GVN.cpp revision 30b4bd4d1099764db4f1e1a955b7f7cc9dafdd97 
1//===- GVN.cpp - Eliminate redundant values and loads ------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This pass performs global value numbering to eliminate fully redundant
11// instructions.  It also performs simple dead load elimination.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "gvn"
16
17#include "llvm/Transforms/Scalar.h"
18#include "llvm/BasicBlock.h"
19#include "llvm/Constants.h"
20#include "llvm/DerivedTypes.h"
21#include "llvm/Function.h"
22#include "llvm/IntrinsicInst.h"
23#include "llvm/Instructions.h"
24#include "llvm/Value.h"
25#include "llvm/ADT/BitVector.h"
26#include "llvm/ADT/DenseMap.h"
27#include "llvm/ADT/DepthFirstIterator.h"
28#include "llvm/ADT/SmallPtrSet.h"
29#include "llvm/ADT/SmallVector.h"
30#include "llvm/ADT/Statistic.h"
31#include "llvm/Analysis/Dominators.h"
32#include "llvm/Analysis/AliasAnalysis.h"
33#include "llvm/Analysis/MemoryDependenceAnalysis.h"
34#include "llvm/Support/CFG.h"
35#include "llvm/Support/Compiler.h"
36using namespace llvm;
37
38//===----------------------------------------------------------------------===//
39//                         ValueTable Class
40//===----------------------------------------------------------------------===//
41
42/// This class holds the mapping between values and value numbers.  It is used
43/// as an efficient mechanism to determine the expression-wise equivalence of
44/// two values.
45namespace {
46  struct VISIBILITY_HIDDEN Expression {
47    enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
48                            FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
49                            ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
50                            ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
51                            FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
52                            FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
53                            FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
54                            SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
55                            FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
56                            PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, EMPTY,
57                            TOMBSTONE };
58
59    ExpressionOpcode opcode;
60    const Type* type;
61    uint32_t firstVN;
62    uint32_t secondVN;
63    uint32_t thirdVN;
64    SmallVector<uint32_t, 4> varargs;
65    Value* function;
66
67    Expression() { }
68    Expression(ExpressionOpcode o) : opcode(o) { }
69
70    bool operator==(const Expression &other) const {
71      if (opcode != other.opcode)
72        return false;
73      else if (opcode == EMPTY || opcode == TOMBSTONE)
74        return true;
75      else if (type != other.type)
76        return false;
77      else if (function != other.function)
78        return false;
79      else if (firstVN != other.firstVN)
80        return false;
81      else if (secondVN != other.secondVN)
82        return false;
83      else if (thirdVN != other.thirdVN)
84        return false;
85      else {
86        if (varargs.size() != other.varargs.size())
87          return false;
88
89        for (size_t i = 0; i < varargs.size(); ++i)
90          if (varargs[i] != other.varargs[i])
91            return false;
92
93        return true;
94      }
95    }
96
97    bool operator!=(const Expression &other) const {
98      if (opcode != other.opcode)
99        return true;
100      else if (opcode == EMPTY || opcode == TOMBSTONE)
101        return false;
102      else if (type != other.type)
103        return true;
104      else if (function != other.function)
105        return true;
106      else if (firstVN != other.firstVN)
107        return true;
108      else if (secondVN != other.secondVN)
109        return true;
110      else if (thirdVN != other.thirdVN)
111        return true;
112      else {
113        if (varargs.size() != other.varargs.size())
114          return true;
115
116        for (size_t i = 0; i < varargs.size(); ++i)
117          if (varargs[i] != other.varargs[i])
118            return true;
119
120          return false;
121      }
122    }
123  };
124
125  class VISIBILITY_HIDDEN ValueTable {
126    private:
127      DenseMap<Value*, uint32_t> valueNumbering;
128      DenseMap<Expression, uint32_t> expressionNumbering;
129      AliasAnalysis* AA;
130
131      uint32_t nextValueNumber;
132
133      Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
134      Expression::ExpressionOpcode getOpcode(CmpInst* C);
135      Expression::ExpressionOpcode getOpcode(CastInst* C);
136      Expression create_expression(BinaryOperator* BO);
137      Expression create_expression(CmpInst* C);
138      Expression create_expression(ShuffleVectorInst* V);
139      Expression create_expression(ExtractElementInst* C);
140      Expression create_expression(InsertElementInst* V);
141      Expression create_expression(SelectInst* V);
142      Expression create_expression(CastInst* C);
