnd(*REI, MF.getFunction())) {
        // This is end of an inlined function.
        const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
        ID = DW->RecordInlinedFnEnd(Subprogram);
        if (ID)
          // Returned ID is 0 if this is unbalanced "end of inlined
          // scope". This could happen if optimizer eats dbg intrinsics
          // or "beginning of inlined scope" is not recoginized due to
          // missing location info. In such cases, ignore this region.end.
          BuildMI(MBB, DL, II).addImm(ID);
      } else {
        const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
        ID =  DW->RecordRegionEnd(cast<GlobalVariable>(REI->getContext()));
        BuildMI(MBB, DL, II).addImm(ID);
      }
    }
    return true;
  }
  case Intrinsic::dbg_func_start: {
    DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
    if (!isValidDebugInfoIntrinsic(*FSI, CodeGenOpt::None) || !DW
        || !DW->ShouldEmitDwarfDebug()) 
      return true;

    if (isInlinedFnStart(*FSI, MF.getFunction())) {
      // This is a beginning of an inlined function.
      
      // If llvm.dbg.func.start is seen in a new block before any
      // llvm.dbg.stoppoint intrinsic then the location info is unknown.
      // FIXME : Why DebugLoc is reset at the beginning of each block ?
      DebugLoc PrevLoc = DL;
      if (PrevLoc.isUnknown())
        return true;
      // Record the source line.
      setCurDebugLoc(ExtractDebugLocation(*FSI, MF.getDebugLocInfo()));
      
      DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
      DISubprogram SP(cast<GlobalVariable>(FSI->getSubprogram()));
      unsigned LabelID = DW->RecordInlinedFnStart(SP,
                                                  DICompileUnit(PrevLocTpl.CompileUnit),
                                                  PrevLocTpl.Line,
                                                  PrevLocTpl.Col);
      const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
      BuildMI(MBB, DL, II).addImm(LabelID);
      return true;
    }
    
    // This is a beginning of a new function.
    MF.setDefaultDebugLoc(ExtractDebugLocation(*FSI, MF.getDebugLocInfo()));
    
    // llvm.dbg.func_start also defines beginning of function scope.
    DW->RecordRegionStart(cast<GlobalVariable>(FSI->getSubprogram()));
    return true;
  }
  case Intrinsic::dbg_declare: {
    DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
    if (!isValidDebugInfoIntrinsic(*DI, CodeGenOpt::None) || !DW
        || !DW->ShouldEmitDwarfDebug())
      return true;

    Value *Variable = DI->getVariable();
    Value *Address = DI->getAddress();
    if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
      Address = BCI->getOperand(0);
    AllocaInst *AI = dyn_cast<AllocaInst>(Address);
    // Don't handle byval struct arguments or VLAs, for example.
    if (!AI) break;
    DenseMap<const AllocaInst*, int>::iterator SI =
      StaticAllocaMap.find(AI);
    if (SI == StaticAllocaMap.end()) break; // VLAs.
    int FI = SI->second;
    
    // Determine the debug globalvariable.
    GlobalValue *GV = cast<GlobalVariable>(Variable);
    
    // Build the DECLARE instruction.
    const TargetInstrDesc &II = TII.get(TargetInstrInfo::DECLARE);
    MachineInstr *DeclareMI 
      = BuildMI(MBB, DL, II).addFrameIndex(FI).addGlobalAddress(GV);
    DIVariable DV(cast<GlobalVariable>(GV));
    DW->RecordVariableScope(DV, DeclareMI);
    return true;
  }
  case Intrinsic::eh_exception: {
    EVT VT = TLI.getValueType(I->getType());
    switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
    default: break;
    case TargetLowering::Expand: {
      assert(MBB->isLandingPad() && "Call to eh.exception not in landing pad!");
      unsigned Reg = TLI.getExceptionAddressRegister();
      const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
      unsigned ResultReg = createResultReg(RC);
      bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                           Reg, RC, RC);
      assert(InsertedCopy && "Can't copy address registers!");
      InsertedCopy = InsertedCopy;
      UpdateValueMap(I, ResultReg);
      return true;
    }
    }
    break;
  }
  case Intrinsic::eh_selector_i32:
  case Intrinsic::eh_selector_i64: {
    EVT VT = TLI.getValueType(I->getType());
    switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
    default: break;
    case TargetLowering::Expand: {
      EVT VT = (IID == Intrinsic::eh_selector_i32 ?
                           MVT::i32 : MVT::i64);

      if (MMI) {
        if (MBB->isLandingPad())
          AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
        else {
#ifndef NDEBUG
          CatchInfoLost.insert(cast<CallInst>(I));
#endif
          // FIXME: Mark exception selector register as live in.  Hack for PR1508.
          unsigned Reg = TLI.getExceptionSelectorRegister();
          if (Reg) MBB->addLiveIn(Reg);
        }

