summaryrefslogtreecommitdiff
path: root/lib/Target/X86/X86ISelDAGToDAG.cpp
blob: bf08dd8c189f316c01dee7b70f577cab3794dbaf (plain)
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
//===- X86ISelDAGToDAG.cpp - A DAG pattern matching inst selector for X86 -===//
//
//                     The LLVM Compiler Infrastructure
//
// This file was developed by the Evan Cheng and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines a DAG pattern matching instruction selector for X86,
// converting from a legalized dag to a X86 dag.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "x86-isel"
#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
#include "X86RegisterInfo.h"
#include "X86Subtarget.h"
#include "X86TargetMachine.h"
#include "llvm/GlobalValue.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/Support/CFG.h"
#include "llvm/Type.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/ADT/Statistic.h"
#include <queue>
#include <set>
using namespace llvm;

STATISTIC(NumFPKill   , "Number of FP_REG_KILL instructions added");
STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");


//===----------------------------------------------------------------------===//
//                      Pattern Matcher Implementation
//===----------------------------------------------------------------------===//

namespace {
  /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
  /// SDOperand's instead of register numbers for the leaves of the matched
  /// tree.
  struct X86ISelAddressMode {
    enum {
      RegBase,
      FrameIndexBase
    } BaseType;

    struct {            // This is really a union, discriminated by BaseType!
      SDOperand Reg;
      int FrameIndex;
    } Base;

    bool isRIPRel;     // RIP relative?
    unsigned Scale;
    SDOperand IndexReg; 
    unsigned Disp;
    GlobalValue *GV;
    Constant *CP;
    const char *ES;
    int JT;
    unsigned Align;    // CP alignment.

    X86ISelAddressMode()
      : BaseType(RegBase), isRIPRel(false), Scale(1), IndexReg(), Disp(0),
        GV(0), CP(0), ES(0), JT(-1), Align(0) {
    }
  };
}

namespace {
  //===--------------------------------------------------------------------===//
  /// ISel - X86 specific code to select X86 machine instructions for
  /// SelectionDAG operations.
  ///
  class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
    /// ContainsFPCode - Every instruction we select that uses or defines a FP
    /// register should set this to true.
    bool ContainsFPCode;

    /// FastISel - Enable fast(er) instruction selection.
    ///
    bool FastISel;

    /// TM - Keep a reference to X86TargetMachine.
    ///
    X86TargetMachine &TM;

    /// X86Lowering - This object fully describes how to lower LLVM code to an
    /// X86-specific SelectionDAG.
    X86TargetLowering X86Lowering;

    /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
    /// make the right decision when generating code for different targets.
    const X86Subtarget *Subtarget;

    /// GlobalBaseReg - keeps track of the virtual register mapped onto global
    /// base register.
    unsigned GlobalBaseReg;

  public:
    X86DAGToDAGISel(X86TargetMachine &tm, bool fast)
      : SelectionDAGISel(X86Lowering),
        ContainsFPCode(false), FastISel(fast), TM(tm),
        X86Lowering(*TM.getTargetLowering()),
        Subtarget(&TM.getSubtarget<X86Subtarget>()) {}

    virtual bool runOnFunction(Function &Fn) {
      // Make sure we re-emit a set of the global base reg if necessary
      GlobalBaseReg = 0;
      return SelectionDAGISel::runOnFunction(Fn);
    }
   
    virtual const char *getPassName() const {
      return "X86 DAG->DAG Instruction Selection";
    }

    /// InstructionSelectBasicBlock - This callback is invoked by
    /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
    virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);

    virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);

    virtual bool CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root);

// Include the pieces autogenerated from the target description.
#include "X86GenDAGISel.inc"

  private:
    SDNode *Select(SDOperand N);

    bool MatchAddress(SDOperand N, X86ISelAddressMode &AM,
                      bool isRoot = true, unsigned Depth = 0);
    bool SelectAddr(SDOperand Op, SDOperand N, SDOperand &Base,
                    SDOperand &Scale, SDOperand &Index, SDOperand &Disp);
    bool SelectLEAAddr(SDOperand Op, SDOperand N, SDOperand &Base,
                       SDOperand &Scale, SDOperand &Index, SDOperand &Disp);
    bool SelectScalarSSELoad(SDOperand Op, SDOperand Pred,
                             SDOperand N, SDOperand &Base, SDOperand &Scale,
                             SDOperand &Index, SDOperand &Disp,
                             SDOperand &InChain, SDOperand &OutChain);
    bool TryFoldLoad(SDOperand P, SDOperand N,
                     SDOperand &Base, SDOperand &Scale,
                     SDOperand &Index, SDOperand &Disp);
    void InstructionSelectPreprocess(SelectionDAG &DAG);

    /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
    /// inline asm expressions.
    virtual bool SelectInlineAsmMemoryOperand(const SDOperand &Op,
                                              char ConstraintCode,
                                              std::vector<SDOperand> &OutOps,
                                              SelectionDAG &DAG);
    
    void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);

    inline void getAddressOperands(X86ISelAddressMode &AM, SDOperand &Base, 
                                   SDOperand &Scale, SDOperand &Index,
                                   SDOperand &Disp) {
      Base  = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
        CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
        AM.Base.Reg;
      Scale = getI8Imm(AM.Scale);
      Index = AM.IndexReg;
      // These are 32-bit even in 64-bit mode since RIP relative offset
      // is 32-bit.
      if (AM.GV)
        Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp);
      else if (AM.CP)
        Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32, AM.Align, AM.Disp);
      else if (AM.ES)
        Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32);
      else if (AM.JT != -1)
        Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32);
      else
        Disp = getI32Imm(AM.Disp);
    }