143      Expression create_expression(GetElementPtrInst* G);
144      Expression create_expression(CallInst* C);
145    public:
146      ValueTable() : nextValueNumber(1) { }
147      uint32_t lookup_or_add(Value* V);
148      uint32_t lookup(Value* V) const;
149      void add(Value* V, uint32_t num);
150      void clear();
151      void erase(Value* v);
152      unsigned size();
153      void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
154      uint32_t hash_operand(Value* v);
155  };
156}
157
158namespace llvm {
159template <> struct DenseMapInfo<Expression> {
160  static inline Expression getEmptyKey() {
161    return Expression(Expression::EMPTY);
162  }
163
164  static inline Expression getTombstoneKey() {
165    return Expression(Expression::TOMBSTONE);
166  }
167
168  static unsigned getHashValue(const Expression e) {
169    unsigned hash = e.opcode;
170
171    hash = e.firstVN + hash * 37;
172    hash = e.secondVN + hash * 37;
173    hash = e.thirdVN + hash * 37;
174
175    hash = (unsigned)((uintptr_t)e.type >> 4) ^
176            (unsigned)((uintptr_t)e.type >> 9) +
177            hash * 37;
178
179    for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
180         E = e.varargs.end(); I != E; ++I)
181      hash = *I + hash * 37;
182
183    hash = (unsigned)((uintptr_t)e.function >> 4) ^
184            (unsigned)((uintptr_t)e.function >> 9) +
185            hash * 37;
186
187    return hash;
188  }
189  static bool isEqual(const Expression &LHS, const Expression &RHS) {
190    return LHS == RHS;
191  }
192  static bool isPod() { return true; }
193};
194}
195
196//===----------------------------------------------------------------------===//
197//                     ValueTable Internal Functions
198//===----------------------------------------------------------------------===//
199Expression::ExpressionOpcode
200                             ValueTable::getOpcode(BinaryOperator* BO) {
201  switch(BO->getOpcode()) {
202    case Instruction::Add:
203      return Expression::ADD;
204    case Instruction::Sub:
205      return Expression::SUB;
206    case Instruction::Mul:
207      return Expression::MUL;
208    case Instruction::UDiv:
209      return Expression::UDIV;
210    case Instruction::SDiv:
211      return Expression::SDIV;
212    case Instruction::FDiv:
213      return Expression::FDIV;
214    case Instruction::URem:
215      return Expression::UREM;
216    case Instruction::SRem:
217      return Expression::SREM;
218    case Instruction::FRem:
219      return Expression::FREM;
220    case Instruction::Shl:
221      return Expression::SHL;
222    case Instruction::LShr:
223      return Expression::LSHR;
224    case Instruction::AShr:
225      return Expression::ASHR;
226    case Instruction::And:
227      return Expression::AND;
228    case Instruction::Or:
229      return Expression::OR;
230    case Instruction::Xor:
231      return Expression::XOR;
232
233    // THIS SHOULD NEVER HAPPEN
234    default:
235      assert(0 && "Binary operator with unknown opcode?");
236      return Expression::ADD;
237  }
238}
239
240Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
241  if (C->getOpcode() == Instruction::ICmp) {
242    switch (C->getPredicate()) {
243      case ICmpInst::ICMP_EQ:
244        return Expression::ICMPEQ;
245      case ICmpInst::ICMP_NE:
246        return Expression::ICMPNE;
247      case ICmpInst::ICMP_UGT:
248        return Expression::ICMPUGT;
249      case ICmpInst::ICMP_UGE:
250        return Expression::ICMPUGE;
251      case ICmpInst::ICMP_ULT:
252        return Expression::ICMPULT;
253      case ICmpInst::ICMP_ULE:
254        return Expression::ICMPULE;
255      case ICmpInst::ICMP_SGT:
256        return Expression::ICMPSGT;
257      case ICmpInst::ICMP_SGE:
258        return Expression::ICMPSGE;
259      case ICmpInst::ICMP_SLT:
260        return Expression::ICMPSLT;
261      case ICmpInst::ICMP_SLE:
262        return Expression::ICMPSLE;
263
264      // THIS SHOULD NEVER HAPPEN
265      default:
266        assert(0 && "Comparison with unknown predicate?");
267        return Expression::ICMPEQ;
268    }
269  } else {
270    switch (C->getPredicate()) {
271      case FCmpInst::FCMP_OEQ:
272        return Expression::FCMPOEQ;
273      case FCmpInst::FCMP_OGT:
274        return Expression::FCMPOGT;
275      case FCmpInst::FCMP_OGE:
276        return Expression::FCMPOGE;
277      case FCmpInst::FCMP_OLT:
278        return Expression::FCMPOLT;
279      case FCmpInst::FCMP_OLE:
280        return Expression::FCMPOLE;
281      case FCmpInst::FCMP_ONE:
282        return Expression::FCMPONE;
283      case FCmpInst::FCMP_ORD:
284        return Expression::FCMPORD;
285      case FCmpInst::FCMP_UNO:
286        return Expression::FCMPUNO;
287      case FCmpInst::FCMP_UEQ:
288        return Expression::FCMPUEQ;
289      case FCmpInst::FCMP_UGT:
290        return Expression::FCMPUGT;
291      case FCmpInst::FCMP_UGE:
292        return Expression::FCMPUGE;
293      case FCmpInst::FCMP_ULT:
294        return Expression::FCMPULT;
295      case FCmpInst::FCMP_ULE:
