1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/ADT/SmallBitVector.h"
69 #include "llvm/ADT/SetVector.h"
70 #include "llvm/ADT/DenseSet.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/ValueHandle.h"
73 #include "llvm/Support/raw_ostream.h"
74 #include "llvm/Target/TargetLowering.h"
80 /// RegSortData - This class holds data which is used to order reuse candidates.
83 /// UsedByIndices - This represents the set of LSRUse indices which reference
84 /// a particular register.
85 SmallBitVector UsedByIndices;
89 void print(raw_ostream &OS) const;
95 void RegSortData::print(raw_ostream &OS) const {
96 OS << "[NumUses=" << UsedByIndices.count() << ']';
99 void RegSortData::dump() const {
100 print(errs()); errs() << '\n';
105 /// RegUseTracker - Map register candidates to information about how they are
107 class RegUseTracker {
108 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
111 SmallVector<const SCEV *, 16> RegSequence;
114 void CountRegister(const SCEV *Reg, size_t LUIdx);
116 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
118 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
122 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
123 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
124 iterator begin() { return RegSequence.begin(); }
125 iterator end() { return RegSequence.end(); }
126 const_iterator begin() const { return RegSequence.begin(); }
127 const_iterator end() const { return RegSequence.end(); }
133 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
134 std::pair<RegUsesTy::iterator, bool> Pair =
135 RegUses.insert(std::make_pair(Reg, RegSortData()));
136 RegSortData &RSD = Pair.first->second;
138 RegSequence.push_back(Reg);
139 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
140 RSD.UsedByIndices.set(LUIdx);
144 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
145 if (!RegUses.count(Reg)) return false;
146 const SmallBitVector &UsedByIndices =
147 RegUses.find(Reg)->second.UsedByIndices;
148 int i = UsedByIndices.find_first();
149 if (i == -1) return false;
150 if ((size_t)i != LUIdx) return true;
151 return UsedByIndices.find_next(i) != -1;
154 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
155 RegUsesTy::const_iterator I = RegUses.find(Reg);
156 assert(I != RegUses.end() && "Unknown register!");
157 return I->second.UsedByIndices;
160 void RegUseTracker::clear() {
167 /// Formula - This class holds information that describes a formula for
168 /// computing satisfying a use. It may include broken-out immediates and scaled
171 /// AM - This is used to represent complex addressing, as well as other kinds
172 /// of interesting uses.
173 TargetLowering::AddrMode AM;
175 /// BaseRegs - The list of "base" registers for this use. When this is
176 /// non-empty, AM.HasBaseReg should be set to true.
177 SmallVector<const SCEV *, 2> BaseRegs;
179 /// ScaledReg - The 'scaled' register for this use. This should be non-null
180 /// when AM.Scale is not zero.
181 const SCEV *ScaledReg;
183 Formula() : ScaledReg(0) {}
185 void InitialMatch(const SCEV *S, Loop *L,
186 ScalarEvolution &SE, DominatorTree &DT);
188 unsigned getNumRegs() const;
189 const Type *getType() const;
191 bool referencesReg(const SCEV *S) const;
192 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
193 const RegUseTracker &RegUses) const;
195 void print(raw_ostream &OS) const;
201 /// DoInitialMatch - Recurrsion helper for InitialMatch.
202 static void DoInitialMatch(const SCEV *S, Loop *L,
203 SmallVectorImpl<const SCEV *> &Good,
204 SmallVectorImpl<const SCEV *> &Bad,
205 ScalarEvolution &SE, DominatorTree &DT) {
206 // Collect expressions which properly dominate the loop header.
207 if (S->properlyDominates(L->getHeader(), &DT)) {
212 // Look at add operands.
213 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
214 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
216 DoInitialMatch(*I, L, Good, Bad, SE, DT);
220 // Look at addrec operands.
221 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
222 if (!AR->getStart()->isZero()) {
223 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
224 DoInitialMatch(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
225 AR->getStepRecurrence(SE),
227 L, Good, Bad, SE, DT);
231 // Handle a multiplication by -1 (negation) if it didn't fold.
232 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
233 if (Mul->getOperand(0)->isAllOnesValue()) {
234 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
235 const SCEV *NewMul = SE.getMulExpr(Ops);
237 SmallVector<const SCEV *, 4> MyGood;
238 SmallVector<const SCEV *, 4> MyBad;
239 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
240 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
241 SE.getEffectiveSCEVType(NewMul->getType())));
242 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
243 E = MyGood.end(); I != E; ++I)
244 Good.push_back(SE.getMulExpr(NegOne, *I));
245 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
246 E = MyBad.end(); I != E; ++I)
247 Bad.push_back(SE.getMulExpr(NegOne, *I));
251 // Ok, we can't do anything interesting. Just stuff the whole thing into a
252 // register and hope for the best.
256 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
257 /// attempting to keep all loop-invariant and loop-computable values in a
258 /// single base register.
259 void Formula::InitialMatch(const SCEV *S, Loop *L,
260 ScalarEvolution &SE, DominatorTree &DT) {
261 SmallVector<const SCEV *, 4> Good;
262 SmallVector<const SCEV *, 4> Bad;
263 DoInitialMatch(S, L, Good, Bad, SE, DT);
265 BaseRegs.push_back(SE.getAddExpr(Good));
266 AM.HasBaseReg = true;
269 BaseRegs.push_back(SE.getAddExpr(Bad));
270 AM.HasBaseReg = true;
274 /// getNumRegs - Return the total number of register operands used by this
275 /// formula. This does not include register uses implied by non-constant
277 unsigned Formula::getNumRegs() const {
278 return !!ScaledReg + BaseRegs.size();
281 /// getType - Return the type of this formula, if it has one, or null
282 /// otherwise. This type is meaningless except for the bit size.
283 const Type *Formula::getType() const {
284 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
285 ScaledReg ? ScaledReg->getType() :
286 AM.BaseGV ? AM.BaseGV->getType() :
290 /// referencesReg - Test if this formula references the given register.
291 bool Formula::referencesReg(const SCEV *S) const {
292 return S == ScaledReg ||
293 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
296 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
297 /// which are used by uses other than the use with the given index.
298 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
299 const RegUseTracker &RegUses) const {
301 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
303 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
304 E = BaseRegs.end(); I != E; ++I)
305 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
310 void Formula::print(raw_ostream &OS) const {
313 if (!First) OS << " + "; else First = false;
314 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
316 if (AM.BaseOffs != 0) {
317 if (!First) OS << " + "; else First = false;
320 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
321 E = BaseRegs.end(); I != E; ++I) {
322 if (!First) OS << " + "; else First = false;
323 OS << "reg(" << **I << ')';
326 if (!First) OS << " + "; else First = false;
327 OS << AM.Scale << "*reg(";
336 void Formula::dump() const {
337 print(errs()); errs() << '\n';
340 /// getSDiv - Return an expression for LHS /s RHS, if it can be determined,
341 /// or null otherwise. If IgnoreSignificantBits is true, expressions like
342 /// (X * Y) /s Y are simplified to Y, ignoring that the multiplication may
343 /// overflow, which is useful when the result will be used in a context where
344 /// the most significant bits are ignored.
345 static const SCEV *getSDiv(const SCEV *LHS, const SCEV *RHS,
347 bool IgnoreSignificantBits = false) {
348 // Handle the trivial case, which works for any SCEV type.
350 return SE.getIntegerSCEV(1, LHS->getType());
352 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
354 if (RHS->isAllOnesValue())
355 return SE.getMulExpr(LHS, RHS);
357 // Check for a division of a constant by a constant.
358 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
359 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
362 if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
364 return SE.getConstant(C->getValue()->getValue()
365 .sdiv(RC->getValue()->getValue()));
368 // Distribute the sdiv over addrec operands.
369 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
370 const SCEV *Start = getSDiv(AR->getStart(), RHS, SE,
371 IgnoreSignificantBits);
372 if (!Start) return 0;
373 const SCEV *Step = getSDiv(AR->getStepRecurrence(SE), RHS, SE,
374 IgnoreSignificantBits);
376 return SE.getAddRecExpr(Start, Step, AR->getLoop());
379 // Distribute the sdiv over add operands.
380 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
381 SmallVector<const SCEV *, 8> Ops;
382 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
384 const SCEV *Op = getSDiv(*I, RHS, SE,
385 IgnoreSignificantBits);
389 return SE.getAddExpr(Ops);
392 // Check for a multiply operand that we can pull RHS out of.
393 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
394 if (IgnoreSignificantBits || Mul->hasNoSignedWrap()) {
395 SmallVector<const SCEV *, 4> Ops;
397 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
400 if (const SCEV *Q = getSDiv(*I, RHS, SE, IgnoreSignificantBits)) {
407 return Found ? SE.getMulExpr(Ops) : 0;
410 // Otherwise we don't know.
414 /// ExtractImmediate - If S involves the addition of a constant integer value,
415 /// return that integer value, and mutate S to point to a new SCEV with that
417 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
418 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
419 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
420 S = SE.getIntegerSCEV(0, C->getType());
421 return C->getValue()->getSExtValue();
423 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
424 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
425 int64_t Result = ExtractImmediate(NewOps.front(), SE);
426 S = SE.getAddExpr(NewOps);
428 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
429 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
430 int64_t Result = ExtractImmediate(NewOps.front(), SE);
431 S = SE.getAddRecExpr(NewOps, AR->getLoop());
437 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
438 /// return that symbol, and mutate S to point to a new SCEV with that
440 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
441 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
442 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
443 S = SE.getIntegerSCEV(0, GV->getType());
446 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
447 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
448 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
449 S = SE.getAddExpr(NewOps);
451 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
452 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
453 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
454 S = SE.getAddRecExpr(NewOps, AR->getLoop());
460 /// isAddressUse - Returns true if the specified instruction is using the
461 /// specified value as an address.
462 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
463 bool isAddress = isa<LoadInst>(Inst);
464 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
465 if (SI->getOperand(1) == OperandVal)
467 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
468 // Addressing modes can also be folded into prefetches and a variety
470 switch (II->getIntrinsicID()) {
472 case Intrinsic::prefetch:
473 case Intrinsic::x86_sse2_loadu_dq:
474 case Intrinsic::x86_sse2_loadu_pd:
475 case Intrinsic::x86_sse_loadu_ps:
476 case Intrinsic::x86_sse_storeu_ps:
477 case Intrinsic::x86_sse2_storeu_pd:
478 case Intrinsic::x86_sse2_storeu_dq:
479 case Intrinsic::x86_sse2_storel_dq:
480 if (II->getOperand(1) == OperandVal)
488 /// getAccessType - Return the type of the memory being accessed.
