GVN.cpp revision 2486686689af91c02357fd0925ed213e4dfe0ed1 
1//===- GVN.cpp - Eliminate redundant values and loads ------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the Owen Anderson and is distributed under
6// the University of Illinois Open Source 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/Instructions.h"
23#include "llvm/Value.h"
24#include "llvm/Analysis/Dominators.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/MemoryDependenceAnalysis.h"
32#include "llvm/Support/CFG.h"
33#include "llvm/Support/Compiler.h"
34using namespace llvm;
35
36//===----------------------------------------------------------------------===//
37//                         ValueTable Class
38//===----------------------------------------------------------------------===//
39
40/// This class holds the mapping between values and value numbers.  It is used
41/// as an efficient mechanism to determine the expression-wise equivalence of
42/// two values.
43namespace {
44  struct VISIBILITY_HIDDEN Expression {
45    enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
46                            FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
47                            ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
48                            ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
49                            FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
50                            FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
51                            FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
52                            SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
53                            FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
54                            PTRTOINT, INTTOPTR, BITCAST, GEP, EMPTY,
55                            TOMBSTONE };
56
57    ExpressionOpcode opcode;
58    const Type* type;
59    uint32_t firstVN;
60    uint32_t secondVN;
61    uint32_t thirdVN;
62    SmallVector<uint32_t, 4> varargs;
63
64    Expression() { }
65    Expression(ExpressionOpcode o) : opcode(o) { }
66
67    bool operator==(const Expression &other) const {
68      if (opcode != other.opcode)
69        return false;
70      else if (opcode == EMPTY || opcode == TOMBSTONE)
71        return true;
72      else if (type != other.type)
73        return false;
74      else if (firstVN != other.firstVN)
75        return false;
76      else if (secondVN != other.secondVN)
77        return false;
78      else if (thirdVN != other.thirdVN)
79        return false;
80      else {
81        if (varargs.size() != other.varargs.size())
82          return false;
83
84        for (size_t i = 0; i < varargs.size(); ++i)
85          if (varargs[i] != other.varargs[i])
86            return false;
87
88        return true;
89      }
90    }
91
92    bool operator!=(const Expression &other) const {
93      if (opcode != other.opcode)
94        return true;
95      else if (opcode == EMPTY || opcode == TOMBSTONE)
96        return false;
97      else if (type != other.type)
98        return true;
99      else if (firstVN != other.firstVN)
100        return true;
101      else if (secondVN != other.secondVN)
102        return true;
103      else if (thirdVN != other.thirdVN)
104        return true;
105      else {
106        if (varargs.size() != other.varargs.size())
107          return true;
108
109        for (size_t i = 0; i < varargs.size(); ++i)
110          if (varargs[i] != other.varargs[i])
111            return true;
112
113          return false;
114      }
115    }
116  };
117
118  class VISIBILITY_HIDDEN ValueTable {
119    private:
120      DenseMap<Value*, uint32_t> valueNumbering;
121      DenseMap<Expression, uint32_t> expressionNumbering;
122
123      uint32_t nextValueNumber;
124
125      Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
126      Expression::ExpressionOpcode getOpcode(CmpInst* C);
127      Expression::ExpressionOpcode getOpcode(CastInst* C);
128      Expression create_expression(BinaryOperator* BO);
129      Expression create_expression(CmpInst* C);
130      Expression create_expression(ShuffleVectorInst* V);
131      Expression create_expression(ExtractElementInst* C);
132      Expression create_expression(InsertElementInst* V);
133      Expression create_expression(SelectInst* V);
134      Expression create_expression(CastInst* C);
135      Expression create_expression(GetElementPtrInst* G);
