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/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/ValueHandle.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Target/TargetLowering.h"
81 /// RegSortData - This class holds data which is used to order reuse candidates.
84 /// UsedByIndices - This represents the set of LSRUse indices which reference
85 /// a particular register.
86 SmallBitVector UsedByIndices;
90 void print(raw_ostream &OS) const;
96 void RegSortData::print(raw_ostream &OS) const {
97 OS << "[NumUses=" << UsedByIndices.count() << ']';
100 void RegSortData::dump() const {
101 print(errs()); errs() << '\n';
106 /// RegUseTracker - Map register candidates to information about how they are
108 class RegUseTracker {
109 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
111 RegUsesTy RegUsesMap;
112 SmallVector<const SCEV *, 16> RegSequence;
115 void CountRegister(const SCEV *Reg, size_t LUIdx);
116 void DropRegister(const SCEV *Reg, size_t LUIdx);
117 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
119 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
121 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
125 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
126 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
127 iterator begin() { return RegSequence.begin(); }
128 iterator end() { return RegSequence.end(); }
129 const_iterator begin() const { return RegSequence.begin(); }
130 const_iterator end() const { return RegSequence.end(); }
136 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
137 std::pair<RegUsesTy::iterator, bool> Pair =
138 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
139 RegSortData &RSD = Pair.first->second;
141 RegSequence.push_back(Reg);
142 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
143 RSD.UsedByIndices.set(LUIdx);
147 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
148 RegUsesTy::iterator It = RegUsesMap.find(Reg);
149 assert(It != RegUsesMap.end());
150 RegSortData &RSD = It->second;
151 assert(RSD.UsedByIndices.size() > LUIdx);
152 RSD.UsedByIndices.reset(LUIdx);
156 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
157 assert(LUIdx <= LastLUIdx);
159 // Update RegUses. The data structure is not optimized for this purpose;
160 // we must iterate through it and update each of the bit vectors.
161 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
163 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
164 if (LUIdx < UsedByIndices.size())
165 UsedByIndices[LUIdx] =
166 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
167 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
172 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
173 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
174 if (I == RegUsesMap.end())
176 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
177 int i = UsedByIndices.find_first();
178 if (i == -1) return false;
179 if ((size_t)i != LUIdx) return true;
180 return UsedByIndices.find_next(i) != -1;
183 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
184 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
185 assert(I != RegUsesMap.end() && "Unknown register!");
186 return I->second.UsedByIndices;
189 void RegUseTracker::clear() {
196 /// Formula - This class holds information that describes a formula for
197 /// computing satisfying a use. It may include broken-out immediates and scaled
200 /// AM - This is used to represent complex addressing, as well as other kinds
201 /// of interesting uses.
202 TargetLowering::AddrMode AM;
204 /// BaseRegs - The list of "base" registers for this use. When this is
205 /// non-empty, AM.HasBaseReg should be set to true.
206 SmallVector<const SCEV *, 2> BaseRegs;
208 /// ScaledReg - The 'scaled' register for this use. This should be non-null
209 /// when AM.Scale is not zero.
210 const SCEV *ScaledReg;
212 Formula() : ScaledReg(0) {}
214 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
216 unsigned getNumRegs() const;
217 const Type *getType() const;
219 void DeleteBaseReg(const SCEV *&S);
221 bool referencesReg(const SCEV *S) const;
222 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
223 const RegUseTracker &RegUses) const;
225 void print(raw_ostream &OS) const;
231 /// DoInitialMatch - Recursion helper for InitialMatch.
232 static void DoInitialMatch(const SCEV *S, Loop *L,
233 SmallVectorImpl<const SCEV *> &Good,
234 SmallVectorImpl<const SCEV *> &Bad,
235 ScalarEvolution &SE) {
236 // Collect expressions which properly dominate the loop header.
237 if (SE.properlyDominates(S, L->getHeader())) {
242 // Look at add operands.
243 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
244 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
246 DoInitialMatch(*I, L, Good, Bad, SE);
250 // Look at addrec operands.
251 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
252 if (!AR->getStart()->isZero()) {
253 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
254 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
255 AR->getStepRecurrence(SE),
261 // Handle a multiplication by -1 (negation) if it didn't fold.
262 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
263 if (Mul->getOperand(0)->isAllOnesValue()) {
264 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
265 const SCEV *NewMul = SE.getMulExpr(Ops);
267 SmallVector<const SCEV *, 4> MyGood;
268 SmallVector<const SCEV *, 4> MyBad;
269 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
270 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
271 SE.getEffectiveSCEVType(NewMul->getType())));
272 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
273 E = MyGood.end(); I != E; ++I)
274 Good.push_back(SE.getMulExpr(NegOne, *I));
275 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
276 E = MyBad.end(); I != E; ++I)
277 Bad.push_back(SE.getMulExpr(NegOne, *I));
281 // Ok, we can't do anything interesting. Just stuff the whole thing into a
282 // register and hope for the best.
286 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
287 /// attempting to keep all loop-invariant and loop-computable values in a
288 /// single base register.
289 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
290 SmallVector<const SCEV *, 4> Good;
291 SmallVector<const SCEV *, 4> Bad;
292 DoInitialMatch(S, L, Good, Bad, SE);
294 const SCEV *Sum = SE.getAddExpr(Good);
296 BaseRegs.push_back(Sum);
297 AM.HasBaseReg = true;
300 const SCEV *Sum = SE.getAddExpr(Bad);
302 BaseRegs.push_back(Sum);
303 AM.HasBaseReg = true;
307 /// getNumRegs - Return the total number of register operands used by this
308 /// formula. This does not include register uses implied by non-constant
310 unsigned Formula::getNumRegs() const {
311 return !!ScaledReg + BaseRegs.size();
314 /// getType - Return the type of this formula, if it has one, or null
315 /// otherwise. This type is meaningless except for the bit size.
316 const Type *Formula::getType() const {
317 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
318 ScaledReg ? ScaledReg->getType() :
319 AM.BaseGV ? AM.BaseGV->getType() :
323 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
324 void Formula::DeleteBaseReg(const SCEV *&S) {
325 if (&S != &BaseRegs.back())
326 std::swap(S, BaseRegs.back());
330 /// referencesReg - Test if this formula references the given register.
331 bool Formula::referencesReg(const SCEV *S) const {
332 return S == ScaledReg ||
333 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
336 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
337 /// which are used by uses other than the use with the given index.
338 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
339 const RegUseTracker &RegUses) const {
341 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
343 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
344 E = BaseRegs.end(); I != E; ++I)
345 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
350 void Formula::print(raw_ostream &OS) const {
353 if (!First) OS << " + "; else First = false;
354 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
356 if (AM.BaseOffs != 0) {
357 if (!First) OS << " + "; else First = false;
360 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
361 E = BaseRegs.end(); I != E; ++I) {
362 if (!First) OS << " + "; else First = false;
363 OS << "reg(" << **I << ')';
365 if (AM.HasBaseReg && BaseRegs.empty()) {
366 if (!First) OS << " + "; else First = false;
367 OS << "**error: HasBaseReg**";
368 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
369 if (!First) OS << " + "; else First = false;
370 OS << "**error: !HasBaseReg**";
373 if (!First) OS << " + "; else First = false;
374 OS << AM.Scale << "*reg(";
383 void Formula::dump() const {
384 print(errs()); errs() << '\n';
387 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
388 /// without changing its value.
389 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
391 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
392 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
395 /// isAddSExtable - Return true if the given add can be sign-extended
396 /// without changing its value.
397 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
399 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
400 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
403 /// isMulSExtable - Return true if the given mul can be sign-extended
404 /// without changing its value.
405 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
407 IntegerType::get(SE.getContext(),
408 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
409 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
412 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
413 /// and if the remainder is known to be zero, or null otherwise. If
414 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
415 /// to Y, ignoring that the multiplication may overflow, which is useful when
416 /// the result will be used in a context where the most significant bits are
418 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
420 bool IgnoreSignificantBits = false) {
421 // Handle the trivial case, which works for any SCEV type.
423 return SE.getConstant(LHS->getType(), 1);
425 // Handle a few RHS special cases.
426 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
428 const APInt &RA = RC->getValue()->getValue();
429 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
431 if (RA.isAllOnesValue())
432 return SE.getMulExpr(LHS, RC);
433 // Handle x /s 1 as x.
438 // Check for a division of a constant by a constant.
439 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
442 const APInt &LA = C->getValue()->getValue();
443 const APInt &RA = RC->getValue()->getValue();
444 if (LA.srem(RA) != 0)
446 return SE.getConstant(LA.sdiv(RA));
449 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
450 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
451 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
452 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
453 IgnoreSignificantBits);
455 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
456 IgnoreSignificantBits);
457 if (!Start) return 0;
458 return SE.getAddRecExpr(Start, Step, AR->getLoop());
463 // Distribute the sdiv over add operands, if the add doesn't overflow.
464 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
465 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
466 SmallVector<const SCEV *, 8> Ops;
467 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
469 const SCEV *Op = getExactSDiv(*I, RHS, SE,
470 IgnoreSignificantBits);
474 return SE.getAddExpr(Ops);
479 // Check for a multiply operand that we can pull RHS out of.
480 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
481 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
482 SmallVector<const SCEV *, 4> Ops;
484 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
488 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
489 IgnoreSignificantBits)) {
495 return Found ? SE.getMulExpr(Ops) : 0;
500 // Otherwise we don't know.
504 /// ExtractImmediate - If S involves the addition of a constant integer value,
505 /// return that integer value, and mutate S to point to a new SCEV with that
507 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
508 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
509 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
510 S = SE.getConstant(C->getType(), 0);
511 return C->getValue()->getSExtValue();
513 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
514 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
515 int64_t Result = ExtractImmediate(NewOps.front(), SE);
517 S = SE.getAddExpr(NewOps);
519 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
520 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
521 int64_t Result = ExtractImmediate(NewOps.front(), SE);
523 S = SE.getAddRecExpr(NewOps, AR->getLoop());
529 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
530 /// return that symbol, and mutate S to point to a new SCEV with that
532 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
533 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
534 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
535 S = SE.getConstant(GV->getType(), 0);
538 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
539 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
540 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
542 S = SE.getAddExpr(NewOps);
544 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
545 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
546 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
548 S = SE.getAddRecExpr(NewOps, AR->getLoop());
554 /// isAddressUse - Returns true if the specified instruction is using the
555 /// specified value as an address.
556 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
557 bool isAddress = isa<LoadInst>(Inst);
558 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
559 if (SI->getOperand(1) == OperandVal)
561 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
562 // Addressing modes can also be folded into prefetches and a variety
564 switch (II->getIntrinsicID()) {
566 case Intrinsic::prefetch:
567 case Intrinsic::x86_sse2_loadu_dq:
568 case Intrinsic::x86_sse2_loadu_pd:
569 case Intrinsic::x86_sse_loadu_ps:
570 case Intrinsic::x86_sse_storeu_ps:
571 case Intrinsic::x86_sse2_storeu_pd:
572 case Intrinsic::x86_sse2_storeu_dq:
573 case Intrinsic::x86_sse2_storel_dq:
574 if (II->getArgOperand(0) == OperandVal)
582 /// getAccessType - Return the type of the memory being accessed.
583 static const Type *getAccessType(const Instruction *Inst) {
584 const Type *AccessTy = Inst->getType();
585 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
586 AccessTy = SI->getOperand(0)->getType();
587 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
588 // Addressing modes can also be folded into prefetches and a variety
590 switch (II->getIntrinsicID()) {
592 case Intrinsic::x86_sse_storeu_ps:
593 case Intrinsic::x86_sse2_storeu_pd:
594 case Intrinsic::x86_sse2_storeu_dq:
595 case Intrinsic::x86_sse2_storel_dq:
596 AccessTy = II->getArgOperand(0)->getType();
601 // All pointers have the same requirements, so canonicalize them to an
602 // arbitrary pointer type to minimize variation.
603 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
604 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
605 PTy->getAddressSpace());
610 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
611 /// specified set are trivially dead, delete them and see if this makes any of
612 /// their operands subsequently dead.
