summaryrefslogtreecommitdiff
path: root/lib/CodeGen/RegAllocGreedy.cpp
blob: cbecc4b2d387a69fb68c92a4352c5c713c555cb0 (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
//===-- RegAllocGreedy.cpp - greedy register allocator --------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the RAGreedy function pass for register allocation in
// optimized builds.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "regalloc"
#include "AllocationOrder.h"
#include "LiveIntervalUnion.h"
#include "LiveRangeEdit.h"
#include "RegAllocBase.h"
#include "Spiller.h"
#include "SpillPlacement.h"
#include "SplitKit.h"
#include "VirtRegMap.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Function.h"
#include "llvm/PassAnalysisSupport.h"
#include "llvm/CodeGen/CalcSpillWeights.h"
#include "llvm/CodeGen/EdgeBundles.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/LiveStackAnalysis.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineLoopRanges.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/RegAllocRegistry.h"
#include "llvm/CodeGen/RegisterCoalescer.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Timer.h"

#include <queue>

using namespace llvm;

STATISTIC(NumGlobalSplits, "Number of split global live ranges");
STATISTIC(NumLocalSplits,  "Number of split local live ranges");
STATISTIC(NumReassigned,   "Number of interferences reassigned");
STATISTIC(NumEvicted,      "Number of interferences evicted");

static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
                                       createGreedyRegisterAllocator);

namespace {
class RAGreedy : public MachineFunctionPass, public RegAllocBase {
  // context
  MachineFunction *MF;
  BitVector ReservedRegs;

  // analyses
  SlotIndexes *Indexes;
  LiveStacks *LS;
  MachineDominatorTree *DomTree;
  MachineLoopInfo *Loops;
  MachineLoopRanges *LoopRanges;
  EdgeBundles *Bundles;
  SpillPlacement *SpillPlacer;

  // state
  std::auto_ptr<Spiller> SpillerInstance;
  std::priority_queue<std::pair<unsigned, unsigned> > Queue;

  // Live ranges pass through a number of stages as we try to allocate them.
  // Some of the stages may also create new live ranges:
  //
  // - Region splitting.
  // - Per-block splitting.
  // - Local splitting.
  // - Spilling.
  //
  // Ranges produced by one of the stages skip the previous stages when they are
  // dequeued. This improves performance because we can skip interference checks
  // that are unlikely to give any results. It also guarantees that the live
  // range splitting algorithm terminates, something that is otherwise hard to
  // ensure.
  enum LiveRangeStage {
    RS_Original, ///< Never seen before, never split.
    RS_Second,   ///< Second time in the queue.
    RS_Region,   ///< Produced by region splitting.
    RS_Block,    ///< Produced by per-block splitting.
    RS_Local,    ///< Produced by local splitting.
    RS_Spill     ///< Produced by spilling.
  };

  IndexedMap<unsigned char, VirtReg2IndexFunctor> LRStage;

  LiveRangeStage getStage(const LiveInterval &VirtReg) const {
    return LiveRangeStage(LRStage[VirtReg.reg]);
  }

  template<typename Iterator>
  void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
    LRStage.resize(MRI->getNumVirtRegs());
    for (;Begin != End; ++Begin)
      LRStage[(*Begin)->reg] = NewStage;
  }

  // splitting state.
  std::auto_ptr<SplitAnalysis> SA;
  std::auto_ptr<SplitEditor> SE;

  /// All basic blocks where the current register is live.
  SmallVector<SpillPlacement::BlockConstraint, 8> SpillConstraints;

  /// For every instruction in SA->UseSlots, store the previous non-copy
  /// instruction.
  SmallVector<SlotIndex, 8> PrevSlot;

public:
  RAGreedy();

  /// Return the pass name.
  virtual const char* getPassName() const {
    return "Greedy Register Allocator";
  }

  /// RAGreedy analysis usage.
  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
  virtual void releaseMemory();
  virtual Spiller &spiller() { return *SpillerInstance; }
  virtual void enqueue(LiveInterval *LI);
  virtual LiveInterval *dequeue();
  virtual unsigned selectOrSplit(LiveInterval&,
                                 SmallVectorImpl<LiveInterval*>&);

  /// Perform register allocation.
  virtual bool runOnMachineFunction(MachineFunction &mf);

  static char ID;

private:
  bool checkUncachedInterference(LiveInterval&, unsigned);
  LiveInterval *getSingleInterference(LiveInterval&, unsigned);
  bool reassignVReg(LiveInterval &InterferingVReg, unsigned OldPhysReg);
  float calcInterferenceInfo(LiveInterval&, unsigned);
  float calcGlobalSplitCost(const BitVector&);
  void splitAroundRegion(LiveInterval&, unsigned, const BitVector&,
                         SmallVectorImpl<LiveInterval*>&);
  void calcGapWeights(unsigned, SmallVectorImpl<float>&);
  SlotIndex getPrevMappedIndex(const MachineInstr*);
  void calcPrevSlots();
  unsigned nextSplitPoint(unsigned);
  bool canEvictInterference(LiveInterval&, unsigned, unsigned, float&);

  unsigned tryReassign(LiveInterval&, AllocationOrder&,
                              SmallVectorImpl<LiveInterval*>&);
  unsigned tryEvict(LiveInterval&, AllocationOrder&,
                    SmallVectorImpl<LiveInterval*>&);
  unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
                          SmallVectorImpl<LiveInterval*>&);
  unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
    SmallVectorImpl<LiveInterval*>&);
  unsigned trySplit(LiveInterval&, AllocationOrder&,
                    SmallVectorImpl<LiveInterval*>&);
};
} // end anonymous namespace

char RAGreedy::ID = 0;

