LoopStrengthReduce.cpp

    // Split the expression into subexprs.
    SeparateSubExprs(SubExprs, Uses[i].Base, SE);

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

/// ValidStride - Check whether the given Scale is valid for all loads and 
/// stores in UsersToProcess.
///
bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
                               int64_t Scale, 
                               const std::vector<BasedUser>& UsersToProcess) {
  if (!TLI)
    return true;

  for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
    // If this is a load or other access, pass the type of the access in.
    const Type *AccessTy = Type::VoidTy;
    if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
      AccessTy = SI->getOperand(0)->getType();
    else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
      AccessTy = LI->getType();
    else if (isa<PHINode>(UsersToProcess[i].Inst))
      continue;
    
    TargetLowering::AddrMode AM;
    if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
      AM.BaseOffs = SC->getValue()->getSExtValue();
    AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
    AM.Scale = Scale;

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

/// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
/// a nop.
bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
                                                const Type *Ty2) {
  if (Ty1 == Ty2)
    return false;
  if (TLI && TLI->isTruncateFree(Ty1, Ty2))
    return false;
  return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
          !(isa<PointerType>(Ty2) &&
            Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
          !(isa<PointerType>(Ty1) &&
            Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
}

/// CheckForIVReuse - Returns the multiple if the stride is the multiple
/// of a previous stride and it is a legal value for the target addressing
/// mode scale component and optional base reg. This allows the users of
/// this stride to be rewritten as prev iv * factor. It returns 0 if no
/// reuse is possible.
unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
                                bool AllUsesAreAddresses,
                                const SCEVHandle &Stride, 
                                IVExpr &IV, const Type *Ty,
                                const std::vector<BasedUser>& UsersToProcess) {
  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
    int64_t SInt = SC->getValue()->getSExtValue();
    for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
         ++NewStride) {
      std::map<SCEVHandle, IVsOfOneStride>::iterator SI = 
                IVsByStride.find(StrideOrder[NewStride]);
      if (SI == IVsByStride.end()) 
        continue;
      int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
      if (SI->first != Stride &&
          (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
        continue;
      int64_t Scale = SInt / SSInt;
      // Check that this stride is valid for all the types used for loads and
      // stores; if it can be used for some and not others, we might as well use
      // the original stride everywhere, since we have to create the IV for it
      // anyway. If the scale is 1, then we don't need to worry about folding
      // multiplications.
      if (Scale == 1 ||
          (AllUsesAreAddresses &&
           ValidStride(HasBaseReg, Scale, UsersToProcess)))
        for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
               IE = SI->second.IVs.end(); II != IE; ++II)
          // FIXME: Only handle base == 0 for now.
          // Only reuse previous IV if it would not require a type conversion.
          if (II->Base->isZero() &&
              !RequiresTypeConversion(II->Base->getType(), Ty)) {
            IV = *II;
            return Scale;
          }
    }
  }
  return 0;
}

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

/// isNonConstantNegative - Return true if the specified scev is negated, but
/// not a constant.
static bool isNonConstantNegative(const SCEVHandle &Expr) {
  SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
  if (!Mul) return false;
  
  // If there is a constant factor, it will be first.
  SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
  if (!SC) return false;
  
  // Return true if the value is negative, this matches things like (-42 * V).
  return SC->getValue()->getValue().isNegative();
}

/// isAddress - Returns true if the specified instruction is using the
/// specified value as an address.
static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
  bool isAddress = isa<LoadInst>(Inst);
  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
    if (SI->getOperand(1) == OperandVal)
      isAddress = true;
  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
    // Addressing modes can also be folded into prefetches and a variety
    // of intrinsics.
    switch (II->getIntrinsicID()) {
      default: break;
      case Intrinsic::prefetch:
      case Intrinsic::x86_sse2_loadu_dq:
      case Intrinsic::x86_sse2_loadu_pd:
      case Intrinsic::x86_sse_loadu_ps:
      case Intrinsic::x86_sse_storeu_ps:
      case Intrinsic::x86_sse2_storeu_pd:
      case Intrinsic::x86_sse2_storeu_dq:
      case Intrinsic::x86_sse2_storel_dq:
        if (II->getOperand(1) == OperandVal)
          isAddress = true;
        break;
    }
  }
  return isAddress;
}