136    public:
137      ValueTable() { nextValueNumber = 1; }
138      uint32_t lookup_or_add(Value* V);
139      uint32_t lookup(Value* V) const;
140      void add(Value* V, uint32_t num);
141      void clear();
142      void erase(Value* v);
143      unsigned size();
144  };
145}
146
147namespace llvm {
148template <> struct DenseMapKeyInfo<Expression> {
149  static inline Expression getEmptyKey() {
150    return Expression(Expression::EMPTY);
151  }
152
153  static inline Expression getTombstoneKey() {
154    return Expression(Expression::TOMBSTONE);
155  }
156
157  static unsigned getHashValue(const Expression e) {
158    unsigned hash = e.opcode;
159
160    hash = e.firstVN + hash * 37;
161    hash = e.secondVN + hash * 37;
162    hash = e.thirdVN + hash * 37;
163
164    hash = (unsigned)((uintptr_t)e.type >> 4) ^
165            (unsigned)((uintptr_t)e.type >> 9) +
166            hash * 37;
167
168    for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
169         E = e.varargs.end(); I != E; ++I)
170      hash = *I + hash * 37;
171
172    return hash;
173  }
174  static bool isPod() { return true; }
175};
176}
177
178//===----------------------------------------------------------------------===//
179//                     ValueTable Internal Functions
180//===----------------------------------------------------------------------===//
181Expression::ExpressionOpcode
182                             ValueTable::getOpcode(BinaryOperator* BO) {
183  switch(BO->getOpcode()) {
184    case Instruction::Add:
185      return Expression::ADD;
186    case Instruction::Sub:
187      return Expression::SUB;
188    case Instruction::Mul:
189      return Expression::MUL;
190    case Instruction::UDiv:
191      return Expression::UDIV;
192    case Instruction::SDiv:
193      return Expression::SDIV;
194    case Instruction::FDiv:
195      return Expression::FDIV;
196    case Instruction::URem:
197      return Expression::UREM;
198    case Instruction::SRem:
199      return Expression::SREM;
200    case Instruction::FRem:
201      return Expression::FREM;
202    case Instruction::Shl:
203      return Expression::SHL;
204    case Instruction::LShr:
205      return Expression::LSHR;
206    case Instruction::AShr:
207      return Expression::ASHR;
208    case Instruction::And:
209      return Expression::AND;
210    case Instruction::Or:
211      return Expression::OR;
212    case Instruction::Xor:
213      return Expression::XOR;
214
215    // THIS SHOULD NEVER HAPPEN
216    default:
217      assert(0 && "Binary operator with unknown opcode?");
218      return Expression::ADD;
219  }
220}
221
222Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
223  if (C->getOpcode() == Instruction::ICmp) {
224    switch (C->getPredicate()) {
225      case ICmpInst::ICMP_EQ:
226        return Expression::ICMPEQ;
227      case ICmpInst::ICMP_NE:
228        return Expression::ICMPNE;
229      case ICmpInst::ICMP_UGT:
230        return Expression::ICMPUGT;
231      case ICmpInst::ICMP_UGE:
232        return Expression::ICMPUGE;
233      case ICmpInst::ICMP_ULT:
234        return Expression::ICMPULT;
235      case ICmpInst::ICMP_ULE:
236        return Expression::ICMPULE;
237      case ICmpInst::ICMP_SGT:
238        return Expression::ICMPSGT;
239      case ICmpInst::ICMP_SGE:
240        return Expression::ICMPSGE;
241      case ICmpInst::ICMP_SLT:
242        return Expression::ICMPSLT;
243      case ICmpInst::ICMP_SLE:
244        return Expression::ICMPSLE;
245
246      // THIS SHOULD NEVER HAPPEN
247      default:
248        assert(0 && "Comparison with unknown predicate?");
249        return Expression::ICMPEQ;
250    }
251  } else {
252    switch (C->getPredicate()) {
253      case FCmpInst::FCMP_OEQ:
254        return Expression::FCMPOEQ;
255      case FCmpInst::FCMP_OGT:
256        return Expression::FCMPOGT;
257      case FCmpInst::FCMP_OGE:
258        return Expression::FCMPOGE;
259      case FCmpInst::FCMP_OLT:
260        return Expression::FCMPOLT;
261      case FCmpInst::FCMP_OLE:
262        return Expression::FCMPOLE;
263      case FCmpInst::FCMP_ONE:
264        return Expression::FCMPONE;
265      case FCmpInst::FCMP_ORD:
266        return Expression::FCMPORD;
267      case FCmpInst::FCMP_UNO:
268        return Expression::FCMPUNO;
269      case FCmpInst::FCMP_UEQ:
270        return Expression::FCMPUEQ;
271      case FCmpInst::FCMP_UGT:
272        return Expression::FCMPUGT;
273      case FCmpInst::FCMP_UGE:
