class RegUseTracker {
typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
- RegUsesTy RegUses;
+ RegUsesTy RegUsesMap;
SmallVector<const SCEV *, 16> RegSequence;
public:
void CountRegister(const SCEV *Reg, size_t LUIdx);
+ void DropRegister(const SCEV *Reg, size_t LUIdx);
+ void DropUse(size_t LUIdx);
bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
void
RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
std::pair<RegUsesTy::iterator, bool> Pair =
- RegUses.insert(std::make_pair(Reg, RegSortData()));
+ RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
RegSortData &RSD = Pair.first->second;
if (Pair.second)
RegSequence.push_back(Reg);
RSD.UsedByIndices.set(LUIdx);
}
+void
+RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
+ RegUsesTy::iterator It = RegUsesMap.find(Reg);
+ assert(It != RegUsesMap.end());
+ RegSortData &RSD = It->second;
+ assert(RSD.UsedByIndices.size() > LUIdx);
+ RSD.UsedByIndices.reset(LUIdx);
+}
+
+void
+RegUseTracker::DropUse(size_t LUIdx) {
+ // Remove the use index from every register's use list.
+ for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
+ I != E; ++I)
+ I->second.UsedByIndices.reset(LUIdx);
+}
+
bool
RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
- if (!RegUses.count(Reg)) return false;
+ if (!RegUsesMap.count(Reg)) return false;
const SmallBitVector &UsedByIndices =
- RegUses.find(Reg)->second.UsedByIndices;
+ RegUsesMap.find(Reg)->second.UsedByIndices;
int i = UsedByIndices.find_first();
if (i == -1) return false;
if ((size_t)i != LUIdx) return true;
}
const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
- RegUsesTy::const_iterator I = RegUses.find(Reg);
- assert(I != RegUses.end() && "Unknown register!");
+ RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
+ assert(I != RegUsesMap.end() && "Unknown register!");
return I->second.UsedByIndices;
}
void RegUseTracker::clear() {
- RegUses.clear();
+ RegUsesMap.clear();
RegSequence.clear();
}
unsigned getNumRegs() const;
const Type *getType() const;
+ void DeleteBaseReg(const SCEV *&S);
+
bool referencesReg(const SCEV *S) const;
bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
const RegUseTracker &RegUses) const;
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
if (!AR->getStart()->isZero()) {
DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
- DoInitialMatch(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
+ DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
AR->getStepRecurrence(SE),
AR->getLoop()),
L, Good, Bad, SE, DT);
0;
}
+/// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
+void Formula::DeleteBaseReg(const SCEV *&S) {
+ if (&S != &BaseRegs.back())
+ std::swap(S, BaseRegs.back());
+ BaseRegs.pop_back();
+}
+
/// referencesReg - Test if this formula references the given register.
bool Formula::referencesReg(const SCEV *S) const {
return S == ScaledReg ||
if (!First) OS << " + "; else First = false;
OS << "reg(" << **I << ')';
}
+ if (AM.HasBaseReg && BaseRegs.empty()) {
+ if (!First) OS << " + "; else First = false;
+ OS << "**error: HasBaseReg**";
+ } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
+ if (!First) OS << " + "; else First = false;
+ OS << "**error: !HasBaseReg**";
+ }
if (AM.Scale != 0) {
if (!First) OS << " + "; else First = false;
OS << AM.Scale << "*reg(";
/// without changing its value.
static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
const Type *WideTy =
- IntegerType::get(SE.getContext(),
- SE.getTypeSizeInBits(AR->getType()) + 1);
+ IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
}
/// without changing its value.
static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
const Type *WideTy =
- IntegerType::get(SE.getContext(),
- SE.getTypeSizeInBits(A->getType()) + 1);
+ IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
}
-/// isMulSExtable - Return true if the given add can be sign-extended
+/// isMulSExtable - Return true if the given mul can be sign-extended
/// without changing its value.
-static bool isMulSExtable(const SCEVMulExpr *A, ScalarEvolution &SE) {
+static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
const Type *WideTy =
IntegerType::get(SE.getContext(),
- SE.getTypeSizeInBits(A->getType()) + 1);
- return isa<SCEVMulExpr>(SE.getSignExtendExpr(A, WideTy));
+ SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
+ return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
}
/// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
bool IgnoreSignificantBits = false) {
// Handle the trivial case, which works for any SCEV type.
if (LHS == RHS)
- return SE.getIntegerSCEV(1, LHS->getType());
+ return SE.getConstant(LHS->getType(), 1);
- // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
- // folding.