// CollectIVUsers - Transform our list of users and offsets to a bit more
// complex table. In this new vector, each 'BasedUser' contains 'Base', the base
// of the strided accesses, as well as the old information from Uses. We
// progressively move information from the Base field to the Imm field, until
// we eventually have the full access expression to rewrite the use.
SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
                                              IVUsersOfOneStride &Uses,
                                              Loop *L,
                                              bool &AllUsesAreAddresses,
                                       std::vector<BasedUser> &UsersToProcess) {
  UsersToProcess.reserve(Uses.Users.size());
  for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
    UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
    
    // Move any loop variant operands from the offset field to the immediate
    // field of the use, so that we don't try to use something before it is
    // computed.
    MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
                                    UsersToProcess.back().Imm, L, SE);
    assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
           "Base value is not loop invariant!");
  }

  // We now have a whole bunch of uses of like-strided induction variables, but
  // they might all have different bases.  We want to emit one PHI node for this
  // stride which we fold as many common expressions (between the IVs) into as
  // possible.  Start by identifying the common expressions in the base values 
  // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
  // "A+B"), emit it to the preheader, then remove the expression from the
  // UsersToProcess base values.
  SCEVHandle CommonExprs =
    RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L);

  // Next, figure out what we can represent in the immediate fields of
  // instructions.  If we can represent anything there, move it to the imm
  // fields of the BasedUsers.  We do this so that it increases the commonality
  // of the remaining uses.
  unsigned NumPHI = 0;
  for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
    // If the user is not in the current loop, this means it is using the exit
    // value of the IV.  Do not put anything in the base, make sure it's all in
    // the immediate field to allow as much factoring as possible.
    if (!L->contains(UsersToProcess[i].Inst->getParent())) {
      UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
                                             UsersToProcess[i].Base);
      UsersToProcess[i].Base = 
        SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
    } else {
      
      // Addressing modes can be folded into loads and stores.  Be careful that
      // the store is through the expression, not of the expression though.
      bool isPHI = false;
      bool isAddress = isAddressUse(UsersToProcess[i].Inst,
                                    UsersToProcess[i].OperandValToReplace);
      if (isa<PHINode>(UsersToProcess[i].Inst)) {
        isPHI = true;
        ++NumPHI;
      }

      // If this use isn't an address, then not all uses are addresses.
      if (!isAddress && !isPHI)
        AllUsesAreAddresses = false;
      
      MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
                          UsersToProcess[i].Imm, isAddress, L, SE);
    }
  }

  // If one of the use if a PHI node and all other uses are addresses, still
  // allow iv reuse. Essentially we are trading one constant multiplication
  // for one fewer iv.
  if (NumPHI > 1)
    AllUsesAreAddresses = false;

  return CommonExprs;
}

/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
/// stride of IV.  All of the users may have different starting values, and this
/// may not be the only stride (we know it is if isOnlyStride is true).
void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
                                                      IVUsersOfOneStride &Uses,
                                                      Loop *L,
                                                      bool isOnlyStride) {
  // If all the users are moved to another stride, then there is nothing to do.
  if (Uses.Users.empty())
    return;

  // Keep track if every use in UsersToProcess is an address. If they all are,
  // we may be able to rewrite the entire collection of them in terms of a
  // smaller-stride IV.
  bool AllUsesAreAddresses = true;

  // Transform our list of users and offsets to a bit more complex table.  In
  // this new vector, each 'BasedUser' contains 'Base' the base of the
  // strided accessas well as the old information from Uses.  We progressively
  // move information from the Base field to the Imm field, until we eventually
  // have the full access expression to rewrite the use.
  std::vector<BasedUser> UsersToProcess;
  SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
                                          UsersToProcess);

