LoopStrengthReduce.cpp

//===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file was developed by Nate Begeman and is distributed under the
// University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass performs a strength reduction on array references inside loops that
// have as one or more of their components the loop induction variable.  This is
// accomplished by creating a new Value to hold the initial value of the array
// access for the first iteration, and then creating a new GEP instruction in
// the loop to increment the value by the appropriate amount.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "loop-reduce"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Type.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Target/TargetLowering.h"
#include <algorithm>
#include <set>
using namespace llvm;

STATISTIC(NumReduced ,    "Number of GEPs strength reduced");
STATISTIC(NumInserted,    "Number of PHIs inserted");
STATISTIC(NumVariable,    "Number of PHIs with variable strides");
STATISTIC(NumEliminated , "Number of strides eliminated");

namespace {
  // Hidden options for help debugging.
  cl::opt<bool> AllowPHIIVReuse("lsr-allow-phi-iv-reuse",
                                cl::init(true), cl::Hidden);
}

namespace {

  struct BasedUser;

  /// IVStrideUse - Keep track of one use of a strided induction variable, where
  /// the stride is stored externally.  The Offset member keeps track of the 
  /// offset from the IV, User is the actual user of the operand, and
  /// 'OperandValToReplace' is the operand of the User that is the use.
  struct VISIBILITY_HIDDEN IVStrideUse {
    SCEVHandle Offset;
    Instruction *User;
    Value *OperandValToReplace;

    // isUseOfPostIncrementedValue - True if this should use the
    // post-incremented version of this IV, not the preincremented version.
    // This can only be set in special cases, such as the terminating setcc
    // instruction for a loop or uses dominated by the loop.
    bool isUseOfPostIncrementedValue;
    
    IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
      : Offset(Offs), User(U), OperandValToReplace(O),
        isUseOfPostIncrementedValue(false) {}
  };
  
  /// IVUsersOfOneStride - This structure keeps track of all instructions that
  /// have an operand that is based on the trip count multiplied by some stride.
  /// The stride for all of these users is common and kept external to this
  /// structure.
  struct VISIBILITY_HIDDEN IVUsersOfOneStride {
    /// Users - Keep track of all of the users of this stride as well as the
    /// initial value and the operand that uses the IV.
    std::vector<IVStrideUse> Users;
    
    void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
      Users.push_back(IVStrideUse(Offset, User, Operand));
    }
  };

  /// IVInfo - This structure keeps track of one IV expression inserted during
  /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
  /// well as the PHI node and increment value created for rewrite.
  struct VISIBILITY_HIDDEN IVExpr {
    SCEVHandle  Stride;
    SCEVHandle  Base;
    PHINode    *PHI;
    Value      *IncV;

    IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
           Value *incv)
      : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
  };

  /// IVsOfOneStride - This structure keeps track of all IV expression inserted
  /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
  struct VISIBILITY_HIDDEN IVsOfOneStride {
    std::vector<IVExpr> IVs;

    void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
               Value *IncV) {
      IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
    }
  };

  class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
    LoopInfo *LI;
    DominatorTree *DT;
    ScalarEvolution *SE;
    const TargetData *TD;
    const Type *UIntPtrTy;
    bool Changed;

    /// IVUsesByStride - Keep track of all uses of induction variables that we
    /// are interested in.  The key of the map is the stride of the access.
    std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;

    /// IVsByStride - Keep track of all IVs that have been inserted for a
    /// particular stride.
    std::map<SCEVHandle, IVsOfOneStride> IVsByStride;

    /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
    /// We use this to iterate over the IVUsesByStride collection without being
    /// dependent on random ordering of pointers in the process.
    SmallVector<SCEVHandle, 16> StrideOrder;

    /// CastedValues - As we need to cast values to uintptr_t, this keeps track
    /// of the casted version of each value.  This is accessed by
    /// getCastedVersionOf.
    DenseMap<Value*, Value*> CastedPointers;