296        return Expression::FCMPULE;
297      case FCmpInst::FCMP_UNE:
298        return Expression::FCMPUNE;
299
300      // THIS SHOULD NEVER HAPPEN
301      default:
302        assert(0 && "Comparison with unknown predicate?");
303        return Expression::FCMPOEQ;
304    }
305  }
306}
307
308Expression::ExpressionOpcode
309                             ValueTable::getOpcode(CastInst* C) {
310  switch(C->getOpcode()) {
311    case Instruction::Trunc:
312      return Expression::TRUNC;
313    case Instruction::ZExt:
314      return Expression::ZEXT;
315    case Instruction::SExt:
316      return Expression::SEXT;
317    case Instruction::FPToUI:
318      return Expression::FPTOUI;
319    case Instruction::FPToSI:
320      return Expression::FPTOSI;
321    case Instruction::UIToFP:
322      return Expression::UITOFP;
323    case Instruction::SIToFP:
324      return Expression::SITOFP;
325    case Instruction::FPTrunc:
326      return Expression::FPTRUNC;
327    case Instruction::FPExt:
328      return Expression::FPEXT;
329    case Instruction::PtrToInt:
330      return Expression::PTRTOINT;
331    case Instruction::IntToPtr:
332      return Expression::INTTOPTR;
333    case Instruction::BitCast:
334      return Expression::BITCAST;
335
336    // THIS SHOULD NEVER HAPPEN
337    default:
338      assert(0 && "Cast operator with unknown opcode?");
339      return Expression::BITCAST;
340  }
341}
342
343uint32_t ValueTable::hash_operand(Value* v) {
344  if (CallInst* CI = dyn_cast<CallInst>(v))
345    if (!AA->doesNotAccessMemory(CI))
346      return nextValueNumber++;
347
348  return lookup_or_add(v);
349}
350
351Expression ValueTable::create_expression(CallInst* C) {
352  Expression e;
353
354  e.type = C->getType();
355  e.firstVN = 0;
356  e.secondVN = 0;
357  e.thirdVN = 0;
358  e.function = C->getCalledFunction();
359  e.opcode = Expression::CALL;
360
361  for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
362       I != E; ++I)
363    e.varargs.push_back(hash_operand(*I));
364
365  return e;
366}
367
368Expression ValueTable::create_expression(BinaryOperator* BO) {
369  Expression e;
370
371  e.firstVN = hash_operand(BO->getOperand(0));
372  e.secondVN = hash_operand(BO->getOperand(1));
373  e.thirdVN = 0;
374  e.function = 0;
375  e.type = BO->getType();
376  e.opcode = getOpcode(BO);
377
378  return e;
379}
380
381Expression ValueTable::create_expression(CmpInst* C) {
382  Expression e;
383
384  e.firstVN = hash_operand(C->getOperand(0));
385  e.secondVN = hash_operand(C->getOperand(1));
386  e.thirdVN = 0;
387  e.function = 0;
388  e.type = C->getType();
389  e.opcode = getOpcode(C);
390
391  return e;
392}
393
394Expression ValueTable::create_expression(CastInst* C) {
395  Expression e;
396
397  e.firstVN = hash_operand(C->getOperand(0));
398  e.secondVN = 0;
399  e.thirdVN = 0;
400  e.function = 0;
401  e.type = C->getType();
402  e.opcode = getOpcode(C);
403
404  return e;
405}
406
407Expression ValueTable::create_expression(ShuffleVectorInst* S) {
408  Expression e;
409
410  e.firstVN = hash_operand(S->getOperand(0));
411  e.secondVN = hash_operand(S->getOperand(1));
412  e.thirdVN = hash_operand(S->getOperand(2));
413  e.function = 0;
414  e.type = S->getType();
415  e.opcode = Expression::SHUFFLE;
416
417  return e;
418}
419
420Expression ValueTable::create_expression(ExtractElementInst* E) {
421  Expression e;
422
423  e.firstVN = hash_operand(E->getOperand(0));
424  e.secondVN = hash_operand(E->getOperand(1));
425  e.thirdVN = 0;
426  e.function = 0;
427  e.type = E->getType();
428  e.opcode = Expression::EXTRACT;
429
430  return e;
431}
432
433Expression ValueTable::create_expression(InsertElementInst* I) {
434  Expression e;
435
436  e.firstVN = hash_operand(I->getOperand(0));
437  e.secondVN = hash_operand(I->getOperand(1));
438  e.thirdVN = hash_operand(I->getOperand(2));
439  e.function = 0;
440  e.type = I->getType();
441  e.opcode = Expression::INSERT;
442
443  return e;
444}
445
446Expression ValueTable::create_expression(SelectInst* I) {
447  Expression e;
448
449  e.firstVN = hash_operand(I->getCondition());
450  e.secondVN = hash_operand(I->getTrueValue());
451  e.thirdVN = hash_operand(I->getFalseValue());
452  e.function = 0;
453  e.type = I->getType();
454  e.opcode = Expression::SELECT;
455
456  return e;
457}
458
459Expression ValueTable::create_expression(GetElementPtrInst* G) {
460  Expression e;
461
462  e.firstVN = hash_operand(G->getPointerOperand());
463  e.secondVN = 0;
464  e.thirdVN = 0;
465  e.function = 0;
466  e.type = G->getType();
467  e.opcode = Expression::GEP;
468
469  for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
470       I != E; ++I)
471    e.varargs.push_back(hash_operand(*I));
472
473  return e;
474}
475
476//===----------------------------------------------------------------------===//
477//                     ValueTable External Functions
478//===----------------------------------------------------------------------===//
479
480/// lookup_or_add - Returns the value number for the specified value, assigning
481/// it a new number if it did not have one before.