614 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
615 bool Changed = false;
617 while (!DeadInsts.empty()) {
618 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
620 if (I == 0 || !isInstructionTriviallyDead(I))
623 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
624 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
627 DeadInsts.push_back(U);
630 I->eraseFromParent();
639 /// Cost - This class is used to measure and compare candidate formulae.
641 /// TODO: Some of these could be merged. Also, a lexical ordering
642 /// isn't always optimal.
646 unsigned NumBaseAdds;
652 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
655 bool operator<(const Cost &Other) const;
659 void RateFormula(const Formula &F,
660 SmallPtrSet<const SCEV *, 16> &Regs,
661 const DenseSet<const SCEV *> &VisitedRegs,
663 const SmallVectorImpl<int64_t> &Offsets,
664 ScalarEvolution &SE, DominatorTree &DT);
666 void print(raw_ostream &OS) const;
670 void RateRegister(const SCEV *Reg,
671 SmallPtrSet<const SCEV *, 16> &Regs,
673 ScalarEvolution &SE, DominatorTree &DT);
674 void RatePrimaryRegister(const SCEV *Reg,
675 SmallPtrSet<const SCEV *, 16> &Regs,
677 ScalarEvolution &SE, DominatorTree &DT);
682 /// RateRegister - Tally up interesting quantities from the given register.
683 void Cost::RateRegister(const SCEV *Reg,
684 SmallPtrSet<const SCEV *, 16> &Regs,
686 ScalarEvolution &SE, DominatorTree &DT) {
687 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
688 if (AR->getLoop() == L)
689 AddRecCost += 1; /// TODO: This should be a function of the stride.
691 // If this is an addrec for a loop that's already been visited by LSR,
692 // don't second-guess its addrec phi nodes. LSR isn't currently smart
693 // enough to reason about more than one loop at a time. Consider these
694 // registers free and leave them alone.
695 else if (L->contains(AR->getLoop()) ||
696 (!AR->getLoop()->contains(L) &&
697 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
698 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
699 PHINode *PN = dyn_cast<PHINode>(I); ++I)
700 if (SE.isSCEVable(PN->getType()) &&
701 (SE.getEffectiveSCEVType(PN->getType()) ==
702 SE.getEffectiveSCEVType(AR->getType())) &&
703 SE.getSCEV(PN) == AR)
706 // If this isn't one of the addrecs that the loop already has, it
707 // would require a costly new phi and add. TODO: This isn't
708 // precisely modeled right now.
710 if (!Regs.count(AR->getStart()))
711 RateRegister(AR->getStart(), Regs, L, SE, DT);
714 // Add the step value register, if it needs one.
715 // TODO: The non-affine case isn't precisely modeled here.
716 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
717 if (!Regs.count(AR->getStart()))
718 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
722 // Rough heuristic; favor registers which don't require extra setup
723 // instructions in the preheader.
724 if (!isa<SCEVUnknown>(Reg) &&
725 !isa<SCEVConstant>(Reg) &&
726 !(isa<SCEVAddRecExpr>(Reg) &&
727 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
728 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
731 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
732 SE.hasComputableLoopEvolution(Reg, L);
735 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
737 void Cost::RatePrimaryRegister(const SCEV *Reg,
738 SmallPtrSet<const SCEV *, 16> &Regs,
740 ScalarEvolution &SE, DominatorTree &DT) {
741 if (Regs.insert(Reg))
742 RateRegister(Reg, Regs, L, SE, DT);
745 void Cost::RateFormula(const Formula &F,
746 SmallPtrSet<const SCEV *, 16> &Regs,
747 const DenseSet<const SCEV *> &VisitedRegs,
749 const SmallVectorImpl<int64_t> &Offsets,
750 ScalarEvolution &SE, DominatorTree &DT) {
751 // Tally up the registers.
752 if (const SCEV *ScaledReg = F.ScaledReg) {
753 if (VisitedRegs.count(ScaledReg)) {
757 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
759 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
760 E = F.BaseRegs.end(); I != E; ++I) {
761 const SCEV *BaseReg = *I;
762 if (VisitedRegs.count(BaseReg)) {
766 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
769 if (F.BaseRegs.size() > 1)
770 NumBaseAdds += F.BaseRegs.size() - 1;
772 // Tally up the non-zero immediates.
773 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
774 E = Offsets.end(); I != E; ++I) {
775 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
777 ImmCost += 64; // Handle symbolic values conservatively.
778 // TODO: This should probably be the pointer size.
779 else if (Offset != 0)
780 ImmCost += APInt(64, Offset, true).getMinSignedBits();
784 /// Loose - Set this cost to a loosing value.
794 /// operator< - Choose the lower cost.
795 bool Cost::operator<(const Cost &Other) const {
796 if (NumRegs != Other.NumRegs)
797 return NumRegs < Other.NumRegs;
798 if (AddRecCost != Other.AddRecCost)
799 return AddRecCost < Other.AddRecCost;
800 if (NumIVMuls != Other.NumIVMuls)
801 return NumIVMuls < Other.NumIVMuls;
802 if (NumBaseAdds != Other.NumBaseAdds)
803 return NumBaseAdds < Other.NumBaseAdds;
804 if (ImmCost != Other.ImmCost)
805 return ImmCost < Other.ImmCost;
806 if (SetupCost != Other.SetupCost)
807 return SetupCost < Other.SetupCost;
811 void Cost::print(raw_ostream &OS) const {
812 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
814 OS << ", with addrec cost " << AddRecCost;
816 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
817 if (NumBaseAdds != 0)
818 OS << ", plus " << NumBaseAdds << " base add"
819 << (NumBaseAdds == 1 ? "" : "s");
821 OS << ", plus " << ImmCost << " imm cost";
823 OS << ", plus " << SetupCost << " setup cost";
826 void Cost::dump() const {
827 print(errs()); errs() << '\n';
832 /// LSRFixup - An operand value in an instruction which is to be replaced
833 /// with some equivalent, possibly strength-reduced, replacement.
835 /// UserInst - The instruction which will be updated.
836 Instruction *UserInst;
838 /// OperandValToReplace - The operand of the instruction which will
839 /// be replaced. The operand may be used more than once; every instance
840 /// will be replaced.
841 Value *OperandValToReplace;
843 /// PostIncLoops - If this user is to use the post-incremented value of an
844 /// induction variable, this variable is non-null and holds the loop
845 /// associated with the induction variable.
846 PostIncLoopSet PostIncLoops;
848 /// LUIdx - The index of the LSRUse describing the expression which
849 /// this fixup needs, minus an offset (below).
852 /// Offset - A constant offset to be added to the LSRUse expression.
853 /// This allows multiple fixups to share the same LSRUse with different
854 /// offsets, for example in an unrolled loop.
857 bool isUseFullyOutsideLoop(const Loop *L) const;
861 void print(raw_ostream &OS) const;
868 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
870 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
871 /// value outside of the given loop.
872 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
873 // PHI nodes use their value in their incoming blocks.
874 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
875 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
876 if (PN->getIncomingValue(i) == OperandValToReplace &&
877 L->contains(PN->getIncomingBlock(i)))
882 return !L->contains(UserInst);
885 void LSRFixup::print(raw_ostream &OS) const {
887 // Store is common and interesting enough to be worth special-casing.
888 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
890 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
891 } else if (UserInst->getType()->isVoidTy())
892 OS << UserInst->getOpcodeName();
894 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
896 OS << ", OperandValToReplace=";
897 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
899 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
900 E = PostIncLoops.end(); I != E; ++I) {
901 OS << ", PostIncLoop=";
902 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
905 if (LUIdx != ~size_t(0))
906 OS << ", LUIdx=" << LUIdx;
909 OS << ", Offset=" << Offset;
912 void LSRFixup::dump() const {
913 print(errs()); errs() << '\n';
918 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
919 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
920 struct UniquifierDenseMapInfo {
921 static SmallVector<const SCEV *, 2> getEmptyKey() {
922 SmallVector<const SCEV *, 2> V;
923 V.push_back(reinterpret_cast<const SCEV *>(-1));
927 static SmallVector<const SCEV *, 2> getTombstoneKey() {
928 SmallVector<const SCEV *, 2> V;
929 V.push_back(reinterpret_cast<const SCEV *>(-2));
933 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
935 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
936 E = V.end(); I != E; ++I)
937 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
941 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
942 const SmallVector<const SCEV *, 2> &RHS) {
947 /// LSRUse - This class holds the state that LSR keeps for each use in
948 /// IVUsers, as well as uses invented by LSR itself. It includes information
949 /// about what kinds of things can be folded into the user, information about
950 /// the user itself, and information about how the use may be satisfied.
951 /// TODO: Represent multiple users of the same expression in common?
953 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
956 /// KindType - An enum for a kind of use, indicating what types of
957 /// scaled and immediate operands it might support.
959 Basic, ///< A normal use, with no folding.
960 Special, ///< A special case of basic, allowing -1 scales.
961 Address, ///< An address use; folding according to TargetLowering
962 ICmpZero ///< An equality icmp with both operands folded into one.
963 // TODO: Add a generic icmp too?
967 const Type *AccessTy;
969 SmallVector<int64_t, 8> Offsets;
973 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
974 /// LSRUse are outside of the loop, in which case some special-case heuristics
976 bool AllFixupsOutsideLoop;
978 /// WidestFixupType - This records the widest use type for any fixup using
979 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
980 /// max fixup widths to be equivalent, because the narrower one may be relying
981 /// on the implicit truncation to truncate away bogus bits.
982 const Type *WidestFixupType;
984 /// Formulae - A list of ways to build a value that can satisfy this user.
985 /// After the list is populated, one of these is selected heuristically and
986 /// used to formulate a replacement for OperandValToReplace in UserInst.
987 SmallVector<Formula, 12> Formulae;
989 /// Regs - The set of register candidates used by all formulae in this LSRUse.
990 SmallPtrSet<const SCEV *, 4> Regs;
992 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
993 MinOffset(INT64_MAX),
994 MaxOffset(INT64_MIN),
995 AllFixupsOutsideLoop(true),
996 WidestFixupType(0) {}
998 bool HasFormulaWithSameRegs(const Formula &F) const;
999 bool InsertFormula(const Formula &F);
1000 void DeleteFormula(Formula &F);
1001 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1003 void print(raw_ostream &OS) const;
1009 /// HasFormula - Test whether this use as a formula which has the same
1010 /// registers as the given formula.
1011 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1012 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1013 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1014 // Unstable sort by host order ok, because this is only used for uniquifying.
1015 std::sort(Key.begin(), Key.end());
1016 return Uniquifier.count(Key);
1019 /// InsertFormula - If the given formula has not yet been inserted, add it to
1020 /// the list, and return true. Return false otherwise.
1021 bool LSRUse::InsertFormula(const Formula &F) {
1022 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1023 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1024 // Unstable sort by host order ok, because this is only used for uniquifying.
1025 std::sort(Key.begin(), Key.end());
1027 if (!Uniquifier.insert(Key).second)
1030 // Using a register to hold the value of 0 is not profitable.
1031 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1032 "Zero allocated in a scaled register!");
1034 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1035 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1036 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1039 // Add the formula to the list.
1040 Formulae.push_back(F);
1042 // Record registers now being used by this use.
1043 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1044 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1049 /// DeleteFormula - Remove the given formula from this use's list.
1050 void LSRUse::DeleteFormula(Formula &F) {
1051 if (&F != &Formulae.back())
1052 std::swap(F, Formulae.back());
1053 Formulae.pop_back();
1054 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1057 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1058 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1059 // Now that we've filtered out some formulae, recompute the Regs set.
1060 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1062 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1063 E = Formulae.end(); I != E; ++I) {
1064 const Formula &F = *I;
1065 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1066 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1069 // Update the RegTracker.