    /// DeadInsts - Keep track of instructions we may have made dead, so that
    /// we can remove them after we are done working.
    SmallPtrSet<Instruction*,16> DeadInsts;

    /// TLI - Keep a pointer of a TargetLowering to consult for determining
    /// transformation profitability.
    const TargetLowering *TLI;

  public:
    static char ID; // Pass ID, replacement for typeid
    explicit LoopStrengthReduce(const TargetLowering *tli = NULL) : 
      LoopPass((intptr_t)&ID), TLI(tli) {
    }

    bool runOnLoop(Loop *L, LPPassManager &LPM);

    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
      // We split critical edges, so we change the CFG.  However, we do update
      // many analyses if they are around.
      AU.addPreservedID(LoopSimplifyID);
      AU.addPreserved<LoopInfo>();
      AU.addPreserved<DominanceFrontier>();
      AU.addPreserved<DominatorTree>();

      AU.addRequiredID(LoopSimplifyID);
      AU.addRequired<LoopInfo>();
      AU.addRequired<DominatorTree>();
      AU.addRequired<TargetData>();
      AU.addRequired<ScalarEvolution>();
    }
    
    /// getCastedVersionOf - Return the specified value casted to uintptr_t.
    ///
    Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
private:
    bool AddUsersIfInteresting(Instruction *I, Loop *L,
                               SmallPtrSet<Instruction*,16> &Processed);
    SCEVHandle GetExpressionSCEV(Instruction *E);
    ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
                                  IVStrideUse* &CondUse,
                                  const SCEVHandle* &CondStride);
    void OptimizeIndvars(Loop *L);
    bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
                       const SCEVHandle *&CondStride);
    bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
    unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
                             IVExpr&, const Type*,
                             const std::vector<BasedUser>& UsersToProcess);
    bool ValidStride(bool, int64_t,
                     const std::vector<BasedUser>& UsersToProcess);
    SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
                              IVUsersOfOneStride &Uses,
                              Loop *L,
                              bool &AllUsesAreAddresses,
                              std::vector<BasedUser> &UsersToProcess);
    void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
                                      IVUsersOfOneStride &Uses,
                                      Loop *L, bool isOnlyStride);
    void DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*,16> &Insts);
  };
  char LoopStrengthReduce::ID = 0;
  RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
}

LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
  return new LoopStrengthReduce(TLI);
}

/// getCastedVersionOf - Return the specified value casted to uintptr_t. This
/// assumes that the Value* V is of integer or pointer type only.
///
Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode, 
                                              Value *V) {
  if (V->getType() == UIntPtrTy) return V;
  if (Constant *CB = dyn_cast<Constant>(V))
    return ConstantExpr::getCast(opcode, CB, UIntPtrTy);

  Value *&New = CastedPointers[V];
  if (New) return New;
  
  New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
  DeadInsts.insert(cast<Instruction>(New));
  return New;
}


/// DeleteTriviallyDeadInstructions - If any of the instructions is the
/// specified set are trivially dead, delete them and see if this makes any of
/// their operands subsequently dead.
void LoopStrengthReduce::
DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*,16> &Insts) {
  while (!Insts.empty()) {
    Instruction *I = *Insts.begin();
    Insts.erase(I);

    if (PHINode *PN = dyn_cast<PHINode>(I)) {
      // If all incoming values to the Phi are the same, we can replace the Phi
      // with that value.
      if (Value *PNV = PN->hasConstantValue()) {
        if (Instruction *U = dyn_cast<Instruction>(PNV))
          Insts.insert(U);
        PN->replaceAllUsesWith(PNV);
        SE->deleteValueFromRecords(PN);
        PN->eraseFromParent();
        Changed = true;
        continue;
      }
    }

    if (isInstructionTriviallyDead(I)) {
      for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
        if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
          Insts.insert(U);
      SE->deleteValueFromRecords(I);
      I->eraseFromParent();
      Changed = true;
    }
  }
}