    /// getI8Imm - Return a target constant with the specified value, of type
    /// i8.
    inline SDOperand getI8Imm(unsigned Imm) {
      return CurDAG->getTargetConstant(Imm, MVT::i8);
    }

    /// getI16Imm - Return a target constant with the specified value, of type
    /// i16.
    inline SDOperand getI16Imm(unsigned Imm) {
      return CurDAG->getTargetConstant(Imm, MVT::i16);
    }

    /// getI32Imm - Return a target constant with the specified value, of type
    /// i32.
    inline SDOperand getI32Imm(unsigned Imm) {
      return CurDAG->getTargetConstant(Imm, MVT::i32);
    }

    /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
    /// base register.  Return the virtual register that holds this value.
    SDNode *getGlobalBaseReg();

#ifndef NDEBUG
    unsigned Indent;
#endif
  };
}

static SDNode *findFlagUse(SDNode *N) {
  unsigned FlagResNo = N->getNumValues()-1;
  for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
    SDNode *User = *I;
    for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
      SDOperand Op = User->getOperand(i);
      if (Op.Val == N && Op.ResNo == FlagResNo)
        return User;
    }
  }
  return NULL;
}

static void findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
                          SDNode *Root, SDNode *Skip, bool &found,
                          std::set<SDNode *> &Visited) {
  if (found ||
      Use->getNodeId() > Def->getNodeId() ||
      !Visited.insert(Use).second)
    return;

  for (unsigned i = 0, e = Use->getNumOperands(); !found && i != e; ++i) {
    SDNode *N = Use->getOperand(i).Val;
    if (N == Skip)
      continue;
    if (N == Def) {
      if (Use == ImmedUse)
        continue; // Immediate use is ok.
      if (Use == Root) {
        assert(Use->getOpcode() == ISD::STORE ||
               Use->getOpcode() == X86ISD::CMP);
        continue;
      }
      found = true;
      break;
    }
    findNonImmUse(N, Def, ImmedUse, Root, Skip, found, Visited);
  }
}

/// isNonImmUse - Start searching from Root up the DAG to check is Def can
/// be reached. Return true if that's the case. However, ignore direct uses
/// by ImmedUse (which would be U in the example illustrated in
/// CanBeFoldedBy) and by Root (which can happen in the store case).
/// FIXME: to be really generic, we should allow direct use by any node
/// that is being folded. But realisticly since we only fold loads which
/// have one non-chain use, we only need to watch out for load/op/store
/// and load/op/cmp case where the root (store / cmp) may reach the load via
/// its chain operand.
static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse,
                               SDNode *Skip = NULL) {
  std::set<SDNode *> Visited;
  bool found = false;
  findNonImmUse(Root, Def, ImmedUse, Root, Skip, found, Visited);
  return found;
}


bool X86DAGToDAGISel::CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root) {
  if (FastISel) return false;

  // If U use can somehow reach N through another path then U can't fold N or
  // it will create a cycle. e.g. In the following diagram, U can reach N
  // through X. If N is folded into into U, then X is both a predecessor and
  // a successor of U.
  //
  //         [ N ]
  //         ^  ^
  //         |  |
  //        /   \---
  //      /        [X]
  //      |         ^
  //     [U]--------|

  if (isNonImmUse(Root, N, U))
    return false;

  // If U produces a flag, then it gets (even more) interesting. Since it
  // would have been "glued" together with its flag use, we need to check if
  // it might reach N:
  //
  //       [ N ]
  //        ^ ^
  //        | |
  //       [U] \--
  //        ^   [TF]
  //        |    ^
  //        |    |
  //         \  /
  //          [FU]
  //
  // If FU (flag use) indirectly reach N (the load), and U fold N (call it
  // NU), then TF is a predecessor of FU and a successor of NU. But since
  // NU and FU are flagged together, this effectively creates a cycle.
  bool HasFlagUse = false;
  MVT::ValueType VT = Root->getValueType(Root->getNumValues()-1);
  while ((VT == MVT::Flag && !Root->use_empty())) {
    SDNode *FU = findFlagUse(Root);
    if (FU == NULL)
      break;
    else {
      Root = FU;
      HasFlagUse = true;
    }
    VT = Root->getValueType(Root->getNumValues()-1);
  }

  if (HasFlagUse)
    return !isNonImmUse(Root, N, Root, U);
  return true;
}

/// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
/// and move load below the TokenFactor. Replace store's chain operand with
/// load's chain result.
static void MoveBelowTokenFactor(SelectionDAG &DAG, SDOperand Load,
                                 SDOperand Store, SDOperand TF) {
  std::vector<SDOperand> Ops;
  for (unsigned i = 0, e = TF.Val->getNumOperands(); i != e; ++i)
    if (Load.Val == TF.Val->getOperand(i).Val)
      Ops.push_back(Load.Val->getOperand(0));
    else
      Ops.push_back(TF.Val->getOperand(i));
  DAG.UpdateNodeOperands(TF, &Ops[0], Ops.size());
  DAG.UpdateNodeOperands(Load, TF, Load.getOperand(1), Load.getOperand(2));
  DAG.UpdateNodeOperands(Store, Load.getValue(1), Store.getOperand(1),
                         Store.getOperand(2), Store.getOperand(3));
}