489 static const Type *getAccessType(const Instruction *Inst) {
490 const Type *AccessTy = Inst->getType();
491 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
492 AccessTy = SI->getOperand(0)->getType();
493 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
494 // Addressing modes can also be folded into prefetches and a variety
496 switch (II->getIntrinsicID()) {
498 case Intrinsic::x86_sse_storeu_ps:
499 case Intrinsic::x86_sse2_storeu_pd:
500 case Intrinsic::x86_sse2_storeu_dq:
501 case Intrinsic::x86_sse2_storel_dq:
502 AccessTy = II->getOperand(1)->getType();
507 // All pointers have the same requirements, so canonicalize them to an
508 // arbitrary pointer type to minimize variation.
509 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
510 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
511 PTy->getAddressSpace());
516 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
517 /// specified set are trivially dead, delete them and see if this makes any of
518 /// their operands subsequently dead.
520 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
521 bool Changed = false;
523 while (!DeadInsts.empty()) {
524 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
526 if (I == 0 || !isInstructionTriviallyDead(I))
529 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
530 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
533 DeadInsts.push_back(U);
536 I->eraseFromParent();
545 /// Cost - This class is used to measure and compare candidate formulae.
547 /// TODO: Some of these could be merged. Also, a lexical ordering
548 /// isn't always optimal.
552 unsigned NumBaseAdds;
558 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
561 unsigned getNumRegs() const { return NumRegs; }
563 bool operator<(const Cost &Other) const;
567 void RateFormula(const Formula &F,
568 SmallPtrSet<const SCEV *, 16> &Regs,
569 const DenseSet<const SCEV *> &VisitedRegs,
571 const SmallVectorImpl<int64_t> &Offsets,
572 ScalarEvolution &SE, DominatorTree &DT);
574 void print(raw_ostream &OS) const;
578 void RateRegister(const SCEV *Reg,
579 SmallPtrSet<const SCEV *, 16> &Regs,
581 ScalarEvolution &SE, DominatorTree &DT);
582 void RatePrimaryRegister(const SCEV *Reg,
583 SmallPtrSet<const SCEV *, 16> &Regs,
585 ScalarEvolution &SE, DominatorTree &DT);
590 /// RateRegister - Tally up interesting quantities from the given register.
591 void Cost::RateRegister(const SCEV *Reg,
592 SmallPtrSet<const SCEV *, 16> &Regs,
594 ScalarEvolution &SE, DominatorTree &DT) {
595 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
596 if (AR->getLoop() == L)
597 AddRecCost += 1; /// TODO: This should be a function of the stride.
599 // If this is an addrec for a loop that's already been visited by LSR,
600 // don't second-guess its addrec phi nodes. LSR isn't currently smart
601 // enough to reason about more than one loop at a time. Consider these
602 // registers free and leave them alone.
603 else if (L->contains(AR->getLoop()) ||
604 (!AR->getLoop()->contains(L) &&
605 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
606 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
607 PHINode *PN = dyn_cast<PHINode>(I); ++I)
608 if (SE.isSCEVable(PN->getType()) &&
609 (SE.getEffectiveSCEVType(PN->getType()) ==
610 SE.getEffectiveSCEVType(AR->getType())) &&
611 SE.getSCEV(PN) == AR)
614 // If this isn't one of the addrecs that the loop already has, it
615 // would require a costly new phi and add. TODO: This isn't
616 // precisely modeled right now.
618 if (!Regs.count(AR->getStart()))
619 RateRegister(AR->getStart(), Regs, L, SE, DT);
622 // Add the step value register, if it needs one.
623 // TODO: The non-affine case isn't precisely modeled here.
624 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
625 if (!Regs.count(AR->getStart()))
626 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
630 // Rough heuristic; favor registers which don't require extra setup
631 // instructions in the preheader.
632 if (!isa<SCEVUnknown>(Reg) &&
633 !isa<SCEVConstant>(Reg) &&
634 !(isa<SCEVAddRecExpr>(Reg) &&
635 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
636 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
640 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
642 void Cost::RatePrimaryRegister(const SCEV *Reg,
643 SmallPtrSet<const SCEV *, 16> &Regs,
645 ScalarEvolution &SE, DominatorTree &DT) {
646 if (Regs.insert(Reg))
647 RateRegister(Reg, Regs, L, SE, DT);
650 void Cost::RateFormula(const Formula &F,
651 SmallPtrSet<const SCEV *, 16> &Regs,
652 const DenseSet<const SCEV *> &VisitedRegs,
654 const SmallVectorImpl<int64_t> &Offsets,
655 ScalarEvolution &SE, DominatorTree &DT) {
656 // Tally up the registers.
657 if (const SCEV *ScaledReg = F.ScaledReg) {
658 if (VisitedRegs.count(ScaledReg)) {
662 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
664 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
665 E = F.BaseRegs.end(); I != E; ++I) {
666 const SCEV *BaseReg = *I;
667 if (VisitedRegs.count(BaseReg)) {
671 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
673 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
674 BaseReg->hasComputableLoopEvolution(L);
677 if (F.BaseRegs.size() > 1)
678 NumBaseAdds += F.BaseRegs.size() - 1;
680 // Tally up the non-zero immediates.
681 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
682 E = Offsets.end(); I != E; ++I) {
683 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
685 ImmCost += 64; // Handle symbolic values conservatively.
686 // TODO: This should probably be the pointer size.
687 else if (Offset != 0)
688 ImmCost += APInt(64, Offset, true).getMinSignedBits();
692 /// Loose - Set this cost to a loosing value.
702 /// operator< - Choose the lower cost.
703 bool Cost::operator<(const Cost &Other) const {
704 if (NumRegs != Other.NumRegs)
705 return NumRegs < Other.NumRegs;
706 if (AddRecCost != Other.AddRecCost)
707 return AddRecCost < Other.AddRecCost;
708 if (NumIVMuls != Other.NumIVMuls)
709 return NumIVMuls < Other.NumIVMuls;
710 if (NumBaseAdds != Other.NumBaseAdds)
711 return NumBaseAdds < Other.NumBaseAdds;
712 if (ImmCost != Other.ImmCost)
713 return ImmCost < Other.ImmCost;
714 if (SetupCost != Other.SetupCost)
715 return SetupCost < Other.SetupCost;
719 void Cost::print(raw_ostream &OS) const {
720 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
722 OS << ", with addrec cost " << AddRecCost;
724 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
725 if (NumBaseAdds != 0)
726 OS << ", plus " << NumBaseAdds << " base add"
727 << (NumBaseAdds == 1 ? "" : "s");
729 OS << ", plus " << ImmCost << " imm cost";
731 OS << ", plus " << SetupCost << " setup cost";
734 void Cost::dump() const {
735 print(errs()); errs() << '\n';
740 /// LSRFixup - An operand value in an instruction which is to be replaced
741 /// with some equivalent, possibly strength-reduced, replacement.
743 /// UserInst - The instruction which will be updated.
744 Instruction *UserInst;
746 /// OperandValToReplace - The operand of the instruction which will
747 /// be replaced. The operand may be used more than once; every instance
748 /// will be replaced.
749 Value *OperandValToReplace;
751 /// PostIncLoop - If this user is to use the post-incremented value of an
752 /// induction variable, this variable is non-null and holds the loop
753 /// associated with the induction variable.
754 const Loop *PostIncLoop;
756 /// LUIdx - The index of the LSRUse describing the expression which
757 /// this fixup needs, minus an offset (below).
760 /// Offset - A constant offset to be added to the LSRUse expression.
761 /// This allows multiple fixups to share the same LSRUse with different
762 /// offsets, for example in an unrolled loop.
767 void print(raw_ostream &OS) const;
774 : UserInst(0), OperandValToReplace(0), PostIncLoop(0),
775 LUIdx(~size_t(0)), Offset(0) {}
777 void LSRFixup::print(raw_ostream &OS) const {
779 // Store is common and interesting enough to be worth special-casing.
780 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
782 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
783 } else if (UserInst->getType()->isVoidTy())
784 OS << UserInst->getOpcodeName();
786 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
788 OS << ", OperandValToReplace=";
789 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
792 OS << ", PostIncLoop=";
793 WriteAsOperand(OS, PostIncLoop->getHeader(), /*PrintType=*/false);
796 if (LUIdx != ~size_t(0))
797 OS << ", LUIdx=" << LUIdx;
800 OS << ", Offset=" << Offset;
803 void LSRFixup::dump() const {
804 print(errs()); errs() << '\n';
809 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
810 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
811 struct UniquifierDenseMapInfo {
812 static SmallVector<const SCEV *, 2> getEmptyKey() {
813 SmallVector<const SCEV *, 2> V;
814 V.push_back(reinterpret_cast<const SCEV *>(-1));
818 static SmallVector<const SCEV *, 2> getTombstoneKey() {
819 SmallVector<const SCEV *, 2> V;
820 V.push_back(reinterpret_cast<const SCEV *>(-2));
824 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
826 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
827 E = V.end(); I != E; ++I)
828 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
832 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
833 const SmallVector<const SCEV *, 2> &RHS) {
838 /// LSRUse - This class holds the state that LSR keeps for each use in
839 /// IVUsers, as well as uses invented by LSR itself. It includes information
840 /// about what kinds of things can be folded into the user, information about
841 /// the user itself, and information about how the use may be satisfied.
842 /// TODO: Represent multiple users of the same expression in common?
844 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
847 /// KindType - An enum for a kind of use, indicating what types of
848 /// scaled and immediate operands it might support.
850 Basic, ///< A normal use, with no folding.
851 Special, ///< A special case of basic, allowing -1 scales.
852 Address, ///< An address use; folding according to TargetLowering
853 ICmpZero ///< An equality icmp with both operands folded into one.
854 // TODO: Add a generic icmp too?
858 const Type *AccessTy;
860 SmallVector<int64_t, 8> Offsets;
864 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
865 /// LSRUse are outside of the loop, in which case some special-case heuristics
867 bool AllFixupsOutsideLoop;
869 /// Formulae - A list of ways to build a value that can satisfy this user.
870 /// After the list is populated, one of these is selected heuristically and
871 /// used to formulate a replacement for OperandValToReplace in UserInst.
872 SmallVector<Formula, 12> Formulae;
874 /// Regs - The set of register candidates used by all formulae in this LSRUse.
875 SmallPtrSet<const SCEV *, 4> Regs;
877 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
878 MinOffset(INT64_MAX),
879 MaxOffset(INT64_MIN),
880 AllFixupsOutsideLoop(true) {}
882 bool InsertFormula(size_t LUIdx, const Formula &F);
886 void print(raw_ostream &OS) const;
890 /// InsertFormula - If the given formula has not yet been inserted, add it to
891 /// the list, and return true. Return false otherwise.
892 bool LSRUse::InsertFormula(size_t LUIdx, const Formula &F) {
893 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
894 if (F.ScaledReg) Key.push_back(F.ScaledReg);
895 // Unstable sort by host order ok, because this is only used for uniquifying.
896 std::sort(Key.begin(), Key.end());
898 if (!Uniquifier.insert(Key).second)
901 // Using a register to hold the value of 0 is not profitable.