/// GetExpressionSCEV - Compute and return the SCEV for the specified
/// instruction.
SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
  // Pointer to pointer bitcast instructions return the same value as their
  // operand.
  if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
    if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
      return SE->getSCEV(BCI);
    SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
    SE->setSCEV(BCI, R);
    return R;
  }

  // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
  // If this is a GEP that SE doesn't know about, compute it now and insert it.
  // If this is not a GEP, or if we have already done this computation, just let
  // SE figure it out.
  GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
  if (!GEP || SE->hasSCEV(GEP))
    return SE->getSCEV(Exp);
    
  // Analyze all of the subscripts of this getelementptr instruction, looking
  // for uses that are determined by the trip count of the loop.  First, skip
  // all operands the are not dependent on the IV.

  // Build up the base expression.  Insert an LLVM cast of the pointer to
  // uintptr_t first.
  SCEVHandle GEPVal = SE->getUnknown(
      getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));

  gep_type_iterator GTI = gep_type_begin(GEP);
  
  for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
    // If this is a use of a recurrence that we can analyze, and it comes before
    // Op does in the GEP operand list, we will handle this when we process this
    // operand.
    if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
      const StructLayout *SL = TD->getStructLayout(STy);
      unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
      uint64_t Offset = SL->getElementOffset(Idx);
      GEPVal = SE->getAddExpr(GEPVal,
                             SE->getIntegerSCEV(Offset, UIntPtrTy));
    } else {
      unsigned GEPOpiBits = 
        GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
      unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
      Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ? 
          Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
            Instruction::BitCast));
      Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
      SCEVHandle Idx = SE->getSCEV(OpVal);

      uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
      if (TypeSize != 1)
        Idx = SE->getMulExpr(Idx,
                            SE->getConstant(ConstantInt::get(UIntPtrTy,
                                                             TypeSize)));
      GEPVal = SE->getAddExpr(GEPVal, Idx);
    }
  }

  SE->setSCEV(GEP, GEPVal);
  return GEPVal;
}

/// getSCEVStartAndStride - Compute the start and stride of this expression,
/// returning false if the expression is not a start/stride pair, or true if it
/// is.  The stride must be a loop invariant expression, but the start may be
/// a mix of loop invariant and loop variant expressions.
static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
                                  SCEVHandle &Start, SCEVHandle &Stride,
                                  ScalarEvolution *SE) {
  SCEVHandle TheAddRec = Start;   // Initialize to zero.

  // If the outer level is an AddExpr, the operands are all start values except
  // for a nested AddRecExpr.
  if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
    for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
      if (SCEVAddRecExpr *AddRec =
             dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
        if (AddRec->getLoop() == L)
          TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
        else
          return false;  // Nested IV of some sort?
      } else {
        Start = SE->getAddExpr(Start, AE->getOperand(i));
      }
        
  } else if (isa<SCEVAddRecExpr>(SH)) {
    TheAddRec = SH;
  } else {
    return false;  // not analyzable.
  }
  
  SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
  if (!AddRec || AddRec->getLoop() != L) return false;
  
  // FIXME: Generalize to non-affine IV's.
  if (!AddRec->isAffine()) return false;

  Start = SE->getAddExpr(Start, AddRec->getOperand(0));
  
  if (!isa<SCEVConstant>(AddRec->getOperand(1)))
    DOUT << "[" << L->getHeader()->getName()
         << "] Variable stride: " << *AddRec << "\n";

  Stride = AddRec->getOperand(1);
  return true;
}

/// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
/// and now we need to decide whether the user should use the preinc or post-inc
/// value.  If this user should use the post-inc version of the IV, return true.
///
/// Choosing wrong here can break dominance properties (if we choose to use the
/// post-inc value when we cannot) or it can end up adding extra live-ranges to
/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
/// should use the post-inc value).
static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
                                       Loop *L, DominatorTree *DT, Pass *P,
                                       SmallPtrSet<Instruction*,16> &DeadInsts){
  // If the user is in the loop, use the preinc value.
  if (L->contains(User->getParent())) return false;
  
