bool
RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
- if (!RegUsesMap.count(Reg)) return false;
- const SmallBitVector &UsedByIndices =
- RegUsesMap.find(Reg)->second.UsedByIndices;
+ RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
+ if (I == RegUsesMap.end())
+ return false;
+ const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
int i = UsedByIndices.find_first();
if (i == -1) return false;
if ((size_t)i != LUIdx) return true;
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));
+ IntegerType::get(SE.getContext(),
+ 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
if (LHS == RHS)
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 (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
- const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
- IgnoreSignificantBits);
- if (!Start) return 0;
const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
IgnoreSignificantBits);
if (!Step) return 0;
+ const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
+ IgnoreSignificantBits);
+ if (!Start) 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;
}
return Found ? SE.getMulExpr(Ops) : 0;
}
+ return 0;
+ }
// Otherwise we don't know.
return 0;
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
int64_t Result = ExtractImmediate(NewOps.front(), SE);
- S = SE.getAddExpr(NewOps);
+ if (Result != 0)
+ S = SE.getAddExpr(NewOps);
return Result;
} else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
int64_t Result = ExtractImmediate(NewOps.front(), SE);
- S = SE.getAddRecExpr(NewOps, AR->getLoop());
+ if (Result != 0)
+ S = SE.getAddRecExpr(NewOps, AR->getLoop());
return Result;
}
return 0;
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
- S = SE.getAddExpr(NewOps);
+ if (Result)
+ S = SE.getAddExpr(NewOps);
return Result;
} else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
- S = SE.getAddRecExpr(NewOps, AR->getLoop());
+ if (Result)
+ S = SE.getAddRecExpr(NewOps, AR->getLoop());
return Result;
}
return 0;
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;
}
}
bool Changed = false;
while (!DeadInsts.empty()) {
- Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
+ Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
if (I == 0 || !isInstructionTriviallyDead(I))
continue;
: NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
SetupCost(0) {}
- unsigned getNumRegs() const { return NumRegs; }
-
bool operator<(const Cost &Other) const;
void Loose();
/// may be used.
bool AllFixupsOutsideLoop;
+ /// WidestFixupType - This records the widest use type for any fixup using
+ /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
+ /// max fixup widths to be equivalent, because the narrower one may be relying
+ /// on the implicit truncation to truncate away bogus bits.
+ const Type *WidestFixupType;
+
/// Formulae - A list of ways to build a value that can satisfy this user.
/// After the list is populated, one of these is selected heuristically and
/// used to formulate a replacement for OperandValToReplace in UserInst.
LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
MinOffset(INT64_MAX),
MaxOffset(INT64_MIN),
- AllFixupsOutsideLoop(true) {}
+ AllFixupsOutsideLoop(true),
+ WidestFixupType(0) {}
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 print(raw_ostream &OS) 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 {
for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
E = Offsets.end(); I != E; ++I) {
OS << *I;
- if (next(I) != E)
+ if (llvm::next(I) != E)
OS << ',';
}
OS << '}';
if (AllFixupsOutsideLoop)
OS << ", all-fixups-outside-loop";
+
+ if (WidestFixupType)
+ OS << ", widest fixup type: " << *WidestFixupType;
}
void LSRUse::dump() const {
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 {
}
// 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 HasBaseReg,
void FilterOutUndesirableDedicatedRegisters();
size_t EstimateSearchSpaceComplexity() const;
+ void NarrowSearchSpaceByDetectingSupersets();
+ void NarrowSearchSpaceByCollapsingUnrolledCode();
+ void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
+ void NarrowSearchSpaceByPickingWinnerRegs();
void NarrowSearchSpaceUsingHeuristics();
void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
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);
+ const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
if (IterationCount != SE.getSCEV(Sel)) return Cond;
// Check for a max calculation that matches the pattern. There's no check
NewRHS = Sel->getOperand(1);
else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
NewRHS = Sel->getOperand(2);
+ else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
+ NewRHS = SU->getValue();
else
- llvm_unreachable("Max doesn't match expected pattern!");
+ // 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.