        unsigned Reg = TLI.getExceptionSelectorRegister();
        const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
        unsigned ResultReg = createResultReg(RC);
        bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                             Reg, RC, RC);
        assert(InsertedCopy && "Can't copy address registers!");
        InsertedCopy = InsertedCopy;
        UpdateValueMap(I, ResultReg);
      } else {
        unsigned ResultReg =
          getRegForValue(Constant::getNullValue(I->getType()));
        UpdateValueMap(I, ResultReg);
      }
      return true;
    }
    }
    break;
  }
  }
  return false;
}

bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
  EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
  EVT DstVT = TLI.getValueType(I->getType());
    
  if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
      DstVT == MVT::Other || !DstVT.isSimple())
    // Unhandled type. Halt "fast" selection and bail.
    return false;
    
  // Check if the destination type is legal. Or as a special case,
  // it may be i1 if we're doing a truncate because that's
  // easy and somewhat common.
  if (!TLI.isTypeLegal(DstVT))
    if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
      // Unhandled type. Halt "fast" selection and bail.
      return false;

  // Check if the source operand is legal. Or as a special case,
  // it may be i1 if we're doing zero-extension because that's
  // easy and somewhat common.
  if (!TLI.isTypeLegal(SrcVT))
    if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
      // Unhandled type. Halt "fast" selection and bail.
      return false;

  unsigned InputReg = getRegForValue(I->getOperand(0));
  if (!InputReg)
    // Unhandled operand.  Halt "fast" selection and bail.
    return false;

  // If the operand is i1, arrange for the high bits in the register to be zero.
  if (SrcVT == MVT::i1) {
   SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
   InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
   if (!InputReg)
     return false;
  }
  // If the result is i1, truncate to the target's type for i1 first.
  if (DstVT == MVT::i1)
    DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);

  unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
                                  DstVT.getSimpleVT(),
                                  Opcode,
                                  InputReg);
  if (!ResultReg)
    return false;
    
  UpdateValueMap(I, ResultReg);
  return true;
}

bool FastISel::SelectBitCast(User *I) {
  // If the bitcast doesn't change the type, just use the operand value.
  if (I->getType() == I->getOperand(0)->getType()) {
    unsigned Reg = getRegForValue(I->getOperand(0));
    if (Reg == 0)
      return false;
    UpdateValueMap(I, Reg);
    return true;
  }

  // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
  EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
  EVT DstVT = TLI.getValueType(I->getType());
  
  if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
      DstVT == MVT::Other || !DstVT.isSimple() ||
      !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
    // Unhandled type. Halt "fast" selection and bail.
    return false;
  
  unsigned Op0 = getRegForValue(I->getOperand(0));
  if (Op0 == 0)
    // Unhandled operand. Halt "fast" selection and bail.
    return false;
  
  // First, try to perform the bitcast by inserting a reg-reg copy.
  unsigned ResultReg = 0;
  if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
    TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
    TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
    ResultReg = createResultReg(DstClass);
    
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         Op0, DstClass, SrcClass);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  
  // If the reg-reg copy failed, select a BIT_CONVERT opcode.
  if (!ResultReg)
    ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
                           ISD::BIT_CONVERT, Op0);
  
  if (!ResultReg)
    return false;
  
  UpdateValueMap(I, ResultReg);
  return true;
}

bool
FastISel::SelectInstruction(Instruction *I) {
  return SelectOperator(I, I->getOpcode());
}

/// FastEmitBranch - Emit an unconditional branch to the given block,
/// unless it is the immediate (fall-through) successor, and update
/// the CFG.
void
FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
  MachineFunction::iterator NextMBB =
     next(MachineFunction::iterator(MBB));

  if (MBB->isLayoutSuccessor(MSucc)) {
    // The unconditional fall-through case, which needs no instructions.
  } else {
    // The unconditional branch case.
    TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
  }
  MBB->addSuccessor(MSucc);
}