1070 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1071 E = OldRegs.end(); I != E; ++I)
1072 if (!Regs.count(*I))
1073 RegUses.DropRegister(*I, LUIdx);
1076 void LSRUse::print(raw_ostream &OS) const {
1077 OS << "LSR Use: Kind=";
1079 case Basic: OS << "Basic"; break;
1080 case Special: OS << "Special"; break;
1081 case ICmpZero: OS << "ICmpZero"; break;
1083 OS << "Address of ";
1084 if (AccessTy->isPointerTy())
1085 OS << "pointer"; // the full pointer type could be really verbose
1090 OS << ", Offsets={";
1091 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1092 E = Offsets.end(); I != E; ++I) {
1094 if (llvm::next(I) != E)
1099 if (AllFixupsOutsideLoop)
1100 OS << ", all-fixups-outside-loop";
1102 if (WidestFixupType)
1103 OS << ", widest fixup type: " << *WidestFixupType;
1106 void LSRUse::dump() const {
1107 print(errs()); errs() << '\n';
1110 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1111 /// be completely folded into the user instruction at isel time. This includes
1112 /// address-mode folding and special icmp tricks.
1113 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1114 LSRUse::KindType Kind, const Type *AccessTy,
1115 const TargetLowering *TLI) {
1117 case LSRUse::Address:
1118 // If we have low-level target information, ask the target if it can
1119 // completely fold this address.
1120 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1122 // Otherwise, just guess that reg+reg addressing is legal.
1123 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1125 case LSRUse::ICmpZero:
1126 // There's not even a target hook for querying whether it would be legal to
1127 // fold a GV into an ICmp.
1131 // ICmp only has two operands; don't allow more than two non-trivial parts.
1132 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1135 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1136 // putting the scaled register in the other operand of the icmp.
1137 if (AM.Scale != 0 && AM.Scale != -1)
1140 // If we have low-level target information, ask the target if it can fold an
1141 // integer immediate on an icmp.
1142 if (AM.BaseOffs != 0) {
1143 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1150 // Only handle single-register values.
1151 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1153 case LSRUse::Special:
1154 // Only handle -1 scales, or no scale.
1155 return AM.Scale == 0 || AM.Scale == -1;
1161 static bool isLegalUse(TargetLowering::AddrMode AM,
1162 int64_t MinOffset, int64_t MaxOffset,
1163 LSRUse::KindType Kind, const Type *AccessTy,
1164 const TargetLowering *TLI) {
1165 // Check for overflow.
1166 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1169 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1170 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1171 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1172 // Check for overflow.
1173 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1176 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1177 return isLegalUse(AM, Kind, AccessTy, TLI);
1182 static bool isAlwaysFoldable(int64_t BaseOffs,
1183 GlobalValue *BaseGV,
1185 LSRUse::KindType Kind, const Type *AccessTy,
1186 const TargetLowering *TLI) {
1187 // Fast-path: zero is always foldable.
1188 if (BaseOffs == 0 && !BaseGV) return true;
1190 // Conservatively, create an address with an immediate and a
1191 // base and a scale.
1192 TargetLowering::AddrMode AM;
1193 AM.BaseOffs = BaseOffs;
1195 AM.HasBaseReg = HasBaseReg;
1196 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1198 // Canonicalize a scale of 1 to a base register if the formula doesn't
1199 // already have a base register.
1200 if (!AM.HasBaseReg && AM.Scale == 1) {
1202 AM.HasBaseReg = true;
1205 return isLegalUse(AM, Kind, AccessTy, TLI);
1208 static bool isAlwaysFoldable(const SCEV *S,
1209 int64_t MinOffset, int64_t MaxOffset,
1211 LSRUse::KindType Kind, const Type *AccessTy,
1212 const TargetLowering *TLI,
1213 ScalarEvolution &SE) {
1214 // Fast-path: zero is always foldable.
1215 if (S->isZero()) return true;
1217 // Conservatively, create an address with an immediate and a
1218 // base and a scale.
1219 int64_t BaseOffs = ExtractImmediate(S, SE);
1220 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1222 // If there's anything else involved, it's not foldable.
1223 if (!S->isZero()) return false;
1225 // Fast-path: zero is always foldable.
1226 if (BaseOffs == 0 && !BaseGV) return true;
1228 // Conservatively, create an address with an immediate and a
1229 // base and a scale.
1230 TargetLowering::AddrMode AM;
1231 AM.BaseOffs = BaseOffs;
1233 AM.HasBaseReg = HasBaseReg;
1234 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1236 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1241 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1242 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1243 struct UseMapDenseMapInfo {
1244 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1245 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1248 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1249 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1253 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1254 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1255 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1259 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1260 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1265 /// LSRInstance - This class holds state for the main loop strength reduction
1269 ScalarEvolution &SE;
1272 const TargetLowering *const TLI;
1276 /// IVIncInsertPos - This is the insert position that the current loop's
1277 /// induction variable increment should be placed. In simple loops, this is
1278 /// the latch block's terminator. But in more complicated cases, this is a
1279 /// position which will dominate all the in-loop post-increment users.
1280 Instruction *IVIncInsertPos;
1282 /// Factors - Interesting factors between use strides.
1283 SmallSetVector<int64_t, 8> Factors;
1285 /// Types - Interesting use types, to facilitate truncation reuse.
1286 SmallSetVector<const Type *, 4> Types;
1288 /// Fixups - The list of operands which are to be replaced.
1289 SmallVector<LSRFixup, 16> Fixups;
1291 /// Uses - The list of interesting uses.
1292 SmallVector<LSRUse, 16> Uses;
1294 /// RegUses - Track which uses use which register candidates.
1295 RegUseTracker RegUses;
1297 void OptimizeShadowIV();
1298 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1299 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1300 void OptimizeLoopTermCond();
1302 void CollectInterestingTypesAndFactors();
1303 void CollectFixupsAndInitialFormulae();
1305 LSRFixup &getNewFixup() {
1306 Fixups.push_back(LSRFixup());
1307 return Fixups.back();
1310 // Support for sharing of LSRUses between LSRFixups.
1311 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1313 UseMapDenseMapInfo> UseMapTy;
1316 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1317 LSRUse::KindType Kind, const Type *AccessTy);
1319 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1320 LSRUse::KindType Kind,
1321 const Type *AccessTy);
1323 void DeleteUse(LSRUse &LU, size_t LUIdx);
1325 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1328 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1329 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1330 void CountRegisters(const Formula &F, size_t LUIdx);
1331 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1333 void CollectLoopInvariantFixupsAndFormulae();
1335 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1336 unsigned Depth = 0);
1337 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1338 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1339 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1340 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1341 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1342 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1343 void GenerateCrossUseConstantOffsets();
1344 void GenerateAllReuseFormulae();
1346 void FilterOutUndesirableDedicatedRegisters();
1348 size_t EstimateSearchSpaceComplexity() const;
1349 void NarrowSearchSpaceByDetectingSupersets();
1350 void NarrowSearchSpaceByCollapsingUnrolledCode();
1351 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1352 void NarrowSearchSpaceByPickingWinnerRegs();
1353 void NarrowSearchSpaceUsingHeuristics();
1355 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1357 SmallVectorImpl<const Formula *> &Workspace,
1358 const Cost &CurCost,
1359 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1360 DenseSet<const SCEV *> &VisitedRegs) const;
1361 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1363 BasicBlock::iterator
1364 HoistInsertPosition(BasicBlock::iterator IP,
1365 const SmallVectorImpl<Instruction *> &Inputs) const;
1366 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1368 const LSRUse &LU) const;
1370 Value *Expand(const LSRFixup &LF,
1372 BasicBlock::iterator IP,
1373 SCEVExpander &Rewriter,
1374 SmallVectorImpl<WeakVH> &DeadInsts) const;
1375 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1377 SCEVExpander &Rewriter,
1378 SmallVectorImpl<WeakVH> &DeadInsts,
1380 void Rewrite(const LSRFixup &LF,
1382 SCEVExpander &Rewriter,
1383 SmallVectorImpl<WeakVH> &DeadInsts,
1385 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1388 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1390 bool getChanged() const { return Changed; }
1392 void print_factors_and_types(raw_ostream &OS) const;
1393 void print_fixups(raw_ostream &OS) const;
1394 void print_uses(raw_ostream &OS) const;
1395 void print(raw_ostream &OS) const;
1401 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1402 /// inside the loop then try to eliminate the cast operation.
1403 void LSRInstance::OptimizeShadowIV() {
1404 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1405 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1408 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1409 UI != E; /* empty */) {
1410 IVUsers::const_iterator CandidateUI = UI;
1412 Instruction *ShadowUse = CandidateUI->getUser();
1413 const Type *DestTy = NULL;
1415 /* If shadow use is a int->float cast then insert a second IV
1416 to eliminate this cast.
1418 for (unsigned i = 0; i < n; ++i)
1424 for (unsigned i = 0; i < n; ++i, ++d)
1427 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1428 DestTy = UCast->getDestTy();
1429 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1430 DestTy = SCast->getDestTy();
1431 if (!DestTy) continue;
1434 // If target does not support DestTy natively then do not apply
1435 // this transformation.
1436 EVT DVT = TLI->getValueType(DestTy);
1437 if (!TLI->isTypeLegal(DVT)) continue;
1440 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1442 if (PH->getNumIncomingValues() != 2) continue;
1444 const Type *SrcTy = PH->getType();
1445 int Mantissa = DestTy->getFPMantissaWidth();
1446 if (Mantissa == -1) continue;
1447 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1450 unsigned Entry, Latch;
1451 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1459 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1460 if (!Init) continue;
1461 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1463 BinaryOperator *Incr =
1464 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1465 if (!Incr) continue;
1466 if (Incr->getOpcode() != Instruction::Add
1467 && Incr->getOpcode() != Instruction::Sub)
1470 /* Initialize new IV, double d = 0.0 in above example. */
1471 ConstantInt *C = NULL;
1472 if (Incr->getOperand(0) == PH)
1473 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1474 else if (Incr->getOperand(1) == PH)
1475 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1481 // Ignore negative constants, as the code below doesn't handle them
1482 // correctly. TODO: Remove this restriction.
1483 if (!C->getValue().isStrictlyPositive()) continue;
1485 /* Add new PHINode. */
1486 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1488 /* create new increment. '++d' in above example. */
1489 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1490 BinaryOperator *NewIncr =
1491 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1492 Instruction::FAdd : Instruction::FSub,
1493 NewPH, CFP, "IV.S.next.", Incr);
1495 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1496 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1498 /* Remove cast operation */
1499 ShadowUse->replaceAllUsesWith(NewPH);
1500 ShadowUse->eraseFromParent();
1506 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1507 /// set the IV user and stride information and return true, otherwise return
1509 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1510 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1511 if (UI->getUser() == Cond) {
1512 // NOTE: we could handle setcc instructions with multiple uses here, but
1513 // InstCombine does it as well for simple uses, it's not clear that it
1514 // occurs enough in real life to handle.
1521 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1522 /// a max computation.
1524 /// This is a narrow solution to a specific, but acute, problem. For loops
1530 /// } while (++i < n);
1532 /// the trip count isn't just 'n', because 'n' might not be positive. And
1533 /// unfortunately this can come up even for loops where the user didn't use
1534 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1535 /// will commonly be lowered like this:
1541 /// } while (++i < n);
1544 /// and then it's possible for subsequent optimization to obscure the if
1545 /// test in such a way that indvars can't find it.
1547 /// When indvars can't find the if test in loops like this, it creates a
1548 /// max expression, which allows it to give the loop a canonical
1549 /// induction variable:
1552 /// max = n < 1 ? 1 : n;
1555 /// } while (++i != max);
1557 /// Canonical induction variables are necessary because the loop passes
1558 /// are designed around them. The most obvious example of this is the
1559 /// LoopInfo analysis, which doesn't remember trip count values. It
1560 /// expects to be able to rediscover the trip count each time it is
1561 /// needed, and it does this using a simple analysis that only succeeds if
1562 /// the loop has a canonical induction variable.
1564 /// However, when it comes time to generate code, the maximum operation
1565 /// can be quite costly, especially if it's inside of an outer loop.
1567 /// This function solves this problem by detecting this type of loop and
1568 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1569 /// the instructions for the maximum computation.
1571 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1572 // Check that the loop matches the pattern we're looking for.