}
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;
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 *
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.
+ // Search all uses for the formula. This could be more clever.
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
LSRUse &LU = Uses[LUIdx];
+ // Check whether this use is close enough to OrigLU, to see whether it's
+ // worthwhile looking through its formulae.
+ // Ignore ICmpZero uses because they may contain formulae generated by
+ // GenerateICmpZeroScales, in which case adding fixup offsets may
+ // be invalid.
if (&LU != &OrigLU &&
LU.Kind != LSRUse::ICmpZero &&
LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
+ LU.WidestFixupType == OrigLU.WidestFixupType &&
LU.HasFormulaWithSameRegs(OrigF)) {
+ // Scan through this use's formulae.
for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
E = LU.Formulae.end(); I != E; ++I) {
const Formula &F = *I;
+ // Check to see if this formula has the same registers and symbols
+ // as OrigF.
if (F.BaseRegs == OrigF.BaseRegs &&
F.ScaledReg == OrigF.ScaledReg &&
F.AM.BaseGV == OrigF.AM.BaseGV &&
- F.AM.Scale == OrigF.AM.Scale &&
- LU.Kind) {
+ F.AM.Scale == OrigF.AM.Scale) {
if (F.AM.BaseOffs == 0)
return &LU;
+ // This is the formula where all the registers and symbols matched;
+ // there aren't going to be any others. Since we declined it, we
+ // can skip the rest of the formulae and procede to the next LSRUse.
break;
}
}
}
}
+ // Nothing looked good.
return 0;
}
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());
}
for (SmallSetVector<const SCEV *, 4>::const_iterator
I = Strides.begin(), E = Strides.end(); I != E; ++I)
for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
- next(I); NewStrideIter != E; ++NewStrideIter) {
+ llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
const SCEV *OldStride = *I;
const SCEV *NewStride = *NewStrideIter;
LF.Offset = P.second;
LSRUse &LU = Uses[LF.LUIdx];
LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
+ if (!LU.WidestFixupType ||
+ SE.getTypeSizeInBits(LU.WidestFixupType) <
+ SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
+ LU.WidestFixupType = LF.OperandValToReplace->getType();
// If this is the first use of this LSRUse, give it a formula.
if (LU.Formulae.empty()) {
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);
LF.Offset = P.second;
LSRUse &LU = Uses[LF.LUIdx];
LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
+ if (!LU.WidestFixupType ||
+ SE.getTypeSizeInBits(LU.WidestFixupType) <
+ SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
+ LU.WidestFixupType = LF.OperandValToReplace->getType();
InsertSupplementalFormula(U, LU, LF.LUIdx);
CountRegisters(LU.Formulae.back(), Uses.size() - 1);
break;
/// 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,
+ 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, 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.getConstant(AR->getType(), 0),
AR->getStepRecurrence(SE),
- AR->getLoop()), C, Ops, SE);
- CollectSubexprs(AR->getStart(), C, Ops, SE);
+ AR->getLoop()),
+ C, Ops, L, SE);
+ CollectSubexprs(AR->getStart(), C, Ops, 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, L, SE);
return;
}
}
- // Otherwise use the value itself.
+ // Otherwise use the value itself, optionally with a scale applied.
Ops.push_back(C ? SE.getMulExpr(C, S) : S);
}
const SCEV *BaseReg = Base.BaseRegs[i];
SmallVector<const SCEV *, 8> AddOps;
- CollectSubexprs(BaseReg, 0, AddOps, SE);
+ CollectSubexprs(BaseReg, 0, AddOps, L, SE);
+
if (AddOps.size() == 1) continue;
for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
JE = AddOps.end(); J != JE; ++J) {
+
+ // Loop-variant "unknown" values are uninteresting; we won't be able to
+ // do anything meaningful with them.
+ if (isa<SCEVUnknown>(*J) && !(*J)->isLoopInvariant(L))
+ continue;
+
// 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,
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
+ (llvm::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.