- if (RHS->isAllOnesValue())
- return SE.getMulExpr(LHS, RHS);
+ // Handle a few RHS special cases.
+ const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
+ if (RC) {
+ const APInt &RA = RC->getValue()->getValue();
+ // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
+ // some folding.
+ if (RA.isAllOnesValue())
+ return SE.getMulExpr(LHS, RC);
+ // Handle x /s 1 as x.
+ if (RA == 1)
+ return LHS;
+ }
// Check for a division of a constant by a constant.
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
- const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
if (!RC)
return 0;
- if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
+ const APInt &LA = C->getValue()->getValue();
+ const APInt &RA = RC->getValue()->getValue();
+ if (LA.srem(RA) != 0)
return 0;
- return SE.getConstant(C->getValue()->getValue()
- .sdiv(RC->getValue()->getValue()));
+ return SE.getConstant(LA.sdiv(RA));
}
// Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
if (!Step) return 0;
return SE.getAddRecExpr(Start, Step, AR->getLoop());
}
+ return 0;
}
// Distribute the sdiv over add operands, if the add doesn't overflow.
}
return SE.getAddExpr(Ops);
}
+ return 0;
}
// Check for a multiply operand that we can pull RHS out of.
- if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
+ if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
SmallVector<const SCEV *, 4> Ops;
bool Found = false;
for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
I != E; ++I) {
+ const SCEV *S = *I;
if (!Found)
- if (const SCEV *Q = getExactSDiv(*I, RHS, SE,
+ if (const SCEV *Q = getExactSDiv(S, RHS, SE,
IgnoreSignificantBits)) {
- Ops.push_back(Q);
+ S = Q;
Found = true;
- continue;
}
- Ops.push_back(*I);
+ Ops.push_back(S);
}
return Found ? SE.getMulExpr(Ops) : 0;
}
+ return 0;
+ }
// Otherwise we don't know.
return 0;
static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
if (C->getValue()->getValue().getMinSignedBits() <= 64) {
- S = SE.getIntegerSCEV(0, C->getType());
+ S = SE.getConstant(C->getType(), 0);
return C->getValue()->getSExtValue();
}
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
- S = SE.getIntegerSCEV(0, GV->getType());
+ S = SE.getConstant(GV->getType(), 0);
return GV;
}
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
case Intrinsic::x86_sse2_storeu_pd:
case Intrinsic::x86_sse2_storeu_dq:
case Intrinsic::x86_sse2_storel_dq:
- if (II->getOperand(1) == OperandVal)
+ if (II->getArgOperand(0) == OperandVal)
isAddress = true;
break;
}
case Intrinsic::x86_sse2_storeu_pd:
case Intrinsic::x86_sse2_storeu_dq:
case Intrinsic::x86_sse2_storel_dq:
- AccessTy = II->getOperand(1)->getType();
+ AccessTy = II->getArgOperand(0)->getType();
break;
}
}
}
LSRFixup::LSRFixup()
- : UserInst(0), OperandValToReplace(0),
- LUIdx(~size_t(0)), Offset(0) {}
+ : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
/// isUseFullyOutsideLoop - Test whether this fixup always uses its
/// value outside of the given loop.
MaxOffset(INT64_MIN),
AllFixupsOutsideLoop(true) {}
+ bool HasFormulaWithSameRegs(const Formula &F) const;
bool InsertFormula(const Formula &F);
+ void DeleteFormula(Formula &F);
+ void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
void check() const;
void dump() const;
};
+}
+
+/// HasFormula - Test whether this use as a formula which has the same
+/// registers as the given formula.
+bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
+ SmallVector<const SCEV *, 2> Key = F.BaseRegs;
+ if (F.ScaledReg) Key.push_back(F.ScaledReg);
+ // Unstable sort by host order ok, because this is only used for uniquifying.
+ std::sort(Key.begin(), Key.end());
+ return Uniquifier.count(Key);
+}
+
/// InsertFormula - If the given formula has not yet been inserted, add it to
/// the list, and return true. Return false otherwise.
bool LSRUse::InsertFormula(const Formula &F) {
return true;
}
+/// DeleteFormula - Remove the given formula from this use's list.
+void LSRUse::DeleteFormula(Formula &F) {
+ if (&F != &Formulae.back())
+ std::swap(F, Formulae.back());
+ Formulae.pop_back();
+ assert(!Formulae.empty() && "LSRUse has no formulae left!");
+}
+
+/// RecomputeRegs - Recompute the Regs field, and update RegUses.
+void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
+ // Now that we've filtered out some formulae, recompute the Regs set.
+ SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
+ Regs.clear();
+ for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
+ E = Formulae.end(); I != E; ++I) {
+ const Formula &F = *I;
+ if (F.ScaledReg) Regs.insert(F.ScaledReg);
+ Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
+ }
+
+ // Update the RegTracker.
+ for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
+ E = OldRegs.end(); I != E; ++I)
+ if (!Regs.count(*I))
+ RegUses.DropRegister(*I, LUIdx);
+}
+
void LSRUse::print(raw_ostream &OS) const {
OS << "LSR Use: Kind=";
switch (Kind) {
AM.HasBaseReg = HasBaseReg;
AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
+ // Canonicalize a scale of 1 to a base register if the formula doesn't
+ // already have a base register.
+ if (!AM.HasBaseReg && AM.Scale == 1) {
+ AM.Scale = 0;
+ AM.HasBaseReg = true;
+ }
+
return isLegalUse(AM, Kind, AccessTy, TLI);
}
return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
}
+namespace {
+
+/// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
+/// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
+struct UseMapDenseMapInfo {
+ static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
+ return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
+ }
+
+ static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
+ return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
+ }
+
+ static unsigned
+ getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
+ unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
+ Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
+ return Result;
+ }
+
+ static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
+ const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
+ return LHS == RHS;
+ }
+};
+
/// FormulaSorter - This class implements an ordering for formulae which sorts
/// the by their standalone cost.
class FormulaSorter {
void OptimizeShadowIV();
bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
- bool OptimizeLoopTermCond();
+ void OptimizeLoopTermCond();
void CollectInterestingTypesAndFactors();
void CollectFixupsAndInitialFormulae();
}
// Support for sharing of LSRUses between LSRFixups.
- typedef DenseMap<const SCEV *, size_t> UseMapTy;
+ typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
+ size_t,
+ UseMapDenseMapInfo> UseMapTy;
UseMapTy UseMap;
- bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
+ bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
LSRUse::KindType Kind, const Type *AccessTy);
std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
LSRUse::KindType Kind,
const Type *AccessTy);
+ void DeleteUse(LSRUse &LU);
+
+ LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
+
public:
void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
void GenerateAllReuseFormulae();
void FilterOutUndesirableDedicatedRegisters();
+
+ size_t EstimateSearchSpaceComplexity() const;
void NarrowSearchSpaceUsingHeuristics();
void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
/* Remove cast operation */
ShadowUse->replaceAllUsesWith(NewPH);
ShadowUse->eraseFromParent();
+ Changed = true;
break;
}
}
/// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
/// set the IV user and stride information and return true, otherwise return
/// false.
-bool LSRInstance::FindIVUserForCond(ICmpInst *Cond,
- IVStrideUse *&CondUse) {
+bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
if (UI->getUser() == Cond) {
// NOTE: we could handle setcc instructions with multiple uses here, but
const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
return Cond;
- const SCEV *One = SE.getIntegerSCEV(1, BackedgeTakenCount->getType());
+ const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
// Add one to the backedge-taken count to get the trip count.
const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
-
- // Check for a max calculation that matches the pattern.
- if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
+ if (IterationCount != SE.getSCEV(Sel)) return Cond;
+
+ // Check for a max calculation that matches the pattern. There's no check
+ // for ICMP_ULE here because the comparison would be with zero, which
+ // isn't interesting.
+ CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
+ const SCEVNAryExpr *Max = 0;
+ if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
+ Pred = ICmpInst::ICMP_SLE;
+ Max = S;
+ } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
+ Pred = ICmpInst::ICMP_SLT;
+ Max = S;
+ } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
+ Pred = ICmpInst::ICMP_ULT;
+ Max = U;
+ } else {
+ // No match; bail.
return Cond;
- const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
- if (Max != SE.getSCEV(Sel)) return Cond;
+ }
// To handle a max with more than two operands, this optimization would
// require additional checking and setup.
const SCEV *MaxLHS = Max->getOperand(0);
const SCEV *MaxRHS = Max->getOperand(1);
- if (!MaxLHS || MaxLHS != One) return Cond;
+
+ // ScalarEvolution canonicalizes constants to the left. For < and >, look
+ // for a comparison with 1. For <= and >=, a comparison with zero.
+ if (!MaxLHS ||
+ (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
+ return Cond;
+
// Check the relevant induction variable for conformance to
// the pattern.
const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
// Check the right operand of the select, and remember it, as it will
// be used in the new comparison instruction.
Value *NewRHS = 0;
- if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
+ if (ICmpInst::isTrueWhenEqual(Pred)) {
+ // Look for n+1, and grab n.
