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 - Recursion 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.getConstant(AR->getType(), 0),
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 const SCEV *Sum = SE.getAddExpr(Good);
267 BaseRegs.push_back(Sum);
268 AM.HasBaseReg = true;
271 const SCEV *Sum = SE.getAddExpr(Bad);
273 BaseRegs.push_back(Sum);
274 AM.HasBaseReg = true;
278 /// getNumRegs - Return the total number of register operands used by this
279 /// formula. This does not include register uses implied by non-constant
281 unsigned Formula::getNumRegs() const {
282 return !!ScaledReg + BaseRegs.size();
285 /// getType - Return the type of this formula, if it has one, or null
286 /// otherwise. This type is meaningless except for the bit size.
287 const Type *Formula::getType() const {
288 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
289 ScaledReg ? ScaledReg->getType() :
290 AM.BaseGV ? AM.BaseGV->getType() :
294 /// referencesReg - Test if this formula references the given register.
295 bool Formula::referencesReg(const SCEV *S) const {
296 return S == ScaledReg ||
297 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
300 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
301 /// which are used by uses other than the use with the given index.
302 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
303 const RegUseTracker &RegUses) const {
305 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
307 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
308 E = BaseRegs.end(); I != E; ++I)
309 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
314 void Formula::print(raw_ostream &OS) const {
317 if (!First) OS << " + "; else First = false;
318 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
320 if (AM.BaseOffs != 0) {
321 if (!First) OS << " + "; else First = false;
324 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
325 E = BaseRegs.end(); I != E; ++I) {
326 if (!First) OS << " + "; else First = false;
327 OS << "reg(" << **I << ')';
330 if (!First) OS << " + "; else First = false;
331 OS << AM.Scale << "*reg(";
340 void Formula::dump() const {
341 print(errs()); errs() << '\n';
344 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
345 /// without changing its value.
346 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
348 IntegerType::get(SE.getContext(),
349 SE.getTypeSizeInBits(AR->getType()) + 1);
350 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
353 /// isAddSExtable - Return true if the given add can be sign-extended
354 /// without changing its value.
355 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
357 IntegerType::get(SE.getContext(),
358 SE.getTypeSizeInBits(A->getType()) + 1);
359 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
362 /// isMulSExtable - Return true if the given add can be sign-extended
363 /// without changing its value.
364 static bool isMulSExtable(const SCEVMulExpr *A, ScalarEvolution &SE) {
366 IntegerType::get(SE.getContext(),
367 SE.getTypeSizeInBits(A->getType()) + 1);
368 return isa<SCEVMulExpr>(SE.getSignExtendExpr(A, WideTy));
371 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
372 /// and if the remainder is known to be zero, or null otherwise. If
373 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
374 /// to Y, ignoring that the multiplication may overflow, which is useful when
375 /// the result will be used in a context where the most significant bits are
377 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
379 bool IgnoreSignificantBits = false) {
380 // Handle the trivial case, which works for any SCEV type.
382 return SE.getConstant(LHS->getType(), 1);
384 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
386 if (RHS->isAllOnesValue())
387 return SE.getMulExpr(LHS, RHS);
389 // Check for a division of a constant by a constant.
390 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
391 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
394 if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
396 return SE.getConstant(C->getValue()->getValue()
397 .sdiv(RC->getValue()->getValue()));
400 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
401 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
402 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
403 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
404 IgnoreSignificantBits);
405 if (!Start) return 0;
406 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
407 IgnoreSignificantBits);
409 return SE.getAddRecExpr(Start, Step, AR->getLoop());
413 // Distribute the sdiv over add operands, if the add doesn't overflow.
414 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
415 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
416 SmallVector<const SCEV *, 8> Ops;
417 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
419 const SCEV *Op = getExactSDiv(*I, RHS, SE,
420 IgnoreSignificantBits);
424 return SE.getAddExpr(Ops);
428 // Check for a multiply operand that we can pull RHS out of.
429 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
430 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
431 SmallVector<const SCEV *, 4> Ops;
433 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
436 if (const SCEV *Q = getExactSDiv(*I, RHS, SE,
437 IgnoreSignificantBits)) {
444 return Found ? SE.getMulExpr(Ops) : 0;
447 // Otherwise we don't know.
451 /// ExtractImmediate - If S involves the addition of a constant integer value,
452 /// return that integer value, and mutate S to point to a new SCEV with that
454 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
455 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
456 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
457 S = SE.getConstant(C->getType(), 0);
458 return C->getValue()->getSExtValue();
460 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
461 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
462 int64_t Result = ExtractImmediate(NewOps.front(), SE);
463 S = SE.getAddExpr(NewOps);
465 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
466 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
467 int64_t Result = ExtractImmediate(NewOps.front(), SE);
468 S = SE.getAddRecExpr(NewOps, AR->getLoop());
474 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
475 /// return that symbol, and mutate S to point to a new SCEV with that
477 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
478 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
479 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
480 S = SE.getConstant(GV->getType(), 0);
483 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
484 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
485 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
486 S = SE.getAddExpr(NewOps);
488 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
489 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
490 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
491 S = SE.getAddRecExpr(NewOps, AR->getLoop());
497 /// isAddressUse - Returns true if the specified instruction is using the
498 /// specified value as an address.
499 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
500 bool isAddress = isa<LoadInst>(Inst);
501 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
502 if (SI->getOperand(1) == OperandVal)
504 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
505 // Addressing modes can also be folded into prefetches and a variety
507 switch (II->getIntrinsicID()) {
509 case Intrinsic::prefetch:
510 case Intrinsic::x86_sse2_loadu_dq:
511 case Intrinsic::x86_sse2_loadu_pd:
512 case Intrinsic::x86_sse_loadu_ps:
513 case Intrinsic::x86_sse_storeu_ps:
514 case Intrinsic::x86_sse2_storeu_pd:
515 case Intrinsic::x86_sse2_storeu_dq:
516 case Intrinsic::x86_sse2_storel_dq:
517 if (II->getOperand(1) == OperandVal)
525 /// getAccessType - Return the type of the memory being accessed.
526 static const Type *getAccessType(const Instruction *Inst) {
527 const Type *AccessTy = Inst->getType();
528 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
529 AccessTy = SI->getOperand(0)->getType();
530 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
531 // Addressing modes can also be folded into prefetches and a variety
533 switch (II->getIntrinsicID()) {
535 case Intrinsic::x86_sse_storeu_ps:
536 case Intrinsic::x86_sse2_storeu_pd:
537 case Intrinsic::x86_sse2_storeu_dq:
538 case Intrinsic::x86_sse2_storel_dq:
539 AccessTy = II->getOperand(1)->getType();
544 // All pointers have the same requirements, so canonicalize them to an
545 // arbitrary pointer type to minimize variation.
546 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
547 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
548 PTy->getAddressSpace());
553 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
554 /// specified set are trivially dead, delete them and see if this makes any of
555 /// their operands subsequently dead.
557 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
558 bool Changed = false;
560 while (!DeadInsts.empty()) {
561 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
563 if (I == 0 || !isInstructionTriviallyDead(I))
566 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
567 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
570 DeadInsts.push_back(U);
573 I->eraseFromParent();
582 /// Cost - This class is used to measure and compare candidate formulae.
584 /// TODO: Some of these could be merged. Also, a lexical ordering
585 /// isn't always optimal.
589 unsigned NumBaseAdds;
595 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
598 unsigned getNumRegs() const { return NumRegs; }
600 bool operator<(const Cost &Other) const;
604 void RateFormula(const Formula &F,
605 SmallPtrSet<const SCEV *, 16> &Regs,
606 const DenseSet<const SCEV *> &VisitedRegs,
608 const SmallVectorImpl<int64_t> &Offsets,
609 ScalarEvolution &SE, DominatorTree &DT);
611 void print(raw_ostream &OS) const;
615 void RateRegister(const SCEV *Reg,
616 SmallPtrSet<const SCEV *, 16> &Regs,
618 ScalarEvolution &SE, DominatorTree &DT);
619 void RatePrimaryRegister(const SCEV *Reg,
620 SmallPtrSet<const SCEV *, 16> &Regs,
622 ScalarEvolution &SE, DominatorTree &DT);
627 /// RateRegister - Tally up interesting quantities from the given register.
628 void Cost::RateRegister(const SCEV *Reg,
629 SmallPtrSet<const SCEV *, 16> &Regs,
631 ScalarEvolution &SE, DominatorTree &DT) {
632 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
633 if (AR->getLoop() == L)
634 AddRecCost += 1; /// TODO: This should be a function of the stride.
636 // If this is an addrec for a loop that's already been visited by LSR,
637 // don't second-guess its addrec phi nodes. LSR isn't currently smart
638 // enough to reason about more than one loop at a time. Consider these
639 // registers free and leave them alone.
640 else if (L->contains(AR->getLoop()) ||
641 (!AR->getLoop()->contains(L) &&
642 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
643 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
644 PHINode *PN = dyn_cast<PHINode>(I); ++I)
645 if (SE.isSCEVable(PN->getType()) &&
646 (SE.getEffectiveSCEVType(PN->getType()) ==
647 SE.getEffectiveSCEVType(AR->getType())) &&
648 SE.getSCEV(PN) == AR)
651 // If this isn't one of the addrecs that the loop already has, it
652 // would require a costly new phi and add. TODO: This isn't
653 // precisely modeled right now.
655 if (!Regs.count(AR->getStart()))
656 RateRegister(AR->getStart(), Regs, L, SE, DT);
659 // Add the step value register, if it needs one.
660 // TODO: The non-affine case isn't precisely modeled here.
661 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
662 if (!Regs.count(AR->getStart()))
663 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
667 // Rough heuristic; favor registers which don't require extra setup
668 // instructions in the preheader.
669 if (!isa<SCEVUnknown>(Reg) &&
670 !isa<SCEVConstant>(Reg) &&
671 !(isa<SCEVAddRecExpr>(Reg) &&
672 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
673 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
677 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
679 void Cost::RatePrimaryRegister(const SCEV *Reg,
680 SmallPtrSet<const SCEV *, 16> &Regs,
682 ScalarEvolution &SE, DominatorTree &DT) {
683 if (Regs.insert(Reg))
684 RateRegister(Reg, Regs, L, SE, DT);
687 void Cost::RateFormula(const Formula &F,
688 SmallPtrSet<const SCEV *, 16> &Regs,
689 const DenseSet<const SCEV *> &VisitedRegs,
691 const SmallVectorImpl<int64_t> &Offsets,
692 ScalarEvolution &SE, DominatorTree &DT) {
693 // Tally up the registers.