FunctionPass* llvm::createGreedyRegisterAllocator() {
  return new RAGreedy();
}

RAGreedy::RAGreedy(): MachineFunctionPass(ID), LRStage(RS_Original) {
  initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
  initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
  initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
  initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
  initializeRegisterCoalescerAnalysisGroup(*PassRegistry::getPassRegistry());
  initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
  initializeLiveStacksPass(*PassRegistry::getPassRegistry());
  initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
  initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
  initializeMachineLoopRangesPass(*PassRegistry::getPassRegistry());
  initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
  initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
  initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
}

void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
  AU.setPreservesCFG();
  AU.addRequired<AliasAnalysis>();
  AU.addPreserved<AliasAnalysis>();
  AU.addRequired<LiveIntervals>();
  AU.addRequired<SlotIndexes>();
  AU.addPreserved<SlotIndexes>();
  if (StrongPHIElim)
    AU.addRequiredID(StrongPHIEliminationID);
  AU.addRequiredTransitive<RegisterCoalescer>();
  AU.addRequired<CalculateSpillWeights>();
  AU.addRequired<LiveStacks>();
  AU.addPreserved<LiveStacks>();
  AU.addRequired<MachineDominatorTree>();
  AU.addPreserved<MachineDominatorTree>();
  AU.addRequired<MachineLoopInfo>();
  AU.addPreserved<MachineLoopInfo>();
  AU.addRequired<MachineLoopRanges>();
  AU.addPreserved<MachineLoopRanges>();
  AU.addRequired<VirtRegMap>();
  AU.addPreserved<VirtRegMap>();
  AU.addRequired<EdgeBundles>();
  AU.addRequired<SpillPlacement>();
  MachineFunctionPass::getAnalysisUsage(AU);
}

void RAGreedy::releaseMemory() {
  SpillerInstance.reset(0);
  LRStage.clear();
  RegAllocBase::releaseMemory();
}

void RAGreedy::enqueue(LiveInterval *LI) {
  // Prioritize live ranges by size, assigning larger ranges first.
  // The queue holds (size, reg) pairs.
  const unsigned Size = LI->getSize();
  const unsigned Reg = LI->reg;
  assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
         "Can only enqueue virtual registers");
  unsigned Prio;

  LRStage.grow(Reg);
  if (LRStage[Reg] == RS_Original)
    // 1st generation ranges are handled first, long -> short.
    Prio = (1u << 31) + Size;
  else
    // Repeat offenders are handled second, short -> long
    Prio = (1u << 30) - Size;

  // Boost ranges that have a physical register hint.
  const unsigned Hint = VRM->getRegAllocPref(Reg);
  if (TargetRegisterInfo::isPhysicalRegister(Hint))
    Prio |= (1u << 30);

  Queue.push(std::make_pair(Prio, Reg));
}

LiveInterval *RAGreedy::dequeue() {
  if (Queue.empty())
    return 0;
  LiveInterval *LI = &LIS->getInterval(Queue.top().second);
  Queue.pop();
  return LI;
}

//===----------------------------------------------------------------------===//
//                         Register Reassignment
//===----------------------------------------------------------------------===//

// Check interference without using the cache.
bool RAGreedy::checkUncachedInterference(LiveInterval &VirtReg,
                                         unsigned PhysReg) {
  for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
    LiveIntervalUnion::Query subQ(&VirtReg, &PhysReg2LiveUnion[*AliasI]);
    if (subQ.checkInterference())
      return true;
  }
  return false;
}

/// getSingleInterference - Return the single interfering virtual register
/// assigned to PhysReg. Return 0 if more than one virtual register is
/// interfering.
LiveInterval *RAGreedy::getSingleInterference(LiveInterval &VirtReg,
                                              unsigned PhysReg) {
  // Check physreg and aliases.
  LiveInterval *Interference = 0;
  for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
    LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
    if (Q.checkInterference()) {
      if (Interference)
        return 0;
      if (Q.collectInterferingVRegs(2) > 1)
        return 0;
      Interference = Q.interferingVRegs().front();
    }
  }
  return Interference;
}

// Attempt to reassign this virtual register to a different physical register.
//
// FIXME: we are not yet caching these "second-level" interferences discovered
// in the sub-queries. These interferences can change with each call to
// selectOrSplit. However, we could implement a "may-interfere" cache that
// could be conservatively dirtied when we reassign or split.
//
// FIXME: This may result in a lot of alias queries. We could summarize alias
// live intervals in their parent register's live union, but it's messy.
bool RAGreedy::reassignVReg(LiveInterval &InterferingVReg,
                            unsigned WantedPhysReg) {
  assert(TargetRegisterInfo::isVirtualRegister(InterferingVReg.reg) &&
         "Can only reassign virtual registers");
  assert(TRI->regsOverlap(WantedPhysReg, VRM->getPhys(InterferingVReg.reg)) &&
         "inconsistent phys reg assigment");

  AllocationOrder Order(InterferingVReg.reg, *VRM, ReservedRegs);
  while (unsigned PhysReg = Order.next()) {
    // Don't reassign to a WantedPhysReg alias.
    if (TRI->regsOverlap(PhysReg, WantedPhysReg))
      continue;

    if (checkUncachedInterference(InterferingVReg, PhysReg))
      continue;

    // Reassign the interfering virtual reg to this physical reg.
    unsigned OldAssign = VRM->getPhys(InterferingVReg.reg);
    DEBUG(dbgs() << "reassigning: " << InterferingVReg << " from " <<
          TRI->getName(OldAssign) << " to " << TRI->getName(PhysReg) << '\n');
    unassign(InterferingVReg, OldAssign);
    assign(InterferingVReg, PhysReg);
    ++NumReassigned;
    return true;
  }
  return false;
}

/// tryReassign - Try to reassign a single interference to a different physreg.
/// @param  VirtReg Currently unassigned virtual register.
/// @param  Order   Physregs to try.
/// @return         Physreg to assign VirtReg, or 0.
unsigned RAGreedy::tryReassign(LiveInterval &VirtReg, AllocationOrder &Order,
                               SmallVectorImpl<LiveInterval*> &NewVRegs){
  NamedRegionTimer T("Reassign", TimerGroupName, TimePassesIsEnabled);

