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.getIntegerSCEV(0, AR->getType()),
225 AR->getStepRecurrence(SE),
227 L, Good, Bad, SE, DT);
231 // Handle a multiplication by -1 (negation) if it didn't fold.
232 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
233 if (Mul->getOperand(0)->isAllOnesValue()) {
234 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
235 const SCEV *NewMul = SE.getMulExpr(Ops);
237 SmallVector<const SCEV *, 4> MyGood;
238 SmallVector<const SCEV *, 4> MyBad;
239 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
240 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
241 SE.getEffectiveSCEVType(NewMul->getType())));
242 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
243 E = MyGood.end(); I != E; ++I)
244 Good.push_back(SE.getMulExpr(NegOne, *I));
245 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
246 E = MyBad.end(); I != E; ++I)
247 Bad.push_back(SE.getMulExpr(NegOne, *I));
251 // Ok, we can't do anything interesting. Just stuff the whole thing into a
252 // register and hope for the best.
256 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
257 /// attempting to keep all loop-invariant and loop-computable values in a
258 /// single base register.
259 void Formula::InitialMatch(const SCEV *S, Loop *L,
260 ScalarEvolution &SE, DominatorTree &DT) {
261 SmallVector<const SCEV *, 4> Good;
262 SmallVector<const SCEV *, 4> Bad;
263 DoInitialMatch(S, L, Good, Bad, SE, DT);
265 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.getIntegerSCEV(1, LHS->getType());
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.getIntegerSCEV(0, C->getType());
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.getIntegerSCEV(0, GV->getType());
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.getIntegerSCEV(1, BackedgeTakenCount->getType());
1462 // Add one to the backedge-taken count to get the trip count.
1463 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1465 // Check for a max calculation that matches the pattern.
1466 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
1468 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
1469 if (Max != SE.getSCEV(Sel)) return Cond;
1471 // To handle a max with more than two operands, this optimization would
1472 // require additional checking and setup.
1473 if (Max->getNumOperands() != 2)
1476 const SCEV *MaxLHS = Max->getOperand(0);
1477 const SCEV *MaxRHS = Max->getOperand(1);
1478 if (!MaxLHS || MaxLHS != One) return Cond;
1479 // Check the relevant induction variable for conformance to
1481 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1482 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1483 if (!AR || !AR->isAffine() ||
1484 AR->getStart() != One ||
1485 AR->getStepRecurrence(SE) != One)
1488 assert(AR->getLoop() == L &&
1489 "Loop condition operand is an addrec in a different loop!");
1491 // Check the right operand of the select, and remember it, as it will
1492 // be used in the new comparison instruction.
1494 if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1495 NewRHS = Sel->getOperand(1);
1496 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1497 NewRHS = Sel->getOperand(2);
1498 if (!NewRHS) return Cond;
1500 // Determine the new comparison opcode. It may be signed or unsigned,
1501 // and the original comparison may be either equality or inequality.
1502 CmpInst::Predicate Pred =
1503 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
1504 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1505 Pred = CmpInst::getInversePredicate(Pred);
1507 // Ok, everything looks ok to change the condition into an SLT or SGE and
1508 // delete the max calculation.
1510 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1512 // Delete the max calculation instructions.
1513 Cond->replaceAllUsesWith(NewCond);
1514 CondUse->setUser(NewCond);
1515 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1516 Cond->eraseFromParent();
1517 Sel->eraseFromParent();
1518 if (Cmp->use_empty())
1519 Cmp->eraseFromParent();
1523 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1524 /// postinc iv when possible.
1526 LSRInstance::OptimizeLoopTermCond() {
1527 SmallPtrSet<Instruction *, 4> PostIncs;
1529 BasicBlock *LatchBlock = L->getLoopLatch();
1530 SmallVector<BasicBlock*, 8> ExitingBlocks;
1531 L->getExitingBlocks(ExitingBlocks);
1533 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1534 BasicBlock *ExitingBlock = ExitingBlocks[i];
1536 // Get the terminating condition for the loop if possible. If we
1537 // can, we want to change it to use a post-incremented version of its
1538 // induction variable, to allow coalescing the live ranges for the IV into
1539 // one register value.
1541 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1544 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1545 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1548 // Search IVUsesByStride to find Cond's IVUse if there is one.
1549 IVStrideUse *CondUse = 0;
1550 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1551 if (!FindIVUserForCond(Cond, CondUse))
1554 // If the trip count is computed in terms of a max (due to ScalarEvolution
1555 // being unable to find a sufficient guard, for example), change the loop
1556 // comparison to use SLT or ULT instead of NE.
1557 // One consequence of doing this now is that it disrupts the count-down
1558 // optimization. That's not always a bad thing though, because in such
1559 // cases it may still be worthwhile to avoid a max.
1560 Cond = OptimizeMax(Cond, CondUse);
1562 // If this exiting block dominates the latch block, it may also use
1563 // the post-inc value if it won't be shared with other uses.
1564 // Check for dominance.
1565 if (!DT.dominates(ExitingBlock, LatchBlock))
1568 // Conservatively avoid trying to use the post-inc value in non-latch
1569 // exits if there may be pre-inc users in intervening blocks.
1570 if (LatchBlock != ExitingBlock)
1571 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1572 // Test if the use is reachable from the exiting block. This dominator
1573 // query is a conservative approximation of reachability.
1574 if (&*UI != CondUse &&
1575 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1576 // Conservatively assume there may be reuse if the quotient of their
1577 // strides could be a legal scale.
1578 const SCEV *A = IU.getStride(*CondUse, L);
1579 const SCEV *B = IU.getStride(*UI, L);
1580 if (!A || !B) continue;
1581 if (SE.getTypeSizeInBits(A->getType()) !=
1582 SE.getTypeSizeInBits(B->getType())) {
1583 if (SE.getTypeSizeInBits(A->getType()) >
1584 SE.getTypeSizeInBits(B->getType()))
1585 B = SE.getSignExtendExpr(B, A->getType());
1587 A = SE.getSignExtendExpr(A, B->getType());
1589 if (const SCEVConstant *D =
1590 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1591 // Stride of one or negative one can have reuse with non-addresses.
1592 if (D->getValue()->isOne() ||
1593 D->getValue()->isAllOnesValue())
1594 goto decline_post_inc;
1595 // Avoid weird situations.
1596 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1597 D->getValue()->getValue().isMinSignedValue())
1598 goto decline_post_inc;
1599 // Without TLI, assume that any stride might be valid, and so any
1600 // use might be shared.
1602 goto decline_post_inc;
1603 // Check for possible scaled-address reuse.
1604 const Type *AccessTy = getAccessType(UI->getUser());
1605 TargetLowering::AddrMode AM;
1606 AM.Scale = D->getValue()->getSExtValue();
1607 if (TLI->isLegalAddressingMode(AM, AccessTy))
1608 goto decline_post_inc;
1609 AM.Scale = -AM.Scale;
1610 if (TLI->isLegalAddressingMode(AM, AccessTy))
1611 goto decline_post_inc;
1615 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1618 // It's possible for the setcc instruction to be anywhere in the loop, and
1619 // possible for it to have multiple users. If it is not immediately before
1620 // the exiting block branch, move it.