694 if (const SCEV *ScaledReg = F.ScaledReg) {
695 if (VisitedRegs.count(ScaledReg)) {
699 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
701 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
702 E = F.BaseRegs.end(); I != E; ++I) {
703 const SCEV *BaseReg = *I;
704 if (VisitedRegs.count(BaseReg)) {
708 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
710 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
711 BaseReg->hasComputableLoopEvolution(L);
714 if (F.BaseRegs.size() > 1)
715 NumBaseAdds += F.BaseRegs.size() - 1;
717 // Tally up the non-zero immediates.
718 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
719 E = Offsets.end(); I != E; ++I) {
720 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
722 ImmCost += 64; // Handle symbolic values conservatively.
723 // TODO: This should probably be the pointer size.
724 else if (Offset != 0)
725 ImmCost += APInt(64, Offset, true).getMinSignedBits();
729 /// Loose - Set this cost to a loosing value.
739 /// operator< - Choose the lower cost.
740 bool Cost::operator<(const Cost &Other) const {
741 if (NumRegs != Other.NumRegs)
742 return NumRegs < Other.NumRegs;
743 if (AddRecCost != Other.AddRecCost)
744 return AddRecCost < Other.AddRecCost;
745 if (NumIVMuls != Other.NumIVMuls)
746 return NumIVMuls < Other.NumIVMuls;
747 if (NumBaseAdds != Other.NumBaseAdds)
748 return NumBaseAdds < Other.NumBaseAdds;
749 if (ImmCost != Other.ImmCost)
750 return ImmCost < Other.ImmCost;
751 if (SetupCost != Other.SetupCost)
752 return SetupCost < Other.SetupCost;
756 void Cost::print(raw_ostream &OS) const {
757 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
759 OS << ", with addrec cost " << AddRecCost;
761 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
762 if (NumBaseAdds != 0)
763 OS << ", plus " << NumBaseAdds << " base add"
764 << (NumBaseAdds == 1 ? "" : "s");
766 OS << ", plus " << ImmCost << " imm cost";
768 OS << ", plus " << SetupCost << " setup cost";
771 void Cost::dump() const {
772 print(errs()); errs() << '\n';
777 /// LSRFixup - An operand value in an instruction which is to be replaced
778 /// with some equivalent, possibly strength-reduced, replacement.
780 /// UserInst - The instruction which will be updated.
781 Instruction *UserInst;
783 /// OperandValToReplace - The operand of the instruction which will
784 /// be replaced. The operand may be used more than once; every instance
785 /// will be replaced.
786 Value *OperandValToReplace;
788 /// PostIncLoops - If this user is to use the post-incremented value of an
789 /// induction variable, this variable is non-null and holds the loop
790 /// associated with the induction variable.
791 PostIncLoopSet PostIncLoops;
793 /// LUIdx - The index of the LSRUse describing the expression which
794 /// this fixup needs, minus an offset (below).
797 /// Offset - A constant offset to be added to the LSRUse expression.
798 /// This allows multiple fixups to share the same LSRUse with different
799 /// offsets, for example in an unrolled loop.
802 bool isUseFullyOutsideLoop(const Loop *L) const;
806 void print(raw_ostream &OS) const;
813 : UserInst(0), OperandValToReplace(0),
814 LUIdx(~size_t(0)), Offset(0) {}
816 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
817 /// value outside of the given loop.
818 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
819 // PHI nodes use their value in their incoming blocks.
820 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
821 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
822 if (PN->getIncomingValue(i) == OperandValToReplace &&
823 L->contains(PN->getIncomingBlock(i)))
828 return !L->contains(UserInst);
831 void LSRFixup::print(raw_ostream &OS) const {
833 // Store is common and interesting enough to be worth special-casing.
834 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
836 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
837 } else if (UserInst->getType()->isVoidTy())
838 OS << UserInst->getOpcodeName();
840 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
842 OS << ", OperandValToReplace=";
843 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
845 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
846 E = PostIncLoops.end(); I != E; ++I) {
847 OS << ", PostIncLoop=";
848 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
851 if (LUIdx != ~size_t(0))
852 OS << ", LUIdx=" << LUIdx;
855 OS << ", Offset=" << Offset;
858 void LSRFixup::dump() const {
859 print(errs()); errs() << '\n';
864 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
865 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
866 struct UniquifierDenseMapInfo {
867 static SmallVector<const SCEV *, 2> getEmptyKey() {
868 SmallVector<const SCEV *, 2> V;
869 V.push_back(reinterpret_cast<const SCEV *>(-1));
873 static SmallVector<const SCEV *, 2> getTombstoneKey() {
874 SmallVector<const SCEV *, 2> V;
875 V.push_back(reinterpret_cast<const SCEV *>(-2));
879 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
881 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
882 E = V.end(); I != E; ++I)
883 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
887 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
888 const SmallVector<const SCEV *, 2> &RHS) {
893 /// LSRUse - This class holds the state that LSR keeps for each use in
894 /// IVUsers, as well as uses invented by LSR itself. It includes information
895 /// about what kinds of things can be folded into the user, information about
896 /// the user itself, and information about how the use may be satisfied.
897 /// TODO: Represent multiple users of the same expression in common?
899 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
902 /// KindType - An enum for a kind of use, indicating what types of
903 /// scaled and immediate operands it might support.
905 Basic, ///< A normal use, with no folding.
906 Special, ///< A special case of basic, allowing -1 scales.
907 Address, ///< An address use; folding according to TargetLowering
908 ICmpZero ///< An equality icmp with both operands folded into one.
909 // TODO: Add a generic icmp too?
913 const Type *AccessTy;
915 SmallVector<int64_t, 8> Offsets;
919 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
920 /// LSRUse are outside of the loop, in which case some special-case heuristics
922 bool AllFixupsOutsideLoop;
924 /// Formulae - A list of ways to build a value that can satisfy this user.
925 /// After the list is populated, one of these is selected heuristically and
926 /// used to formulate a replacement for OperandValToReplace in UserInst.
927 SmallVector<Formula, 12> Formulae;
929 /// Regs - The set of register candidates used by all formulae in this LSRUse.
930 SmallPtrSet<const SCEV *, 4> Regs;
932 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
933 MinOffset(INT64_MAX),
934 MaxOffset(INT64_MIN),
935 AllFixupsOutsideLoop(true) {}
937 bool InsertFormula(const Formula &F);
941 void print(raw_ostream &OS) const;
945 /// InsertFormula - If the given formula has not yet been inserted, add it to
946 /// the list, and return true. Return false otherwise.
947 bool LSRUse::InsertFormula(const Formula &F) {
948 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
949 if (F.ScaledReg) Key.push_back(F.ScaledReg);
950 // Unstable sort by host order ok, because this is only used for uniquifying.
951 std::sort(Key.begin(), Key.end());
953 if (!Uniquifier.insert(Key).second)
956 // Using a register to hold the value of 0 is not profitable.
957 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
958 "Zero allocated in a scaled register!");
960 for (SmallVectorImpl<const SCEV *>::const_iterator I =
961 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
962 assert(!(*I)->isZero() && "Zero allocated in a base register!");
965 // Add the formula to the list.
966 Formulae.push_back(F);
968 // Record registers now being used by this use.
969 if (F.ScaledReg) Regs.insert(F.ScaledReg);
970 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
975 void LSRUse::print(raw_ostream &OS) const {
976 OS << "LSR Use: Kind=";
978 case Basic: OS << "Basic"; break;
979 case Special: OS << "Special"; break;
980 case ICmpZero: OS << "ICmpZero"; break;
983 if (AccessTy->isPointerTy())
984 OS << "pointer"; // the full pointer type could be really verbose
990 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
991 E = Offsets.end(); I != E; ++I) {
998 if (AllFixupsOutsideLoop)
999 OS << ", all-fixups-outside-loop";
1002 void LSRUse::dump() const {
1003 print(errs()); errs() << '\n';
1006 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1007 /// be completely folded into the user instruction at isel time. This includes
1008 /// address-mode folding and special icmp tricks.
1009 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1010 LSRUse::KindType Kind, const Type *AccessTy,
1011 const TargetLowering *TLI) {
1013 case LSRUse::Address:
1014 // If we have low-level target information, ask the target if it can
1015 // completely fold this address.
1016 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1018 // Otherwise, just guess that reg+reg addressing is legal.
1019 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1021 case LSRUse::ICmpZero:
1022 // There's not even a target hook for querying whether it would be legal to
1023 // fold a GV into an ICmp.
1027 // ICmp only has two operands; don't allow more than two non-trivial parts.
1028 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1031 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1032 // putting the scaled register in the other operand of the icmp.
1033 if (AM.Scale != 0 && AM.Scale != -1)
1036 // If we have low-level target information, ask the target if it can fold an
1037 // integer immediate on an icmp.
1038 if (AM.BaseOffs != 0) {
1039 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1046 // Only handle single-register values.
1047 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1049 case LSRUse::Special:
1050 // Only handle -1 scales, or no scale.
1051 return AM.Scale == 0 || AM.Scale == -1;
1057 static bool isLegalUse(TargetLowering::AddrMode AM,
1058 int64_t MinOffset, int64_t MaxOffset,
1059 LSRUse::KindType Kind, const Type *AccessTy,
1060 const TargetLowering *TLI) {
1061 // Check for overflow.
1062 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1065 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1066 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1067 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1068 // Check for overflow.
1069 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1072 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1073 return isLegalUse(AM, Kind, AccessTy, TLI);
1078 static bool isAlwaysFoldable(int64_t BaseOffs,
1079 GlobalValue *BaseGV,
1081 LSRUse::KindType Kind, const Type *AccessTy,
1082 const TargetLowering *TLI) {
1083 // Fast-path: zero is always foldable.
1084 if (BaseOffs == 0 && !BaseGV) return true;
1086 // Conservatively, create an address with an immediate and a
1087 // base and a scale.
1088 TargetLowering::AddrMode AM;
1089 AM.BaseOffs = BaseOffs;
1091 AM.HasBaseReg = HasBaseReg;
1092 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1094 return isLegalUse(AM, Kind, AccessTy, TLI);
1097 static bool isAlwaysFoldable(const SCEV *S,
1098 int64_t MinOffset, int64_t MaxOffset,
1100 LSRUse::KindType Kind, const Type *AccessTy,
1101 const TargetLowering *TLI,
1102 ScalarEvolution &SE) {
1103 // Fast-path: zero is always foldable.