/// InstructionSelectPreprocess - Preprocess the DAG to allow the instruction
/// selector to pick more load-modify-store instructions. This is a common
/// case:
///
///     [Load chain]
///         ^
///         |
///       [Load]
///       ^    ^
///       |    |
///      /      \-
///     /         |
/// [TokenFactor] [Op]
///     ^          ^
///     |          |
///      \        /
///       \      /
///       [Store]
///
/// The fact the store's chain operand != load's chain will prevent the
/// (store (op (load))) instruction from being selected. We can transform it to:
///
///     [Load chain]
///         ^
///         |
///    [TokenFactor]
///         ^
///         |
///       [Load]
///       ^    ^
///       |    |
///       |     \- 
///       |       | 
///       |     [Op]
///       |       ^
///       |       |
///       \      /
///        \    /
///       [Store]
void X86DAGToDAGISel::InstructionSelectPreprocess(SelectionDAG &DAG) {
  for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
         E = DAG.allnodes_end(); I != E; ++I) {
    if (!ISD::isNON_TRUNCStore(I))
      continue;
    SDOperand Chain = I->getOperand(0);
    if (Chain.Val->getOpcode() != ISD::TokenFactor)
      continue;

    SDOperand N1 = I->getOperand(1);
    SDOperand N2 = I->getOperand(2);
    if (MVT::isFloatingPoint(N1.getValueType()) ||
        MVT::isVector(N1.getValueType()) ||
        !N1.hasOneUse())
      continue;

    bool RModW = false;
    SDOperand Load;
    unsigned Opcode = N1.Val->getOpcode();
    switch (Opcode) {
      case ISD::ADD:
      case ISD::MUL:
      case ISD::AND:
      case ISD::OR:
      case ISD::XOR:
      case ISD::ADDC:
      case ISD::ADDE: {
        SDOperand N10 = N1.getOperand(0);
        SDOperand N11 = N1.getOperand(1);
        if (ISD::isNON_EXTLoad(N10.Val))
          RModW = true;
        else if (ISD::isNON_EXTLoad(N11.Val)) {
          RModW = true;
          std::swap(N10, N11);
        }
        RModW = RModW && N10.Val->isOperand(Chain.Val) && N10.hasOneUse() &&
          (N10.getOperand(1) == N2) &&
          (N10.Val->getValueType(0) == N1.getValueType());
        if (RModW)
          Load = N10;
        break;
      }
      case ISD::SUB:
      case ISD::SHL:
      case ISD::SRA:
      case ISD::SRL:
      case ISD::ROTL:
      case ISD::ROTR:
      case ISD::SUBC:
      case ISD::SUBE:
      case X86ISD::SHLD:
      case X86ISD::SHRD: {
        SDOperand N10 = N1.getOperand(0);
        if (ISD::isNON_EXTLoad(N10.Val))
          RModW = N10.Val->isOperand(Chain.Val) && N10.hasOneUse() &&
            (N10.getOperand(1) == N2) &&
            (N10.Val->getValueType(0) == N1.getValueType());
        if (RModW)
          Load = N10;
        break;
      }
    }

    if (RModW) {
      MoveBelowTokenFactor(DAG, Load, SDOperand(I, 0), Chain);
      ++NumLoadMoved;
    }
  }
}

/// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
/// when it has created a SelectionDAG for us to codegen.
void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
  DEBUG(BB->dump());
  MachineFunction::iterator FirstMBB = BB;

  if (!FastISel)
    InstructionSelectPreprocess(DAG);

  // Codegen the basic block.
#ifndef NDEBUG
  DOUT << "===== Instruction selection begins:\n";
  Indent = 0;
#endif
  DAG.setRoot(SelectRoot(DAG.getRoot()));
#ifndef NDEBUG
  DOUT << "===== Instruction selection ends:\n";
#endif

  DAG.RemoveDeadNodes();

  // Emit machine code to BB. 
  ScheduleAndEmitDAG(DAG);
  
  // If we are emitting FP stack code, scan the basic block to determine if this
  // block defines any FP values.  If so, put an FP_REG_KILL instruction before
  // the terminator of the block.
  if (!Subtarget->hasSSE2()) {
    // Note that FP stack instructions *are* used in SSE code when returning
    // values, but these are not live out of the basic block, so we don't need
    // an FP_REG_KILL in this case either.
    bool ContainsFPCode = false;
    
    // Scan all of the machine instructions in these MBBs, checking for FP
    // stores.
    MachineFunction::iterator MBBI = FirstMBB;
    do {
      for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end();
           !ContainsFPCode && I != E; ++I) {
        if (I->getNumOperands() != 0 && I->getOperand(0).isRegister()) {
          for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
            if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
                MRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
                RegMap->getRegClass(I->getOperand(0).getReg()) == 
                X86::RFPRegisterClass) {
              ContainsFPCode = true;
              break;
            }
          }
        }
      }
    } while (!ContainsFPCode && &*(MBBI++) != BB);
    
    // Check PHI nodes in successor blocks.  These PHI's will be lowered to have
    // a copy of the input value in this block.
    if (!ContainsFPCode) {
      // Final check, check LLVM BB's that are successors to the LLVM BB
      // corresponding to BB for FP PHI nodes.
      const BasicBlock *LLVMBB = BB->getBasicBlock();
      const PHINode *PN;
      for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
           !ContainsFPCode && SI != E; ++SI) {
        for (BasicBlock::const_iterator II = SI->begin();
             (PN = dyn_cast<PHINode>(II)); ++II) {
          if (PN->getType()->isFloatingPoint()) {
            ContainsFPCode = true;
            break;
          }
        }
      }
    }

    // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
    if (ContainsFPCode) {
      BuildMI(*BB, BB->getFirstTerminator(),
              TM.getInstrInfo()->get(X86::FP_REG_KILL));
      ++NumFPKill;
    }
  }
}

/// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
/// the main function.
void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
                                             MachineFrameInfo *MFI) {
  const TargetInstrInfo *TII = TM.getInstrInfo();
  if (Subtarget->isTargetCygMing())
    BuildMI(BB, TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");

  // Switch the FPU to 64-bit precision mode for better compatibility and speed.
  int CWFrameIdx = MFI->CreateStackObject(2, 2);
  addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);

  // Set the high part to be 64-bit precision.
  addFrameReference(BuildMI(BB, TII->get(X86::MOV8mi)),
                    CWFrameIdx, 1).addImm(2);

  // Reload the modified control word now.
  addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
}

void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
  // If this is main, emit special code for main.
  MachineBasicBlock *BB = MF.begin();
  if (Fn.hasExternalLinkage() && Fn.getName() == "main")
    EmitSpecialCodeForMain(BB, MF.getFrameInfo());
}

/// MatchAddress - Add the specified node to the specified addressing mode,
/// returning true if it cannot be done.  This just pattern matches for the
/// addressing mode
bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
                                   bool isRoot, unsigned Depth) {
  if (Depth > 5) {
    // Default, generate it as a register.
    AM.BaseType = X86ISelAddressMode::RegBase;
    AM.Base.Reg = N;
    return false;
  }
  
  // RIP relative addressing: %rip + 32-bit displacement!
  if (AM.isRIPRel) {
    if (!AM.ES && AM.JT != -1 && N.getOpcode() == ISD::Constant) {
      int64_t Val = cast<ConstantSDNode>(N)->getSignExtended();
      if (isInt32(AM.Disp + Val)) {
        AM.Disp += Val;
        return false;
      }
    }
    return true;
  }

  int id = N.Val->getNodeId();
  bool Available = isSelected(id);

  switch (N.getOpcode()) {
  default: break;
  case ISD::Constant: {
    int64_t Val = cast<ConstantSDNode>(N)->getSignExtended();
    if (isInt32(AM.Disp + Val)) {
      AM.Disp += Val;
      return false;
    }
    break;
  }

  case X86ISD::Wrapper: {
    bool is64Bit = Subtarget->is64Bit();
    // Under X86-64 non-small code model, GV (and friends) are 64-bits.
    if (is64Bit && TM.getCodeModel() != CodeModel::Small)
      break;
    if (AM.GV != 0 || AM.CP != 0 || AM.ES != 0 || AM.JT != -1)
      break;
    // If value is available in a register both base and index components have
    // been picked, we can't fit the result available in the register in the
    // addressing mode. Duplicate GlobalAddress or ConstantPool as displacement.
    if (!Available || (AM.Base.Reg.Val && AM.IndexReg.Val)) {
      bool isStatic = TM.getRelocationModel() == Reloc::Static;
      SDOperand N0 = N.getOperand(0);
      if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
        GlobalValue *GV = G->getGlobal();
        bool isAbs32 = !is64Bit || isStatic;
        if (isAbs32 || isRoot) {
          AM.GV = GV;
          AM.Disp += G->getOffset();
          AM.isRIPRel = !isAbs32;
          return false;
        }
      } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
        if (!is64Bit || isStatic || isRoot) {
          AM.CP = CP->getConstVal();
          AM.Align = CP->getAlignment();
          AM.Disp += CP->getOffset();
          AM.isRIPRel = !isStatic;
          return false;
        }
      } else if (ExternalSymbolSDNode *S =dyn_cast<ExternalSymbolSDNode>(N0)) {
        if (isStatic || isRoot) {
          AM.ES = S->getSymbol();
          AM.isRIPRel = !isStatic;
          return false;
        }
      } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
        if (isStatic || isRoot) {
          AM.JT = J->getIndex();
          AM.isRIPRel = !isStatic;
          return false;
        }
      }
    }
    break;
  }

  case ISD::FrameIndex:
    if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
      AM.BaseType = X86ISelAddressMode::FrameIndexBase;
      AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
      return false;
    }
    break;

  case ISD::SHL:
    if (!Available && AM.IndexReg.Val == 0 && AM.Scale == 1)
      if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
        unsigned Val = CN->getValue();
        if (Val == 1 || Val == 2 || Val == 3) {
          AM.Scale = 1 << Val;
          SDOperand ShVal = N.Val->getOperand(0);

          // Okay, we know that we have a scale by now.  However, if the scaled
          // value is an add of something and a constant, we can fold the
          // constant into the disp field here.
          if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
              isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
            AM.IndexReg = ShVal.Val->getOperand(0);
            ConstantSDNode *AddVal =
              cast<ConstantSDNode>(ShVal.Val->getOperand(1));
            uint64_t Disp = AM.Disp + (AddVal->getValue() << Val);
            if (isInt32(Disp))
              AM.Disp = Disp;
            else
              AM.IndexReg = ShVal;
          } else {
            AM.IndexReg = ShVal;
          }
          return false;
        }
      }
    break;

  case ISD::MUL:
    // X*[3,5,9] -> X+X*[2,4,8]
    if (!Available &&
        AM.BaseType == X86ISelAddressMode::RegBase &&
        AM.Base.Reg.Val == 0 &&
        AM.IndexReg.Val == 0) {
      if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
        if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
          AM.Scale = unsigned(CN->getValue())-1;

          SDOperand MulVal = N.Val->getOperand(0);
          SDOperand Reg;

          // Okay, we know that we have a scale by now.  However, if the scaled
          // value is an add of something and a constant, we can fold the
          // constant into the disp field here.
          if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
              isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
            Reg = MulVal.Val->getOperand(0);
            ConstantSDNode *AddVal =
              cast<ConstantSDNode>(MulVal.Val->getOperand(1));
            uint64_t Disp = AM.Disp + AddVal->getValue() * CN->getValue();
            if (isInt32(Disp))
              AM.Disp = Disp;
            else
              Reg = N.Val->getOperand(0);
          } else {
            Reg = N.Val->getOperand(0);
          }