  // If we managed to find some expressions in common, we'll need to carry
  // their value in a register and add it in for each use. This will take up
  // a register operand, which potentially restricts what stride values are
  // valid.
  bool HaveCommonExprs = !CommonExprs->isZero();
  
  // If all uses are addresses, check if it is possible to reuse an IV with a
  // stride that is a factor of this stride. And that the multiple is a number
  // that can be encoded in the scale field of the target addressing mode. And
  // that we will have a valid instruction after this substition, including the
  // immediate field, if any.
  PHINode *NewPHI = NULL;
  Value   *IncV   = NULL;
  IVExpr   ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
                   SE->getIntegerSCEV(0, Type::Int32Ty),
                   0, 0);
  unsigned RewriteFactor = 0;
  RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
                                  Stride, ReuseIV, CommonExprs->getType(),
                                  UsersToProcess);
  if (RewriteFactor != 0) {
    DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
         << " and BASE " << *ReuseIV.Base << " :\n";
    NewPHI = ReuseIV.PHI;
    IncV   = ReuseIV.IncV;
  }

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

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


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

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

    // If the stride is negative, insert a sub instead of an add for the
    // increment.
    bool isNegative = isNonConstantNegative(Stride);
    SCEVHandle IncAmount = Stride;
    if (isNegative)
      IncAmount = SE->getNegativeSCEV(Stride);
    
    // Insert the stride into the preheader.
    Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
    if (!isa<ConstantInt>(StrideV)) ++NumVariable;

    // Emit the increment of the base value before the terminator of the loop
    // latch block, and add it to the Phi node.
    SCEVHandle IncExp = SE->getUnknown(StrideV);
    if (isNegative)
      IncExp = SE->getNegativeSCEV(IncExp);
    IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
  
    IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
    IncV->setName(NewPHI->getName()+".inc");
    NewPHI->addIncoming(IncV, LatchBlock);

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

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

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

    // Emit the code for Base into the preheader.
    Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);

    DOUT << "  INSERTING code for BASE = " << *Base << ":";
    if (BaseV->hasName())
      DOUT << " Result value name = %" << BaseV->getNameStr();
    DOUT << "\n";

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

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

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

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

      SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);

      // If we had to insert new instrutions for RewriteOp, we have to
      // consider that they may not have been able to end up immediately
      // next to RewriteOp, because non-PHI instructions may never precede
      // PHI instructions in a block. In this case, remember where the last
      // instruction was inserted so that if we're replacing a different
      // PHI node, we can use the later point to expand the final
      // RewriteExpr.
      Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
      if (RewriteOp == NewPHI) NewBasePt = 0;

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

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

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

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

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

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

      UsersToProcess.pop_back();
      ++NumReduced;

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

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

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

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

/// ChangeCompareStride - If a loop termination compare instruction is the
/// only use of its stride, and the compaison is against a constant value,
/// try eliminate the stride by moving the compare instruction to another
/// stride and change its constant operand accordingly. e.g.
///
/// loop:
/// ...
/// v1 = v1 + 3
/// v2 = v2 + 1
/// if (v2 < 10) goto loop
/// =>
/// loop:
/// ...
/// v1 = v1 + 3
/// if (v1 < 30) goto loop
ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
                                                IVStrideUse* &CondUse,
                                                const SCEVHandle* &CondStride) {
  if (StrideOrder.size() < 2 ||
      IVUsesByStride[*CondStride].Users.size() != 1)
    return Cond;
  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
  if (!SC) return Cond;
  ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
  if (!C) return Cond;

  ICmpInst::Predicate Predicate = Cond->getPredicate();
  int64_t CmpSSInt = SC->getValue()->getSExtValue();
  int64_t CmpVal = C->getValue().getSExtValue();
  unsigned BitWidth = C->getValue().getBitWidth();
  uint64_t SignBit = 1ULL << (BitWidth-1);
  const Type *CmpTy = C->getType();
  const Type *NewCmpTy = NULL;
  unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
  unsigned NewTyBits = 0;
  int64_t NewCmpVal = CmpVal;
  SCEVHandle *NewStride = NULL;
  Value *NewIncV = NULL;
  int64_t Scale = 1;