1104 if (S->isZero()) return true;
1106 // Conservatively, create an address with an immediate and a
1107 // base and a scale.
1108 int64_t BaseOffs = ExtractImmediate(S, SE);
1109 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1111 // If there's anything else involved, it's not foldable.
1112 if (!S->isZero()) return false;
1114 // Fast-path: zero is always foldable.
1115 if (BaseOffs == 0 && !BaseGV) return true;
1117 // Conservatively, create an address with an immediate and a
1118 // base and a scale.
1119 TargetLowering::AddrMode AM;
1120 AM.BaseOffs = BaseOffs;
1122 AM.HasBaseReg = HasBaseReg;
1123 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1125 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1128 /// FormulaSorter - This class implements an ordering for formulae which sorts
1129 /// the by their standalone cost.
1130 class FormulaSorter {
1131 /// These two sets are kept empty, so that we compute standalone costs.
1132 DenseSet<const SCEV *> VisitedRegs;
1133 SmallPtrSet<const SCEV *, 16> Regs;
1136 ScalarEvolution &SE;
1140 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1141 : L(l), LU(&lu), SE(se), DT(dt) {}
1143 bool operator()(const Formula &A, const Formula &B) {
1145 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1148 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1150 return CostA < CostB;
1154 /// LSRInstance - This class holds state for the main loop strength reduction
1158 ScalarEvolution &SE;
1161 const TargetLowering *const TLI;
1165 /// IVIncInsertPos - This is the insert position that the current loop's
1166 /// induction variable increment should be placed. In simple loops, this is
1167 /// the latch block's terminator. But in more complicated cases, this is a
1168 /// position which will dominate all the in-loop post-increment users.
1169 Instruction *IVIncInsertPos;
1171 /// Factors - Interesting factors between use strides.
1172 SmallSetVector<int64_t, 8> Factors;
1174 /// Types - Interesting use types, to facilitate truncation reuse.
1175 SmallSetVector<const Type *, 4> Types;
1177 /// Fixups - The list of operands which are to be replaced.
1178 SmallVector<LSRFixup, 16> Fixups;
1180 /// Uses - The list of interesting uses.
1181 SmallVector<LSRUse, 16> Uses;
1183 /// RegUses - Track which uses use which register candidates.
1184 RegUseTracker RegUses;
1186 void OptimizeShadowIV();
1187 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1188 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1189 bool OptimizeLoopTermCond();
1191 void CollectInterestingTypesAndFactors();
1192 void CollectFixupsAndInitialFormulae();
1194 LSRFixup &getNewFixup() {
1195 Fixups.push_back(LSRFixup());
1196 return Fixups.back();
1199 // Support for sharing of LSRUses between LSRFixups.
1200 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1203 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1204 LSRUse::KindType Kind, const Type *AccessTy);
1206 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1207 LSRUse::KindType Kind,
1208 const Type *AccessTy);
1211 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1212 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1213 void CountRegisters(const Formula &F, size_t LUIdx);
1214 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1216 void CollectLoopInvariantFixupsAndFormulae();
1218 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1219 unsigned Depth = 0);
1220 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1221 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1222 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1223 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1224 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1225 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1226 void GenerateCrossUseConstantOffsets();
1227 void GenerateAllReuseFormulae();
1229 void FilterOutUndesirableDedicatedRegisters();
1230 void NarrowSearchSpaceUsingHeuristics();
1232 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1234 SmallVectorImpl<const Formula *> &Workspace,
1235 const Cost &CurCost,
1236 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1237 DenseSet<const SCEV *> &VisitedRegs) const;
1238 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1240 BasicBlock::iterator
1241 HoistInsertPosition(BasicBlock::iterator IP,
1242 const SmallVectorImpl<Instruction *> &Inputs) const;
1243 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1245 const LSRUse &LU) const;
1247 Value *Expand(const LSRFixup &LF,
1249 BasicBlock::iterator IP,
1250 SCEVExpander &Rewriter,
1251 SmallVectorImpl<WeakVH> &DeadInsts) const;
1252 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1254 SCEVExpander &Rewriter,
1255 SmallVectorImpl<WeakVH> &DeadInsts,
1257 void Rewrite(const LSRFixup &LF,
1259 SCEVExpander &Rewriter,
1260 SmallVectorImpl<WeakVH> &DeadInsts,
1262 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1265 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1267 bool getChanged() const { return Changed; }
1269 void print_factors_and_types(raw_ostream &OS) const;
1270 void print_fixups(raw_ostream &OS) const;
1271 void print_uses(raw_ostream &OS) const;
1272 void print(raw_ostream &OS) const;
1278 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1279 /// inside the loop then try to eliminate the cast operation.
1280 void LSRInstance::OptimizeShadowIV() {
1281 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1282 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1285 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1286 UI != E; /* empty */) {
1287 IVUsers::const_iterator CandidateUI = UI;
1289 Instruction *ShadowUse = CandidateUI->getUser();
1290 const Type *DestTy = NULL;
1292 /* If shadow use is a int->float cast then insert a second IV
1293 to eliminate this cast.
1295 for (unsigned i = 0; i < n; ++i)
1301 for (unsigned i = 0; i < n; ++i, ++d)
1304 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1305 DestTy = UCast->getDestTy();
1306 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1307 DestTy = SCast->getDestTy();
1308 if (!DestTy) continue;
1311 // If target does not support DestTy natively then do not apply
1312 // this transformation.
1313 EVT DVT = TLI->getValueType(DestTy);
1314 if (!TLI->isTypeLegal(DVT)) continue;
1317 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1319 if (PH->getNumIncomingValues() != 2) continue;
1321 const Type *SrcTy = PH->getType();
1322 int Mantissa = DestTy->getFPMantissaWidth();
1323 if (Mantissa == -1) continue;
1324 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1327 unsigned Entry, Latch;
1328 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1336 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1337 if (!Init) continue;
1338 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1340 BinaryOperator *Incr =
1341 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1342 if (!Incr) continue;
1343 if (Incr->getOpcode() != Instruction::Add
1344 && Incr->getOpcode() != Instruction::Sub)
1347 /* Initialize new IV, double d = 0.0 in above example. */
1348 ConstantInt *C = NULL;
1349 if (Incr->getOperand(0) == PH)
1350 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1351 else if (Incr->getOperand(1) == PH)
1352 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1358 // Ignore negative constants, as the code below doesn't handle them
1359 // correctly. TODO: Remove this restriction.
1360 if (!C->getValue().isStrictlyPositive()) continue;
1362 /* Add new PHINode. */
1363 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1365 /* create new increment. '++d' in above example. */
1366 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1367 BinaryOperator *NewIncr =
1368 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1369 Instruction::FAdd : Instruction::FSub,
1370 NewPH, CFP, "IV.S.next.", Incr);
1372 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1373 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1375 /* Remove cast operation */
1376 ShadowUse->replaceAllUsesWith(NewPH);
1377 ShadowUse->eraseFromParent();
1382 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1383 /// set the IV user and stride information and return true, otherwise return
1385 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond,
1386 IVStrideUse *&CondUse) {
1387 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1388 if (UI->getUser() == Cond) {
1389 // NOTE: we could handle setcc instructions with multiple uses here, but
1390 // InstCombine does it as well for simple uses, it's not clear that it
1391 // occurs enough in real life to handle.
1398 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1399 /// a max computation.
1401 /// This is a narrow solution to a specific, but acute, problem. For loops
1407 /// } while (++i < n);
1409 /// the trip count isn't just 'n', because 'n' might not be positive. And
1410 /// unfortunately this can come up even for loops where the user didn't use
1411 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1412 /// will commonly be lowered like this:
1418 /// } while (++i < n);
1421 /// and then it's possible for subsequent optimization to obscure the if
1422 /// test in such a way that indvars can't find it.
1424 /// When indvars can't find the if test in loops like this, it creates a
1425 /// max expression, which allows it to give the loop a canonical
1426 /// induction variable:
1429 /// max = n < 1 ? 1 : n;
1432 /// } while (++i != max);
1434 /// Canonical induction variables are necessary because the loop passes
1435 /// are designed around them. The most obvious example of this is the
1436 /// LoopInfo analysis, which doesn't remember trip count values. It
1437 /// expects to be able to rediscover the trip count each time it is
1438 /// needed, and it does this using a simple analysis that only succeeds if
1439 /// the loop has a canonical induction variable.
1441 /// However, when it comes time to generate code, the maximum operation
1442 /// can be quite costly, especially if it's inside of an outer loop.
1444 /// This function solves this problem by detecting this type of loop and
1445 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1446 /// the instructions for the maximum computation.
1448 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1449 // Check that the loop matches the pattern we're looking for.
1450 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1451 Cond->getPredicate() != CmpInst::ICMP_NE)
1454 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1455 if (!Sel || !Sel->hasOneUse()) return Cond;
1457 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1458 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1460 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1462 // Add one to the backedge-taken count to get the trip count.
1463 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1464 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1466 // Check for a max calculation that matches the pattern. There's no check
1467 // for ICMP_ULE here because the comparison would be with zero, which
1468 // isn't interesting.
1469 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1470 const SCEVNAryExpr *Max = 0;
1471 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1472 Pred = ICmpInst::ICMP_SLE;
1474 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1475 Pred = ICmpInst::ICMP_SLT;
1477 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1478 Pred = ICmpInst::ICMP_ULT;
1485 // To handle a max with more than two operands, this optimization would
1486 // require additional checking and setup.
1487 if (Max->getNumOperands() != 2)
1490 const SCEV *MaxLHS = Max->getOperand(0);
1491 const SCEV *MaxRHS = Max->getOperand(1);
1493 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1494 // for a comparison with 1. For <= and >=, a comparison with zero.
1496 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1499 // Check the relevant induction variable for conformance to
1501 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1502 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1503 if (!AR || !AR->isAffine() ||
1504 AR->getStart() != One ||
1505 AR->getStepRecurrence(SE) != One)
1508 assert(AR->getLoop() == L &&
1509 "Loop condition operand is an addrec in a different loop!");
1511 // Check the right operand of the select, and remember it, as it will
1512 // be used in the new comparison instruction.
1514 if (ICmpInst::isTrueWhenEqual(Pred)) {
1515 // Look for n+1, and grab n.
