1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/ADT/SmallBitVector.h"
69 #include "llvm/ADT/SetVector.h"
70 #include "llvm/ADT/DenseSet.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/ValueHandle.h"
73 #include "llvm/Support/raw_ostream.h"
74 #include "llvm/Target/TargetLowering.h"
80 /// RegSortData - This class holds data which is used to order reuse candidates.
83 /// UsedByIndices - This represents the set of LSRUse indices which reference
84 /// a particular register.
85 SmallBitVector UsedByIndices;
89 void print(raw_ostream &OS) const;
95 void RegSortData::print(raw_ostream &OS) const {
96 OS << "[NumUses=" << UsedByIndices.count() << ']';
99 void RegSortData::dump() const {
100 print(errs()); errs() << '\n';
105 /// RegUseTracker - Map register candidates to information about how they are
107 class RegUseTracker {
108 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
110 RegUsesTy RegUsesMap;
111 SmallVector<const SCEV *, 16> RegSequence;
114 void CountRegister(const SCEV *Reg, size_t LUIdx);
115 void DropRegister(const SCEV *Reg, size_t LUIdx);
116 void DropUse(size_t LUIdx);
118 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
120 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
124 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
125 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
126 iterator begin() { return RegSequence.begin(); }
127 iterator end() { return RegSequence.end(); }
128 const_iterator begin() const { return RegSequence.begin(); }
129 const_iterator end() const { return RegSequence.end(); }
135 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
136 std::pair<RegUsesTy::iterator, bool> Pair =
137 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
138 RegSortData &RSD = Pair.first->second;
140 RegSequence.push_back(Reg);
141 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
142 RSD.UsedByIndices.set(LUIdx);
146 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
147 RegUsesTy::iterator It = RegUsesMap.find(Reg);
148 assert(It != RegUsesMap.end());
149 RegSortData &RSD = It->second;
150 assert(RSD.UsedByIndices.size() > LUIdx);
151 RSD.UsedByIndices.reset(LUIdx);
155 RegUseTracker::DropUse(size_t LUIdx) {
156 // Remove the use index from every register's use list.
157 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
159 I->second.UsedByIndices.reset(LUIdx);
163 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
164 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
165 if (I == RegUsesMap.end())
167 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
168 int i = UsedByIndices.find_first();
169 if (i == -1) return false;
170 if ((size_t)i != LUIdx) return true;
171 return UsedByIndices.find_next(i) != -1;
174 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
175 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
176 assert(I != RegUsesMap.end() && "Unknown register!");
177 return I->second.UsedByIndices;
180 void RegUseTracker::clear() {
187 /// Formula - This class holds information that describes a formula for
188 /// computing satisfying a use. It may include broken-out immediates and scaled
191 /// AM - This is used to represent complex addressing, as well as other kinds
192 /// of interesting uses.
193 TargetLowering::AddrMode AM;
195 /// BaseRegs - The list of "base" registers for this use. When this is
196 /// non-empty, AM.HasBaseReg should be set to true.
197 SmallVector<const SCEV *, 2> BaseRegs;
199 /// ScaledReg - The 'scaled' register for this use. This should be non-null
200 /// when AM.Scale is not zero.
201 const SCEV *ScaledReg;
203 Formula() : ScaledReg(0) {}
205 void InitialMatch(const SCEV *S, Loop *L,
206 ScalarEvolution &SE, DominatorTree &DT);
208 unsigned getNumRegs() const;
209 const Type *getType() const;
211 void DeleteBaseReg(const SCEV *&S);
213 bool referencesReg(const SCEV *S) const;
214 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
215 const RegUseTracker &RegUses) const;
217 void print(raw_ostream &OS) const;
223 /// DoInitialMatch - Recursion helper for InitialMatch.
224 static void DoInitialMatch(const SCEV *S, Loop *L,
225 SmallVectorImpl<const SCEV *> &Good,
226 SmallVectorImpl<const SCEV *> &Bad,
227 ScalarEvolution &SE, DominatorTree &DT) {
228 // Collect expressions which properly dominate the loop header.