1621 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1622 if (Cond->hasOneUse()) {
1623 Cond->moveBefore(TermBr);
1625 // Clone the terminating condition and insert into the loopend.
1626 ICmpInst *OldCond = Cond;
1627 Cond = cast<ICmpInst>(Cond->clone());
1628 Cond->setName(L->getHeader()->getName() + ".termcond");
1629 ExitingBlock->getInstList().insert(TermBr, Cond);
1631 // Clone the IVUse, as the old use still exists!
1632 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1633 TermBr->replaceUsesOfWith(OldCond, Cond);
1637 // If we get to here, we know that we can transform the setcc instruction to
1638 // use the post-incremented version of the IV, allowing us to coalesce the
1639 // live ranges for the IV correctly.
1640 CondUse->transformToPostInc(L);
1643 PostIncs.insert(Cond);
1647 // Determine an insertion point for the loop induction variable increment. It
1648 // must dominate all the post-inc comparisons we just set up, and it must
1649 // dominate the loop latch edge.
1650 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1651 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1652 E = PostIncs.end(); I != E; ++I) {
1654 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1656 if (BB == (*I)->getParent())
1657 IVIncInsertPos = *I;
1658 else if (BB != IVIncInsertPos->getParent())
1659 IVIncInsertPos = BB->getTerminator();
1666 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1667 LSRUse::KindType Kind, const Type *AccessTy) {
1668 int64_t NewMinOffset = LU.MinOffset;
1669 int64_t NewMaxOffset = LU.MaxOffset;
1670 const Type *NewAccessTy = AccessTy;
1672 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1673 // something conservative, however this can pessimize in the case that one of
1674 // the uses will have all its uses outside the loop, for example.
1675 if (LU.Kind != Kind)
1677 // Conservatively assume HasBaseReg is true for now.
1678 if (NewOffset < LU.MinOffset) {
1679 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true,
1680 Kind, AccessTy, TLI))
1682 NewMinOffset = NewOffset;
1683 } else if (NewOffset > LU.MaxOffset) {
1684 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true,
1685 Kind, AccessTy, TLI))
1687 NewMaxOffset = NewOffset;
1689 // Check for a mismatched access type, and fall back conservatively as needed.
1690 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1691 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1694 LU.MinOffset = NewMinOffset;
1695 LU.MaxOffset = NewMaxOffset;
1696 LU.AccessTy = NewAccessTy;
1697 if (NewOffset != LU.Offsets.back())
1698 LU.Offsets.push_back(NewOffset);
1702 /// getUse - Return an LSRUse index and an offset value for a fixup which
1703 /// needs the given expression, with the given kind and optional access type.
1704 /// Either reuse an existing use or create a new one, as needed.
1705 std::pair<size_t, int64_t>
1706 LSRInstance::getUse(const SCEV *&Expr,
1707 LSRUse::KindType Kind, const Type *AccessTy) {
1708 const SCEV *Copy = Expr;
1709 int64_t Offset = ExtractImmediate(Expr, SE);
1711 // Basic uses can't accept any offset, for example.
1712 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1717 std::pair<UseMapTy::iterator, bool> P =
1718 UseMap.insert(std::make_pair(Expr, 0));
1720 // A use already existed with this base.
1721 size_t LUIdx = P.first->second;
1722 LSRUse &LU = Uses[LUIdx];
1723 if (reconcileNewOffset(LU, Offset, Kind, AccessTy))
1725 return std::make_pair(LUIdx, Offset);
1728 // Create a new use.
1729 size_t LUIdx = Uses.size();
1730 P.first->second = LUIdx;
1731 Uses.push_back(LSRUse(Kind, AccessTy));
1732 LSRUse &LU = Uses[LUIdx];
1734 // We don't need to track redundant offsets, but we don't need to go out
1735 // of our way here to avoid them.
1736 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1737 LU.Offsets.push_back(Offset);
1739 LU.MinOffset = Offset;
1740 LU.MaxOffset = Offset;
1741 return std::make_pair(LUIdx, Offset);
1744 void LSRInstance::CollectInterestingTypesAndFactors() {
1745 SmallSetVector<const SCEV *, 4> Strides;
1747 // Collect interesting types and strides.
1748 SmallVector<const SCEV *, 4> Worklist;
1749 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1750 const SCEV *Expr = IU.getExpr(*UI);
1752 // Collect interesting types.
1753 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1755 // Add strides for mentioned loops.
1756 Worklist.push_back(Expr);
1758 const SCEV *S = Worklist.pop_back_val();
1759 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1760 Strides.insert(AR->getStepRecurrence(SE));
1761 Worklist.push_back(AR->getStart());
1762 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1763 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1765 } while (!Worklist.empty());
1768 // Compute interesting factors from the set of interesting strides.
1769 for (SmallSetVector<const SCEV *, 4>::const_iterator
1770 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1771 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1772 next(I); NewStrideIter != E; ++NewStrideIter) {
1773 const SCEV *OldStride = *I;
1774 const SCEV *NewStride = *NewStrideIter;
1776 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1777 SE.getTypeSizeInBits(NewStride->getType())) {
1778 if (SE.getTypeSizeInBits(OldStride->getType()) >
1779 SE.getTypeSizeInBits(NewStride->getType()))
1780 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1782 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1784 if (const SCEVConstant *Factor =
1785 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1787 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1788 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1789 } else if (const SCEVConstant *Factor =
1790 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1793 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1794 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1798 // If all uses use the same type, don't bother looking for truncation-based
1800 if (Types.size() == 1)
1803 DEBUG(print_factors_and_types(dbgs()));
1806 void LSRInstance::CollectFixupsAndInitialFormulae() {
1807 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1809 LSRFixup &LF = getNewFixup();
1810 LF.UserInst = UI->getUser();
1811 LF.OperandValToReplace = UI->getOperandValToReplace();
1812 LF.PostIncLoops = UI->getPostIncLoops();
1814 LSRUse::KindType Kind = LSRUse::Basic;
1815 const Type *AccessTy = 0;
1816 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1817 Kind = LSRUse::Address;
1818 AccessTy = getAccessType(LF.UserInst);
1821 const SCEV *S = IU.getExpr(*UI);
1823 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1824 // (N - i == 0), and this allows (N - i) to be the expression that we work
1825 // with rather than just N or i, so we can consider the register
1826 // requirements for both N and i at the same time. Limiting this code to
1827 // equality icmps is not a problem because all interesting loops use
1828 // equality icmps, thanks to IndVarSimplify.
1829 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1830 if (CI->isEquality()) {
1831 // Swap the operands if needed to put the OperandValToReplace on the
1832 // left, for consistency.