274        return Expression::FCMPUGE;
275      case FCmpInst::FCMP_ULT:
276        return Expression::FCMPULT;
277      case FCmpInst::FCMP_ULE:
278        return Expression::FCMPULE;
279      case FCmpInst::FCMP_UNE:
280        return Expression::FCMPUNE;
281
282      // THIS SHOULD NEVER HAPPEN
283      default:
284        assert(0 && "Comparison with unknown predicate?");
285        return Expression::FCMPOEQ;
286    }
287  }
288}
289
290Expression::ExpressionOpcode
291                             ValueTable::getOpcode(CastInst* C) {
292  switch(C->getOpcode()) {
293    case Instruction::Trunc:
294      return Expression::TRUNC;
295    case Instruction::ZExt:
296      return Expression::ZEXT;
297    case Instruction::SExt:
298      return Expression::SEXT;
299    case Instruction::FPToUI:
300      return Expression::FPTOUI;
301    case Instruction::FPToSI:
302      return Expression::FPTOSI;
303    case Instruction::UIToFP:
304      return Expression::UITOFP;
305    case Instruction::SIToFP:
306      return Expression::SITOFP;
307    case Instruction::FPTrunc:
308      return Expression::FPTRUNC;
309    case Instruction::FPExt:
310      return Expression::FPEXT;
311    case Instruction::PtrToInt:
312      return Expression::PTRTOINT;
313    case Instruction::IntToPtr:
314      return Expression::INTTOPTR;
315    case Instruction::BitCast:
316      return Expression::BITCAST;
317
318    // THIS SHOULD NEVER HAPPEN
319    default:
320      assert(0 && "Cast operator with unknown opcode?");
321      return Expression::BITCAST;
322  }
323}
324
325Expression ValueTable::create_expression(BinaryOperator* BO) {
326  Expression e;
327
328  e.firstVN = lookup_or_add(BO->getOperand(0));
329  e.secondVN = lookup_or_add(BO->getOperand(1));
330  e.thirdVN = 0;
331  e.type = BO->getType();
332  e.opcode = getOpcode(BO);
333
334  return e;
335}
336
337Expression ValueTable::create_expression(CmpInst* C) {
338  Expression e;
339
340  e.firstVN = lookup_or_add(C->getOperand(0));
341  e.secondVN = lookup_or_add(C->getOperand(1));
342  e.thirdVN = 0;
343  e.type = C->getType();
344  e.opcode = getOpcode(C);
345
346  return e;
347}
348
349Expression ValueTable::create_expression(CastInst* C) {
350  Expression e;
351
352  e.firstVN = lookup_or_add(C->getOperand(0));
353  e.secondVN = 0;
354  e.thirdVN = 0;
355  e.type = C->getType();
356  e.opcode = getOpcode(C);
357
358  return e;
359}
360
361Expression ValueTable::create_expression(ShuffleVectorInst* S) {
362  Expression e;
363
364  e.firstVN = lookup_or_add(S->getOperand(0));
365  e.secondVN = lookup_or_add(S->getOperand(1));
366  e.thirdVN = lookup_or_add(S->getOperand(2));
367  e.type = S->getType();
368  e.opcode = Expression::SHUFFLE;
369
370  return e;
371}
372
373Expression ValueTable::create_expression(ExtractElementInst* E) {
374  Expression e;
375
376  e.firstVN = lookup_or_add(E->getOperand(0));
377  e.secondVN = lookup_or_add(E->getOperand(1));
378  e.thirdVN = 0;
379  e.type = E->getType();
380  e.opcode = Expression::EXTRACT;
381
382  return e;
383}
384
385Expression ValueTable::create_expression(InsertElementInst* I) {
386  Expression e;
387
388  e.firstVN = lookup_or_add(I->getOperand(0));
389  e.secondVN = lookup_or_add(I->getOperand(1));
390  e.thirdVN = lookup_or_add(I->getOperand(2));
391  e.type = I->getType();
392  e.opcode = Expression::INSERT;
393
394  return e;
395}
396
397Expression ValueTable::create_expression(SelectInst* I) {
398  Expression e;
399
400  e.firstVN = lookup_or_add(I->getCondition());
401  e.secondVN = lookup_or_add(I->getTrueValue());
402  e.thirdVN = lookup_or_add(I->getFalseValue());
403  e.type = I->getType();
404  e.opcode = Expression::SELECT;
405
406  return e;
407}
408
409Expression ValueTable::create_expression(GetElementPtrInst* G) {
410  Expression e;
411
412  e.firstVN = lookup_or_add(G->getPointerOperand());
413  e.secondVN = 0;
414  e.thirdVN = 0;
415  e.type = G->getType();
416  e.opcode = Expression::GEP;
417
418  for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
419       I != E; ++I)
420    e.varargs.push_back(lookup_or_add(*I));
421
422  return e;
423}
424
425//===----------------------------------------------------------------------===//
426//                     ValueTable External Functions
427//===----------------------------------------------------------------------===//
428
429/// lookup_or_add - Returns the value number for the specified value, assigning