  Order.rewind();
  while (unsigned PhysReg = Order.next()) {
    LiveInterval *InterferingVReg = getSingleInterference(VirtReg, PhysReg);
    if (!InterferingVReg)
      continue;
    if (TargetRegisterInfo::isPhysicalRegister(InterferingVReg->reg))
      continue;
    if (reassignVReg(*InterferingVReg, PhysReg))
      return PhysReg;
  }
  return 0;
}


//===----------------------------------------------------------------------===//
//                         Interference eviction
//===----------------------------------------------------------------------===//

/// canEvict - Return true if all interferences between VirtReg and PhysReg can
/// be evicted. Set maxWeight to the maximal spill weight of an interference.
bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
                                    unsigned Size, float &MaxWeight) {
  float Weight = 0;
  for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
    LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
    // If there is 10 or more interferences, chances are one is smaller.
    if (Q.collectInterferingVRegs(10) >= 10)
      return false;

    // CHeck if any interfering live range is shorter than VirtReg.
    for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
      LiveInterval *Intf = Q.interferingVRegs()[i];
      if (TargetRegisterInfo::isPhysicalRegister(Intf->reg))
        return false;
      if (Intf->getSize() <= Size)
        return false;
      Weight = std::max(Weight, Intf->weight);
    }
  }
  MaxWeight = Weight;
  return true;
}

/// tryEvict - Try to evict all interferences for a physreg.
/// @param  VirtReg Currently unassigned virtual register.
/// @param  Order   Physregs to try.
/// @return         Physreg to assign VirtReg, or 0.
unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
                            AllocationOrder &Order,
                            SmallVectorImpl<LiveInterval*> &NewVRegs){
  NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);

  // We can only evict interference if all interfering registers are virtual and
  // longer than VirtReg.
  const unsigned Size = VirtReg.getSize();

  // Keep track of the lightest single interference seen so far.
  float BestWeight = 0;
  unsigned BestPhys = 0;

  Order.rewind();
  while (unsigned PhysReg = Order.next()) {
    float Weight = 0;
    if (!canEvictInterference(VirtReg, PhysReg, Size, Weight))
      continue;

    // This is an eviction candidate.
    DEBUG(dbgs() << "max " << PrintReg(PhysReg, TRI) << " interference = "
                 << Weight << '\n');
    if (BestPhys && Weight >= BestWeight)
      continue;

    // Best so far.
    BestPhys = PhysReg;
    BestWeight = Weight;
    // Stop if the hint can be used.
    if (Order.isHint(PhysReg))
      break;
  }

  if (!BestPhys)
    return 0;

  DEBUG(dbgs() << "evicting " << PrintReg(BestPhys, TRI) << " interference\n");
  for (const unsigned *AliasI = TRI->getOverlaps(BestPhys); *AliasI; ++AliasI) {
    LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
    assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
    for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
      LiveInterval *Intf = Q.interferingVRegs()[i];
      unassign(*Intf, VRM->getPhys(Intf->reg));
      ++NumEvicted;
      NewVRegs.push_back(Intf);
    }
  }
  return BestPhys;
}


//===----------------------------------------------------------------------===//
//                              Region Splitting
//===----------------------------------------------------------------------===//

/// calcInterferenceInfo - Compute per-block outgoing and ingoing constraints
/// when considering interference from PhysReg. Also compute an optimistic local
/// cost of this interference pattern.
///
/// The final cost of a split is the local cost + global cost of preferences
/// broken by SpillPlacement.
///
float RAGreedy::calcInterferenceInfo(LiveInterval &VirtReg, unsigned PhysReg) {
  // Reset interference dependent info.
  SpillConstraints.resize(SA->LiveBlocks.size());
  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
    SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
    BC.Number = BI.MBB->getNumber();
    BC.Entry = (BI.Uses && BI.LiveIn) ?
      SpillPlacement::PrefReg : SpillPlacement::DontCare;
    BC.Exit = (BI.Uses && BI.LiveOut) ?
      SpillPlacement::PrefReg : SpillPlacement::DontCare;
    BI.OverlapEntry = BI.OverlapExit = false;
  }

  // Add interference info from each PhysReg alias.
  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
    if (!query(VirtReg, *AI).checkInterference())
      continue;
    LiveIntervalUnion::SegmentIter IntI =
      PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
    if (!IntI.valid())
      continue;

    // Determine which blocks have interference live in or after the last split
    // point.
    for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
      SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
      SpillPlacement::BlockConstraint &BC = SpillConstraints[i];

      // Skip interference-free blocks.
      if (IntI.start() >= BI.Stop)
        continue;

      // Is the interference live-in?
      if (BI.LiveIn) {
        IntI.advanceTo(BI.Start);
        if (!IntI.valid())
          break;
        if (IntI.start() <= BI.Start)
          BC.Entry = SpillPlacement::MustSpill;
      }

      // Is the interference overlapping the last split point?
      if (BI.LiveOut) {
        if (IntI.stop() < BI.LastSplitPoint)
          IntI.advanceTo(BI.LastSplitPoint.getPrevSlot());
        if (!IntI.valid())
          break;
        if (IntI.start() < BI.Stop)
          BC.Exit = SpillPlacement::MustSpill;
      }
    }

    // Rewind iterator and check other interferences.
    IntI.find(VirtReg.beginIndex());
    for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
      SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
      SpillPlacement::BlockConstraint &BC = SpillConstraints[i];

      // Skip interference-free blocks.
      if (IntI.start() >= BI.Stop)
        continue;

