IndVarSimplify.cpp

      }
      else {
        // A wider extend was hidden behind a narrower one. This may induce
        // another round of IV widening in which the intermediate IV becomes
        // dead. It should be very rare.
        DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
              << " not wide enough to subsume " << *DU.NarrowUse << "\n");
        DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
        NewDef = DU.NarrowUse;
      }
    }
    if (NewDef != DU.NarrowUse) {
      DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
            << " replaced by " << *DU.WideDef << "\n");
      ++NumElimExt;
      DU.NarrowUse->replaceAllUsesWith(NewDef);
      DeadInsts.push_back(DU.NarrowUse);
    }
    // Now that the extend is gone, we want to expose it's uses for potential
    // further simplification. We don't need to directly inform SimplifyIVUsers
    // of the new users, because their parent IV will be processed later as a
    // new loop phi. If we preserved IVUsers analysis, we would also want to
    // push the uses of WideDef here.

    // No further widening is needed. The deceased [sz]ext had done it for us.
    return 0;
  }

  // Does this user itself evaluate to a recurrence after widening?
  const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
  if (!WideAddRec) {
      WideAddRec = GetExtendedOperandRecurrence(DU);
  }
  if (!WideAddRec) {
    // This user does not evaluate to a recurence after widening, so don't
    // follow it. Instead insert a Trunc to kill off the original use,
    // eventually isolating the original narrow IV so it can be removed.
    IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
    Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
    DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
    return 0;
  }
  // Assume block terminators cannot evaluate to a recurrence. We can't to
  // insert a Trunc after a terminator if there happens to be a critical edge.
  assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
         "SCEV is not expected to evaluate a block terminator");

  // Reuse the IV increment that SCEVExpander created as long as it dominates
  // NarrowUse.
  Instruction *WideUse = 0;
  if (WideAddRec == WideIncExpr
      && SCEVExpander::hoistStep(WideInc, DU.NarrowUse, DT))
    WideUse = WideInc;
  else {
    WideUse = CloneIVUser(DU);
    if (!WideUse)
      return 0;
  }
  // Evaluation of WideAddRec ensured that the narrow expression could be
  // extended outside the loop without overflow. This suggests that the wide use
  // evaluates to the same expression as the extended narrow use, but doesn't
  // absolutely guarantee it. Hence the following failsafe check. In rare cases
  // where it fails, we simply throw away the newly created wide use.
  if (WideAddRec != SE->getSCEV(WideUse)) {
    DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
          << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
    DeadInsts.push_back(WideUse);
    return 0;
  }

  // Returning WideUse pushes it on the worklist.
  return WideUse;
}

/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
///
void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
  for (Value::use_iterator UI = NarrowDef->use_begin(),
         UE = NarrowDef->use_end(); UI != UE; ++UI) {
    Instruction *NarrowUse = cast<Instruction>(*UI);

    // Handle data flow merges and bizarre phi cycles.
    if (!Widened.insert(NarrowUse))
      continue;

    NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
  }
}

/// CreateWideIV - Process a single induction variable. First use the
/// SCEVExpander to create a wide induction variable that evaluates to the same
/// recurrence as the original narrow IV. Then use a worklist to forward
/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
/// interesting IV users, the narrow IV will be isolated for removal by
/// DeleteDeadPHIs.
///
/// It would be simpler to delete uses as they are processed, but we must avoid
/// invalidating SCEV expressions.
///
PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
  // Is this phi an induction variable?
  const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
  if (!AddRec)
    return NULL;

  // Widen the induction variable expression.
  const SCEV *WideIVExpr = IsSigned ?
    SE->getSignExtendExpr(AddRec, WideType) :
    SE->getZeroExtendExpr(AddRec, WideType);

  assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
         "Expect the new IV expression to preserve its type");

  // Can the IV be extended outside the loop without overflow?
  AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
  if (!AddRec || AddRec->getLoop() != L)
    return NULL;