229 if (S->properlyDominates(L->getHeader(), &DT)) {
234 // Look at add operands.
235 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
236 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
238 DoInitialMatch(*I, L, Good, Bad, SE, DT);
242 // Look at addrec operands.
243 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
244 if (!AR->getStart()->isZero()) {
245 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
246 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
247 AR->getStepRecurrence(SE),
249 L, Good, Bad, SE, DT);
253 // Handle a multiplication by -1 (negation) if it didn't fold.
254 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
255 if (Mul->getOperand(0)->isAllOnesValue()) {
256 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
257 const SCEV *NewMul = SE.getMulExpr(Ops);
259 SmallVector<const SCEV *, 4> MyGood;
260 SmallVector<const SCEV *, 4> MyBad;
261 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
262 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
263 SE.getEffectiveSCEVType(NewMul->getType())));
264 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
265 E = MyGood.end(); I != E; ++I)
266 Good.push_back(SE.getMulExpr(NegOne, *I));
267 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
268 E = MyBad.end(); I != E; ++I)
269 Bad.push_back(SE.getMulExpr(NegOne, *I));
273 // Ok, we can't do anything interesting. Just stuff the whole thing into a
274 // register and hope for the best.
278 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
279 /// attempting to keep all loop-invariant and loop-computable values in a
280 /// single base register.
281 void Formula::InitialMatch(const SCEV *S, Loop *L,
282 ScalarEvolution &SE, DominatorTree &DT) {
283 SmallVector<const SCEV *, 4> Good;
284 SmallVector<const SCEV *, 4> Bad;
285 DoInitialMatch(S, L, Good, Bad, SE, DT);
287 const SCEV *Sum = SE.getAddExpr(Good);
289 BaseRegs.push_back(Sum);
290 AM.HasBaseReg = true;
293 const SCEV *Sum = SE.getAddExpr(Bad);
295 BaseRegs.push_back(Sum);
296 AM.HasBaseReg = true;
300 /// getNumRegs - Return the total number of register operands used by this
301 /// formula. This does not include register uses implied by non-constant
303 unsigned Formula::getNumRegs() const {
304 return !!ScaledReg + BaseRegs.size();
307 /// getType - Return the type of this formula, if it has one, or null
308 /// otherwise. This type is meaningless except for the bit size.
309 const Type *Formula::getType() const {
310 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
311 ScaledReg ? ScaledReg->getType() :
312 AM.BaseGV ? AM.BaseGV->getType() :
316 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
317 void Formula::DeleteBaseReg(const SCEV *&S) {
318 if (&S != &BaseRegs.back())
319 std::swap(S, BaseRegs.back());
323 /// referencesReg - Test if this formula references the given register.
324 bool Formula::referencesReg(const SCEV *S) const {
325 return S == ScaledReg ||
326 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
329 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
330 /// which are used by uses other than the use with the given index.
331 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
332 const RegUseTracker &RegUses) const {
334 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
336 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
337 E = BaseRegs.end(); I != E; ++I)
338 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
343 void Formula::print(raw_ostream &OS) const {
346 if (!First) OS << " + "; else First = false;
347 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
349 if (AM.BaseOffs != 0) {
350 if (!First) OS << " + "; else First = false;
353 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
354 E = BaseRegs.end(); I != E; ++I) {
355 if (!First) OS << " + "; else First = false;
356 OS << "reg(" << **I << ')';
358 if (AM.HasBaseReg && BaseRegs.empty()) {
359 if (!First) OS << " + "; else First = false;
360 OS << "**error: HasBaseReg**";
361 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
362 if (!First) OS << " + "; else First = false;
363 OS << "**error: !HasBaseReg**";
366 if (!First) OS << " + "; else First = false;
367 OS << AM.Scale << "*reg(";
376 void Formula::dump() const {
377 print(errs()); errs() << '\n';
380 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
381 /// without changing its value.