  // Check stride constant and the comparision constant signs to detect
  // overflow.
  if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
    return Cond;

  // Look for a suitable stride / iv as replacement.
  std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
  for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
    std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI = 
      IVUsesByStride.find(StrideOrder[i]);
    if (!isa<SCEVConstant>(SI->first))
      continue;
    int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
    if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
      continue;

    Scale = SSInt / CmpSSInt;
    NewCmpVal = CmpVal * Scale;
    APInt Mul = APInt(BitWidth, NewCmpVal);
    // Check for overflow.
    if (Mul.getSExtValue() != NewCmpVal) {
      NewCmpVal = CmpVal;
      continue;
    }

    // Watch out for overflow.
    if (ICmpInst::isSignedPredicate(Predicate) &&
        (CmpVal & SignBit) != (NewCmpVal & SignBit))
      NewCmpVal = CmpVal;

    if (NewCmpVal != CmpVal) {
      // Pick the best iv to use trying to avoid a cast.
      NewIncV = NULL;
      for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
             E = SI->second.Users.end(); UI != E; ++UI) {
        NewIncV = UI->OperandValToReplace;
        if (NewIncV->getType() == CmpTy)
          break;
      }
      if (!NewIncV) {
        NewCmpVal = CmpVal;
        continue;
      }

      NewCmpTy = NewIncV->getType();
      NewTyBits = isa<PointerType>(NewCmpTy)
        ? UIntPtrTy->getPrimitiveSizeInBits()
        : NewCmpTy->getPrimitiveSizeInBits();
      if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
        // Check if it is possible to rewrite it using
        // an iv / stride of a smaller integer type.
        bool TruncOk = false;
        if (NewCmpTy->isInteger()) {
          unsigned Bits = NewTyBits;
          if (ICmpInst::isSignedPredicate(Predicate))
            --Bits;
          uint64_t Mask = (1ULL << Bits) - 1;
          if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
            TruncOk = true;
        }
        if (!TruncOk) {
          NewCmpVal = CmpVal;
          continue;
        }
      }

      // Don't rewrite if use offset is non-constant and the new type is
      // of a different type.
      // FIXME: too conservative?
      if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
        NewCmpVal = CmpVal;
        continue;
      }

      bool AllUsesAreAddresses = true;
      std::vector<BasedUser> UsersToProcess;
      SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
                                              AllUsesAreAddresses,
                                              UsersToProcess);
      // Avoid rewriting the compare instruction with an iv of new stride
      // if it's likely the new stride uses will be rewritten using the
      if (AllUsesAreAddresses &&
          ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
        NewCmpVal = CmpVal;
        continue;
      }

      // If scale is negative, use swapped predicate unless it's testing
      // for equality.
      if (Scale < 0 && !Cond->isEquality())
        Predicate = ICmpInst::getSwappedPredicate(Predicate);

      NewStride = &StrideOrder[i];
      break;
    }
  }

  // Forgo this transformation if it the increment happens to be
  // unfortunately positioned after the condition, and the condition
  // has multiple uses which prevent it from being moved immediately
  // before the branch. See
  // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
  // for an example of this situation.
  if (!Cond->hasOneUse()) {
    for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
         I != E; ++I)
      if (I == NewIncV)
        return Cond;
  }

  if (NewCmpVal != CmpVal) {
    // Create a new compare instruction using new stride / iv.
    ICmpInst *OldCond = Cond;
    Value *RHS;
    if (!isa<PointerType>(NewCmpTy))
      RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
    else {
      RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
      RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
    }
    // Insert new compare instruction.
    Cond = new ICmpInst(Predicate, NewIncV, RHS,
                        L->getHeader()->getName() + ".termcond",
                        OldCond);