+ if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
+ if (isa<ConstantInt>(BO->getOperand(1)) &&
+ cast<ConstantInt>(BO->getOperand(1))->isOne() &&
+ SE.getSCEV(BO->getOperand(0)) == MaxRHS)
+ NewRHS = BO->getOperand(0);
+ if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
+ if (isa<ConstantInt>(BO->getOperand(1)) &&
+ cast<ConstantInt>(BO->getOperand(1))->isOne() &&
+ SE.getSCEV(BO->getOperand(0)) == MaxRHS)
+ NewRHS = BO->getOperand(0);
+ if (!NewRHS)
+ return Cond;
+ } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
NewRHS = Sel->getOperand(1);
else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
NewRHS = Sel->getOperand(2);
- if (!NewRHS) return Cond;
+ else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
+ NewRHS = SU->getValue();
+ else
+ // Max doesn't match expected pattern.
+ return Cond;
// Determine the new comparison opcode. It may be signed or unsigned,
// and the original comparison may be either equality or inequality.
- CmpInst::Predicate Pred =
- isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
if (Cond->getPredicate() == CmpInst::ICMP_EQ)
Pred = CmpInst::getInversePredicate(Pred);
/// OptimizeLoopTermCond - Change loop terminating condition to use the
/// postinc iv when possible.
-bool
+void
LSRInstance::OptimizeLoopTermCond() {
SmallPtrSet<Instruction *, 4> PostIncs;
}
if (const SCEVConstant *D =
dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
+ const ConstantInt *C = D->getValue();
// Stride of one or negative one can have reuse with non-addresses.
- if (D->getValue()->isOne() ||
- D->getValue()->isAllOnesValue())
+ if (C->isOne() || C->isAllOnesValue())
goto decline_post_inc;
// Avoid weird situations.
- if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
- D->getValue()->getValue().isMinSignedValue())
+ if (C->getValue().getMinSignedBits() >= 64 ||
+ C->getValue().isMinSignedValue())
goto decline_post_inc;
// Without TLI, assume that any stride might be valid, and so any
// use might be shared.
// Check for possible scaled-address reuse.
const Type *AccessTy = getAccessType(UI->getUser());
TargetLowering::AddrMode AM;
- AM.Scale = D->getValue()->getSExtValue();
+ AM.Scale = C->getSExtValue();
if (TLI->isLegalAddressingMode(AM, AccessTy))
goto decline_post_inc;
AM.Scale = -AM.Scale;
else if (BB != IVIncInsertPos->getParent())
IVIncInsertPos = BB->getTerminator();
}
-
- return Changed;
}
+/// reconcileNewOffset - Determine if the given use can accomodate a fixup
+/// at the given offset and other details. If so, update the use and
+/// return true.
bool
-LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
+LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
LSRUse::KindType Kind, const Type *AccessTy) {
int64_t NewMinOffset = LU.MinOffset;
int64_t NewMaxOffset = LU.MaxOffset;
return false;
// Conservatively assume HasBaseReg is true for now.
if (NewOffset < LU.MinOffset) {
- if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true,
+ if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
Kind, AccessTy, TLI))
return false;
NewMinOffset = NewOffset;
} else if (NewOffset > LU.MaxOffset) {
- if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true,
+ if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
Kind, AccessTy, TLI))
return false;
NewMaxOffset = NewOffset;
}
// Check for a mismatched access type, and fall back conservatively as needed.
+ // TODO: Be less conservative when the type is similar and can use the same
+ // addressing modes.
if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
NewAccessTy = Type::getVoidTy(AccessTy->getContext());
}
std::pair<UseMapTy::iterator, bool> P =
- UseMap.insert(std::make_pair(Expr, 0));
+ UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
if (!P.second) {
// A use already existed with this base.
size_t LUIdx = P.first->second;
LSRUse &LU = Uses[LUIdx];
- if (reconcileNewOffset(LU, Offset, Kind, AccessTy))
+ if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
// Reuse this use.
return std::make_pair(LUIdx, Offset);
}
return std::make_pair(LUIdx, Offset);
}
+/// DeleteUse - Delete the given use from the Uses list.
+void LSRInstance::DeleteUse(LSRUse &LU) {
+ if (&LU != &Uses.back())
+ std::swap(LU, Uses.back());
+ Uses.pop_back();
+}
+
+/// FindUseWithFormula - Look for a use distinct from OrigLU which is has
+/// a formula that has the same registers as the given formula.
+LSRUse *
+LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
+ const LSRUse &OrigLU) {
+ // Search all uses for the formula. This could be more clever. Ignore
+ // ICmpZero uses because they may contain formulae generated by
+ // GenerateICmpZeroScales, in which case adding fixup offsets may
+ // be invalid.
