summaryrefslogtreecommitdiff
path: root/lib/Transforms/Scalar/LoopStrengthReduce.cpp
blob: 5f9562b6b3f785ed3ae7fccbe76144e06fc2f509 (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
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
//===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file was developed by Nate Begeman and is distributed under the
// University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass performs a strength reduction on array references inside loops that
// have as one or more of their components the loop induction variable.  This is
// accomplished by creating a new Value to hold the initial value of the array
// access for the first iteration, and then creating a new GEP instruction in
// the loop to increment the value by the appropriate amount.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "loop-reduce"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#include "llvm/Type.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Target/TargetLowering.h"
#include <algorithm>
#include <set>
using namespace llvm;

STATISTIC(NumReduced , "Number of GEPs strength reduced");
STATISTIC(NumInserted, "Number of PHIs inserted");
STATISTIC(NumVariable, "Number of PHIs with variable strides");

namespace {

  struct BasedUser;

  /// IVStrideUse - Keep track of one use of a strided induction variable, where
  /// the stride is stored externally.  The Offset member keeps track of the 
  /// offset from the IV, User is the actual user of the operand, and 'Operand'
  /// is the operand # of the User that is the use.
  struct VISIBILITY_HIDDEN IVStrideUse {
    SCEVHandle Offset;
    Instruction *User;
    Value *OperandValToReplace;

    // isUseOfPostIncrementedValue - True if this should use the
    // post-incremented version of this IV, not the preincremented version.
    // This can only be set in special cases, such as the terminating setcc
    // instruction for a loop or uses dominated by the loop.
    bool isUseOfPostIncrementedValue;
    
    IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
      : Offset(Offs), User(U), OperandValToReplace(O),
        isUseOfPostIncrementedValue(false) {}
  };
  
  /// IVUsersOfOneStride - This structure keeps track of all instructions that
  /// have an operand that is based on the trip count multiplied by some stride.
  /// The stride for all of these users is common and kept external to this
  /// structure.
  struct VISIBILITY_HIDDEN IVUsersOfOneStride {
    /// Users - Keep track of all of the users of this stride as well as the
    /// initial value and the operand that uses the IV.
    std::vector<IVStrideUse> Users;
    
    void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
      Users.push_back(IVStrideUse(Offset, User, Operand));
    }
  };

  /// IVInfo - This structure keeps track of one IV expression inserted during
  /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
  /// well as the PHI node and increment value created for rewrite.
  struct VISIBILITY_HIDDEN IVExpr {
    SCEVHandle  Stride;
    SCEVHandle  Base;
    PHINode    *PHI;
    Value      *IncV;

    IVExpr()
      : Stride(SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)),
        Base  (SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)) {}
    IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
           Value *incv)
      : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
  };

  /// IVsOfOneStride - This structure keeps track of all IV expression inserted
  /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
  struct VISIBILITY_HIDDEN IVsOfOneStride {
    std::vector<IVExpr> IVs;

    void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
               Value *IncV) {
      IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
    }
  };

  class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
    LoopInfo *LI;
    ETForest *EF;
    ScalarEvolution *SE;
    const TargetData *TD;
    const Type *UIntPtrTy;
    bool Changed;

    /// IVUsesByStride - Keep track of all uses of induction variables that we
    /// are interested in.  The key of the map is the stride of the access.
    std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;

    /// IVsByStride - Keep track of all IVs that have been inserted for a
    /// particular stride.
    std::map<SCEVHandle, IVsOfOneStride> IVsByStride;

    /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
    /// We use this to iterate over the IVUsesByStride collection without being
    /// dependent on random ordering of pointers in the process.
    std::vector<SCEVHandle> StrideOrder;

    /// CastedValues - As we need to cast values to uintptr_t, this keeps track
    /// of the casted version of each value.  This is accessed by
    /// getCastedVersionOf.
    std::map<Value*, Value*> CastedPointers;

    /// DeadInsts - Keep track of instructions we may have made dead, so that
    /// we can remove them after we are done working.
    std::set<Instruction*> DeadInsts;

    /// TLI - Keep a pointer of a TargetLowering to consult for determining
    /// transformation profitability.
    const TargetLowering *TLI;

  public:
    static const int ID; // Pass ID, replacement for typeid
    LoopStrengthReduce(const TargetLowering *tli = NULL) : 
      LoopPass((intptr_t)&ID), TLI(tli) {
    }

    bool runOnLoop(Loop *L, LPPassManager &LPM);

    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
      // We split critical edges, so we change the CFG.  However, we do update
      // many analyses if they are around.
      AU.addPreservedID(LoopSimplifyID);
      AU.addPreserved<LoopInfo>();
      AU.addPreserved<ETForest>();
      AU.addPreserved<DominanceFrontier>();
      AU.addPreserved<DominatorTree>();

      AU.addRequiredID(LoopSimplifyID);
      AU.addRequired<LoopInfo>();
      AU.addRequired<ETForest>();
      AU.addRequired<TargetData>();
      AU.addRequired<ScalarEvolution>();
    }
    
    /// getCastedVersionOf - Return the specified value casted to uintptr_t.
    ///
    Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
private:
    bool AddUsersIfInteresting(Instruction *I, Loop *L,
                               std::set<Instruction*> &Processed);
    SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);

    void OptimizeIndvars(Loop *L);
    bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
                       const SCEVHandle *&CondStride);

    unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*,
                             const std::vector<BasedUser>& UsersToProcess);

    bool ValidStride(int64_t, const std::vector<BasedUser>& UsersToProcess);

    void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
                                      IVUsersOfOneStride &Uses,
                                      Loop *L, bool isOnlyStride);
    void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
  };
  const int LoopStrengthReduce::ID = 0;
  RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
}

LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
  return new LoopStrengthReduce(TLI);
}