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;
+ SmallVector<int64_t, 2> Worklist;
Worklist.push_back(LU.MinOffset);
if (LU.MaxOffset != LU.MinOffset)
Worklist.push_back(LU.MaxOffset);
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.getConstant(G->getType(), *I));
+ // Add the offset to the base register.
+ const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
+ // If it cancelled out, drop the base register, otherwise update it.
+ if (NewG->isZero()) {
+ std::swap(F.BaseRegs[i], F.BaseRegs.back());
+ F.BaseRegs.pop_back();
+ } else
+ F.BaseRegs[i] = NewG;
(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;
// 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 +
}
GenerateCrossUseConstantOffsets();
+
+ DEBUG(dbgs() << "\n"
+ "After generating reuse formulae:\n";
+ print_uses(dbgs()));
}
/// If their are multiple formulae with the same set of registers used
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() {
+/// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
+/// of the registers of another formula, it won't help reduce register
+/// pressure (though it may not necessarily hurt register pressure); remove
+/// it to simplify the system.
+void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
DEBUG(dbgs() << "The search space is too complex.\n");
DEBUG(dbgs() << "After pre-selection:\n";
print_uses(dbgs()));
}
+}
+/// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
+/// for expressions like A, A+1, A+2, etc., allocate a single register for
+/// them.
+void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
DEBUG(dbgs() << "The search space is too complex.\n");
if (Fixup.LUIdx == LUIdx) {
Fixup.LUIdx = LUThatHas - &Uses.front();
Fixup.Offset += F.AM.BaseOffs;
- DEBUG(errs() << "New fixup has offset "
+ DEBUG(dbgs() << "New fixup has offset "
<< Fixup.Offset << '\n');
}
if (Fixup.LUIdx == NumUses-1)
DEBUG(dbgs() << "After pre-selection:\n";
print_uses(dbgs()));
}
+}
+
+/// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
+/// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
+/// we've done more filtering, as it may be able to find more formulae to
+/// eliminate.
+void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
+ if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
+ DEBUG(dbgs() << "The search space is too complex.\n");
+ DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
+ "undesirable dedicated registers.\n");
+
+ FilterOutUndesirableDedicatedRegisters();
+
+ DEBUG(dbgs() << "After pre-selection:\n";
+ print_uses(dbgs()));
+ }
+}
+
+/// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
+/// to be profitable, and then in any use which has any reference to that
+/// register, delete all formulae which do not reference that register.
+void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
// With all other options exhausted, loop until the system is simple
// enough to handle.
SmallPtrSet<const SCEV *, 4> Taken;
}
}
+/// 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() {
+ NarrowSearchSpaceByDetectingSupersets();
+ NarrowSearchSpaceByCollapsingUnrolledCode();
+ NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
+ NarrowSearchSpaceByPickingWinnerRegs();
+}
+
/// SolveRecurse - This is the recursive solver.
void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
Cost &SolutionCost,
Solution[i]->print(dbgs());
dbgs() << '\n';
});
+
+ assert(Solution.size() == Uses.size() && "Malformed solution!");
}
/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
BasicBlock *IDom;
for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
- assert(Rung && "Block has no DomTreeNode!");
+ if (!Rung) return IP;
Rung = Rung->getIDom();
if (!Rung) return IP;
IDom = Rung->getBlock();
// to formulate the values needed for the uses.
GenerateAllReuseFormulae();
- DEBUG(dbgs() << "\n"
- "After generating reuse formulae:\n";
- print_uses(dbgs()));
-
FilterOutUndesirableDedicatedRegisters();
NarrowSearchSpaceUsingHeuristics();
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';
}
}
}
char LoopStrengthReduce::ID = 0;
-static RegisterPass<LoopStrengthReduce>
-X("loop-reduce", "Loop Strength Reduction");
+INITIALIZE_PASS(LoopStrengthReduce, "loop-reduce",
+ "Loop Strength Reduction", false, false);
Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
return new LoopStrengthReduce(TLI);
}
LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
- : LoopPass(&ID), TLI(tli) {}
+ : LoopPass(ID), TLI(tli) {}
void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
// We split critical edges, so we change the CFG. However, we do update