1833 Value *NV = CI->getOperand(1);
1834 if (NV == LF.OperandValToReplace) {
1835 CI->setOperand(1, CI->getOperand(0));
1836 CI->setOperand(0, NV);
1839 // x == y --> x - y == 0
1840 const SCEV *N = SE.getSCEV(NV);
1841 if (N->isLoopInvariant(L)) {
1842 Kind = LSRUse::ICmpZero;
1843 S = SE.getMinusSCEV(N, S);
1846 // -1 and the negations of all interesting strides (except the negation
1847 // of -1) are now also interesting.
1848 for (size_t i = 0, e = Factors.size(); i != e; ++i)
1849 if (Factors[i] != -1)
1850 Factors.insert(-(uint64_t)Factors[i]);
1854 // Set up the initial formula for this use.
1855 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
1857 LF.Offset = P.second;
1858 LSRUse &LU = Uses[LF.LUIdx];
1859 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
1861 // If this is the first use of this LSRUse, give it a formula.
1862 if (LU.Formulae.empty()) {
1863 InsertInitialFormula(S, LU, LF.LUIdx);
1864 CountRegisters(LU.Formulae.back(), LF.LUIdx);
1868 DEBUG(print_fixups(dbgs()));
1872 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
1874 F.InitialMatch(S, L, SE, DT);
1875 bool Inserted = InsertFormula(LU, LUIdx, F);
1876 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
1880 LSRInstance::InsertSupplementalFormula(const SCEV *S,
1881 LSRUse &LU, size_t LUIdx) {
1883 F.BaseRegs.push_back(S);
1884 F.AM.HasBaseReg = true;
1885 bool Inserted = InsertFormula(LU, LUIdx, F);
1886 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
1889 /// CountRegisters - Note which registers are used by the given formula,
1890 /// updating RegUses.
1891 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
1893 RegUses.CountRegister(F.ScaledReg, LUIdx);
1894 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1895 E = F.BaseRegs.end(); I != E; ++I)
1896 RegUses.CountRegister(*I, LUIdx);
1899 /// InsertFormula - If the given formula has not yet been inserted, add it to
1900 /// the list, and return true. Return false otherwise.
1901 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
1902 if (!LU.InsertFormula(F))
1905 CountRegisters(F, LUIdx);
1909 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
1910 /// loop-invariant values which we're tracking. These other uses will pin these
1911 /// values in registers, making them less profitable for elimination.
1912 /// TODO: This currently misses non-constant addrec step registers.
1913 /// TODO: Should this give more weight to users inside the loop?
1915 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
1916 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
1917 SmallPtrSet<const SCEV *, 8> Inserted;
1919 while (!Worklist.empty()) {
1920 const SCEV *S = Worklist.pop_back_val();
1922 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
1923 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
1924 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1925 Worklist.push_back(C->getOperand());
1926 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1927 Worklist.push_back(D->getLHS());
1928 Worklist.push_back(D->getRHS());
1929 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
1930 if (!Inserted.insert(U)) continue;
1931 const Value *V = U->getValue();
1932 if (const Instruction *Inst = dyn_cast<Instruction>(V))
1933 if (L->contains(Inst)) continue;
1934 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1936 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
1937 // Ignore non-instructions.
1940 // Ignore instructions in other functions (as can happen with
1942 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
1944 // Ignore instructions not dominated by the loop.
1945 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
1946 UserInst->getParent() :
1947 cast<PHINode>(UserInst)->getIncomingBlock(
1948 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
1949 if (!DT.dominates(L->getHeader(), UseBB))
1951 // Ignore uses which are part of other SCEV expressions, to avoid
1952 // analyzing them multiple times.
1953 if (SE.isSCEVable(UserInst->getType())) {
1954 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
1955 // If the user is a no-op, look through to its uses.
1956 if (!isa<SCEVUnknown>(UserS))
1960 SE.getUnknown(const_cast<Instruction *>(UserInst)));
1964 // Ignore icmp instructions which are already being analyzed.
1965 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
1966 unsigned OtherIdx = !UI.getOperandNo();
1967 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
1968 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
1972 LSRFixup &LF = getNewFixup();
1973 LF.UserInst = const_cast<Instruction *>(UserInst);
1974 LF.OperandValToReplace = UI.getUse();
1975 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
1977 LF.Offset = P.second;
1978 LSRUse &LU = Uses[LF.LUIdx];
1979 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
1980 InsertSupplementalFormula(U, LU, LF.LUIdx);
1981 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
1988 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
1989 /// separate registers. If C is non-null, multiply each subexpression by C.
1990 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
1991 SmallVectorImpl<const SCEV *> &Ops,
1992 ScalarEvolution &SE) {
1993 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1994 // Break out add operands.
1995 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1997 CollectSubexprs(*I, C, Ops, SE);
1999 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2000 // Split a non-zero base out of an addrec.
2001 if (!AR->getStart()->isZero()) {
2002 CollectSubexprs(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
2003 AR->getStepRecurrence(SE),
2004 AR->getLoop()), C, Ops, SE);
2005 CollectSubexprs(AR->getStart(), C, Ops, SE);
2008 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2009 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2010 if (Mul->getNumOperands() == 2)
2011 if (const SCEVConstant *Op0 =
2012 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2013 CollectSubexprs(Mul->getOperand(1),
2014 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2020 // Otherwise use the value itself.
2021 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2024 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2026 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2029 // Arbitrarily cap recursion to protect compile time.
2030 if (Depth >= 3) return;
2032 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2033 const SCEV *BaseReg = Base.BaseRegs[i];
2035 SmallVector<const SCEV *, 8> AddOps;
2036 CollectSubexprs(BaseReg, 0, AddOps, SE);
2037 if (AddOps.size() == 1) continue;
2039 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2040 JE = AddOps.end(); J != JE; ++J) {
2041 // Don't pull a constant into a register if the constant could be folded
2042 // into an immediate field.
2043 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2044 Base.getNumRegs() > 1,
2045 LU.Kind, LU.AccessTy, TLI, SE))
2048 // Collect all operands except *J.
2049 SmallVector<const SCEV *, 8> InnerAddOps;
2050 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2051 KE = AddOps.end(); K != KE; ++K)
2053 InnerAddOps.push_back(*K);
2055 // Don't leave just a constant behind in a register if the constant could
2056 // be folded into an immediate field.
2057 if (InnerAddOps.size() == 1 &&
2058 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2059 Base.getNumRegs() > 1,
2060 LU.Kind, LU.AccessTy, TLI, SE))
2063 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2064 if (InnerSum->isZero())
2067 F.BaseRegs[i] = InnerSum;
2068 F.BaseRegs.push_back(*J);
2069 if (InsertFormula(LU, LUIdx, F))
2070 // If that formula hadn't been seen before, recurse to find more like
2072 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2077 /// GenerateCombinations - Generate a formula consisting of all of the
2078 /// loop-dominating registers added into a single register.