  // An AddRec must have loop-invariant operands. Since this AddRec is
  // materialized by a loop header phi, the expression cannot have any post-loop
  // operands, so they must dominate the loop header.
  assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
         SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
         && "Loop header phi recurrence inputs do not dominate the loop");

  // The rewriter provides a value for the desired IV expression. This may
  // either find an existing phi or materialize a new one. Either way, we
  // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
  // of the phi-SCC dominates the loop entry.
  Instruction *InsertPt = L->getHeader()->begin();
  WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));

  // Remembering the WideIV increment generated by SCEVExpander allows
  // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
  // employ a general reuse mechanism because the call above is the only call to
  // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
  if (BasicBlock *LatchBlock = L->getLoopLatch()) {
    WideInc =
      cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
    WideIncExpr = SE->getSCEV(WideInc);
  }

  DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
  ++NumWidened;

  // Traverse the def-use chain using a worklist starting at the original IV.
  assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );

  Widened.insert(OrigPhi);
  pushNarrowIVUsers(OrigPhi, WidePhi);

  while (!NarrowIVUsers.empty()) {
    NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();

    // Process a def-use edge. This may replace the use, so don't hold a
    // use_iterator across it.
    Instruction *WideUse = WidenIVUse(DU);

    // Follow all def-use edges from the previous narrow use.
    if (WideUse)
      pushNarrowIVUsers(DU.NarrowUse, WideUse);

    // WidenIVUse may have removed the def-use edge.
    if (DU.NarrowDef->use_empty())
      DeadInsts.push_back(DU.NarrowDef);
  }
  return WidePhi;
}

//===----------------------------------------------------------------------===//
//  Simplification of IV users based on SCEV evaluation.
//===----------------------------------------------------------------------===//


/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
/// users. Each successive simplification may push more users which may
/// themselves be candidates for simplification.
///
/// Sign/Zero extend elimination is interleaved with IV simplification.
///
void IndVarSimplify::SimplifyAndExtend(Loop *L,
                                       SCEVExpander &Rewriter,
                                       LPPassManager &LPM) {
  SmallVector<WideIVInfo, 8> WideIVs;

  SmallVector<PHINode*, 8> LoopPhis;
  for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
    LoopPhis.push_back(cast<PHINode>(I));
  }
  // Each round of simplification iterates through the SimplifyIVUsers worklist
  // for all current phis, then determines whether any IVs can be
  // widened. Widening adds new phis to LoopPhis, inducing another round of
  // simplification on the wide IVs.
  while (!LoopPhis.empty()) {
    // Evaluate as many IV expressions as possible before widening any IVs. This
    // forces SCEV to set no-wrap flags before evaluating sign/zero
    // extension. The first time SCEV attempts to normalize sign/zero extension,
    // the result becomes final. So for the most predictable results, we delay
    // evaluation of sign/zero extend evaluation until needed, and avoid running
    // other SCEV based analysis prior to SimplifyAndExtend.
    do {
      PHINode *CurrIV = LoopPhis.pop_back_val();

      // Information about sign/zero extensions of CurrIV.
      WideIVVisitor WIV(CurrIV, SE, TD);

      Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);

      if (WIV.WI.WidestNativeType) {
        WideIVs.push_back(WIV.WI);
      }
    } while(!LoopPhis.empty());

    for (; !WideIVs.empty(); WideIVs.pop_back()) {
      WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
      if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
        Changed = true;
        LoopPhis.push_back(WidePhi);
      }
    }
  }
}

//===----------------------------------------------------------------------===//
//  LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
//===----------------------------------------------------------------------===//