  BasicBlock *LatchBlock = L->getLoopLatch();
  
  // Ok, the user is outside of the loop.  If it is dominated by the latch
  // block, use the post-inc value.
  if (DT->dominates(LatchBlock, User->getParent()))
    return true;

  // There is one case we have to be careful of: PHI nodes.  These little guys
  // can live in blocks that do not dominate the latch block, but (since their
  // uses occur in the predecessor block, not the block the PHI lives in) should
  // still use the post-inc value.  Check for this case now.
  PHINode *PN = dyn_cast<PHINode>(User);
  if (!PN) return false;  // not a phi, not dominated by latch block.
  
  // Look at all of the uses of IV by the PHI node.  If any use corresponds to
  // a block that is not dominated by the latch block, give up and use the
  // preincremented value.
  unsigned NumUses = 0;
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    if (PN->getIncomingValue(i) == IV) {
      ++NumUses;
      if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
        return false;
    }

  // Okay, all uses of IV by PN are in predecessor blocks that really are
  // dominated by the latch block.  Split the critical edges and use the
  // post-incremented value.
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    if (PN->getIncomingValue(i) == IV) {
      SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
      // Splitting the critical edge can reduce the number of entries in this
      // PHI.
      e = PN->getNumIncomingValues();
      if (--NumUses == 0) break;
    }

  // PHI node might have become a constant value after SplitCriticalEdge.
  DeadInsts.insert(User);
  
  return true;
}

  

/// AddUsersIfInteresting - Inspect the specified instruction.  If it is a
/// reducible SCEV, recursively add its users to the IVUsesByStride set and
/// return true.  Otherwise, return false.
bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
                                      SmallPtrSet<Instruction*,16> &Processed) {
  if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
      return false;   // Void and FP expressions cannot be reduced.
  if (!Processed.insert(I))
    return true;    // Instruction already handled.
  
  // Get the symbolic expression for this instruction.
  SCEVHandle ISE = GetExpressionSCEV(I);
  if (isa<SCEVCouldNotCompute>(ISE)) return false;
  
  // Get the start and stride for this expression.
  SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
  SCEVHandle Stride = Start;
  if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
    return false;  // Non-reducible symbolic expression, bail out.

  std::vector<Instruction *> IUsers;
  // Collect all I uses now because IVUseShouldUsePostIncValue may 
  // invalidate use_iterator.
  for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
    IUsers.push_back(cast<Instruction>(*UI));

  for (unsigned iused_index = 0, iused_size = IUsers.size(); 
       iused_index != iused_size; ++iused_index) {

    Instruction *User = IUsers[iused_index];

    // Do not infinitely recurse on PHI nodes.
    if (isa<PHINode>(User) && Processed.count(User))
      continue;

    // If this is an instruction defined in a nested loop, or outside this loop,
    // don't recurse into it.
    bool AddUserToIVUsers = false;
    if (LI->getLoopFor(User->getParent()) != L) {
      DOUT << "FOUND USER in other loop: " << *User
           << "   OF SCEV: " << *ISE << "\n";
      AddUserToIVUsers = true;
    } else if (!AddUsersIfInteresting(User, L, Processed)) {
      DOUT << "FOUND USER: " << *User
           << "   OF SCEV: " << *ISE << "\n";
      AddUserToIVUsers = true;
    }

    if (AddUserToIVUsers) {
      IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
      if (StrideUses.Users.empty())     // First occurance of this stride?
        StrideOrder.push_back(Stride);
      