430/// it a new number if it did not have one before.
431uint32_t ValueTable::lookup_or_add(Value* V) {
432  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
433  if (VI != valueNumbering.end())
434    return VI->second;
435
436
437  if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
438    Expression e = create_expression(BO);
439
440    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
441    if (EI != expressionNumbering.end()) {
442      valueNumbering.insert(std::make_pair(V, EI->second));
443      return EI->second;
444    } else {
445      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
446      valueNumbering.insert(std::make_pair(V, nextValueNumber));
447
448      return nextValueNumber++;
449    }
450  } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
451    Expression e = create_expression(C);
452
453    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
454    if (EI != expressionNumbering.end()) {
455      valueNumbering.insert(std::make_pair(V, EI->second));
456      return EI->second;
457    } else {
458      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
459      valueNumbering.insert(std::make_pair(V, nextValueNumber));
460
461      return nextValueNumber++;
462    }
463  } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
464    Expression e = create_expression(U);
465
466    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
467    if (EI != expressionNumbering.end()) {
468      valueNumbering.insert(std::make_pair(V, EI->second));
469      return EI->second;
470    } else {
471      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
472      valueNumbering.insert(std::make_pair(V, nextValueNumber));
473
474      return nextValueNumber++;
475    }
476  } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
477    Expression e = create_expression(U);
478
479    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
480    if (EI != expressionNumbering.end()) {
481      valueNumbering.insert(std::make_pair(V, EI->second));
482      return EI->second;
483    } else {
484      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
485      valueNumbering.insert(std::make_pair(V, nextValueNumber));
486
487      return nextValueNumber++;
488    }
489  } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
490    Expression e = create_expression(U);
491
492    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
493    if (EI != expressionNumbering.end()) {
494      valueNumbering.insert(std::make_pair(V, EI->second));
495      return EI->second;
496    } else {
497      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
498      valueNumbering.insert(std::make_pair(V, nextValueNumber));
499
500      return nextValueNumber++;
501    }
502  } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
503    Expression e = create_expression(U);
504
505    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
506    if (EI != expressionNumbering.end()) {
507      valueNumbering.insert(std::make_pair(V, EI->second));
508      return EI->second;
509    } else {
510      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
511      valueNumbering.insert(std::make_pair(V, nextValueNumber));
512
513      return nextValueNumber++;
514    }
515  } else if (CastInst* U = dyn_cast<CastInst>(V)) {
516    Expression e = create_expression(U);
517
518    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
519    if (EI != expressionNumbering.end()) {
520      valueNumbering.insert(std::make_pair(V, EI->second));
521      return EI->second;
522    } else {
523      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
524      valueNumbering.insert(std::make_pair(V, nextValueNumber));
525
526      return nextValueNumber++;
527    }
528  } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
529    Expression e = create_expression(U);
530
531    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
532    if (EI != expressionNumbering.end()) {
533      valueNumbering.insert(std::make_pair(V, EI->second));
534      return EI->second;
535    } else {
536      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
537      valueNumbering.insert(std::make_pair(V, nextValueNumber));
538
539      return nextValueNumber++;
540    }
541  } else {
542    valueNumbering.insert(std::make_pair(V, nextValueNumber));
543    return nextValueNumber++;
544  }
545}
546
547/// lookup - Returns the value number of the specified value. Fails if
548/// the value has not yet been numbered.