1516 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1517 if (isa<ConstantInt>(BO->getOperand(1)) &&
1518 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1519 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1520 NewRHS = BO->getOperand(0);
1521 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1522 if (isa<ConstantInt>(BO->getOperand(1)) &&
1523 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1524 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1525 NewRHS = BO->getOperand(0);
1528 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1529 NewRHS = Sel->getOperand(1);
1530 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1531 NewRHS = Sel->getOperand(2);
1533 llvm_unreachable("Max doesn't match expected pattern!");
1535 // Determine the new comparison opcode. It may be signed or unsigned,
1536 // and the original comparison may be either equality or inequality.
1537 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1538 Pred = CmpInst::getInversePredicate(Pred);
1540 // Ok, everything looks ok to change the condition into an SLT or SGE and
1541 // delete the max calculation.
1543 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1545 // Delete the max calculation instructions.
1546 Cond->replaceAllUsesWith(NewCond);
1547 CondUse->setUser(NewCond);
1548 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1549 Cond->eraseFromParent();
1550 Sel->eraseFromParent();
1551 if (Cmp->use_empty())
1552 Cmp->eraseFromParent();
1556 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1557 /// postinc iv when possible.
1559 LSRInstance::OptimizeLoopTermCond() {
1560 SmallPtrSet<Instruction *, 4> PostIncs;
1562 BasicBlock *LatchBlock = L->getLoopLatch();
1563 SmallVector<BasicBlock*, 8> ExitingBlocks;
1564 L->getExitingBlocks(ExitingBlocks);
1566 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1567 BasicBlock *ExitingBlock = ExitingBlocks[i];
1569 // Get the terminating condition for the loop if possible. If we
1570 // can, we want to change it to use a post-incremented version of its
1571 // induction variable, to allow coalescing the live ranges for the IV into
1572 // one register value.
1574 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1577 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1578 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1581 // Search IVUsesByStride to find Cond's IVUse if there is one.
1582 IVStrideUse *CondUse = 0;
1583 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1584 if (!FindIVUserForCond(Cond, CondUse))
1587 // If the trip count is computed in terms of a max (due to ScalarEvolution
1588 // being unable to find a sufficient guard, for example), change the loop
1589 // comparison to use SLT or ULT instead of NE.
1590 // One consequence of doing this now is that it disrupts the count-down
1591 // optimization. That's not always a bad thing though, because in such
1592 // cases it may still be worthwhile to avoid a max.
1593 Cond = OptimizeMax(Cond, CondUse);
1595 // If this exiting block dominates the latch block, it may also use
1596 // the post-inc value if it won't be shared with other uses.
1597 // Check for dominance.
1598 if (!DT.dominates(ExitingBlock, LatchBlock))
1601 // Conservatively avoid trying to use the post-inc value in non-latch
1602 // exits if there may be pre-inc users in intervening blocks.
1603 if (LatchBlock != ExitingBlock)
1604 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1605 // Test if the use is reachable from the exiting block. This dominator
1606 // query is a conservative approximation of reachability.
1607 if (&*UI != CondUse &&
1608 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1609 // Conservatively assume there may be reuse if the quotient of their
1610 // strides could be a legal scale.
1611 const SCEV *A = IU.getStride(*CondUse, L);
1612 const SCEV *B = IU.getStride(*UI, L);
1613 if (!A || !B) continue;
1614 if (SE.getTypeSizeInBits(A->getType()) !=
1615 SE.getTypeSizeInBits(B->getType())) {
1616 if (SE.getTypeSizeInBits(A->getType()) >
1617 SE.getTypeSizeInBits(B->getType()))
1618 B = SE.getSignExtendExpr(B, A->getType());
1620 A = SE.getSignExtendExpr(A, B->getType());
1622 if (const SCEVConstant *D =
1623 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1624 // Stride of one or negative one can have reuse with non-addresses.
1625 if (D->getValue()->isOne() ||
1626 D->getValue()->isAllOnesValue())
1627 goto decline_post_inc;
1628 // Avoid weird situations.
1629 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1630 D->getValue()->getValue().isMinSignedValue())
1631 goto decline_post_inc;
1632 // Without TLI, assume that any stride might be valid, and so any
1633 // use might be shared.
1635 goto decline_post_inc;
1636 // Check for possible scaled-address reuse.
1637 const Type *AccessTy = getAccessType(UI->getUser());
1638 TargetLowering::AddrMode AM;
1639 AM.Scale = D->getValue()->getSExtValue();
1640 if (TLI->isLegalAddressingMode(AM, AccessTy))
1641 goto decline_post_inc;
1642 AM.Scale = -AM.Scale;
1643 if (TLI->isLegalAddressingMode(AM, AccessTy))
1644 goto decline_post_inc;
1648 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1651 // It's possible for the setcc instruction to be anywhere in the loop, and
1652 // possible for it to have multiple users. If it is not immediately before
1653 // the exiting block branch, move it.
1654 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1655 if (Cond->hasOneUse()) {
1656 Cond->moveBefore(TermBr);
1658 // Clone the terminating condition and insert into the loopend.
1659 ICmpInst *OldCond = Cond;
1660 Cond = cast<ICmpInst>(Cond->clone());
1661 Cond->setName(L->getHeader()->getName() + ".termcond");
1662 ExitingBlock->getInstList().insert(TermBr, Cond);
1664 // Clone the IVUse, as the old use still exists!
1665 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1666 TermBr->replaceUsesOfWith(OldCond, Cond);
1670 // If we get to here, we know that we can transform the setcc instruction to
1671 // use the post-incremented version of the IV, allowing us to coalesce the
1672 // live ranges for the IV correctly.
1673 CondUse->transformToPostInc(L);
1676 PostIncs.insert(Cond);
1680 // Determine an insertion point for the loop induction variable increment. It
1681 // must dominate all the post-inc comparisons we just set up, and it must
1682 // dominate the loop latch edge.
1683 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1684 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1685 E = PostIncs.end(); I != E; ++I) {
1687 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1689 if (BB == (*I)->getParent())
1690 IVIncInsertPos = *I;
1691 else if (BB != IVIncInsertPos->getParent())
1692 IVIncInsertPos = BB->getTerminator();
1699 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1700 LSRUse::KindType Kind, const Type *AccessTy) {
1701 int64_t NewMinOffset = LU.MinOffset;
1702 int64_t NewMaxOffset = LU.MaxOffset;
1703 const Type *NewAccessTy = AccessTy;
1705 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1706 // something conservative, however this can pessimize in the case that one of
1707 // the uses will have all its uses outside the loop, for example.
1708 if (LU.Kind != Kind)
1710 // Conservatively assume HasBaseReg is true for now.
1711 if (NewOffset < LU.MinOffset) {
1712 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true,
1713 Kind, AccessTy, TLI))
1715 NewMinOffset = NewOffset;
1716 } else if (NewOffset > LU.MaxOffset) {
1717 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true,
1718 Kind, AccessTy, TLI))
1720 NewMaxOffset = NewOffset;
1722 // Check for a mismatched access type, and fall back conservatively as needed.
1723 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1724 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1727 LU.MinOffset = NewMinOffset;
1728 LU.MaxOffset = NewMaxOffset;
1729 LU.AccessTy = NewAccessTy;
1730 if (NewOffset != LU.Offsets.back())
1731 LU.Offsets.push_back(NewOffset);
1735 /// getUse - Return an LSRUse index and an offset value for a fixup which
1736 /// needs the given expression, with the given kind and optional access type.
1737 /// Either reuse an existing use or create a new one, as needed.
1738 std::pair<size_t, int64_t>
1739 LSRInstance::getUse(const SCEV *&Expr,
1740 LSRUse::KindType Kind, const Type *AccessTy) {
1741 const SCEV *Copy = Expr;
1742 int64_t Offset = ExtractImmediate(Expr, SE);
1744 // Basic uses can't accept any offset, for example.
1745 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1750 std::pair<UseMapTy::iterator, bool> P =
1751 UseMap.insert(std::make_pair(Expr, 0));
1753 // A use already existed with this base.
1754 size_t LUIdx = P.first->second;
1755 LSRUse &LU = Uses[LUIdx];
1756 if (reconcileNewOffset(LU, Offset, Kind, AccessTy))
1758 return std::make_pair(LUIdx, Offset);
1761 // Create a new use.
1762 size_t LUIdx = Uses.size();
1763 P.first->second = LUIdx;
1764 Uses.push_back(LSRUse(Kind, AccessTy));
1765 LSRUse &LU = Uses[LUIdx];
1767 // We don't need to track redundant offsets, but we don't need to go out
1768 // of our way here to avoid them.
1769 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1770 LU.Offsets.push_back(Offset);
1772 LU.MinOffset = Offset;
1773 LU.MaxOffset = Offset;
1774 return std::make_pair(LUIdx, Offset);
1777 void LSRInstance::CollectInterestingTypesAndFactors() {
1778 SmallSetVector<const SCEV *, 4> Strides;
1780 // Collect interesting types and strides.
1781 SmallVector<const SCEV *, 4> Worklist;
1782 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1783 const SCEV *Expr = IU.getExpr(*UI);
1785 // Collect interesting types.
1786 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1788 // Add strides for mentioned loops.
1789 Worklist.push_back(Expr);
1791 const SCEV *S = Worklist.pop_back_val();
1792 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1793 Strides.insert(AR->getStepRecurrence(SE));
1794 Worklist.push_back(AR->getStart());
1795 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1796 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1798 } while (!Worklist.empty());
1801 // Compute interesting factors from the set of interesting strides.
1802 for (SmallSetVector<const SCEV *, 4>::const_iterator
1803 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1804 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1805 next(I); NewStrideIter != E; ++NewStrideIter) {
1806 const SCEV *OldStride = *I;
1807 const SCEV *NewStride = *NewStrideIter;
1809 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1810 SE.getTypeSizeInBits(NewStride->getType())) {
1811 if (SE.getTypeSizeInBits(OldStride->getType()) >
1812 SE.getTypeSizeInBits(NewStride->getType()))
1813 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1815 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1817 if (const SCEVConstant *Factor =
1818 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1820 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1821 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1822 } else if (const SCEVConstant *Factor =
1823 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1826 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1827 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1831 // If all uses use the same type, don't bother looking for truncation-based
1833 if (Types.size() == 1)
1836 DEBUG(print_factors_and_types(dbgs()));
1839 void LSRInstance::CollectFixupsAndInitialFormulae() {
1840 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1842 LSRFixup &LF = getNewFixup();
1843 LF.UserInst = UI->getUser();
1844 LF.OperandValToReplace = UI->getOperandValToReplace();
1845 LF.PostIncLoops = UI->getPostIncLoops();
1847 LSRUse::KindType Kind = LSRUse::Basic;
1848 const Type *AccessTy = 0;
1849 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1850 Kind = LSRUse::Address;
1851 AccessTy = getAccessType(LF.UserInst);
1854 const SCEV *S = IU.getExpr(*UI);
1856 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1857 // (N - i == 0), and this allows (N - i) to be the expression that we work
1858 // with rather than just N or i, so we can consider the register
1859 // requirements for both N and i at the same time. Limiting this code to
1860 // equality icmps is not a problem because all interesting loops use
1861 // equality icmps, thanks to IndVarSimplify.