902 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
903 "Zero allocated in a scaled register!");
905 for (SmallVectorImpl<const SCEV *>::const_iterator I =
906 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
907 assert(!(*I)->isZero() && "Zero allocated in a base register!");
910 // Add the formula to the list.
911 Formulae.push_back(F);
913 // Record registers now being used by this use.
914 if (F.ScaledReg) Regs.insert(F.ScaledReg);
915 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
920 void LSRUse::print(raw_ostream &OS) const {
921 OS << "LSR Use: Kind=";
923 case Basic: OS << "Basic"; break;
924 case Special: OS << "Special"; break;
925 case ICmpZero: OS << "ICmpZero"; break;
928 if (isa<PointerType>(AccessTy))
929 OS << "pointer"; // the full pointer type could be really verbose
935 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
936 E = Offsets.end(); I != E; ++I) {
943 if (AllFixupsOutsideLoop)
944 OS << ", all-fixups-outside-loop";
947 void LSRUse::dump() const {
948 print(errs()); errs() << '\n';
951 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
952 /// be completely folded into the user instruction at isel time. This includes
953 /// address-mode folding and special icmp tricks.
954 static bool isLegalUse(const TargetLowering::AddrMode &AM,
955 LSRUse::KindType Kind, const Type *AccessTy,
956 const TargetLowering *TLI) {
958 case LSRUse::Address:
959 // If we have low-level target information, ask the target if it can
960 // completely fold this address.
961 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
963 // Otherwise, just guess that reg+reg addressing is legal.
964 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
966 case LSRUse::ICmpZero:
967 // There's not even a target hook for querying whether it would be legal to
968 // fold a GV into an ICmp.
972 // ICmp only has two operands; don't allow more than two non-trivial parts.
973 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
976 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
977 // putting the scaled register in the other operand of the icmp.
978 if (AM.Scale != 0 && AM.Scale != -1)
981 // If we have low-level target information, ask the target if it can fold an
982 // integer immediate on an icmp.
983 if (AM.BaseOffs != 0) {
984 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
991 // Only handle single-register values.
992 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
994 case LSRUse::Special:
995 // Only handle -1 scales, or no scale.
996 return AM.Scale == 0 || AM.Scale == -1;
1002 static bool isLegalUse(TargetLowering::AddrMode AM,
1003 int64_t MinOffset, int64_t MaxOffset,
1004 LSRUse::KindType Kind, const Type *AccessTy,
1005 const TargetLowering *TLI) {
1006 // Check for overflow.
1007 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1010 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1011 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1012 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1013 // Check for overflow.
1014 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1017 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1018 return isLegalUse(AM, Kind, AccessTy, TLI);
1023 static bool isAlwaysFoldable(int64_t BaseOffs,
1024 GlobalValue *BaseGV,
1026 LSRUse::KindType Kind, const Type *AccessTy,
1027 const TargetLowering *TLI,
1028 ScalarEvolution &SE) {
1029 // Fast-path: zero is always foldable.
1030 if (BaseOffs == 0 && !BaseGV) return true;
1032 // Conservatively, create an address with an immediate and a
1033 // base and a scale.
1034 TargetLowering::AddrMode AM;
1035 AM.BaseOffs = BaseOffs;
1037 AM.HasBaseReg = HasBaseReg;
1038 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1040 return isLegalUse(AM, Kind, AccessTy, TLI);
1043 static bool isAlwaysFoldable(const SCEV *S,
1044 int64_t MinOffset, int64_t MaxOffset,
1046 LSRUse::KindType Kind, const Type *AccessTy,
1047 const TargetLowering *TLI,
1048 ScalarEvolution &SE) {
1049 // Fast-path: zero is always foldable.
1050 if (S->isZero()) return true;
1052 // Conservatively, create an address with an immediate and a
1053 // base and a scale.
1054 int64_t BaseOffs = ExtractImmediate(S, SE);
1055 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1057 // If there's anything else involved, it's not foldable.
1058 if (!S->isZero()) return false;
1060 // Fast-path: zero is always foldable.
1061 if (BaseOffs == 0 && !BaseGV) return true;
1063 // Conservatively, create an address with an immediate and a
1064 // base and a scale.
1065 TargetLowering::AddrMode AM;
1066 AM.BaseOffs = BaseOffs;
1068 AM.HasBaseReg = HasBaseReg;
1069 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1071 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1074 /// FormulaSorter - This class implements an ordering for formulae which sorts
1075 /// the by their standalone cost.
1076 class FormulaSorter {
1077 /// These two sets are kept empty, so that we compute standalone costs.
1078 DenseSet<const SCEV *> VisitedRegs;
1079 SmallPtrSet<const SCEV *, 16> Regs;
1082 ScalarEvolution &SE;
1086 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1087 : L(l), LU(&lu), SE(se), DT(dt) {}
1089 bool operator()(const Formula &A, const Formula &B) {
1091 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1094 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1096 return CostA < CostB;
1100 /// LSRInstance - This class holds state for the main loop strength reduction
1104 ScalarEvolution &SE;
1106 const TargetLowering *const TLI;
1110 /// IVIncInsertPos - This is the insert position that the current loop's
1111 /// induction variable increment should be placed. In simple loops, this is
1112 /// the latch block's terminator. But in more complicated cases, this is a
1113 /// position which will dominate all the in-loop post-increment users.
1114 Instruction *IVIncInsertPos;
1116 /// Factors - Interesting factors between use strides.
1117 SmallSetVector<int64_t, 8> Factors;
1119 /// Types - Interesting use types, to facilitate truncation reuse.
1120 SmallSetVector<const Type *, 4> Types;
1122 /// Fixups - The list of operands which are to be replaced.
1123 SmallVector<LSRFixup, 16> Fixups;
1125 /// Uses - The list of interesting uses.
1126 SmallVector<LSRUse, 16> Uses;
1128 /// RegUses - Track which uses use which register candidates.
1129 RegUseTracker RegUses;
1131 void OptimizeShadowIV();
1132 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1133 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1134 bool OptimizeLoopTermCond();
1136 void CollectInterestingTypesAndFactors();
1137 void CollectFixupsAndInitialFormulae();
1139 LSRFixup &getNewFixup() {
1140 Fixups.push_back(LSRFixup());
1141 return Fixups.back();
1144 // Support for sharing of LSRUses between LSRFixups.
1145 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1148 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1149 LSRUse::KindType Kind, const Type *AccessTy);
1151 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1152 LSRUse::KindType Kind,
1153 const Type *AccessTy);
1156 void InsertInitialFormula(const SCEV *S, Loop *L, LSRUse &LU, size_t LUIdx);
1157 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1158 void CountRegisters(const Formula &F, size_t LUIdx);
1159 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1161 void CollectLoopInvariantFixupsAndFormulae();
1163 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1164 unsigned Depth = 0);
1165 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1166 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1167 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1168 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1169 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1170 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1171 void GenerateCrossUseConstantOffsets();
1172 void GenerateAllReuseFormulae();
1174 void FilterOutUndesirableDedicatedRegisters();
1175 void NarrowSearchSpaceUsingHeuristics();
1177 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1179 SmallVectorImpl<const Formula *> &Workspace,
1180 const Cost &CurCost,
1181 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1182 DenseSet<const SCEV *> &VisitedRegs) const;
1183 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1185 Value *Expand(const LSRFixup &LF,
1187 BasicBlock::iterator IP, Loop *L, Instruction *IVIncInsertPos,
1188 SCEVExpander &Rewriter,
1189 SmallVectorImpl<WeakVH> &DeadInsts,
1190 ScalarEvolution &SE, DominatorTree &DT) const;
1191 void Rewrite(const LSRFixup &LF,
1193 Loop *L, Instruction *IVIncInsertPos,
1194 SCEVExpander &Rewriter,
1195 SmallVectorImpl<WeakVH> &DeadInsts,
1196 ScalarEvolution &SE, DominatorTree &DT,
1198 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1201 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1203 bool getChanged() const { return Changed; }
1205 void print_factors_and_types(raw_ostream &OS) const;
1206 void print_fixups(raw_ostream &OS) const;
1207 void print_uses(raw_ostream &OS) const;
1208 void print(raw_ostream &OS) const;
1214 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1215 /// inside the loop then try to eliminate the cast opeation.
1216 void LSRInstance::OptimizeShadowIV() {
1217 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1218 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1221 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1222 UI != E; /* empty */) {
1223 IVUsers::const_iterator CandidateUI = UI;
1225 Instruction *ShadowUse = CandidateUI->getUser();
1226 const Type *DestTy = NULL;
1228 /* If shadow use is a int->float cast then insert a second IV
1229 to eliminate this cast.
1231 for (unsigned i = 0; i < n; ++i)
1237 for (unsigned i = 0; i < n; ++i, ++d)
1240 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1241 DestTy = UCast->getDestTy();
1242 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1243 DestTy = SCast->getDestTy();
1244 if (!DestTy) continue;
1247 // If target does not support DestTy natively then do not apply
1248 // this transformation.
1249 EVT DVT = TLI->getValueType(DestTy);
1250 if (!TLI->isTypeLegal(DVT)) continue;
1253 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1255 if (PH->getNumIncomingValues() != 2) continue;
1257 const Type *SrcTy = PH->getType();
1258 int Mantissa = DestTy->getFPMantissaWidth();
1259 if (Mantissa == -1) continue;
1260 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1263 unsigned Entry, Latch;
1264 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1272 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1273 if (!Init) continue;
1274 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1276 BinaryOperator *Incr =
1277 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1278 if (!Incr) continue;
1279 if (Incr->getOpcode() != Instruction::Add
1280 && Incr->getOpcode() != Instruction::Sub)
1283 /* Initialize new IV, double d = 0.0 in above example. */
1284 ConstantInt *C = NULL;
1285 if (Incr->getOperand(0) == PH)
1286 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1287 else if (Incr->getOperand(1) == PH)
1288 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1294 // Ignore negative constants, as the code below doesn't handle them
1295 // correctly. TODO: Remove this restriction.
1296 if (!C->getValue().isStrictlyPositive()) continue;
1298 /* Add new PHINode. */
1299 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1301 /* create new increment. '++d' in above example. */
1302 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1303 BinaryOperator *NewIncr =
1304 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1305 Instruction::FAdd : Instruction::FSub,
1306 NewPH, CFP, "IV.S.next.", Incr);
1308 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1309 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1311 /* Remove cast operation */
1312 ShadowUse->replaceAllUsesWith(NewPH);
1313 ShadowUse->eraseFromParent();
1318 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1319 /// set the IV user and stride information and return true, otherwise return
1321 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond,
1322 IVStrideUse *&CondUse) {
1323 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1324 if (UI->getUser() == Cond) {
1325 // NOTE: we could handle setcc instructions with multiple uses here, but
1326 // InstCombine does it as well for simple uses, it's not clear that it
1327 // occurs enough in real life to handle.