2079 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2081 // This method is only interesting on a plurality of registers.
2082 if (Base.BaseRegs.size() <= 1) return;
2086 SmallVector<const SCEV *, 4> Ops;
2087 for (SmallVectorImpl<const SCEV *>::const_iterator
2088 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2089 const SCEV *BaseReg = *I;
2090 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2091 !BaseReg->hasComputableLoopEvolution(L))
2092 Ops.push_back(BaseReg);
2094 F.BaseRegs.push_back(BaseReg);
2096 if (Ops.size() > 1) {
2097 const SCEV *Sum = SE.getAddExpr(Ops);
2098 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2099 // opportunity to fold something. For now, just ignore such cases
2100 // rather than proceed with zero in a register.
2101 if (!Sum->isZero()) {
2102 F.BaseRegs.push_back(Sum);
2103 (void)InsertFormula(LU, LUIdx, F);
2108 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2109 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2111 // We can't add a symbolic offset if the address already contains one.
2112 if (Base.AM.BaseGV) return;
2114 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2115 const SCEV *G = Base.BaseRegs[i];
2116 GlobalValue *GV = ExtractSymbol(G, SE);
2117 if (G->isZero() || !GV)
2121 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2122 LU.Kind, LU.AccessTy, TLI))
2125 (void)InsertFormula(LU, LUIdx, F);
2129 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2130 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2132 // TODO: For now, just add the min and max offset, because it usually isn't
2133 // worthwhile looking at everything inbetween.
2134 SmallVector<int64_t, 4> Worklist;
2135 Worklist.push_back(LU.MinOffset);
2136 if (LU.MaxOffset != LU.MinOffset)
2137 Worklist.push_back(LU.MaxOffset);
2139 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2140 const SCEV *G = Base.BaseRegs[i];
2142 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2143 E = Worklist.end(); I != E; ++I) {
2145 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2146 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2147 LU.Kind, LU.AccessTy, TLI)) {
2148 F.BaseRegs[i] = SE.getAddExpr(G, SE.getIntegerSCEV(*I, G->getType()));
2150 (void)InsertFormula(LU, LUIdx, F);
2154 int64_t Imm = ExtractImmediate(G, SE);
2155 if (G->isZero() || Imm == 0)
2158 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2159 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2160 LU.Kind, LU.AccessTy, TLI))
2163 (void)InsertFormula(LU, LUIdx, F);
2167 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2168 /// the comparison. For example, x == y -> x*c == y*c.
2169 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2171 if (LU.Kind != LSRUse::ICmpZero) return;
2173 // Determine the integer type for the base formula.
2174 const Type *IntTy = Base.getType();
2176 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2178 // Don't do this if there is more than one offset.
2179 if (LU.MinOffset != LU.MaxOffset) return;
2181 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2183 // Check each interesting stride.
2184 for (SmallSetVector<int64_t, 8>::const_iterator
2185 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2186 int64_t Factor = *I;
2189 // Check that the multiplication doesn't overflow.
2190 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2192 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2193 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2196 // Check that multiplying with the use offset doesn't overflow.
2197 int64_t Offset = LU.MinOffset;
2198 if (Offset == INT64_MIN && Factor == -1)
2200 Offset = (uint64_t)Offset * Factor;
2201 if (Offset / Factor != LU.MinOffset)
2204 // Check that this scale is legal.
2205 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2208 // Compensate for the use having MinOffset built into it.
2209 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2211 const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
2213 // Check that multiplying with each base register doesn't overflow.
2214 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2215 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2216 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2220 // Check that multiplying with the scaled register doesn't overflow.
2222 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2223 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2227 // If we make it here and it's legal, add it.
2228 (void)InsertFormula(LU, LUIdx, F);
2233 /// GenerateScales - Generate stride factor reuse formulae by making use of
2234 /// scaled-offset address modes, for example.
2235 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
2237 // Determine the integer type for the base formula.
2238 const Type *IntTy = Base.getType();
2241 // If this Formula already has a scaled register, we can't add another one.
2242 if (Base.AM.Scale != 0) return;
2244 // Check each interesting stride.
2245 for (SmallSetVector<int64_t, 8>::const_iterator
2246 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2247 int64_t Factor = *I;
2249 Base.AM.Scale = Factor;
2250 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2251 // Check whether this scale is going to be legal.
2252 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2253 LU.Kind, LU.AccessTy, TLI)) {
2254 // As a special-case, handle special out-of-loop Basic users specially.
2255 // TODO: Reconsider this special case.
2256 if (LU.Kind == LSRUse::Basic &&
2257 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2258 LSRUse::Special, LU.AccessTy, TLI) &&
2259 LU.AllFixupsOutsideLoop)
2260 LU.Kind = LSRUse::Special;
2264 // For an ICmpZero, negating a solitary base register won't lead to
2266 if (LU.Kind == LSRUse::ICmpZero &&
2267 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2269 // For each addrec base reg, apply the scale, if possible.
2270 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2271 if (const SCEVAddRecExpr *AR =
2272 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2273 const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
2274 if (FactorS->isZero())
2276 // Divide out the factor, ignoring high bits, since we'll be
2277 // scaling the value back up in the end.
2278 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2279 // TODO: This could be optimized to avoid all the copying.
2281 F.ScaledReg = Quotient;
2282 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2283 F.BaseRegs.pop_back();
2284 (void)InsertFormula(LU, LUIdx, F);
2290 /// GenerateTruncates - Generate reuse formulae from different IV types.
2291 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
2293 // This requires TargetLowering to tell us which truncates are free.
2296 // Don't bother truncating symbolic values.
2297 if (Base.AM.BaseGV) return;
2299 // Determine the integer type for the base formula.
2300 const Type *DstTy = Base.getType();
2302 DstTy = SE.getEffectiveSCEVType(DstTy);
2304 for (SmallSetVector<const Type *, 4>::const_iterator
2305 I = Types.begin(), E = Types.end(); I != E; ++I) {
2306 const Type *SrcTy = *I;
2307 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2310 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2311 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2312 JE = F.BaseRegs.end(); J != JE; ++J)
2313 *J = SE.getAnyExtendExpr(*J, SrcTy);
2315 // TODO: This assumes we've done basic processing on all uses and
2316 // have an idea what the register usage is.
2317 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2320 (void)InsertFormula(LU, LUIdx, F);
2327 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2328 /// defer modifications so that the search phase doesn't have to worry about
2329 /// the data structures moving underneath it.
2333 const SCEV *OrigReg;
2335 WorkItem(size_t LI, int64_t I, const SCEV *R)
2336 : LUIdx(LI), Imm(I), OrigReg(R) {}
2338 void print(raw_ostream &OS) const;
2344 void WorkItem::print(raw_ostream &OS) const {
2345 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2346 << " , add offset " << Imm;
2349 void WorkItem::dump() const {
2350 print(errs()); errs() << '\n';
2353 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2354 /// distance apart and try to form reuse opportunities between them.