549uint32_t ValueTable::lookup(Value* V) const {
550  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
551  if (VI != valueNumbering.end())
552    return VI->second;
553  else
554    assert(0 && "Value not numbered?");
555
556  return 0;
557}
558
559/// clear - Remove all entries from the ValueTable
560void ValueTable::clear() {
561  valueNumbering.clear();
562  expressionNumbering.clear();
563  nextValueNumber = 1;
564}
565
566/// erase - Remove a value from the value numbering
567void ValueTable::erase(Value* V) {
568  valueNumbering.erase(V);
569}
570
571//===----------------------------------------------------------------------===//
572//                       ValueNumberedSet Class
573//===----------------------------------------------------------------------===//
574namespace {
575class ValueNumberedSet {
576  private:
577    SmallPtrSet<Value*, 8> contents;
578    BitVector numbers;
579  public:
580    ValueNumberedSet() { numbers.resize(1); }
581    ValueNumberedSet(const ValueNumberedSet& other) {
582      numbers = other.numbers;
583      contents = other.contents;
584    }
585
586    typedef SmallPtrSet<Value*, 8>::iterator iterator;
587
588    iterator begin() { return contents.begin(); }
589    iterator end() { return contents.end(); }
590
591    bool insert(Value* v) { return contents.insert(v); }
592    void insert(iterator I, iterator E) { contents.insert(I, E); }
593    void erase(Value* v) { contents.erase(v); }
594    unsigned count(Value* v) { return contents.count(v); }
595    size_t size() { return contents.size(); }
596
597    void set(unsigned i)  {
598      if (i >= numbers.size())
599        numbers.resize(i+1);
600
601      numbers.set(i);
602    }
603
604    void operator=(const ValueNumberedSet& other) {
605      contents = other.contents;
606      numbers = other.numbers;
607    }
608
609    void reset(unsigned i)  {
610      if (i < numbers.size())
611        numbers.reset(i);
612    }
613
614    bool test(unsigned i)  {
615      if (i >= numbers.size())
616        return false;
617
618      return numbers.test(i);
619    }
620
621    void clear() {
622      contents.clear();
623      numbers.clear();
624    }
625};
626}
627
628//===----------------------------------------------------------------------===//
629//                         GVN Pass
630//===----------------------------------------------------------------------===//
631
632namespace {
633
634  class VISIBILITY_HIDDEN GVN : public FunctionPass {
635    bool runOnFunction(Function &F);
636  public:
637    static char ID; // Pass identification, replacement for typeid
638    GVN() : FunctionPass((intptr_t)&ID) { }
639
640  private:
641    ValueTable VN;
642
643    DenseMap<BasicBlock*, ValueNumberedSet> availableOut;
644
645    typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
646    PhiMapType phiMap;
647
648
649    // This transformation requires dominator postdominator info
650    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
651      AU.setPreservesCFG();
652      AU.addRequired<DominatorTree>();
653      AU.addRequired<MemoryDependenceAnalysis>();
654      AU.addPreserved<MemoryDependenceAnalysis>();
655    }
656
657    // Helper fuctions
658    // FIXME: eliminate or document these better
659    Value* find_leader(ValueNumberedSet& vals, uint32_t v) ;
660    void val_insert(ValueNumberedSet& s, Value* v);
661    bool processLoad(LoadInst* L,
662                     DenseMap<Value*, LoadInst*>& lastLoad,
663                     SmallVector<Instruction*, 4>& toErase);
664    bool processInstruction(Instruction* I,
665                            ValueNumberedSet& currAvail,
666                            DenseMap<Value*, LoadInst*>& lastSeenLoad,
667                            SmallVector<Instruction*, 4>& toErase);
668    bool processNonLocalLoad(LoadInst* L,
669                             SmallVector<Instruction*, 4>& toErase);
670    Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
671                            DenseMap<BasicBlock*, Value*> &Phis,
672                            bool top_level = false);
673    void dump(DenseMap<BasicBlock*, Value*>& d);
674    bool iterateOnFunction(Function &F);
675    Value* CollapsePhi(PHINode* p);
676    bool isSafeReplacement(PHINode* p, Instruction* inst);
677  };
678
679  char GVN::ID = 0;
680
681}
682
683// createGVNPass - The public interface to this file...