+ for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
+ LSRUse &LU = Uses[LUIdx];
+ if (&LU != &OrigLU &&
+ LU.Kind != LSRUse::ICmpZero &&
+ LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
+ LU.HasFormulaWithSameRegs(OrigF)) {
+ for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
+ E = LU.Formulae.end(); I != E; ++I) {
+ const Formula &F = *I;
+ if (F.BaseRegs == OrigF.BaseRegs &&
+ F.ScaledReg == OrigF.ScaledReg &&
+ F.AM.BaseGV == OrigF.AM.BaseGV &&
+ F.AM.Scale == OrigF.AM.Scale &&
+ LU.Kind) {
+ if (F.AM.BaseOffs == 0)
+ return &LU;
+ break;
+ }
+ }
+ }
+ }
+
+ return 0;
+}
+
void LSRInstance::CollectInterestingTypesAndFactors() {
SmallSetVector<const SCEV *, 4> Strides;
Strides.insert(AR->getStepRecurrence(SE));
Worklist.push_back(AR->getStart());
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
- Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
+ Worklist.append(Add->op_begin(), Add->op_end());
}
} while (!Worklist.empty());
}
if (NV == LF.OperandValToReplace) {
CI->setOperand(1, CI->getOperand(0));
CI->setOperand(0, NV);
+ NV = CI->getOperand(1);
+ Changed = true;
}
// x == y --> x - y == 0
DEBUG(print_fixups(dbgs()));
}
+/// InsertInitialFormula - Insert a formula for the given expression into
+/// the given use, separating out loop-variant portions from loop-invariant
+/// and loop-computable portions.
void
LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
Formula F;
assert(Inserted && "Initial formula already exists!"); (void)Inserted;
}
+/// InsertSupplementalFormula - Insert a simple single-register formula for
+/// the given expression into the given use.
void
LSRInstance::InsertSupplementalFormula(const SCEV *S,
LSRUse &LU, size_t LUIdx) {
const SCEV *S = Worklist.pop_back_val();
if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
- Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
+ Worklist.append(N->op_begin(), N->op_end());
else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
Worklist.push_back(C->getOperand());
else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
} else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
if (!Inserted.insert(U)) continue;
const Value *V = U->getValue();
- if (const Instruction *Inst = dyn_cast<Instruction>(V))
+ if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
+ // Look for instructions defined outside the loop.
if (L->contains(Inst)) continue;
+ } else if (isa<UndefValue>(V))
+ // Undef doesn't have a live range, so it doesn't matter.
+ continue;
for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
UI != UE; ++UI) {
const Instruction *UserInst = dyn_cast<Instruction>(*UI);
/// separate registers. If C is non-null, multiply each subexpression by C.
static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
SmallVectorImpl<const SCEV *> &Ops,
+ SmallVectorImpl<const SCEV *> &UninterestingOps,
+ const Loop *L,
ScalarEvolution &SE) {
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
// Break out add operands.
for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
I != E; ++I)
- CollectSubexprs(*I, C, Ops, SE);
+ CollectSubexprs(*I, C, Ops, UninterestingOps, L, SE);
return;
} else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
// Split a non-zero base out of an addrec.
if (!AR->getStart()->isZero()) {
- CollectSubexprs(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
+ CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
AR->getStepRecurrence(SE),
- AR->getLoop()), C, Ops, SE);
- CollectSubexprs(AR->getStart(), C, Ops, SE);
+ AR->getLoop()),
+ C, Ops, UninterestingOps, L, SE);
+ CollectSubexprs(AR->getStart(), C, Ops, UninterestingOps, L, SE);
return;
}
} else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
CollectSubexprs(Mul->getOperand(1),
C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
- Ops, SE);
+ Ops, UninterestingOps, L, SE);
return;
}
}
- // Otherwise use the value itself.
- Ops.push_back(C ? SE.getMulExpr(C, S) : S);
+ // Otherwise use the value itself. Loop-variant "unknown" values are
+ // uninteresting; we won't be able to do anything meaningful with them.
+ if (!C && isa<SCEVUnknown>(S) && !S->isLoopInvariant(L))
+ UninterestingOps.push_back(S);
+ else
+ Ops.push_back(C ? SE.getMulExpr(C, S) : S);
}
/// GenerateReassociations - Split out subexpressions from adds and the bases of
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);
+ SmallVector<const SCEV *, 8> AddOps, UninterestingAddOps;
+ CollectSubexprs(BaseReg, 0, AddOps, UninterestingAddOps, L, SE);
+
+ // Add any uninteresting values as one register, as we won't be able to
+ // form any interesting reassociation opportunities with them. They'll
+ // just have to be added inside the loop no matter what we do.