1862 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1863 if (CI->isEquality()) {
1864 // Swap the operands if needed to put the OperandValToReplace on the
1865 // left, for consistency.
1866 Value *NV = CI->getOperand(1);
1867 if (NV == LF.OperandValToReplace) {
1868 CI->setOperand(1, CI->getOperand(0));
1869 CI->setOperand(0, NV);
1872 // x == y --> x - y == 0
1873 const SCEV *N = SE.getSCEV(NV);
1874 if (N->isLoopInvariant(L)) {
1875 Kind = LSRUse::ICmpZero;
1876 S = SE.getMinusSCEV(N, S);
1879 // -1 and the negations of all interesting strides (except the negation
1880 // of -1) are now also interesting.
1881 for (size_t i = 0, e = Factors.size(); i != e; ++i)
1882 if (Factors[i] != -1)
1883 Factors.insert(-(uint64_t)Factors[i]);
1887 // Set up the initial formula for this use.
1888 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
1890 LF.Offset = P.second;
1891 LSRUse &LU = Uses[LF.LUIdx];
1892 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
1894 // If this is the first use of this LSRUse, give it a formula.
1895 if (LU.Formulae.empty()) {
1896 InsertInitialFormula(S, LU, LF.LUIdx);
1897 CountRegisters(LU.Formulae.back(), LF.LUIdx);
1901 DEBUG(print_fixups(dbgs()));
1905 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
1907 F.InitialMatch(S, L, SE, DT);
1908 bool Inserted = InsertFormula(LU, LUIdx, F);
1909 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
1913 LSRInstance::InsertSupplementalFormula(const SCEV *S,
1914 LSRUse &LU, size_t LUIdx) {
1916 F.BaseRegs.push_back(S);
1917 F.AM.HasBaseReg = true;
1918 bool Inserted = InsertFormula(LU, LUIdx, F);
1919 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
1922 /// CountRegisters - Note which registers are used by the given formula,
1923 /// updating RegUses.
1924 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
1926 RegUses.CountRegister(F.ScaledReg, LUIdx);
1927 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1928 E = F.BaseRegs.end(); I != E; ++I)
1929 RegUses.CountRegister(*I, LUIdx);
1932 /// InsertFormula - If the given formula has not yet been inserted, add it to
1933 /// the list, and return true. Return false otherwise.
1934 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
1935 if (!LU.InsertFormula(F))
1938 CountRegisters(F, LUIdx);
1942 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
1943 /// loop-invariant values which we're tracking. These other uses will pin these
1944 /// values in registers, making them less profitable for elimination.
1945 /// TODO: This currently misses non-constant addrec step registers.
1946 /// TODO: Should this give more weight to users inside the loop?
1948 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
1949 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
1950 SmallPtrSet<const SCEV *, 8> Inserted;
1952 while (!Worklist.empty()) {
1953 const SCEV *S = Worklist.pop_back_val();
1955 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
1956 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
1957 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1958 Worklist.push_back(C->getOperand());
1959 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1960 Worklist.push_back(D->getLHS());
1961 Worklist.push_back(D->getRHS());
1962 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
1963 if (!Inserted.insert(U)) continue;
1964 const Value *V = U->getValue();
1965 if (const Instruction *Inst = dyn_cast<Instruction>(V))
1966 if (L->contains(Inst)) continue;
1967 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1969 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
1970 // Ignore non-instructions.
1973 // Ignore instructions in other functions (as can happen with
1975 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
1977 // Ignore instructions not dominated by the loop.
1978 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
1979 UserInst->getParent() :
1980 cast<PHINode>(UserInst)->getIncomingBlock(
1981 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
1982 if (!DT.dominates(L->getHeader(), UseBB))
1984 // Ignore uses which are part of other SCEV expressions, to avoid
1985 // analyzing them multiple times.
1986 if (SE.isSCEVable(UserInst->getType())) {
1987 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
1988 // If the user is a no-op, look through to its uses.
1989 if (!isa<SCEVUnknown>(UserS))
1993 SE.getUnknown(const_cast<Instruction *>(UserInst)));
1997 // Ignore icmp instructions which are already being analyzed.
1998 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
1999 unsigned OtherIdx = !UI.getOperandNo();
2000 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2001 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2005 LSRFixup &LF = getNewFixup();
2006 LF.UserInst = const_cast<Instruction *>(UserInst);
2007 LF.OperandValToReplace = UI.getUse();
2008 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2010 LF.Offset = P.second;
2011 LSRUse &LU = Uses[LF.LUIdx];
2012 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2013 InsertSupplementalFormula(U, LU, LF.LUIdx);
2014 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2021 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2022 /// separate registers. If C is non-null, multiply each subexpression by C.
2023 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2024 SmallVectorImpl<const SCEV *> &Ops,
2025 ScalarEvolution &SE) {
2026 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2027 // Break out add operands.
2028 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2030 CollectSubexprs(*I, C, Ops, SE);
2032 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2033 // Split a non-zero base out of an addrec.
2034 if (!AR->getStart()->isZero()) {
2035 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2036 AR->getStepRecurrence(SE),
2037 AR->getLoop()), C, Ops, SE);
2038 CollectSubexprs(AR->getStart(), C, Ops, SE);
2041 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2042 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2043 if (Mul->getNumOperands() == 2)
2044 if (const SCEVConstant *Op0 =
2045 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2046 CollectSubexprs(Mul->getOperand(1),
2047 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2053 // Otherwise use the value itself.
2054 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2057 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2059 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2062 // Arbitrarily cap recursion to protect compile time.
2063 if (Depth >= 3) return;
2065 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2066 const SCEV *BaseReg = Base.BaseRegs[i];
2068 SmallVector<const SCEV *, 8> AddOps;
2069 CollectSubexprs(BaseReg, 0, AddOps, SE);
2070 if (AddOps.size() == 1) continue;
2072 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2073 JE = AddOps.end(); J != JE; ++J) {
2074 // Don't pull a constant into a register if the constant could be folded
2075 // into an immediate field.
2076 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2077 Base.getNumRegs() > 1,
2078 LU.Kind, LU.AccessTy, TLI, SE))
2081 // Collect all operands except *J.
2082 SmallVector<const SCEV *, 8> InnerAddOps;
2083 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2084 KE = AddOps.end(); K != KE; ++K)
2086 InnerAddOps.push_back(*K);
2088 // Don't leave just a constant behind in a register if the constant could
2089 // be folded into an immediate field.
2090 if (InnerAddOps.size() == 1 &&
2091 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2092 Base.getNumRegs() > 1,
2093 LU.Kind, LU.AccessTy, TLI, SE))
2096 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2097 if (InnerSum->isZero())
2100 F.BaseRegs[i] = InnerSum;
2101 F.BaseRegs.push_back(*J);
2102 if (InsertFormula(LU, LUIdx, F))
2103 // If that formula hadn't been seen before, recurse to find more like
2105 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2110 /// GenerateCombinations - Generate a formula consisting of all of the
2111 /// loop-dominating registers added into a single register.
2112 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2114 // This method is only interesting on a plurality of registers.
2115 if (Base.BaseRegs.size() <= 1) return;
2119 SmallVector<const SCEV *, 4> Ops;
2120 for (SmallVectorImpl<const SCEV *>::const_iterator
2121 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2122 const SCEV *BaseReg = *I;
2123 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2124 !BaseReg->hasComputableLoopEvolution(L))
2125 Ops.push_back(BaseReg);
2127 F.BaseRegs.push_back(BaseReg);
2129 if (Ops.size() > 1) {
2130 const SCEV *Sum = SE.getAddExpr(Ops);
2131 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2132 // opportunity to fold something. For now, just ignore such cases
2133 // rather than proceed with zero in a register.
2134 if (!Sum->isZero()) {
2135 F.BaseRegs.push_back(Sum);
2136 (void)InsertFormula(LU, LUIdx, F);
2141 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2142 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2144 // We can't add a symbolic offset if the address already contains one.
2145 if (Base.AM.BaseGV) return;
2147 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2148 const SCEV *G = Base.BaseRegs[i];
2149 GlobalValue *GV = ExtractSymbol(G, SE);
2150 if (G->isZero() || !GV)
2154 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2155 LU.Kind, LU.AccessTy, TLI))
2158 (void)InsertFormula(LU, LUIdx, F);
2162 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2163 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2165 // TODO: For now, just add the min and max offset, because it usually isn't
2166 // worthwhile looking at everything inbetween.
2167 SmallVector<int64_t, 4> Worklist;
2168 Worklist.push_back(LU.MinOffset);
2169 if (LU.MaxOffset != LU.MinOffset)
2170 Worklist.push_back(LU.MaxOffset);
2172 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2173 const SCEV *G = Base.BaseRegs[i];
2175 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2176 E = Worklist.end(); I != E; ++I) {
2178 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2179 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2180 LU.Kind, LU.AccessTy, TLI)) {
2181 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2183 (void)InsertFormula(LU, LUIdx, F);
2187 int64_t Imm = ExtractImmediate(G, SE);
2188 if (G->isZero() || Imm == 0)
2191 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2192 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2193 LU.Kind, LU.AccessTy, TLI))
2196 (void)InsertFormula(LU, LUIdx, F);
2200 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2201 /// the comparison. For example, x == y -> x*c == y*c.
2202 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2204 if (LU.Kind != LSRUse::ICmpZero) return;
2206 // Determine the integer type for the base formula.
2207 const Type *IntTy = Base.getType();
2209 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2211 // Don't do this if there is more than one offset.
2212 if (LU.MinOffset != LU.MaxOffset) return;
2214 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2216 // Check each interesting stride.
2217 for (SmallSetVector<int64_t, 8>::const_iterator
2218 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2219 int64_t Factor = *I;
2222 // Check that the multiplication doesn't overflow.
