GVN.cpp

//===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass performs global value numbering to eliminate fully redundant
// instructions.  It also performs simple dead load elimination.
//
// Note that this pass does the value numbering itself, it does not use the
// ValueNumbering analysis passes.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "gvn"
#include "llvm/Transforms/Scalar.h"
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Value.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include <cstdio>
using namespace llvm;

STATISTIC(NumGVNInstr, "Number of instructions deleted");
STATISTIC(NumGVNLoad, "Number of loads deleted");
STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
STATISTIC(NumGVNBlocks, "Number of blocks merged");
STATISTIC(NumPRELoad, "Number of loads PRE'd");

static cl::opt<bool> EnablePRE("enable-pre",
                               cl::init(true), cl::Hidden);
cl::opt<bool> EnableLoadPRE("enable-load-pre"/*, cl::init(true)*/);

//===----------------------------------------------------------------------===//
//                         ValueTable Class
//===----------------------------------------------------------------------===//

/// This class holds the mapping between values and value numbers.  It is used
/// as an efficient mechanism to determine the expression-wise equivalence of
/// two values.
namespace {
  struct VISIBILITY_HIDDEN Expression {
    enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM, 
                            FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ, 
                            ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE, 
                            ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ, 
                            FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE, 
                            FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE, 
                            FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
                            SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
                            FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT, 
                            PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
                            EMPTY, TOMBSTONE };

    ExpressionOpcode opcode;
    const Type* type;
    uint32_t firstVN;
    uint32_t secondVN;
    uint32_t thirdVN;
    SmallVector<uint32_t, 4> varargs;
    Value* function;
  
    Expression() { }
    Expression(ExpressionOpcode o) : opcode(o) { }
  
    bool operator==(const Expression &other) const {
      if (opcode != other.opcode)
        return false;
      else if (opcode == EMPTY || opcode == TOMBSTONE)
        return true;
      else if (type != other.type)
        return false;
      else if (function != other.function)
        return false;
      else if (firstVN != other.firstVN)
        return false;
      else if (secondVN != other.secondVN)
        return false;
      else if (thirdVN != other.thirdVN)
        return false;
      else {
        if (varargs.size() != other.varargs.size())
          return false;
      
        for (size_t i = 0; i < varargs.size(); ++i)
          if (varargs[i] != other.varargs[i])
            return false;
    
        return true;
      }
    }
  
    bool operator!=(const Expression &other) const {
      if (opcode != other.opcode)
        return true;
      else if (opcode == EMPTY || opcode == TOMBSTONE)
        return false;
      else if (type != other.type)
        return true;
      else if (function != other.function)
        return true;
      else if (firstVN != other.firstVN)
        return true;
      else if (secondVN != other.secondVN)
        return true;
      else if (thirdVN != other.thirdVN)
        return true;
      else {
        if (varargs.size() != other.varargs.size())
          return true;
      
        for (size_t i = 0; i < varargs.size(); ++i)
          if (varargs[i] != other.varargs[i])
            return true;
    
          return false;
      }
    }
  };
  
  class VISIBILITY_HIDDEN ValueTable {
    private:
      DenseMap<Value*, uint32_t> valueNumbering;
      DenseMap<Expression, uint32_t> expressionNumbering;
      AliasAnalysis* AA;
      MemoryDependenceAnalysis* MD;
      DominatorTree* DT;
  
      uint32_t nextValueNumber;
    
      Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
      Expression::ExpressionOpcode getOpcode(CmpInst* C);
      Expression::ExpressionOpcode getOpcode(CastInst* C);
      Expression create_expression(BinaryOperator* BO);
      Expression create_expression(CmpInst* C);
      Expression create_expression(ShuffleVectorInst* V);
      Expression create_expression(ExtractElementInst* C);
      Expression create_expression(InsertElementInst* V);
      Expression create_expression(SelectInst* V);
      Expression create_expression(CastInst* C);
      Expression create_expression(GetElementPtrInst* G);
      Expression create_expression(CallInst* C);
      Expression create_expression(Constant* C);
    public:
      ValueTable() : nextValueNumber(1) { }
      uint32_t lookup_or_add(Value* V);
      uint32_t lookup(Value* V) const;
      void add(Value* V, uint32_t num);
      void clear();
      void erase(Value* v);
      unsigned size();
      void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
      AliasAnalysis *getAliasAnalysis() const { return AA; }
      void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
      void setDomTree(DominatorTree* D) { DT = D; }
      uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
  };
}

namespace llvm {
template <> struct DenseMapInfo<Expression> {
  static inline Expression getEmptyKey() {
    return Expression(Expression::EMPTY);
  }
  
  static inline Expression getTombstoneKey() {
    return Expression(Expression::TOMBSTONE);
  }
  
  static unsigned getHashValue(const Expression e) {
    unsigned hash = e.opcode;
    
    hash = e.firstVN + hash * 37;
    hash = e.secondVN + hash * 37;
    hash = e.thirdVN + hash * 37;
    
    hash = ((unsigned)((uintptr_t)e.type >> 4) ^
            (unsigned)((uintptr_t)e.type >> 9)) +
           hash * 37;
    
    for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
         E = e.varargs.end(); I != E; ++I)
      hash = *I + hash * 37;
    
    hash = ((unsigned)((uintptr_t)e.function >> 4) ^
            (unsigned)((uintptr_t)e.function >> 9)) +
           hash * 37;
    
    return hash;
  }
  static bool isEqual(const Expression &LHS, const Expression &RHS) {
    return LHS == RHS;
  }
  static bool isPod() { return true; }
};
}

//===----------------------------------------------------------------------===//
//                     ValueTable Internal Functions
//===----------------------------------------------------------------------===//
Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
  switch(BO->getOpcode()) {
  default: // THIS SHOULD NEVER HAPPEN
    assert(0 && "Binary operator with unknown opcode?");
  case Instruction::Add:  return Expression::ADD;
  case Instruction::Sub:  return Expression::SUB;
  case Instruction::Mul:  return Expression::MUL;
  case Instruction::UDiv: return Expression::UDIV;
  case Instruction::SDiv: return Expression::SDIV;
  case Instruction::FDiv: return Expression::FDIV;
  case Instruction::URem: return Expression::UREM;
  case Instruction::SRem: return Expression::SREM;
  case Instruction::FRem: return Expression::FREM;
  case Instruction::Shl:  return Expression::SHL;
  case Instruction::LShr: return Expression::LSHR;
  case Instruction::AShr: return Expression::ASHR;
  case Instruction::And:  return Expression::AND;
  case Instruction::Or:   return Expression::OR;
  case Instruction::Xor:  return Expression::XOR;
  }
}

Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
  if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
    switch (C->getPredicate()) {
    default:  // THIS SHOULD NEVER HAPPEN
      assert(0 && "Comparison with unknown predicate?");
    case ICmpInst::ICMP_EQ:  return Expression::ICMPEQ;
    case ICmpInst::ICMP_NE:  return Expression::ICMPNE;
    case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
    case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
    case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
    case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
    case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
    case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
    case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
    case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
    }
  }
  assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
  switch (C->getPredicate()) {
  default: // THIS SHOULD NEVER HAPPEN
    assert(0 && "Comparison with unknown predicate?");
  case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
  case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
  case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
  case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
  case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
  case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
  case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
  case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
  case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
  case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
  case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
  case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
  case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
  case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
  }
}

Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
  switch(C->getOpcode()) {
  default: // THIS SHOULD NEVER HAPPEN
    assert(0 && "Cast operator with unknown opcode?");
  case Instruction::Trunc:    return Expression::TRUNC;
  case Instruction::ZExt:     return Expression::ZEXT;
  case Instruction::SExt:     return Expression::SEXT;
  case Instruction::FPToUI:   return Expression::FPTOUI;
  case Instruction::FPToSI:   return Expression::FPTOSI;
  case Instruction::UIToFP:   return Expression::UITOFP;
  case Instruction::SIToFP:   return Expression::SITOFP;
  case Instruction::FPTrunc:  return Expression::FPTRUNC;
  case Instruction::FPExt:    return Expression::FPEXT;
  case Instruction::PtrToInt: return Expression::PTRTOINT;
  case Instruction::IntToPtr: return Expression::INTTOPTR;
  case Instruction::BitCast:  return Expression::BITCAST;
  }
}