2355 void LSRInstance::GenerateCrossUseConstantOffsets() {
2356 // Group the registers by their value without any added constant offset.
2357 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2358 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2360 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2361 SmallVector<const SCEV *, 8> Sequence;
2362 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2364 const SCEV *Reg = *I;
2365 int64_t Imm = ExtractImmediate(Reg, SE);
2366 std::pair<RegMapTy::iterator, bool> Pair =
2367 Map.insert(std::make_pair(Reg, ImmMapTy()));
2369 Sequence.push_back(Reg);
2370 Pair.first->second.insert(std::make_pair(Imm, *I));
2371 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2374 // Now examine each set of registers with the same base value. Build up
2375 // a list of work to do and do the work in a separate step so that we're
2376 // not adding formulae and register counts while we're searching.
2377 SmallVector<WorkItem, 32> WorkItems;
2378 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2379 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2380 E = Sequence.end(); I != E; ++I) {
2381 const SCEV *Reg = *I;
2382 const ImmMapTy &Imms = Map.find(Reg)->second;
2384 // It's not worthwhile looking for reuse if there's only one offset.
2385 if (Imms.size() == 1)
2388 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2389 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2391 dbgs() << ' ' << J->first;
2394 // Examine each offset.
2395 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2397 const SCEV *OrigReg = J->second;
2399 int64_t JImm = J->first;
2400 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2402 if (!isa<SCEVConstant>(OrigReg) &&
2403 UsedByIndicesMap[Reg].count() == 1) {
2404 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2408 // Conservatively examine offsets between this orig reg a few selected
2410 ImmMapTy::const_iterator OtherImms[] = {
2411 Imms.begin(), prior(Imms.end()),
2412 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2414 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2415 ImmMapTy::const_iterator M = OtherImms[i];
2416 if (M == J || M == JE) continue;
2418 // Compute the difference between the two.
2419 int64_t Imm = (uint64_t)JImm - M->first;
2420 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2421 LUIdx = UsedByIndices.find_next(LUIdx))
2422 // Make a memo of this use, offset, and register tuple.
2423 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2424 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2431 UsedByIndicesMap.clear();
2432 UniqueItems.clear();
2434 // Now iterate through the worklist and add new formulae.
2435 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2436 E = WorkItems.end(); I != E; ++I) {
2437 const WorkItem &WI = *I;
2438 size_t LUIdx = WI.LUIdx;
2439 LSRUse &LU = Uses[LUIdx];
2440 int64_t Imm = WI.Imm;
2441 const SCEV *OrigReg = WI.OrigReg;
2443 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2444 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2445 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2447 // TODO: Use a more targeted data structure.
2448 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2449 Formula F = LU.Formulae[L];
2450 // Use the immediate in the scaled register.
2451 if (F.ScaledReg == OrigReg) {
2452 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2453 Imm * (uint64_t)F.AM.Scale;
2454 // Don't create 50 + reg(-50).
2455 if (F.referencesReg(SE.getSCEV(
2456 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2459 NewF.AM.BaseOffs = Offs;
2460 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2461 LU.Kind, LU.AccessTy, TLI))
2463 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2465 // If the new scale is a constant in a register, and adding the constant
2466 // value to the immediate would produce a value closer to zero than the
2467 // immediate itself, then the formula isn't worthwhile.
2468 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2469 if (C->getValue()->getValue().isNegative() !=
2470 (NewF.AM.BaseOffs < 0) &&
2471 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2472 .ule(abs64(NewF.AM.BaseOffs)))
2476 (void)InsertFormula(LU, LUIdx, NewF);
2478 // Use the immediate in a base register.
2479 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2480 const SCEV *BaseReg = F.BaseRegs[N];
2481 if (BaseReg != OrigReg)
2484 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2485 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2486 LU.Kind, LU.AccessTy, TLI))
2488 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2490 // If the new formula has a constant in a register, and adding the
2491 // constant value to the immediate would produce a value closer to
2492 // zero than the immediate itself, then the formula isn't worthwhile.
2493 for (SmallVectorImpl<const SCEV *>::const_iterator
2494 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2496 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2497 if (C->getValue()->getValue().isNegative() !=
2498 (NewF.AM.BaseOffs < 0) &&
2499 C->getValue()->getValue().abs()
2500 .ule(abs64(NewF.AM.BaseOffs)))
2504 (void)InsertFormula(LU, LUIdx, NewF);
2513 /// GenerateAllReuseFormulae - Generate formulae for each use.
2515 LSRInstance::GenerateAllReuseFormulae() {
2516 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2517 // queries are more precise.
2518 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2519 LSRUse &LU = Uses[LUIdx];
2520 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2521 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2522 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2523 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2525 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2526 LSRUse &LU = Uses[LUIdx];
2527 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2528 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2529 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2530 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2531 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2532 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2533 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2534 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2536 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2537 LSRUse &LU = Uses[LUIdx];
2538 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2539 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2542 GenerateCrossUseConstantOffsets();
2545 /// If their are multiple formulae with the same set of registers used
2546 /// by other uses, pick the best one and delete the others.
2547 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2549 bool Changed = false;
2552 // Collect the best formula for each unique set of shared registers. This
2553 // is reset for each use.
2554 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2556 BestFormulaeTy BestFormulae;
2558 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2559 LSRUse &LU = Uses[LUIdx];
2560 FormulaSorter Sorter(L, LU, SE, DT);
2562 // Clear out the set of used regs; it will be recomputed.
2565 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2566 FIdx != NumForms; ++FIdx) {
2567 Formula &F = LU.Formulae[FIdx];
2569 SmallVector<const SCEV *, 2> Key;
2570 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2571 JE = F.BaseRegs.end(); J != JE; ++J) {
2572 const SCEV *Reg = *J;
2573 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2577 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2578 Key.push_back(F.ScaledReg);
2579 // Unstable sort by host order ok, because this is only used for
2581 std::sort(Key.begin(), Key.end());
2583 std::pair<BestFormulaeTy::const_iterator, bool> P =
2584 BestFormulae.insert(std::make_pair(Key, FIdx));
2586 Formula &Best = LU.Formulae[P.first->second];
2587 if (Sorter.operator()(F, Best))
2589 DEBUG(dbgs() << "Filtering out "; F.print(dbgs());
2591 " in favor of "; Best.print(dbgs());
2596 std::swap(F, LU.Formulae.back());
2597 LU.Formulae.pop_back();
2602 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2603 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2605 BestFormulae.clear();
2608 DEBUG(if (Changed) {
2610 "After filtering out undesirable candidates:\n";
2615 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2616 /// formulae to choose from, use some rough heuristics to prune down the number
2617 /// of formulae. This keeps the main solver from taking an extraordinary amount
2618 /// of time in some worst-case scenarios.
2619 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2620 // This is a rough guess that seems to work fairly well.