684FunctionPass *llvm::createGVNPass() { return new GVN(); }
685
686static RegisterPass<GVN> X("gvn",
687                           "Global Value Numbering");
688
689STATISTIC(NumGVNInstr, "Number of instructions deleted");
690STATISTIC(NumGVNLoad, "Number of loads deleted");
691
692/// find_leader - Given a set and a value number, return the first
693/// element of the set with that value number, or 0 if no such element
694/// is present
695Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) {
696  if (!vals.test(v))
697    return 0;
698
699  for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end();
700       I != E; ++I)
701    if (v == VN.lookup(*I))
702      return *I;
703
704  assert(0 && "No leader found, but present bit is set?");
705  return 0;
706}
707
708/// val_insert - Insert a value into a set only if there is not a value
709/// with the same value number already in the set
710void GVN::val_insert(ValueNumberedSet& s, Value* v) {
711  uint32_t num = VN.lookup(v);
712  if (!s.test(num))
713    s.insert(v);
714}
715
716void GVN::dump(DenseMap<BasicBlock*, Value*>& d) {
717  printf("{\n");
718  for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(),
719       E = d.end(); I != E; ++I) {
720    if (I->second == MemoryDependenceAnalysis::None)
721      printf("None\n");
722    else
723      I->second->dump();
724  }
725  printf("}\n");
726}
727
728Value* GVN::CollapsePhi(PHINode* p) {
729  DominatorTree &DT = getAnalysis<DominatorTree>();
730  Value* constVal = p->hasConstantValue();
731
732  if (constVal) {
733    if (Instruction* inst = dyn_cast<Instruction>(constVal)) {
734      if (DT.dominates(inst, p))
735        if (isSafeReplacement(p, inst))
736          return inst;
737    } else {
738      return constVal;
739    }
740  }
741
742  return 0;
743}
744
745bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
746  if (!isa<PHINode>(inst))
747    return true;
748
749  for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
750       UI != E; ++UI)
751    if (PHINode* use_phi = dyn_cast<PHINode>(UI))
752      if (use_phi->getParent() == inst->getParent())
753        return false;
754
755  return true;
756}
757
758/// GetValueForBlock - Get the value to use within the specified basic block.
759/// available values are in Phis.
760Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
761                               DenseMap<BasicBlock*, Value*> &Phis,
762                               bool top_level) {
763
764  // If we have already computed this value, return the previously computed val.
765  DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
766  if (V != Phis.end() && !top_level) return V->second;
767
768  BasicBlock* singlePred = BB->getSinglePredecessor();
769  if (singlePred) {
770    Value *ret = GetValueForBlock(singlePred, orig, Phis);
771    Phis[BB] = ret;
772    return ret;
773  }
774  // Otherwise, the idom is the loop, so we need to insert a PHI node.  Do so
775  // now, then get values to fill in the incoming values for the PHI.
776  PHINode *PN = new PHINode(orig->getType(), orig->getName()+".rle",
777                            BB->begin());
778  PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
779
780  if (Phis.count(BB) == 0)
781    Phis.insert(std::make_pair(BB, PN));
782
783  // Fill in the incoming values for the block.