      // Handle transparent blocks with interference separately.
      // Transparent blocks never incur any fixed cost.
      if (BI.LiveThrough && !BI.Uses) {
        IntI.advanceTo(BI.Start);
        if (!IntI.valid())
          break;
        if (IntI.start() >= BI.Stop)
          continue;

        if (BC.Entry != SpillPlacement::MustSpill)
          BC.Entry = SpillPlacement::PrefSpill;
        if (BC.Exit != SpillPlacement::MustSpill)
          BC.Exit = SpillPlacement::PrefSpill;
        continue;
      }

      // Now we only have blocks with uses left.
      // Check if the interference overlaps the uses.
      assert(BI.Uses && "Non-transparent block without any uses");

      // Check interference on entry.
      if (BI.LiveIn && BC.Entry != SpillPlacement::MustSpill) {
        IntI.advanceTo(BI.Start);
        if (!IntI.valid())
          break;
        // Not live in, but before the first use.
        if (IntI.start() < BI.FirstUse) {
          BC.Entry = SpillPlacement::PrefSpill;
          // If the block contains a kill from an earlier split, never split
          // again in the same block.
          if (!BI.LiveThrough && !SA->isOriginalEndpoint(BI.Kill))
            BC.Entry = SpillPlacement::MustSpill;
        }
      }

      // Does interference overlap the uses in the entry segment
      // [FirstUse;Kill)?
      if (BI.LiveIn && !BI.OverlapEntry) {
        IntI.advanceTo(BI.FirstUse);
        if (!IntI.valid())
          break;
        // A live-through interval has no kill.
        // Check [FirstUse;LastUse) instead.
        if (IntI.start() < (BI.LiveThrough ? BI.LastUse : BI.Kill))
          BI.OverlapEntry = true;
      }

      // Does interference overlap the uses in the exit segment [Def;LastUse)?
      if (BI.LiveOut && !BI.LiveThrough && !BI.OverlapExit) {
        IntI.advanceTo(BI.Def);
        if (!IntI.valid())
          break;
        if (IntI.start() < BI.LastUse)
          BI.OverlapExit = true;
      }

      // Check interference on exit.
      if (BI.LiveOut && BC.Exit != SpillPlacement::MustSpill) {
        // Check interference between LastUse and Stop.
        if (BC.Exit != SpillPlacement::PrefSpill) {
          IntI.advanceTo(BI.LastUse);
          if (!IntI.valid())
            break;
          if (IntI.start() < BI.Stop) {
            BC.Exit = SpillPlacement::PrefSpill;
            // Avoid splitting twice in the same block.
            if (!BI.LiveThrough && !SA->isOriginalEndpoint(BI.Def))
              BC.Exit = SpillPlacement::MustSpill;
          }
        }
      }
    }
  }

  // Accumulate a local cost of this interference pattern.
  float LocalCost = 0;
  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
    if (!BI.Uses)
      continue;
    SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
    unsigned Inserts = 0;

    // Do we need spill code for the entry segment?
    if (BI.LiveIn)
      Inserts += BI.OverlapEntry || BC.Entry != SpillPlacement::PrefReg;

    // For the exit segment?
    if (BI.LiveOut)
      Inserts += BI.OverlapExit || BC.Exit != SpillPlacement::PrefReg;

    // The local cost of spill code in this block is the block frequency times
    // the number of spill instructions inserted.
    if (Inserts)
      LocalCost += Inserts * SpillPlacer->getBlockFrequency(BI.MBB);
  }
  DEBUG(dbgs() << "Local cost of " << PrintReg(PhysReg, TRI) << " = "
               << LocalCost << '\n');
  return LocalCost;
}

/// calcGlobalSplitCost - Return the global split cost of following the split
/// pattern in LiveBundles. This cost should be added to the local cost of the
/// interference pattern in SpillConstraints.
///
float RAGreedy::calcGlobalSplitCost(const BitVector &LiveBundles) {
  float GlobalCost = 0;
  for (unsigned i = 0, e = SpillConstraints.size(); i != e; ++i) {
    SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
    unsigned Inserts = 0;
    // Broken entry preference?
    Inserts += LiveBundles[Bundles->getBundle(BC.Number, 0)] !=
                 (BC.Entry == SpillPlacement::PrefReg);
    // Broken exit preference?
    Inserts += LiveBundles[Bundles->getBundle(BC.Number, 1)] !=
                 (BC.Exit == SpillPlacement::PrefReg);
    if (Inserts)
      GlobalCost +=
        Inserts * SpillPlacer->getBlockFrequency(SA->LiveBlocks[i].MBB);
  }
  DEBUG(dbgs() << "Global cost = " << GlobalCost << '\n');
  return GlobalCost;
}

/// splitAroundRegion - Split VirtReg around the region determined by
/// LiveBundles. Make an effort to avoid interference from PhysReg.
///
/// The 'register' interval is going to contain as many uses as possible while
/// avoiding interference. The 'stack' interval is the complement constructed by
/// SplitEditor. It will contain the rest.
///
void RAGreedy::splitAroundRegion(LiveInterval &VirtReg, unsigned PhysReg,
                                 const BitVector &LiveBundles,
                                 SmallVectorImpl<LiveInterval*> &NewVRegs) {
  DEBUG({
    dbgs() << "Splitting around region for " << PrintReg(PhysReg, TRI)
           << " with bundles";
    for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
      dbgs() << " EB#" << i;
    dbgs() << ".\n";
  });

  // First compute interference ranges in the live blocks.
  typedef std::pair<SlotIndex, SlotIndex> IndexPair;
  SmallVector<IndexPair, 8> InterferenceRanges;
  InterferenceRanges.resize(SA->LiveBlocks.size());
  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
    if (!query(VirtReg, *AI).checkInterference())
      continue;
    LiveIntervalUnion::SegmentIter IntI =
      PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
    if (!IntI.valid())
      continue;
    for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
      const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
      IndexPair &IP = InterferenceRanges[i];