2223 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2225 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2226 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2229 // Check that multiplying with the use offset doesn't overflow.
2230 int64_t Offset = LU.MinOffset;
2231 if (Offset == INT64_MIN && Factor == -1)
2233 Offset = (uint64_t)Offset * Factor;
2234 if (Offset / Factor != LU.MinOffset)
2237 // Check that this scale is legal.
2238 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2241 // Compensate for the use having MinOffset built into it.
2242 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2244 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2246 // Check that multiplying with each base register doesn't overflow.
2247 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2248 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2249 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2253 // Check that multiplying with the scaled register doesn't overflow.
2255 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2256 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2260 // If we make it here and it's legal, add it.
2261 (void)InsertFormula(LU, LUIdx, F);
2266 /// GenerateScales - Generate stride factor reuse formulae by making use of
2267 /// scaled-offset address modes, for example.
2268 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
2270 // Determine the integer type for the base formula.
2271 const Type *IntTy = Base.getType();
2274 // If this Formula already has a scaled register, we can't add another one.
2275 if (Base.AM.Scale != 0) return;
2277 // Check each interesting stride.
2278 for (SmallSetVector<int64_t, 8>::const_iterator
2279 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2280 int64_t Factor = *I;
2282 Base.AM.Scale = Factor;
2283 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2284 // Check whether this scale is going to be legal.
2285 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2286 LU.Kind, LU.AccessTy, TLI)) {
2287 // As a special-case, handle special out-of-loop Basic users specially.
2288 // TODO: Reconsider this special case.
2289 if (LU.Kind == LSRUse::Basic &&
2290 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2291 LSRUse::Special, LU.AccessTy, TLI) &&
2292 LU.AllFixupsOutsideLoop)
2293 LU.Kind = LSRUse::Special;
2297 // For an ICmpZero, negating a solitary base register won't lead to
2299 if (LU.Kind == LSRUse::ICmpZero &&
2300 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2302 // For each addrec base reg, apply the scale, if possible.
2303 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2304 if (const SCEVAddRecExpr *AR =
2305 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2306 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2307 if (FactorS->isZero())
2309 // Divide out the factor, ignoring high bits, since we'll be
2310 // scaling the value back up in the end.
2311 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2312 // TODO: This could be optimized to avoid all the copying.
2314 F.ScaledReg = Quotient;
2315 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2316 F.BaseRegs.pop_back();
2317 (void)InsertFormula(LU, LUIdx, F);
2323 /// GenerateTruncates - Generate reuse formulae from different IV types.
2324 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
2326 // This requires TargetLowering to tell us which truncates are free.
2329 // Don't bother truncating symbolic values.
2330 if (Base.AM.BaseGV) return;
2332 // Determine the integer type for the base formula.
2333 const Type *DstTy = Base.getType();
2335 DstTy = SE.getEffectiveSCEVType(DstTy);
2337 for (SmallSetVector<const Type *, 4>::const_iterator
2338 I = Types.begin(), E = Types.end(); I != E; ++I) {
2339 const Type *SrcTy = *I;
2340 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2343 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2344 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2345 JE = F.BaseRegs.end(); J != JE; ++J)
2346 *J = SE.getAnyExtendExpr(*J, SrcTy);
2348 // TODO: This assumes we've done basic processing on all uses and
2349 // have an idea what the register usage is.
2350 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2353 (void)InsertFormula(LU, LUIdx, F);
2360 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2361 /// defer modifications so that the search phase doesn't have to worry about
2362 /// the data structures moving underneath it.
2366 const SCEV *OrigReg;
2368 WorkItem(size_t LI, int64_t I, const SCEV *R)
2369 : LUIdx(LI), Imm(I), OrigReg(R) {}
2371 void print(raw_ostream &OS) const;
2377 void WorkItem::print(raw_ostream &OS) const {
2378 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2379 << " , add offset " << Imm;
2382 void WorkItem::dump() const {
2383 print(errs()); errs() << '\n';
2386 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2387 /// distance apart and try to form reuse opportunities between them.
2388 void LSRInstance::GenerateCrossUseConstantOffsets() {
2389 // Group the registers by their value without any added constant offset.
2390 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2391 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2393 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2394 SmallVector<const SCEV *, 8> Sequence;
2395 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2397 const SCEV *Reg = *I;
2398 int64_t Imm = ExtractImmediate(Reg, SE);
2399 std::pair<RegMapTy::iterator, bool> Pair =
2400 Map.insert(std::make_pair(Reg, ImmMapTy()));
2402 Sequence.push_back(Reg);
2403 Pair.first->second.insert(std::make_pair(Imm, *I));
2404 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2407 // Now examine each set of registers with the same base value. Build up
2408 // a list of work to do and do the work in a separate step so that we're
2409 // not adding formulae and register counts while we're searching.
2410 SmallVector<WorkItem, 32> WorkItems;
2411 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2412 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2413 E = Sequence.end(); I != E; ++I) {
2414 const SCEV *Reg = *I;
2415 const ImmMapTy &Imms = Map.find(Reg)->second;
2417 // It's not worthwhile looking for reuse if there's only one offset.
2418 if (Imms.size() == 1)
2421 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2422 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2424 dbgs() << ' ' << J->first;
2427 // Examine each offset.
2428 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2430 const SCEV *OrigReg = J->second;
2432 int64_t JImm = J->first;
2433 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2435 if (!isa<SCEVConstant>(OrigReg) &&
2436 UsedByIndicesMap[Reg].count() == 1) {
2437 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2441 // Conservatively examine offsets between this orig reg a few selected
2443 ImmMapTy::const_iterator OtherImms[] = {
2444 Imms.begin(), prior(Imms.end()),
2445 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2447 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2448 ImmMapTy::const_iterator M = OtherImms[i];
2449 if (M == J || M == JE) continue;
2451 // Compute the difference between the two.
2452 int64_t Imm = (uint64_t)JImm - M->first;
2453 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2454 LUIdx = UsedByIndices.find_next(LUIdx))
2455 // Make a memo of this use, offset, and register tuple.
2456 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2457 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2464 UsedByIndicesMap.clear();
2465 UniqueItems.clear();
2467 // Now iterate through the worklist and add new formulae.
2468 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2469 E = WorkItems.end(); I != E; ++I) {
2470 const WorkItem &WI = *I;
2471 size_t LUIdx = WI.LUIdx;
2472 LSRUse &LU = Uses[LUIdx];
2473 int64_t Imm = WI.Imm;
2474 const SCEV *OrigReg = WI.OrigReg;
2476 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2477 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2478 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2480 // TODO: Use a more targeted data structure.
2481 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2482 Formula F = LU.Formulae[L];
2483 // Use the immediate in the scaled register.
2484 if (F.ScaledReg == OrigReg) {
2485 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2486 Imm * (uint64_t)F.AM.Scale;
2487 // Don't create 50 + reg(-50).
2488 if (F.referencesReg(SE.getSCEV(
2489 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2492 NewF.AM.BaseOffs = Offs;
2493 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2494 LU.Kind, LU.AccessTy, TLI))
2496 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2498 // If the new scale is a constant in a register, and adding the constant
2499 // value to the immediate would produce a value closer to zero than the
2500 // immediate itself, then the formula isn't worthwhile.
2501 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2502 if (C->getValue()->getValue().isNegative() !=
2503 (NewF.AM.BaseOffs < 0) &&
2504 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2505 .ule(abs64(NewF.AM.BaseOffs)))
2509 (void)InsertFormula(LU, LUIdx, NewF);
2511 // Use the immediate in a base register.
2512 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2513 const SCEV *BaseReg = F.BaseRegs[N];
2514 if (BaseReg != OrigReg)
2517 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2518 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2519 LU.Kind, LU.AccessTy, TLI))
2521 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2523 // If the new formula has a constant in a register, and adding the
2524 // constant value to the immediate would produce a value closer to
2525 // zero than the immediate itself, then the formula isn't worthwhile.
2526 for (SmallVectorImpl<const SCEV *>::const_iterator
2527 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2529 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2530 if (C->getValue()->getValue().isNegative() !=
2531 (NewF.AM.BaseOffs < 0) &&
2532 C->getValue()->getValue().abs()
2533 .ule(abs64(NewF.AM.BaseOffs)))
2537 (void)InsertFormula(LU, LUIdx, NewF);
2546 /// GenerateAllReuseFormulae - Generate formulae for each use.
2548 LSRInstance::GenerateAllReuseFormulae() {
2549 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2550 // queries are more precise.
2551 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2552 LSRUse &LU = Uses[LUIdx];
2553 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2554 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2555 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2556 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2558 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2559 LSRUse &LU = Uses[LUIdx];
2560 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2561 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2562 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2563 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2564 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2565 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2566 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2567 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2569 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2570 LSRUse &LU = Uses[LUIdx];
2571 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2572 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2575 GenerateCrossUseConstantOffsets();
2578 /// If their are multiple formulae with the same set of registers used
2579 /// by other uses, pick the best one and delete the others.
2580 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2582 bool Changed = false;
2585 // Collect the best formula for each unique set of shared registers. This
2586 // is reset for each use.
2587 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2589 BestFormulaeTy BestFormulae;
2591 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2592 LSRUse &LU = Uses[LUIdx];
2593 FormulaSorter Sorter(L, LU, SE, DT);
2595 // Clear out the set of used regs; it will be recomputed.
2598 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2599 FIdx != NumForms; ++FIdx) {
2600 Formula &F = LU.Formulae[FIdx];
2602 SmallVector<const SCEV *, 2> Key;
2603 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2604 JE = F.BaseRegs.end(); J != JE; ++J) {
2605 const SCEV *Reg = *J;
2606 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2610 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2611 Key.push_back(F.ScaledReg);
2612 // Unstable sort by host order ok, because this is only used for
2614 std::sort(Key.begin(), Key.end());
2616 std::pair<BestFormulaeTy::const_iterator, bool> P =
2617 BestFormulae.insert(std::make_pair(Key, FIdx));
2619 Formula &Best = LU.Formulae[P.first->second];
2620 if (Sorter.operator()(F, Best))
2622 DEBUG(dbgs() << "Filtering out "; F.print(dbgs());
2624 " in favor of "; Best.print(dbgs());
2629 std::swap(F, LU.Formulae.back());
2630 LU.Formulae.pop_back();
2635 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2636 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2638 BestFormulae.clear();
2641 DEBUG(if (Changed) {
2643 "After filtering out undesirable candidates:\n";
2648 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2649 /// formulae to choose from, use some rough heuristics to prune down the number
2650 /// of formulae. This keeps the main solver from taking an extraordinary amount
2651 /// of time in some worst-case scenarios.