      // Okay, we found a user that we cannot reduce.  Analyze the instruction
      // and decide what to do with it.  If we are a use inside of the loop, use
      // the value before incrementation, otherwise use it after incrementation.
      if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
        // The value used will be incremented by the stride more than we are
        // expecting, so subtract this off.
        SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
        StrideUses.addUser(NewStart, User, I);
        StrideUses.Users.back().isUseOfPostIncrementedValue = true;
        DOUT << "   USING POSTINC SCEV, START=" << *NewStart<< "\n";
      } else {        
        StrideUses.addUser(Start, User, I);
      }
    }
  }
  return true;
}

namespace {
  /// BasedUser - For a particular base value, keep information about how we've
  /// partitioned the expression so far.
  struct BasedUser {
    /// SE - The current ScalarEvolution object.
    ScalarEvolution *SE;

    /// Base - The Base value for the PHI node that needs to be inserted for
    /// this use.  As the use is processed, information gets moved from this
    /// field to the Imm field (below).  BasedUser values are sorted by this
    /// field.
    SCEVHandle Base;
    
    /// Inst - The instruction using the induction variable.
    Instruction *Inst;

    /// OperandValToReplace - The operand value of Inst to replace with the
    /// EmittedBase.
    Value *OperandValToReplace;

    /// Imm - The immediate value that should be added to the base immediately
    /// before Inst, because it will be folded into the imm field of the
    /// instruction.
    SCEVHandle Imm;

    /// EmittedBase - The actual value* to use for the base value of this
    /// operation.  This is null if we should just use zero so far.
    Value *EmittedBase;

    // isUseOfPostIncrementedValue - True if this should use the
    // post-incremented version of this IV, not the preincremented version.
    // This can only be set in special cases, such as the terminating setcc
    // instruction for a loop and uses outside the loop that are dominated by
    // the loop.
    bool isUseOfPostIncrementedValue;
    
    BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
      : SE(se), Base(IVSU.Offset), Inst(IVSU.User), 
        OperandValToReplace(IVSU.OperandValToReplace), 
        Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
        isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}

    // Once we rewrite the code to insert the new IVs we want, update the
    // operands of Inst to use the new expression 'NewBase', with 'Imm' added
    // to it.
    void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
                                       SCEVExpander &Rewriter, Loop *L, Pass *P,
                                       SmallPtrSet<Instruction*,16> &DeadInsts);
    
    Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase, 
                                       SCEVExpander &Rewriter,
                                       Instruction *IP, Loop *L);
    void dump() const;
  };
}

void BasedUser::dump() const {
  cerr << " Base=" << *Base;
  cerr << " Imm=" << *Imm;
  if (EmittedBase)
    cerr << "  EB=" << *EmittedBase;

  cerr << "   Inst: " << *Inst;
}

Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase, 
                                              SCEVExpander &Rewriter,
                                              Instruction *IP, Loop *L) {
  // Figure out where we *really* want to insert this code.  In particular, if
  // the user is inside of a loop that is nested inside of L, we really don't
  // want to insert this expression before the user, we'd rather pull it out as
  // many loops as possible.
  LoopInfo &LI = Rewriter.getLoopInfo();
  Instruction *BaseInsertPt = IP;
  
  // Figure out the most-nested loop that IP is in.
  Loop *InsertLoop = LI.getLoopFor(IP->getParent());
  
  // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
  // the preheader of the outer-most loop where NewBase is not loop invariant.
  while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
    BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
    InsertLoop = InsertLoop->getParentLoop();
  }
  
  // If there is no immediate value, skip the next part.
  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
    if (SC->getValue()->isZero())
      return Rewriter.expandCodeFor(NewBase, BaseInsertPt);

  Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);

  // If we are inserting the base and imm values in the same block, make sure to
  // adjust the IP position if insertion reused a result.
  if (IP == BaseInsertPt)
    IP = Rewriter.getInsertionPoint();
  
  // Always emit the immediate (if non-zero) into the same block as the user.
  SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
  return Rewriter.expandCodeFor(NewValSCEV, IP);
  