      // Skip interference-free blocks.
      if (IntI.start() >= BI.Stop)
        continue;

      // First interference in block.
      if (BI.LiveIn) {
        IntI.advanceTo(BI.Start);
        if (!IntI.valid())
          break;
        if (IntI.start() >= BI.Stop)
          continue;
        if (!IP.first.isValid() || IntI.start() < IP.first)
          IP.first = IntI.start();
      }

      // Last interference in block.
      if (BI.LiveOut) {
        IntI.advanceTo(BI.Stop);
        if (!IntI.valid() || IntI.start() >= BI.Stop)
          --IntI;
        if (IntI.stop() <= BI.Start)
          continue;
        if (!IP.second.isValid() || IntI.stop() > IP.second)
          IP.second = IntI.stop();
      }
    }
  }

  SmallVector<LiveInterval*, 4> SpillRegs;
  LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
  SE->reset(LREdit);

  // Create the main cross-block interval.
  SE->openIntv();

  // First add all defs that are live out of a block.
  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
    bool RegIn  = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
    bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];

    // Should the register be live out?
    if (!BI.LiveOut || !RegOut)
      continue;

    IndexPair &IP = InterferenceRanges[i];
    DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " -> EB#"
                 << Bundles->getBundle(BI.MBB->getNumber(), 1)
                 << " intf [" << IP.first << ';' << IP.second << ')');

    // The interference interval should either be invalid or overlap MBB.
    assert((!IP.first.isValid() || IP.first < BI.Stop) && "Bad interference");
    assert((!IP.second.isValid() || IP.second > BI.Start)
           && "Bad interference");

    // Check interference leaving the block.
    if (!IP.second.isValid()) {
      // Block is interference-free.
      DEBUG(dbgs() << ", no interference");
      if (!BI.Uses) {
        assert(BI.LiveThrough && "No uses, but not live through block?");
        // Block is live-through without interference.
        DEBUG(dbgs() << ", no uses"
                     << (RegIn ? ", live-through.\n" : ", stack in.\n"));
        if (!RegIn)
          SE->enterIntvAtEnd(*BI.MBB);
        continue;
      }
      if (!BI.LiveThrough) {
        DEBUG(dbgs() << ", not live-through.\n");
        SE->useIntv(SE->enterIntvBefore(BI.Def), BI.Stop);
        continue;
      }
      if (!RegIn) {
        // Block is live-through, but entry bundle is on the stack.
        // Reload just before the first use.
        DEBUG(dbgs() << ", not live-in, enter before first use.\n");
        SE->useIntv(SE->enterIntvBefore(BI.FirstUse), BI.Stop);
        continue;
      }
      DEBUG(dbgs() << ", live-through.\n");
      continue;
    }

    // Block has interference.
    DEBUG(dbgs() << ", interference to " << IP.second);

    if (!BI.LiveThrough && IP.second <= BI.Def) {
      // The interference doesn't reach the outgoing segment.
      DEBUG(dbgs() << " doesn't affect def from " << BI.Def << '\n');
      SE->useIntv(BI.Def, BI.Stop);
      continue;
    }


    if (!BI.Uses) {
      // No uses in block, avoid interference by reloading as late as possible.
      DEBUG(dbgs() << ", no uses.\n");
      SlotIndex SegStart = SE->enterIntvAtEnd(*BI.MBB);
      assert(SegStart >= IP.second && "Couldn't avoid interference");
      continue;
    }

    if (IP.second.getBoundaryIndex() < BI.LastUse) {
      // There are interference-free uses at the end of the block.
      // Find the first use that can get the live-out register.
      SmallVectorImpl<SlotIndex>::const_iterator UI =
        std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
                         IP.second.getBoundaryIndex());
      assert(UI != SA->UseSlots.end() && "Couldn't find last use");
      SlotIndex Use = *UI;
      assert(Use <= BI.LastUse && "Couldn't find last use");
      // Only attempt a split befroe the last split point.
      if (Use.getBaseIndex() <= BI.LastSplitPoint) {
        DEBUG(dbgs() << ", free use at " << Use << ".\n");
        SlotIndex SegStart = SE->enterIntvBefore(Use);
        assert(SegStart >= IP.second && "Couldn't avoid interference");
        assert(SegStart < BI.LastSplitPoint && "Impossible split point");
        SE->useIntv(SegStart, BI.Stop);
        continue;
      }
    }

    // Interference is after the last use.
    DEBUG(dbgs() << " after last use.\n");
    SlotIndex SegStart = SE->enterIntvAtEnd(*BI.MBB);
    assert(SegStart >= IP.second && "Couldn't avoid interference");
  }

  // Now all defs leading to live bundles are handled, do everything else.
  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
    bool RegIn  = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
    bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];

    // Is the register live-in?
    if (!BI.LiveIn || !RegIn)
      continue;

    // We have an incoming register. Check for interference.
    IndexPair &IP = InterferenceRanges[i];

    DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
                 << " -> BB#" << BI.MBB->getNumber());