482uint32_t ValueTable::lookup_or_add(Value* V) {
483  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
484  if (VI != valueNumbering.end())
485    return VI->second;
486
487  if (CallInst* C = dyn_cast<CallInst>(V)) {
488    if (AA->onlyReadsMemory(C)) { // includes doesNotAccessMemory
489      Expression e = create_expression(C);
490
491      DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
492      if (EI != expressionNumbering.end()) {
493        valueNumbering.insert(std::make_pair(V, EI->second));
494        return EI->second;
495      } else {
496        expressionNumbering.insert(std::make_pair(e, nextValueNumber));
497        valueNumbering.insert(std::make_pair(V, nextValueNumber));
498
499        return nextValueNumber++;
500      }
501    } else {
502      valueNumbering.insert(std::make_pair(V, nextValueNumber));
503      return nextValueNumber++;
504    }
505  } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
506    Expression e = create_expression(BO);
507
508    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
509    if (EI != expressionNumbering.end()) {
510      valueNumbering.insert(std::make_pair(V, EI->second));
511      return EI->second;
512    } else {
513      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
514      valueNumbering.insert(std::make_pair(V, nextValueNumber));
515
516      return nextValueNumber++;
517    }
518  } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
519    Expression e = create_expression(C);
520
521    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
522    if (EI != expressionNumbering.end()) {
523      valueNumbering.insert(std::make_pair(V, EI->second));
524      return EI->second;
525    } else {
526      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
527      valueNumbering.insert(std::make_pair(V, nextValueNumber));
528
529      return nextValueNumber++;
530    }
531  } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
532    Expression e = create_expression(U);
533
534    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
535    if (EI != expressionNumbering.end()) {
536      valueNumbering.insert(std::make_pair(V, EI->second));
537      return EI->second;
538    } else {
539      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
540      valueNumbering.insert(std::make_pair(V, nextValueNumber));
541
542      return nextValueNumber++;
543    }
544  } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
545    Expression e = create_expression(U);
546
547    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
548    if (EI != expressionNumbering.end()) {
549      valueNumbering.insert(std::make_pair(V, EI->second));
550      return EI->second;
551    } else {
552      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
553      valueNumbering.insert(std::make_pair(V, nextValueNumber));
554
555      return nextValueNumber++;
556    }
557  } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
558    Expression e = create_expression(U);
559
560    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
561    if (EI != expressionNumbering.end()) {
562      valueNumbering.insert(std::make_pair(V, EI->second));
563      return EI->second;
564    } else {
565      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
566      valueNumbering.insert(std::make_pair(V, nextValueNumber));
567
568      return nextValueNumber++;
569    }
570  } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
571    Expression e = create_expression(U);
572
573    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
574    if (EI != expressionNumbering.end()) {
575      valueNumbering.insert(std::make_pair(V, EI->second));
576      return EI->second;
577    } else {
578      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
579      valueNumbering.insert(std::make_pair(V, nextValueNumber));
580
581      return nextValueNumber++;
582    }
583  } else if (CastInst* U = dyn_cast<CastInst>(V)) {
584    Expression e = create_expression(U);
585
586    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
587    if (EI != expressionNumbering.end()) {
588      valueNumbering.insert(std::make_pair(V, EI->second));
589      return EI->second;
590    } else {
591      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
592      valueNumbering.insert(std::make_pair(V, nextValueNumber));
593
594      return nextValueNumber++;
595    }
596  } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
597    Expression e = create_expression(U);
598
599    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
600    if (EI != expressionNumbering.end()) {
601      valueNumbering.insert(std::make_pair(V, EI->second));
602      return EI->second;
603    } else {
604      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
605      valueNumbering.insert(std::make_pair(V, nextValueNumber));
606
607      return nextValueNumber++;
608    }
609  } else {
610    valueNumbering.insert(std::make_pair(V, nextValueNumber));
611    return nextValueNumber++;
612  }
613}
614
615/// lookup - Returns the value number of the specified value. Fails if
616/// the value has not yet been numbered.