}


// Once we rewrite the code to insert the new IVs we want, update the
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
// to it.
void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
                                      SCEVExpander &Rewriter, Loop *L, Pass *P,
                                      SmallPtrSet<Instruction*,16> &DeadInsts) {
  if (!isa<PHINode>(Inst)) {
    // By default, insert code at the user instruction.
    BasicBlock::iterator InsertPt = Inst;
    
    // However, if the Operand is itself an instruction, the (potentially
    // complex) inserted code may be shared by many users.  Because of this, we
    // want to emit code for the computation of the operand right before its old
    // computation.  This is usually safe, because we obviously used to use the
    // computation when it was computed in its current block.  However, in some
    // cases (e.g. use of a post-incremented induction variable) the NewBase
    // value will be pinned to live somewhere after the original computation.
    // In this case, we have to back off.
    if (!isUseOfPostIncrementedValue) {
      if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) { 
        InsertPt = OpInst;
        while (isa<PHINode>(InsertPt)) ++InsertPt;
      }
    }
    Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
    // Adjust the type back to match the Inst. Note that we can't use InsertPt
    // here because the SCEVExpander may have inserted the instructions after
    // that point, in its efforts to avoid inserting redundant expressions.
    if (isa<PointerType>(OperandValToReplace->getType())) {
      NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
                                            NewVal,
                                            OperandValToReplace->getType());
    }
    // Replace the use of the operand Value with the new Phi we just created.
    Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
    DOUT << "    CHANGED: IMM =" << *Imm;
    DOUT << "  \tNEWBASE =" << *NewBase;
    DOUT << "  \tInst = " << *Inst;
    return;
  }
  
  // PHI nodes are more complex.  We have to insert one copy of the NewBase+Imm
  // expression into each operand block that uses it.  Note that PHI nodes can
  // have multiple entries for the same predecessor.  We use a map to make sure
  // that a PHI node only has a single Value* for each predecessor (which also
  // prevents us from inserting duplicate code in some blocks).
  DenseMap<BasicBlock*, Value*> InsertedCode;
  PHINode *PN = cast<PHINode>(Inst);
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
    if (PN->getIncomingValue(i) == OperandValToReplace) {
      // If this is a critical edge, split the edge so that we do not insert the
      // code on all predecessor/successor paths.  We do this unless this is the
      // canonical backedge for this loop, as this can make some inserted code
      // be in an illegal position.
      BasicBlock *PHIPred = PN->getIncomingBlock(i);
      if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
          (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
        
        // First step, split the critical edge.
        SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
            
        // Next step: move the basic block.  In particular, if the PHI node
        // is outside of the loop, and PredTI is in the loop, we want to
        // move the block to be immediately before the PHI block, not
        // immediately after PredTI.
        if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
          BasicBlock *NewBB = PN->getIncomingBlock(i);
          NewBB->moveBefore(PN->getParent());
        }
        
        // Splitting the edge can reduce the number of PHI entries we have.
        e = PN->getNumIncomingValues();
      }

      Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
      if (!Code) {
        // Insert the code into the end of the predecessor block.
        Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
        Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);

        // Adjust the type back to match the PHI. Note that we can't use
        // InsertPt here because the SCEVExpander may have inserted its
        // instructions after that point, in its efforts to avoid inserting
        // redundant expressions.
        if (isa<PointerType>(PN->getType())) {
          Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
                                              Code,
                                              PN->getType());
        }
      }
      
      // Replace the use of the operand Value with the new Phi we just created.
      PN->setIncomingValue(i, Code);
      Rewriter.clear();
    }
  }

  // PHI node might have become a constant value after SplitCriticalEdge.
  DeadInsts.insert(Inst);

