LoopStrengthReduce.cpp

  // Otherwise use the value itself.
  Ops.push_back(C ? SE.getMulExpr(C, S) : S);
}

/// GenerateReassociations - Split out subexpressions from adds and the bases of
/// addrecs.
void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
                                         Formula Base,
                                         unsigned Depth) {
  // Arbitrarily cap recursion to protect compile time.
  if (Depth >= 3) return;

  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
    const SCEV *BaseReg = Base.BaseRegs[i];

    SmallVector<const SCEV *, 8> AddOps;
    CollectSubexprs(BaseReg, 0, AddOps, SE);
    if (AddOps.size() == 1) continue;

    for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
         JE = AddOps.end(); J != JE; ++J) {
      // Don't pull a constant into a register if the constant could be folded
      // into an immediate field.
      if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
                           Base.getNumRegs() > 1,
                           LU.Kind, LU.AccessTy, TLI, SE))
        continue;

      // Collect all operands except *J.
      SmallVector<const SCEV *, 8> InnerAddOps;
      for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
           KE = AddOps.end(); K != KE; ++K)
        if (K != J)
          InnerAddOps.push_back(*K);

      // Don't leave just a constant behind in a register if the constant could
      // be folded into an immediate field.
      if (InnerAddOps.size() == 1 &&
          isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
                           Base.getNumRegs() > 1,
                           LU.Kind, LU.AccessTy, TLI, SE))
        continue;

      Formula F = Base;
      F.BaseRegs[i] = SE.getAddExpr(InnerAddOps);
      F.BaseRegs.push_back(*J);
      if (InsertFormula(LU, LUIdx, F))
        // If that formula hadn't been seen before, recurse to find more like
        // it.
        GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
    }
  }
}

/// GenerateCombinations - Generate a formula consisting of all of the
/// loop-dominating registers added into a single register.
void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
                                       Formula Base) {
  // This method is only interesting on a plurality of registers.
  if (Base.BaseRegs.size() <= 1) return;

  Formula F = Base;
  F.BaseRegs.clear();
  SmallVector<const SCEV *, 4> Ops;
  for (SmallVectorImpl<const SCEV *>::const_iterator
       I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
    const SCEV *BaseReg = *I;
    if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
        !BaseReg->hasComputableLoopEvolution(L))
      Ops.push_back(BaseReg);
    else
      F.BaseRegs.push_back(BaseReg);
  }
  if (Ops.size() > 1) {
    const SCEV *Sum = SE.getAddExpr(Ops);
    // TODO: If Sum is zero, it probably means ScalarEvolution missed an
    // opportunity to fold something. For now, just ignore such cases
    // rather than proceed with zero in a register.
    if (!Sum->isZero()) {
      F.BaseRegs.push_back(Sum);
      (void)InsertFormula(LU, LUIdx, F);
    }
  }
}

/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
                                          Formula Base) {
  // We can't add a symbolic offset if the address already contains one.
  if (Base.AM.BaseGV) return;

  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
    const SCEV *G = Base.BaseRegs[i];
    GlobalValue *GV = ExtractSymbol(G, SE);
    if (G->isZero() || !GV)
      continue;
    Formula F = Base;
    F.AM.BaseGV = GV;
    if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
                    LU.Kind, LU.AccessTy, TLI))
      continue;
    F.BaseRegs[i] = G;
    (void)InsertFormula(LU, LUIdx, F);
  }
}

/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
                                          Formula Base) {
  // TODO: For now, just add the min and max offset, because it usually isn't
  // worthwhile looking at everything inbetween.
  SmallVector<int64_t, 4> Worklist;
  Worklist.push_back(LU.MinOffset);
  if (LU.MaxOffset != LU.MinOffset)
    Worklist.push_back(LU.MaxOffset);

  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
    const SCEV *G = Base.BaseRegs[i];

    for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
         E = Worklist.end(); I != E; ++I) {
      Formula F = Base;
      F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
      if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
                     LU.Kind, LU.AccessTy, TLI)) {
        F.BaseRegs[i] = SE.getAddExpr(G, SE.getIntegerSCEV(*I, G->getType()));

        (void)InsertFormula(LU, LUIdx, F);
      }
    }

    int64_t Imm = ExtractImmediate(G, SE);
    if (G->isZero() || Imm == 0)
      continue;
    Formula F = Base;
    F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
    if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
                    LU.Kind, LU.AccessTy, TLI))
      continue;
    F.BaseRegs[i] = G;
    (void)InsertFormula(LU, LUIdx, F);
  }
}