/// getCastedVersionOf - Return the specified value casted to uintptr_t. This
/// assumes that the Value* V is of integer or pointer type only.
///
Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode, 
                                              Value *V) {
  if (V->getType() == UIntPtrTy) return V;
  if (Constant *CB = dyn_cast<Constant>(V))
    return ConstantExpr::getCast(opcode, CB, UIntPtrTy);

  Value *&New = CastedPointers[V];
  if (New) return New;
  
  New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
  DeadInsts.insert(cast<Instruction>(New));
  return New;
}


/// DeleteTriviallyDeadInstructions - If any of the instructions is the
/// specified set are trivially dead, delete them and see if this makes any of
/// their operands subsequently dead.
void LoopStrengthReduce::
DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
  while (!Insts.empty()) {
    Instruction *I = *Insts.begin();
    Insts.erase(Insts.begin());
    if (isInstructionTriviallyDead(I)) {
      for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
        if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
          Insts.insert(U);
      SE->deleteInstructionFromRecords(I);
      I->eraseFromParent();
      Changed = true;
    }
  }
}


/// GetExpressionSCEV - Compute and return the SCEV for the specified
/// instruction.
SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
  // Pointer to pointer bitcast instructions return the same value as their
  // operand.
  if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
    if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
      return SE->getSCEV(BCI);
    SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)), L);
    SE->setSCEV(BCI, R);
    return R;
  }

  // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
  // If this is a GEP that SE doesn't know about, compute it now and insert it.
  // If this is not a GEP, or if we have already done this computation, just let
  // SE figure it out.
  GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
  if (!GEP || SE->hasSCEV(GEP))
    return SE->getSCEV(Exp);
    
  // Analyze all of the subscripts of this getelementptr instruction, looking
  // for uses that are determined by the trip count of L.  First, skip all
  // operands the are not dependent on the IV.

  // Build up the base expression.  Insert an LLVM cast of the pointer to
  // uintptr_t first.
  SCEVHandle GEPVal = SCEVUnknown::get(
      getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));

  gep_type_iterator GTI = gep_type_begin(GEP);
  
  for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
    // If this is a use of a recurrence that we can analyze, and it comes before
    // Op does in the GEP operand list, we will handle this when we process this
    // operand.
    if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
      const StructLayout *SL = TD->getStructLayout(STy);
      unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
      uint64_t Offset = SL->getElementOffset(Idx);
      GEPVal = SCEVAddExpr::get(GEPVal,
                                SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
    } else {
      unsigned GEPOpiBits = 
        GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
      unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
      Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ? 
          Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
            Instruction::BitCast));
      Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
      SCEVHandle Idx = SE->getSCEV(OpVal);

      uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
      if (TypeSize != 1)
        Idx = SCEVMulExpr::get(Idx,
                               SCEVConstant::get(ConstantInt::get(UIntPtrTy,
                                                                   TypeSize)));
      GEPVal = SCEVAddExpr::get(GEPVal, Idx);
    }
  }

  SE->setSCEV(GEP, GEPVal);
  return GEPVal;
}

/// getSCEVStartAndStride - Compute the start and stride of this expression,
/// returning false if the expression is not a start/stride pair, or true if it
/// is.  The stride must be a loop invariant expression, but the start may be
/// a mix of loop invariant and loop variant expressions.
static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
                                  SCEVHandle &Start, SCEVHandle &Stride) {
  SCEVHandle TheAddRec = Start;   // Initialize to zero.

  // If the outer level is an AddExpr, the operands are all start values except
  // for a nested AddRecExpr.
  if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
    for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
      if (SCEVAddRecExpr *AddRec =
             dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
        if (AddRec->getLoop() == L)
          TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
        else
          return false;  // Nested IV of some sort?
      } else {
        Start = SCEVAddExpr::get(Start, AE->getOperand(i));
      }
        
  } else if (isa<SCEVAddRecExpr>(SH)) {
    TheAddRec = SH;
  } else {
    return false;  // not analyzable.
  }
  
  SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
  if (!AddRec || AddRec->getLoop() != L) return false;
  
  // FIXME: Generalize to non-affine IV's.
  if (!AddRec->isAffine()) return false;

  Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
  
  if (!isa<SCEVConstant>(AddRec->getOperand(1)))
    DOUT << "[" << L->getHeader()->getName()
         << "] Variable stride: " << *AddRec << "\n";

  Stride = AddRec->getOperand(1);
  return true;
}

/// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
/// and now we need to decide whether the user should use the preinc or post-inc
/// value.  If this user should use the post-inc version of the IV, return true.
///
/// Choosing wrong here can break dominance properties (if we choose to use the
/// post-inc value when we cannot) or it can end up adding extra live-ranges to
/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
/// should use the post-inc value).
static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
                                       Loop *L, ETForest *EF, Pass *P) {
  // If the user is in the loop, use the preinc value.
  if (L->contains(User->getParent())) return false;
  
  BasicBlock *LatchBlock = L->getLoopLatch();
  
  // Ok, the user is outside of the loop.  If it is dominated by the latch
  // block, use the post-inc value.
  if (EF->dominates(LatchBlock, User->getParent()))
    return true;

  // There is one case we have to be careful of: PHI nodes.  These little guys
  // can live in blocks that do not dominate the latch block, but (since their
  // uses occur in the predecessor block, not the block the PHI lives in) should
  // still use the post-inc value.  Check for this case now.
  PHINode *PN = dyn_cast<PHINode>(User);
  if (!PN) return false;  // not a phi, not dominated by latch block.
  