1573 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1574 Cond->getPredicate() != CmpInst::ICMP_NE)
1577 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1578 if (!Sel || !Sel->hasOneUse()) return Cond;
1580 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1581 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1583 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1585 // Add one to the backedge-taken count to get the trip count.
1586 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1587 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1589 // Check for a max calculation that matches the pattern. There's no check
1590 // for ICMP_ULE here because the comparison would be with zero, which
1591 // isn't interesting.
1592 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1593 const SCEVNAryExpr *Max = 0;
1594 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1595 Pred = ICmpInst::ICMP_SLE;
1597 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1598 Pred = ICmpInst::ICMP_SLT;
1600 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1601 Pred = ICmpInst::ICMP_ULT;
1608 // To handle a max with more than two operands, this optimization would
1609 // require additional checking and setup.
1610 if (Max->getNumOperands() != 2)
1613 const SCEV *MaxLHS = Max->getOperand(0);
1614 const SCEV *MaxRHS = Max->getOperand(1);
1616 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1617 // for a comparison with 1. For <= and >=, a comparison with zero.
1619 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1622 // Check the relevant induction variable for conformance to
1624 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1625 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1626 if (!AR || !AR->isAffine() ||
1627 AR->getStart() != One ||
1628 AR->getStepRecurrence(SE) != One)
1631 assert(AR->getLoop() == L &&
1632 "Loop condition operand is an addrec in a different loop!");
1634 // Check the right operand of the select, and remember it, as it will
1635 // be used in the new comparison instruction.
1637 if (ICmpInst::isTrueWhenEqual(Pred)) {
1638 // Look for n+1, and grab n.
1639 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1640 if (isa<ConstantInt>(BO->getOperand(1)) &&
1641 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1642 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1643 NewRHS = BO->getOperand(0);
1644 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1645 if (isa<ConstantInt>(BO->getOperand(1)) &&
1646 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1647 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1648 NewRHS = BO->getOperand(0);
1651 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1652 NewRHS = Sel->getOperand(1);
1653 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1654 NewRHS = Sel->getOperand(2);
1655 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1656 NewRHS = SU->getValue();
1658 // Max doesn't match expected pattern.
1661 // Determine the new comparison opcode. It may be signed or unsigned,
1662 // and the original comparison may be either equality or inequality.
1663 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1664 Pred = CmpInst::getInversePredicate(Pred);
1666 // Ok, everything looks ok to change the condition into an SLT or SGE and
1667 // delete the max calculation.
1669 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1671 // Delete the max calculation instructions.
1672 Cond->replaceAllUsesWith(NewCond);
1673 CondUse->setUser(NewCond);
1674 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1675 Cond->eraseFromParent();
1676 Sel->eraseFromParent();
1677 if (Cmp->use_empty())
1678 Cmp->eraseFromParent();
1682 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1683 /// postinc iv when possible.
1685 LSRInstance::OptimizeLoopTermCond() {
1686 SmallPtrSet<Instruction *, 4> PostIncs;
1688 BasicBlock *LatchBlock = L->getLoopLatch();
1689 SmallVector<BasicBlock*, 8> ExitingBlocks;
1690 L->getExitingBlocks(ExitingBlocks);
1692 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1693 BasicBlock *ExitingBlock = ExitingBlocks[i];
1695 // Get the terminating condition for the loop if possible. If we
1696 // can, we want to change it to use a post-incremented version of its
1697 // induction variable, to allow coalescing the live ranges for the IV into
1698 // one register value.
1700 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1703 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1704 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1707 // Search IVUsesByStride to find Cond's IVUse if there is one.
1708 IVStrideUse *CondUse = 0;
1709 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1710 if (!FindIVUserForCond(Cond, CondUse))
1713 // If the trip count is computed in terms of a max (due to ScalarEvolution
1714 // being unable to find a sufficient guard, for example), change the loop
1715 // comparison to use SLT or ULT instead of NE.
1716 // One consequence of doing this now is that it disrupts the count-down
1717 // optimization. That's not always a bad thing though, because in such
1718 // cases it may still be worthwhile to avoid a max.
1719 Cond = OptimizeMax(Cond, CondUse);
1721 // If this exiting block dominates the latch block, it may also use
1722 // the post-inc value if it won't be shared with other uses.
1723 // Check for dominance.
1724 if (!DT.dominates(ExitingBlock, LatchBlock))
1727 // Conservatively avoid trying to use the post-inc value in non-latch
1728 // exits if there may be pre-inc users in intervening blocks.
1729 if (LatchBlock != ExitingBlock)
1730 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1731 // Test if the use is reachable from the exiting block. This dominator
1732 // query is a conservative approximation of reachability.
1733 if (&*UI != CondUse &&
1734 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1735 // Conservatively assume there may be reuse if the quotient of their
1736 // strides could be a legal scale.
1737 const SCEV *A = IU.getStride(*CondUse, L);
1738 const SCEV *B = IU.getStride(*UI, L);
1739 if (!A || !B) continue;
1740 if (SE.getTypeSizeInBits(A->getType()) !=
1741 SE.getTypeSizeInBits(B->getType())) {
1742 if (SE.getTypeSizeInBits(A->getType()) >
1743 SE.getTypeSizeInBits(B->getType()))
1744 B = SE.getSignExtendExpr(B, A->getType());
1746 A = SE.getSignExtendExpr(A, B->getType());
1748 if (const SCEVConstant *D =
1749 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1750 const ConstantInt *C = D->getValue();
1751 // Stride of one or negative one can have reuse with non-addresses.
1752 if (C->isOne() || C->isAllOnesValue())
1753 goto decline_post_inc;
1754 // Avoid weird situations.
1755 if (C->getValue().getMinSignedBits() >= 64 ||
1756 C->getValue().isMinSignedValue())
1757 goto decline_post_inc;
1758 // Without TLI, assume that any stride might be valid, and so any
1759 // use might be shared.
1761 goto decline_post_inc;
1762 // Check for possible scaled-address reuse.
1763 const Type *AccessTy = getAccessType(UI->getUser());
1764 TargetLowering::AddrMode AM;
1765 AM.Scale = C->getSExtValue();
1766 if (TLI->isLegalAddressingMode(AM, AccessTy))
1767 goto decline_post_inc;
1768 AM.Scale = -AM.Scale;
1769 if (TLI->isLegalAddressingMode(AM, AccessTy))
1770 goto decline_post_inc;
1774 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1777 // It's possible for the setcc instruction to be anywhere in the loop, and
1778 // possible for it to have multiple users. If it is not immediately before
1779 // the exiting block branch, move it.
1780 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1781 if (Cond->hasOneUse()) {
1782 Cond->moveBefore(TermBr);
1784 // Clone the terminating condition and insert into the loopend.
1785 ICmpInst *OldCond = Cond;
1786 Cond = cast<ICmpInst>(Cond->clone());
1787 Cond->setName(L->getHeader()->getName() + ".termcond");
1788 ExitingBlock->getInstList().insert(TermBr, Cond);
1790 // Clone the IVUse, as the old use still exists!
1791 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1792 TermBr->replaceUsesOfWith(OldCond, Cond);
1796 // If we get to here, we know that we can transform the setcc instruction to
1797 // use the post-incremented version of the IV, allowing us to coalesce the
1798 // live ranges for the IV correctly.
1799 CondUse->transformToPostInc(L);
1802 PostIncs.insert(Cond);
1806 // Determine an insertion point for the loop induction variable increment. It
1807 // must dominate all the post-inc comparisons we just set up, and it must
1808 // dominate the loop latch edge.
1809 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1810 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1811 E = PostIncs.end(); I != E; ++I) {
1813 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1815 if (BB == (*I)->getParent())
1816 IVIncInsertPos = *I;
1817 else if (BB != IVIncInsertPos->getParent())
1818 IVIncInsertPos = BB->getTerminator();
1822 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1823 /// at the given offset and other details. If so, update the use and
1826 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1827 LSRUse::KindType Kind, const Type *AccessTy) {
1828 int64_t NewMinOffset = LU.MinOffset;
1829 int64_t NewMaxOffset = LU.MaxOffset;
1830 const Type *NewAccessTy = AccessTy;
1832 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1833 // something conservative, however this can pessimize in the case that one of
1834 // the uses will have all its uses outside the loop, for example.
1835 if (LU.Kind != Kind)
1837 // Conservatively assume HasBaseReg is true for now.
1838 if (NewOffset < LU.MinOffset) {
1839 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1840 Kind, AccessTy, TLI))
1842 NewMinOffset = NewOffset;
1843 } else if (NewOffset > LU.MaxOffset) {
1844 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1845 Kind, AccessTy, TLI))
1847 NewMaxOffset = NewOffset;
1849 // Check for a mismatched access type, and fall back conservatively as needed.
1850 // TODO: Be less conservative when the type is similar and can use the same
1851 // addressing modes.
1852 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1853 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1856 LU.MinOffset = NewMinOffset;
1857 LU.MaxOffset = NewMaxOffset;
1858 LU.AccessTy = NewAccessTy;
1859 if (NewOffset != LU.Offsets.back())
1860 LU.Offsets.push_back(NewOffset);
1864 /// getUse - Return an LSRUse index and an offset value for a fixup which
1865 /// needs the given expression, with the given kind and optional access type.
1866 /// Either reuse an existing use or create a new one, as needed.
1867 std::pair<size_t, int64_t>
1868 LSRInstance::getUse(const SCEV *&Expr,
1869 LSRUse::KindType Kind, const Type *AccessTy) {
1870 const SCEV *Copy = Expr;
1871 int64_t Offset = ExtractImmediate(Expr, SE);
1873 // Basic uses can't accept any offset, for example.
1874 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1879 std::pair<UseMapTy::iterator, bool> P =
1880 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1882 // A use already existed with this base.
1883 size_t LUIdx = P.first->second;
1884 LSRUse &LU = Uses[LUIdx];
1885 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1887 return std::make_pair(LUIdx, Offset);
1890 // Create a new use.
1891 size_t LUIdx = Uses.size();
1892 P.first->second = LUIdx;
1893 Uses.push_back(LSRUse(Kind, AccessTy));
1894 LSRUse &LU = Uses[LUIdx];
1896 // We don't need to track redundant offsets, but we don't need to go out
1897 // of our way here to avoid them.
1898 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1899 LU.Offsets.push_back(Offset);
1901 LU.MinOffset = Offset;
1902 LU.MaxOffset = Offset;
1903 return std::make_pair(LUIdx, Offset);
1906 /// DeleteUse - Delete the given use from the Uses list.
1907 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1908 if (&LU != &Uses.back())
1909 std::swap(LU, Uses.back());
1913 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1916 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1917 /// a formula that has the same registers as the given formula.
1919 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1920 const LSRUse &OrigLU) {
1921 // Search all uses for the formula. This could be more clever.
1922 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1923 LSRUse &LU = Uses[LUIdx];
1924 // Check whether this use is close enough to OrigLU, to see whether it's
1925 // worthwhile looking through its formulae.
1926 // Ignore ICmpZero uses because they may contain formulae generated by
1927 // GenerateICmpZeroScales, in which case adding fixup offsets may
1929 if (&LU != &OrigLU &&
1930 LU.Kind != LSRUse::ICmpZero &&
1931 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1932 LU.WidestFixupType == OrigLU.WidestFixupType &&
1933 LU.HasFormulaWithSameRegs(OrigF)) {
1934 // Scan through this use's formulae.
1935 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1936 E = LU.Formulae.end(); I != E; ++I) {
1937 const Formula &F = *I;
1938 // Check to see if this formula has the same registers and symbols
1940 if (F.BaseRegs == OrigF.BaseRegs &&
1941 F.ScaledReg == OrigF.ScaledReg &&
1942 F.AM.BaseGV == OrigF.AM.BaseGV &&
1943 F.AM.Scale == OrigF.AM.Scale) {
1944 if (F.AM.BaseOffs == 0)
1946 // This is the formula where all the registers and symbols matched;
1947 // there aren't going to be any others. Since we declined it, we
1948 // can skip the rest of the formulae and procede to the next LSRUse.
1955 // Nothing looked good.