    // Check interference entering the block.
    if (!IP.first.isValid()) {
      // Block is interference-free.
      DEBUG(dbgs() << ", no interference");
      if (!BI.Uses) {
        assert(BI.LiveThrough && "No uses, but not live through block?");
        // Block is live-through without interference.
        if (RegOut) {
          DEBUG(dbgs() << ", no uses, live-through.\n");
          SE->useIntv(BI.Start, BI.Stop);
        } else {
          DEBUG(dbgs() << ", no uses, stack-out.\n");
          SE->leaveIntvAtTop(*BI.MBB);
        }
        continue;
      }
      if (!BI.LiveThrough) {
        DEBUG(dbgs() << ", killed in block.\n");
        SE->useIntv(BI.Start, SE->leaveIntvAfter(BI.Kill));
        continue;
      }
      if (!RegOut) {
        // Block is live-through, but exit bundle is on the stack.
        // Spill immediately after the last use.
        if (BI.LastUse < BI.LastSplitPoint) {
          DEBUG(dbgs() << ", uses, stack-out.\n");
          SE->useIntv(BI.Start, SE->leaveIntvAfter(BI.LastUse));
          continue;
        }
        // The last use is after the last split point, it is probably an
        // indirect jump.
        DEBUG(dbgs() << ", uses at " << BI.LastUse << " after split point "
                     << BI.LastSplitPoint << ", stack-out.\n");
        SlotIndex SegEnd = SE->leaveIntvBefore(BI.LastSplitPoint);
        SE->useIntv(BI.Start, SegEnd);
        // Run a double interval from the split to the last use.
        // This makes it possible to spill the complement without affecting the
        // indirect branch.
        SE->overlapIntv(SegEnd, BI.LastUse);
        continue;
      }
      // Register is live-through.
      DEBUG(dbgs() << ", uses, live-through.\n");
      SE->useIntv(BI.Start, BI.Stop);
      continue;
    }

    // Block has interference.
    DEBUG(dbgs() << ", interference from " << IP.first);

    if (!BI.LiveThrough && IP.first >= BI.Kill) {
      // The interference doesn't reach the outgoing segment.
      DEBUG(dbgs() << " doesn't affect kill at " << BI.Kill << '\n');
      SE->useIntv(BI.Start, BI.Kill);
      continue;
    }

    if (!BI.Uses) {
      // No uses in block, avoid interference by spilling as soon as possible.
      DEBUG(dbgs() << ", no uses.\n");
      SlotIndex SegEnd = SE->leaveIntvAtTop(*BI.MBB);
      assert(SegEnd <= IP.first && "Couldn't avoid interference");
      continue;
    }
    if (IP.first.getBaseIndex() > BI.FirstUse) {
      // There are interference-free uses at the beginning of the block.
      // Find the last use that can get the register.
      SmallVectorImpl<SlotIndex>::const_iterator UI =
        std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
                         IP.first.getBaseIndex());
      assert(UI != SA->UseSlots.begin() && "Couldn't find first use");
      SlotIndex Use = (--UI)->getBoundaryIndex();
      DEBUG(dbgs() << ", free use at " << *UI << ".\n");
      SlotIndex SegEnd = SE->leaveIntvAfter(Use);
      assert(SegEnd <= IP.first && "Couldn't avoid interference");
      SE->useIntv(BI.Start, SegEnd);
      continue;
    }

    // Interference is before the first use.
    DEBUG(dbgs() << " before first use.\n");
    SlotIndex SegEnd = SE->leaveIntvAtTop(*BI.MBB);
    assert(SegEnd <= IP.first && "Couldn't avoid interference");
  }

  SE->closeIntv();

  // FIXME: Should we be more aggressive about splitting the stack region into
  // per-block segments? The current approach allows the stack region to
  // separate into connected components. Some components may be allocatable.
  SE->finish();
  ++NumGlobalSplits;

  if (VerifyEnabled) {
    MF->verify(this, "After splitting live range around region");

#ifndef NDEBUG
    // Make sure that at least one of the new intervals can allocate to PhysReg.
    // That was the whole point of splitting the live range.
    bool found = false;
    for (LiveRangeEdit::iterator I = LREdit.begin(), E = LREdit.end(); I != E;
         ++I)
      if (!checkUncachedInterference(**I, PhysReg)) {
        found = true;
        break;
      }
    assert(found && "No allocatable intervals after pointless splitting");
#endif
  }
}

unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
                                  SmallVectorImpl<LiveInterval*> &NewVRegs) {
  BitVector LiveBundles, BestBundles;
  float BestCost = 0;
  unsigned BestReg = 0;
  Order.rewind();
  while (unsigned PhysReg = Order.next()) {
    float Cost = calcInterferenceInfo(VirtReg, PhysReg);
    if (BestReg && Cost >= BestCost)
      continue;

    SpillPlacer->placeSpills(SpillConstraints, LiveBundles);
    // No live bundles, defer to splitSingleBlocks().
    if (!LiveBundles.any())
      continue;

    Cost += calcGlobalSplitCost(LiveBundles);
    if (!BestReg || Cost < BestCost) {
      BestReg = PhysReg;
      BestCost = Cost;
      BestBundles.swap(LiveBundles);
    }
  }

  if (!BestReg)
    return 0;

  splitAroundRegion(VirtReg, BestReg, BestBundles, NewVRegs);
  setStage(NewVRegs.begin(), NewVRegs.end(), RS_Region);
  return 0;
}


//===----------------------------------------------------------------------===//
//                             Local Splitting
//===----------------------------------------------------------------------===//


/// calcGapWeights - Compute the maximum spill weight that needs to be evicted
/// in order to use PhysReg between two entries in SA->UseSlots.
///
/// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
///
void RAGreedy::calcGapWeights(unsigned PhysReg,
                              SmallVectorImpl<float> &GapWeight) {
  assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
  const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
  const unsigned NumGaps = Uses.size()-1;

  // Start and end points for the interference check.
  SlotIndex StartIdx = BI.LiveIn ? BI.FirstUse.getBaseIndex() : BI.FirstUse;
  SlotIndex StopIdx = BI.LiveOut ? BI.LastUse.getBoundaryIndex() : BI.LastUse;

  GapWeight.assign(NumGaps, 0.0f);

  // Add interference from each overlapping register.
  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
    if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
           .checkInterference())
      continue;

    // We know that VirtReg is a continuous interval from FirstUse to LastUse,
    // so we don't need InterferenceQuery.
    //
    // Interference that overlaps an instruction is counted in both gaps
    // surrounding the instruction. The exception is interference before
    // StartIdx and after StopIdx.
    //
    LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
    for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
      // Skip the gaps before IntI.
      while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
        if (++Gap == NumGaps)
          break;
      if (Gap == NumGaps)
        break;