Expression ValueTable::create_expression(CallInst* C) {
  Expression e;
  
  e.type = C->getType();
  e.firstVN = 0;
  e.secondVN = 0;
  e.thirdVN = 0;
  e.function = C->getCalledFunction();
  e.opcode = Expression::CALL;
  
  for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
       I != E; ++I)
    e.varargs.push_back(lookup_or_add(*I));
  
  return e;
}

Expression ValueTable::create_expression(BinaryOperator* BO) {
  Expression e;
    
  e.firstVN = lookup_or_add(BO->getOperand(0));
  e.secondVN = lookup_or_add(BO->getOperand(1));
  e.thirdVN = 0;
  e.function = 0;
  e.type = BO->getType();
  e.opcode = getOpcode(BO);
  
  return e;
}

Expression ValueTable::create_expression(CmpInst* C) {
  Expression e;
    
  e.firstVN = lookup_or_add(C->getOperand(0));
  e.secondVN = lookup_or_add(C->getOperand(1));
  e.thirdVN = 0;
  e.function = 0;
  e.type = C->getType();
  e.opcode = getOpcode(C);
  
  return e;
}

Expression ValueTable::create_expression(CastInst* C) {
  Expression e;
    
  e.firstVN = lookup_or_add(C->getOperand(0));
  e.secondVN = 0;
  e.thirdVN = 0;
  e.function = 0;
  e.type = C->getType();
  e.opcode = getOpcode(C);
  
  return e;
}

Expression ValueTable::create_expression(ShuffleVectorInst* S) {
  Expression e;
    
  e.firstVN = lookup_or_add(S->getOperand(0));
  e.secondVN = lookup_or_add(S->getOperand(1));
  e.thirdVN = lookup_or_add(S->getOperand(2));
  e.function = 0;
  e.type = S->getType();
  e.opcode = Expression::SHUFFLE;
  
  return e;
}

Expression ValueTable::create_expression(ExtractElementInst* E) {
  Expression e;
    
  e.firstVN = lookup_or_add(E->getOperand(0));
  e.secondVN = lookup_or_add(E->getOperand(1));
  e.thirdVN = 0;
  e.function = 0;
  e.type = E->getType();
  e.opcode = Expression::EXTRACT;
  
  return e;
}

Expression ValueTable::create_expression(InsertElementInst* I) {
  Expression e;
    
  e.firstVN = lookup_or_add(I->getOperand(0));
  e.secondVN = lookup_or_add(I->getOperand(1));
  e.thirdVN = lookup_or_add(I->getOperand(2));
  e.function = 0;
  e.type = I->getType();
  e.opcode = Expression::INSERT;
  
  return e;
}

Expression ValueTable::create_expression(SelectInst* I) {
  Expression e;
    
  e.firstVN = lookup_or_add(I->getCondition());
  e.secondVN = lookup_or_add(I->getTrueValue());
  e.thirdVN = lookup_or_add(I->getFalseValue());
  e.function = 0;
  e.type = I->getType();
  e.opcode = Expression::SELECT;
  
  return e;
}

Expression ValueTable::create_expression(GetElementPtrInst* G) {
  Expression e;
  
  e.firstVN = lookup_or_add(G->getPointerOperand());
  e.secondVN = 0;
  e.thirdVN = 0;
  e.function = 0;
  e.type = G->getType();
  e.opcode = Expression::GEP;
  
  for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
       I != E; ++I)
    e.varargs.push_back(lookup_or_add(*I));
  
  return e;
}

//===----------------------------------------------------------------------===//
//                     ValueTable External Functions
//===----------------------------------------------------------------------===//