1959 void LSRInstance::CollectInterestingTypesAndFactors() {
1960 SmallSetVector<const SCEV *, 4> Strides;
1962 // Collect interesting types and strides.
1963 SmallVector<const SCEV *, 4> Worklist;
1964 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1965 const SCEV *Expr = IU.getExpr(*UI);
1967 // Collect interesting types.
1968 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1970 // Add strides for mentioned loops.
1971 Worklist.push_back(Expr);
1973 const SCEV *S = Worklist.pop_back_val();
1974 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1975 Strides.insert(AR->getStepRecurrence(SE));
1976 Worklist.push_back(AR->getStart());
1977 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1978 Worklist.append(Add->op_begin(), Add->op_end());
1980 } while (!Worklist.empty());
1983 // Compute interesting factors from the set of interesting strides.
1984 for (SmallSetVector<const SCEV *, 4>::const_iterator
1985 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1986 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1987 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
1988 const SCEV *OldStride = *I;
1989 const SCEV *NewStride = *NewStrideIter;
1991 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1992 SE.getTypeSizeInBits(NewStride->getType())) {
1993 if (SE.getTypeSizeInBits(OldStride->getType()) >
1994 SE.getTypeSizeInBits(NewStride->getType()))
1995 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1997 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1999 if (const SCEVConstant *Factor =
2000 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2002 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2003 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2004 } else if (const SCEVConstant *Factor =
2005 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2008 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2009 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2013 // If all uses use the same type, don't bother looking for truncation-based
2015 if (Types.size() == 1)
2018 DEBUG(print_factors_and_types(dbgs()));
2021 void LSRInstance::CollectFixupsAndInitialFormulae() {
2022 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2024 LSRFixup &LF = getNewFixup();
2025 LF.UserInst = UI->getUser();
2026 LF.OperandValToReplace = UI->getOperandValToReplace();
2027 LF.PostIncLoops = UI->getPostIncLoops();
2029 LSRUse::KindType Kind = LSRUse::Basic;
2030 const Type *AccessTy = 0;
2031 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2032 Kind = LSRUse::Address;
2033 AccessTy = getAccessType(LF.UserInst);
2036 const SCEV *S = IU.getExpr(*UI);
2038 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2039 // (N - i == 0), and this allows (N - i) to be the expression that we work
2040 // with rather than just N or i, so we can consider the register
2041 // requirements for both N and i at the same time. Limiting this code to
2042 // equality icmps is not a problem because all interesting loops use
2043 // equality icmps, thanks to IndVarSimplify.
2044 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2045 if (CI->isEquality()) {
2046 // Swap the operands if needed to put the OperandValToReplace on the
2047 // left, for consistency.
2048 Value *NV = CI->getOperand(1);
2049 if (NV == LF.OperandValToReplace) {
2050 CI->setOperand(1, CI->getOperand(0));
2051 CI->setOperand(0, NV);
2052 NV = CI->getOperand(1);
2056 // x == y --> x - y == 0
2057 const SCEV *N = SE.getSCEV(NV);
2058 if (SE.isLoopInvariant(N, L)) {
2059 Kind = LSRUse::ICmpZero;
2060 S = SE.getMinusSCEV(N, S);
2063 // -1 and the negations of all interesting strides (except the negation
2064 // of -1) are now also interesting.
2065 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2066 if (Factors[i] != -1)
2067 Factors.insert(-(uint64_t)Factors[i]);
2071 // Set up the initial formula for this use.
2072 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2074 LF.Offset = P.second;
2075 LSRUse &LU = Uses[LF.LUIdx];
2076 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2077 if (!LU.WidestFixupType ||
2078 SE.getTypeSizeInBits(LU.WidestFixupType) <
2079 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2080 LU.WidestFixupType = LF.OperandValToReplace->getType();
2082 // If this is the first use of this LSRUse, give it a formula.
2083 if (LU.Formulae.empty()) {
2084 InsertInitialFormula(S, LU, LF.LUIdx);
2085 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2089 DEBUG(print_fixups(dbgs()));
2092 /// InsertInitialFormula - Insert a formula for the given expression into
2093 /// the given use, separating out loop-variant portions from loop-invariant
2094 /// and loop-computable portions.
2096 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2098 F.InitialMatch(S, L, SE);
2099 bool Inserted = InsertFormula(LU, LUIdx, F);
2100 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2103 /// InsertSupplementalFormula - Insert a simple single-register formula for
2104 /// the given expression into the given use.
2106 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2107 LSRUse &LU, size_t LUIdx) {
2109 F.BaseRegs.push_back(S);
2110 F.AM.HasBaseReg = true;
2111 bool Inserted = InsertFormula(LU, LUIdx, F);
2112 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2115 /// CountRegisters - Note which registers are used by the given formula,
2116 /// updating RegUses.
2117 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2119 RegUses.CountRegister(F.ScaledReg, LUIdx);
2120 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2121 E = F.BaseRegs.end(); I != E; ++I)
2122 RegUses.CountRegister(*I, LUIdx);
2125 /// InsertFormula - If the given formula has not yet been inserted, add it to
2126 /// the list, and return true. Return false otherwise.
2127 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2128 if (!LU.InsertFormula(F))
2131 CountRegisters(F, LUIdx);
2135 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2136 /// loop-invariant values which we're tracking. These other uses will pin these
2137 /// values in registers, making them less profitable for elimination.
2138 /// TODO: This currently misses non-constant addrec step registers.
2139 /// TODO: Should this give more weight to users inside the loop?
2141 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2142 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2143 SmallPtrSet<const SCEV *, 8> Inserted;
2145 while (!Worklist.empty()) {
2146 const SCEV *S = Worklist.pop_back_val();
2148 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2149 Worklist.append(N->op_begin(), N->op_end());
2150 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2151 Worklist.push_back(C->getOperand());
2152 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2153 Worklist.push_back(D->getLHS());
2154 Worklist.push_back(D->getRHS());
2155 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2156 if (!Inserted.insert(U)) continue;
2157 const Value *V = U->getValue();
2158 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2159 // Look for instructions defined outside the loop.
2160 if (L->contains(Inst)) continue;
2161 } else if (isa<UndefValue>(V))
2162 // Undef doesn't have a live range, so it doesn't matter.
2164 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2166 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2167 // Ignore non-instructions.
2170 // Ignore instructions in other functions (as can happen with
2172 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2174 // Ignore instructions not dominated by the loop.
2175 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2176 UserInst->getParent() :
2177 cast<PHINode>(UserInst)->getIncomingBlock(
2178 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2179 if (!DT.dominates(L->getHeader(), UseBB))
2181 // Ignore uses which are part of other SCEV expressions, to avoid
2182 // analyzing them multiple times.
2183 if (SE.isSCEVable(UserInst->getType())) {
2184 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2185 // If the user is a no-op, look through to its uses.
2186 if (!isa<SCEVUnknown>(UserS))
2190 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2194 // Ignore icmp instructions which are already being analyzed.
2195 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2196 unsigned OtherIdx = !UI.getOperandNo();
2197 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2198 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2202 LSRFixup &LF = getNewFixup();
2203 LF.UserInst = const_cast<Instruction *>(UserInst);
2204 LF.OperandValToReplace = UI.getUse();
2205 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2207 LF.Offset = P.second;
2208 LSRUse &LU = Uses[LF.LUIdx];
2209 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2210 if (!LU.WidestFixupType ||
2211 SE.getTypeSizeInBits(LU.WidestFixupType) <
2212 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2213 LU.WidestFixupType = LF.OperandValToReplace->getType();
2214 InsertSupplementalFormula(U, LU, LF.LUIdx);
2215 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2222 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2223 /// separate registers. If C is non-null, multiply each subexpression by C.
2224 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2225 SmallVectorImpl<const SCEV *> &Ops,
2227 ScalarEvolution &SE) {
2228 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2229 // Break out add operands.
2230 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2232 CollectSubexprs(*I, C, Ops, L, SE);
2234 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2235 // Split a non-zero base out of an addrec.
2236 if (!AR->getStart()->isZero()) {
2237 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2238 AR->getStepRecurrence(SE),
2241 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2244 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2245 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2246 if (Mul->getNumOperands() == 2)
2247 if (const SCEVConstant *Op0 =
2248 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2249 CollectSubexprs(Mul->getOperand(1),
2250 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2256 // Otherwise use the value itself, optionally with a scale applied.
2257 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2260 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2262 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2265 // Arbitrarily cap recursion to protect compile time.
2266 if (Depth >= 3) return;
2268 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2269 const SCEV *BaseReg = Base.BaseRegs[i];
2271 SmallVector<const SCEV *, 8> AddOps;
2272 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2274 if (AddOps.size() == 1) continue;
2276 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2277 JE = AddOps.end(); J != JE; ++J) {
2279 // Loop-variant "unknown" values are uninteresting; we won't be able to
2280 // do anything meaningful with them.
2281 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2284 // Don't pull a constant into a register if the constant could be folded
2285 // into an immediate field.
2286 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2287 Base.getNumRegs() > 1,
2288 LU.Kind, LU.AccessTy, TLI, SE))
2291 // Collect all operands except *J.
2292 SmallVector<const SCEV *, 8> InnerAddOps
2293 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2295 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2297 // Don't leave just a constant behind in a register if the constant could
2298 // be folded into an immediate field.
2299 if (InnerAddOps.size() == 1 &&
2300 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2301 Base.getNumRegs() > 1,
2302 LU.Kind, LU.AccessTy, TLI, SE))
2305 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2306 if (InnerSum->isZero())
2309 F.BaseRegs[i] = InnerSum;
2310 F.BaseRegs.push_back(*J);
2311 if (InsertFormula(LU, LUIdx, F))
2312 // If that formula hadn't been seen before, recurse to find more like
2314 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2319 /// GenerateCombinations - Generate a formula consisting of all of the
2320 /// loop-dominating registers added into a single register.
2321 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2323 // This method is only interesting on a plurality of registers.
2324 if (Base.BaseRegs.size() <= 1) return;
2328 SmallVector<const SCEV *, 4> Ops;
2329 for (SmallVectorImpl<const SCEV *>::const_iterator
2330 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2331 const SCEV *BaseReg = *I;
2332 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2333 !SE.hasComputableLoopEvolution(BaseReg, L))
2334 Ops.push_back(BaseReg);
2336 F.BaseRegs.push_back(BaseReg);
2338 if (Ops.size() > 1) {
2339 const SCEV *Sum = SE.getAddExpr(Ops);
2340 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2341 // opportunity to fold something. For now, just ignore such cases
2342 // rather than proceed with zero in a register.
2343 if (!Sum->isZero()) {
2344 F.BaseRegs.push_back(Sum);
2345 (void)InsertFormula(LU, LUIdx, F);
2350 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2351 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2353 // We can't add a symbolic offset if the address already contains one.
2354 if (Base.AM.BaseGV) return;
2356 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2357 const SCEV *G = Base.BaseRegs[i];
2358 GlobalValue *GV = ExtractSymbol(G, SE);
2359 if (G->isZero() || !GV)
2363 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2364 LU.Kind, LU.AccessTy, TLI))
2367 (void)InsertFormula(LU, LUIdx, F);
2371 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2372 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2374 // TODO: For now, just add the min and max offset, because it usually isn't
2375 // worthwhile looking at everything inbetween.
2376 SmallVector<int64_t, 2> Worklist;
2377 Worklist.push_back(LU.MinOffset);
2378 if (LU.MaxOffset != LU.MinOffset)
2379 Worklist.push_back(LU.MaxOffset);
2381 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2382 const SCEV *G = Base.BaseRegs[i];
2384 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2385 E = Worklist.end(); I != E; ++I) {
2387 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2388 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2389 LU.Kind, LU.AccessTy, TLI)) {
2390 // Add the offset to the base register.
2391 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2392 // If it cancelled out, drop the base register, otherwise update it.
2393 if (NewG->isZero()) {
2394 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2395 F.BaseRegs.pop_back();
2397 F.BaseRegs[i] = NewG;
2399 (void)InsertFormula(LU, LUIdx, F);
2403 int64_t Imm = ExtractImmediate(G, SE);
2404 if (G->isZero() || Imm == 0)
2407 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2408 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2409 LU.Kind, LU.AccessTy, TLI))
2412 (void)InsertFormula(LU, LUIdx, F);
2416 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2417 /// the comparison. For example, x == y -> x*c == y*c.