  // Look at all of the uses of IV by the PHI node.  If any use corresponds to
  // a block that is not dominated by the latch block, give up and use the
  // preincremented value.
  unsigned NumUses = 0;
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    if (PN->getIncomingValue(i) == IV) {
      ++NumUses;
      if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
        return false;
    }

  // Okay, all uses of IV by PN are in predecessor blocks that really are
  // dominated by the latch block.  Split the critical edges and use the
  // post-incremented value.
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    if (PN->getIncomingValue(i) == IV) {
      SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
                        true);
      // Splitting the critical edge can reduce the number of entries in this
      // PHI.
      e = PN->getNumIncomingValues();
      if (--NumUses == 0) break;
    }
  
  return true;
}

  

/// AddUsersIfInteresting - Inspect the specified instruction.  If it is a
/// reducible SCEV, recursively add its users to the IVUsesByStride set and
/// return true.  Otherwise, return false.
bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
                                            std::set<Instruction*> &Processed) {
  if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
      return false;   // Void and FP expressions cannot be reduced.
  if (!Processed.insert(I).second)
    return true;    // Instruction already handled.
  
  // Get the symbolic expression for this instruction.
  SCEVHandle ISE = GetExpressionSCEV(I, L);
  if (isa<SCEVCouldNotCompute>(ISE)) return false;
  
  // Get the start and stride for this expression.
  SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
  SCEVHandle Stride = Start;
  if (!getSCEVStartAndStride(ISE, L, Start, Stride))
    return false;  // Non-reducible symbolic expression, bail out.

  std::vector<Instruction *> IUsers;
  // Collect all I uses now because IVUseShouldUsePostIncValue may 
  // invalidate use_iterator.
  for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
    IUsers.push_back(cast<Instruction>(*UI));

  for (unsigned iused_index = 0, iused_size = IUsers.size(); 
       iused_index != iused_size; ++iused_index) {

    Instruction *User = IUsers[iused_index];

    // Do not infinitely recurse on PHI nodes.
    if (isa<PHINode>(User) && Processed.count(User))
      continue;

    // If this is an instruction defined in a nested loop, or outside this loop,
    // don't recurse into it.
    bool AddUserToIVUsers = false;
    if (LI->getLoopFor(User->getParent()) != L) {
      DOUT << "FOUND USER in other loop: " << *User
           << "   OF SCEV: " << *ISE << "\n";
      AddUserToIVUsers = true;
    } else if (!AddUsersIfInteresting(User, L, Processed)) {
      DOUT << "FOUND USER: " << *User
           << "   OF SCEV: " << *ISE << "\n";
      AddUserToIVUsers = true;
    }

    if (AddUserToIVUsers) {
      IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
      if (StrideUses.Users.empty())     // First occurance of this stride?
        StrideOrder.push_back(Stride);
      
      // Okay, we found a user that we cannot reduce.  Analyze the instruction
      // and decide what to do with it.  If we are a use inside of the loop, use
      // the value before incrementation, otherwise use it after incrementation.
      if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
        // The value used will be incremented by the stride more than we are
        // expecting, so subtract this off.
        SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
        StrideUses.addUser(NewStart, User, I);
        StrideUses.Users.back().isUseOfPostIncrementedValue = true;
        DOUT << "   USING POSTINC SCEV, START=" << *NewStart<< "\n";
      } else {        
        StrideUses.addUser(Start, User, I);
      }
    }
  }
  return true;
}

namespace {
  /// BasedUser - For a particular base value, keep information about how we've
  /// partitioned the expression so far.
  struct BasedUser {
    /// Base - The Base value for the PHI node that needs to be inserted for
    /// this use.  As the use is processed, information gets moved from this
    /// field to the Imm field (below).  BasedUser values are sorted by this
    /// field.
    SCEVHandle Base;
    
    /// Inst - The instruction using the induction variable.
    Instruction *Inst;

    /// OperandValToReplace - The operand value of Inst to replace with the
    /// EmittedBase.
    Value *OperandValToReplace;

    /// Imm - The immediate value that should be added to the base immediately
    /// before Inst, because it will be folded into the imm field of the
    /// instruction.
    SCEVHandle Imm;

    /// EmittedBase - The actual value* to use for the base value of this
    /// operation.  This is null if we should just use zero so far.
    Value *EmittedBase;

    // isUseOfPostIncrementedValue - True if this should use the
    // post-incremented version of this IV, not the preincremented version.
    // This can only be set in special cases, such as the terminating setcc
    // instruction for a loop and uses outside the loop that are dominated by
    // the loop.
    bool isUseOfPostIncrementedValue;
    
    BasedUser(IVStrideUse &IVSU)
      : Base(IVSU.Offset), Inst(IVSU.User), 
        OperandValToReplace(IVSU.OperandValToReplace), 
        Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
        isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}

    // Once we rewrite the code to insert the new IVs we want, update the
    // operands of Inst to use the new expression 'NewBase', with 'Imm' added
    // to it.
    void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
                                        SCEVExpander &Rewriter, Loop *L,
                                        Pass *P);
    
    Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase, 
                                       SCEVExpander &Rewriter,
                                       Instruction *IP, Loop *L);
    void dump() const;
  };
}

void BasedUser::dump() const {
  cerr << " Base=" << *Base;
  cerr << " Imm=" << *Imm;
  if (EmittedBase)
    cerr << "  EB=" << *EmittedBase;

  cerr << "   Inst: " << *Inst;
}

Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase, 
                                              SCEVExpander &Rewriter,
                                              Instruction *IP, Loop *L) {
  // Figure out where we *really* want to insert this code.  In particular, if
  // the user is inside of a loop that is nested inside of L, we really don't
  // want to insert this expression before the user, we'd rather pull it out as
  // many loops as possible.
  LoopInfo &LI = Rewriter.getLoopInfo();
  Instruction *BaseInsertPt = IP;
  
  // Figure out the most-nested loop that IP is in.
  Loop *InsertLoop = LI.getLoopFor(IP->getParent());
  
  // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
  // the preheader of the outer-most loop where NewBase is not loop invariant.
  while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
    BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
    InsertLoop = InsertLoop->getParentLoop();
  }
  
  // If there is no immediate value, skip the next part.
  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
    if (SC->getValue()->isZero())
      return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
                                    OperandValToReplace->getType());

  Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
  
  // Always emit the immediate (if non-zero) into the same block as the user.
  SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
  return Rewriter.expandCodeFor(NewValSCEV, IP,
                                OperandValToReplace->getType());
}


// Once we rewrite the code to insert the new IVs we want, update the
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
// to it.
void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
                                               SCEVExpander &Rewriter,
                                               Loop *L, Pass *P) {
  if (!isa<PHINode>(Inst)) {
    // By default, insert code at the user instruction.
    BasicBlock::iterator InsertPt = Inst;
    
    // However, if the Operand is itself an instruction, the (potentially
    // complex) inserted code may be shared by many users.  Because of this, we
    // want to emit code for the computation of the operand right before its old
    // computation.  This is usually safe, because we obviously used to use the
    // computation when it was computed in its current block.  However, in some
    // cases (e.g. use of a post-incremented induction variable) the NewBase
    // value will be pinned to live somewhere after the original computation.
    // In this case, we have to back off.
    if (!isUseOfPostIncrementedValue) {
      if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) { 
        InsertPt = OpInst;
        while (isa<PHINode>(InsertPt)) ++InsertPt;
      }
    }
    
    Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
    // Replace the use of the operand Value with the new Phi we just created.
    Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
    DOUT << "    CHANGED: IMM =" << *Imm << "  Inst = " << *Inst;
    return;
  }
  
  // PHI nodes are more complex.  We have to insert one copy of the NewBase+Imm
  // expression into each operand block that uses it.  Note that PHI nodes can
  // have multiple entries for the same predecessor.  We use a map to make sure
  // that a PHI node only has a single Value* for each predecessor (which also
  // prevents us from inserting duplicate code in some blocks).
  std::map<BasicBlock*, Value*> InsertedCode;
  PHINode *PN = cast<PHINode>(Inst);
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
    if (PN->getIncomingValue(i) == OperandValToReplace) {
      // If this is a critical edge, split the edge so that we do not insert the
      // code on all predecessor/successor paths.  We do this unless this is the
      // canonical backedge for this loop, as this can make some inserted code
      // be in an illegal position.
      BasicBlock *PHIPred = PN->getIncomingBlock(i);
      if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
          (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
        
        // First step, split the critical edge.
        SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
            
        // Next step: move the basic block.  In particular, if the PHI node
        // is outside of the loop, and PredTI is in the loop, we want to
        // move the block to be immediately before the PHI block, not
        // immediately after PredTI.
        if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
          BasicBlock *NewBB = PN->getIncomingBlock(i);
          NewBB->moveBefore(PN->getParent());
        }
        
        // Splitting the edge can reduce the number of PHI entries we have.
        e = PN->getNumIncomingValues();
      }

      Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
      if (!Code) {
        // Insert the code into the end of the predecessor block.
        Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
        Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
      }
      
      // Replace the use of the operand Value with the new Phi we just created.
      PN->setIncomingValue(i, Code);
      Rewriter.clear();
    }
  }
  DOUT << "    CHANGED: IMM =" << *Imm << "  Inst = " << *Inst;
}


/// isTargetConstant - Return true if the following can be referenced by the
/// immediate field of a target instruction.
static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
                             const TargetLowering *TLI) {
  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
    int64_t VC = SC->getValue()->getSExtValue();
    if (TLI) {
      TargetLowering::AddrMode AM;
      AM.BaseOffs = VC;
      return TLI->isLegalAddressingMode(AM, UseTy);
    } else {
      // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
      return (VC > -(1 << 16) && VC < (1 << 16)-1);
    }
  }

  if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
      if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
        Constant *Op0 = CE->getOperand(0);
        if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
          TargetLowering::AddrMode AM;
          AM.BaseGV = GV;
          return TLI->isLegalAddressingMode(AM, UseTy);
        }
      }
  return false;
}

/// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
/// loop varying to the Imm operand.
static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
                                            Loop *L) {
  if (Val->isLoopInvariant(L)) return;  // Nothing to do.
  
  if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
    std::vector<SCEVHandle> NewOps;
    NewOps.reserve(SAE->getNumOperands());
    
    for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
      if (!SAE->getOperand(i)->isLoopInvariant(L)) {
        // If this is a loop-variant expression, it must stay in the immediate
        // field of the expression.
        Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
      } else {
        NewOps.push_back(SAE->getOperand(i));
      }

    if (NewOps.empty())
      Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
    else
      Val = SCEVAddExpr::get(NewOps);
  } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
    // Try to pull immediates out of the start value of nested addrec's.
    SCEVHandle Start = SARE->getStart();
    MoveLoopVariantsToImediateField(Start, Imm, L);
    
    std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
    Ops[0] = Start;
    Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
  } else {
    // Otherwise, all of Val is variant, move the whole thing over.
    Imm = SCEVAddExpr::get(Imm, Val);
    Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
  }
}


/// MoveImmediateValues - Look at Val, and pull out any additions of constants
/// that can fit into the immediate field of instructions in the target.
/// Accumulate these immediate values into the Imm value.
static void MoveImmediateValues(const TargetLowering *TLI,
                                Instruction *User,
                                SCEVHandle &Val, SCEVHandle &Imm,
                                bool isAddress, Loop *L) {
  const Type *UseTy = User->getType();
  if (StoreInst *SI = dyn_cast<StoreInst>(User))
    UseTy = SI->getOperand(0)->getType();

  if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
    std::vector<SCEVHandle> NewOps;
    NewOps.reserve(SAE->getNumOperands());
    
    for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
      SCEVHandle NewOp = SAE->getOperand(i);
      MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L);
      
      if (!NewOp->isLoopInvariant(L)) {
        // If this is a loop-variant expression, it must stay in the immediate
        // field of the expression.
        Imm = SCEVAddExpr::get(Imm, NewOp);
      } else {
        NewOps.push_back(NewOp);
      }
    }

    if (NewOps.empty())
      Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
    else
      Val = SCEVAddExpr::get(NewOps);
    return;
  } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
    // Try to pull immediates out of the start value of nested addrec's.
    SCEVHandle Start = SARE->getStart();
    MoveImmediateValues(TLI, User, Start, Imm, isAddress, L);
    
    if (Start != SARE->getStart()) {
      std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
      Ops[0] = Start;
      Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
    }
    return;
  } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
    // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
    if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
        SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {

      SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
      SCEVHandle NewOp = SME->getOperand(1);
      MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L);
      
      // If we extracted something out of the subexpressions, see if we can 
      // simplify this!
      if (NewOp != SME->getOperand(1)) {
        // Scale SubImm up by "8".  If the result is a target constant, we are
        // good.
        SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
        if (isTargetConstant(SubImm, UseTy, TLI)) {
          // Accumulate the immediate.
          Imm = SCEVAddExpr::get(Imm, SubImm);
          
          // Update what is left of 'Val'.
          Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
          return;
        }
      }
    }
  }

  // Loop-variant expressions must stay in the immediate field of the
  // expression.
  if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
      !Val->isLoopInvariant(L)) {
    Imm = SCEVAddExpr::get(Imm, Val);
    Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
    return;
  }

  // Otherwise, no immediates to move.
}


/// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
/// added together.  This is used to reassociate common addition subexprs
/// together for maximal sharing when rewriting bases.
static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
                             SCEVHandle Expr) {
  if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
    for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
      SeparateSubExprs(SubExprs, AE->getOperand(j));
  } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
    SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
    if (SARE->getOperand(0) == Zero) {
      SubExprs.push_back(Expr);
    } else {
      // Compute the addrec with zero as its base.
      std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
      Ops[0] = Zero;   // Start with zero base.
      SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
      

      SeparateSubExprs(SubExprs, SARE->getOperand(0));
    }
  } else if (!isa<SCEVConstant>(Expr) ||
             !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
    // Do not add zero.
    SubExprs.push_back(Expr);
  }
}


/// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
/// removing any common subexpressions from it.  Anything truly common is
/// removed, accumulated, and returned.  This looks for things like (a+b+c) and
/// (a+c+d) -> (a+c).  The common expression is *removed* from the Bases.
static SCEVHandle 
RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
  unsigned NumUses = Uses.size();

  // Only one use?  Use its base, regardless of what it is!
  SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
  SCEVHandle Result = Zero;
  if (NumUses == 1) {
    std::swap(Result, Uses[0].Base);
    return Result;
  }

  // To find common subexpressions, count how many of Uses use each expression.
  // If any subexpressions are used Uses.size() times, they are common.
  std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
  
  // UniqueSubExprs - Keep track of all of the subexpressions we see in the
  // order we see them.
  std::vector<SCEVHandle> UniqueSubExprs;

  std::vector<SCEVHandle> SubExprs;
  for (unsigned i = 0; i != NumUses; ++i) {
    // If the base is zero (which is common), return zero now, there are no
    // CSEs we can find.
    if (Uses[i].Base == Zero) return Zero;

    // Split the expression into subexprs.
    SeparateSubExprs(SubExprs, Uses[i].Base);
    // Add one to SubExpressionUseCounts for each subexpr present.
    for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
      if (++SubExpressionUseCounts[SubExprs[j]] == 1)
        UniqueSubExprs.push_back(SubExprs[j]);
    SubExprs.clear();
  }

  // Now that we know how many times each is used, build Result.  Iterate over
  // UniqueSubexprs so that we have a stable ordering.
  for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
    std::map<SCEVHandle, unsigned>::iterator I = 
       SubExpressionUseCounts.find(UniqueSubExprs[i]);
    assert(I != SubExpressionUseCounts.end() && "Entry not found?");
    if (I->second == NumUses) {  // Found CSE!
      Result = SCEVAddExpr::get(Result, I->first);
    } else {
      // Remove non-cse's from SubExpressionUseCounts.
      SubExpressionUseCounts.erase(I);
    }
  }
  
  // If we found no CSE's, return now.
  if (Result == Zero) return Result;
  
  // Otherwise, remove all of the CSE's we found from each of the base values.
  for (unsigned i = 0; i != NumUses; ++i) {
    // Split the expression into subexprs.
    SeparateSubExprs(SubExprs, Uses[i].Base);

    // Remove any common subexpressions.
    for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
      if (SubExpressionUseCounts.count(SubExprs[j])) {
        SubExprs.erase(SubExprs.begin()+j);
        --j; --e;
      }
    
    // Finally, the non-shared expressions together.
    if (SubExprs.empty())
      Uses[i].Base = Zero;
    else
      Uses[i].Base = SCEVAddExpr::get(SubExprs);
    SubExprs.clear();
  }
 
  return Result;
}

/// isZero - returns true if the scalar evolution expression is zero.
///
static bool isZero(SCEVHandle &V) {
  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
    return SC->getValue()->isZero();
  return false;
}

/// ValidStride - Check whether the given Scale is valid for all loads and 
/// stores in UsersToProcess.
///
bool LoopStrengthReduce::ValidStride(int64_t Scale, 
                               const std::vector<BasedUser>& UsersToProcess) {
  for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
    // If this is a load or other access, pass the type of the access in.
    const Type *AccessTy = Type::VoidTy;
    if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
      AccessTy = SI->getOperand(0)->getType();
    else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
      AccessTy = LI->getType();
    
    TargetLowering::AddrMode AM;
    if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
      AM.BaseOffs = SC->getValue()->getSExtValue();
    AM.Scale = Scale;

    // If load[imm+r*scale] is illegal, bail out.
    if (!TLI->isLegalAddressingMode(AM, AccessTy))
      return false;
  }
  return true;
}