617uint32_t ValueTable::lookup(Value* V) const {
618  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
619  if (VI != valueNumbering.end())
620    return VI->second;
621  else
622    assert(0 && "Value not numbered?");
623
624  return 0;
625}
626
627/// clear - Remove all entries from the ValueTable
628void ValueTable::clear() {
629  valueNumbering.clear();
630  expressionNumbering.clear();
631  nextValueNumber = 1;
632}
633
634/// erase - Remove a value from the value numbering
635void ValueTable::erase(Value* V) {
636  valueNumbering.erase(V);
637}
638
639//===----------------------------------------------------------------------===//
640//                       ValueNumberedSet Class
641//===----------------------------------------------------------------------===//
642namespace {
643class ValueNumberedSet {
644  private:
645    SmallPtrSet<Value*, 8> contents;
646    BitVector numbers;
647  public:
648    ValueNumberedSet() { numbers.resize(1); }
649    ValueNumberedSet(const ValueNumberedSet& other) {
650      numbers = other.numbers;
651      contents = other.contents;
652    }
653
654    typedef SmallPtrSet<Value*, 8>::iterator iterator;
655
656    iterator begin() { return contents.begin(); }
657    iterator end() { return contents.end(); }
658
659    bool insert(Value* v) { return contents.insert(v); }
660    void insert(iterator I, iterator E) { contents.insert(I, E); }
661    void erase(Value* v) { contents.erase(v); }
662    unsigned count(Value* v) { return contents.count(v); }
663    size_t size() { return contents.size(); }
664
665    void set(unsigned i)  {
666      if (i >= numbers.size())
667        numbers.resize(i+1);
668
669      numbers.set(i);
670    }
671
672    void operator=(const ValueNumberedSet& other) {
673      contents = other.contents;
674      numbers = other.numbers;
675    }
676
677    void reset(unsigned i)  {
678      if (i < numbers.size())
679        numbers.reset(i);
680    }
681
682    bool test(unsigned i)  {
683      if (i >= numbers.size())
684        return false;
685
686      return numbers.test(i);
687    }
688
689    void clear() {
690      contents.clear();
691      numbers.clear();
692    }
693};
694}
695
696//===----------------------------------------------------------------------===//
697//                         GVN Pass
698//===----------------------------------------------------------------------===//
699
700namespace {
701
702  class VISIBILITY_HIDDEN GVN : public FunctionPass {
703    bool runOnFunction(Function &F);
704  public:
705    static char ID; // Pass identification, replacement for typeid
706    GVN() : FunctionPass((intptr_t)&ID) { }
707
708  private:
709    ValueTable VN;
710
711    DenseMap<BasicBlock*, ValueNumberedSet> availableOut;
712
713    typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
714    PhiMapType phiMap;
715
716
717    // This transformation requires dominator postdominator info
718    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
719      AU.setPreservesCFG();
720      AU.addRequired<DominatorTree>();
721      AU.addRequired<MemoryDependenceAnalysis>();
722      AU.addRequired<AliasAnalysis>();
723      AU.addPreserved<AliasAnalysis>();
724      AU.addPreserved<MemoryDependenceAnalysis>();
725    }
726
727    // Helper fuctions
728    // FIXME: eliminate or document these better
729    Value* find_leader(ValueNumberedSet& vals, uint32_t v) ;
730    void val_insert(ValueNumberedSet& s, Value* v);
731    bool processLoad(LoadInst* L,
732                     DenseMap<Value*, LoadInst*>& lastLoad,
733                     SmallVector<Instruction*, 4>& toErase);
734    bool processInstruction(Instruction* I,
735                            ValueNumberedSet& currAvail,
736                            DenseMap<Value*, LoadInst*>& lastSeenLoad,
737                            SmallVector<Instruction*, 4>& toErase);
738    bool processNonLocalLoad(LoadInst* L,
739                             SmallVector<Instruction*, 4>& toErase);
740    bool processMemCpy(MemCpyInst* M, SmallVector<Instruction*, 4>& toErase);
741    Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
742                            DenseMap<BasicBlock*, Value*> &Phis,
743                            bool top_level = false);
744    void dump(DenseMap<BasicBlock*, Value*>& d);
745    bool iterateOnFunction(Function &F);
746    Value* CollapsePhi(PHINode* p);
747    bool isSafeReplacement(PHINode* p, Instruction* inst);
748  };
749
750  char GVN::ID = 0;
751
752}
753
754// createGVNPass - The public interface to this file...