/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
/// the comparison. For example, x == y -> x*c == y*c.
void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
                                         Formula Base) {
  if (LU.Kind != LSRUse::ICmpZero) return;

  // Determine the integer type for the base formula.
  const Type *IntTy = Base.getType();
  if (!IntTy) return;
  if (SE.getTypeSizeInBits(IntTy) > 64) return;

  // Don't do this if there is more than one offset.
  if (LU.MinOffset != LU.MaxOffset) return;

  assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");

  // Check each interesting stride.
  for (SmallSetVector<int64_t, 8>::const_iterator
       I = Factors.begin(), E = Factors.end(); I != E; ++I) {
    int64_t Factor = *I;
    Formula F = Base;

    // Check that the multiplication doesn't overflow.
    if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
      continue;
    F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
    if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
      continue;

    // Check that multiplying with the use offset doesn't overflow.
    int64_t Offset = LU.MinOffset;
    if (Offset == INT64_MIN && Factor == -1)
      continue;
    Offset = (uint64_t)Offset * Factor;
    if (Offset / Factor != LU.MinOffset)
      continue;

    // Check that this scale is legal.
    if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
      continue;

    // Compensate for the use having MinOffset built into it.
    F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;

    const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);

    // Check that multiplying with each base register doesn't overflow.
    for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
      F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
      if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
        goto next;
    }

    // Check that multiplying with the scaled register doesn't overflow.
    if (F.ScaledReg) {
      F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
      if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
        continue;
    }

    // If we make it here and it's legal, add it.
    (void)InsertFormula(LU, LUIdx, F);
  next:;
  }
}

/// GenerateScales - Generate stride factor reuse formulae by making use of
/// scaled-offset address modes, for example.
void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
                                 Formula Base) {
  // Determine the integer type for the base formula.
  const Type *IntTy = Base.getType();
  if (!IntTy) return;

  // If this Formula already has a scaled register, we can't add another one.
  if (Base.AM.Scale != 0) return;

  // Check each interesting stride.
  for (SmallSetVector<int64_t, 8>::const_iterator
       I = Factors.begin(), E = Factors.end(); I != E; ++I) {
    int64_t Factor = *I;

    Base.AM.Scale = Factor;
    Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
    // Check whether this scale is going to be legal.
    if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
                    LU.Kind, LU.AccessTy, TLI)) {
      // As a special-case, handle special out-of-loop Basic users specially.
      // TODO: Reconsider this special case.
      if (LU.Kind == LSRUse::Basic &&
          isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
                     LSRUse::Special, LU.AccessTy, TLI) &&
          LU.AllFixupsOutsideLoop)
        LU.Kind = LSRUse::Special;
      else
        continue;
    }
    // For an ICmpZero, negating a solitary base register won't lead to
    // new solutions.
    if (LU.Kind == LSRUse::ICmpZero &&
        !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
      continue;
    // For each addrec base reg, apply the scale, if possible.
    for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
      if (const SCEVAddRecExpr *AR =
            dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
        const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
        if (FactorS->isZero())
          continue;
        // Divide out the factor, ignoring high bits, since we'll be
        // scaling the value back up in the end.
        if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
          // TODO: This could be optimized to avoid all the copying.
          Formula F = Base;
          F.ScaledReg = Quotient;
          std::swap(F.BaseRegs[i], F.BaseRegs.back());
          F.BaseRegs.pop_back();
          (void)InsertFormula(LU, LUIdx, F);
        }
      }
  }
}

/// GenerateTruncates - Generate reuse formulae from different IV types.
void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
                                    Formula Base) {
  // This requires TargetLowering to tell us which truncates are free.
  if (!TLI) return;

  // Don't bother truncating symbolic values.
  if (Base.AM.BaseGV) return;

  // Determine the integer type for the base formula.
  const Type *DstTy = Base.getType();
  if (!DstTy) return;
  DstTy = SE.getEffectiveSCEVType(DstTy);

  for (SmallSetVector<const Type *, 4>::const_iterator
       I = Types.begin(), E = Types.end(); I != E; ++I) {
    const Type *SrcTy = *I;
    if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
      Formula F = Base;

      if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
      for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
           JE = F.BaseRegs.end(); J != JE; ++J)
        *J = SE.getAnyExtendExpr(*J, SrcTy);