1334 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1335 /// a max computation.
1337 /// This is a narrow solution to a specific, but acute, problem. For loops
1343 /// } while (++i < n);
1345 /// the trip count isn't just 'n', because 'n' might not be positive. And
1346 /// unfortunately this can come up even for loops where the user didn't use
1347 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1348 /// will commonly be lowered like this:
1354 /// } while (++i < n);
1357 /// and then it's possible for subsequent optimization to obscure the if
1358 /// test in such a way that indvars can't find it.
1360 /// When indvars can't find the if test in loops like this, it creates a
1361 /// max expression, which allows it to give the loop a canonical
1362 /// induction variable:
1365 /// max = n < 1 ? 1 : n;
1368 /// } while (++i != max);
1370 /// Canonical induction variables are necessary because the loop passes
1371 /// are designed around them. The most obvious example of this is the
1372 /// LoopInfo analysis, which doesn't remember trip count values. It
1373 /// expects to be able to rediscover the trip count each time it is
1374 /// needed, and it does this using a simple analysis that only succeeds if
1375 /// the loop has a canonical induction variable.
1377 /// However, when it comes time to generate code, the maximum operation
1378 /// can be quite costly, especially if it's inside of an outer loop.
1380 /// This function solves this problem by detecting this type of loop and
1381 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1382 /// the instructions for the maximum computation.
1384 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1385 // Check that the loop matches the pattern we're looking for.
1386 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1387 Cond->getPredicate() != CmpInst::ICMP_NE)
1390 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1391 if (!Sel || !Sel->hasOneUse()) return Cond;
1393 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1394 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1396 const SCEV *One = SE.getIntegerSCEV(1, BackedgeTakenCount->getType());
1398 // Add one to the backedge-taken count to get the trip count.
1399 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1401 // Check for a max calculation that matches the pattern.
1402 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
1404 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
1405 if (Max != SE.getSCEV(Sel)) return Cond;
1407 // To handle a max with more than two operands, this optimization would
1408 // require additional checking and setup.
1409 if (Max->getNumOperands() != 2)
1412 const SCEV *MaxLHS = Max->getOperand(0);
1413 const SCEV *MaxRHS = Max->getOperand(1);
1414 if (!MaxLHS || MaxLHS != One) return Cond;
1415 // Check the relevant induction variable for conformance to
1417 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1418 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1419 if (!AR || !AR->isAffine() ||
1420 AR->getStart() != One ||
1421 AR->getStepRecurrence(SE) != One)
1424 assert(AR->getLoop() == L &&
1425 "Loop condition operand is an addrec in a different loop!");
1427 // Check the right operand of the select, and remember it, as it will
1428 // be used in the new comparison instruction.
1430 if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1431 NewRHS = Sel->getOperand(1);
1432 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1433 NewRHS = Sel->getOperand(2);
1434 if (!NewRHS) return Cond;
1436 // Determine the new comparison opcode. It may be signed or unsigned,
1437 // and the original comparison may be either equality or inequality.
1438 CmpInst::Predicate Pred =
1439 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
1440 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1441 Pred = CmpInst::getInversePredicate(Pred);
1443 // Ok, everything looks ok to change the condition into an SLT or SGE and
1444 // delete the max calculation.
1446 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1448 // Delete the max calculation instructions.
1449 Cond->replaceAllUsesWith(NewCond);
1450 CondUse->setUser(NewCond);
1451 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1452 Cond->eraseFromParent();
1453 Sel->eraseFromParent();
1454 if (Cmp->use_empty())
1455 Cmp->eraseFromParent();
1459 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1460 /// postinc iv when possible.
1462 LSRInstance::OptimizeLoopTermCond() {
1463 SmallPtrSet<Instruction *, 4> PostIncs;
1465 BasicBlock *LatchBlock = L->getLoopLatch();
1466 SmallVector<BasicBlock*, 8> ExitingBlocks;
1467 L->getExitingBlocks(ExitingBlocks);
1469 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1470 BasicBlock *ExitingBlock = ExitingBlocks[i];
1472 // Get the terminating condition for the loop if possible. If we
1473 // can, we want to change it to use a post-incremented version of its
1474 // induction variable, to allow coalescing the live ranges for the IV into
1475 // one register value.
1477 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1480 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1481 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1484 // Search IVUsesByStride to find Cond's IVUse if there is one.
1485 IVStrideUse *CondUse = 0;
1486 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1487 if (!FindIVUserForCond(Cond, CondUse))
1490 // If the trip count is computed in terms of a max (due to ScalarEvolution
1491 // being unable to find a sufficient guard, for example), change the loop
1492 // comparison to use SLT or ULT instead of NE.
1493 // One consequence of doing this now is that it disrupts the count-down
1494 // optimization. That's not always a bad thing though, because in such
1495 // cases it may still be worthwhile to avoid a max.
1496 Cond = OptimizeMax(Cond, CondUse);
1498 // If this exiting block dominates the latch block, it may also use
1499 // the post-inc value if it won't be shared with other uses.
1500 // Check for dominance.
1501 if (!DT.dominates(ExitingBlock, LatchBlock))
1504 // Conservatively avoid trying to use the post-inc value in non-latch
1505 // exits if there may be pre-inc users in intervening blocks.
1506 if (LatchBlock != ExitingBlock)
1507 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1508 // Test if the use is reachable from the exiting block. This dominator
1509 // query is a conservative approximation of reachability.
1510 if (&*UI != CondUse &&
1511 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1512 // Conservatively assume there may be reuse if the quotient of their
1513 // strides could be a legal scale.
1514 const SCEV *A = CondUse->getStride();
1515 const SCEV *B = UI->getStride();
1516 if (SE.getTypeSizeInBits(A->getType()) !=
1517 SE.getTypeSizeInBits(B->getType())) {
1518 if (SE.getTypeSizeInBits(A->getType()) >
1519 SE.getTypeSizeInBits(B->getType()))
1520 B = SE.getSignExtendExpr(B, A->getType());
1522 A = SE.getSignExtendExpr(A, B->getType());
1524 if (const SCEVConstant *D =
1525 dyn_cast_or_null<SCEVConstant>(getSDiv(B, A, SE))) {
1526 // Stride of one or negative one can have reuse with non-addresses.
1527 if (D->getValue()->isOne() ||
1528 D->getValue()->isAllOnesValue())
1529 goto decline_post_inc;
1530 // Avoid weird situations.
1531 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1532 D->getValue()->getValue().isMinSignedValue())
1533 goto decline_post_inc;
1534 // Check for possible scaled-address reuse.
1535 const Type *AccessTy = getAccessType(UI->getUser());
1536 TargetLowering::AddrMode AM;
1537 AM.Scale = D->getValue()->getSExtValue();
1538 if (TLI && TLI->isLegalAddressingMode(AM, AccessTy))
1539 goto decline_post_inc;
1540 AM.Scale = -AM.Scale;
1541 if (TLI && TLI->isLegalAddressingMode(AM, AccessTy))
1542 goto decline_post_inc;
1546 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1549 // It's possible for the setcc instruction to be anywhere in the loop, and
1550 // possible for it to have multiple users. If it is not immediately before
1551 // the exiting block branch, move it.
1552 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1553 if (Cond->hasOneUse()) {
1554 Cond->moveBefore(TermBr);
1556 // Clone the terminating condition and insert into the loopend.
1557 ICmpInst *OldCond = Cond;
1558 Cond = cast<ICmpInst>(Cond->clone());
1559 Cond->setName(L->getHeader()->getName() + ".termcond");
1560 ExitingBlock->getInstList().insert(TermBr, Cond);
1562 // Clone the IVUse, as the old use still exists!
1563 CondUse = &IU.AddUser(CondUse->getStride(), CondUse->getOffset(),
1564 Cond, CondUse->getOperandValToReplace());
1565 TermBr->replaceUsesOfWith(OldCond, Cond);
1569 // If we get to here, we know that we can transform the setcc instruction to
1570 // use the post-incremented version of the IV, allowing us to coalesce the
1571 // live ranges for the IV correctly.
1572 CondUse->setOffset(SE.getMinusSCEV(CondUse->getOffset(),
1573 CondUse->getStride()));
1574 CondUse->setIsUseOfPostIncrementedValue(true);
1577 PostIncs.insert(Cond);
1581 // Determine an insertion point for the loop induction variable increment. It
1582 // must dominate all the post-inc comparisons we just set up, and it must
1583 // dominate the loop latch edge.
1584 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1585 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1586 E = PostIncs.end(); I != E; ++I) {
1588 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1590 if (BB == (*I)->getParent())
1591 IVIncInsertPos = *I;
1592 else if (BB != IVIncInsertPos->getParent())
1593 IVIncInsertPos = BB->getTerminator();
1600 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1601 LSRUse::KindType Kind, const Type *AccessTy) {
1602 int64_t NewMinOffset = LU.MinOffset;
1603 int64_t NewMaxOffset = LU.MaxOffset;
1604 const Type *NewAccessTy = AccessTy;
1606 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1607 // something conservative, however this can pessimize in the case that one of
1608 // the uses will have all its uses outside the loop, for example.
1609 if (LU.Kind != Kind)
1611 // Conservatively assume HasBaseReg is true for now.
1612 if (NewOffset < LU.MinOffset) {
1613 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true,
1614 Kind, AccessTy, TLI, SE))
1616 NewMinOffset = NewOffset;
1617 } else if (NewOffset > LU.MaxOffset) {
1618 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true,
1619 Kind, AccessTy, TLI, SE))
1621 NewMaxOffset = NewOffset;
1623 // Check for a mismatched access type, and fall back conservatively as needed.
1624 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1625 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1628 LU.MinOffset = NewMinOffset;
1629 LU.MaxOffset = NewMaxOffset;
1630 LU.AccessTy = NewAccessTy;
1631 if (NewOffset != LU.Offsets.back())
1632 LU.Offsets.push_back(NewOffset);
1636 /// getUse - Return an LSRUse index and an offset value for a fixup which
1637 /// needs the given expression, with the given kind and optional access type.
1638 /// Either reuse an exisitng use or create a new one, as needed.
1639 std::pair<size_t, int64_t>
1640 LSRInstance::getUse(const SCEV *&Expr,
1641 LSRUse::KindType Kind, const Type *AccessTy) {
1642 const SCEV *Copy = Expr;
1643 int64_t Offset = ExtractImmediate(Expr, SE);
1645 // Basic uses can't accept any offset, for example.
1646 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true,
1647 Kind, AccessTy, TLI, SE)) {
1652 std::pair<UseMapTy::iterator, bool> P =
1653 UseMap.insert(std::make_pair(Expr, 0));
1655 // A use already existed with this base.