755FunctionPass *llvm::createGVNPass() { return new GVN(); }
756
757static RegisterPass<GVN> X("gvn",
758                           "Global Value Numbering");
759
760STATISTIC(NumGVNInstr, "Number of instructions deleted");
761STATISTIC(NumGVNLoad, "Number of loads deleted");
762
763/// find_leader - Given a set and a value number, return the first
764/// element of the set with that value number, or 0 if no such element
765/// is present
766Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) {
767  if (!vals.test(v))
768    return 0;
769
770  for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end();
771       I != E; ++I)
772    if (v == VN.lookup(*I))
773      return *I;
774
775  assert(0 && "No leader found, but present bit is set?");
776  return 0;
777}
778
779/// val_insert - Insert a value into a set only if there is not a value
780/// with the same value number already in the set
781void GVN::val_insert(ValueNumberedSet& s, Value* v) {
782  uint32_t num = VN.lookup(v);
783  if (!s.test(num))
784    s.insert(v);
785}
786
787void GVN::dump(DenseMap<BasicBlock*, Value*>& d) {
788  printf("{\n");
789  for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(),
790       E = d.end(); I != E; ++I) {
791    if (I->second == MemoryDependenceAnalysis::None)
792      printf("None\n");
793    else
794      I->second->dump();
795  }
796  printf("}\n");
797}
798
799Value* GVN::CollapsePhi(PHINode* p) {
800  DominatorTree &DT = getAnalysis<DominatorTree>();
801  Value* constVal = p->hasConstantValue();
802
803  if (constVal) {
804    if (Instruction* inst = dyn_cast<Instruction>(constVal)) {
805      if (DT.dominates(inst, p))
806        if (isSafeReplacement(p, inst))
807          return inst;
808    } else {
809      return constVal;
810    }
811  }
812
813  return 0;
814}
815
816bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
817  if (!isa<PHINode>(inst))
818    return true;
819
820  for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
821       UI != E; ++UI)
822    if (PHINode* use_phi = dyn_cast<PHINode>(UI))
823      if (use_phi->getParent() == inst->getParent())
824        return false;
825
826  return true;
827}
828
829/// GetValueForBlock - Get the value to use within the specified basic block.
830/// available values are in Phis.
831Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
832                               DenseMap<BasicBlock*, Value*> &Phis,
833                               bool top_level) {
834
835  // If we have already computed this value, return the previously computed val.
836  DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
837  if (V != Phis.end() && !top_level) return V->second;
838
839  BasicBlock* singlePred = BB->getSinglePredecessor();
840  if (singlePred) {
841    Value *ret = GetValueForBlock(singlePred, orig, Phis);
842    Phis[BB] = ret;
843    return ret;
844  }
845  // Otherwise, the idom is the loop, so we need to insert a PHI node.  Do so
846  // now, then get values to fill in the incoming values for the PHI.
847  PHINode *PN = new PHINode(orig->getType(), orig->getName()+".rle",
848                            BB->begin());
849  PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
850
851  if (Phis.count(BB) == 0)
852    Phis.insert(std::make_pair(BB, PN));
853
854  // Fill in the incoming values for the block.
855  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
856    Value* val = GetValueForBlock(*PI, orig, Phis);
857
858    PN->addIncoming(val, *PI);
859  }
860
861  // Attempt to collapse PHI nodes that are trivially redundant
862  Value* v = CollapsePhi(PN);
863  if (v) {
864    MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
865
866    MD.removeInstruction(PN);
867    PN->replaceAllUsesWith(v);
868
869    for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
870         E = Phis.end(); I != E; ++I)
871      if (I->second == PN)
872        I->second = v;
873
874    PN->eraseFromParent();
875
876    Phis[BB] = v;
877
878    return v;
879  }
880
881  // Cache our phi construction results
882  phiMap[orig->getPointerOperand()].insert(PN);
883  return PN;
884}
885
886/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
887/// non-local by performing PHI construction.
888bool GVN::processNonLocalLoad(LoadInst* L,
889                              SmallVector<Instruction*, 4>& toErase) {
890  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
891
892  // Find the non-local dependencies of the load
893  DenseMap<BasicBlock*, Value*> deps;
894  MD.getNonLocalDependency(L, deps);
895
896  DenseMap<BasicBlock*, Value*> repl;
897
898  // Filter out useless results (non-locals, etc)
899  for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end();
900       I != E; ++I)
901    if (I->second == MemoryDependenceAnalysis::None) {
902      return false;
903    } else if (I->second == MemoryDependenceAnalysis::NonLocal) {
904      continue;
905    } else if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
906      if (S->getPointerOperand() == L->getPointerOperand())
907        repl[I->first] = S->getOperand(0);
908      else
909        return false;
910    } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) {
911      if (LD->getPointerOperand() == L->getPointerOperand())
912        repl[I->first] = LD;
913      else
914        return false;
915    } else {
916      return false;
917    }
918
919  // Use cached PHI construction information from previous runs
920  SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
921  for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
922       I != E; ++I) {
923    if ((*I)->getParent() == L->getParent()) {
924      MD.removeInstruction(L);
925      L->replaceAllUsesWith(*I);
926      toErase.push_back(L);
927      NumGVNLoad++;
928
929      return true;
930    } else {
931      repl.insert(std::make_pair((*I)->getParent(), *I));
932    }
933  }
934
935  // Perform PHI construction
936  SmallPtrSet<BasicBlock*, 4> visited;
937  Value* v = GetValueForBlock(L->getParent(), L, repl, true);
938
939  MD.removeInstruction(L);
940  L->replaceAllUsesWith(v);
941  toErase.push_back(L);
942  NumGVNLoad++;
943
944  return true;
945}
946
947/// processLoad - Attempt to eliminate a load, first by eliminating it
948/// locally, and then attempting non-local elimination if that fails.