/// CheckForIVReuse - Returns the multiple if the stride is the multiple
/// of a previous stride and it is a legal value for the target addressing
/// mode scale component. This allows the users of this stride to be rewritten
/// as prev iv * factor. It returns 0 if no reuse is possible.
unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride, 
                                IVExpr &IV, const Type *Ty,
                                const std::vector<BasedUser>& UsersToProcess) {
  if (!TLI) return 0;

  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
    int64_t SInt = SC->getValue()->getSExtValue();
    if (SInt == 1) return 0;

    for (std::map<SCEVHandle, IVsOfOneStride>::iterator SI= IVsByStride.begin(),
           SE = IVsByStride.end(); SI != SE; ++SI) {
      int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
      if (SInt != -SSInt &&
          (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
        continue;
      int64_t Scale = SInt / SSInt;
      // Check that this stride is valid for all the types used for loads and
      // stores; if it can be used for some and not others, we might as well use
      // the original stride everywhere, since we have to create the IV for it
      // anyway.
      if (ValidStride(Scale, UsersToProcess))
        for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
               IE = SI->second.IVs.end(); II != IE; ++II)
          // FIXME: Only handle base == 0 for now.
          // Only reuse previous IV if it would not require a type conversion.
          if (isZero(II->Base) && II->Base->getType() == Ty) {
            IV = *II;
            return Scale;
          }
    }
  }
  return 0;
}

/// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
/// returns true if Val's isUseOfPostIncrementedValue is true.
static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
  return Val.isUseOfPostIncrementedValue;
}

/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
/// stride of IV.  All of the users may have different starting values, and this
/// may not be the only stride (we know it is if isOnlyStride is true).
void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
                                                      IVUsersOfOneStride &Uses,
                                                      Loop *L,
                                                      bool isOnlyStride) {
  // Transform our list of users and offsets to a bit more complex table.  In
  // this new vector, each 'BasedUser' contains 'Base' the base of the
  // strided accessas well as the old information from Uses.  We progressively
  // move information from the Base field to the Imm field, until we eventually
  // have the full access expression to rewrite the use.
  std::vector<BasedUser> UsersToProcess;
  UsersToProcess.reserve(Uses.Users.size());
  for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
    UsersToProcess.push_back(Uses.Users[i]);
    
    // Move any loop invariant operands from the offset field to the immediate
    // field of the use, so that we don't try to use something before it is
    // computed.
    MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
                                    UsersToProcess.back().Imm, L);
    assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
           "Base value is not loop invariant!");
  }

  // We now have a whole bunch of uses of like-strided induction variables, but
  // they might all have different bases.  We want to emit one PHI node for this
  // stride which we fold as many common expressions (between the IVs) into as
  // possible.  Start by identifying the common expressions in the base values 
  // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
  // "A+B"), emit it to the preheader, then remove the expression from the
  // UsersToProcess base values.
  SCEVHandle CommonExprs =
    RemoveCommonExpressionsFromUseBases(UsersToProcess);
  
  // Next, figure out what we can represent in the immediate fields of
  // instructions.  If we can represent anything there, move it to the imm
  // fields of the BasedUsers.  We do this so that it increases the commonality
  // of the remaining uses.
  for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
    // If the user is not in the current loop, this means it is using the exit
    // value of the IV.  Do not put anything in the base, make sure it's all in
    // the immediate field to allow as much factoring as possible.
    if (!L->contains(UsersToProcess[i].Inst->getParent())) {
      UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
                                               UsersToProcess[i].Base);
      UsersToProcess[i].Base = 
        SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
    } else {
      
      // Addressing modes can be folded into loads and stores.  Be careful that
      // the store is through the expression, not of the expression though.
      bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
      if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
        if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
          isAddress = true;
      
      MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
                          UsersToProcess[i].Imm, isAddress, L);
    }
  }

  // Check if it is possible to reuse a IV with stride that is factor of this
  // stride. And the multiple is a number that can be encoded in the scale
  // field of the target addressing mode.  And we will have a valid
  // instruction after this substition, including the immediate field, if any.
  PHINode *NewPHI = NULL;
  Value   *IncV   = NULL;
  IVExpr   ReuseIV;
  unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
                                           CommonExprs->getType(),
                                           UsersToProcess);
  if (RewriteFactor != 0) {
    DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
         << " and BASE " << *ReuseIV.Base << " :\n";
    NewPHI = ReuseIV.PHI;
    IncV   = ReuseIV.IncV;
  }

  const Type *ReplacedTy = CommonExprs->getType();
  
  // Now that we know what we need to do, insert the PHI node itself.
  //
  DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
       << *Stride << " and BASE " << *CommonExprs << " :\n";

  SCEVExpander Rewriter(*SE, *LI);
  SCEVExpander PreheaderRewriter(*SE, *LI);
  
  BasicBlock  *Preheader = L->getLoopPreheader();
  Instruction *PreInsertPt = Preheader->getTerminator();
  Instruction *PhiInsertBefore = L->getHeader()->begin();
  
  BasicBlock *LatchBlock = L->getLoopLatch();


  // Emit the initial base value into the loop preheader.
  Value *CommonBaseV
    = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
                                      ReplacedTy);

  if (RewriteFactor == 0) {
    // Create a new Phi for this base, and stick it in the loop header.
    NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
    ++NumInserted;
  
    // Add common base to the new Phi node.
    NewPHI->addIncoming(CommonBaseV, Preheader);

    // Insert the stride into the preheader.
    Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
                                                     ReplacedTy);
    if (!isa<ConstantInt>(StrideV)) ++NumVariable;

    // Emit the increment of the base value before the terminator of the loop
    // latch block, and add it to the Phi node.
    SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
                                         SCEVUnknown::get(StrideV));
  
    IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
                                  ReplacedTy);
    IncV->setName(NewPHI->getName()+".inc");
    NewPHI->addIncoming(IncV, LatchBlock);