2621 const size_t Limit = UINT16_MAX;
2623 SmallPtrSet<const SCEV *, 4> Taken;
2625 // Estimate the worst-case number of solutions we might consider. We almost
2626 // never consider this many solutions because we prune the search space,
2627 // but the pruning isn't always sufficient.
2629 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2630 E = Uses.end(); I != E; ++I) {
2631 size_t FSize = I->Formulae.size();
2632 if (FSize >= Limit) {
2643 // Ok, we have too many of formulae on our hands to conveniently handle.
2644 // Use a rough heuristic to thin out the list.
2646 // Pick the register which is used by the most LSRUses, which is likely
2647 // to be a good reuse register candidate.
2648 const SCEV *Best = 0;
2649 unsigned BestNum = 0;
2650 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2652 const SCEV *Reg = *I;
2653 if (Taken.count(Reg))
2658 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2659 if (Count > BestNum) {
2666 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2667 << " will yield profitable reuse.\n");
2670 // In any use with formulae which references this register, delete formulae
2671 // which don't reference it.
2672 for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(),
2673 E = Uses.end(); I != E; ++I) {
2675 if (!LU.Regs.count(Best)) continue;
2677 // Clear out the set of used regs; it will be recomputed.
2680 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2681 Formula &F = LU.Formulae[i];
2682 if (!F.referencesReg(Best)) {
2683 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2684 std::swap(LU.Formulae.back(), F);
2685 LU.Formulae.pop_back();
2691 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2692 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2696 DEBUG(dbgs() << "After pre-selection:\n";
2697 print_uses(dbgs()));
2701 /// SolveRecurse - This is the recursive solver.
2702 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2704 SmallVectorImpl<const Formula *> &Workspace,
2705 const Cost &CurCost,
2706 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2707 DenseSet<const SCEV *> &VisitedRegs) const {
2710 // - use more aggressive filtering
2711 // - sort the formula so that the most profitable solutions are found first
2712 // - sort the uses too
2714 // - don't compute a cost, and then compare. compare while computing a cost
2716 // - track register sets with SmallBitVector
2718 const LSRUse &LU = Uses[Workspace.size()];
2720 // If this use references any register that's already a part of the
2721 // in-progress solution, consider it a requirement that a formula must
2722 // reference that register in order to be considered. This prunes out
2723 // unprofitable searching.
2724 SmallSetVector<const SCEV *, 4> ReqRegs;
2725 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
2726 E = CurRegs.end(); I != E; ++I)
2727 if (LU.Regs.count(*I))
2730 bool AnySatisfiedReqRegs = false;
2731 SmallPtrSet<const SCEV *, 16> NewRegs;
2734 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2735 E = LU.Formulae.end(); I != E; ++I) {
2736 const Formula &F = *I;
2738 // Ignore formulae which do not use any of the required registers.
2739 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
2740 JE = ReqRegs.end(); J != JE; ++J) {
2741 const SCEV *Reg = *J;
2742 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
2743 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
2747 AnySatisfiedReqRegs = true;
2749 // Evaluate the cost of the current formula. If it's already worse than
2750 // the current best, prune the search at that point.
2753 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
2754 if (NewCost < SolutionCost) {
2755 Workspace.push_back(&F);
2756 if (Workspace.size() != Uses.size()) {
2757 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
2758 NewRegs, VisitedRegs);
2759 if (F.getNumRegs() == 1 && Workspace.size() == 1)
2760 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
2762 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
2763 dbgs() << ". Regs:";
2764 for (SmallPtrSet<const SCEV *, 16>::const_iterator
2765 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
2766 dbgs() << ' ' << **I;
2769 SolutionCost = NewCost;
2770 Solution = Workspace;
2772 Workspace.pop_back();
2777 // If none of the formulae had all of the required registers, relax the
2778 // constraint so that we don't exclude all formulae.
2779 if (!AnySatisfiedReqRegs) {
2785 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
2786 SmallVector<const Formula *, 8> Workspace;
2788 SolutionCost.Loose();
2790 SmallPtrSet<const SCEV *, 16> CurRegs;
2791 DenseSet<const SCEV *> VisitedRegs;
2792 Workspace.reserve(Uses.size());
2794 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
2795 CurRegs, VisitedRegs);
2797 // Ok, we've now made all our decisions.
2798 DEBUG(dbgs() << "\n"
2799 "The chosen solution requires "; SolutionCost.print(dbgs());
2801 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
2803 Uses[i].print(dbgs());
2806 Solution[i]->print(dbgs());
2811 /// getImmediateDominator - A handy utility for the specific DominatorTree
2812 /// query that we need here.
2814 static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
2815 DomTreeNode *Node = DT.getNode(BB);
2816 if (!Node) return 0;
2817 Node = Node->getIDom();
2818 if (!Node) return 0;
2819 return Node->getBlock();
2822 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
2823 /// the dominator tree far as we can go while still being dominated by the
2824 /// input positions. This helps canonicalize the insert position, which
2825 /// encourages sharing.
2826 BasicBlock::iterator
2827 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
2828 const SmallVectorImpl<Instruction *> &Inputs)
2831 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
2832 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
2835 for (BasicBlock *Rung = IP->getParent(); ; Rung = IDom) {
2836 IDom = getImmediateDominator(Rung, DT);
2837 if (!IDom) return IP;
2839 // Don't climb into a loop though.
2840 const Loop *IDomLoop = LI.getLoopFor(IDom);
2841 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
2842 if (IDomDepth <= IPLoopDepth &&
2843 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
2847 bool AllDominate = true;
2848 Instruction *BetterPos = 0;
2849 Instruction *Tentative = IDom->getTerminator();
2850 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
2851 E = Inputs.end(); I != E; ++I) {
2852 Instruction *Inst = *I;
2853 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
2854 AllDominate = false;
2857 // Attempt to find an insert position in the middle of the block,
2858 // instead of at the end, so that it can be used for other expansions.
2859 if (IDom == Inst->getParent() &&
2860 (!BetterPos || DT.dominates(BetterPos, Inst)))
2861 BetterPos = next(BasicBlock::iterator(Inst));
2874 /// AdjustInsertPositionForExpand - Determine an input position which will be
2875 /// dominated by the operands and which will dominate the result.
2876 BasicBlock::iterator
2877 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2879 const LSRUse &LU) const {
2880 // Collect some instructions which must be dominated by the
2881 // expanding replacement. These must be dominated by any operands that
2882 // will be required in the expansion.
2883 SmallVector<Instruction *, 4> Inputs;
2884 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
2885 Inputs.push_back(I);
2886 if (LU.Kind == LSRUse::ICmpZero)
2887 if (Instruction *I =
2888 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
2889 Inputs.push_back(I);
2890 if (LF.PostIncLoops.count(L)) {
2891 if (LF.isUseFullyOutsideLoop(L))
2892 Inputs.push_back(L->getLoopLatch()->getTerminator());
2894 Inputs.push_back(IVIncInsertPos);
2896 // The expansion must also be dominated by the increment positions of any
2897 // loops it for which it is using post-inc mode.