/// add - Insert a value into the table with a specified value number.
void ValueTable::add(Value* V, uint32_t num) {
  valueNumbering.insert(std::make_pair(V, num));
}

/// lookup_or_add - Returns the value number for the specified value, assigning
/// it a new number if it did not have one before.
uint32_t ValueTable::lookup_or_add(Value* V) {
  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
  if (VI != valueNumbering.end())
    return VI->second;
  
  if (CallInst* C = dyn_cast<CallInst>(V)) {
    if (AA->doesNotAccessMemory(C)) {
      Expression e = create_expression(C);
    
      DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
      if (EI != expressionNumbering.end()) {
        valueNumbering.insert(std::make_pair(V, EI->second));
        return EI->second;
      } else {
        expressionNumbering.insert(std::make_pair(e, nextValueNumber));
        valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
        return nextValueNumber++;
      }
    } else if (AA->onlyReadsMemory(C)) {
      Expression e = create_expression(C);
      
      if (expressionNumbering.find(e) == expressionNumbering.end()) {
        expressionNumbering.insert(std::make_pair(e, nextValueNumber));
        valueNumbering.insert(std::make_pair(V, nextValueNumber));
        return nextValueNumber++;
      }
      
      MemDepResult local_dep = MD->getDependency(C);
      
      if (local_dep.isNone()) {
        valueNumbering.insert(std::make_pair(V, nextValueNumber));
        return nextValueNumber++;
      }
      
      if (Instruction *LocalDepInst = local_dep.getInst()) {
        if (!isa<CallInst>(LocalDepInst)) {
          valueNumbering.insert(std::make_pair(V, nextValueNumber));
          return nextValueNumber++;
        }
        
        CallInst* local_cdep = cast<CallInst>(LocalDepInst);
        
        if (local_cdep->getCalledFunction() != C->getCalledFunction() ||
            local_cdep->getNumOperands() != C->getNumOperands()) {
          valueNumbering.insert(std::make_pair(V, nextValueNumber));
          return nextValueNumber++;
        }
        
        if (!C->getCalledFunction()) { 
          valueNumbering.insert(std::make_pair(V, nextValueNumber));
          return nextValueNumber++;
        }
        
        for (unsigned i = 1; i < C->getNumOperands(); ++i) {
          uint32_t c_vn = lookup_or_add(C->getOperand(i));
          uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
          if (c_vn != cd_vn) {
            valueNumbering.insert(std::make_pair(V, nextValueNumber));
            return nextValueNumber++;
          }
        }
      
        uint32_t v = lookup_or_add(local_cdep);
        valueNumbering.insert(std::make_pair(V, v));
        return v;
      }
      

      const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 
        MD->getNonLocalDependency(C);
      CallInst* cdep = 0;
      
      // Check to see if we have a single dominating call instruction that is
      // identical to C.
      for (unsigned i = 0, e = deps.size(); i != e; ++i) {
        const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
        // Ignore non-local dependencies.
        if (I->second.isNonLocal())
          continue;

        // We don't handle non-depedencies.  If we already have a call, reject
        // instruction dependencies.
        if (I->second.isNone() || cdep != 0) {
          cdep = 0;
          break;
        }
        
        CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
        // FIXME: All duplicated with non-local case.
        if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
          cdep = NonLocalDepCall;
          continue;
        }
        
        cdep = 0;
        break;
      }
      
      if (!cdep) {
        valueNumbering.insert(std::make_pair(V, nextValueNumber));
        return nextValueNumber++;
      }
      
      if (cdep->getCalledFunction() != C->getCalledFunction() ||
          cdep->getNumOperands() != C->getNumOperands()) {
        valueNumbering.insert(std::make_pair(V, nextValueNumber));
        return nextValueNumber++;
      }
      if (!C->getCalledFunction()) { 
        valueNumbering.insert(std::make_pair(V, nextValueNumber));
        return nextValueNumber++;
      }
      for (unsigned i = 1; i < C->getNumOperands(); ++i) {
        uint32_t c_vn = lookup_or_add(C->getOperand(i));
        uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
        if (c_vn != cd_vn) {
          valueNumbering.insert(std::make_pair(V, nextValueNumber));
          return nextValueNumber++;
        }
      }
      
      uint32_t v = lookup_or_add(cdep);
      valueNumbering.insert(std::make_pair(V, v));
      return v;
      