      // TODO: This assumes we've done basic processing on all uses and
      // have an idea what the register usage is.
      if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
        continue;

      (void)InsertFormula(LU, LUIdx, F);
    }
  }
}

namespace {

/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
/// defer modifications so that the search phase doesn't have to worry about
/// the data structures moving underneath it.
struct WorkItem {
  size_t LUIdx;
  int64_t Imm;
  const SCEV *OrigReg;

  WorkItem(size_t LI, int64_t I, const SCEV *R)
    : LUIdx(LI), Imm(I), OrigReg(R) {}

  void print(raw_ostream &OS) const;
  void dump() const;
};

}

void WorkItem::print(raw_ostream &OS) const {
  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
     << " , add offset " << Imm;
}

void WorkItem::dump() const {
  print(errs()); errs() << '\n';
}

/// GenerateCrossUseConstantOffsets - Look for registers which are a constant
/// distance apart and try to form reuse opportunities between them.
void LSRInstance::GenerateCrossUseConstantOffsets() {
  // Group the registers by their value without any added constant offset.
  typedef std::map<int64_t, const SCEV *> ImmMapTy;
  typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
  RegMapTy Map;
  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
  SmallVector<const SCEV *, 8> Sequence;
  for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
       I != E; ++I) {
    const SCEV *Reg = *I;
    int64_t Imm = ExtractImmediate(Reg, SE);
    std::pair<RegMapTy::iterator, bool> Pair =
      Map.insert(std::make_pair(Reg, ImmMapTy()));
    if (Pair.second)
      Sequence.push_back(Reg);
    Pair.first->second.insert(std::make_pair(Imm, *I));
    UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
  }

  // Now examine each set of registers with the same base value. Build up
  // a list of work to do and do the work in a separate step so that we're
  // not adding formulae and register counts while we're searching.
  SmallVector<WorkItem, 32> WorkItems;
  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
  for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
       E = Sequence.end(); I != E; ++I) {
    const SCEV *Reg = *I;
    const ImmMapTy &Imms = Map.find(Reg)->second;

    // It's not worthwhile looking for reuse if there's only one offset.
    if (Imms.size() == 1)
      continue;

    DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
          for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
               J != JE; ++J)
            dbgs() << ' ' << J->first;
          dbgs() << '\n');

    // Examine each offset.
    for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
         J != JE; ++J) {
      const SCEV *OrigReg = J->second;

      int64_t JImm = J->first;
      const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);

      if (!isa<SCEVConstant>(OrigReg) &&
          UsedByIndicesMap[Reg].count() == 1) {
        DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
        continue;
      }

      // Conservatively examine offsets between this orig reg a few selected
      // other orig regs.
      ImmMapTy::const_iterator OtherImms[] = {
        Imms.begin(), prior(Imms.end()),
        Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
      };
      for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
        ImmMapTy::const_iterator M = OtherImms[i];
        if (M == J || M == JE) continue;

        // Compute the difference between the two.
        int64_t Imm = (uint64_t)JImm - M->first;
        for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
             LUIdx = UsedByIndices.find_next(LUIdx))
          // Make a memo of this use, offset, and register tuple.
          if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
            WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
      }
    }
  }

  Map.clear();
  Sequence.clear();
  UsedByIndicesMap.clear();
  UniqueItems.clear();

  // Now iterate through the worklist and add new formulae.
  for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
       E = WorkItems.end(); I != E; ++I) {
    const WorkItem &WI = *I;
    size_t LUIdx = WI.LUIdx;
    LSRUse &LU = Uses[LUIdx];
    int64_t Imm = WI.Imm;
    const SCEV *OrigReg = WI.OrigReg;

    const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
    const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
    unsigned BitWidth = SE.getTypeSizeInBits(IntTy);

    // TODO: Use a more targeted data structure.
    for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
      Formula F = LU.Formulae[L];
      // Use the immediate in the scaled register.
      if (F.ScaledReg == OrigReg) {
        int64_t Offs = (uint64_t)F.AM.BaseOffs +
                       Imm * (uint64_t)F.AM.Scale;
        // Don't create 50 + reg(-50).
        if (F.referencesReg(SE.getSCEV(
                   ConstantInt::get(IntTy, -(uint64_t)Offs))))
          continue;
        Formula NewF = F;
        NewF.AM.BaseOffs = Offs;
        if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
                        LU.Kind, LU.AccessTy, TLI))
          continue;
        NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);