/// Check for expressions that ScalarEvolution generates to compute
/// BackedgeTakenInfo. If these expressions have not been reduced, then
/// expanding them may incur additional cost (albeit in the loop preheader).
static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
                                ScalarEvolution *SE) {
  // If the backedge-taken count is a UDiv, it's very likely a UDiv that
  // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
  // precise expression, rather than a UDiv from the user's code. If we can't
  // find a UDiv in the code with some simple searching, assume the former and
  // forego rewriting the loop.
  if (isa<SCEVUDivExpr>(S)) {
    ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
    if (!OrigCond) return true;
    const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
    R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
    if (R != S) {
      const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
      L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
      if (L != S)
        return true;
    }
  }

  if (EnableIVRewrite)
    return false;

  // Recurse past add expressions, which commonly occur in the
  // BackedgeTakenCount. They may already exist in program code, and if not,
  // they are not too expensive rematerialize.
  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
    for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
         I != E; ++I) {
      if (isHighCostExpansion(*I, BI, SE))
        return true;
    }
    return false;
  }

  // HowManyLessThans uses a Max expression whenever the loop is not guarded by
  // the exit condition.
  if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
    return true;

  // If we haven't recognized an expensive SCEV patter, assume its an expression
  // produced by program code.
  return false;
}

/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
/// count expression can be safely and cheaply expanded into an instruction
/// sequence that can be used by LinearFunctionTestReplace.
///
/// TODO: This fails for pointer-type loop counters with greater than one byte
/// strides, consequently preventing LFTR from running. For the purpose of LFTR
/// we could skip this check in the case that the LFTR loop counter (chosen by
/// FindLoopCounter) is also pointer type. Instead, we could directly convert
/// the loop test to an inequality test by checking the target data's alignment
/// of element types (given that the initial pointer value originates from or is
/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
/// However, we don't yet have a strong motivation for converting loop tests
/// into inequality tests.
static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
  const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
      BackedgeTakenCount->isZero())
    return false;

  if (!L->getExitingBlock())
    return false;

  // Can't rewrite non-branch yet.
  BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
  if (!BI)
    return false;

  if (isHighCostExpansion(BackedgeTakenCount, BI, SE))
    return false;

  return true;
}

/// getBackedgeIVType - Get the widest type used by the loop test after peeking
/// through Truncs.
///
/// TODO: Unnecessary when ForceLFTR is removed.
static Type *getBackedgeIVType(Loop *L) {
  if (!L->getExitingBlock())
    return 0;

  // Can't rewrite non-branch yet.
  BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
  if (!BI)
    return 0;

  ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
  if (!Cond)
    return 0;

  Type *Ty = 0;
  for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
      OI != OE; ++OI) {
    assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
    TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
    if (!Trunc)
      continue;

    return Trunc->getSrcTy();
  }
  return Ty;
}

/// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
/// invariant value to the phi.
static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
  Instruction *IncI = dyn_cast<Instruction>(IncV);
  if (!IncI)
    return 0;

  switch (IncI->getOpcode()) {
  case Instruction::Add:
  case Instruction::Sub:
    break;
  case Instruction::GetElementPtr:
    // An IV counter must preserve its type.
    if (IncI->getNumOperands() == 2)
      break;
  default:
    return 0;
  }

  PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
  if (Phi && Phi->getParent() == L->getHeader()) {
    if (isLoopInvariant(IncI->getOperand(1), L, DT))
      return Phi;
    return 0;
  }
  if (IncI->getOpcode() == Instruction::GetElementPtr)
    return 0;

  // Allow add/sub to be commuted.
  Phi = dyn_cast<PHINode>(IncI->getOperand(1));
  if (Phi && Phi->getParent() == L->getHeader()) {
    if (isLoopInvariant(IncI->getOperand(0), L, DT))
      return Phi;
  }
  return 0;
}

/// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
/// that the current exit test is already sufficiently canonical.
static bool needsLFTR(Loop *L, DominatorTree *DT) {
  assert(L->getExitingBlock() && "expected loop exit");

  BasicBlock *LatchBlock = L->getLoopLatch();
  // Don't bother with LFTR if the loop is not properly simplified.
  if (!LatchBlock)
    return false;

  BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
  assert(BI && "expected exit branch");

  // Do LFTR to simplify the exit condition to an ICMP.
  ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
  if (!Cond)
    return true;