    // Remove the old compare instruction. The old indvar is probably dead too.
    DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
    SE->deleteValueFromRecords(OldCond);
    OldCond->replaceAllUsesWith(Cond);
    OldCond->eraseFromParent();

    IVUsesByStride[*CondStride].Users.pop_back();
    SCEVHandle NewOffset = TyBits == NewTyBits
      ? SE->getMulExpr(CondUse->Offset,
                       SE->getConstant(ConstantInt::get(CmpTy, Scale)))
      : SE->getConstant(ConstantInt::get(NewCmpTy,
        cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
    IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
    CondUse = &IVUsesByStride[*NewStride].Users.back();
    CondStride = NewStride;
    ++NumEliminated;
  }

  return Cond;
}

/// OptimizeSMax - Rewrite the loop's terminating condition if it uses
/// an smax computation.
///
/// This is a narrow solution to a specific, but acute, problem. For loops
/// like this:
///
///   i = 0;
///   do {
///     p[i] = 0.0;
///   } while (++i < n);
///
/// where the comparison is signed, the trip count isn't just 'n', because
/// 'n' could be negative. And unfortunately this can come up even for loops
/// where the user didn't use a C do-while loop. For example, seemingly
/// well-behaved top-test loops will commonly be lowered like this:
//
///   if (n > 0) {
///     i = 0;
///     do {
///       p[i] = 0.0;
///     } while (++i < n);
///   }
///
/// and then it's possible for subsequent optimization to obscure the if
/// test in such a way that indvars can't find it.
///
/// When indvars can't find the if test in loops like this, it creates a
/// signed-max expression, which allows it to give the loop a canonical
/// induction variable:
///
///   i = 0;
///   smax = n < 1 ? 1 : n;
///   do {
///     p[i] = 0.0;
///   } while (++i != smax);
///
/// Canonical induction variables are necessary because the loop passes
/// are designed around them. The most obvious example of this is the
/// LoopInfo analysis, which doesn't remember trip count values. It
/// expects to be able to rediscover the trip count each time it is
/// needed, and it does this using a simple analyis that only succeeds if
/// the loop has a canonical induction variable.
///
/// However, when it comes time to generate code, the maximum operation
/// can be quite costly, especially if it's inside of an outer loop.
///
/// This function solves this problem by detecting this type of loop and
/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
/// the instructions for the maximum computation.
///
ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
                                           IVStrideUse* &CondUse) {
  // Check that the loop matches the pattern we're looking for.
  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
      Cond->getPredicate() != CmpInst::ICMP_NE)
    return Cond;

  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
  if (!Sel || !Sel->hasOneUse()) return Cond;

  SCEVHandle IterationCount = SE->getIterationCount(L);
  if (isa<SCEVCouldNotCompute>(IterationCount))
    return Cond;
  SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());

  // Adjust for an annoying getIterationCount quirk.
  IterationCount = SE->getAddExpr(IterationCount, One);

  // Check for a max calculation that matches the pattern.
  SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
  if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;

  SCEVHandle SMaxLHS = SMax->getOperand(0);
  SCEVHandle SMaxRHS = SMax->getOperand(1);
  if (!SMaxLHS || SMaxLHS != One) return Cond;

  // Check the relevant induction variable for conformance to
  // the pattern.
  SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
  SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
  if (!AR || !AR->isAffine() ||
      AR->getStart() != One ||
      AR->getStepRecurrence(*SE) != One)
    return Cond;

  // Check the right operand of the select, and remember it, as it will
  // be used in the new comparison instruction.
  Value *NewRHS = 0;
  if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
    NewRHS = Sel->getOperand(1);
  else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
    NewRHS = Sel->getOperand(2);
  if (!NewRHS) return Cond;

  // Ok, everything looks ok to change the condition into an SLT or SGE and
  // delete the max calculation.
  ICmpInst *NewCond =
    new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
                   CmpInst::ICMP_SLT :
                   CmpInst::ICMP_SGE,
                 Cond->getOperand(0), NewRHS, "scmp", Cond);