        // If the new scale is a constant in a register, and adding the constant
        // value to the immediate would produce a value closer to zero than the
        // immediate itself, then the formula isn't worthwhile.
        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
          if (C->getValue()->getValue().isNegative() !=
                (NewF.AM.BaseOffs < 0) &&
              (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
                .ule(APInt(BitWidth, NewF.AM.BaseOffs).abs()))
            continue;

        // OK, looks good.
        (void)InsertFormula(LU, LUIdx, NewF);
      } else {
        // Use the immediate in a base register.
        for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
          const SCEV *BaseReg = F.BaseRegs[N];
          if (BaseReg != OrigReg)
            continue;
          Formula NewF = F;
          NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
          if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
                          LU.Kind, LU.AccessTy, TLI))
            continue;
          NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);

          // If the new formula has a constant in a register, and adding the
          // constant value to the immediate would produce a value closer to
          // zero than the immediate itself, then the formula isn't worthwhile.
          for (SmallVectorImpl<const SCEV *>::const_iterator
               J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
               J != JE; ++J)
            if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
              if (C->getValue()->getValue().isNegative() !=
                    (NewF.AM.BaseOffs < 0) &&
                  C->getValue()->getValue().abs()
                    .ule(APInt(BitWidth, NewF.AM.BaseOffs).abs()))
                goto skip_formula;

          // Ok, looks good.
          (void)InsertFormula(LU, LUIdx, NewF);
          break;
        skip_formula:;
        }
      }
    }
  }
}

/// GenerateAllReuseFormulae - Generate formulae for each use.
void
LSRInstance::GenerateAllReuseFormulae() {
  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
  // queries are more precise.
  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
    LSRUse &LU = Uses[LUIdx];
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
  }
  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
    LSRUse &LU = Uses[LUIdx];
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateScales(LU, LUIdx, LU.Formulae[i]);
  }
  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
    LSRUse &LU = Uses[LUIdx];
    for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
      GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
  }

  GenerateCrossUseConstantOffsets();
}

/// If their are multiple formulae with the same set of registers used
/// by other uses, pick the best one and delete the others.
void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
#ifndef NDEBUG
  bool Changed = false;
#endif

  // Collect the best formula for each unique set of shared registers. This
  // is reset for each use.
  typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
    BestFormulaeTy;
  BestFormulaeTy BestFormulae;

  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
    LSRUse &LU = Uses[LUIdx];
    FormulaSorter Sorter(L, LU, SE, DT);

    // Clear out the set of used regs; it will be recomputed.
    LU.Regs.clear();

    for (size_t FIdx = 0, NumForms = LU.Formulae.size();
         FIdx != NumForms; ++FIdx) {
      Formula &F = LU.Formulae[FIdx];

      SmallVector<const SCEV *, 2> Key;
      for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
           JE = F.BaseRegs.end(); J != JE; ++J) {
        const SCEV *Reg = *J;
        if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
          Key.push_back(Reg);
      }
      if (F.ScaledReg &&
          RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
        Key.push_back(F.ScaledReg);
      // Unstable sort by host order ok, because this is only used for
      // uniquifying.
      std::sort(Key.begin(), Key.end());

      std::pair<BestFormulaeTy::const_iterator, bool> P =
        BestFormulae.insert(std::make_pair(Key, FIdx));
      if (!P.second) {
        Formula &Best = LU.Formulae[P.first->second];
        if (Sorter.operator()(F, Best))
          std::swap(F, Best);
        DEBUG(dbgs() << "Filtering out "; F.print(dbgs());
              dbgs() << "\n"
                        "  in favor of "; Best.print(dbgs());
              dbgs() << '\n');
#ifndef NDEBUG
        Changed = true;
#endif
        std::swap(F, LU.Formulae.back());
        LU.Formulae.pop_back();
        --FIdx;
        --NumForms;
        continue;
      }
      if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
      LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
    }
    BestFormulae.clear();
  }

  DEBUG(if (Changed) {
          dbgs() << "\n"
                    "After filtering out undesirable candidates:\n";
          print_uses(dbgs());
        });
}

/// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
/// formulae to choose from, use some rough heuristics to prune down the number
/// of formulae. This keeps the main solver from taking an extraordinary amount
/// of time in some worst-case scenarios.
void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
  // This is a rough guess that seems to work fairly well.
  const size_t Limit = UINT16_MAX;

  SmallPtrSet<const SCEV *, 4> Taken;
  for (;;) {
    // Estimate the worst-case number of solutions we might consider. We almost
    // never consider this many solutions because we prune the search space,
    // but the pruning isn't always sufficient.
    uint32_t Power = 1;
    for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
         E = Uses.end(); I != E; ++I) {
      size_t FSize = I->Formulae.size();
      if (FSize >= Limit) {
        Power = Limit;
        break;
      }
      Power *= FSize;
      if (Power >= Limit)
        break;
    }
    if (Power < Limit)
      break;

    // Ok, we have too many of formulae on our hands to conveniently handle.
    // Use a rough heuristic to thin out the list.

    // Pick the register which is used by the most LSRUses, which is likely
    // to be a good reuse register candidate.
    const SCEV *Best = 0;
    unsigned BestNum = 0;
    for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
         I != E; ++I) {
      const SCEV *Reg = *I;
      if (Taken.count(Reg))
        continue;
      if (!Best)
        Best = Reg;
      else {
        unsigned Count = RegUses.getUsedByIndices(Reg).count();
        if (Count > BestNum) {
          Best = Reg;
          BestNum = Count;
        }
      }
    }

    DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
                 << " will yield profitable reuse.\n");
    Taken.insert(Best);

    // In any use with formulae which references this register, delete formulae
    // which don't reference it.
    for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(),
         E = Uses.end(); I != E; ++I) {
      LSRUse &LU = *I;
      if (!LU.Regs.count(Best)) continue;

      // Clear out the set of used regs; it will be recomputed.
      LU.Regs.clear();

      for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
        Formula &F = LU.Formulae[i];
        if (!F.referencesReg(Best)) {
          DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n');
          std::swap(LU.Formulae.back(), F);
          LU.Formulae.pop_back();
          --e;
          --i;
          continue;
        }

        if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
        LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
      }
    }

    DEBUG(dbgs() << "After pre-selection:\n";
          print_uses(dbgs()));
  }
}

/// SolveRecurse - This is the recursive solver.
void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
                               Cost &SolutionCost,
                               SmallVectorImpl<const Formula *> &Workspace,
                               const Cost &CurCost,
                               const SmallPtrSet<const SCEV *, 16> &CurRegs,
                               DenseSet<const SCEV *> &VisitedRegs) const {
  // Some ideas:
  //  - prune more:
  //    - use more aggressive filtering
  //    - sort the formula so that the most profitable solutions are found first
  //    - sort the uses too
  //  - search faster:
  //    - don't compute a cost, and then compare. compare while computing a cost
  //      and bail early.
  //    - track register sets with SmallBitVector

  const LSRUse &LU = Uses[Workspace.size()];

  // If this use references any register that's already a part of the
  // in-progress solution, consider it a requirement that a formula must
  // reference that register in order to be considered. This prunes out
  // unprofitable searching.
  SmallSetVector<const SCEV *, 4> ReqRegs;
  for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
       E = CurRegs.end(); I != E; ++I)
    if (LU.Regs.count(*I))
      ReqRegs.insert(*I);

  bool AnySatisfiedReqRegs = false;
  SmallPtrSet<const SCEV *, 16> NewRegs;
  Cost NewCost;
retry:
  for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
       E = LU.Formulae.end(); I != E; ++I) {
    const Formula &F = *I;

    // Ignore formulae which do not use any of the required registers.
    for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
         JE = ReqRegs.end(); J != JE; ++J) {
      const SCEV *Reg = *J;
      if ((!F.ScaledReg || F.ScaledReg != Reg) &&
          std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
          F.BaseRegs.end())
        goto skip;
    }
    AnySatisfiedReqRegs = true;

    // Evaluate the cost of the current formula. If it's already worse than
    // the current best, prune the search at that point.
    NewCost = CurCost;
    NewRegs = CurRegs;
    NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
    if (NewCost < SolutionCost) {
      Workspace.push_back(&F);
      if (Workspace.size() != Uses.size()) {
        SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
                     NewRegs, VisitedRegs);
        if (F.getNumRegs() == 1 && Workspace.size() == 1)
          VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
      } else {
        DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
              dbgs() << ". Regs:";
              for (SmallPtrSet<const SCEV *, 16>::const_iterator
                   I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
                dbgs() << ' ' << **I;
              dbgs() << '\n');