  // Do LFTR to simplify the exit ICMP to EQ/NE
  ICmpInst::Predicate Pred = Cond->getPredicate();
  if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
    return true;

  // Look for a loop invariant RHS
  Value *LHS = Cond->getOperand(0);
  Value *RHS = Cond->getOperand(1);
  if (!isLoopInvariant(RHS, L, DT)) {
    if (!isLoopInvariant(LHS, L, DT))
      return true;
    std::swap(LHS, RHS);
  }
  // Look for a simple IV counter LHS
  PHINode *Phi = dyn_cast<PHINode>(LHS);
  if (!Phi)
    Phi = getLoopPhiForCounter(LHS, L, DT);

  if (!Phi)
    return true;

  // Do LFTR if the exit condition's IV is *not* a simple counter.
  Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch());
  return Phi != getLoopPhiForCounter(IncV, L, DT);
}

/// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
/// be rewritten) loop exit test.
static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
  int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
  Value *IncV = Phi->getIncomingValue(LatchIdx);

  for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
       UI != UE; ++UI) {
    if (*UI != Cond && *UI != IncV) return false;
  }

  for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
       UI != UE; ++UI) {
    if (*UI != Cond && *UI != Phi) return false;
  }
  return true;
}

/// FindLoopCounter - Find an affine IV in canonical form.
///
/// BECount may be an i8* pointer type. The pointer difference is already
/// valid count without scaling the address stride, so it remains a pointer
/// expression as far as SCEV is concerned.
///
/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
///
/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
/// This is difficult in general for SCEV because of potential overflow. But we
/// could at least handle constant BECounts.
static PHINode *
FindLoopCounter(Loop *L, const SCEV *BECount,
                ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) {
  uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());

  Value *Cond =
    cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();

  // Loop over all of the PHI nodes, looking for a simple counter.
  PHINode *BestPhi = 0;
  const SCEV *BestInit = 0;
  BasicBlock *LatchBlock = L->getLoopLatch();
  assert(LatchBlock && "needsLFTR should guarantee a loop latch");

  for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
    PHINode *Phi = cast<PHINode>(I);
    if (!SE->isSCEVable(Phi->getType()))
      continue;

    // Avoid comparing an integer IV against a pointer Limit.
    if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
      continue;

    const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
    if (!AR || AR->getLoop() != L || !AR->isAffine())
      continue;

    // AR may be a pointer type, while BECount is an integer type.
    // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
    // AR may not be a narrower type, or we may never exit.
    uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
    if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
      continue;

    const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
    if (!Step || !Step->isOne())
      continue;

    int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
    Value *IncV = Phi->getIncomingValue(LatchIdx);
    if (getLoopPhiForCounter(IncV, L, DT) != Phi)
      continue;

    const SCEV *Init = AR->getStart();

    if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
      // Don't force a live loop counter if another IV can be used.
      if (AlmostDeadIV(Phi, LatchBlock, Cond))
        continue;

      // Prefer to count-from-zero. This is a more "canonical" counter form. It
      // also prefers integer to pointer IVs.
      if (BestInit->isZero() != Init->isZero()) {
        if (BestInit->isZero())
          continue;
      }
      // If two IVs both count from zero or both count from nonzero then the
      // narrower is likely a dead phi that has been widened. Use the wider phi
      // to allow the other to be eliminated.
      if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
        continue;
    }
    BestPhi = Phi;
    BestInit = Init;
  }
  return BestPhi;
}

/// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
/// holds the RHS of the new loop test.
static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
                           SCEVExpander &Rewriter, ScalarEvolution *SE) {
  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
  assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
  const SCEV *IVInit = AR->getStart();

  // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
  // finds a valid pointer IV. Sign extend BECount in order to materialize a
  // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
  // the existing GEPs whenever possible.
  if (IndVar->getType()->isPointerTy()
      && !IVCount->getType()->isPointerTy()) {

    Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
    const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);

    // Expand the code for the iteration count.
    assert(SE->isLoopInvariant(IVOffset, L) &&
           "Computed iteration count is not loop invariant!");
    BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
    Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);

    Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
    assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
    // We could handle pointer IVs other than i8*, but we need to compensate for
    // gep index scaling. See canExpandBackedgeTakenCount comments.
    assert(SE->getSizeOfExpr(
             cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
           && "unit stride pointer IV must be i8*");

    IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
    return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
  }
  else {
    // In any other case, convert both IVInit and IVCount to integers before
    // comparing. This may result in SCEV expension of pointers, but in practice
    // SCEV will fold the pointer arithmetic away as such:
    // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
    //
    // Valid Cases: (1) both integers is most common; (2) both may be pointers
    // for simple memset-style loops; (3) IVInit is an integer and IVCount is a
    // pointer may occur when enable-iv-rewrite generates a canonical IV on top
    // of case #2.

    const SCEV *IVLimit = 0;
    // For unit stride, IVCount = Start + BECount with 2's complement overflow.
    // For non-zero Start, compute IVCount here.
    if (AR->getStart()->isZero())
      IVLimit = IVCount;
    else {
      assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
      const SCEV *IVInit = AR->getStart();

      // For integer IVs, truncate the IV before computing IVInit + BECount.
      if (SE->getTypeSizeInBits(IVInit->getType())
          > SE->getTypeSizeInBits(IVCount->getType()))
        IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());

      IVLimit = SE->getAddExpr(IVInit, IVCount);
    }
    // Expand the code for the iteration count.
    BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
    IRBuilder<> Builder(BI);
    assert(SE->isLoopInvariant(IVLimit, L) &&
           "Computed iteration count is not loop invariant!");
    // Ensure that we generate the same type as IndVar, or a smaller integer
    // type. In the presence of null pointer values, we have an integer type
    // SCEV expression (IVInit) for a pointer type IV value (IndVar).
    Type *LimitTy = IVCount->getType()->isPointerTy() ?
      IndVar->getType() : IVCount->getType();
    return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
  }
}

/// LinearFunctionTestReplace - This method rewrites the exit condition of the
/// loop to be a canonical != comparison against the incremented loop induction
/// variable.  This pass is able to rewrite the exit tests of any loop where the
/// SCEV analysis can determine a loop-invariant trip count of the loop, which
/// is actually a much broader range than just linear tests.
Value *IndVarSimplify::
LinearFunctionTestReplace(Loop *L,
                          const SCEV *BackedgeTakenCount,
                          PHINode *IndVar,
                          SCEVExpander &Rewriter) {
  assert(canExpandBackedgeTakenCount(L, SE) && "precondition");

  // LFTR can ignore IV overflow and truncate to the width of
  // BECount. This avoids materializing the add(zext(add)) expression.
  Type *CntTy = !EnableIVRewrite ?
    BackedgeTakenCount->getType() : IndVar->getType();

  const SCEV *IVCount = BackedgeTakenCount;

  // If the exiting block is the same as the backedge block, we prefer to
  // compare against the post-incremented value, otherwise we must compare
  // against the preincremented value.
  Value *CmpIndVar;
  if (L->getExitingBlock() == L->getLoopLatch()) {
    // Add one to the "backedge-taken" count to get the trip count.
    // If this addition may overflow, we have to be more pessimistic and
    // cast the induction variable before doing the add.
    const SCEV *N =
      SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1));
    if (CntTy == IVCount->getType())
      IVCount = N;
    else {
      const SCEV *Zero = SE->getConstant(IVCount->getType(), 0);
      if ((isa<SCEVConstant>(N) && !N->isZero()) ||
          SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
        // No overflow. Cast the sum.
        IVCount = SE->getTruncateOrZeroExtend(N, CntTy);
      } else {
        // Potential overflow. Cast before doing the add.
        IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
        IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1));
      }
    }
    // The BackedgeTaken expression contains the number of times that the
    // backedge branches to the loop header.  This is one less than the
    // number of times the loop executes, so use the incremented indvar.
    CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
  } else {
    // We must use the preincremented value...
    IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
    CmpIndVar = IndVar;
  }

  Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
  assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
         && "genLoopLimit missed a cast");

  // Insert a new icmp_ne or icmp_eq instruction before the branch.
  BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
  ICmpInst::Predicate P;
  if (L->contains(BI->getSuccessor(0)))
    P = ICmpInst::ICMP_NE;
  else
    P = ICmpInst::ICMP_EQ;

  DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
               << "      LHS:" << *CmpIndVar << '\n'
               << "       op:\t"
               << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
               << "      RHS:\t" << *ExitCnt << "\n"
               << "  IVCount:\t" << *IVCount << "\n");

  IRBuilder<> Builder(BI);
  if (SE->getTypeSizeInBits(CmpIndVar->getType())
      > SE->getTypeSizeInBits(ExitCnt->getType())) {
    CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
                                    "lftr.wideiv");
  }

  Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
  Value *OrigCond = BI->getCondition();
  // It's tempting to use replaceAllUsesWith here to fully replace the old
  // comparison, but that's not immediately safe, since users of the old
  // comparison may not be dominated by the new comparison. Instead, just
  // update the branch to use the new comparison; in the common case this
  // will make old comparison dead.
  BI->setCondition(Cond);
  DeadInsts.push_back(OrigCond);

  ++NumLFTR;
  Changed = true;
  return Cond;
}

//===----------------------------------------------------------------------===//
//  SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
//===----------------------------------------------------------------------===//

/// If there's a single exit block, sink any loop-invariant values that
/// were defined in the preheader but not used inside the loop into the
/// exit block to reduce register pressure in the loop.
void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
  BasicBlock *ExitBlock = L->getExitBlock();
  if (!ExitBlock) return;

  BasicBlock *Preheader = L->getLoopPreheader();
  if (!Preheader) return;

  Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
  BasicBlock::iterator I = Preheader->getTerminator();
  while (I != Preheader->begin()) {
    --I;
    // New instructions were inserted at the end of the preheader.
    if (isa<PHINode>(I))
      break;

    // Don't move instructions which might have side effects, since the side
    // effects need to complete before instructions inside the loop.  Also don't
    // move instructions which might read memory, since the loop may modify
    // memory. Note that it's okay if the instruction might have undefined
    // behavior: LoopSimplify guarantees that the preheader dominates the exit
    // block.
    if (I->mayHaveSideEffects() || I->mayReadFromMemory())
      continue;

    // Skip debug info intrinsics.
    if (isa<DbgInfoIntrinsic>(I))
      continue;

    // Skip landingpad instructions.
    if (isa<LandingPadInst>(I))
      continue;

    // Don't sink alloca: we never want to sink static alloca's out of the
    // entry block, and correctly sinking dynamic alloca's requires
    // checks for stacksave/stackrestore intrinsics.
    // FIXME: Refactor this check somehow?
    if (isa<AllocaInst>(I))
      continue;

    // Determine if there is a use in or before the loop (direct or
    // otherwise).
    bool UsedInLoop = false;
    for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
         UI != UE; ++UI) {
      User *U = *UI;
      BasicBlock *UseBB = cast<Instruction>(U)->getParent();
      if (PHINode *P = dyn_cast<PHINode>(U)) {
        unsigned i =
          PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
        UseBB = P->getIncomingBlock(i);
      }
      if (UseBB == Preheader || L->contains(UseBB)) {
        UsedInLoop = true;
        break;
      }
    }

    // If there is, the def must remain in the preheader.
    if (UsedInLoop)
      continue;

    // Otherwise, sink it to the exit block.
    Instruction *ToMove = I;
    bool Done = false;

    if (I != Preheader->begin()) {
      // Skip debug info intrinsics.
      do {
        --I;
      } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());

      if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
        Done = true;
    } else {
      Done = true;
    }