382 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
384 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
385 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
388 /// isAddSExtable - Return true if the given add can be sign-extended
389 /// without changing its value.
390 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
392 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
393 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
396 /// isMulSExtable - Return true if the given mul can be sign-extended
397 /// without changing its value.
398 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
400 IntegerType::get(SE.getContext(),
401 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
402 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
405 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
406 /// and if the remainder is known to be zero, or null otherwise. If
407 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
408 /// to Y, ignoring that the multiplication may overflow, which is useful when
409 /// the result will be used in a context where the most significant bits are
411 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
413 bool IgnoreSignificantBits = false) {
414 // Handle the trivial case, which works for any SCEV type.
416 return SE.getConstant(LHS->getType(), 1);
418 // Handle a few RHS special cases.
419 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
421 const APInt &RA = RC->getValue()->getValue();
422 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
424 if (RA.isAllOnesValue())
425 return SE.getMulExpr(LHS, RC);
426 // Handle x /s 1 as x.
431 // Check for a division of a constant by a constant.
432 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
435 const APInt &LA = C->getValue()->getValue();
436 const APInt &RA = RC->getValue()->getValue();
437 if (LA.srem(RA) != 0)
439 return SE.getConstant(LA.sdiv(RA));
442 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
443 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
444 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
445 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
446 IgnoreSignificantBits);
448 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
449 IgnoreSignificantBits);
450 if (!Start) return 0;
451 return SE.getAddRecExpr(Start, Step, AR->getLoop());
456 // Distribute the sdiv over add operands, if the add doesn't overflow.
457 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
458 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
459 SmallVector<const SCEV *, 8> Ops;
460 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
462 const SCEV *Op = getExactSDiv(*I, RHS, SE,
463 IgnoreSignificantBits);
467 return SE.getAddExpr(Ops);
472 // Check for a multiply operand that we can pull RHS out of.
473 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
474 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
475 SmallVector<const SCEV *, 4> Ops;
477 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
481 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
482 IgnoreSignificantBits)) {
488 return Found ? SE.getMulExpr(Ops) : 0;
493 // Otherwise we don't know.
497 /// ExtractImmediate - If S involves the addition of a constant integer value,
498 /// return that integer value, and mutate S to point to a new SCEV with that
500 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
501 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
502 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
503 S = SE.getConstant(C->getType(), 0);
504 return C->getValue()->getSExtValue();
506 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
507 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
508 int64_t Result = ExtractImmediate(NewOps.front(), SE);
510 S = SE.getAddExpr(NewOps);
512 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
513 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
514 int64_t Result = ExtractImmediate(NewOps.front(), SE);
516 S = SE.getAddRecExpr(NewOps, AR->getLoop());
522 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
523 /// return that symbol, and mutate S to point to a new SCEV with that
525 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
526 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
527 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
528 S = SE.getConstant(GV->getType(), 0);
531 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
532 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
533 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
535 S = SE.getAddExpr(NewOps);
537 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
538 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
539 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
541 S = SE.getAddRecExpr(NewOps, AR->getLoop());
547 /// isAddressUse - Returns true if the specified instruction is using the
548 /// specified value as an address.
549 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
550 bool isAddress = isa<LoadInst>(Inst);
551 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
552 if (SI->getOperand(1) == OperandVal)
554 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
555 // Addressing modes can also be folded into prefetches and a variety
557 switch (II->getIntrinsicID()) {
559 case Intrinsic::prefetch:
560 case Intrinsic::x86_sse2_loadu_dq:
561 case Intrinsic::x86_sse2_loadu_pd:
562 case Intrinsic::x86_sse_loadu_ps:
563 case Intrinsic::x86_sse_storeu_ps:
564 case Intrinsic::x86_sse2_storeu_pd:
565 case Intrinsic::x86_sse2_storeu_dq:
566 case Intrinsic::x86_sse2_storel_dq:
567 if (II->getArgOperand(0) == OperandVal)
575 /// getAccessType - Return the type of the memory being accessed.