1656 size_t LUIdx = P.first->second;
1657 LSRUse &LU = Uses[LUIdx];
1658 if (reconcileNewOffset(LU, Offset, Kind, AccessTy))
1660 return std::make_pair(LUIdx, Offset);
1663 // Create a new use.
1664 size_t LUIdx = Uses.size();
1665 P.first->second = LUIdx;
1666 Uses.push_back(LSRUse(Kind, AccessTy));
1667 LSRUse &LU = Uses[LUIdx];
1669 // We don't need to track redundant offsets, but we don't need to go out
1670 // of our way here to avoid them.
1671 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1672 LU.Offsets.push_back(Offset);
1674 LU.MinOffset = Offset;
1675 LU.MaxOffset = Offset;
1676 return std::make_pair(LUIdx, Offset);
1679 void LSRInstance::CollectInterestingTypesAndFactors() {
1680 SmallSetVector<const SCEV *, 4> Strides;
1682 // Collect interesting types and factors.
1683 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1684 const SCEV *Stride = UI->getStride();
1686 // Collect interesting types.
1687 Types.insert(SE.getEffectiveSCEVType(Stride->getType()));
1689 // Collect interesting factors.
1690 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1691 Strides.begin(), SEnd = Strides.end(); NewStrideIter != SEnd;
1693 const SCEV *OldStride = Stride;
1694 const SCEV *NewStride = *NewStrideIter;
1695 if (OldStride == NewStride)
1698 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1699 SE.getTypeSizeInBits(NewStride->getType())) {
1700 if (SE.getTypeSizeInBits(OldStride->getType()) >
1701 SE.getTypeSizeInBits(NewStride->getType()))
1702 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1704 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1706 if (const SCEVConstant *Factor =
1707 dyn_cast_or_null<SCEVConstant>(getSDiv(NewStride, OldStride,
1709 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1710 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1711 } else if (const SCEVConstant *Factor =
1712 dyn_cast_or_null<SCEVConstant>(getSDiv(OldStride, NewStride,
1714 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1715 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1718 Strides.insert(Stride);
1721 // If all uses use the same type, don't bother looking for truncation-based
1723 if (Types.size() == 1)
1726 DEBUG(print_factors_and_types(dbgs()));
1729 void LSRInstance::CollectFixupsAndInitialFormulae() {
1730 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1732 LSRFixup &LF = getNewFixup();
1733 LF.UserInst = UI->getUser();
1734 LF.OperandValToReplace = UI->getOperandValToReplace();
1735 if (UI->isUseOfPostIncrementedValue())
1738 LSRUse::KindType Kind = LSRUse::Basic;
1739 const Type *AccessTy = 0;
1740 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1741 Kind = LSRUse::Address;
1742 AccessTy = getAccessType(LF.UserInst);
1745 const SCEV *S = IU.getCanonicalExpr(*UI);
1747 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1748 // (N - i == 0), and this allows (N - i) to be the expression that we work
1749 // with rather than just N or i, so we can consider the register
1750 // requirements for both N and i at the same time. Limiting this code to
1751 // equality icmps is not a problem because all interesting loops use
1752 // equality icmps, thanks to IndVarSimplify.
1753 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1754 if (CI->isEquality()) {
1755 // Swap the operands if needed to put the OperandValToReplace on the
1756 // left, for consistency.
1757 Value *NV = CI->getOperand(1);
1758 if (NV == LF.OperandValToReplace) {
1759 CI->setOperand(1, CI->getOperand(0));
1760 CI->setOperand(0, NV);
1763 // x == y --> x - y == 0
1764 const SCEV *N = SE.getSCEV(NV);
1765 if (N->isLoopInvariant(L)) {
1766 Kind = LSRUse::ICmpZero;
1767 S = SE.getMinusSCEV(N, S);
1770 // -1 and the negations of all interesting strides (except the negation
1771 // of -1) are now also interesting.
1772 for (size_t i = 0, e = Factors.size(); i != e; ++i)
1773 if (Factors[i] != -1)
1774 Factors.insert(-(uint64_t)Factors[i]);
1778 // Set up the initial formula for this use.
1779 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
1781 LF.Offset = P.second;
1782 LSRUse &LU = Uses[LF.LUIdx];
1783 LU.AllFixupsOutsideLoop &= !L->contains(LF.UserInst);
1785 // If this is the first use of this LSRUse, give it a formula.
1786 if (LU.Formulae.empty()) {
1787 InsertInitialFormula(S, L, LU, LF.LUIdx);
1788 CountRegisters(LU.Formulae.back(), LF.LUIdx);
1792 DEBUG(print_fixups(dbgs()));
1796 LSRInstance::InsertInitialFormula(const SCEV *S, Loop *L,
1797 LSRUse &LU, size_t LUIdx) {
1799 F.InitialMatch(S, L, SE, DT);
1800 bool Inserted = InsertFormula(LU, LUIdx, F);
1801 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
1805 LSRInstance::InsertSupplementalFormula(const SCEV *S,
1806 LSRUse &LU, size_t LUIdx) {
1808 F.BaseRegs.push_back(S);
1809 F.AM.HasBaseReg = true;
1810 bool Inserted = InsertFormula(LU, LUIdx, F);
1811 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
1814 /// CountRegisters - Note which registers are used by the given formula,
1815 /// updating RegUses.
1816 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
1818 RegUses.CountRegister(F.ScaledReg, LUIdx);
1819 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1820 E = F.BaseRegs.end(); I != E; ++I)
1821 RegUses.CountRegister(*I, LUIdx);
1824 /// InsertFormula - If the given formula has not yet been inserted, add it to
1825 /// the list, and return true. Return false otherwise.
1826 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
1827 if (!LU.InsertFormula(LUIdx, F))
1830 CountRegisters(F, LUIdx);
1834 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
1835 /// loop-invariant values which we're tracking. These other uses will pin these
1836 /// values in registers, making them less profitable for elimination.
1837 /// TODO: This currently misses non-constant addrec step registers.
1838 /// TODO: Should this give more weight to users inside the loop?
1840 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
1841 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
1842 SmallPtrSet<const SCEV *, 8> Inserted;
1844 while (!Worklist.empty()) {
1845 const SCEV *S = Worklist.pop_back_val();
1847 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
1848 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
1849 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1850 Worklist.push_back(C->getOperand());
1851 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1852 Worklist.push_back(D->getLHS());
1853 Worklist.push_back(D->getRHS());
1854 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
1855 if (!Inserted.insert(U)) continue;
1856 const Value *V = U->getValue();
1857 if (const Instruction *Inst = dyn_cast<Instruction>(V))
1858 if (L->contains(Inst)) continue;
1859 for (Value::use_const_iterator UI = V->use_begin(), UE = V->use_end();
1861 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
1862 // Ignore non-instructions.
1865 // Ignore instructions in other functions (as can happen with
1867 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
1869 // Ignore instructions not dominated by the loop.
1870 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
1871 UserInst->getParent() :
1872 cast<PHINode>(UserInst)->getIncomingBlock(
1873 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
1874 if (!DT.dominates(L->getHeader(), UseBB))
1876 // Ignore uses which are part of other SCEV expressions, to avoid
1877 // analyzing them multiple times.
1878 if (SE.isSCEVable(UserInst->getType()) &&
1879 !isa<SCEVUnknown>(SE.getSCEV(const_cast<Instruction *>(UserInst))))
1881 // Ignore icmp instructions which are already being analyzed.
1882 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
1883 unsigned OtherIdx = !UI.getOperandNo();
1884 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
1885 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
1889 LSRFixup &LF = getNewFixup();
1890 LF.UserInst = const_cast<Instruction *>(UserInst);
1891 LF.OperandValToReplace = UI.getUse();
1892 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
1894 LF.Offset = P.second;
1895 LSRUse &LU = Uses[LF.LUIdx];
1896 LU.AllFixupsOutsideLoop &= L->contains(LF.UserInst);
1897 InsertSupplementalFormula(U, LU, LF.LUIdx);
1898 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
1905 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
1906 /// separate registers. If C is non-null, multiply each subexpression by C.
1907 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
1908 SmallVectorImpl<const SCEV *> &Ops,
1909 ScalarEvolution &SE) {
1910 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1911 // Break out add operands.
1912 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1914 CollectSubexprs(*I, C, Ops, SE);
1916 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1917 // Split a non-zero base out of an addrec.
1918 if (!AR->getStart()->isZero()) {
1919 CollectSubexprs(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
1920 AR->getStepRecurrence(SE),
1921 AR->getLoop()), C, Ops, SE);
1922 CollectSubexprs(AR->getStart(), C, Ops, SE);
1925 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
1926 // Break (C * (a + b + c)) into C*a + C*b + C*c.
1927 if (Mul->getNumOperands() == 2)
1928 if (const SCEVConstant *Op0 =
1929 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
1930 CollectSubexprs(Mul->getOperand(1),
1931 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
1937 // Otherwise use the value itself.
1938 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
1941 /// GenerateReassociations - Split out subexpressions from adds and the bases of
1943 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
1946 // Arbitrarily cap recursion to protect compile time.
1947 if (Depth >= 3) return;
1949 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
1950 const SCEV *BaseReg = Base.BaseRegs[i];
1952 SmallVector<const SCEV *, 8> AddOps;
1953 CollectSubexprs(BaseReg, 0, AddOps, SE);
1954 if (AddOps.size() == 1) continue;
1956 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
1957 JE = AddOps.end(); J != JE; ++J) {
1958 // Don't pull a constant into a register if the constant could be folded
1959 // into an immediate field.
1960 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
1961 Base.getNumRegs() > 1,
1962 LU.Kind, LU.AccessTy, TLI, SE))
1965 // Collect all operands except *J.
1966 SmallVector<const SCEV *, 8> InnerAddOps;
1967 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
1968 KE = AddOps.end(); K != KE; ++K)
1970 InnerAddOps.push_back(*K);
1972 // Don't leave just a constant behind in a register if the constant could
1973 // be folded into an immediate field.
1974 if (InnerAddOps.size() == 1 &&
1975 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
1976 Base.getNumRegs() > 1,
1977 LU.Kind, LU.AccessTy, TLI, SE))
1981 F.BaseRegs[i] = SE.getAddExpr(InnerAddOps);
1982 F.BaseRegs.push_back(*J);
1983 if (InsertFormula(LU, LUIdx, F))
1984 // If that formula hadn't been seen before, recurse to find more like
1986 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
1991 /// GenerateCombinations - Generate a formula consisting of all of the
1992 /// loop-dominating registers added into a single register.
1993 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
1995 // This method is only intersting on a plurality of registers.
1996 if (Base.BaseRegs.size() <= 1) return;
2000 SmallVector<const SCEV *, 4> Ops;
2001 for (SmallVectorImpl<const SCEV *>::const_iterator
2002 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2003 const SCEV *BaseReg = *I;
2004 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2005 !BaseReg->hasComputableLoopEvolution(L))
2006 Ops.push_back(BaseReg);
2008 F.BaseRegs.push_back(BaseReg);
2010 if (Ops.size() > 1) {
2011 F.BaseRegs.push_back(SE.getAddExpr(Ops));
2012 (void)InsertFormula(LU, LUIdx, F);
2016 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2017 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2019 // We can't add a symbolic offset if the address already contains one.