949bool GVN::processLoad(LoadInst* L,
950                         DenseMap<Value*, LoadInst*>& lastLoad,
951                         SmallVector<Instruction*, 4>& toErase) {
952  if (L->isVolatile()) {
953    lastLoad[L->getPointerOperand()] = L;
954    return false;
955  }
956
957  Value* pointer = L->getPointerOperand();
958  LoadInst*& last = lastLoad[pointer];
959
960  // ... to a pointer that has been loaded from before...
961  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
962  bool removedNonLocal = false;
963  Instruction* dep = MD.getDependency(L);
964  if (dep == MemoryDependenceAnalysis::NonLocal &&
965      L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
966    removedNonLocal = processNonLocalLoad(L, toErase);
967
968    if (!removedNonLocal)
969      last = L;
970
971    return removedNonLocal;
972  }
973
974
975  bool deletedLoad = false;
976
977  // Walk up the dependency chain until we either find
978  // a dependency we can use, or we can't walk any further
979  while (dep != MemoryDependenceAnalysis::None &&
980         dep != MemoryDependenceAnalysis::NonLocal &&
981         (isa<LoadInst>(dep) || isa<StoreInst>(dep))) {
982    // ... that depends on a store ...
983    if (StoreInst* S = dyn_cast<StoreInst>(dep)) {
984      if (S->getPointerOperand() == pointer) {
985        // Remove it!
986        MD.removeInstruction(L);
987
988        L->replaceAllUsesWith(S->getOperand(0));
989        toErase.push_back(L);
990        deletedLoad = true;
991        NumGVNLoad++;
992      }
993
994      // Whether we removed it or not, we can't
995      // go any further
996      break;
997    } else if (!last) {
998      // If we don't depend on a store, and we haven't
999      // been loaded before, bail.
1000      break;
1001    } else if (dep == last) {
1002      // Remove it!
1003      MD.removeInstruction(L);
1004
1005      L->replaceAllUsesWith(last);
1006      toErase.push_back(L);
1007      deletedLoad = true;
1008      NumGVNLoad++;
1009
1010      break;
1011    } else {
1012      dep = MD.getDependency(L, dep);
1013    }
1014  }
1015
1016  if (dep != MemoryDependenceAnalysis::None &&
1017      dep != MemoryDependenceAnalysis::NonLocal &&
1018      isa<AllocationInst>(dep)) {
1019    // Check that this load is actually from the
1020    // allocation we found
1021    Value* v = L->getOperand(0);
1022    while (true) {
1023      if (BitCastInst *BC = dyn_cast<BitCastInst>(v))
1024        v = BC->getOperand(0);
1025      else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(v))
1026        v = GEP->getOperand(0);
1027      else
1028        break;
1029    }
1030    if (v == dep) {
1031      // If this load depends directly on an allocation, there isn't
1032      // anything stored there; therefore, we can optimize this load
1033      // to undef.
1034      MD.removeInstruction(L);
1035
1036      L->replaceAllUsesWith(UndefValue::get(L->getType()));
1037      toErase.push_back(L);
1038      deletedLoad = true;
1039      NumGVNLoad++;
1040    }
1041  }
1042
1043  if (!deletedLoad)
1044    last = L;
1045
1046  return deletedLoad;
1047}
1048
1049/// processMemCpy - perform simplication of memcpy's.  If we have memcpy A which
1050/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
1051/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
1052///  This allows later passes to remove the first memcpy altogether.
1053bool GVN::processMemCpy(MemCpyInst* M,
1054                        SmallVector<Instruction*, 4>& toErase) {
1055  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1056
1057  // First, we have to check that the dependency is another memcpy
1058  Instruction* dep = MD.getDependency(M);
1059  if  (dep == MemoryDependenceAnalysis::None ||
1060       dep == MemoryDependenceAnalysis::NonLocal ||
1061       !isa<MemCpyInst>(dep))
1062    return false;
1063
1064  // We can only transforms memcpy's where the dest of one is the source of the
1065  // other
1066  MemCpyInst* MDep = cast<MemCpyInst>(dep);
1067  if (M->getSource() != MDep->getDest())
1068    return false;
1069
1070  // Second, the length of the memcpy's must be the same, or the preceeding one
1071  // must be larger than the following one.