    } else {
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      return nextValueNumber++;
    }
  } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
    Expression e = create_expression(BO);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
    Expression e = create_expression(C);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
    Expression e = create_expression(U);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
    Expression e = create_expression(U);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
    Expression e = create_expression(U);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
    Expression e = create_expression(U);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (CastInst* U = dyn_cast<CastInst>(V)) {
    Expression e = create_expression(U);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
    Expression e = create_expression(U);
    
    DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
    if (EI != expressionNumbering.end()) {
      valueNumbering.insert(std::make_pair(V, EI->second));
      return EI->second;
    } else {
      expressionNumbering.insert(std::make_pair(e, nextValueNumber));
      valueNumbering.insert(std::make_pair(V, nextValueNumber));
      
      return nextValueNumber++;
    }
  } else {
    valueNumbering.insert(std::make_pair(V, nextValueNumber));
    return nextValueNumber++;
  }
}

/// lookup - Returns the value number of the specified value. Fails if
/// the value has not yet been numbered.
uint32_t ValueTable::lookup(Value* V) const {
  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
  assert(VI != valueNumbering.end() && "Value not numbered?");
  return VI->second;
}

/// clear - Remove all entries from the ValueTable
void ValueTable::clear() {
  valueNumbering.clear();
  expressionNumbering.clear();
  nextValueNumber = 1;
}

/// erase - Remove a value from the value numbering
void ValueTable::erase(Value* V) {
  valueNumbering.erase(V);
}

//===----------------------------------------------------------------------===//
//                         GVN Pass
//===----------------------------------------------------------------------===//

namespace {
  struct VISIBILITY_HIDDEN ValueNumberScope {
    ValueNumberScope* parent;
    DenseMap<uint32_t, Value*> table;
    
    ValueNumberScope(ValueNumberScope* p) : parent(p) { }
  };
}

namespace {

  class VISIBILITY_HIDDEN GVN : public FunctionPass {
    bool runOnFunction(Function &F);
  public:
    static char ID; // Pass identification, replacement for typeid
    GVN() : FunctionPass(&ID) { }

  private:
    MemoryDependenceAnalysis *MD;
    DominatorTree *DT;

    ValueTable VN;
    DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
    
    typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
    PhiMapType phiMap;
    
    
    // This transformation requires dominator postdominator info
    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
      AU.addRequired<DominatorTree>();
      AU.addRequired<MemoryDependenceAnalysis>();
      AU.addRequired<AliasAnalysis>();
      
      AU.addPreserved<DominatorTree>();
      AU.addPreserved<AliasAnalysis>();
    }
  
    // Helper fuctions
    // FIXME: eliminate or document these better
    bool processLoad(LoadInst* L,
                     DenseMap<Value*, LoadInst*> &lastLoad,
                     SmallVectorImpl<Instruction*> &toErase);
    bool processInstruction(Instruction* I,
                            DenseMap<Value*, LoadInst*>& lastSeenLoad,
                            SmallVectorImpl<Instruction*> &toErase);
    bool processNonLocalLoad(LoadInst* L,
                             SmallVectorImpl<Instruction*> &toErase);
    bool processBlock(DomTreeNode* DTN);
    Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
                            DenseMap<BasicBlock*, Value*> &Phis,
                            bool top_level = false);
    void dump(DenseMap<uint32_t, Value*>& d);
    bool iterateOnFunction(Function &F);
    Value* CollapsePhi(PHINode* p);
    bool isSafeReplacement(PHINode* p, Instruction* inst);
    bool performPRE(Function& F);
    Value* lookupNumber(BasicBlock* BB, uint32_t num);
    bool mergeBlockIntoPredecessor(BasicBlock* BB);
    void cleanupGlobalSets();
  };
  