2020 if (Base.AM.BaseGV) return;
2022 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2023 const SCEV *G = Base.BaseRegs[i];
2024 GlobalValue *GV = ExtractSymbol(G, SE);
2025 if (G->isZero() || !GV)
2029 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2030 LU.Kind, LU.AccessTy, TLI))
2033 (void)InsertFormula(LU, LUIdx, F);
2037 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2038 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2040 // TODO: For now, just add the min and max offset, because it usually isn't
2041 // worthwhile looking at everything inbetween.
2042 SmallVector<int64_t, 4> Worklist;
2043 Worklist.push_back(LU.MinOffset);
2044 if (LU.MaxOffset != LU.MinOffset)
2045 Worklist.push_back(LU.MaxOffset);
2047 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2048 const SCEV *G = Base.BaseRegs[i];
2050 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2051 E = Worklist.end(); I != E; ++I) {
2053 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2054 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2055 LU.Kind, LU.AccessTy, TLI)) {
2056 F.BaseRegs[i] = SE.getAddExpr(G, SE.getIntegerSCEV(*I, G->getType()));
2058 (void)InsertFormula(LU, LUIdx, F);
2062 int64_t Imm = ExtractImmediate(G, SE);
2063 if (G->isZero() || Imm == 0)
2066 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2067 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2068 LU.Kind, LU.AccessTy, TLI))
2071 (void)InsertFormula(LU, LUIdx, F);
2075 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2076 /// the comparison. For example, x == y -> x*c == y*c.
2077 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2079 if (LU.Kind != LSRUse::ICmpZero) return;
2081 // Determine the integer type for the base formula.
2082 const Type *IntTy = Base.getType();
2084 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2086 // Don't do this if there is more than one offset.
2087 if (LU.MinOffset != LU.MaxOffset) return;
2089 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2091 // Check each interesting stride.
2092 for (SmallSetVector<int64_t, 8>::const_iterator
2093 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2094 int64_t Factor = *I;
2097 // Check that the multiplication doesn't overflow.
2098 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2099 if ((int64_t)F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2102 // Check that multiplying with the use offset doesn't overflow.
2103 int64_t Offset = LU.MinOffset;
2104 Offset = (uint64_t)Offset * Factor;
2105 if ((int64_t)Offset / Factor != LU.MinOffset)
2108 // Check that this scale is legal.
2109 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2112 // Compensate for the use having MinOffset built into it.
2113 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2115 const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
2117 // Check that multiplying with each base register doesn't overflow.
2118 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2119 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2120 if (getSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2124 // Check that multiplying with the scaled register doesn't overflow.
2126 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2127 if (getSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2131 // If we make it here and it's legal, add it.
2132 (void)InsertFormula(LU, LUIdx, F);
2137 /// GenerateScales - Generate stride factor reuse formulae by making use of
2138 /// scaled-offset address modes, for example.
2139 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
2141 // Determine the integer type for the base formula.
2142 const Type *IntTy = Base.getType();
2145 // If this Formula already has a scaled register, we can't add another one.
2146 if (Base.AM.Scale != 0) return;
2148 // Check each interesting stride.
2149 for (SmallSetVector<int64_t, 8>::const_iterator
2150 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2151 int64_t Factor = *I;
2153 Base.AM.Scale = Factor;
2154 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2155 // Check whether this scale is going to be legal.
2156 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2157 LU.Kind, LU.AccessTy, TLI)) {
2158 // As a special-case, handle special out-of-loop Basic users specially.
2159 // TODO: Reconsider this special case.
2160 if (LU.Kind == LSRUse::Basic &&
2161 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2162 LSRUse::Special, LU.AccessTy, TLI) &&
2163 LU.AllFixupsOutsideLoop)
2164 LU.Kind = LSRUse::Special;
2168 // For an ICmpZero, negating a solitary base register won't lead to
2170 if (LU.Kind == LSRUse::ICmpZero &&
2171 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2173 // For each addrec base reg, apply the scale, if possible.
2174 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2175 if (const SCEVAddRecExpr *AR =
2176 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2177 const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
2178 if (FactorS->isZero())
2180 // Divide out the factor, ignoring high bits, since we'll be
2181 // scaling the value back up in the end.
2182 if (const SCEV *Quotient = getSDiv(AR, FactorS, SE, true)) {
2183 // TODO: This could be optimized to avoid all the copying.
2185 F.ScaledReg = Quotient;
2186 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2187 F.BaseRegs.pop_back();
2188 (void)InsertFormula(LU, LUIdx, F);
2194 /// GenerateTruncates - Generate reuse formulae from different IV types.
2195 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
2197 // This requires TargetLowering to tell us which truncates are free.
2200 // Don't bother truncating symbolic values.
2201 if (Base.AM.BaseGV) return;
2203 // Determine the integer type for the base formula.
2204 const Type *DstTy = Base.getType();
2206 DstTy = SE.getEffectiveSCEVType(DstTy);
2208 for (SmallSetVector<const Type *, 4>::const_iterator
2209 I = Types.begin(), E = Types.end(); I != E; ++I) {
2210 const Type *SrcTy = *I;
2211 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2214 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2215 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2216 JE = F.BaseRegs.end(); J != JE; ++J)
2217 *J = SE.getAnyExtendExpr(*J, SrcTy);
2219 // TODO: This assumes we've done basic processing on all uses and
2220 // have an idea what the register usage is.
2221 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2224 (void)InsertFormula(LU, LUIdx, F);
2231 /// WorkItem - Helper class for GenerateConstantOffsetReuse. It's used to
2232 /// defer modifications so that the search phase doesn't have to worry about
2233 /// the data structures moving underneath it.
2237 const SCEV *OrigReg;
2239 WorkItem(size_t LI, int64_t I, const SCEV *R)
2240 : LUIdx(LI), Imm(I), OrigReg(R) {}
2242 void print(raw_ostream &OS) const;
2248 void WorkItem::print(raw_ostream &OS) const {
2249 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2250 << " , add offset " << Imm;
2253 void WorkItem::dump() const {
2254 print(errs()); errs() << '\n';
2257 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2258 /// distance apart and try to form reuse opportunities between them.
2259 void LSRInstance::GenerateCrossUseConstantOffsets() {
2260 // Group the registers by their value without any added constant offset.
2261 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2262 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2264 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2265 SmallVector<const SCEV *, 8> Sequence;
2266 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2268 const SCEV *Reg = *I;
2269 int64_t Imm = ExtractImmediate(Reg, SE);
2270 std::pair<RegMapTy::iterator, bool> Pair =
2271 Map.insert(std::make_pair(Reg, ImmMapTy()));
2273 Sequence.push_back(Reg);
2274 Pair.first->second.insert(std::make_pair(Imm, *I));
2275 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2278 // Now examine each set of registers with the same base value. Build up
2279 // a list of work to do and do the work in a separate step so that we're
2280 // not adding formulae and register counts while we're searching.
2281 SmallVector<WorkItem, 32> WorkItems;
2282 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2283 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2284 E = Sequence.end(); I != E; ++I) {
2285 const SCEV *Reg = *I;
2286 const ImmMapTy &Imms = Map.find(Reg)->second;
2288 // It's not worthwhile looking for reuse if there's only one offset.
2289 if (Imms.size() == 1)
2292 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2293 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2295 dbgs() << ' ' << J->first;
2298 // Examine each offset.
2299 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2301 const SCEV *OrigReg = J->second;
2303 int64_t JImm = J->first;
2304 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2306 if (!isa<SCEVConstant>(OrigReg) &&
2307 UsedByIndicesMap[Reg].count() == 1) {
2308 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2312 // Conservatively examine offsets between this orig reg a few selected
2314 ImmMapTy::const_iterator OtherImms[] = {
2315 Imms.begin(), prior(Imms.end()),
2316 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2318 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2319 ImmMapTy::const_iterator M = OtherImms[i];
2320 if (M == J || M == JE) continue;
2322 // Compute the difference between the two.
2323 int64_t Imm = (uint64_t)JImm - M->first;
2324 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2325 LUIdx = UsedByIndices.find_next(LUIdx))
2326 // Make a memo of this use, offset, and register tuple.
2327 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2328 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2335 UsedByIndicesMap.clear();
2336 UniqueItems.clear();
2338 // Now iterate through the worklist and add new formulae.
2339 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2340 E = WorkItems.end(); I != E; ++I) {
2341 const WorkItem &WI = *I;
2342 size_t LUIdx = WI.LUIdx;
2343 LSRUse &LU = Uses[LUIdx];
2344 int64_t Imm = WI.Imm;
2345 const SCEV *OrigReg = WI.OrigReg;
2347 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2348 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2349 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2351 // TODO: Use a more targetted data structure.
2352 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2353 Formula F = LU.Formulae[L];
2354 // Use the immediate in the scaled register.
2355 if (F.ScaledReg == OrigReg) {
2356 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2357 Imm * (uint64_t)F.AM.Scale;
2358 // Don't create 50 + reg(-50).
2359 if (F.referencesReg(SE.getSCEV(
2360 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2363 NewF.AM.BaseOffs = Offs;
2364 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2365 LU.Kind, LU.AccessTy, TLI))
2367 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2369 // If the new scale is a constant in a register, and adding the constant
2370 // value to the immediate would produce a value closer to zero than the
2371 // immediate itself, then the formula isn't worthwhile.
2372 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2373 if (C->getValue()->getValue().isNegative() !=
2374 (NewF.AM.BaseOffs < 0) &&
2375 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2376 .ule(APInt(BitWidth, NewF.AM.BaseOffs).abs()))
2380 (void)InsertFormula(LU, LUIdx, NewF);
2382 // Use the immediate in a base register.
2383 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2384 const SCEV *BaseReg = F.BaseRegs[N];
2385 if (BaseReg != OrigReg)
2388 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2389 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2390 LU.Kind, LU.AccessTy, TLI))
2392 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2394 // If the new formula has a constant in a register, and adding the
2395 // constant value to the immediate would produce a value closer to
2396 // zero than the immediate itself, then the formula isn't worthwhile.
2397 for (SmallVectorImpl<const SCEV *>::const_iterator
2398 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2400 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2401 if (C->getValue()->getValue().isNegative() !=
2402 (NewF.AM.BaseOffs < 0) &&
2403 C->getValue()->getValue().abs()
2404 .ule(APInt(BitWidth, NewF.AM.BaseOffs).abs()))
2408 (void)InsertFormula(LU, LUIdx, NewF);
2417 /// GenerateAllReuseFormulae - Generate formulae for each use.