          AM.IndexReg = AM.Base.Reg = Reg;
          return false;
        }
    }
    break;

  case ISD::ADD:
    if (!Available) {
      X86ISelAddressMode Backup = AM;
      if (!MatchAddress(N.Val->getOperand(0), AM, false, Depth+1) &&
          !MatchAddress(N.Val->getOperand(1), AM, false, Depth+1))
        return false;
      AM = Backup;
      if (!MatchAddress(N.Val->getOperand(1), AM, false, Depth+1) &&
          !MatchAddress(N.Val->getOperand(0), AM, false, Depth+1))
        return false;
      AM = Backup;
    }
    break;

  case ISD::OR:
    // Handle "X | C" as "X + C" iff X is known to have C bits clear.
    if (!Available) {
      if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
        X86ISelAddressMode Backup = AM;
        // Start with the LHS as an addr mode.
        if (!MatchAddress(N.getOperand(0), AM, false) &&
            // Address could not have picked a GV address for the displacement.
            AM.GV == NULL &&
            // On x86-64, the resultant disp must fit in 32-bits.
            isInt32(AM.Disp + CN->getSignExtended()) &&
            // Check to see if the LHS & C is zero.
            TLI.MaskedValueIsZero(N.getOperand(0), CN->getValue())) {
          AM.Disp += CN->getValue();
          return false;
        }
        AM = Backup;
      }
    }
    break;
  }

  // Is the base register already occupied?
  if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
    // If so, check to see if the scale index register is set.
    if (AM.IndexReg.Val == 0) {
      AM.IndexReg = N;
      AM.Scale = 1;
      return false;
    }

    // Otherwise, we cannot select it.
    return true;
  }

  // Default, generate it as a register.
  AM.BaseType = X86ISelAddressMode::RegBase;
  AM.Base.Reg = N;
  return false;
}

/// SelectAddr - returns true if it is able pattern match an addressing mode.
/// It returns the operands which make up the maximal addressing mode it can
/// match by reference.
bool X86DAGToDAGISel::SelectAddr(SDOperand Op, SDOperand N, SDOperand &Base,
                                 SDOperand &Scale, SDOperand &Index,
                                 SDOperand &Disp) {
  X86ISelAddressMode AM;
  if (MatchAddress(N, AM))
    return false;

  MVT::ValueType VT = N.getValueType();
  if (AM.BaseType == X86ISelAddressMode::RegBase) {
    if (!AM.Base.Reg.Val)
      AM.Base.Reg = CurDAG->getRegister(0, VT);
  }

  if (!AM.IndexReg.Val)
    AM.IndexReg = CurDAG->getRegister(0, VT);

  getAddressOperands(AM, Base, Scale, Index, Disp);
  return true;
}

/// isZeroNode - Returns true if Elt is a constant zero or a floating point
/// constant +0.0.
static inline bool isZeroNode(SDOperand Elt) {
  return ((isa<ConstantSDNode>(Elt) &&
  cast<ConstantSDNode>(Elt)->getValue() == 0) ||
  (isa<ConstantFPSDNode>(Elt) &&
  cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
}


/// SelectScalarSSELoad - Match a scalar SSE load.  In particular, we want to
/// match a load whose top elements are either undef or zeros.  The load flavor
/// is derived from the type of N, which is either v4f32 or v2f64.
bool X86DAGToDAGISel::SelectScalarSSELoad(SDOperand Op, SDOperand Pred,
                                          SDOperand N, SDOperand &Base,
                                          SDOperand &Scale, SDOperand &Index,
                                          SDOperand &Disp, SDOperand &InChain,
                                          SDOperand &OutChain) {
  if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
    InChain = N.getOperand(0).getValue(1);
    if (ISD::isNON_EXTLoad(InChain.Val) &&
        InChain.getValue(0).hasOneUse() &&
        N.hasOneUse() &&
        CanBeFoldedBy(N.Val, Pred.Val, Op.Val)) {
      LoadSDNode *LD = cast<LoadSDNode>(InChain);
      if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
        return false;
      OutChain = LD->getChain();
      return true;
    }
  }

  // Also handle the case where we explicitly require zeros in the top
  // elements.  This is a vector shuffle from the zero vector.
  if (N.getOpcode() == ISD::VECTOR_SHUFFLE && N.Val->hasOneUse() &&
      N.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
      N.getOperand(1).getOpcode() == ISD::SCALAR_TO_VECTOR && 
      N.getOperand(1).Val->hasOneUse() &&
      ISD::isNON_EXTLoad(N.getOperand(1).getOperand(0).Val) &&
      N.getOperand(1).getOperand(0).hasOneUse()) {
    // Check to see if the BUILD_VECTOR is building a zero vector.
    SDOperand BV = N.getOperand(0);
    for (unsigned i = 0, e = BV.getNumOperands(); i != e; ++i)
      if (!isZeroNode(BV.getOperand(i)) &&
          BV.getOperand(i).getOpcode() != ISD::UNDEF)
        return false;  // Not a zero/undef vector.
    // Check to see if the shuffle mask is 4/L/L/L or 2/L, where L is something
    // from the LHS.
    unsigned VecWidth = BV.getNumOperands();
    SDOperand ShufMask = N.getOperand(2);
    assert(ShufMask.getOpcode() == ISD::BUILD_VECTOR && "Invalid shuf mask!");
    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(ShufMask.getOperand(0))) {
      if (C->getValue() == VecWidth) {
        for (unsigned i = 1; i != VecWidth; ++i) {
          if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF) {
            // ok.
          } else {
            ConstantSDNode *C = cast<ConstantSDNode>(ShufMask.getOperand(i));
            if (C->getValue() >= VecWidth) return false;
          }
        }
      }
      