784  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
785    Value* val = GetValueForBlock(*PI, orig, Phis);
786
787    PN->addIncoming(val, *PI);
788  }
789
790  // Attempt to collapse PHI nodes that are trivially redundant
791  Value* v = CollapsePhi(PN);
792  if (v) {
793    MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
794
795    MD.removeInstruction(PN);
796    PN->replaceAllUsesWith(v);
797
798    for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
799         E = Phis.end(); I != E; ++I)
800      if (I->second == PN)
801        I->second = v;
802
803    PN->eraseFromParent();
804
805    Phis[BB] = v;
806
807    return v;
808  }
809
810  // Cache our phi construction results
811  phiMap[orig->getPointerOperand()].insert(PN);
812  return PN;
813}
814
815/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
816/// non-local by performing PHI construction.
817bool GVN::processNonLocalLoad(LoadInst* L,
818                              SmallVector<Instruction*, 4>& toErase) {
819  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
820
821  // Find the non-local dependencies of the load
822  DenseMap<BasicBlock*, Value*> deps;
823  MD.getNonLocalDependency(L, deps);
824
825  DenseMap<BasicBlock*, Value*> repl;
826
827  // Filter out useless results (non-locals, etc)
828  for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end();
829       I != E; ++I)
830    if (I->second == MemoryDependenceAnalysis::None) {
831      return false;
832    } else if (I->second == MemoryDependenceAnalysis::NonLocal) {
833      continue;
834    }else if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
835      if (S->getPointerOperand() == L->getPointerOperand())
836        repl[I->first] = S->getOperand(0);
837      else
838        return false;
839    } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) {
840      if (LD->getPointerOperand() == L->getPointerOperand())
841        repl[I->first] = LD;
842      else
843        return false;
844    } else {
845      return false;
846    }
847
848  // Use cached PHI construction information from previous runs
849  SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
850  for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
851       I != E; ++I) {
852    if ((*I)->getParent() == L->getParent()) {
853      MD.removeInstruction(L);
854      L->replaceAllUsesWith(*I);
855      toErase.push_back(L);
856      NumGVNLoad++;
857
858      return true;
859    } else {
860      repl.insert(std::make_pair((*I)->getParent(), *I));
861    }
862  }
863
864  // Perform PHI construction
865  SmallPtrSet<BasicBlock*, 4> visited;
866  Value* v = GetValueForBlock(L->getParent(), L, repl, true);
867
868  MD.removeInstruction(L);
869  L->replaceAllUsesWith(v);
870  toErase.push_back(L);
871  NumGVNLoad++;
872
873  return true;
874}
875
876/// processLoad - Attempt to eliminate a load, first by eliminating it
877/// locally, and then attempting non-local elimination if that fails.
878bool GVN::processLoad(LoadInst* L,
879                         DenseMap<Value*, LoadInst*>& lastLoad,
880                         SmallVector<Instruction*, 4>& toErase) {
881  if (L->isVolatile()) {
882    lastLoad[L->getPointerOperand()] = L;
883    return false;
884  }
885
886  Value* pointer = L->getPointerOperand();
887  LoadInst*& last = lastLoad[pointer];
888
889  // ... to a pointer that has been loaded from before...
890  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
891  bool removedNonLocal = false;
892  Instruction* dep = MD.getDependency(L);
893  if (dep == MemoryDependenceAnalysis::NonLocal &&
894      L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
895    removedNonLocal = processNonLocalLoad(L, toErase);
896
897    if (!removedNonLocal)
898      last = L;
899
900    return removedNonLocal;
901  }
902
903
904  bool deletedLoad = false;
905
906  // Walk up the dependency chain until we either find
907  // a dependency we can use, or we can't walk any further
908  while (dep != MemoryDependenceAnalysis::None &&
909         dep != MemoryDependenceAnalysis::NonLocal &&
910         (isa<LoadInst>(dep) || isa<StoreInst>(dep))) {
911    // ... that depends on a store ...
912    if (StoreInst* S = dyn_cast<StoreInst>(dep)) {
913      if (S->getPointerOperand() == pointer) {
914        // Remove it!