2418 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2420 if (LU.Kind != LSRUse::ICmpZero) return;
2422 // Determine the integer type for the base formula.
2423 const Type *IntTy = Base.getType();
2425 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2427 // Don't do this if there is more than one offset.
2428 if (LU.MinOffset != LU.MaxOffset) return;
2430 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2432 // Check each interesting stride.
2433 for (SmallSetVector<int64_t, 8>::const_iterator
2434 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2435 int64_t Factor = *I;
2437 // Check that the multiplication doesn't overflow.
2438 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2440 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2441 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2444 // Check that multiplying with the use offset doesn't overflow.
2445 int64_t Offset = LU.MinOffset;
2446 if (Offset == INT64_MIN && Factor == -1)
2448 Offset = (uint64_t)Offset * Factor;
2449 if (Offset / Factor != LU.MinOffset)
2453 F.AM.BaseOffs = NewBaseOffs;
2455 // Check that this scale is legal.
2456 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2459 // Compensate for the use having MinOffset built into it.
2460 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2462 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2464 // Check that multiplying with each base register doesn't overflow.
2465 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2466 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2467 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2471 // Check that multiplying with the scaled register doesn't overflow.
2473 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2474 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2478 // If we make it here and it's legal, add it.
2479 (void)InsertFormula(LU, LUIdx, F);
2484 /// GenerateScales - Generate stride factor reuse formulae by making use of
2485 /// scaled-offset address modes, for example.
2486 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2487 // Determine the integer type for the base formula.
2488 const Type *IntTy = Base.getType();
2491 // If this Formula already has a scaled register, we can't add another one.
2492 if (Base.AM.Scale != 0) return;
2494 // Check each interesting stride.
2495 for (SmallSetVector<int64_t, 8>::const_iterator
2496 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2497 int64_t Factor = *I;
2499 Base.AM.Scale = Factor;
2500 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2501 // Check whether this scale is going to be legal.
2502 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2503 LU.Kind, LU.AccessTy, TLI)) {
2504 // As a special-case, handle special out-of-loop Basic users specially.
2505 // TODO: Reconsider this special case.
2506 if (LU.Kind == LSRUse::Basic &&
2507 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2508 LSRUse::Special, LU.AccessTy, TLI) &&
2509 LU.AllFixupsOutsideLoop)
2510 LU.Kind = LSRUse::Special;
2514 // For an ICmpZero, negating a solitary base register won't lead to
2516 if (LU.Kind == LSRUse::ICmpZero &&
2517 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2519 // For each addrec base reg, apply the scale, if possible.
2520 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2521 if (const SCEVAddRecExpr *AR =
2522 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2523 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2524 if (FactorS->isZero())
2526 // Divide out the factor, ignoring high bits, since we'll be
2527 // scaling the value back up in the end.
2528 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2529 // TODO: This could be optimized to avoid all the copying.
2531 F.ScaledReg = Quotient;
2532 F.DeleteBaseReg(F.BaseRegs[i]);
2533 (void)InsertFormula(LU, LUIdx, F);
2539 /// GenerateTruncates - Generate reuse formulae from different IV types.
2540 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2541 // This requires TargetLowering to tell us which truncates are free.
2544 // Don't bother truncating symbolic values.
2545 if (Base.AM.BaseGV) return;
2547 // Determine the integer type for the base formula.
2548 const Type *DstTy = Base.getType();
2550 DstTy = SE.getEffectiveSCEVType(DstTy);
2552 for (SmallSetVector<const Type *, 4>::const_iterator
2553 I = Types.begin(), E = Types.end(); I != E; ++I) {
2554 const Type *SrcTy = *I;
2555 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2558 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2559 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2560 JE = F.BaseRegs.end(); J != JE; ++J)
2561 *J = SE.getAnyExtendExpr(*J, SrcTy);
2563 // TODO: This assumes we've done basic processing on all uses and
2564 // have an idea what the register usage is.
2565 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2568 (void)InsertFormula(LU, LUIdx, F);
2575 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2576 /// defer modifications so that the search phase doesn't have to worry about
2577 /// the data structures moving underneath it.
2581 const SCEV *OrigReg;
2583 WorkItem(size_t LI, int64_t I, const SCEV *R)
2584 : LUIdx(LI), Imm(I), OrigReg(R) {}
2586 void print(raw_ostream &OS) const;
2592 void WorkItem::print(raw_ostream &OS) const {
2593 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2594 << " , add offset " << Imm;
2597 void WorkItem::dump() const {
2598 print(errs()); errs() << '\n';
2601 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2602 /// distance apart and try to form reuse opportunities between them.
2603 void LSRInstance::GenerateCrossUseConstantOffsets() {
2604 // Group the registers by their value without any added constant offset.
2605 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2606 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2608 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2609 SmallVector<const SCEV *, 8> Sequence;
2610 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2612 const SCEV *Reg = *I;
2613 int64_t Imm = ExtractImmediate(Reg, SE);
2614 std::pair<RegMapTy::iterator, bool> Pair =
2615 Map.insert(std::make_pair(Reg, ImmMapTy()));
2617 Sequence.push_back(Reg);
2618 Pair.first->second.insert(std::make_pair(Imm, *I));
2619 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2622 // Now examine each set of registers with the same base value. Build up
2623 // a list of work to do and do the work in a separate step so that we're
2624 // not adding formulae and register counts while we're searching.
2625 SmallVector<WorkItem, 32> WorkItems;
2626 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2627 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2628 E = Sequence.end(); I != E; ++I) {
2629 const SCEV *Reg = *I;
2630 const ImmMapTy &Imms = Map.find(Reg)->second;
2632 // It's not worthwhile looking for reuse if there's only one offset.
2633 if (Imms.size() == 1)
2636 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2637 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2639 dbgs() << ' ' << J->first;
2642 // Examine each offset.
2643 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2645 const SCEV *OrigReg = J->second;
2647 int64_t JImm = J->first;
2648 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2650 if (!isa<SCEVConstant>(OrigReg) &&
2651 UsedByIndicesMap[Reg].count() == 1) {
2652 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2656 // Conservatively examine offsets between this orig reg a few selected
2658 ImmMapTy::const_iterator OtherImms[] = {
2659 Imms.begin(), prior(Imms.end()),
2660 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2662 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2663 ImmMapTy::const_iterator M = OtherImms[i];
2664 if (M == J || M == JE) continue;
2666 // Compute the difference between the two.
2667 int64_t Imm = (uint64_t)JImm - M->first;
2668 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2669 LUIdx = UsedByIndices.find_next(LUIdx))
2670 // Make a memo of this use, offset, and register tuple.
2671 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2672 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2679 UsedByIndicesMap.clear();
2680 UniqueItems.clear();
2682 // Now iterate through the worklist and add new formulae.
2683 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2684 E = WorkItems.end(); I != E; ++I) {
2685 const WorkItem &WI = *I;
2686 size_t LUIdx = WI.LUIdx;
2687 LSRUse &LU = Uses[LUIdx];
2688 int64_t Imm = WI.Imm;
2689 const SCEV *OrigReg = WI.OrigReg;
2691 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2692 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2693 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2695 // TODO: Use a more targeted data structure.
2696 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2697 const Formula &F = LU.Formulae[L];
2698 // Use the immediate in the scaled register.
2699 if (F.ScaledReg == OrigReg) {
2700 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2701 Imm * (uint64_t)F.AM.Scale;
2702 // Don't create 50 + reg(-50).
2703 if (F.referencesReg(SE.getSCEV(
2704 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2707 NewF.AM.BaseOffs = Offs;
2708 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2709 LU.Kind, LU.AccessTy, TLI))
2711 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2713 // If the new scale is a constant in a register, and adding the constant
2714 // value to the immediate would produce a value closer to zero than the
2715 // immediate itself, then the formula isn't worthwhile.
2716 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2717 if (C->getValue()->getValue().isNegative() !=
2718 (NewF.AM.BaseOffs < 0) &&
2719 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2720 .ule(abs64(NewF.AM.BaseOffs)))
2724 (void)InsertFormula(LU, LUIdx, NewF);
2726 // Use the immediate in a base register.
2727 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2728 const SCEV *BaseReg = F.BaseRegs[N];
2729 if (BaseReg != OrigReg)
2732 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2733 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2734 LU.Kind, LU.AccessTy, TLI))
2736 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2738 // If the new formula has a constant in a register, and adding the
2739 // constant value to the immediate would produce a value closer to
2740 // zero than the immediate itself, then the formula isn't worthwhile.
2741 for (SmallVectorImpl<const SCEV *>::const_iterator
2742 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2744 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2745 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2746 abs64(NewF.AM.BaseOffs)) &&
2747 (C->getValue()->getValue() +
2748 NewF.AM.BaseOffs).countTrailingZeros() >=
2749 CountTrailingZeros_64(NewF.AM.BaseOffs))
2753 (void)InsertFormula(LU, LUIdx, NewF);
2762 /// GenerateAllReuseFormulae - Generate formulae for each use.
2764 LSRInstance::GenerateAllReuseFormulae() {
2765 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2766 // queries are more precise.
2767 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2768 LSRUse &LU = Uses[LUIdx];
2769 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2770 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2771 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2772 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2774 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2775 LSRUse &LU = Uses[LUIdx];
2776 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2777 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2778 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2779 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2780 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2781 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2782 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2783 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2785 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2786 LSRUse &LU = Uses[LUIdx];
2787 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2788 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2791 GenerateCrossUseConstantOffsets();
2793 DEBUG(dbgs() << "\n"
2794 "After generating reuse formulae:\n";
2795 print_uses(dbgs()));
2798 /// If there are multiple formulae with the same set of registers used
2799 /// by other uses, pick the best one and delete the others.
2800 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2801 DenseSet<const SCEV *> VisitedRegs;
2802 SmallPtrSet<const SCEV *, 16> Regs;
2804 bool ChangedFormulae = false;
2807 // Collect the best formula for each unique set of shared registers. This
2808 // is reset for each use.
2809 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2811 BestFormulaeTy BestFormulae;
2813 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2814 LSRUse &LU = Uses[LUIdx];
2815 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2818 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2819 FIdx != NumForms; ++FIdx) {
2820 Formula &F = LU.Formulae[FIdx];
2822 SmallVector<const SCEV *, 2> Key;
2823 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2824 JE = F.BaseRegs.end(); J != JE; ++J) {
2825 const SCEV *Reg = *J;
2826 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2830 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2831 Key.push_back(F.ScaledReg);
2832 // Unstable sort by host order ok, because this is only used for
2834 std::sort(Key.begin(), Key.end());
2836 std::pair<BestFormulaeTy::const_iterator, bool> P =
2837 BestFormulae.insert(std::make_pair(Key, FIdx));
2839 Formula &Best = LU.Formulae[P.first->second];
2842 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2845 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2847 if (CostF < CostBest)
2849 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2851 " in favor of formula "; Best.print(dbgs());
2854 ChangedFormulae = true;
2856 LU.DeleteFormula(F);
2864 // Now that we've filtered out some formulae, recompute the Regs set.
2866 LU.RecomputeRegs(LUIdx, RegUses);
2868 // Reset this to prepare for the next use.
2869 BestFormulae.clear();
2872 DEBUG(if (ChangedFormulae) {
2874 "After filtering out undesirable candidates:\n";
2879 // This is a rough guess that seems to work fairly well.
2880 static const size_t ComplexityLimit = UINT16_MAX;
2882 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2883 /// solutions the solver might have to consider. It almost never considers
2884 /// this many solutions because it prune the search space, but the pruning
2885 /// isn't always sufficient.
2886 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2888 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2889 E = Uses.end(); I != E; ++I) {
2890 size_t FSize = I->Formulae.size();
2891 if (FSize >= ComplexityLimit) {
2892 Power = ComplexityLimit;
2896 if (Power >= ComplexityLimit)
2902 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2903 /// of the registers of another formula, it won't help reduce register
2904 /// pressure (though it may not necessarily hurt register pressure); remove
2905 /// it to simplify the system.