bool
FastISel::SelectOperator(User *I, unsigned Opcode) {
  switch (Opcode) {
  case Instruction::Add:
    return SelectBinaryOp(I, ISD::ADD);
  case Instruction::FAdd:
    return SelectBinaryOp(I, ISD::FADD);
  case Instruction::Sub:
    return SelectBinaryOp(I, ISD::SUB);
  case Instruction::FSub:
    return SelectBinaryOp(I, ISD::FSUB);
  case Instruction::Mul:
    return SelectBinaryOp(I, ISD::MUL);
  case Instruction::FMul:
    return SelectBinaryOp(I, ISD::FMUL);
  case Instruction::SDiv:
    return SelectBinaryOp(I, ISD::SDIV);
  case Instruction::UDiv:
    return SelectBinaryOp(I, ISD::UDIV);
  case Instruction::FDiv:
    return SelectBinaryOp(I, ISD::FDIV);
  case Instruction::SRem:
    return SelectBinaryOp(I, ISD::SREM);
  case Instruction::URem:
    return SelectBinaryOp(I, ISD::UREM);
  case Instruction::FRem:
    return SelectBinaryOp(I, ISD::FREM);
  case Instruction::Shl:
    return SelectBinaryOp(I, ISD::SHL);
  case Instruction::LShr:
    return SelectBinaryOp(I, ISD::SRL);
  case Instruction::AShr:
    return SelectBinaryOp(I, ISD::SRA);
  case Instruction::And:
    return SelectBinaryOp(I, ISD::AND);
  case Instruction::Or:
    return SelectBinaryOp(I, ISD::OR);
  case Instruction::Xor:
    return SelectBinaryOp(I, ISD::XOR);

  case Instruction::GetElementPtr:
    return SelectGetElementPtr(I);

  case Instruction::Br: {
    BranchInst *BI = cast<BranchInst>(I);

    if (BI->isUnconditional()) {
      BasicBlock *LLVMSucc = BI->getSuccessor(0);
      MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
      FastEmitBranch(MSucc);
      return true;
    }

    // Conditional branches are not handed yet.
    // Halt "fast" selection and bail.
    return false;
  }

  case Instruction::Unreachable:
    // Nothing to emit.
    return true;

  case Instruction::PHI:
    // PHI nodes are already emitted.
    return true;

  case Instruction::Alloca:
    // FunctionLowering has the static-sized case covered.
    if (StaticAllocaMap.count(cast<AllocaInst>(I)))
      return true;

    // Dynamic-sized alloca is not handled yet.
    return false;
    
  case Instruction::Call:
    return SelectCall(I);
  
  case Instruction::BitCast:
    return SelectBitCast(I);

  case Instruction::FPToSI:
    return SelectCast(I, ISD::FP_TO_SINT);
  case Instruction::ZExt:
    return SelectCast(I, ISD::ZERO_EXTEND);
  case Instruction::SExt:
    return SelectCast(I, ISD::SIGN_EXTEND);
  case Instruction::Trunc:
    return SelectCast(I, ISD::TRUNCATE);
  case Instruction::SIToFP:
    return SelectCast(I, ISD::SINT_TO_FP);

  case Instruction::IntToPtr: // Deliberate fall-through.
  case Instruction::PtrToInt: {
    EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
    EVT DstVT = TLI.getValueType(I->getType());
    if (DstVT.bitsGT(SrcVT))
      return SelectCast(I, ISD::ZERO_EXTEND);
    if (DstVT.bitsLT(SrcVT))
      return SelectCast(I, ISD::TRUNCATE);
    unsigned Reg = getRegForValue(I->getOperand(0));
    if (Reg == 0) return false;
    UpdateValueMap(I, Reg);
    return true;
  }

  default:
    // Unhandled instruction. Halt "fast" selection and bail.
    return false;
  }
}

FastISel::FastISel(MachineFunction &mf,
                   MachineModuleInfo *mmi,
                   DwarfWriter *dw,
                   DenseMap<const Value *, unsigned> &vm,
                   DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
                   DenseMap<const AllocaInst *, int> &am
#ifndef NDEBUG
                   , SmallSet<Instruction*, 8> &cil
#endif
                   )
  : MBB(0),
    ValueMap(vm),
    MBBMap(bm),
    StaticAllocaMap(am),
#ifndef NDEBUG
    CatchInfoLost(cil),
#endif
    MF(mf),
    MMI(mmi),
    DW(dw),
    MRI(MF.getRegInfo()),
    MFI(*MF.getFrameInfo()),
    MCP(*MF.getConstantPool()),
    TM(MF.getTarget()),
    TD(*TM.getTargetData()),
    TII(*TM.getInstrInfo()),
    TLI(*TM.getTargetLowering()) {
}