915        MD.removeInstruction(L);
916
917        L->replaceAllUsesWith(S->getOperand(0));
918        toErase.push_back(L);
919        deletedLoad = true;
920        NumGVNLoad++;
921      }
922
923      // Whether we removed it or not, we can't
924      // go any further
925      break;
926    } else if (!last) {
927      // If we don't depend on a store, and we haven't
928      // been loaded before, bail.
929      break;
930    } else if (dep == last) {
931      // Remove it!
932      MD.removeInstruction(L);
933
934      L->replaceAllUsesWith(last);
935      toErase.push_back(L);
936      deletedLoad = true;
937      NumGVNLoad++;
938
939      break;
940    } else {
941      dep = MD.getDependency(L, dep);
942    }
943  }
944
945  if (!deletedLoad)
946    last = L;
947
948  return deletedLoad;
949}
950
951/// processInstruction - When calculating availability, handle an instruction
952/// by inserting it into the appropriate sets
953bool GVN::processInstruction(Instruction* I,
954                                ValueNumberedSet& currAvail,
955                                DenseMap<Value*, LoadInst*>& lastSeenLoad,
956                                SmallVector<Instruction*, 4>& toErase) {
957  if (LoadInst* L = dyn_cast<LoadInst>(I)) {
958    return processLoad(L, lastSeenLoad, toErase);
959  }
960
961  unsigned num = VN.lookup_or_add(I);
962
963  // Collapse PHI nodes
964  if (PHINode* p = dyn_cast<PHINode>(I)) {
965    Value* constVal = CollapsePhi(p);
966
967    if (constVal) {
968      for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
969           PI != PE; ++PI)
970        if (PI->second.count(p))
971          PI->second.erase(p);
972
973      p->replaceAllUsesWith(constVal);
974      toErase.push_back(p);
975    }
976  // Perform value-number based elimination
977  } else if (currAvail.test(num)) {
978    Value* repl = find_leader(currAvail, num);
979
980    VN.erase(I);
981    I->replaceAllUsesWith(repl);
982    toErase.push_back(I);
983    return true;
984  } else if (!I->isTerminator()) {
985    currAvail.set(num);
986    currAvail.insert(I);
987  }
988
989  return false;
990}
991
992// GVN::runOnFunction - This is the main transformation entry point for a
993// function.
994//
995bool GVN::runOnFunction(Function& F) {
996  bool changed = false;
997  bool shouldContinue = true;
998
999  while (shouldContinue) {
1000    shouldContinue = iterateOnFunction(F);
1001    changed |= shouldContinue;
1002  }
1003
1004  return changed;
1005}
1006
1007
1008// GVN::iterateOnFunction - Executes one iteration of GVN
1009bool GVN::iterateOnFunction(Function &F) {
1010  // Clean out global sets from any previous functions
1011  VN.clear();
1012  availableOut.clear();
1013  phiMap.clear();
1014
1015  bool changed_function = false;
1016
1017  DominatorTree &DT = getAnalysis<DominatorTree>();
1018
1019  SmallVector<Instruction*, 4> toErase;
1020
1021  // Top-down walk of the dominator tree
1022  for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
1023         E = df_end(DT.getRootNode()); DI != E; ++DI) {
1024
1025    // Get the set to update for this block
1026    ValueNumberedSet& currAvail = availableOut[DI->getBlock()];
1027    DenseMap<Value*, LoadInst*> lastSeenLoad;
1028
1029    BasicBlock* BB = DI->getBlock();
1030
1031    // A block inherits AVAIL_OUT from its dominator
1032    if (DI->getIDom() != 0)
1033      currAvail = availableOut[DI->getIDom()->getBlock()];
1034
1035    for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1036         BI != BE; ) {
1037      changed_function |= processInstruction(BI, currAvail,
1038                                             lastSeenLoad, toErase);
1039
1040      NumGVNInstr += toErase.size();
1041
1042      // Avoid iterator invalidation
1043      ++BI;
1044
1045      for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1046           E = toErase.end(); I != E; ++I)
1047        (*I)->eraseFromParent();
1048
1049      toErase.clear();
1050    }
1051  }
1052
1053  return changed_function;
1054}
1055