      // Okay, this is a zero extending load.  Fold it.
      LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(1).getOperand(0));
      if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
        return false;
      OutChain = LD->getChain();
      InChain = SDOperand(LD, 1);
      return true;
    }
  }
  return false;
}


/// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
/// mode it matches can be cost effectively emitted as an LEA instruction.
bool X86DAGToDAGISel::SelectLEAAddr(SDOperand Op, SDOperand N,
                                    SDOperand &Base, SDOperand &Scale,
                                    SDOperand &Index, SDOperand &Disp) {
  X86ISelAddressMode AM;
  if (MatchAddress(N, AM))
    return false;

  MVT::ValueType VT = N.getValueType();
  unsigned Complexity = 0;
  if (AM.BaseType == X86ISelAddressMode::RegBase)
    if (AM.Base.Reg.Val)
      Complexity = 1;
    else
      AM.Base.Reg = CurDAG->getRegister(0, VT);
  else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
    Complexity = 4;

  if (AM.IndexReg.Val)
    Complexity++;
  else
    AM.IndexReg = CurDAG->getRegister(0, VT);

  // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
  // a simple shift.
  if (AM.Scale > 1)
    Complexity++;

  // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
  // to a LEA. This is determined with some expermentation but is by no means
  // optimal (especially for code size consideration). LEA is nice because of
  // its three-address nature. Tweak the cost function again when we can run
  // convertToThreeAddress() at register allocation time.
  if (AM.GV || AM.CP || AM.ES || AM.JT != -1) {
    // For X86-64, we should always use lea to materialize RIP relative
    // addresses.
    if (Subtarget->is64Bit())
      Complexity = 4;
    else
      Complexity += 2;
  }

  if (AM.Disp && (AM.Base.Reg.Val || AM.IndexReg.Val))
    Complexity++;

  if (Complexity > 2) {
    getAddressOperands(AM, Base, Scale, Index, Disp);
    return true;
  }
  return false;
}

bool X86DAGToDAGISel::TryFoldLoad(SDOperand P, SDOperand N,
                                  SDOperand &Base, SDOperand &Scale,
                                  SDOperand &Index, SDOperand &Disp) {
  if (ISD::isNON_EXTLoad(N.Val) &&
      N.hasOneUse() &&
      CanBeFoldedBy(N.Val, P.Val, P.Val))
    return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp);
  return false;
}

/// getGlobalBaseReg - Output the instructions required to put the
/// base address to use for accessing globals into a register.
///
SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
  assert(!Subtarget->is64Bit() && "X86-64 PIC uses RIP relative addressing");
  if (!GlobalBaseReg) {
    // Insert the set of GlobalBaseReg into the first MBB of the function
    MachineBasicBlock &FirstMBB = BB->getParent()->front();
    MachineBasicBlock::iterator MBBI = FirstMBB.begin();
    SSARegMap *RegMap = BB->getParent()->getSSARegMap();
    unsigned PC = RegMap->createVirtualRegister(X86::GR32RegisterClass);
    
    const TargetInstrInfo *TII = TM.getInstrInfo();
    BuildMI(FirstMBB, MBBI, TII->get(X86::MovePCtoStack));
    BuildMI(FirstMBB, MBBI, TII->get(X86::POP32r), PC);
    
    // If we're using vanilla 'GOT' PIC style, we should use relative addressing
    // not to pc, but to _GLOBAL_ADDRESS_TABLE_ external
    if (TM.getRelocationModel() == Reloc::PIC_ &&
        Subtarget->isPICStyleGOT()) {
      GlobalBaseReg = RegMap->createVirtualRegister(X86::GR32RegisterClass);
      BuildMI(FirstMBB, MBBI, TII->get(X86::ADD32ri), GlobalBaseReg).
        addReg(PC).
        addExternalSymbol("_GLOBAL_OFFSET_TABLE_");
    } else {
      GlobalBaseReg = PC;
    }
    
  }
  return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).Val;
}

static SDNode *FindCallStartFromCall(SDNode *Node) {
  if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
    assert(Node->getOperand(0).getValueType() == MVT::Other &&
         "Node doesn't have a token chain argument!");
  return FindCallStartFromCall(Node->getOperand(0).Val);
}

SDNode *X86DAGToDAGISel::Select(SDOperand N) {
  SDNode *Node = N.Val;
  MVT::ValueType NVT = Node->getValueType(0);
  unsigned Opc, MOpc;
  unsigned Opcode = Node->getOpcode();