576 static const Type *getAccessType(const Instruction *Inst) {
577 const Type *AccessTy = Inst->getType();
578 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
579 AccessTy = SI->getOperand(0)->getType();
580 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
581 // Addressing modes can also be folded into prefetches and a variety
583 switch (II->getIntrinsicID()) {
585 case Intrinsic::x86_sse_storeu_ps:
586 case Intrinsic::x86_sse2_storeu_pd:
587 case Intrinsic::x86_sse2_storeu_dq:
588 case Intrinsic::x86_sse2_storel_dq:
589 AccessTy = II->getArgOperand(0)->getType();
594 // All pointers have the same requirements, so canonicalize them to an
595 // arbitrary pointer type to minimize variation.
596 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
597 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
598 PTy->getAddressSpace());
603 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
604 /// specified set are trivially dead, delete them and see if this makes any of
605 /// their operands subsequently dead.
607 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
608 bool Changed = false;
610 while (!DeadInsts.empty()) {
611 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
613 if (I == 0 || !isInstructionTriviallyDead(I))
616 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
617 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
620 DeadInsts.push_back(U);
623 I->eraseFromParent();
632 /// Cost - This class is used to measure and compare candidate formulae.
634 /// TODO: Some of these could be merged. Also, a lexical ordering
635 /// isn't always optimal.
639 unsigned NumBaseAdds;
645 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
648 unsigned getNumRegs() const { return NumRegs; }
650 bool operator<(const Cost &Other) const;
654 void RateFormula(const Formula &F,
655 SmallPtrSet<const SCEV *, 16> &Regs,
656 const DenseSet<const SCEV *> &VisitedRegs,
658 const SmallVectorImpl<int64_t> &Offsets,
659 ScalarEvolution &SE, DominatorTree &DT);
661 void print(raw_ostream &OS) const;
665 void RateRegister(const SCEV *Reg,
666 SmallPtrSet<const SCEV *, 16> &Regs,
668 ScalarEvolution &SE, DominatorTree &DT);
669 void RatePrimaryRegister(const SCEV *Reg,
670 SmallPtrSet<const SCEV *, 16> &Regs,
672 ScalarEvolution &SE, DominatorTree &DT);
677 /// RateRegister - Tally up interesting quantities from the given register.
678 void Cost::RateRegister(const SCEV *Reg,
679 SmallPtrSet<const SCEV *, 16> &Regs,
681 ScalarEvolution &SE, DominatorTree &DT) {
682 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
683 if (AR->getLoop() == L)
684 AddRecCost += 1; /// TODO: This should be a function of the stride.
686 // If this is an addrec for a loop that's already been visited by LSR,
687 // don't second-guess its addrec phi nodes. LSR isn't currently smart
688 // enough to reason about more than one loop at a time. Consider these
689 // registers free and leave them alone.
690 else if (L->contains(AR->getLoop()) ||
691 (!AR->getLoop()->contains(L) &&
692 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
693 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
694 PHINode *PN = dyn_cast<PHINode>(I); ++I)
695 if (SE.isSCEVable(PN->getType()) &&
696 (SE.getEffectiveSCEVType(PN->getType()) ==
697 SE.getEffectiveSCEVType(AR->getType())) &&
698 SE.getSCEV(PN) == AR)
701 // If this isn't one of the addrecs that the loop already has, it
702 // would require a costly new phi and add. TODO: This isn't
703 // precisely modeled right now.
705 if (!Regs.count(AR->getStart()))
706 RateRegister(AR->getStart(), Regs, L, SE, DT);
709 // Add the step value register, if it needs one.
710 // TODO: The non-affine case isn't precisely modeled here.
711 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
712 if (!Regs.count(AR->getStart()))
713 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
717 // Rough heuristic; favor registers which don't require extra setup
718 // instructions in the preheader.