    ToMove->moveBefore(InsertPt);
    if (Done) break;
    InsertPt = ToMove;
  }
}

//===----------------------------------------------------------------------===//
//  IndVarSimplify driver. Manage several subpasses of IV simplification.
//===----------------------------------------------------------------------===//

bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
  // If LoopSimplify form is not available, stay out of trouble. Some notes:
  //  - LSR currently only supports LoopSimplify-form loops. Indvars'
  //    canonicalization can be a pessimization without LSR to "clean up"
  //    afterwards.
  //  - We depend on having a preheader; in particular,
  //    Loop::getCanonicalInductionVariable only supports loops with preheaders,
  //    and we're in trouble if we can't find the induction variable even when
  //    we've manually inserted one.
  if (!L->isLoopSimplifyForm())
    return false;

  if (EnableIVRewrite)
    IU = &getAnalysis<IVUsers>();
  LI = &getAnalysis<LoopInfo>();
  SE = &getAnalysis<ScalarEvolution>();
  DT = &getAnalysis<DominatorTree>();
  TD = getAnalysisIfAvailable<TargetData>();

  DeadInsts.clear();
  Changed = false;

  // If there are any floating-point recurrences, attempt to
  // transform them to use integer recurrences.
  RewriteNonIntegerIVs(L);

  const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);

  // Create a rewriter object which we'll use to transform the code with.
  SCEVExpander Rewriter(*SE, "indvars");
#ifndef NDEBUG
  Rewriter.setDebugType(DEBUG_TYPE);
#endif

  // Eliminate redundant IV users.
  //
  // Simplification works best when run before other consumers of SCEV. We
  // attempt to avoid evaluating SCEVs for sign/zero extend operations until
  // other expressions involving loop IVs have been evaluated. This helps SCEV
  // set no-wrap flags before normalizing sign/zero extension.
  if (!EnableIVRewrite) {
    Rewriter.disableCanonicalMode();
    SimplifyAndExtend(L, Rewriter, LPM);
  }

  // Check to see if this loop has a computable loop-invariant execution count.
  // If so, this means that we can compute the final value of any expressions
  // that are recurrent in the loop, and substitute the exit values from the
  // loop into any instructions outside of the loop that use the final values of
  // the current expressions.
  //
  if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
    RewriteLoopExitValues(L, Rewriter);

  // Eliminate redundant IV users.
  if (EnableIVRewrite)
    Changed |= simplifyIVUsers(IU, SE, &LPM, DeadInsts);

  // Eliminate redundant IV cycles.
  if (!EnableIVRewrite)
    NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);

  // Compute the type of the largest recurrence expression, and decide whether
  // a canonical induction variable should be inserted.
  Type *LargestType = 0;
  bool NeedCannIV = false;
  bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
  if (EnableIVRewrite && ExpandBECount) {
    // If we have a known trip count and a single exit block, we'll be
    // rewriting the loop exit test condition below, which requires a
    // canonical induction variable.
    NeedCannIV = true;
    Type *Ty = BackedgeTakenCount->getType();
    if (!EnableIVRewrite) {
      // In this mode, SimplifyIVUsers may have already widened the IV used by
      // the backedge test and inserted a Trunc on the compare's operand. Get
      // the wider type to avoid creating a redundant narrow IV only used by the
      // loop test.
      LargestType = getBackedgeIVType(L);
    }
    if (!LargestType ||
        SE->getTypeSizeInBits(Ty) >
        SE->getTypeSizeInBits(LargestType))
      LargestType = SE->getEffectiveSCEVType(Ty);
  }
  if (EnableIVRewrite) {
    for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
      NeedCannIV = true;
      Type *Ty =
        SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
      if (!LargestType ||
          SE->getTypeSizeInBits(Ty) >
          SE->getTypeSizeInBits(LargestType))
        LargestType = Ty;
    }
  }