  char GVN::ID = 0;
}

// createGVNPass - The public interface to this file...
FunctionPass *llvm::createGVNPass() { return new GVN(); }

static RegisterPass<GVN> X("gvn",
                           "Global Value Numbering");

void GVN::dump(DenseMap<uint32_t, Value*>& d) {
  printf("{\n");
  for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
       E = d.end(); I != E; ++I) {
      printf("%d\n", I->first);
      I->second->dump();
  }
  printf("}\n");
}

Value* GVN::CollapsePhi(PHINode* p) {
  Value* constVal = p->hasConstantValue();
  if (!constVal) return 0;
  
  Instruction* inst = dyn_cast<Instruction>(constVal);
  if (!inst)
    return constVal;
    
  if (DT->dominates(inst, p))
    if (isSafeReplacement(p, inst))
      return inst;
  return 0;
}

bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
  if (!isa<PHINode>(inst))
    return true;
  
  for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
       UI != E; ++UI)
    if (PHINode* use_phi = dyn_cast<PHINode>(UI))
      if (use_phi->getParent() == inst->getParent())
        return false;
  
  return true;
}

/// GetValueForBlock - Get the value to use within the specified basic block.
/// available values are in Phis.
Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
                             DenseMap<BasicBlock*, Value*> &Phis,
                             bool top_level) { 
                                 
  // If we have already computed this value, return the previously computed val.
  DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
  if (V != Phis.end() && !top_level) return V->second;
  
  // If the block is unreachable, just return undef, since this path
  // can't actually occur at runtime.
  if (!DT->isReachableFromEntry(BB))
    return Phis[BB] = UndefValue::get(orig->getType());
  
  BasicBlock* singlePred = BB->getSinglePredecessor();
  if (singlePred) {
    Value *ret = GetValueForBlock(singlePred, orig, Phis);
    Phis[BB] = ret;
    return ret;
  }
  
  // Otherwise, the idom is the loop, so we need to insert a PHI node.  Do so
  // now, then get values to fill in the incoming values for the PHI.
  PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
                                BB->begin());
  PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
  
  if (Phis.count(BB) == 0)
    Phis.insert(std::make_pair(BB, PN));
  
  // Fill in the incoming values for the block.
  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
    Value* val = GetValueForBlock(*PI, orig, Phis);
    PN->addIncoming(val, *PI);
  }
  
  VN.getAliasAnalysis()->copyValue(orig, PN);
  
  // Attempt to collapse PHI nodes that are trivially redundant
  Value* v = CollapsePhi(PN);
  if (!v) {
    // Cache our phi construction results
    phiMap[orig->getPointerOperand()].insert(PN);
    return PN;
  }
    
  PN->replaceAllUsesWith(v);

  for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
       E = Phis.end(); I != E; ++I)
    if (I->second == PN)
      I->second = v;

  DEBUG(cerr << "GVN removed: " << *PN);
  MD->removeInstruction(PN);
  PN->eraseFromParent();

  Phis[BB] = v;
  return v;
}

/// IsValueFullyAvailableInBlock - Return true if we can prove that the value
/// we're analyzing is fully available in the specified block.  As we go, keep
/// track of which blocks we know are fully alive in FullyAvailableBlocks.  This
/// map is actually a tri-state map with the following values:
///   0) we know the block *is not* fully available.
///   1) we know the block *is* fully available.
///   2) we do not know whether the block is fully available or not, but we are
///      currently speculating that it will be.
///   3) we are speculating for this block and have used that to speculate for
///      other blocks.
static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 
                            DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
  // Optimistically assume that the block is fully available and check to see
  // if we already know about this block in one lookup.
  std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 
    FullyAvailableBlocks.insert(std::make_pair(BB, 2));