    // Remember this in case a later stride is multiple of this.
    IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
  } else {
    Constant *C = dyn_cast<Constant>(CommonBaseV);
    if (!C ||
        (!C->isNullValue() &&
         !isTargetConstant(SCEVUnknown::get(CommonBaseV), ReplacedTy, TLI)))
      // We want the common base emitted into the preheader! This is just
      // using cast as a copy so BitCast (no-op cast) is appropriate
      CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(), 
                                    "commonbase", PreInsertPt);
  }

  // We want to emit code for users inside the loop first.  To do this, we
  // rearrange BasedUser so that the entries at the end have
  // isUseOfPostIncrementedValue = false, because we pop off the end of the
  // vector (so we handle them first).
  std::partition(UsersToProcess.begin(), UsersToProcess.end(),
                 PartitionByIsUseOfPostIncrementedValue);
  
  // Sort this by base, so that things with the same base are handled
  // together.  By partitioning first and stable-sorting later, we are
  // guaranteed that within each base we will pop off users from within the
  // loop before users outside of the loop with a particular base.
  //
  // We would like to use stable_sort here, but we can't.  The problem is that
  // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
  // we don't have anything to do a '<' comparison on.  Because we think the
  // number of uses is small, do a horrible bubble sort which just relies on
  // ==.
  for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
    // Get a base value.
    SCEVHandle Base = UsersToProcess[i].Base;
    
    // Compact everything with this base to be consequetive with this one.
    for (unsigned j = i+1; j != e; ++j) {
      if (UsersToProcess[j].Base == Base) {
        std::swap(UsersToProcess[i+1], UsersToProcess[j]);
        ++i;
      }
    }
  }

  // Process all the users now.  This outer loop handles all bases, the inner
  // loop handles all users of a particular base.
  while (!UsersToProcess.empty()) {
    SCEVHandle Base = UsersToProcess.back().Base;

    DOUT << "  INSERTING code for BASE = " << *Base << ":\n";
   
    // Emit the code for Base into the preheader.
    Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
                                                   ReplacedTy);
    
    // If BaseV is a constant other than 0, make sure that it gets inserted into
    // the preheader, instead of being forward substituted into the uses.  We do
    // this by forcing a BitCast (noop cast) to be inserted into the preheader 
    // in this case.
    if (Constant *C = dyn_cast<Constant>(BaseV)) {
      if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
        // We want this constant emitted into the preheader! This is just
        // using cast as a copy so BitCast (no-op cast) is appropriate
        BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
                             PreInsertPt);       
      }
    }

    // Emit the code to add the immediate offset to the Phi value, just before
    // the instructions that we identified as using this stride and base.
    do {
      // FIXME: Use emitted users to emit other users.
      BasedUser &User = UsersToProcess.back();

      // If this instruction wants to use the post-incremented value, move it
      // after the post-inc and use its value instead of the PHI.
      Value *RewriteOp = NewPHI;
      if (User.isUseOfPostIncrementedValue) {
        RewriteOp = IncV;

        // If this user is in the loop, make sure it is the last thing in the
        // loop to ensure it is dominated by the increment.
        if (L->contains(User.Inst->getParent()))
          User.Inst->moveBefore(LatchBlock->getTerminator());
      }
      if (RewriteOp->getType() != ReplacedTy) {
        Instruction::CastOps opcode = Instruction::Trunc;
        if (ReplacedTy->getPrimitiveSizeInBits() ==
            RewriteOp->getType()->getPrimitiveSizeInBits())
          opcode = Instruction::BitCast;
        RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
      }

      SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);

      // Clear the SCEVExpander's expression map so that we are guaranteed
      // to have the code emitted where we expect it.
      Rewriter.clear();

      // If we are reusing the iv, then it must be multiplied by a constant
      // factor take advantage of addressing mode scale component.
      if (RewriteFactor != 0) {
        RewriteExpr =
          SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
                                                       RewriteExpr->getType()),
                           RewriteExpr);

        // The common base is emitted in the loop preheader. But since we
        // are reusing an IV, it has not been used to initialize the PHI node.
        // Add it to the expression used to rewrite the uses.
        if (!isa<ConstantInt>(CommonBaseV) ||
            !cast<ConstantInt>(CommonBaseV)->isZero())
          RewriteExpr = SCEVAddExpr::get(RewriteExpr,
                                         SCEVUnknown::get(CommonBaseV));
      }

      // Now that we know what we need to do, insert code before User for the
      // immediate and any loop-variant expressions.
      if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
        // Add BaseV to the PHI value if needed.
        RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));

      User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);

      // Mark old value we replaced as possibly dead, so that it is elminated
      // if we just replaced the last use of that value.
      DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));

      UsersToProcess.pop_back();
      ++NumReduced;

      // If there are any more users to process with the same base, process them
      // now.  We sorted by base above, so we just have to check the last elt.
    } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
    // TODO: Next, find out which base index is the most common, pull it out.
  }

  // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
  // different starting values, into different PHIs.
}

/// FindIVForUser - If Cond has an operand that is an expression of an IV,
/// set the IV user and stride information and return true, otherwise return
/// false.
bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
                                       const SCEVHandle *&CondStride) {
  for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
       ++Stride) {
    std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI = 
    IVUsesByStride.find(StrideOrder[Stride]);
    assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
    
    for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
         E = SI->second.Users.end(); UI != E; ++UI)
      if (UI->User == Cond) {
        // NOTE: we could handle setcc instructions with multiple uses here, but
        // InstCombine does it as well for simple uses, it's not clear that it
        // occurs enough in real life to handle.
        CondUse = &*UI;
        CondStride = &SI->first;
        return true;
      }
  }
  return false;
}    

// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
// uses in the loop, look to see if we can eliminate some, in favor of using
// common indvars for the different uses.
void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
  // TODO: implement optzns here.