2906 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2907 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2908 DEBUG(dbgs() << "The search space is too complex.\n");
2910 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2911 "which use a superset of registers used by other "
2914 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2915 LSRUse &LU = Uses[LUIdx];
2917 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2918 Formula &F = LU.Formulae[i];
2919 // Look for a formula with a constant or GV in a register. If the use
2920 // also has a formula with that same value in an immediate field,
2921 // delete the one that uses a register.
2922 for (SmallVectorImpl<const SCEV *>::const_iterator
2923 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2924 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2926 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2927 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2928 (I - F.BaseRegs.begin()));
2929 if (LU.HasFormulaWithSameRegs(NewF)) {
2930 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2931 LU.DeleteFormula(F);
2937 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2938 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2941 NewF.AM.BaseGV = GV;
2942 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2943 (I - F.BaseRegs.begin()));
2944 if (LU.HasFormulaWithSameRegs(NewF)) {
2945 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2947 LU.DeleteFormula(F);
2958 LU.RecomputeRegs(LUIdx, RegUses);
2961 DEBUG(dbgs() << "After pre-selection:\n";
2962 print_uses(dbgs()));
2966 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
2967 /// for expressions like A, A+1, A+2, etc., allocate a single register for
2969 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
2970 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2971 DEBUG(dbgs() << "The search space is too complex.\n");
2973 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2974 "separated by a constant offset will use the same "
2977 // This is especially useful for unrolled loops.
2979 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2980 LSRUse &LU = Uses[LUIdx];
2981 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2982 E = LU.Formulae.end(); I != E; ++I) {
2983 const Formula &F = *I;
2984 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2985 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2986 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2987 /*HasBaseReg=*/false,
2988 LU.Kind, LU.AccessTy)) {
2989 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2992 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2994 // Update the relocs to reference the new use.
2995 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
2996 E = Fixups.end(); I != E; ++I) {
2997 LSRFixup &Fixup = *I;
2998 if (Fixup.LUIdx == LUIdx) {
2999 Fixup.LUIdx = LUThatHas - &Uses.front();
3000 Fixup.Offset += F.AM.BaseOffs;
3001 // Add the new offset to LUThatHas' offset list.
3002 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3003 LUThatHas->Offsets.push_back(Fixup.Offset);
3004 if (Fixup.Offset > LUThatHas->MaxOffset)
3005 LUThatHas->MaxOffset = Fixup.Offset;
3006 if (Fixup.Offset < LUThatHas->MinOffset)
3007 LUThatHas->MinOffset = Fixup.Offset;
3009 DEBUG(dbgs() << "New fixup has offset "
3010 << Fixup.Offset << '\n');
3012 if (Fixup.LUIdx == NumUses-1)
3013 Fixup.LUIdx = LUIdx;
3016 // Delete formulae from the new use which are no longer legal.
3018 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3019 Formula &F = LUThatHas->Formulae[i];
3020 if (!isLegalUse(F.AM,
3021 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3022 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3023 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3025 LUThatHas->DeleteFormula(F);
3032 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3034 // Delete the old use.
3035 DeleteUse(LU, LUIdx);
3045 DEBUG(dbgs() << "After pre-selection:\n";
3046 print_uses(dbgs()));
3050 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3051 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3052 /// we've done more filtering, as it may be able to find more formulae to
3054 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3055 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3056 DEBUG(dbgs() << "The search space is too complex.\n");
3058 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3059 "undesirable dedicated registers.\n");
3061 FilterOutUndesirableDedicatedRegisters();
3063 DEBUG(dbgs() << "After pre-selection:\n";
3064 print_uses(dbgs()));
3068 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3069 /// to be profitable, and then in any use which has any reference to that
3070 /// register, delete all formulae which do not reference that register.
3071 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3072 // With all other options exhausted, loop until the system is simple
3073 // enough to handle.
3074 SmallPtrSet<const SCEV *, 4> Taken;
3075 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3076 // Ok, we have too many of formulae on our hands to conveniently handle.
3077 // Use a rough heuristic to thin out the list.
3078 DEBUG(dbgs() << "The search space is too complex.\n");
3080 // Pick the register which is used by the most LSRUses, which is likely
3081 // to be a good reuse register candidate.
3082 const SCEV *Best = 0;
3083 unsigned BestNum = 0;
3084 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3086 const SCEV *Reg = *I;
3087 if (Taken.count(Reg))
3092 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3093 if (Count > BestNum) {
3100 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3101 << " will yield profitable reuse.\n");
3104 // In any use with formulae which references this register, delete formulae
3105 // which don't reference it.
3106 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3107 LSRUse &LU = Uses[LUIdx];
3108 if (!LU.Regs.count(Best)) continue;
3111 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3112 Formula &F = LU.Formulae[i];
3113 if (!F.referencesReg(Best)) {
3114 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3115 LU.DeleteFormula(F);
3119 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3125 LU.RecomputeRegs(LUIdx, RegUses);
3128 DEBUG(dbgs() << "After pre-selection:\n";
3129 print_uses(dbgs()));
3133 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3134 /// formulae to choose from, use some rough heuristics to prune down the number
3135 /// of formulae. This keeps the main solver from taking an extraordinary amount
3136 /// of time in some worst-case scenarios.
3137 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3138 NarrowSearchSpaceByDetectingSupersets();
3139 NarrowSearchSpaceByCollapsingUnrolledCode();
3140 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3141 NarrowSearchSpaceByPickingWinnerRegs();
3144 /// SolveRecurse - This is the recursive solver.
3145 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3147 SmallVectorImpl<const Formula *> &Workspace,
3148 const Cost &CurCost,
3149 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3150 DenseSet<const SCEV *> &VisitedRegs) const {
3153 // - use more aggressive filtering
3154 // - sort the formula so that the most profitable solutions are found first
3155 // - sort the uses too
3157 // - don't compute a cost, and then compare. compare while computing a cost
3159 // - track register sets with SmallBitVector
3161 const LSRUse &LU = Uses[Workspace.size()];
3163 // If this use references any register that's already a part of the
3164 // in-progress solution, consider it a requirement that a formula must
3165 // reference that register in order to be considered. This prunes out
3166 // unprofitable searching.
3167 SmallSetVector<const SCEV *, 4> ReqRegs;
3168 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3169 E = CurRegs.end(); I != E; ++I)
3170 if (LU.Regs.count(*I))
3173 bool AnySatisfiedReqRegs = false;
3174 SmallPtrSet<const SCEV *, 16> NewRegs;
3177 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3178 E = LU.Formulae.end(); I != E; ++I) {
3179 const Formula &F = *I;
3181 // Ignore formulae which do not use any of the required registers.
3182 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3183 JE = ReqRegs.end(); J != JE; ++J) {
3184 const SCEV *Reg = *J;
3185 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3186 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3190 AnySatisfiedReqRegs = true;
3192 // Evaluate the cost of the current formula. If it's already worse than
3193 // the current best, prune the search at that point.
3196 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3197 if (NewCost < SolutionCost) {
3198 Workspace.push_back(&F);
3199 if (Workspace.size() != Uses.size()) {
3200 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3201 NewRegs, VisitedRegs);
3202 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3203 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3205 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3206 dbgs() << ". Regs:";
3207 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3208 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3209 dbgs() << ' ' << **I;
3212 SolutionCost = NewCost;
3213 Solution = Workspace;
3215 Workspace.pop_back();
3220 // If none of the formulae had all of the required registers, relax the
3221 // constraint so that we don't exclude all formulae.
3222 if (!AnySatisfiedReqRegs) {
3223 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3229 /// Solve - Choose one formula from each use. Return the results in the given
3230 /// Solution vector.
3231 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3232 SmallVector<const Formula *, 8> Workspace;
3234 SolutionCost.Loose();
3236 SmallPtrSet<const SCEV *, 16> CurRegs;
3237 DenseSet<const SCEV *> VisitedRegs;
3238 Workspace.reserve(Uses.size());
3240 // SolveRecurse does all the work.
3241 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3242 CurRegs, VisitedRegs);
3244 // Ok, we've now made all our decisions.
3245 DEBUG(dbgs() << "\n"
3246 "The chosen solution requires "; SolutionCost.print(dbgs());
3248 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3250 Uses[i].print(dbgs());
3253 Solution[i]->print(dbgs());
3257 assert(Solution.size() == Uses.size() && "Malformed solution!");
3260 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3261 /// the dominator tree far as we can go while still being dominated by the
3262 /// input positions. This helps canonicalize the insert position, which
3263 /// encourages sharing.
3264 BasicBlock::iterator
3265 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3266 const SmallVectorImpl<Instruction *> &Inputs)
3269 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3270 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3273 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3274 if (!Rung) return IP;
3275 Rung = Rung->getIDom();
3276 if (!Rung) return IP;
3277 IDom = Rung->getBlock();
3279 // Don't climb into a loop though.
3280 const Loop *IDomLoop = LI.getLoopFor(IDom);
3281 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3282 if (IDomDepth <= IPLoopDepth &&
3283 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3287 bool AllDominate = true;
3288 Instruction *BetterPos = 0;
3289 Instruction *Tentative = IDom->getTerminator();
3290 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3291 E = Inputs.end(); I != E; ++I) {
3292 Instruction *Inst = *I;
3293 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3294 AllDominate = false;
3297 // Attempt to find an insert position in the middle of the block,
3298 // instead of at the end, so that it can be used for other expansions.
3299 if (IDom == Inst->getParent() &&
3300 (!BetterPos || DT.dominates(BetterPos, Inst)))
3301 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3314 /// AdjustInsertPositionForExpand - Determine an input position which will be
3315 /// dominated by the operands and which will dominate the result.
3316 BasicBlock::iterator
3317 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3319 const LSRUse &LU) const {
3320 // Collect some instructions which must be dominated by the
3321 // expanding replacement. These must be dominated by any operands that
3322 // will be required in the expansion.
3323 SmallVector<Instruction *, 4> Inputs;
3324 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3325 Inputs.push_back(I);
3326 if (LU.Kind == LSRUse::ICmpZero)
3327 if (Instruction *I =
3328 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3329 Inputs.push_back(I);
3330 if (LF.PostIncLoops.count(L)) {
3331 if (LF.isUseFullyOutsideLoop(L))
3332 Inputs.push_back(L->getLoopLatch()->getTerminator());
3334 Inputs.push_back(IVIncInsertPos);
3336 // The expansion must also be dominated by the increment positions of any
3337 // loops it for which it is using post-inc mode.
3338 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3339 E = LF.PostIncLoops.end(); I != E; ++I) {
3340 const Loop *PIL = *I;
3341 if (PIL == L) continue;
3343 // Be dominated by the loop exit.
3344 SmallVector<BasicBlock *, 4> ExitingBlocks;
3345 PIL->getExitingBlocks(ExitingBlocks);
3346 if (!ExitingBlocks.empty()) {
3347 BasicBlock *BB = ExitingBlocks[0];
3348 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3349 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3350 Inputs.push_back(BB->getTerminator());
3354 // Then, climb up the immediate dominator tree as far as we can go while
3355 // still being dominated by the input positions.
3356 IP = HoistInsertPosition(IP, Inputs);
3358 // Don't insert instructions before PHI nodes.
3359 while (isa<PHINode>(IP)) ++IP;
3361 // Ignore debug intrinsics.
3362 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3367 /// Expand - Emit instructions for the leading candidate expression for this
3368 /// LSRUse (this is called "expanding").
3369 Value *LSRInstance::Expand(const LSRFixup &LF,
3371 BasicBlock::iterator IP,
3372 SCEVExpander &Rewriter,
3373 SmallVectorImpl<WeakVH> &DeadInsts) const {
3374 const LSRUse &LU = Uses[LF.LUIdx];
3376 // Determine an input position which will be dominated by the operands and
3377 // which will dominate the result.
3378 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3380 // Inform the Rewriter if we have a post-increment use, so that it can
3381 // perform an advantageous expansion.