719 if (!isa<SCEVUnknown>(Reg) &&
720 !isa<SCEVConstant>(Reg) &&
721 !(isa<SCEVAddRecExpr>(Reg) &&
722 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
723 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
727 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
729 void Cost::RatePrimaryRegister(const SCEV *Reg,
730 SmallPtrSet<const SCEV *, 16> &Regs,
732 ScalarEvolution &SE, DominatorTree &DT) {
733 if (Regs.insert(Reg))
734 RateRegister(Reg, Regs, L, SE, DT);
737 void Cost::RateFormula(const Formula &F,
738 SmallPtrSet<const SCEV *, 16> &Regs,
739 const DenseSet<const SCEV *> &VisitedRegs,
741 const SmallVectorImpl<int64_t> &Offsets,
742 ScalarEvolution &SE, DominatorTree &DT) {
743 // Tally up the registers.
744 if (const SCEV *ScaledReg = F.ScaledReg) {
745 if (VisitedRegs.count(ScaledReg)) {
749 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
751 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
752 E = F.BaseRegs.end(); I != E; ++I) {
753 const SCEV *BaseReg = *I;
754 if (VisitedRegs.count(BaseReg)) {
758 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
760 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
761 BaseReg->hasComputableLoopEvolution(L);
764 if (F.BaseRegs.size() > 1)
765 NumBaseAdds += F.BaseRegs.size() - 1;
767 // Tally up the non-zero immediates.
768 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
769 E = Offsets.end(); I != E; ++I) {
770 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
772 ImmCost += 64; // Handle symbolic values conservatively.
773 // TODO: This should probably be the pointer size.
774 else if (Offset != 0)
775 ImmCost += APInt(64, Offset, true).getMinSignedBits();
779 /// Loose - Set this cost to a loosing value.
789 /// operator< - Choose the lower cost.
790 bool Cost::operator<(const Cost &Other) const {
791 if (NumRegs != Other.NumRegs)
792 return NumRegs < Other.NumRegs;
793 if (AddRecCost != Other.AddRecCost)
794 return AddRecCost < Other.AddRecCost;
795 if (NumIVMuls != Other.NumIVMuls)
796 return NumIVMuls < Other.NumIVMuls;
797 if (NumBaseAdds != Other.NumBaseAdds)
798 return NumBaseAdds < Other.NumBaseAdds;
799 if (ImmCost != Other.ImmCost)
800 return ImmCost < Other.ImmCost;
801 if (SetupCost != Other.SetupCost)
802 return SetupCost < Other.SetupCost;
806 void Cost::print(raw_ostream &OS) const {
807 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
809 OS << ", with addrec cost " << AddRecCost;
811 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
812 if (NumBaseAdds != 0)
813 OS << ", plus " << NumBaseAdds << " base add"
814 << (NumBaseAdds == 1 ? "" : "s");
816 OS << ", plus " << ImmCost << " imm cost";
818 OS << ", plus " << SetupCost << " setup cost";
821 void Cost::dump() const {
822 print(errs()); errs() << '\n';
827 /// LSRFixup - An operand value in an instruction which is to be replaced
828 /// with some equivalent, possibly strength-reduced, replacement.
830 /// UserInst - The instruction which will be updated.
831 Instruction *UserInst;
833 /// OperandValToReplace - The operand of the instruction which will
834 /// be replaced. The operand may be used more than once; every instance
835 /// will be replaced.
836 Value *OperandValToReplace;
838 /// PostIncLoops - If this user is to use the post-incremented value of an
839 /// induction variable, this variable is non-null and holds the loop
840 /// associated with the induction variable.
841 PostIncLoopSet PostIncLoops;
843 /// LUIdx - The index of the LSRUse describing the expression which
844 /// this fixup needs, minus an offset (below).
847 /// Offset - A constant offset to be added to the LSRUse expression.
848 /// This allows multiple fixups to share the same LSRUse with different
849 /// offsets, for example in an unrolled loop.
852 bool isUseFullyOutsideLoop(const Loop *L) const;
856 void print(raw_ostream &OS) const;
863 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
865 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
866 /// value outside of the given loop.
867 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
868 // PHI nodes use their value in their incoming blocks.