  // Finally, get the terminating condition for the loop if possible.  If we
  // can, we want to change it to use a post-incremented version of its
  // induction variable, to allow coalescing the live ranges for the IV into
  // one register value.
  PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
  BasicBlock  *Preheader = L->getLoopPreheader();
  BasicBlock *LatchBlock =
   SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
  BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
  if (!TermBr || TermBr->isUnconditional() || 
      !isa<ICmpInst>(TermBr->getCondition()))
    return;
  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());

  // Search IVUsesByStride to find Cond's IVUse if there is one.
  IVStrideUse *CondUse = 0;
  const SCEVHandle *CondStride = 0;

  if (!FindIVForUser(Cond, CondUse, CondStride))
    return; // setcc doesn't use the IV.
  

  // It's possible for the setcc instruction to be anywhere in the loop, and
  // possible for it to have multiple users.  If it is not immediately before
  // the latch block branch, move it.
  if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
    if (Cond->hasOneUse()) {   // Condition has a single use, just move it.
      Cond->moveBefore(TermBr);
    } else {
      // Otherwise, clone the terminating condition and insert into the loopend.
      Cond = cast<ICmpInst>(Cond->clone());
      Cond->setName(L->getHeader()->getName() + ".termcond");
      LatchBlock->getInstList().insert(TermBr, Cond);
      
      // Clone the IVUse, as the old use still exists!
      IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
                                         CondUse->OperandValToReplace);
      CondUse = &IVUsesByStride[*CondStride].Users.back();
    }
  }

  // If we get to here, we know that we can transform the setcc instruction to
  // use the post-incremented version of the IV, allowing us to coalesce the
  // live ranges for the IV correctly.
  CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
  CondUse->isUseOfPostIncrementedValue = true;
}

namespace {
  // Constant strides come first which in turns are sorted by their absolute
  // values. If absolute values are the same, then positive strides comes first.
  // e.g.
  // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
  struct StrideCompare {
    bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
      SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
      SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
      if (LHSC && RHSC) {
        int64_t  LV = LHSC->getValue()->getSExtValue();
        int64_t  RV = RHSC->getValue()->getSExtValue();
        uint64_t ALV = (LV < 0) ? -LV : LV;
        uint64_t ARV = (RV < 0) ? -RV : RV;
        if (ALV == ARV)
          return LV > RV;
        else
          return ALV < ARV;
      }
      return (LHSC && !RHSC);
    }
  };
}

bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {

  LI = &getAnalysis<LoopInfo>();
  EF = &getAnalysis<ETForest>();
  SE = &getAnalysis<ScalarEvolution>();
  TD = &getAnalysis<TargetData>();
  UIntPtrTy = TD->getIntPtrType();

  // Find all uses of induction variables in this loop, and catagorize
  // them by stride.  Start by finding all of the PHI nodes in the header for
  // this loop.  If they are induction variables, inspect their uses.
  std::set<Instruction*> Processed;   // Don't reprocess instructions.
  for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
    AddUsersIfInteresting(I, L, Processed);

  // If we have nothing to do, return.
  if (IVUsesByStride.empty()) return false;

  // Optimize induction variables.  Some indvar uses can be transformed to use
  // strides that will be needed for other purposes.  A common example of this
  // is the exit test for the loop, which can often be rewritten to use the
  // computation of some other indvar to decide when to terminate the loop.
  OptimizeIndvars(L);


  // FIXME: We can widen subreg IV's here for RISC targets.  e.g. instead of
  // doing computation in byte values, promote to 32-bit values if safe.

  // FIXME: Attempt to reuse values across multiple IV's.  In particular, we
  // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
  // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.  Need
  // to be careful that IV's are all the same type.  Only works for intptr_t
  // indvars.

  // If we only have one stride, we can more aggressively eliminate some things.
  bool HasOneStride = IVUsesByStride.size() == 1;

#ifndef NDEBUG
  DOUT << "\nLSR on ";
  DEBUG(L->dump());
#endif

  // IVsByStride keeps IVs for one particular loop.
  IVsByStride.clear();

  // Sort the StrideOrder so we process larger strides first.
  std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());

  // Note: this processes each stride/type pair individually.  All users passed
  // into StrengthReduceStridedIVUsers have the same type AND stride.  Also,
  // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
  // This extra layer of indirection makes the ordering of strides deterministic
  // - not dependent on map order.
  for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
    std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI = 
      IVUsesByStride.find(StrideOrder[Stride]);
    assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
    StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
  }

  // Clean up after ourselves
  if (!DeadInsts.empty()) {
    DeleteTriviallyDeadInstructions(DeadInsts);

    BasicBlock::iterator I = L->getHeader()->begin();
    PHINode *PN;
    while ((PN = dyn_cast<PHINode>(I))) {
      ++I;  // Preincrement iterator to avoid invalidating it when deleting PN.
      
      // At this point, we know that we have killed one or more GEP
      // instructions.  It is worth checking to see if the cann indvar is also
      // dead, so that we can remove it as well.  The requirements for the cann
      // indvar to be considered dead are:
      // 1. the cann indvar has one use
      // 2. the use is an add instruction
      // 3. the add has one use
      // 4. the add is used by the cann indvar
      // If all four cases above are true, then we can remove both the add and
      // the cann indvar.
      // FIXME: this needs to eliminate an induction variable even if it's being
      // compared against some value to decide loop termination.
      if (PN->hasOneUse()) {
        Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
        if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
          if (BO->hasOneUse() && PN == *(BO->use_begin())) {
            DeadInsts.insert(BO);
            // Break the cycle, then delete the PHI.
            PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
            SE->deleteInstructionFromRecords(PN);
            PN->eraseFromParent();
          }
        }
      }
    }
    DeleteTriviallyDeadInstructions(DeadInsts);
  }

  CastedPointers.clear();
  IVUsesByStride.clear();
  StrideOrder.clear();
  return false;
}