  // Now that we know the largest of the induction variable expressions
  // in this loop, insert a canonical induction variable of the largest size.
  PHINode *IndVar = 0;
  if (NeedCannIV) {
    // Check to see if the loop already has any canonical-looking induction
    // variables. If any are present and wider than the planned canonical
    // induction variable, temporarily remove them, so that the Rewriter
    // doesn't attempt to reuse them.
    SmallVector<PHINode *, 2> OldCannIVs;
    while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
      if (SE->getTypeSizeInBits(OldCannIV->getType()) >
          SE->getTypeSizeInBits(LargestType))
        OldCannIV->removeFromParent();
      else
        break;
      OldCannIVs.push_back(OldCannIV);
    }

    IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);

    ++NumInserted;
    Changed = true;
    DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');

    // Now that the official induction variable is established, reinsert
    // any old canonical-looking variables after it so that the IR remains
    // consistent. They will be deleted as part of the dead-PHI deletion at
    // the end of the pass.
    while (!OldCannIVs.empty()) {
      PHINode *OldCannIV = OldCannIVs.pop_back_val();
      OldCannIV->insertBefore(L->getHeader()->getFirstInsertionPt());
    }
  }
  else if (!EnableIVRewrite && ExpandBECount && needsLFTR(L, DT)) {
    IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
  }
  // If we have a trip count expression, rewrite the loop's exit condition
  // using it.  We can currently only handle loops with a single exit.
  Value *NewICmp = 0;
  if (ExpandBECount && IndVar) {
    // Check preconditions for proper SCEVExpander operation. SCEV does not
    // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
    // pass that uses the SCEVExpander must do it. This does not work well for
    // loop passes because SCEVExpander makes assumptions about all loops, while
    // LoopPassManager only forces the current loop to be simplified.
    //
    // FIXME: SCEV expansion has no way to bail out, so the caller must
    // explicitly check any assumptions made by SCEV. Brittle.
    const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
    if (!AR || AR->getLoop()->getLoopPreheader())
      NewICmp =
        LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter);
  }
  // Rewrite IV-derived expressions.
  if (EnableIVRewrite)
    RewriteIVExpressions(L, Rewriter);

  // Clear the rewriter cache, because values that are in the rewriter's cache
  // can be deleted in the loop below, causing the AssertingVH in the cache to
  // trigger.
  Rewriter.clear();

  // Now that we're done iterating through lists, clean up any instructions
  // which are now dead.
  while (!DeadInsts.empty())
    if (Instruction *Inst =
          dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
      RecursivelyDeleteTriviallyDeadInstructions(Inst);

  // The Rewriter may not be used from this point on.

  // Loop-invariant instructions in the preheader that aren't used in the
  // loop may be sunk below the loop to reduce register pressure.
  SinkUnusedInvariants(L);

  // For completeness, inform IVUsers of the IV use in the newly-created
  // loop exit test instruction.
  if (IU && NewICmp) {
    ICmpInst *NewICmpInst = dyn_cast<ICmpInst>(NewICmp);
    if (NewICmpInst)
      IU->AddUsersIfInteresting(cast<Instruction>(NewICmpInst->getOperand(0)));
  }
  // Clean up dead instructions.
  Changed |= DeleteDeadPHIs(L->getHeader());
  // Check a post-condition.
  assert(L->isLCSSAForm(*DT) &&
         "Indvars did not leave the loop in lcssa form!");

  // Verify that LFTR, and any other change have not interfered with SCEV's
  // ability to compute trip count.
#ifndef NDEBUG
  if (!EnableIVRewrite && VerifyIndvars &&
      !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
    SE->forgetLoop(L);
    const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
    if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
        SE->getTypeSizeInBits(NewBECount->getType()))
      NewBECount = SE->getTruncateOrNoop(NewBECount,
                                         BackedgeTakenCount->getType());
    else
      BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
                                                 NewBECount->getType());
    assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");
  }
#endif

  return Changed;
}