        SolutionCost = NewCost;
        Solution = Workspace;
      }
      Workspace.pop_back();
    }
  skip:;
  }

  // If none of the formulae had all of the required registers, relax the
  // constraint so that we don't exclude all formulae.
  if (!AnySatisfiedReqRegs) {
    ReqRegs.clear();
    goto retry;
  }
}

void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
  SmallVector<const Formula *, 8> Workspace;
  Cost SolutionCost;
  SolutionCost.Loose();
  Cost CurCost;
  SmallPtrSet<const SCEV *, 16> CurRegs;
  DenseSet<const SCEV *> VisitedRegs;
  Workspace.reserve(Uses.size());

  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
               CurRegs, VisitedRegs);

  // Ok, we've now made all our decisions.
  DEBUG(dbgs() << "\n"
                  "The chosen solution requires "; SolutionCost.print(dbgs());
        dbgs() << ":\n";
        for (size_t i = 0, e = Uses.size(); i != e; ++i) {
          dbgs() << "  ";
          Uses[i].print(dbgs());
          dbgs() << "\n"
                    "    ";
          Solution[i]->print(dbgs());
          dbgs() << '\n';
        });
}

/// getImmediateDominator - A handy utility for the specific DominatorTree
/// query that we need here.
///
static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
  DomTreeNode *Node = DT.getNode(BB);
  if (!Node) return 0;
  Node = Node->getIDom();
  if (!Node) return 0;
  return Node->getBlock();
}

Value *LSRInstance::Expand(const LSRFixup &LF,
                           const Formula &F,
                           BasicBlock::iterator IP,
                           SCEVExpander &Rewriter,
                           SmallVectorImpl<WeakVH> &DeadInsts) const {
  const LSRUse &LU = Uses[LF.LUIdx];

  // Then, collect some instructions which must be dominated by the
  // expanding replacement. These must be dominated by any operands that
  // will be required in the expansion.
  SmallVector<Instruction *, 4> Inputs;
  if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
    Inputs.push_back(I);
  if (LU.Kind == LSRUse::ICmpZero)
    if (Instruction *I =
          dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
      Inputs.push_back(I);
  if (LF.PostIncLoops.count(L)) {
    if (LF.isUseFullyOutsideLoop(L))
      Inputs.push_back(L->getLoopLatch()->getTerminator());
    else
      Inputs.push_back(IVIncInsertPos);
  }
  // The expansion must also be dominated by the increment positions of any
  // loops it for which it is using post-inc mode.
  for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
       E = LF.PostIncLoops.end(); I != E; ++I) {
    const Loop *PIL = *I;
    if (PIL == L) continue;

    SmallVector<BasicBlock *, 4> ExitingBlocks;
    PIL->getExitingBlocks(ExitingBlocks);
    if (!ExitingBlocks.empty()) {
      BasicBlock *BB = ExitingBlocks[0];
      for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
        BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
      Inputs.push_back(BB->getTerminator());
    }
  }

  // Then, climb up the immediate dominator tree as far as we can go while
  // still being dominated by the input positions.
  for (;;) {
    bool AllDominate = true;
    Instruction *BetterPos = 0;
    BasicBlock *IDom = getImmediateDominator(IP->getParent(), DT);
    if (!IDom) break;
    Instruction *Tentative = IDom->getTerminator();
    for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
         E = Inputs.end(); I != E; ++I) {
      Instruction *Inst = *I;
      if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
        AllDominate = false;
        break;
      }
      if (IDom == Inst->getParent() &&
          (!BetterPos || DT.dominates(BetterPos, Inst)))
        BetterPos = next(BasicBlock::iterator(Inst));
    }
    if (!AllDominate)
      break;
    if (BetterPos)
      IP = BetterPos;
    else
      IP = Tentative;
  }
  while (isa<PHINode>(IP)) ++IP;
  while (isa<DbgInfoIntrinsic>(IP)) ++IP;

  // Inform the Rewriter if we have a post-increment use, so that it can
  // perform an advantageous expansion.
  Rewriter.setPostInc(LF.PostIncLoops);