2652 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2653 // This is a rough guess that seems to work fairly well.
2654 const size_t Limit = UINT16_MAX;
2656 SmallPtrSet<const SCEV *, 4> Taken;
2658 // Estimate the worst-case number of solutions we might consider. We almost
2659 // never consider this many solutions because we prune the search space,
2660 // but the pruning isn't always sufficient.
2662 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2663 E = Uses.end(); I != E; ++I) {
2664 size_t FSize = I->Formulae.size();
2665 if (FSize >= Limit) {
2676 // Ok, we have too many of formulae on our hands to conveniently handle.
2677 // Use a rough heuristic to thin out the list.
2679 // Pick the register which is used by the most LSRUses, which is likely
2680 // to be a good reuse register candidate.
2681 const SCEV *Best = 0;
2682 unsigned BestNum = 0;
2683 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2685 const SCEV *Reg = *I;
2686 if (Taken.count(Reg))
2691 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2692 if (Count > BestNum) {
2699 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2700 << " will yield profitable reuse.\n");
2703 // In any use with formulae which references this register, delete formulae
2704 // which don't reference it.
2705 for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(),
2706 E = Uses.end(); I != E; ++I) {
2708 if (!LU.Regs.count(Best)) continue;
2710 // Clear out the set of used regs; it will be recomputed.
2713 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2714 Formula &F = LU.Formulae[i];
2715 if (!F.referencesReg(Best)) {
2716 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2717 std::swap(LU.Formulae.back(), F);
2718 LU.Formulae.pop_back();
2724 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2725 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2729 DEBUG(dbgs() << "After pre-selection:\n";
2730 print_uses(dbgs()));
2734 /// SolveRecurse - This is the recursive solver.
2735 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2737 SmallVectorImpl<const Formula *> &Workspace,
2738 const Cost &CurCost,
2739 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2740 DenseSet<const SCEV *> &VisitedRegs) const {
2743 // - use more aggressive filtering
2744 // - sort the formula so that the most profitable solutions are found first
2745 // - sort the uses too
2747 // - don't compute a cost, and then compare. compare while computing a cost
2749 // - track register sets with SmallBitVector
2751 const LSRUse &LU = Uses[Workspace.size()];
2753 // If this use references any register that's already a part of the
2754 // in-progress solution, consider it a requirement that a formula must
2755 // reference that register in order to be considered. This prunes out
2756 // unprofitable searching.
2757 SmallSetVector<const SCEV *, 4> ReqRegs;
2758 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
2759 E = CurRegs.end(); I != E; ++I)
2760 if (LU.Regs.count(*I))
2763 bool AnySatisfiedReqRegs = false;
2764 SmallPtrSet<const SCEV *, 16> NewRegs;
2767 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2768 E = LU.Formulae.end(); I != E; ++I) {
2769 const Formula &F = *I;
2771 // Ignore formulae which do not use any of the required registers.
2772 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
2773 JE = ReqRegs.end(); J != JE; ++J) {
2774 const SCEV *Reg = *J;
2775 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
2776 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
2780 AnySatisfiedReqRegs = true;
2782 // Evaluate the cost of the current formula. If it's already worse than
2783 // the current best, prune the search at that point.
2786 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
2787 if (NewCost < SolutionCost) {
2788 Workspace.push_back(&F);
2789 if (Workspace.size() != Uses.size()) {
2790 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
2791 NewRegs, VisitedRegs);
2792 if (F.getNumRegs() == 1 && Workspace.size() == 1)
2793 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
2795 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
2796 dbgs() << ". Regs:";
2797 for (SmallPtrSet<const SCEV *, 16>::const_iterator
2798 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
2799 dbgs() << ' ' << **I;
2802 SolutionCost = NewCost;
2803 Solution = Workspace;
2805 Workspace.pop_back();
2810 // If none of the formulae had all of the required registers, relax the
2811 // constraint so that we don't exclude all formulae.
2812 if (!AnySatisfiedReqRegs) {
2818 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
2819 SmallVector<const Formula *, 8> Workspace;
2821 SolutionCost.Loose();
2823 SmallPtrSet<const SCEV *, 16> CurRegs;
2824 DenseSet<const SCEV *> VisitedRegs;
2825 Workspace.reserve(Uses.size());
2827 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
2828 CurRegs, VisitedRegs);
2830 // Ok, we've now made all our decisions.
2831 DEBUG(dbgs() << "\n"
2832 "The chosen solution requires "; SolutionCost.print(dbgs());
2834 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
2836 Uses[i].print(dbgs());
2839 Solution[i]->print(dbgs());
2844 /// getImmediateDominator - A handy utility for the specific DominatorTree
2845 /// query that we need here.
2847 static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
2848 DomTreeNode *Node = DT.getNode(BB);
2849 if (!Node) return 0;
2850 Node = Node->getIDom();
2851 if (!Node) return 0;
2852 return Node->getBlock();
2855 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
2856 /// the dominator tree far as we can go while still being dominated by the
2857 /// input positions. This helps canonicalize the insert position, which
2858 /// encourages sharing.
2859 BasicBlock::iterator
2860 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
2861 const SmallVectorImpl<Instruction *> &Inputs)
2864 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
2865 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
2868 for (BasicBlock *Rung = IP->getParent(); ; Rung = IDom) {
2869 IDom = getImmediateDominator(Rung, DT);
2870 if (!IDom) return IP;
2872 // Don't climb into a loop though.
2873 const Loop *IDomLoop = LI.getLoopFor(IDom);
2874 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
2875 if (IDomDepth <= IPLoopDepth &&
2876 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
2880 bool AllDominate = true;
2881 Instruction *BetterPos = 0;
2882 Instruction *Tentative = IDom->getTerminator();
2883 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
2884 E = Inputs.end(); I != E; ++I) {
2885 Instruction *Inst = *I;
2886 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
2887 AllDominate = false;
2890 // Attempt to find an insert position in the middle of the block,
2891 // instead of at the end, so that it can be used for other expansions.
2892 if (IDom == Inst->getParent() &&
2893 (!BetterPos || DT.dominates(BetterPos, Inst)))
2894 BetterPos = next(BasicBlock::iterator(Inst));
2907 /// AdjustInsertPositionForExpand - Determine an input position which will be
2908 /// dominated by the operands and which will dominate the result.
2909 BasicBlock::iterator
2910 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2912 const LSRUse &LU) const {
2913 // Collect some instructions which must be dominated by the
2914 // expanding replacement. These must be dominated by any operands that
2915 // will be required in the expansion.
2916 SmallVector<Instruction *, 4> Inputs;
2917 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
2918 Inputs.push_back(I);
2919 if (LU.Kind == LSRUse::ICmpZero)
2920 if (Instruction *I =
2921 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
2922 Inputs.push_back(I);
2923 if (LF.PostIncLoops.count(L)) {
2924 if (LF.isUseFullyOutsideLoop(L))
2925 Inputs.push_back(L->getLoopLatch()->getTerminator());
2927 Inputs.push_back(IVIncInsertPos);
2929 // The expansion must also be dominated by the increment positions of any
2930 // loops it for which it is using post-inc mode.
2931 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
2932 E = LF.PostIncLoops.end(); I != E; ++I) {
2933 const Loop *PIL = *I;
2934 if (PIL == L) continue;
2936 // Be dominated by the loop exit.
2937 SmallVector<BasicBlock *, 4> ExitingBlocks;
2938 PIL->getExitingBlocks(ExitingBlocks);
2939 if (!ExitingBlocks.empty()) {
2940 BasicBlock *BB = ExitingBlocks[0];
2941 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
2942 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
2943 Inputs.push_back(BB->getTerminator());
2947 // Then, climb up the immediate dominator tree as far as we can go while
2948 // still being dominated by the input positions.
2949 IP = HoistInsertPosition(IP, Inputs);
2951 // Don't insert instructions before PHI nodes.
2952 while (isa<PHINode>(IP)) ++IP;
2954 // Ignore debug intrinsics.
2955 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
2960 Value *LSRInstance::Expand(const LSRFixup &LF,
2962 BasicBlock::iterator IP,
2963 SCEVExpander &Rewriter,
2964 SmallVectorImpl<WeakVH> &DeadInsts) const {
2965 const LSRUse &LU = Uses[LF.LUIdx];
2967 // Determine an input position which will be dominated by the operands and
2968 // which will dominate the result.
2969 IP = AdjustInsertPositionForExpand(IP, LF, LU);
2971 // Inform the Rewriter if we have a post-increment use, so that it can
2972 // perform an advantageous expansion.
2973 Rewriter.setPostInc(LF.PostIncLoops);
2975 // This is the type that the user actually needs.
2976 const Type *OpTy = LF.OperandValToReplace->getType();
2977 // This will be the type that we'll initially expand to.
2978 const Type *Ty = F.getType();
2980 // No type known; just expand directly to the ultimate type.
2982 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
2983 // Expand directly to the ultimate type if it's the right size.
2985 // This is the type to do integer arithmetic in.
2986 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
2988 // Build up a list of operands to add together to form the full base.
2989 SmallVector<const SCEV *, 8> Ops;
2991 // Expand the BaseRegs portion.
2992 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2993 E = F.BaseRegs.end(); I != E; ++I) {
2994 const SCEV *Reg = *I;
2995 assert(!Reg->isZero() && "Zero allocated in a base register!");
2997 // If we're expanding for a post-inc user, make the post-inc adjustment.
2998 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
2999 Reg = TransformForPostIncUse(Denormalize, Reg,
3000 LF.UserInst, LF.OperandValToReplace,
3003 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3006 // Flush the operand list to suppress SCEVExpander hoisting.
3008 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3010 Ops.push_back(SE.getUnknown(FullV));
3013 // Expand the ScaledReg portion.
3014 Value *ICmpScaledV = 0;
3015 if (F.AM.Scale != 0) {
3016 const SCEV *ScaledS = F.ScaledReg;
3018 // If we're expanding for a post-inc user, make the post-inc adjustment.