  DOUT << "    CHANGED: IMM =" << *Imm << "  Inst = " << *Inst;
}


/// isTargetConstant - Return true if the following can be referenced by the
/// immediate field of a target instruction.
static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
                             const TargetLowering *TLI) {
  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
    int64_t VC = SC->getValue()->getSExtValue();
    if (TLI) {
      TargetLowering::AddrMode AM;
      AM.BaseOffs = VC;
      return TLI->isLegalAddressingMode(AM, UseTy);
    } else {
      // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
      return (VC > -(1 << 16) && VC < (1 << 16)-1);
    }
  }

  if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
      if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
        Constant *Op0 = CE->getOperand(0);
        if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
          TargetLowering::AddrMode AM;
          AM.BaseGV = GV;
          return TLI->isLegalAddressingMode(AM, UseTy);
        }
      }
  return false;
}

/// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
/// loop varying to the Imm operand.
static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
                                            Loop *L, ScalarEvolution *SE) {
  if (Val->isLoopInvariant(L)) return;  // Nothing to do.
  
  if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
    std::vector<SCEVHandle> NewOps;
    NewOps.reserve(SAE->getNumOperands());
    
    for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
      if (!SAE->getOperand(i)->isLoopInvariant(L)) {
        // If this is a loop-variant expression, it must stay in the immediate
        // field of the expression.
        Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
      } else {
        NewOps.push_back(SAE->getOperand(i));
      }

    if (NewOps.empty())
      Val = SE->getIntegerSCEV(0, Val->getType());
    else
      Val = SE->getAddExpr(NewOps);
  } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
    // Try to pull immediates out of the start value of nested addrec's.
    SCEVHandle Start = SARE->getStart();
    MoveLoopVariantsToImediateField(Start, Imm, L, SE);
    
    std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
    Ops[0] = Start;
    Val = SE->getAddRecExpr(Ops, SARE->getLoop());
  } else {
    // Otherwise, all of Val is variant, move the whole thing over.
    Imm = SE->getAddExpr(Imm, Val);
    Val = SE->getIntegerSCEV(0, Val->getType());
  }
}


/// MoveImmediateValues - Look at Val, and pull out any additions of constants
/// that can fit into the immediate field of instructions in the target.
/// Accumulate these immediate values into the Imm value.
static void MoveImmediateValues(const TargetLowering *TLI,
                                Instruction *User,
                                SCEVHandle &Val, SCEVHandle &Imm,
                                bool isAddress, Loop *L,
                                ScalarEvolution *SE) {
  const Type *UseTy = User->getType();
  if (StoreInst *SI = dyn_cast<StoreInst>(User))
    UseTy = SI->getOperand(0)->getType();

  if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
    std::vector<SCEVHandle> NewOps;
    NewOps.reserve(SAE->getNumOperands());
    
    for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
      SCEVHandle NewOp = SAE->getOperand(i);
      MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
      
      if (!NewOp->isLoopInvariant(L)) {
        // If this is a loop-variant expression, it must stay in the immediate
        // field of the expression.
        Imm = SE->getAddExpr(Imm, NewOp);
      } else {
        NewOps.push_back(NewOp);
      }
    }

    if (NewOps.empty())
      Val = SE->getIntegerSCEV(0, Val->getType());
    else
      Val = SE->getAddExpr(NewOps);
    return;
  } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
    // Try to pull immediates out of the start value of nested addrec's.
    SCEVHandle Start = SARE->getStart();
    MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
    
    if (Start != SARE->getStart()) {
      std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
      Ops[0] = Start;
      Val = SE->getAddRecExpr(Ops, SARE->getLoop());
    }
    return;
  } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
    // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
    if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
        SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {

      SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
      SCEVHandle NewOp = SME->getOperand(1);
      MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
      
      // If we extracted something out of the subexpressions, see if we can 
      // simplify this!
      if (NewOp != SME->getOperand(1)) {
        // Scale SubImm up by "8".  If the result is a target constant, we are
        // good.
        SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
        if (isTargetConstant(SubImm, UseTy, TLI)) {
          // Accumulate the immediate.
          Imm = SE->getAddExpr(Imm, SubImm);
          