869 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
870 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
871 if (PN->getIncomingValue(i) == OperandValToReplace &&
872 L->contains(PN->getIncomingBlock(i)))
877 return !L->contains(UserInst);
880 void LSRFixup::print(raw_ostream &OS) const {
882 // Store is common and interesting enough to be worth special-casing.
883 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
885 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
886 } else if (UserInst->getType()->isVoidTy())
887 OS << UserInst->getOpcodeName();
889 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
891 OS << ", OperandValToReplace=";
892 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
894 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
895 E = PostIncLoops.end(); I != E; ++I) {
896 OS << ", PostIncLoop=";
897 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
900 if (LUIdx != ~size_t(0))
901 OS << ", LUIdx=" << LUIdx;
904 OS << ", Offset=" << Offset;
907 void LSRFixup::dump() const {
908 print(errs()); errs() << '\n';
913 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
914 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
915 struct UniquifierDenseMapInfo {
916 static SmallVector<const SCEV *, 2> getEmptyKey() {
917 SmallVector<const SCEV *, 2> V;
918 V.push_back(reinterpret_cast<const SCEV *>(-1));
922 static SmallVector<const SCEV *, 2> getTombstoneKey() {
923 SmallVector<const SCEV *, 2> V;
924 V.push_back(reinterpret_cast<const SCEV *>(-2));
928 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
930 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
931 E = V.end(); I != E; ++I)
932 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
936 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
937 const SmallVector<const SCEV *, 2> &RHS) {
942 /// LSRUse - This class holds the state that LSR keeps for each use in
943 /// IVUsers, as well as uses invented by LSR itself. It includes information
944 /// about what kinds of things can be folded into the user, information about
945 /// the user itself, and information about how the use may be satisfied.
946 /// TODO: Represent multiple users of the same expression in common?
948 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
951 /// KindType - An enum for a kind of use, indicating what types of
952 /// scaled and immediate operands it might support.
954 Basic, ///< A normal use, with no folding.
955 Special, ///< A special case of basic, allowing -1 scales.
956 Address, ///< An address use; folding according to TargetLowering
957 ICmpZero ///< An equality icmp with both operands folded into one.
958 // TODO: Add a generic icmp too?
962 const Type *AccessTy;
964 SmallVector<int64_t, 8> Offsets;
968 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
969 /// LSRUse are outside of the loop, in which case some special-case heuristics
971 bool AllFixupsOutsideLoop;
973 /// WidestFixupType - This records the widest use type for any fixup using
974 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
975 /// max fixup widths to be equivalent, because the narrower one may be relying
976 /// on the implicit truncation to truncate away bogus bits.
977 const Type *WidestFixupType;
979 /// Formulae - A list of ways to build a value that can satisfy this user.
980 /// After the list is populated, one of these is selected heuristically and
981 /// used to formulate a replacement for OperandValToReplace in UserInst.
982 SmallVector<Formula, 12> Formulae;
984 /// Regs - The set of register candidates used by all formulae in this LSRUse.
985 SmallPtrSet<const SCEV *, 4> Regs;
987 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
988 MinOffset(INT64_MAX),
989 MaxOffset(INT64_MIN),
990 AllFixupsOutsideLoop(true),
991 WidestFixupType(0) {}
993 bool HasFormulaWithSameRegs(const Formula &F) const;
994 bool InsertFormula(const Formula &F);
995 void DeleteFormula(Formula &F);
996 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
998 void print(raw_ostream &OS) const;
1004 /// HasFormula - Test whether this use as a formula which has the same
1005 /// registers as the given formula.
1006 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1007 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1008 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1009 // Unstable sort by host order ok, because this is only used for uniquifying.
1010 std::sort(Key.begin(), Key.end());
1011 return Uniquifier.count(Key);
1014 /// InsertFormula - If the given formula has not yet been inserted, add it to
1015 /// the list, and return true. Return false otherwise.