#ifndef NDEBUG
  DOUT << std::string(Indent, ' ') << "Selecting: ";
  DEBUG(Node->dump(CurDAG));
  DOUT << "\n";
  Indent += 2;
#endif

  if (Opcode >= ISD::BUILTIN_OP_END && Opcode < X86ISD::FIRST_NUMBER) {
#ifndef NDEBUG
    DOUT << std::string(Indent-2, ' ') << "== ";
    DEBUG(Node->dump(CurDAG));
    DOUT << "\n";
    Indent -= 2;
#endif
    return NULL;   // Already selected.
  }

  switch (Opcode) {
    default: break;
    case X86ISD::GlobalBaseReg: 
      return getGlobalBaseReg();

    case ISD::ADD: {
      // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
      // code and is matched first so to prevent it from being turned into
      // LEA32r X+c.
      // In 64-bit mode, use LEA to take advantage of RIP-relative addressing.
      MVT::ValueType PtrVT = TLI.getPointerTy();
      SDOperand N0 = N.getOperand(0);
      SDOperand N1 = N.getOperand(1);
      if (N.Val->getValueType(0) == PtrVT &&
          N0.getOpcode() == X86ISD::Wrapper &&
          N1.getOpcode() == ISD::Constant) {
        unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getValue();
        SDOperand C(0, 0);
        // TODO: handle ExternalSymbolSDNode.
        if (GlobalAddressSDNode *G =
            dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
          C = CurDAG->getTargetGlobalAddress(G->getGlobal(), PtrVT,
                                             G->getOffset() + Offset);
        } else if (ConstantPoolSDNode *CP =
                   dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
          C = CurDAG->getTargetConstantPool(CP->getConstVal(), PtrVT,
                                            CP->getAlignment(),
                                            CP->getOffset()+Offset);
        }

        if (C.Val) {
          if (Subtarget->is64Bit()) {
            SDOperand Ops[] = { CurDAG->getRegister(0, PtrVT), getI8Imm(1),
                                CurDAG->getRegister(0, PtrVT), C };
            return CurDAG->SelectNodeTo(N.Val, X86::LEA64r, MVT::i64, Ops, 4);
          } else
            return CurDAG->SelectNodeTo(N.Val, X86::MOV32ri, PtrVT, C);
        }
      }

      // Other cases are handled by auto-generated code.
      break;
    }

    case ISD::MULHU:
    case ISD::MULHS: {
      if (Opcode == ISD::MULHU)
        switch (NVT) {
        default: assert(0 && "Unsupported VT!");
        case MVT::i8:  Opc = X86::MUL8r;  MOpc = X86::MUL8m;  break;
        case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
        case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
        case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
        }
      else
        switch (NVT) {
        default: assert(0 && "Unsupported VT!");
        case MVT::i8:  Opc = X86::IMUL8r;  MOpc = X86::IMUL8m;  break;
        case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
        case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
        case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
        }

      unsigned LoReg, HiReg;
      switch (NVT) {
      default: assert(0 && "Unsupported VT!");
      case MVT::i8:  LoReg = X86::AL;  HiReg = X86::AH;  break;
      case MVT::i16: LoReg = X86::AX;  HiReg = X86::DX;  break;
      case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
      case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
      }

      SDOperand N0 = Node->getOperand(0);
      SDOperand N1 = Node->getOperand(1);

      bool foldedLoad = false;
      SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
      foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
      // MULHU and MULHS are commmutative
      if (!foldedLoad) {
        foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
        if (foldedLoad) {
          N0 = Node->getOperand(1);
          N1 = Node->getOperand(0);
        }
      }

      SDOperand Chain;
      if (foldedLoad) {
        Chain = N1.getOperand(0);
        AddToISelQueue(Chain);
      } else
        Chain = CurDAG->getEntryNode();

      SDOperand InFlag(0, 0);
      AddToISelQueue(N0);
      Chain  = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
                                    N0, InFlag);
      InFlag = Chain.getValue(1);

      if (foldedLoad) {
        AddToISelQueue(Tmp0);
        AddToISelQueue(Tmp1);
        AddToISelQueue(Tmp2);
        AddToISelQueue(Tmp3);
        SDOperand Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Chain, InFlag };
        SDNode *CNode =
          CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
        Chain  = SDOperand(CNode, 0);
        InFlag = SDOperand(CNode, 1);
      } else {
        AddToISelQueue(N1);
        InFlag =
          SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
      }

      SDOperand Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
      ReplaceUses(N.getValue(0), Result);
      if (foldedLoad)
        ReplaceUses(N1.getValue(1), Result.getValue(1));

#ifndef NDEBUG
      DOUT << std::string(Indent-2, ' ') << "=> ";
      DEBUG(Result.Val->dump(CurDAG));
      DOUT << "\n";
      Indent -= 2;
#endif
      return NULL;
    }
      
    case ISD::SDIV:
    case ISD::UDIV:
    case ISD::SREM:
    case ISD::UREM: {
      bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
      bool isDiv    = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
      if (!isSigned)
        switch (NVT) {
        default: assert(0 && "Unsupported VT!");
        case MVT::i8:  Opc = X86::DIV8r;  MOpc = X86::DIV8m;  break;
        case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
        case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
        case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
        }
      else
        switch (NVT) {
        default: assert(0 && "Unsupported VT!");
        case MVT::i8:  Opc = X86::IDIV8r;  MOpc = X86::IDIV8m;  break;
        case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
        case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
        case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
        }

      unsigned LoReg, HiReg;
      unsigned ClrOpcode, SExtOpcode;
      switch (NVT) {
      default: assert(0 && "Unsupported VT!");
      case MVT::i8:
        LoReg = X86::AL;  HiReg = X86::AH;
        ClrOpcode  = 0;
        SExtOpcode = X86::CBW;
        break;
      case MVT::i16:
        LoReg = X86::AX;  HiReg = X86::DX;
        ClrOpcode  = X86::MOV16r0;
        SExtOpcode = X86::CWD;
        break;
      case MVT::i32:
        LoReg = X86::EAX; HiReg = X86::EDX;
        ClrOpcode  = X86::MOV32r0;
        SExtOpcode = X86::CDQ;
        break;
      case MVT::i64:
        LoReg = X86::RAX; HiReg = X86::RDX;
        ClrOpcode  = X86::MOV64r0;
        SExtOpcode = X86::CQO;
        break;
      }