  // This is the type that the user actually needs.
  const Type *OpTy = LF.OperandValToReplace->getType();
  // This will be the type that we'll initially expand to.
  const Type *Ty = F.getType();
  if (!Ty)
    // No type known; just expand directly to the ultimate type.
    Ty = OpTy;
  else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
    // Expand directly to the ultimate type if it's the right size.
    Ty = OpTy;
  // This is the type to do integer arithmetic in.
  const Type *IntTy = SE.getEffectiveSCEVType(Ty);

  // Build up a list of operands to add together to form the full base.
  SmallVector<const SCEV *, 8> Ops;

  // Expand the BaseRegs portion.
  for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
       E = F.BaseRegs.end(); I != E; ++I) {
    const SCEV *Reg = *I;
    assert(!Reg->isZero() && "Zero allocated in a base register!");

    // If we're expanding for a post-inc user, make the post-inc adjustment.
    PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
    Reg = TransformForPostIncUse(Denormalize, Reg,
                                 LF.UserInst, LF.OperandValToReplace,
                                 Loops, SE, DT);

    Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
  }

  // Flush the operand list to suppress SCEVExpander hoisting.
  if (!Ops.empty()) {
    Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
    Ops.clear();
    Ops.push_back(SE.getUnknown(FullV));
  }

  // Expand the ScaledReg portion.
  Value *ICmpScaledV = 0;
  if (F.AM.Scale != 0) {
    const SCEV *ScaledS = F.ScaledReg;

    // If we're expanding for a post-inc user, make the post-inc adjustment.
    PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
    ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
                                     LF.UserInst, LF.OperandValToReplace,
                                     Loops, SE, DT);

    if (LU.Kind == LSRUse::ICmpZero) {
      // An interesting way of "folding" with an icmp is to use a negated
      // scale, which we'll implement by inserting it into the other operand
      // of the icmp.
      assert(F.AM.Scale == -1 &&
             "The only scale supported by ICmpZero uses is -1!");
      ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
    } else {
      // Otherwise just expand the scaled register and an explicit scale,
      // which is expected to be matched as part of the address.
      ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
      ScaledS = SE.getMulExpr(ScaledS,
                              SE.getIntegerSCEV(F.AM.Scale,
                                                ScaledS->getType()));
      Ops.push_back(ScaledS);

      // Flush the operand list to suppress SCEVExpander hoisting.
      Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
      Ops.clear();
      Ops.push_back(SE.getUnknown(FullV));
    }
  }

  // Expand the GV portion.
  if (F.AM.BaseGV) {
    Ops.push_back(SE.getUnknown(F.AM.BaseGV));

    // Flush the operand list to suppress SCEVExpander hoisting.
    Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
    Ops.clear();
    Ops.push_back(SE.getUnknown(FullV));
  }

  // Expand the immediate portion.
  int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
  if (Offset != 0) {
    if (LU.Kind == LSRUse::ICmpZero) {
      // The other interesting way of "folding" with an ICmpZero is to use a
      // negated immediate.
      if (!ICmpScaledV)
        ICmpScaledV = ConstantInt::get(IntTy, -Offset);
      else {
        Ops.push_back(SE.getUnknown(ICmpScaledV));
        ICmpScaledV = ConstantInt::get(IntTy, Offset);
      }
    } else {
      // Just add the immediate values. These again are expected to be matched
      // as part of the address.
      Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
    }
  }

  // Emit instructions summing all the operands.
  const SCEV *FullS = Ops.empty() ?
                      SE.getIntegerSCEV(0, IntTy) :
                      SE.getAddExpr(Ops);
  Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);

  // We're done expanding now, so reset the rewriter.
  Rewriter.clearPostInc();

  // An ICmpZero Formula represents an ICmp which we're handling as a
  // comparison against zero. Now that we've expanded an expression for that
  // form, update the ICmp's other operand.
  if (LU.Kind == LSRUse::ICmpZero) {
    ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
    DeadInsts.push_back(CI->getOperand(1));
    assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
                           "a scale at the same time!");
    if (F.AM.Scale == -1) {
      if (ICmpScaledV->getType() != OpTy) {
        Instruction *Cast =
          CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
                                                   OpTy, false),
                           ICmpScaledV, OpTy, "tmp", CI);
        ICmpScaledV = Cast;
      }
      CI->setOperand(1, ICmpScaledV);
    } else {