3019 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3020 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3021 LF.UserInst, LF.OperandValToReplace,
3024 if (LU.Kind == LSRUse::ICmpZero) {
3025 // An interesting way of "folding" with an icmp is to use a negated
3026 // scale, which we'll implement by inserting it into the other operand
3028 assert(F.AM.Scale == -1 &&
3029 "The only scale supported by ICmpZero uses is -1!");
3030 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3032 // Otherwise just expand the scaled register and an explicit scale,
3033 // which is expected to be matched as part of the address.
3034 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3035 ScaledS = SE.getMulExpr(ScaledS,
3036 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3037 Ops.push_back(ScaledS);
3039 // Flush the operand list to suppress SCEVExpander hoisting.
3040 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3042 Ops.push_back(SE.getUnknown(FullV));
3046 // Expand the GV portion.
3048 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3050 // Flush the operand list to suppress SCEVExpander hoisting.
3051 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3053 Ops.push_back(SE.getUnknown(FullV));
3056 // Expand the immediate portion.
3057 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3059 if (LU.Kind == LSRUse::ICmpZero) {
3060 // The other interesting way of "folding" with an ICmpZero is to use a
3061 // negated immediate.
3063 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3065 Ops.push_back(SE.getUnknown(ICmpScaledV));
3066 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3069 // Just add the immediate values. These again are expected to be matched
3070 // as part of the address.
3071 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3075 // Emit instructions summing all the operands.
3076 const SCEV *FullS = Ops.empty() ?
3077 SE.getConstant(IntTy, 0) :
3079 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3081 // We're done expanding now, so reset the rewriter.
3082 Rewriter.clearPostInc();
3084 // An ICmpZero Formula represents an ICmp which we're handling as a
3085 // comparison against zero. Now that we've expanded an expression for that
3086 // form, update the ICmp's other operand.
3087 if (LU.Kind == LSRUse::ICmpZero) {
3088 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3089 DeadInsts.push_back(CI->getOperand(1));
3090 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3091 "a scale at the same time!");
3092 if (F.AM.Scale == -1) {
3093 if (ICmpScaledV->getType() != OpTy) {
3095 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3097 ICmpScaledV, OpTy, "tmp", CI);
3100 CI->setOperand(1, ICmpScaledV);
3102 assert(F.AM.Scale == 0 &&
3103 "ICmp does not support folding a global value and "
3104 "a scale at the same time!");
3105 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3107 if (C->getType() != OpTy)
3108 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3112 CI->setOperand(1, C);
3119 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3120 /// of their operands effectively happens in their predecessor blocks, so the
3121 /// expression may need to be expanded in multiple places.
3122 void LSRInstance::RewriteForPHI(PHINode *PN,
3125 SCEVExpander &Rewriter,
3126 SmallVectorImpl<WeakVH> &DeadInsts,
3128 DenseMap<BasicBlock *, Value *> Inserted;
3129 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3130 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3131 BasicBlock *BB = PN->getIncomingBlock(i);
3133 // If this is a critical edge, split the edge so that we do not insert
3134 // the code on all predecessor/successor paths. We do this unless this
3135 // is the canonical backedge for this loop, which complicates post-inc
3137 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3138 !isa<IndirectBrInst>(BB->getTerminator()) &&
3139 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3140 // Split the critical edge.
3141 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3143 // If PN is outside of the loop and BB is in the loop, we want to
3144 // move the block to be immediately before the PHI block, not
3145 // immediately after BB.
3146 if (L->contains(BB) && !L->contains(PN))
3147 NewBB->moveBefore(PN->getParent());
3149 // Splitting the edge can reduce the number of PHI entries we have.
3150 e = PN->getNumIncomingValues();
3152 i = PN->getBasicBlockIndex(BB);
3155 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3156 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3158 PN->setIncomingValue(i, Pair.first->second);
3160 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3162 // If this is reuse-by-noop-cast, insert the noop cast.
3163 const Type *OpTy = LF.OperandValToReplace->getType();
3164 if (FullV->getType() != OpTy)
3166 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3168 FullV, LF.OperandValToReplace->getType(),
3169 "tmp", BB->getTerminator());
3171 PN->setIncomingValue(i, FullV);
3172 Pair.first->second = FullV;
3177 /// Rewrite - Emit instructions for the leading candidate expression for this
3178 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3179 /// the newly expanded value.
3180 void LSRInstance::Rewrite(const LSRFixup &LF,
3182 SCEVExpander &Rewriter,
3183 SmallVectorImpl<WeakVH> &DeadInsts,
3185 // First, find an insertion point that dominates UserInst. For PHI nodes,
3186 // find the nearest block which dominates all the relevant uses.
3187 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3188 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3190 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3192 // If this is reuse-by-noop-cast, insert the noop cast.
3193 const Type *OpTy = LF.OperandValToReplace->getType();
3194 if (FullV->getType() != OpTy) {
3196 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3197 FullV, OpTy, "tmp", LF.UserInst);
3201 // Update the user. ICmpZero is handled specially here (for now) because
3202 // Expand may have updated one of the operands of the icmp already, and
3203 // its new value may happen to be equal to LF.OperandValToReplace, in
3204 // which case doing replaceUsesOfWith leads to replacing both operands
3205 // with the same value. TODO: Reorganize this.
3206 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3207 LF.UserInst->setOperand(0, FullV);
3209 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3212 DeadInsts.push_back(LF.OperandValToReplace);
3216 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3218 // Keep track of instructions we may have made dead, so that
3219 // we can remove them after we are done working.
3220 SmallVector<WeakVH, 16> DeadInsts;
3222 SCEVExpander Rewriter(SE);
3223 Rewriter.disableCanonicalMode();
3224 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3226 // Expand the new value definitions and update the users.
3227 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3228 size_t LUIdx = Fixups[i].LUIdx;
3230 Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
3235 // Clean up after ourselves. This must be done before deleting any
3239 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3242 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3243 : IU(P->getAnalysis<IVUsers>()),
3244 SE(P->getAnalysis<ScalarEvolution>()),
3245 DT(P->getAnalysis<DominatorTree>()),
3246 LI(P->getAnalysis<LoopInfo>()),
3247 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3249 // If LoopSimplify form is not available, stay out of trouble.
3250 if (!L->isLoopSimplifyForm()) return;
3252 // If there's no interesting work to be done, bail early.
3253 if (IU.empty()) return;
3255 DEBUG(dbgs() << "\nLSR on loop ";
3256 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3259 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3260 /// inside the loop then try to eliminate the cast operation.
3263 // Change loop terminating condition to use the postinc iv when possible.
3264 Changed |= OptimizeLoopTermCond();
3266 CollectInterestingTypesAndFactors();
3267 CollectFixupsAndInitialFormulae();
3268 CollectLoopInvariantFixupsAndFormulae();
3270 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3271 print_uses(dbgs()));
3273 // Now use the reuse data to generate a bunch of interesting ways
3274 // to formulate the values needed for the uses.
3275 GenerateAllReuseFormulae();
3277 DEBUG(dbgs() << "\n"
3278 "After generating reuse formulae:\n";
3279 print_uses(dbgs()));
3281 FilterOutUndesirableDedicatedRegisters();
3282 NarrowSearchSpaceUsingHeuristics();
3284 SmallVector<const Formula *, 8> Solution;
3286 assert(Solution.size() == Uses.size() && "Malformed solution!");
3288 // Release memory that is no longer needed.
3294 // Formulae should be legal.
3295 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3296 E = Uses.end(); I != E; ++I) {
3297 const LSRUse &LU = *I;
3298 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3299 JE = LU.Formulae.end(); J != JE; ++J)
3300 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3301 LU.Kind, LU.AccessTy, TLI) &&
3302 "Illegal formula generated!");
3306 // Now that we've decided what we want, make it so.
3307 ImplementSolution(Solution, P);
3310 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3311 if (Factors.empty() && Types.empty()) return;
3313 OS << "LSR has identified the following interesting factors and types: ";
3316 for (SmallSetVector<int64_t, 8>::const_iterator
3317 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3318 if (!First) OS << ", ";
3323 for (SmallSetVector<const Type *, 4>::const_iterator
3324 I = Types.begin(), E = Types.end(); I != E; ++I) {
3325 if (!First) OS << ", ";
3327 OS << '(' << **I << ')';
3332 void LSRInstance::print_fixups(raw_ostream &OS) const {
3333 OS << "LSR is examining the following fixup sites:\n";
3334 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3335 E = Fixups.end(); I != E; ++I) {
3336 const LSRFixup &LF = *I;
3343 void LSRInstance::print_uses(raw_ostream &OS) const {
3344 OS << "LSR is examining the following uses:\n";
3345 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3346 E = Uses.end(); I != E; ++I) {
3347 const LSRUse &LU = *I;
3351 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3352 JE = LU.Formulae.end(); J != JE; ++J) {
3360 void LSRInstance::print(raw_ostream &OS) const {
3361 print_factors_and_types(OS);
3366 void LSRInstance::dump() const {
3367 print(errs()); errs() << '\n';
3372 class LoopStrengthReduce : public LoopPass {
3373 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3374 /// transformation profitability.
3375 const TargetLowering *const TLI;
3378 static char ID; // Pass ID, replacement for typeid
3379 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3382 bool runOnLoop(Loop *L, LPPassManager &LPM);
3383 void getAnalysisUsage(AnalysisUsage &AU) const;
3388 char LoopStrengthReduce::ID = 0;
3389 static RegisterPass<LoopStrengthReduce>
3390 X("loop-reduce", "Loop Strength Reduction");
3392 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3393 return new LoopStrengthReduce(TLI);
3396 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3397 : LoopPass(&ID), TLI(tli) {}
3399 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3400 // We split critical edges, so we change the CFG. However, we do update
3401 // many analyses if they are around.
3402 AU.addPreservedID(LoopSimplifyID);
3403 AU.addPreserved("domfrontier");
3405 AU.addRequired<LoopInfo>();
3406 AU.addPreserved<LoopInfo>();
3407 AU.addRequiredID(LoopSimplifyID);
3408 AU.addRequired<DominatorTree>();
3409 AU.addPreserved<DominatorTree>();
3410 AU.addRequired<ScalarEvolution>();
3411 AU.addPreserved<ScalarEvolution>();
3412 AU.addRequired<IVUsers>();
3413 AU.addPreserved<IVUsers>();
3416 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3417 bool Changed = false;
3419 // Run the main LSR transformation.
3420 Changed |= LSRInstance(TLI, L, this).getChanged();
3422 // At this point, it is worth checking to see if any recurrence PHIs are also
3423 // dead, so that we can remove them as well.
3424 Changed |= DeleteDeadPHIs(L->getHeader());