      SDOperand N0 = Node->getOperand(0);
      SDOperand N1 = Node->getOperand(1);
      SDOperand InFlag(0, 0);
      if (NVT == MVT::i8 && !isSigned) {
        // Special case for div8, just use a move with zero extension to AX to
        // clear the upper 8 bits (AH).
        SDOperand Tmp0, Tmp1, Tmp2, Tmp3, Move, Chain;
        if (TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3)) {
          SDOperand Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N0.getOperand(0) };
          AddToISelQueue(N0.getOperand(0));
          AddToISelQueue(Tmp0);
          AddToISelQueue(Tmp1);
          AddToISelQueue(Tmp2);
          AddToISelQueue(Tmp3);
          Move =
            SDOperand(CurDAG->getTargetNode(X86::MOVZX16rm8, MVT::i16, MVT::Other,
                                            Ops, 5), 0);
          Chain = Move.getValue(1);
          ReplaceUses(N0.getValue(1), Chain);
        } else {
          AddToISelQueue(N0);
          Move =
            SDOperand(CurDAG->getTargetNode(X86::MOVZX16rr8, MVT::i16, N0), 0);
          Chain = CurDAG->getEntryNode();
        }
        Chain  = CurDAG->getCopyToReg(Chain, X86::AX, Move, InFlag);
        InFlag = Chain.getValue(1);
      } else {
        AddToISelQueue(N0);
        InFlag =
          CurDAG->getCopyToReg(CurDAG->getEntryNode(), LoReg, N0,
                               InFlag).getValue(1);
        if (isSigned) {
          // Sign extend the low part into the high part.
          InFlag =
            SDOperand(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
        } else {
          // Zero out the high part, effectively zero extending the input.
          SDOperand ClrNode = SDOperand(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
          InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), HiReg, ClrNode,
                                        InFlag).getValue(1);
        }
      }

      SDOperand Tmp0, Tmp1, Tmp2, Tmp3, Chain;
      bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
      if (foldedLoad) {
        AddToISelQueue(N1.getOperand(0));
        AddToISelQueue(Tmp0);
        AddToISelQueue(Tmp1);
        AddToISelQueue(Tmp2);
        AddToISelQueue(Tmp3);
        SDOperand Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N1.getOperand(0), InFlag };
        SDNode *CNode =
          CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
        Chain  = SDOperand(CNode, 0);
        InFlag = SDOperand(CNode, 1);
      } else {
        AddToISelQueue(N1);
        Chain = CurDAG->getEntryNode();
        InFlag =
          SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
      }

      SDOperand Result =
        CurDAG->getCopyFromReg(Chain, isDiv ? LoReg : HiReg, NVT, InFlag);
      ReplaceUses(N.getValue(0), Result);
      if (foldedLoad)
        ReplaceUses(N1.getValue(1), Result.getValue(1));

#ifndef NDEBUG
      DOUT << std::string(Indent-2, ' ') << "=> ";
      DEBUG(Result.Val->dump(CurDAG));
      DOUT << "\n";
      Indent -= 2;
#endif

      return NULL;
    }

    case ISD::TRUNCATE: {
      if (!Subtarget->is64Bit() && NVT == MVT::i8) {
        unsigned Opc2;
        MVT::ValueType VT;
        switch (Node->getOperand(0).getValueType()) {
        default: assert(0 && "Unknown truncate!");
        case MVT::i16:
          Opc = X86::MOV16to16_;
          VT = MVT::i16;
          Opc2 = X86::TRUNC_16_to8;
          break;
        case MVT::i32:
          Opc = X86::MOV32to32_;
          VT = MVT::i32;
          Opc2 = X86::TRUNC_32_to8;
          break;
        }

        AddToISelQueue(Node->getOperand(0));
        SDOperand Tmp =
          SDOperand(CurDAG->getTargetNode(Opc, VT, Node->getOperand(0)), 0);
        SDNode *ResNode = CurDAG->getTargetNode(Opc2, NVT, Tmp);
      
#ifndef NDEBUG
        DOUT << std::string(Indent-2, ' ') << "=> ";
        DEBUG(ResNode->dump(CurDAG));
        DOUT << "\n";
        Indent -= 2;
#endif
        return ResNode;
      }

      break;
    }
  }

  SDNode *ResNode = SelectCode(N);

#ifndef NDEBUG
  DOUT << std::string(Indent-2, ' ') << "=> ";
  if (ResNode == NULL || ResNode == N.Val)
    DEBUG(N.Val->dump(CurDAG));
  else
    DEBUG(ResNode->dump(CurDAG));
  DOUT << "\n";
  Indent -= 2;
#endif

  return ResNode;
}

bool X86DAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDOperand &Op, char ConstraintCode,
                             std::vector<SDOperand> &OutOps, SelectionDAG &DAG){
  SDOperand Op0, Op1, Op2, Op3;
  switch (ConstraintCode) {
  case 'o':   // offsetable        ??
  case 'v':   // not offsetable    ??
  default: return true;
  case 'm':   // memory
    if (!SelectAddr(Op, Op, Op0, Op1, Op2, Op3))
      return true;
    break;
  }
  
  OutOps.push_back(Op0);
  OutOps.push_back(Op1);
  OutOps.push_back(Op2);
  OutOps.push_back(Op3);
  AddToISelQueue(Op0);
  AddToISelQueue(Op1);
  AddToISelQueue(Op2);
  AddToISelQueue(Op3);
  return false;
}

/// createX86ISelDag - This pass converts a legalized DAG into a 
/// X86-specific DAG, ready for instruction scheduling.
///
FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM, bool Fast) {
  return new X86DAGToDAGISel(TM, Fast);
}