  // Delete the max calculation instructions.
  SE->deleteValueFromRecords(Cond);
  Cond->replaceAllUsesWith(NewCond);
  Cond->eraseFromParent();
  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
  SE->deleteValueFromRecords(Sel);
  Sel->eraseFromParent();
  if (Cmp->use_empty()) {
    SE->deleteValueFromRecords(Cmp);
    Cmp->eraseFromParent();
  }
  CondUse->User = NewCond;
  return NewCond;
}

/// OptimizeShadowIV - If IV is used in a int-to-float cast
/// inside the loop then try to eliminate the cast opeation.
void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {

  SCEVHandle IterationCount = SE->getIterationCount(L);
  if (isa<SCEVCouldNotCompute>(IterationCount))
    return;

  for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
       ++Stride) {
    std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI = 
      IVUsesByStride.find(StrideOrder[Stride]);
    assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
    if (!isa<SCEVConstant>(SI->first))
      continue;

    for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
           E = SI->second.Users.end(); UI != E; /* empty */) {
      std::vector<IVStrideUse>::iterator CandidateUI = UI;
      ++UI;
      Instruction *ShadowUse = CandidateUI->User;
      const Type *DestTy = NULL;

      /* If shadow use is a int->float cast then insert a second IV
         to eliminate this cast.

           for (unsigned i = 0; i < n; ++i) 
             foo((double)i);

         is transformed into

           double d = 0.0;
           for (unsigned i = 0; i < n; ++i, ++d) 
             foo(d);
      */
      if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
        DestTy = UCast->getDestTy();
      else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
        DestTy = SCast->getDestTy();
      if (!DestTy) continue;

      if (TLI) {
        /* If target does not support DestTy natively then do not apply
           this transformation. */
        MVT DVT = TLI->getValueType(DestTy);
        if (!TLI->isTypeLegal(DVT)) continue;
      }

      PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
      if (!PH) continue;
      if (PH->getNumIncomingValues() != 2) continue;

      const Type *SrcTy = PH->getType();
      int Mantissa = DestTy->getFPMantissaWidth();
      if (Mantissa == -1) continue; 
      if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
        continue;

      unsigned Entry, Latch;
      if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
        Entry = 0;
        Latch = 1;
      } else {
        Entry = 1;
        Latch = 0;
      }
        
      ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
      if (!Init) continue;
      ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());

      BinaryOperator *Incr = 
        dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
      if (!Incr) continue;
      if (Incr->getOpcode() != Instruction::Add
          && Incr->getOpcode() != Instruction::Sub)
        continue;

      /* Initialize new IV, double d = 0.0 in above example. */
      ConstantInt *C = NULL;
      if (Incr->getOperand(0) == PH)
        C = dyn_cast<ConstantInt>(Incr->getOperand(1));
      else if (Incr->getOperand(1) == PH)
        C = dyn_cast<ConstantInt>(Incr->getOperand(0));
      else
        continue;

      if (!C) continue;

      /* Add new PHINode. */
      PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);

      /* create new increment. '++d' in above example. */
      ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
      BinaryOperator *NewIncr = 
        BinaryOperator::Create(Incr->getOpcode(),
                               NewPH, CFP, "IV.S.next.", Incr);

      NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
      NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));

      /* Remove cast operation */
      SE->deleteValueFromRecords(ShadowUse);
      ShadowUse->replaceAllUsesWith(NewPH);
      ShadowUse->eraseFromParent();
      SI->second.Users.erase(CandidateUI);
      NumShadow++;
      break;
    }
  }
}

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

  OptimizeShadowIV(L);

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

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

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

  // If the trip count is computed in terms of an smax (due to ScalarEvolution
  // being unable to find a sufficient guard, for example), change the loop
  // comparison to use SLT instead of NE.
  Cond = OptimizeSMax(L, Cond, CondUse);

  // If possible, change stride and operands of the compare instruction to
  // eliminate one stride.
  Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);

  // It's possible for the setcc instruction to be anywhere in the loop, and