FastISel::~FastISel() {}

unsigned FastISel::FastEmit_(MVT, MVT,
                             ISD::NodeType) {
  return 0;
}

unsigned FastISel::FastEmit_r(MVT, MVT,
                              ISD::NodeType, unsigned /*Op0*/) {
  return 0;
}

unsigned FastISel::FastEmit_rr(MVT, MVT, 
                               ISD::NodeType, unsigned /*Op0*/,
                               unsigned /*Op0*/) {
  return 0;
}

unsigned FastISel::FastEmit_i(MVT, MVT, ISD::NodeType, uint64_t /*Imm*/) {
  return 0;
}

unsigned FastISel::FastEmit_f(MVT, MVT,
                              ISD::NodeType, ConstantFP * /*FPImm*/) {
  return 0;
}

unsigned FastISel::FastEmit_ri(MVT, MVT,
                               ISD::NodeType, unsigned /*Op0*/,
                               uint64_t /*Imm*/) {
  return 0;
}

unsigned FastISel::FastEmit_rf(MVT, MVT,
                               ISD::NodeType, unsigned /*Op0*/,
                               ConstantFP * /*FPImm*/) {
  return 0;
}

unsigned FastISel::FastEmit_rri(MVT, MVT,
                                ISD::NodeType,
                                unsigned /*Op0*/, unsigned /*Op1*/,
                                uint64_t /*Imm*/) {
  return 0;
}

/// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
/// to emit an instruction with an immediate operand using FastEmit_ri.
/// If that fails, it materializes the immediate into a register and try
/// FastEmit_rr instead.
unsigned FastISel::FastEmit_ri_(MVT VT, ISD::NodeType Opcode,
                                unsigned Op0, uint64_t Imm,
                                MVT ImmType) {
  // First check if immediate type is legal. If not, we can't use the ri form.
  unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
  if (ResultReg != 0)
    return ResultReg;
  unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
  if (MaterialReg == 0)
    return 0;
  return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
}

/// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
/// to emit an instruction with a floating-point immediate operand using
/// FastEmit_rf. If that fails, it materializes the immediate into a register
/// and try FastEmit_rr instead.
unsigned FastISel::FastEmit_rf_(MVT VT, ISD::NodeType Opcode,
                                unsigned Op0, ConstantFP *FPImm,
                                MVT ImmType) {
  // First check if immediate type is legal. If not, we can't use the rf form.
  unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
  if (ResultReg != 0)
    return ResultReg;

  // Materialize the constant in a register.
  unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
  if (MaterialReg == 0) {
    // If the target doesn't have a way to directly enter a floating-point
    // value into a register, use an alternate approach.
    // TODO: The current approach only supports floating-point constants
    // that can be constructed by conversion from integer values. This should
    // be replaced by code that creates a load from a constant-pool entry,
    // which will require some target-specific work.
    const APFloat &Flt = FPImm->getValueAPF();
    EVT IntVT = TLI.getPointerTy();

    uint64_t x[2];
    uint32_t IntBitWidth = IntVT.getSizeInBits();
    bool isExact;
    (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
                             APFloat::rmTowardZero, &isExact);
    if (!isExact)
      return 0;
    APInt IntVal(IntBitWidth, 2, x);

    unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
                                     ISD::Constant, IntVal.getZExtValue());
    if (IntegerReg == 0)
      return 0;
    MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
                             ISD::SINT_TO_FP, IntegerReg);
    if (MaterialReg == 0)
      return 0;
  }
  return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
}

unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
  return MRI.createVirtualRegister(RC);
}

unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
                                 const TargetRegisterClass* RC) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  BuildMI(MBB, DL, II, ResultReg);
  return ResultReg;
}

unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
                                  const TargetRegisterClass *RC,
                                  unsigned Op0) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
  else {
    BuildMI(MBB, DL, II).addReg(Op0);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }

  return ResultReg;
}

unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
                                   const TargetRegisterClass *RC,
                                   unsigned Op0, unsigned Op1) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
  else {
    BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  return ResultReg;
}

unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
                                   const TargetRegisterClass *RC,
                                   unsigned Op0, uint64_t Imm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
  else {
    BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  return ResultReg;
}

unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
                                   const TargetRegisterClass *RC,
                                   unsigned Op0, ConstantFP *FPImm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
  else {
    BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  return ResultReg;
}

unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
                                    const TargetRegisterClass *RC,
                                    unsigned Op0, unsigned Op1, uint64_t Imm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
  else {
    BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  return ResultReg;
}

unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
                                  const TargetRegisterClass *RC,
                                  uint64_t Imm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);
  
  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
  else {
    BuildMI(MBB, DL, II).addImm(Imm);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  return ResultReg;
}

unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
                                              unsigned Op0, uint32_t Idx) {
  const TargetRegisterClass* RC = MRI.getRegClass(Op0);
  
  unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
  const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
  
  if (II.getNumDefs() >= 1)
    BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
  else {
    BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
    bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                         II.ImplicitDefs[0], RC, RC);
    if (!InsertedCopy)
      ResultReg = 0;
  }
  return ResultReg;
}

/// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
/// with all but the least significant bit set to zero.
unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op) {
  return FastEmit_ri(VT, VT, ISD::AND, Op, 1);
}