1072  ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength());
1073  ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength());
1074  if (!C1 || !C2)
1075    return false;
1076
1077  uint64_t CpySize = C1->getValue().getZExtValue();
1078  uint64_t DepSize = C2->getValue().getZExtValue();
1079
1080  if (DepSize < CpySize)
1081    return false;
1082
1083  // Finally, we have to make sure that the dest of the second does not
1084  // alias the source of the first
1085  AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
1086  if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
1087      AliasAnalysis::NoAlias)
1088    return false;
1089  else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
1090           AliasAnalysis::NoAlias)
1091    return false;
1092  else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
1093           != AliasAnalysis::NoAlias)
1094    return false;
1095
1096  // If all checks passed, then we can transform these memcpy's
1097  bool is32bit = M->getIntrinsicID() == Intrinsic::memcpy_i32;
1098  Function* MemMoveFun = Intrinsic::getDeclaration(
1099                                 M->getParent()->getParent()->getParent(),
1100                                 is32bit ? Intrinsic::memcpy_i32 :
1101                                           Intrinsic::memcpy_i64);
1102
1103  std::vector<Value*> args;
1104  args.push_back(M->getRawDest());
1105  args.push_back(MDep->getRawSource());
1106  args.push_back(M->getLength());
1107  args.push_back(M->getAlignment());
1108
1109  CallInst* C = new CallInst(MemMoveFun, args.begin(), args.end(), "", M);
1110
1111  if (MD.getDependency(C) == MDep) {
1112    MD.dropInstruction(M);
1113    toErase.push_back(M);
1114    return true;
1115  } else {
1116    MD.removeInstruction(C);
1117    toErase.push_back(C);
1118    return false;
1119  }
1120}
1121
1122/// processInstruction - When calculating availability, handle an instruction
1123/// by inserting it into the appropriate sets
1124bool GVN::processInstruction(Instruction* I,
1125                                ValueNumberedSet& currAvail,
1126                                DenseMap<Value*, LoadInst*>& lastSeenLoad,
1127                                SmallVector<Instruction*, 4>& toErase) {
1128  if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1129    return processLoad(L, lastSeenLoad, toErase);
1130  } else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
1131    return processMemCpy(M, toErase);
1132  }
1133
1134  unsigned num = VN.lookup_or_add(I);
1135
1136  // Collapse PHI nodes
1137  if (PHINode* p = dyn_cast<PHINode>(I)) {
1138    Value* constVal = CollapsePhi(p);
1139
1140    if (constVal) {
1141      for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1142           PI != PE; ++PI)
1143        if (PI->second.count(p))
1144          PI->second.erase(p);
1145
1146      p->replaceAllUsesWith(constVal);
1147      toErase.push_back(p);
1148    }
1149  // Perform value-number based elimination
1150  } else if (currAvail.test(num)) {
1151    Value* repl = find_leader(currAvail, num);
1152
1153    if (CallInst* CI = dyn_cast<CallInst>(I)) {
1154      AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
1155      if (!AA.doesNotAccessMemory(CI)) {
1156        MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1157        if (cast<Instruction>(repl)->getParent() != CI->getParent() ||
1158            MD.getDependency(CI) != MD.getDependency(cast<CallInst>(repl))) {
1159          // There must be an intervening may-alias store, so nothing from
1160          // this point on will be able to be replaced with the preceding call
1161          currAvail.erase(repl);
1162          currAvail.insert(I);
1163
1164          return false;
1165        }
1166      }
1167    }
1168
1169    // Remove it!
1170    MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1171    MD.removeInstruction(I);
1172
1173    VN.erase(I);
1174    I->replaceAllUsesWith(repl);
1175    toErase.push_back(I);
1176    return true;
1177  } else if (!I->isTerminator()) {
1178    currAvail.set(num);
1179    currAvail.insert(I);
1180  }
1181
1182  return false;
1183}
1184
1185// GVN::runOnFunction - This is the main transformation entry point for a
1186// function.
1187//
1188bool GVN::runOnFunction(Function& F) {
1189  VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1190
1191  bool changed = false;
1192  bool shouldContinue = true;
1193
1194  while (shouldContinue) {
1195    shouldContinue = iterateOnFunction(F);
1196    changed |= shouldContinue;
1197  }
1198
1199  return changed;
1200}
1201
1202
1203// GVN::iterateOnFunction - Executes one iteration of GVN
1204bool GVN::iterateOnFunction(Function &F) {
1205  // Clean out global sets from any previous functions
1206  VN.clear();
1207  availableOut.clear();
1208  phiMap.clear();
1209
1210  bool changed_function = false;
1211
1212  DominatorTree &DT = getAnalysis<DominatorTree>();
1213
1214  SmallVector<Instruction*, 4> toErase;
1215
1216  // Top-down walk of the dominator tree
1217  for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
1218         E = df_end(DT.getRootNode()); DI != E; ++DI) {
1219
1220    // Get the set to update for this block
1221    ValueNumberedSet& currAvail = availableOut[DI->getBlock()];
1222    DenseMap<Value*, LoadInst*> lastSeenLoad;
1223
1224    BasicBlock* BB = DI->getBlock();
1225
1226    // A block inherits AVAIL_OUT from its dominator
1227    if (DI->getIDom() != 0)
1228      currAvail = availableOut[DI->getIDom()->getBlock()];
1229
1230    for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1231         BI != BE; ) {
1232      changed_function |= processInstruction(BI, currAvail,
1233                                             lastSeenLoad, toErase);
1234
1235      NumGVNInstr += toErase.size();
1236
1237      // Avoid iterator invalidation
1238      ++BI;
1239
1240      for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1241           E = toErase.end(); I != E; ++I) {
1242        (*I)->eraseFromParent();
1243      }
1244
1245      toErase.clear();
1246    }
1247  }
1248
1249  return changed_function;
1250}
1251