          // Update what is left of 'Val'.
          Val = SE->getMulExpr(SME->getOperand(0), NewOp);
          return;
        }
      }
    }
  }

  // Loop-variant expressions must stay in the immediate field of the
  // expression.
  if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
      !Val->isLoopInvariant(L)) {
    Imm = SE->getAddExpr(Imm, Val);
    Val = SE->getIntegerSCEV(0, Val->getType());
    return;
  }

  // Otherwise, no immediates to move.
}


/// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
/// added together.  This is used to reassociate common addition subexprs
/// together for maximal sharing when rewriting bases.
static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
                             SCEVHandle Expr,
                             ScalarEvolution *SE) {
  if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
    for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
      SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
  } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
    SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
    if (SARE->getOperand(0) == Zero) {
      SubExprs.push_back(Expr);
    } else {
      // Compute the addrec with zero as its base.
      std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
      Ops[0] = Zero;   // Start with zero base.
      SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
      

      SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
    }
  } else if (!isa<SCEVConstant>(Expr) ||
             !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
    // Do not add zero.
    SubExprs.push_back(Expr);
  }
}


/// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
/// removing any common subexpressions from it.  Anything truly common is
/// removed, accumulated, and returned.  This looks for things like (a+b+c) and
/// (a+c+d) -> (a+c).  The common expression is *removed* from the Bases.
static SCEVHandle 
RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
                                    ScalarEvolution *SE) {
  unsigned NumUses = Uses.size();

  // Only one use?  Use its base, regardless of what it is!
  SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
  SCEVHandle Result = Zero;
  if (NumUses == 1) {
    std::swap(Result, Uses[0].Base);
    return Result;
  }

  // To find common subexpressions, count how many of Uses use each expression.
  // If any subexpressions are used Uses.size() times, they are common.
  std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
  
  // UniqueSubExprs - Keep track of all of the subexpressions we see in the
  // order we see them.
  std::vector<SCEVHandle> UniqueSubExprs;

  std::vector<SCEVHandle> SubExprs;
  for (unsigned i = 0; i != NumUses; ++i) {
    // If the base is zero (which is common), return zero now, there are no
    // CSEs we can find.
    if (Uses[i].Base == Zero) return Zero;

    // Split the expression into subexprs.
    SeparateSubExprs(SubExprs, Uses[i].Base, SE);
    // Add one to SubExpressionUseCounts for each subexpr present.
    for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
      if (++SubExpressionUseCounts[SubExprs[j]] == 1)
        UniqueSubExprs.push_back(SubExprs[j]);
    SubExprs.clear();
  }

  // Now that we know how many times each is used, build Result.  Iterate over
  // UniqueSubexprs so that we have a stable ordering.
  for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
    std::map<SCEVHandle, unsigned>::iterator I = 
       SubExpressionUseCounts.find(UniqueSubExprs[i]);
    assert(I != SubExpressionUseCounts.end() && "Entry not found?");
    if (I->second == NumUses) {  // Found CSE!
      Result = SE->getAddExpr(Result, I->first);
    } else {
      // Remove non-cse's from SubExpressionUseCounts.
      SubExpressionUseCounts.erase(I);
    }
  }
  
  // If we found no CSE's, return now.
  if (Result == Zero) return Result;
  
  // Otherwise, remove all of the CSE's we found from each of the base values.
  for (unsigned i = 0; i != NumUses; ++i) {
    // 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, the non-shared expressions together.
    if (SubExprs.empty())
      Uses[i].Base = Zero;
    else
      Uses[i].Base = SE->getAddExpr(SubExprs);
    SubExprs.clear();
  }
 
  return Result;
}

/// isZero - returns true if the scalar evolution expression is zero.
///
static bool isZero(const SCEVHandle &V) {
  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
    return SC->getValue()->isZero();
  return false;
}

/// 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)) {