+ if (!UninterestingAddOps.empty())
+ AddOps.push_back(SE.getAddExpr(UninterestingAddOps));
+
if (AddOps.size() == 1) continue;
for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
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);
+ SmallVector<const SCEV *, 8> InnerAddOps
+ ( ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
+ InnerAddOps.append
+ (next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
// Don't leave just a constant behind in a register if the constant could
// be folded into an immediate field.
LU.Kind, LU.AccessTy, TLI, SE))
continue;
+ const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
+ if (InnerSum->isZero())
+ continue;
Formula F = Base;
- F.BaseRegs[i] = SE.getAddExpr(InnerAddOps);
+ F.BaseRegs[i] = InnerSum;
F.BaseRegs.push_back(*J);
if (InsertFormula(LU, LUIdx, F))
// If that formula hadn't been seen before, recurse to find more like
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()));
+ F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
(void)InsertFormula(LU, LUIdx, F);
}
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)
+ if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
continue;
- F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
- if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
+ int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
+ if (NewBaseOffs / Factor != Base.AM.BaseOffs)
continue;
// Check that multiplying with the use offset doesn't overflow.
if (Offset / Factor != LU.MinOffset)
continue;
+ Formula F = Base;
+ F.AM.BaseOffs = NewBaseOffs;
+
// 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);
+ const SCEV *FactorS = SE.getConstant(IntTy, Factor);
// Check that multiplying with each base register doesn't overflow.
for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
/// 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) {
+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;
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);
+ const SCEV *FactorS = SE.getConstant(IntTy, Factor);
if (FactorS->isZero())
continue;
// Divide out the factor, ignoring high bits, since we'll be
// 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();
+ F.DeleteBaseReg(F.BaseRegs[i]);
(void)InsertFormula(LU, LUIdx, F);
}
}
}
/// GenerateTruncates - Generate reuse formulae from different IV types.
-void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
- Formula Base) {
+void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
// This requires TargetLowering to tell us which truncates are free.
if (!TLI) return;
// TODO: Use a more targeted data structure.
for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
- Formula F = LU.Formulae[L];
+ const Formula &F = LU.Formulae[L];
// Use the immediate in the scaled register.
if (F.ScaledReg == OrigReg) {
int64_t Offs = (uint64_t)F.AM.BaseOffs +
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(abs64(NewF.AM.BaseOffs)))
+ if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
+ abs64(NewF.AM.BaseOffs)) &&
+ (C->getValue()->getValue() +
+ NewF.AM.BaseOffs).countTrailingZeros() >=
+ CountTrailingZeros_64(NewF.AM.BaseOffs))
goto skip_formula;
// Ok, looks good.
/// by other uses, pick the best one and delete the others.
void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
#ifndef NDEBUG
- bool Changed = false;
+ bool ChangedFormulae = false;
#endif
// Collect the best formula for each unique set of shared registers. This
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
LSRUse &LU = Uses[LUIdx];
FormulaSorter Sorter(L, LU, SE, DT);
+ DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
- // Clear out the set of used regs; it will be recomputed.
- LU.Regs.clear();
-
+ bool Any = false;
for (size_t FIdx = 0, NumForms = LU.Formulae.size();
FIdx != NumForms; ++FIdx) {
Formula &F = LU.Formulae[FIdx];
Formula &Best = LU.Formulae[P.first->second];
if (Sorter.operator()(F, Best))
std::swap(F, Best);
- DEBUG(dbgs() << "Filtering out "; F.print(dbgs());
+ DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
dbgs() << "\n"
- " in favor of "; Best.print(dbgs());
+ " in favor of formula "; Best.print(dbgs());
dbgs() << '\n');
#ifndef NDEBUG
- Changed = true;
+ ChangedFormulae = true;
#endif
- std::swap(F, LU.Formulae.back());
- LU.Formulae.pop_back();
+ LU.DeleteFormula(F);
--FIdx;
--NumForms;
+ Any = true;
continue;
}
- if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
- LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
}
+
+ // Now that we've filtered out some formulae, recompute the Regs set.
+ if (Any)
+ LU.RecomputeRegs(LUIdx, RegUses);
+
+ // Reset this to prepare for the next use.
BestFormulae.clear();
}
- DEBUG(if (Changed) {
+ DEBUG(if (ChangedFormulae) {
dbgs() << "\n"
"After filtering out undesirable candidates:\n";
print_uses(dbgs());
});
}
+// This is a rough guess that seems to work fairly well.