      // Update the gaps covered by IntI.
      const float weight = IntI.value()->weight;
      for (; Gap != NumGaps; ++Gap) {
        GapWeight[Gap] = std::max(GapWeight[Gap], weight);
        if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
          break;
      }
      if (Gap == NumGaps)
        break;
    }
  }
}

/// getPrevMappedIndex - Return the slot index of the last non-copy instruction
/// before MI that has a slot index. If MI is the first mapped instruction in
/// its block, return the block start index instead.
///
SlotIndex RAGreedy::getPrevMappedIndex(const MachineInstr *MI) {
  assert(MI && "Missing MachineInstr");
  const MachineBasicBlock *MBB = MI->getParent();
  MachineBasicBlock::const_iterator B = MBB->begin(), I = MI;
  while (I != B)
    if (!(--I)->isDebugValue() && !I->isCopy())
      return Indexes->getInstructionIndex(I);
  return Indexes->getMBBStartIdx(MBB);
}

/// calcPrevSlots - Fill in the PrevSlot array with the index of the previous
/// real non-copy instruction for each instruction in SA->UseSlots.
///
void RAGreedy::calcPrevSlots() {
  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
  PrevSlot.clear();
  PrevSlot.reserve(Uses.size());
  for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
    const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]);
    PrevSlot.push_back(getPrevMappedIndex(MI).getDefIndex());
  }
}

/// nextSplitPoint - Find the next index into SA->UseSlots > i such that it may
/// be beneficial to split before UseSlots[i].
///
/// 0 is always a valid split point
unsigned RAGreedy::nextSplitPoint(unsigned i) {
  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
  const unsigned Size = Uses.size();
  assert(i != Size && "No split points after the end");
  // Allow split before i when Uses[i] is not adjacent to the previous use.
  while (++i != Size && PrevSlot[i].getBaseIndex() <= Uses[i-1].getBaseIndex())
    ;
  return i;
}

/// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
/// basic block.
///
unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
                                 SmallVectorImpl<LiveInterval*> &NewVRegs) {
  assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
  const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();

  // Note that it is possible to have an interval that is live-in or live-out
  // while only covering a single block - A phi-def can use undef values from
  // predecessors, and the block could be a single-block loop.
  // We don't bother doing anything clever about such a case, we simply assume
  // that the interval is continuous from FirstUse to LastUse. We should make
  // sure that we don't do anything illegal to such an interval, though.

  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
  if (Uses.size() <= 2)
    return 0;
  const unsigned NumGaps = Uses.size()-1;

  DEBUG({
    dbgs() << "tryLocalSplit: ";
    for (unsigned i = 0, e = Uses.size(); i != e; ++i)
      dbgs() << ' ' << SA->UseSlots[i];
    dbgs() << '\n';
  });

  // For every use, find the previous mapped non-copy instruction.
  // We use this to detect valid split points, and to estimate new interval
  // sizes.
  calcPrevSlots();

  unsigned BestBefore = NumGaps;
  unsigned BestAfter = 0;
  float BestDiff = 0;

  const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB);
  SmallVector<float, 8> GapWeight;

  Order.rewind();
  while (unsigned PhysReg = Order.next()) {
    // Keep track of the largest spill weight that would need to be evicted in
    // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
    calcGapWeights(PhysReg, GapWeight);

    // Try to find the best sequence of gaps to close.
    // The new spill weight must be larger than any gap interference.

    // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
    unsigned SplitBefore = 0, SplitAfter = nextSplitPoint(1) - 1;

    // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
    // It is the spill weight that needs to be evicted.
    float MaxGap = GapWeight[0];
    for (unsigned i = 1; i != SplitAfter; ++i)
      MaxGap = std::max(MaxGap, GapWeight[i]);

    for (;;) {
      // Live before/after split?
      const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
      const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;

      DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
                   << Uses[SplitBefore] << '-' << Uses[SplitAfter]
                   << " i=" << MaxGap);

      // Stop before the interval gets so big we wouldn't be making progress.
      if (!LiveBefore && !LiveAfter) {
        DEBUG(dbgs() << " all\n");
        break;
      }
      // Should the interval be extended or shrunk?
      bool Shrink = true;
      if (MaxGap < HUGE_VALF) {
        // Estimate the new spill weight.
        //
        // Each instruction reads and writes the register, except the first
        // instr doesn't read when !FirstLive, and the last instr doesn't write
        // when !LastLive.
        //
        // We will be inserting copies before and after, so the total number of
        // reads and writes is 2 * EstUses.
        //
        const unsigned EstUses = 2*(SplitAfter - SplitBefore) +
                                 2*(LiveBefore + LiveAfter);

        // Try to guess the size of the new interval. This should be trivial,
        // but the slot index of an inserted copy can be a lot smaller than the
        // instruction it is inserted before if there are many dead indexes
        // between them.
        //
        // We measure the distance from the instruction before SplitBefore to
        // get a conservative estimate.
        //
        // The final distance can still be different if inserting copies
        // triggers a slot index renumbering.
        //
        const float EstWeight = normalizeSpillWeight(blockFreq * EstUses,
                              PrevSlot[SplitBefore].distance(Uses[SplitAfter]));
        // Would this split be possible to allocate?
        // Never allocate all gaps, we wouldn't be making progress.
        float Diff = EstWeight - MaxGap;
        DEBUG(dbgs() << " w=" << EstWeight << " d=" << Diff);
        if (Diff > 0) {
          Shrink = false;
          if (Diff > BestDiff) {
            DEBUG(dbgs() << " (best)");
            BestDiff = Diff;
            BestBefore = SplitBefore;
            BestAfter = SplitAfter;
          }
        }
      }