2898 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
2899 E = LF.PostIncLoops.end(); I != E; ++I) {
2900 const Loop *PIL = *I;
2901 if (PIL == L) continue;
2903 // Be dominated by the loop exit.
2904 SmallVector<BasicBlock *, 4> ExitingBlocks;
2905 PIL->getExitingBlocks(ExitingBlocks);
2906 if (!ExitingBlocks.empty()) {
2907 BasicBlock *BB = ExitingBlocks[0];
2908 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
2909 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
2910 Inputs.push_back(BB->getTerminator());
2914 // Then, climb up the immediate dominator tree as far as we can go while
2915 // still being dominated by the input positions.
2916 IP = HoistInsertPosition(IP, Inputs);
2918 // Don't insert instructions before PHI nodes.
2919 while (isa<PHINode>(IP)) ++IP;
2921 // Ignore debug intrinsics.
2922 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
2927 Value *LSRInstance::Expand(const LSRFixup &LF,
2929 BasicBlock::iterator IP,
2930 SCEVExpander &Rewriter,
2931 SmallVectorImpl<WeakVH> &DeadInsts) const {
2932 const LSRUse &LU = Uses[LF.LUIdx];
2934 // Determine an input position which will be dominated by the operands and
2935 // which will dominate the result.
2936 IP = AdjustInsertPositionForExpand(IP, LF, LU);
2938 // Inform the Rewriter if we have a post-increment use, so that it can
2939 // perform an advantageous expansion.
2940 Rewriter.setPostInc(LF.PostIncLoops);
2942 // This is the type that the user actually needs.
2943 const Type *OpTy = LF.OperandValToReplace->getType();
2944 // This will be the type that we'll initially expand to.
2945 const Type *Ty = F.getType();
2947 // No type known; just expand directly to the ultimate type.
2949 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
2950 // Expand directly to the ultimate type if it's the right size.
2952 // This is the type to do integer arithmetic in.
2953 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
2955 // Build up a list of operands to add together to form the full base.
2956 SmallVector<const SCEV *, 8> Ops;
2958 // Expand the BaseRegs portion.
2959 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2960 E = F.BaseRegs.end(); I != E; ++I) {
2961 const SCEV *Reg = *I;
2962 assert(!Reg->isZero() && "Zero allocated in a base register!");
2964 // If we're expanding for a post-inc user, make the post-inc adjustment.
2965 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
2966 Reg = TransformForPostIncUse(Denormalize, Reg,
2967 LF.UserInst, LF.OperandValToReplace,
2970 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
2973 // Flush the operand list to suppress SCEVExpander hoisting.
2975 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
2977 Ops.push_back(SE.getUnknown(FullV));
2980 // Expand the ScaledReg portion.
2981 Value *ICmpScaledV = 0;
2982 if (F.AM.Scale != 0) {
2983 const SCEV *ScaledS = F.ScaledReg;
2985 // If we're expanding for a post-inc user, make the post-inc adjustment.
2986 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
2987 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
2988 LF.UserInst, LF.OperandValToReplace,
2991 if (LU.Kind == LSRUse::ICmpZero) {
2992 // An interesting way of "folding" with an icmp is to use a negated
2993 // scale, which we'll implement by inserting it into the other operand
2995 assert(F.AM.Scale == -1 &&
2996 "The only scale supported by ICmpZero uses is -1!");
2997 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
2999 // Otherwise just expand the scaled register and an explicit scale,
3000 // which is expected to be matched as part of the address.
3001 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3002 ScaledS = SE.getMulExpr(ScaledS,
3003 SE.getIntegerSCEV(F.AM.Scale,
3004 ScaledS->getType()));
3005 Ops.push_back(ScaledS);
3007 // 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));
3014 // Expand the GV portion.
3016 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3018 // Flush the operand list to suppress SCEVExpander hoisting.
3019 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3021 Ops.push_back(SE.getUnknown(FullV));
3024 // Expand the immediate portion.
3025 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3027 if (LU.Kind == LSRUse::ICmpZero) {
3028 // The other interesting way of "folding" with an ICmpZero is to use a
3029 // negated immediate.
3031 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3033 Ops.push_back(SE.getUnknown(ICmpScaledV));
3034 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3037 // Just add the immediate values. These again are expected to be matched
3038 // as part of the address.
3039 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3043 // Emit instructions summing all the operands.
3044 const SCEV *FullS = Ops.empty() ?
3045 SE.getIntegerSCEV(0, IntTy) :
3047 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3049 // We're done expanding now, so reset the rewriter.
3050 Rewriter.clearPostInc();
3052 // An ICmpZero Formula represents an ICmp which we're handling as a
3053 // comparison against zero. Now that we've expanded an expression for that
3054 // form, update the ICmp's other operand.
3055 if (LU.Kind == LSRUse::ICmpZero) {
3056 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3057 DeadInsts.push_back(CI->getOperand(1));
3058 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3059 "a scale at the same time!");
3060 if (F.AM.Scale == -1) {
3061 if (ICmpScaledV->getType() != OpTy) {
3063 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3065 ICmpScaledV, OpTy, "tmp", CI);
3068 CI->setOperand(1, ICmpScaledV);
3070 assert(F.AM.Scale == 0 &&
3071 "ICmp does not support folding a global value and "
3072 "a scale at the same time!");
3073 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3075 if (C->getType() != OpTy)
3076 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3080 CI->setOperand(1, C);
3087 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3088 /// of their operands effectively happens in their predecessor blocks, so the
3089 /// expression may need to be expanded in multiple places.
3090 void LSRInstance::RewriteForPHI(PHINode *PN,
3093 SCEVExpander &Rewriter,
3094 SmallVectorImpl<WeakVH> &DeadInsts,
3096 DenseMap<BasicBlock *, Value *> Inserted;
3097 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3098 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3099 BasicBlock *BB = PN->getIncomingBlock(i);
3101 // If this is a critical edge, split the edge so that we do not insert
3102 // the code on all predecessor/successor paths. We do this unless this
3103 // is the canonical backedge for this loop, which complicates post-inc
3105 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3106 !isa<IndirectBrInst>(BB->getTerminator()) &&
3107 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3108 // Split the critical edge.
3109 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3111 // If PN is outside of the loop and BB is in the loop, we want to
3112 // move the block to be immediately before the PHI block, not
3113 // immediately after BB.
3114 if (L->contains(BB) && !L->contains(PN))
3115 NewBB->moveBefore(PN->getParent());
3117 // Splitting the edge can reduce the number of PHI entries we have.
3118 e = PN->getNumIncomingValues();
3120 i = PN->getBasicBlockIndex(BB);
3123 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3124 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3126 PN->setIncomingValue(i, Pair.first->second);
3128 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3130 // If this is reuse-by-noop-cast, insert the noop cast.