+static const size_t ComplexityLimit = UINT16_MAX;
+
+/// EstimateSearchSpaceComplexity - Estimate the worst-case number of
+/// solutions the solver might have to consider. It almost never considers
+/// this many solutions because it prune the search space, but the pruning
+/// isn't always sufficient.
+size_t LSRInstance::EstimateSearchSpaceComplexity() const {
+ 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 >= ComplexityLimit) {
+ Power = ComplexityLimit;
+ break;
+ }
+ Power *= FSize;
+ if (Power >= ComplexityLimit)
+ break;
+ }
+ return Power;
+}
+
/// 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;
+ if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
+ DEBUG(dbgs() << "The search space is too complex.\n");
- 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;
+ DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
+ "which use a superset of registers used by other "
+ "formulae.\n");
+
+ for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
+ LSRUse &LU = Uses[LUIdx];
+ bool Any = false;
+ for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
+ Formula &F = LU.Formulae[i];
+ // Look for a formula with a constant or GV in a register. If the use
+ // also has a formula with that same value in an immediate field,
+ // delete the one that uses a register.
+ for (SmallVectorImpl<const SCEV *>::const_iterator
+ I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
+ Formula NewF = F;
+ NewF.AM.BaseOffs += C->getValue()->getSExtValue();
+ NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
+ (I - F.BaseRegs.begin()));
+ if (LU.HasFormulaWithSameRegs(NewF)) {
+ DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
+ LU.DeleteFormula(F);
+ --i;
+ --e;
+ Any = true;
+ break;
+ }
+ } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
+ if (!F.AM.BaseGV) {
+ Formula NewF = F;
+ NewF.AM.BaseGV = GV;
+ NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
+ (I - F.BaseRegs.begin()));
+ if (LU.HasFormulaWithSameRegs(NewF)) {
+ DEBUG(dbgs() << " Deleting "; F.print(dbgs());
+ dbgs() << '\n');
+ LU.DeleteFormula(F);
+ --i;
+ --e;
+ Any = true;
+ break;
+ }
+ }
+ }
+ }
}
- Power *= FSize;
- if (Power >= Limit)
- break;
+ if (Any)
+ LU.RecomputeRegs(LUIdx, RegUses);
}
- if (Power < Limit)
- break;
+ DEBUG(dbgs() << "After pre-selection:\n";
+ print_uses(dbgs()));
+ }
+
+ if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
+ DEBUG(dbgs() << "The search space is too complex.\n");
+
+ DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
+ "separated by a constant offset will use the same "
+ "registers.\n");
+
+ // This is especially useful for unrolled loops.
+
+ for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
+ LSRUse &LU = Uses[LUIdx];
+ for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
+ E = LU.Formulae.end(); I != E; ++I) {
+ const Formula &F = *I;
+ if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
+ if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
+ if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
+ /*HasBaseReg=*/false,
+ LU.Kind, LU.AccessTy)) {
+ DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
+ dbgs() << '\n');
+
+ LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
+
+ // Delete formulae from the new use which are no longer legal.
+ bool Any = false;
+ for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
+ Formula &F = LUThatHas->Formulae[i];
+ if (!isLegalUse(F.AM,
+ LUThatHas->MinOffset, LUThatHas->MaxOffset,
+ LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
+ DEBUG(dbgs() << " Deleting "; F.print(dbgs());
+ dbgs() << '\n');
+ LUThatHas->DeleteFormula(F);
+ --i;
+ --e;
+ Any = true;
+ }
+ }
+ if (Any)
+ LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
+
+ // Update the relocs to reference the new use.
+ for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
+ E = Fixups.end(); I != E; ++I) {
+ LSRFixup &Fixup = *I;
+ if (Fixup.LUIdx == LUIdx) {
+ Fixup.LUIdx = LUThatHas - &Uses.front();
+ Fixup.Offset += F.AM.BaseOffs;
+ DEBUG(errs() << "New fixup has offset "
+ << Fixup.Offset << '\n');
+ }
+ if (Fixup.LUIdx == NumUses-1)
+ Fixup.LUIdx = LUIdx;
+ }
+
+ // Delete the old use.
+ DeleteUse(LU);
+ --LUIdx;
+ --NumUses;
+ break;
+ }
+ }
+ }
+ }
+ }
+
+ DEBUG(dbgs() << "After pre-selection:\n";
+ print_uses(dbgs()));
+ }
+
+ // With all other options exhausted, loop until the system is simple
+ // enough to handle.