      // Try to shrink.
      if (Shrink) {
        SplitBefore = nextSplitPoint(SplitBefore);
        if (SplitBefore < SplitAfter) {
          DEBUG(dbgs() << " shrink\n");
          // Recompute the max when necessary.
          if (GapWeight[SplitBefore - 1] >= MaxGap) {
            MaxGap = GapWeight[SplitBefore];
            for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
              MaxGap = std::max(MaxGap, GapWeight[i]);
          }
          continue;
        }
        MaxGap = 0;
      }

      // Try to extend the interval.
      if (SplitAfter >= NumGaps) {
        DEBUG(dbgs() << " end\n");
        break;
      }

      DEBUG(dbgs() << " extend\n");
      for (unsigned e = nextSplitPoint(SplitAfter + 1) - 1;
           SplitAfter != e; ++SplitAfter)
        MaxGap = std::max(MaxGap, GapWeight[SplitAfter]);
          continue;
    }
  }

  // Didn't find any candidates?
  if (BestBefore == NumGaps)
    return 0;

  DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
               << '-' << Uses[BestAfter] << ", " << BestDiff
               << ", " << (BestAfter - BestBefore + 1) << " instrs\n");

  SmallVector<LiveInterval*, 4> SpillRegs;
  LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
  SE->reset(LREdit);

  SE->openIntv();
  SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
  SlotIndex SegStop  = SE->leaveIntvAfter(Uses[BestAfter]);
  SE->useIntv(SegStart, SegStop);
  SE->closeIntv();
  SE->finish();
  setStage(NewVRegs.begin(), NewVRegs.end(), RS_Local);
  ++NumLocalSplits;

  return 0;
}

//===----------------------------------------------------------------------===//
//                          Live Range Splitting
//===----------------------------------------------------------------------===//

/// trySplit - Try to split VirtReg or one of its interferences, making it
/// assignable.
/// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
                            SmallVectorImpl<LiveInterval*>&NewVRegs) {
  // Local intervals are handled separately.
  if (LIS->intervalIsInOneMBB(VirtReg)) {
    NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
    SA->analyze(&VirtReg);
    return tryLocalSplit(VirtReg, Order, NewVRegs);
  }

  NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);

  // Don't iterate global splitting.
  // Move straight to spilling if this range was produced by a global split.
  LiveRangeStage Stage = getStage(VirtReg);
  if (Stage >= RS_Block)
    return 0;

  SA->analyze(&VirtReg);

  // First try to split around a region spanning multiple blocks.
  if (Stage < RS_Region) {
    unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
    if (PhysReg || !NewVRegs.empty())
      return PhysReg;
  }

  // Then isolate blocks with multiple uses.
  if (Stage < RS_Block) {
    SplitAnalysis::BlockPtrSet Blocks;
    if (SA->getMultiUseBlocks(Blocks)) {
      SmallVector<LiveInterval*, 4> SpillRegs;
      LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
      SE->reset(LREdit);
      SE->splitSingleBlocks(Blocks);
      setStage(NewVRegs.begin(), NewVRegs.end(), RS_Block);
      if (VerifyEnabled)
        MF->verify(this, "After splitting live range around basic blocks");
    }
  }

  // Don't assign any physregs.
  return 0;
}


//===----------------------------------------------------------------------===//
//                            Main Entry Point
//===----------------------------------------------------------------------===//

unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
                                 SmallVectorImpl<LiveInterval*> &NewVRegs) {
  LiveRangeStage Stage = getStage(VirtReg);
  if (Stage == RS_Original)
    LRStage[VirtReg.reg] = RS_Second;

  // First try assigning a free register.
  AllocationOrder Order(VirtReg.reg, *VRM, ReservedRegs);
  while (unsigned PhysReg = Order.next()) {
    if (!checkPhysRegInterference(VirtReg, PhysReg))
      return PhysReg;
  }

  if (unsigned PhysReg = tryReassign(VirtReg, Order, NewVRegs))
    return PhysReg;

  if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
    return PhysReg;

  assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");

  // The first time we see a live range, don't try to split or spill.
  // Wait until the second time, when all smaller ranges have been allocated.
  // This gives a better picture of the interference to split around.
  if (Stage == RS_Original) {
    NewVRegs.push_back(&VirtReg);
    return 0;
  }

  assert(Stage < RS_Spill && "Cannot allocate after spilling");

  // Try splitting VirtReg or interferences.
  unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
  if (PhysReg || !NewVRegs.empty())
    return PhysReg;

  // Finally spill VirtReg itself.
  NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
  SmallVector<LiveInterval*, 1> pendingSpills;
  spiller().spill(&VirtReg, NewVRegs, pendingSpills);

  // The live virtual register requesting allocation was spilled, so tell
  // the caller not to allocate anything during this round.
  return 0;
}

bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
  DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
               << "********** Function: "
               << ((Value*)mf.getFunction())->getName() << '\n');

  MF = &mf;
  if (VerifyEnabled)
    MF->verify(this, "Before greedy register allocator");

  RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
  Indexes = &getAnalysis<SlotIndexes>();
  DomTree = &getAnalysis<MachineDominatorTree>();
  ReservedRegs = TRI->getReservedRegs(*MF);
  SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
  Loops = &getAnalysis<MachineLoopInfo>();
  LoopRanges = &getAnalysis<MachineLoopRanges>();
  Bundles = &getAnalysis<EdgeBundles>();
  SpillPlacer = &getAnalysis<SpillPlacement>();

  SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
  SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
  LRStage.clear();
  LRStage.resize(MRI->getNumVirtRegs());

  allocatePhysRegs();
  addMBBLiveIns(MF);
  LIS->addKillFlags();

  // Run rewriter
  {
    NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
    VRM->rewrite(Indexes);
  }

  // The pass output is in VirtRegMap. Release all the transient data.
  releaseMemory();

  return true;
}