2419 LSRInstance::GenerateAllReuseFormulae() {
2420 // This is split into two loops so that hasRegsUsedByUsesOtherThan
2421 // queries are more precise.
2422 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2423 LSRUse &LU = Uses[LUIdx];
2424 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2425 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2426 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2427 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2429 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2430 LSRUse &LU = Uses[LUIdx];
2431 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2432 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2433 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2434 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2435 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2436 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2437 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2438 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2439 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2440 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2443 GenerateCrossUseConstantOffsets();
2446 /// If their are multiple formulae with the same set of registers used
2447 /// by other uses, pick the best one and delete the others.
2448 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2450 bool Changed = false;
2453 // Collect the best formula for each unique set of shared registers. This
2454 // is reset for each use.
2455 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2457 BestFormulaeTy BestFormulae;
2459 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2460 LSRUse &LU = Uses[LUIdx];
2461 FormulaSorter Sorter(L, LU, SE, DT);
2463 // Clear out the set of used regs; it will be recomputed.
2466 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2467 FIdx != NumForms; ++FIdx) {
2468 Formula &F = LU.Formulae[FIdx];
2470 SmallVector<const SCEV *, 2> Key;
2471 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2472 JE = F.BaseRegs.end(); J != JE; ++J) {
2473 const SCEV *Reg = *J;
2474 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2478 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2479 Key.push_back(F.ScaledReg);
2480 // Unstable sort by host order ok, because this is only used for
2482 std::sort(Key.begin(), Key.end());
2484 std::pair<BestFormulaeTy::const_iterator, bool> P =
2485 BestFormulae.insert(std::make_pair(Key, FIdx));
2487 Formula &Best = LU.Formulae[P.first->second];
2488 if (Sorter.operator()(F, Best))
2490 DEBUG(dbgs() << "Filtering out "; F.print(dbgs());
2492 " in favor of "; Best.print(dbgs());
2497 std::swap(F, LU.Formulae.back());
2498 LU.Formulae.pop_back();
2503 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2504 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2506 BestFormulae.clear();
2509 DEBUG(if (Changed) {
2511 "After filtering out undesirable candidates:\n";
2516 /// NarrowSearchSpaceUsingHeuristics - If there are an extrordinary number of
2517 /// formulae to choose from, use some rough heuristics to prune down the number
2518 /// of formulae. This keeps the main solver from taking an extrordinary amount
2519 /// of time in some worst-case scenarios.
2520 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2521 // This is a rough guess that seems to work fairly well.
2522 const size_t Limit = UINT16_MAX;
2524 SmallPtrSet<const SCEV *, 4> Taken;
2526 // Estimate the worst-case number of solutions we might consider. We almost
2527 // never consider this many solutions because we prune the search space,
2528 // but the pruning isn't always sufficient.
2530 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2531 E = Uses.end(); I != E; ++I) {
2532 size_t FSize = I->Formulae.size();
2533 if (FSize >= Limit) {
2544 // Ok, we have too many of formulae on our hands to conveniently handle.
2545 // Use a rough heuristic to thin out the list.
2547 // Pick the register which is used by the most LSRUses, which is likely
2548 // to be a good reuse register candidate.
2549 const SCEV *Best = 0;
2550 unsigned BestNum = 0;
2551 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2553 const SCEV *Reg = *I;
2554 if (Taken.count(Reg))
2559 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2560 if (Count > BestNum) {
2567 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2568 << " will yeild profitable reuse.\n");
2571 // In any use with formulae which references this register, delete formulae
2572 // which don't reference it.
2573 for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(),
2574 E = Uses.end(); I != E; ++I) {
2576 if (!LU.Regs.count(Best)) continue;
2578 // Clear out the set of used regs; it will be recomputed.
2581 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2582 Formula &F = LU.Formulae[i];
2583 if (!F.referencesReg(Best)) {
2584 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2585 std::swap(LU.Formulae.back(), F);
2586 LU.Formulae.pop_back();
2592 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2593 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2597 DEBUG(dbgs() << "After pre-selection:\n";
2598 print_uses(dbgs()));
2602 /// SolveRecurse - This is the recursive solver.
2603 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2605 SmallVectorImpl<const Formula *> &Workspace,
2606 const Cost &CurCost,
2607 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2608 DenseSet<const SCEV *> &VisitedRegs) const {
2611 // - use more aggressive filtering
2612 // - sort the formula so that the most profitable solutions are found first
2613 // - sort the uses too
2615 // - dont compute a cost, and then compare. compare while computing a cost
2617 // - track register sets with SmallBitVector
2619 const LSRUse &LU = Uses[Workspace.size()];
2621 // If this use references any register that's already a part of the
2622 // in-progress solution, consider it a requirement that a formula must
2623 // reference that register in order to be considered. This prunes out
2624 // unprofitable searching.
2625 SmallSetVector<const SCEV *, 4> ReqRegs;
2626 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
2627 E = CurRegs.end(); I != E; ++I)
2628 if (LU.Regs.count(*I))
2631 bool AnySatisfiedReqRegs = false;
2632 SmallPtrSet<const SCEV *, 16> NewRegs;
2635 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2636 E = LU.Formulae.end(); I != E; ++I) {
2637 const Formula &F = *I;
2639 // Ignore formulae which do not use any of the required registers.
2640 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
2641 JE = ReqRegs.end(); J != JE; ++J) {
2642 const SCEV *Reg = *J;
2643 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
2644 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
2648 AnySatisfiedReqRegs = true;
2650 // Evaluate the cost of the current formula. If it's already worse than
2651 // the current best, prune the search at that point.
2654 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
2655 if (NewCost < SolutionCost) {
2656 Workspace.push_back(&F);
2657 if (Workspace.size() != Uses.size()) {
2658 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
2659 NewRegs, VisitedRegs);
2660 if (F.getNumRegs() == 1 && Workspace.size() == 1)
2661 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
2663 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
2664 dbgs() << ". Regs:";
2665 for (SmallPtrSet<const SCEV *, 16>::const_iterator
2666 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
2667 dbgs() << ' ' << **I;
2670 SolutionCost = NewCost;
2671 Solution = Workspace;
2673 Workspace.pop_back();
2678 // If none of the formulae had all of the required registers, relax the
2679 // constraint so that we don't exclude all formulae.
2680 if (!AnySatisfiedReqRegs) {
2686 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
2687 SmallVector<const Formula *, 8> Workspace;
2689 SolutionCost.Loose();
2691 SmallPtrSet<const SCEV *, 16> CurRegs;
2692 DenseSet<const SCEV *> VisitedRegs;
2693 Workspace.reserve(Uses.size());
2695 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
2696 CurRegs, VisitedRegs);
2698 // Ok, we've now made all our decisions.
2699 DEBUG(dbgs() << "\n"
2700 "The chosen solution requires "; SolutionCost.print(dbgs());
2702 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
2704 Uses[i].print(dbgs());
2707 Solution[i]->print(dbgs());
2712 /// getImmediateDominator - A handy utility for the specific DominatorTree
2713 /// query that we need here.
2715 static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
2716 DomTreeNode *Node = DT.getNode(BB);
2717 if (!Node) return 0;
2718 Node = Node->getIDom();
2719 if (!Node) return 0;
2720 return Node->getBlock();
2723 Value *LSRInstance::Expand(const LSRFixup &LF,
2725 BasicBlock::iterator IP,
2726 Loop *L, Instruction *IVIncInsertPos,
2727 SCEVExpander &Rewriter,
2728 SmallVectorImpl<WeakVH> &DeadInsts,
2729 ScalarEvolution &SE, DominatorTree &DT) const {
2730 const LSRUse &LU = Uses[LF.LUIdx];
2732 // Then, collect some instructions which we will remain dominated by when
2733 // expanding the replacement. These must be dominated by any operands that
2734 // will be required in the expansion.
2735 SmallVector<Instruction *, 4> Inputs;
2736 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
2737 Inputs.push_back(I);
2738 if (LU.Kind == LSRUse::ICmpZero)
2739 if (Instruction *I =
2740 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
2741 Inputs.push_back(I);
2742 if (LF.PostIncLoop && !L->contains(LF.UserInst))
2743 Inputs.push_back(L->getLoopLatch()->getTerminator());
2745 // Then, climb up the immediate dominator tree as far as we can go while
2746 // still being dominated by the input positions.
2748 bool AllDominate = true;
2749 Instruction *BetterPos = 0;
2750 BasicBlock *IDom = getImmediateDominator(IP->getParent(), DT);
2752 Instruction *Tentative = IDom->getTerminator();
2753 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
2754 E = Inputs.end(); I != E; ++I) {
2755 Instruction *Inst = *I;
2756 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
2757 AllDominate = false;
2760 if (IDom == Inst->getParent() &&
2761 (!BetterPos || DT.dominates(BetterPos, Inst)))
2762 BetterPos = next(BasicBlock::iterator(Inst));
2771 while (isa<PHINode>(IP)) ++IP;
2773 // Inform the Rewriter if we have a post-increment use, so that it can
2774 // perform an advantageous expansion.
2775 Rewriter.setPostInc(LF.PostIncLoop);
2777 // This is the type that the user actually needs.
2778 const Type *OpTy = LF.OperandValToReplace->getType();
2779 // This will be the type that we'll initially expand to.
2780 const Type *Ty = F.getType();
2782 // No type known; just expand directly to the ultimate type.
2784 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
2785 // Expand directly to the ultimate type if it's the right size.
2787 // This is the type to do integer arithmetic in.
2788 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
2790 // Build up a list of operands to add together to form the full base.
2791 SmallVector<const SCEV *, 8> Ops;
2793 // Expand the BaseRegs portion.
2794 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2795 E = F.BaseRegs.end(); I != E; ++I) {
2796 const SCEV *Reg = *I;
2797 assert(!Reg->isZero() && "Zero allocated in a base register!");
2799 // If we're expanding for a post-inc user for the add-rec's loop, make the
2800 // post-inc adjustment.
2801 const SCEV *Start = Reg;
2802 while (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Start)) {
2803 if (AR->getLoop() == LF.PostIncLoop) {
2804 Reg = SE.getAddExpr(Reg, AR->getStepRecurrence(SE));
2805 // If the user is inside the loop, insert the code after the increment
2806 // so that it is dominated by its operand.
2807 if (L->contains(LF.UserInst))
2808 IP = IVIncInsertPos;
2811 Start = AR->getStart();
2814 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
2817 // Expand the ScaledReg portion.
2818 Value *ICmpScaledV = 0;
2819 if (F.AM.Scale != 0) {
2820 const SCEV *ScaledS = F.ScaledReg;
2822 // If we're expanding for a post-inc user for the add-rec's loop, make the
2823 // post-inc adjustment.