3131 const Type *OpTy = LF.OperandValToReplace->getType();
3132 if (FullV->getType() != OpTy)
3134 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3136 FullV, LF.OperandValToReplace->getType(),
3137 "tmp", BB->getTerminator());
3139 PN->setIncomingValue(i, FullV);
3140 Pair.first->second = FullV;
3145 /// Rewrite - Emit instructions for the leading candidate expression for this
3146 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3147 /// the newly expanded value.
3148 void LSRInstance::Rewrite(const LSRFixup &LF,
3150 SCEVExpander &Rewriter,
3151 SmallVectorImpl<WeakVH> &DeadInsts,
3153 // First, find an insertion point that dominates UserInst. For PHI nodes,
3154 // find the nearest block which dominates all the relevant uses.
3155 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3156 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3158 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3160 // If this is reuse-by-noop-cast, insert the noop cast.
3161 const Type *OpTy = LF.OperandValToReplace->getType();
3162 if (FullV->getType() != OpTy) {
3164 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3165 FullV, OpTy, "tmp", LF.UserInst);
3169 // Update the user. ICmpZero is handled specially here (for now) because
3170 // Expand may have updated one of the operands of the icmp already, and
3171 // its new value may happen to be equal to LF.OperandValToReplace, in
3172 // which case doing replaceUsesOfWith leads to replacing both operands
3173 // with the same value. TODO: Reorganize this.
3174 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3175 LF.UserInst->setOperand(0, FullV);
3177 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3180 DeadInsts.push_back(LF.OperandValToReplace);
3184 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3186 // Keep track of instructions we may have made dead, so that
3187 // we can remove them after we are done working.
3188 SmallVector<WeakVH, 16> DeadInsts;
3190 SCEVExpander Rewriter(SE);
3191 Rewriter.disableCanonicalMode();
3192 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3194 // Expand the new value definitions and update the users.
3195 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3196 size_t LUIdx = Fixups[i].LUIdx;
3198 Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
3203 // Clean up after ourselves. This must be done before deleting any
3207 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3210 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3211 : IU(P->getAnalysis<IVUsers>()),
3212 SE(P->getAnalysis<ScalarEvolution>()),
3213 DT(P->getAnalysis<DominatorTree>()),
3214 LI(P->getAnalysis<LoopInfo>()),
3215 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3217 // If LoopSimplify form is not available, stay out of trouble.
3218 if (!L->isLoopSimplifyForm()) return;
3220 // If there's no interesting work to be done, bail early.
3221 if (IU.empty()) return;
3223 DEBUG(dbgs() << "\nLSR on loop ";
3224 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3227 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3228 /// inside the loop then try to eliminate the cast operation.
3231 // Change loop terminating condition to use the postinc iv when possible.
3232 Changed |= OptimizeLoopTermCond();
3234 CollectInterestingTypesAndFactors();
3235 CollectFixupsAndInitialFormulae();
3236 CollectLoopInvariantFixupsAndFormulae();
3238 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3239 print_uses(dbgs()));
3241 // Now use the reuse data to generate a bunch of interesting ways
3242 // to formulate the values needed for the uses.
3243 GenerateAllReuseFormulae();
3245 DEBUG(dbgs() << "\n"
3246 "After generating reuse formulae:\n";
3247 print_uses(dbgs()));
3249 FilterOutUndesirableDedicatedRegisters();
3250 NarrowSearchSpaceUsingHeuristics();
3252 SmallVector<const Formula *, 8> Solution;
3254 assert(Solution.size() == Uses.size() && "Malformed solution!");
3256 // Release memory that is no longer needed.
3262 // Formulae should be legal.
3263 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3264 E = Uses.end(); I != E; ++I) {
3265 const LSRUse &LU = *I;
3266 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3267 JE = LU.Formulae.end(); J != JE; ++J)
3268 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3269 LU.Kind, LU.AccessTy, TLI) &&
3270 "Illegal formula generated!");
3274 // Now that we've decided what we want, make it so.
3275 ImplementSolution(Solution, P);
3278 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3279 if (Factors.empty() && Types.empty()) return;
3281 OS << "LSR has identified the following interesting factors and types: ";
3284 for (SmallSetVector<int64_t, 8>::const_iterator
3285 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3286 if (!First) OS << ", ";
3291 for (SmallSetVector<const Type *, 4>::const_iterator
3292 I = Types.begin(), E = Types.end(); I != E; ++I) {
3293 if (!First) OS << ", ";
3295 OS << '(' << **I << ')';
3300 void LSRInstance::print_fixups(raw_ostream &OS) const {
3301 OS << "LSR is examining the following fixup sites:\n";
3302 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3303 E = Fixups.end(); I != E; ++I) {
3304 const LSRFixup &LF = *I;
3311 void LSRInstance::print_uses(raw_ostream &OS) const {
3312 OS << "LSR is examining the following uses:\n";
3313 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3314 E = Uses.end(); I != E; ++I) {
3315 const LSRUse &LU = *I;
3319 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3320 JE = LU.Formulae.end(); J != JE; ++J) {
3328 void LSRInstance::print(raw_ostream &OS) const {
3329 print_factors_and_types(OS);
3334 void LSRInstance::dump() const {
3335 print(errs()); errs() << '\n';
3340 class LoopStrengthReduce : public LoopPass {
3341 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3342 /// transformation profitability.
3343 const TargetLowering *const TLI;
3346 static char ID; // Pass ID, replacement for typeid
3347 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3350 bool runOnLoop(Loop *L, LPPassManager &LPM);
3351 void getAnalysisUsage(AnalysisUsage &AU) const;
3356 char LoopStrengthReduce::ID = 0;
3357 static RegisterPass<LoopStrengthReduce>
3358 X("loop-reduce", "Loop Strength Reduction");
3360 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3361 return new LoopStrengthReduce(TLI);
3364 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3365 : LoopPass(&ID), TLI(tli) {}
3367 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3368 // We split critical edges, so we change the CFG. However, we do update
3369 // many analyses if they are around.
3370 AU.addPreservedID(LoopSimplifyID);
3371 AU.addPreserved("domfrontier");
3373 AU.addRequired<LoopInfo>();
3374 AU.addPreserved<LoopInfo>();
3375 AU.addRequiredID(LoopSimplifyID);
3376 AU.addRequired<DominatorTree>();
3377 AU.addPreserved<DominatorTree>();
3378 AU.addRequired<ScalarEvolution>();
3379 AU.addPreserved<ScalarEvolution>();
3380 AU.addRequired<IVUsers>();
3381 AU.addPreserved<IVUsers>();
3384 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3385 bool Changed = false;
3387 // Run the main LSR transformation.
3388 Changed |= LSRInstance(TLI, L, this).getChanged();
3390 // At this point, it is worth checking to see if any recurrence PHIs are also
3391 // dead, so that we can remove them as well.
3392 Changed |= DeleteDeadPHIs(L->getHeader());