+ SmallPtrSet<const SCEV *, 4> Taken;
+ while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
// Ok, we have too many of formulae on our hands to conveniently handle.
// Use a rough heuristic to thin out the list.
+ DEBUG(dbgs() << "The search space is too complex.\n");
// Pick the register which is used by the most LSRUses, which is likely
// to be a good reuse register candidate.
// 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;
+ for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
+ LSRUse &LU = Uses[LUIdx];
if (!LU.Regs.count(Best)) continue;
- // Clear out the set of used regs; it will be recomputed.
- LU.Regs.clear();
-
+ bool Any = false;
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();
+ LU.DeleteFormula(F);
--e;
--i;
+ Any = true;
+ assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
continue;
}
-
- if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
- LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
}
+
+ if (Any)
+ LU.RecomputeRegs(LUIdx, RegUses);
}
DEBUG(dbgs() << "After pre-selection:\n";
// If none of the formulae had all of the required registers, relax the
// constraint so that we don't exclude all formulae.
if (!AnySatisfiedReqRegs) {
+ assert(!ReqRegs.empty() && "Solver failed even without required registers");
ReqRegs.clear();
goto retry;
}
}
+/// Solve - Choose one formula from each use. Return the results in the given
+/// Solution vector.
void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
SmallVector<const Formula *, 8> Workspace;
Cost SolutionCost;
DenseSet<const SCEV *> VisitedRegs;
Workspace.reserve(Uses.size());
+ // SolveRecurse does all the work.
SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
CurRegs, VisitedRegs);
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();
+ assert(Solution.size() == Uses.size() && "Malformed solution!");
}
/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
BasicBlock *IDom;
- for (BasicBlock *Rung = IP->getParent(); ; Rung = IDom) {
- IDom = getImmediateDominator(Rung, DT);
- if (!IDom) return IP;
+ for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
+ if (!Rung) return IP;
+ Rung = Rung->getIDom();
+ if (!Rung) return IP;
+ IDom = Rung->getBlock();
// Don't climb into a loop though.
const Loop *IDomLoop = LI.getLoopFor(IDom);
// instead of at the end, so that it can be used for other expansions.
if (IDom == Inst->getParent() &&
(!BetterPos || DT.dominates(BetterPos, Inst)))
- BetterPos = next(BasicBlock::iterator(Inst));
+ BetterPos = llvm::next(BasicBlock::iterator(Inst));
}
if (!AllDominate)
break;
return IP;
}
+/// Expand - Emit instructions for the leading candidate expression for this
+/// LSRUse (this is called "expanding").
Value *LSRInstance::Expand(const LSRFixup &LF,
const Formula &F,
BasicBlock::iterator IP,
// 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()));
+ SE.getConstant(ScaledS->getType(), F.AM.Scale));
Ops.push_back(ScaledS);
// Flush the operand list to suppress SCEVExpander hoisting.
// Emit instructions summing all the operands.
const SCEV *FullS = Ops.empty() ?
- SE.getIntegerSCEV(0, IntTy) :
+ SE.getConstant(IntTy, 0) :
SE.getAddExpr(Ops);
Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
DeadInsts.push_back(LF.OperandValToReplace);
}
+/// ImplementSolution - Rewrite all the fixup locations with new values,
+/// following the chosen solution.
void
LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
Pass *P) {
Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
// Expand the new value definitions and update the users.
- for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
- size_t LUIdx = Fixups[i].LUIdx;
+ for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
+ E = Fixups.end(); I != E; ++I) {
+ const LSRFixup &Fixup = *I;
- Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
+ Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
Changed = true;
}
WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
dbgs() << ":\n");
- /// OptimizeShadowIV - If IV is used in a int-to-float cast
- /// inside the loop then try to eliminate the cast operation.
+ // First, perform some low-level loop optimizations.
OptimizeShadowIV();
+ OptimizeLoopTermCond();
- // Change loop terminating condition to use the postinc iv when possible.
- Changed |= OptimizeLoopTermCond();
-
+ // Start collecting data and preparing for the solver.
CollectInterestingTypesAndFactors();
CollectFixupsAndInitialFormulae();
CollectLoopInvariantFixupsAndFormulae();
SmallVector<const Formula *, 8> Solution;
Solve(Solution);
- assert(Solution.size() == Uses.size() && "Malformed solution!");
// Release memory that is no longer needed.
Factors.clear();
OS << "LSR is examining the following fixup sites:\n";
for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
E = Fixups.end(); I != E; ++I) {
- const LSRFixup &LF = *I;
dbgs() << " ";
- LF.print(OS);
+ I->print(OS);
OS << '\n';
}
}