  // If the entry already existed for this block, return the precomputed value.
  if (!IV.second) {
    // If this is a speculative "available" value, mark it as being used for
    // speculation of other blocks.
    if (IV.first->second == 2)
      IV.first->second = 3;
    return IV.first->second != 0;
  }
  
  // Otherwise, see if it is fully available in all predecessors.
  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
  
  // If this block has no predecessors, it isn't live-in here.
  if (PI == PE)
    goto SpeculationFailure;
  
  for (; PI != PE; ++PI)
    // If the value isn't fully available in one of our predecessors, then it
    // isn't fully available in this block either.  Undo our previous
    // optimistic assumption and bail out.
    if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
      goto SpeculationFailure;
  
  return true;
  
// SpeculationFailure - If we get here, we found out that this is not, after
// all, a fully-available block.  We have a problem if we speculated on this and
// used the speculation to mark other blocks as available.
SpeculationFailure:
  char &BBVal = FullyAvailableBlocks[BB];
  
  // If we didn't speculate on this, just return with it set to false.
  if (BBVal == 2) {
    BBVal = 0;
    return false;
  }

  // If we did speculate on this value, we could have blocks set to 1 that are
  // incorrect.  Walk the (transitive) successors of this block and mark them as
  // 0 if set to one.
  SmallVector<BasicBlock*, 32> BBWorklist;
  BBWorklist.push_back(BB);
  
  while (!BBWorklist.empty()) {
    BasicBlock *Entry = BBWorklist.pop_back_val();
    // Note that this sets blocks to 0 (unavailable) if they happen to not
    // already be in FullyAvailableBlocks.  This is safe.
    char &EntryVal = FullyAvailableBlocks[Entry];
    if (EntryVal == 0) continue;  // Already unavailable.

    // Mark as unavailable.
    EntryVal = 0;
    
    for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
      BBWorklist.push_back(*I);
  }
  
  return false;
}

/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
/// non-local by performing PHI construction.
bool GVN::processNonLocalLoad(LoadInst *LI,
                              SmallVectorImpl<Instruction*> &toErase) {
  // Find the non-local dependencies of the load.
  const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 
    MD->getNonLocalDependency(LI);
  //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << deps.size() << *LI);
  
  // If we had to process more than one hundred blocks to find the
  // dependencies, this load isn't worth worrying about.  Optimizing
  // it will be too expensive.
  if (deps.size() > 100)
    return false;
  
  BasicBlock *EntryBlock = &LI->getParent()->getParent()->getEntryBlock();
  
  // Filter out useless results (non-locals, etc).  Keep track of the blocks
  // where we have a value available in repl, also keep track of whether we see
  // dependencies that produce an unknown value for the load (such as a call
  // that could potentially clobber the load).
  SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
  SmallVector<BasicBlock*, 16> UnavailableBlocks;
  
  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
    BasicBlock *DepBB = deps[i].first;
    MemDepResult DepInfo = deps[i].second;
    
    if (DepInfo.isNonLocal()) {
      // If this is a non-local dependency in the entry block, then we depend on
      // the value live-in at the start of the function.  We could insert a load
      // in the entry block to get this, but for now we'll just bail out.
      if (DepBB == EntryBlock)
        UnavailableBlocks.push_back(DepBB);
      continue;
    }
    
    if (DepInfo.isNone()) {
      ValuesPerBlock.push_back(std::make_pair(DepBB, 
                                              UndefValue::get(LI->getType())));
      continue;
    }
  
    if (StoreInst* S = dyn_cast<StoreInst>(DepInfo.getInst())) {
      // Reject loads and stores that are to the same address but are of 
      // different types.
      // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because
      // of bitfield access, it would be interesting to optimize for it at some
      // point.
      if (S->getOperand(0)->getType() != LI->getType()) {
        UnavailableBlocks.push_back(DepBB);
        continue;
      }
      
      if (S->getPointerOperand() != LI->getPointerOperand() &&
          VN.getAliasAnalysis()->alias(S->getPointerOperand(), 1,