3382 Rewriter.setPostInc(LF.PostIncLoops);
3384 // This is the type that the user actually needs.
3385 const Type *OpTy = LF.OperandValToReplace->getType();
3386 // This will be the type that we'll initially expand to.
3387 const Type *Ty = F.getType();
3389 // No type known; just expand directly to the ultimate type.
3391 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3392 // Expand directly to the ultimate type if it's the right size.
3394 // This is the type to do integer arithmetic in.
3395 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3397 // Build up a list of operands to add together to form the full base.
3398 SmallVector<const SCEV *, 8> Ops;
3400 // Expand the BaseRegs portion.
3401 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3402 E = F.BaseRegs.end(); I != E; ++I) {
3403 const SCEV *Reg = *I;
3404 assert(!Reg->isZero() && "Zero allocated in a base register!");
3406 // If we're expanding for a post-inc user, make the post-inc adjustment.
3407 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3408 Reg = TransformForPostIncUse(Denormalize, Reg,
3409 LF.UserInst, LF.OperandValToReplace,
3412 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3415 // Flush the operand list to suppress SCEVExpander hoisting.
3417 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3419 Ops.push_back(SE.getUnknown(FullV));
3422 // Expand the ScaledReg portion.
3423 Value *ICmpScaledV = 0;
3424 if (F.AM.Scale != 0) {
3425 const SCEV *ScaledS = F.ScaledReg;
3427 // If we're expanding for a post-inc user, make the post-inc adjustment.
3428 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3429 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3430 LF.UserInst, LF.OperandValToReplace,
3433 if (LU.Kind == LSRUse::ICmpZero) {
3434 // An interesting way of "folding" with an icmp is to use a negated
3435 // scale, which we'll implement by inserting it into the other operand
3437 assert(F.AM.Scale == -1 &&
3438 "The only scale supported by ICmpZero uses is -1!");
3439 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3441 // Otherwise just expand the scaled register and an explicit scale,
3442 // which is expected to be matched as part of the address.
3443 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3444 ScaledS = SE.getMulExpr(ScaledS,
3445 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3446 Ops.push_back(ScaledS);
3448 // Flush the operand list to suppress SCEVExpander hoisting.
3449 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3451 Ops.push_back(SE.getUnknown(FullV));
3455 // Expand the GV portion.
3457 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3459 // Flush the operand list to suppress SCEVExpander hoisting.
3460 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3462 Ops.push_back(SE.getUnknown(FullV));
3465 // Expand the immediate portion.
3466 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3468 if (LU.Kind == LSRUse::ICmpZero) {
3469 // The other interesting way of "folding" with an ICmpZero is to use a
3470 // negated immediate.
3472 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3474 Ops.push_back(SE.getUnknown(ICmpScaledV));
3475 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3478 // Just add the immediate values. These again are expected to be matched
3479 // as part of the address.
3480 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3484 // Emit instructions summing all the operands.
3485 const SCEV *FullS = Ops.empty() ?
3486 SE.getConstant(IntTy, 0) :
3488 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3490 // We're done expanding now, so reset the rewriter.
3491 Rewriter.clearPostInc();
3493 // An ICmpZero Formula represents an ICmp which we're handling as a
3494 // comparison against zero. Now that we've expanded an expression for that
3495 // form, update the ICmp's other operand.
3496 if (LU.Kind == LSRUse::ICmpZero) {
3497 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3498 DeadInsts.push_back(CI->getOperand(1));
3499 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3500 "a scale at the same time!");
3501 if (F.AM.Scale == -1) {
3502 if (ICmpScaledV->getType() != OpTy) {
3504 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3506 ICmpScaledV, OpTy, "tmp", CI);
3509 CI->setOperand(1, ICmpScaledV);
3511 assert(F.AM.Scale == 0 &&
3512 "ICmp does not support folding a global value and "
3513 "a scale at the same time!");
3514 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3516 if (C->getType() != OpTy)
3517 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3521 CI->setOperand(1, C);
3528 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3529 /// of their operands effectively happens in their predecessor blocks, so the
3530 /// expression may need to be expanded in multiple places.
3531 void LSRInstance::RewriteForPHI(PHINode *PN,
3534 SCEVExpander &Rewriter,
3535 SmallVectorImpl<WeakVH> &DeadInsts,
3537 DenseMap<BasicBlock *, Value *> Inserted;
3538 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3539 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3540 BasicBlock *BB = PN->getIncomingBlock(i);
3542 // If this is a critical edge, split the edge so that we do not insert
3543 // the code on all predecessor/successor paths. We do this unless this
3544 // is the canonical backedge for this loop, which complicates post-inc
3546 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3547 !isa<IndirectBrInst>(BB->getTerminator())) {
3548 Loop *PNLoop = LI.getLoopFor(PN->getParent());
3549 if (!PNLoop || PN->getParent() != PNLoop->getHeader()) {
3550 // Split the critical edge.
3551 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3553 // If PN is outside of the loop and BB is in the loop, we want to
3554 // move the block to be immediately before the PHI block, not
3555 // immediately after BB.
3556 if (L->contains(BB) && !L->contains(PN))
3557 NewBB->moveBefore(PN->getParent());
3559 // Splitting the edge can reduce the number of PHI entries we have.
3560 e = PN->getNumIncomingValues();
3562 i = PN->getBasicBlockIndex(BB);
3566 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3567 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3569 PN->setIncomingValue(i, Pair.first->second);
3571 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3573 // If this is reuse-by-noop-cast, insert the noop cast.
3574 const Type *OpTy = LF.OperandValToReplace->getType();
3575 if (FullV->getType() != OpTy)
3577 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3579 FullV, LF.OperandValToReplace->getType(),
3580 "tmp", BB->getTerminator());
3582 PN->setIncomingValue(i, FullV);
3583 Pair.first->second = FullV;
3588 /// Rewrite - Emit instructions for the leading candidate expression for this
3589 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3590 /// the newly expanded value.
3591 void LSRInstance::Rewrite(const LSRFixup &LF,
3593 SCEVExpander &Rewriter,
3594 SmallVectorImpl<WeakVH> &DeadInsts,
3596 // First, find an insertion point that dominates UserInst. For PHI nodes,
3597 // find the nearest block which dominates all the relevant uses.
3598 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3599 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3601 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3603 // If this is reuse-by-noop-cast, insert the noop cast.
3604 const Type *OpTy = LF.OperandValToReplace->getType();
3605 if (FullV->getType() != OpTy) {
3607 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3608 FullV, OpTy, "tmp", LF.UserInst);
3612 // Update the user. ICmpZero is handled specially here (for now) because
3613 // Expand may have updated one of the operands of the icmp already, and
3614 // its new value may happen to be equal to LF.OperandValToReplace, in
3615 // which case doing replaceUsesOfWith leads to replacing both operands
3616 // with the same value. TODO: Reorganize this.
3617 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3618 LF.UserInst->setOperand(0, FullV);
3620 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3623 DeadInsts.push_back(LF.OperandValToReplace);
3626 /// ImplementSolution - Rewrite all the fixup locations with new values,
3627 /// following the chosen solution.
3629 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3631 // Keep track of instructions we may have made dead, so that
3632 // we can remove them after we are done working.
3633 SmallVector<WeakVH, 16> DeadInsts;
3635 SCEVExpander Rewriter(SE);
3636 Rewriter.disableCanonicalMode();
3637 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3639 // Expand the new value definitions and update the users.
3640 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3641 E = Fixups.end(); I != E; ++I) {
3642 const LSRFixup &Fixup = *I;
3644 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3649 // Clean up after ourselves. This must be done before deleting any
3653 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3656 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3657 : IU(P->getAnalysis<IVUsers>()),
3658 SE(P->getAnalysis<ScalarEvolution>()),
3659 DT(P->getAnalysis<DominatorTree>()),
3660 LI(P->getAnalysis<LoopInfo>()),
3661 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3663 // If LoopSimplify form is not available, stay out of trouble.
3664 if (!L->isLoopSimplifyForm()) return;
3666 // If there's no interesting work to be done, bail early.
3667 if (IU.empty()) return;
3669 DEBUG(dbgs() << "\nLSR on loop ";
3670 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3673 // First, perform some low-level loop optimizations.
3675 OptimizeLoopTermCond();
3677 // Start collecting data and preparing for the solver.
3678 CollectInterestingTypesAndFactors();
3679 CollectFixupsAndInitialFormulae();
3680 CollectLoopInvariantFixupsAndFormulae();
3682 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3683 print_uses(dbgs()));
3685 // Now use the reuse data to generate a bunch of interesting ways
3686 // to formulate the values needed for the uses.
3687 GenerateAllReuseFormulae();
3689 FilterOutUndesirableDedicatedRegisters();
3690 NarrowSearchSpaceUsingHeuristics();
3692 SmallVector<const Formula *, 8> Solution;
3695 // Release memory that is no longer needed.
3701 // Formulae should be legal.
3702 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3703 E = Uses.end(); I != E; ++I) {
3704 const LSRUse &LU = *I;
3705 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3706 JE = LU.Formulae.end(); J != JE; ++J)
3707 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3708 LU.Kind, LU.AccessTy, TLI) &&
3709 "Illegal formula generated!");
3713 // Now that we've decided what we want, make it so.
3714 ImplementSolution(Solution, P);
3717 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3718 if (Factors.empty() && Types.empty()) return;
3720 OS << "LSR has identified the following interesting factors and types: ";
3723 for (SmallSetVector<int64_t, 8>::const_iterator
3724 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3725 if (!First) OS << ", ";
3730 for (SmallSetVector<const Type *, 4>::const_iterator
3731 I = Types.begin(), E = Types.end(); I != E; ++I) {
3732 if (!First) OS << ", ";
3734 OS << '(' << **I << ')';
3739 void LSRInstance::print_fixups(raw_ostream &OS) const {
3740 OS << "LSR is examining the following fixup sites:\n";
3741 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3742 E = Fixups.end(); I != E; ++I) {
3749 void LSRInstance::print_uses(raw_ostream &OS) const {
3750 OS << "LSR is examining the following uses:\n";
3751 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3752 E = Uses.end(); I != E; ++I) {
3753 const LSRUse &LU = *I;
3757 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3758 JE = LU.Formulae.end(); J != JE; ++J) {
3766 void LSRInstance::print(raw_ostream &OS) const {
3767 print_factors_and_types(OS);
3772 void LSRInstance::dump() const {
3773 print(errs()); errs() << '\n';
3778 class LoopStrengthReduce : public LoopPass {
3779 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3780 /// transformation profitability.
3781 const TargetLowering *const TLI;
3784 static char ID; // Pass ID, replacement for typeid
3785 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3788 bool runOnLoop(Loop *L, LPPassManager &LPM);
3789 void getAnalysisUsage(AnalysisUsage &AU) const;
3794 char LoopStrengthReduce::ID = 0;
3795 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3796 "Loop Strength Reduction", false, false)
3797 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3798 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3799 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3800 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3801 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3802 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3803 "Loop Strength Reduction", false, false)
3806 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3807 return new LoopStrengthReduce(TLI);
3810 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3811 : LoopPass(ID), TLI(tli) {
3812 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3815 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3816 // We split critical edges, so we change the CFG. However, we do update
3817 // many analyses if they are around.
3818 AU.addPreservedID(LoopSimplifyID);
3820 AU.addRequired<LoopInfo>();
3821 AU.addPreserved<LoopInfo>();
3822 AU.addRequiredID(LoopSimplifyID);
3823 AU.addRequired<DominatorTree>();
3824 AU.addPreserved<DominatorTree>();
3825 AU.addRequired<ScalarEvolution>();
3826 AU.addPreserved<ScalarEvolution>();
3827 AU.addRequired<IVUsers>();
3828 AU.addPreserved<IVUsers>();
3831 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3832 bool Changed = false;
3834 // Run the main LSR transformation.
3835 Changed |= LSRInstance(TLI, L, this).getChanged();
3837 // At this point, it is worth checking to see if any recurrence PHIs are also
3838 // dead, so that we can remove them as well.
3839 Changed |= DeleteDeadPHIs(L->getHeader());