2824 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(ScaledS))
2825 if (AR->getLoop() == LF.PostIncLoop)
2826 ScaledS = SE.getAddExpr(ScaledS, AR->getStepRecurrence(SE));
2828 if (LU.Kind == LSRUse::ICmpZero) {
2829 // An interesting way of "folding" with an icmp is to use a negated
2830 // scale, which we'll implement by inserting it into the other operand
2832 assert(F.AM.Scale == -1 &&
2833 "The only scale supported by ICmpZero uses is -1!");
2834 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
2836 // Otherwise just expand the scaled register and an explicit scale,
2837 // which is expected to be matched as part of the address.
2838 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
2839 ScaledS = SE.getMulExpr(ScaledS,
2840 SE.getIntegerSCEV(F.AM.Scale,
2841 ScaledS->getType()));
2842 Ops.push_back(ScaledS);
2846 // Expand the immediate portions.
2848 Ops.push_back(SE.getSCEV(F.AM.BaseGV));
2849 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
2851 if (LU.Kind == LSRUse::ICmpZero) {
2852 // The other interesting way of "folding" with an ICmpZero is to use a
2853 // negated immediate.
2855 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
2857 Ops.push_back(SE.getUnknown(ICmpScaledV));
2858 ICmpScaledV = ConstantInt::get(IntTy, Offset);
2861 // Just add the immediate values. These again are expected to be matched
2862 // as part of the address.
2863 Ops.push_back(SE.getIntegerSCEV(Offset, IntTy));
2867 // Emit instructions summing all the operands.
2868 const SCEV *FullS = Ops.empty() ?
2869 SE.getIntegerSCEV(0, IntTy) :
2871 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
2873 // We're done expanding now, so reset the rewriter.
2874 Rewriter.setPostInc(0);
2876 // An ICmpZero Formula represents an ICmp which we're handling as a
2877 // comparison against zero. Now that we've expanded an expression for that
2878 // form, update the ICmp's other operand.
2879 if (LU.Kind == LSRUse::ICmpZero) {
2880 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
2881 DeadInsts.push_back(CI->getOperand(1));
2882 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
2883 "a scale at the same time!");
2884 if (F.AM.Scale == -1) {
2885 if (ICmpScaledV->getType() != OpTy) {
2887 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
2889 ICmpScaledV, OpTy, "tmp", CI);
2892 CI->setOperand(1, ICmpScaledV);
2894 assert(F.AM.Scale == 0 &&
2895 "ICmp does not support folding a global value and "
2896 "a scale at the same time!");
2897 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
2899 if (C->getType() != OpTy)
2900 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
2904 CI->setOperand(1, C);
2911 /// Rewrite - Emit instructions for the leading candidate expression for this
2912 /// LSRUse (this is called "expanding"), and update the UserInst to reference
2913 /// the newly expanded value.
2914 void LSRInstance::Rewrite(const LSRFixup &LF,
2916 Loop *L, Instruction *IVIncInsertPos,
2917 SCEVExpander &Rewriter,
2918 SmallVectorImpl<WeakVH> &DeadInsts,
2919 ScalarEvolution &SE, DominatorTree &DT,
2921 const Type *OpTy = LF.OperandValToReplace->getType();
2923 // First, find an insertion point that dominates UserInst. For PHI nodes,
2924 // find the nearest block which dominates all the relevant uses.
2925 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
2926 DenseMap<BasicBlock *, Value *> Inserted;
2927 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2928 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
2929 BasicBlock *BB = PN->getIncomingBlock(i);
2931 // If this is a critical edge, split the edge so that we do not insert
2932 // the code on all predecessor/successor paths. We do this unless this
2933 // is the canonical backedge for this loop, which complicates post-inc
2935 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
2936 !isa<IndirectBrInst>(BB->getTerminator()) &&
2937 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
2938 // Split the critical edge.
2939 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
2941 // If PN is outside of the loop and BB is in the loop, we want to
2942 // move the block to be immediately before the PHI block, not
2943 // immediately after BB.
2944 if (L->contains(BB) && !L->contains(PN))
2945 NewBB->moveBefore(PN->getParent());
2947 // Splitting the edge can reduce the number of PHI entries we have.
2948 e = PN->getNumIncomingValues();
2950 i = PN->getBasicBlockIndex(BB);
2953 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
2954 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
2956 PN->setIncomingValue(i, Pair.first->second);
2958 Value *FullV = Expand(LF, F, BB->getTerminator(), L, IVIncInsertPos,
2959 Rewriter, DeadInsts, SE, DT);
2961 // If this is reuse-by-noop-cast, insert the noop cast.
2962 if (FullV->getType() != OpTy)
2964 CastInst::Create(CastInst::getCastOpcode(FullV, false,
2966 FullV, LF.OperandValToReplace->getType(),
2967 "tmp", BB->getTerminator());
2969 PN->setIncomingValue(i, FullV);
2970 Pair.first->second = FullV;
2974 Value *FullV = Expand(LF, F, LF.UserInst, L, IVIncInsertPos,
2975 Rewriter, DeadInsts, SE, DT);
2977 // If this is reuse-by-noop-cast, insert the noop cast.
2978 if (FullV->getType() != OpTy) {
2980 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
2981 FullV, OpTy, "tmp", LF.UserInst);
2985 // Update the user. ICmpZero is handled specially here (for now) because
2986 // Expand may have updated one of the operands of the icmp already, and
2987 // its new value may happen to be equal to LF.OperandValToReplace, in
2988 // which case doing replaceUsesOfWith leads to replacing both operands
2989 // with the same value. TODO: Reorganize this.
2990 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
2991 LF.UserInst->setOperand(0, FullV);
2993 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
2996 DeadInsts.push_back(LF.OperandValToReplace);
3000 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3002 // Keep track of instructions we may have made dead, so that
3003 // we can remove them after we are done working.
3004 SmallVector<WeakVH, 16> DeadInsts;
3006 SCEVExpander Rewriter(SE);
3007 Rewriter.disableCanonicalMode();
3008 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3010 // Expand the new value definitions and update the users.
3011 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3012 size_t LUIdx = Fixups[i].LUIdx;
3014 Rewrite(Fixups[i], *Solution[LUIdx], L, IVIncInsertPos, Rewriter,
3015 DeadInsts, SE, DT, P);
3020 // Clean up after ourselves. This must be done before deleting any
3024 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3027 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3028 : IU(P->getAnalysis<IVUsers>()),
3029 SE(P->getAnalysis<ScalarEvolution>()),
3030 DT(P->getAnalysis<DominatorTree>()),
3031 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3033 // If LoopSimplify form is not available, stay out of trouble.
3034 if (!L->isLoopSimplifyForm()) return;
3036 // If there's no interesting work to be done, bail early.
3037 if (IU.empty()) return;
3039 DEBUG(dbgs() << "\nLSR on loop ";
3040 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3043 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3044 /// inside the loop then try to eliminate the cast opeation.
3047 // Change loop terminating condition to use the postinc iv when possible.
3048 Changed |= OptimizeLoopTermCond();
3050 CollectInterestingTypesAndFactors();
3051 CollectFixupsAndInitialFormulae();
3052 CollectLoopInvariantFixupsAndFormulae();
3054 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3055 print_uses(dbgs()));
3057 // Now use the reuse data to generate a bunch of interesting ways
3058 // to formulate the values needed for the uses.
3059 GenerateAllReuseFormulae();
3061 DEBUG(dbgs() << "\n"
3062 "After generating reuse formulae:\n";
3063 print_uses(dbgs()));
3065 FilterOutUndesirableDedicatedRegisters();
3066 NarrowSearchSpaceUsingHeuristics();
3068 SmallVector<const Formula *, 8> Solution;
3070 assert(Solution.size() == Uses.size() && "Malformed solution!");
3072 // Release memory that is no longer needed.
3078 // Formulae should be legal.
3079 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3080 E = Uses.end(); I != E; ++I) {
3081 const LSRUse &LU = *I;
3082 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3083 JE = LU.Formulae.end(); J != JE; ++J)
3084 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3085 LU.Kind, LU.AccessTy, TLI) &&
3086 "Illegal formula generated!");
3090 // Now that we've decided what we want, make it so.
3091 ImplementSolution(Solution, P);
3094 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3095 if (Factors.empty() && Types.empty()) return;
3097 OS << "LSR has identified the following interesting factors and types: ";
3100 for (SmallSetVector<int64_t, 8>::const_iterator
3101 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3102 if (!First) OS << ", ";
3107 for (SmallSetVector<const Type *, 4>::const_iterator
3108 I = Types.begin(), E = Types.end(); I != E; ++I) {
3109 if (!First) OS << ", ";
3111 OS << '(' << **I << ')';
3116 void LSRInstance::print_fixups(raw_ostream &OS) const {
3117 OS << "LSR is examining the following fixup sites:\n";
3118 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3119 E = Fixups.end(); I != E; ++I) {
3120 const LSRFixup &LF = *I;
3127 void LSRInstance::print_uses(raw_ostream &OS) const {
3128 OS << "LSR is examining the following uses:\n";
3129 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3130 E = Uses.end(); I != E; ++I) {
3131 const LSRUse &LU = *I;
3135 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3136 JE = LU.Formulae.end(); J != JE; ++J) {
3144 void LSRInstance::print(raw_ostream &OS) const {
3145 print_factors_and_types(OS);
3150 void LSRInstance::dump() const {
3151 print(errs()); errs() << '\n';
3156 class LoopStrengthReduce : public LoopPass {
3157 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3158 /// transformation profitability.
3159 const TargetLowering *const TLI;
3162 static char ID; // Pass ID, replacement for typeid
3163 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3166 bool runOnLoop(Loop *L, LPPassManager &LPM);
3167 void getAnalysisUsage(AnalysisUsage &AU) const;
3172 char LoopStrengthReduce::ID = 0;
3173 static RegisterPass<LoopStrengthReduce>
3174 X("loop-reduce", "Loop Strength Reduction");
3176 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3177 return new LoopStrengthReduce(TLI);
3180 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3181 : LoopPass(&ID), TLI(tli) {}
3183 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3184 // We split critical edges, so we change the CFG. However, we do update
3185 // many analyses if they are around.
3186 AU.addPreservedID(LoopSimplifyID);
3187 AU.addPreserved<LoopInfo>();
3188 AU.addPreserved("domfrontier");
3190 AU.addRequiredID(LoopSimplifyID);
3191 AU.addRequired<DominatorTree>();
3192 AU.addPreserved<DominatorTree>();
3193 AU.addRequired<ScalarEvolution>();
3194 AU.addPreserved<ScalarEvolution>();
3195 AU.addRequired<IVUsers>();
3196 AU.addPreserved<IVUsers>();
3199 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3200 bool Changed = false;
3202 // Run the main LSR transformation.
3203 Changed |= LSRInstance(TLI, L, this).getChanged();
3205 // At this point, it is worth checking to see if any recurrence PHIs are also
3206 // dead, so that we can remove them as well.
3207 Changed |= DeleteDeadPHIs(L->getHeader());