1016 bool LSRUse::InsertFormula(const Formula &F) {
1017 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1018 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1019 // Unstable sort by host order ok, because this is only used for uniquifying.
1020 std::sort(Key.begin(), Key.end());
1022 if (!Uniquifier.insert(Key).second)
1025 // Using a register to hold the value of 0 is not profitable.
1026 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1027 "Zero allocated in a scaled register!");
1029 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1030 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1031 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1034 // Add the formula to the list.
1035 Formulae.push_back(F);
1037 // Record registers now being used by this use.
1038 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1039 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1044 /// DeleteFormula - Remove the given formula from this use's list.
1045 void LSRUse::DeleteFormula(Formula &F) {
1046 if (&F != &Formulae.back())
1047 std::swap(F, Formulae.back());
1048 Formulae.pop_back();
1049 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1052 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1053 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1054 // Now that we've filtered out some formulae, recompute the Regs set.
1055 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1057 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1058 E = Formulae.end(); I != E; ++I) {
1059 const Formula &F = *I;
1060 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1061 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1064 // Update the RegTracker.
1065 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1066 E = OldRegs.end(); I != E; ++I)
1067 if (!Regs.count(*I))
1068 RegUses.DropRegister(*I, LUIdx);
1071 void LSRUse::print(raw_ostream &OS) const {
1072 OS << "LSR Use: Kind=";
1074 case Basic: OS << "Basic"; break;
1075 case Special: OS << "Special"; break;
1076 case ICmpZero: OS << "ICmpZero"; break;
1078 OS << "Address of ";
1079 if (AccessTy->isPointerTy())
1080 OS << "pointer"; // the full pointer type could be really verbose
1085 OS << ", Offsets={";
1086 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1087 E = Offsets.end(); I != E; ++I) {
1089 if (llvm::next(I) != E)
1094 if (AllFixupsOutsideLoop)
1095 OS << ", all-fixups-outside-loop";
1097 if (WidestFixupType)
1098 OS << ", widest fixup type: " << *WidestFixupType;
1101 void LSRUse::dump() const {
1102 print(errs()); errs() << '\n';
1105 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1106 /// be completely folded into the user instruction at isel time. This includes
1107 /// address-mode folding and special icmp tricks.
1108 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1109 LSRUse::KindType Kind, const Type *AccessTy,
1110 const TargetLowering *TLI) {
1112 case LSRUse::Address:
1113 // If we have low-level target information, ask the target if it can
1114 // completely fold this address.
1115 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1117 // Otherwise, just guess that reg+reg addressing is legal.
1118 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1120 case LSRUse::ICmpZero:
1121 // There's not even a target hook for querying whether it would be legal to
1122 // fold a GV into an ICmp.
1126 // ICmp only has two operands; don't allow more than two non-trivial parts.
1127 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1130 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1131 // putting the scaled register in the other operand of the icmp.
1132 if (AM.Scale != 0 && AM.Scale != -1)
1135 // If we have low-level target information, ask the target if it can fold an
1136 // integer immediate on an icmp.
1137 if (AM.BaseOffs != 0) {
1138 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1145 // Only handle single-register values.
1146 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1148 case LSRUse::Special:
1149 // Only handle -1 scales, or no scale.
1150 return AM.Scale == 0 || AM.Scale == -1;
1156 static bool isLegalUse(TargetLowering::AddrMode AM,
1157 int64_t MinOffset, int64_t MaxOffset,
1158 LSRUse::KindType Kind, const Type *AccessTy,
1159 const TargetLowering *TLI) {
1160 // Check for overflow.
1161 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1164 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1165 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1166 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1167 // Check for overflow.
1168 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1171 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1172 return isLegalUse(AM, Kind, AccessTy, TLI);
1177 static bool isAlwaysFoldable(int64_t BaseOffs,
1178 GlobalValue *BaseGV,
1180 LSRUse::KindType Kind, const Type *AccessTy,
1181 const TargetLowering *TLI) {
1182 // Fast-path: zero is always foldable.
1183 if (BaseOffs == 0 && !BaseGV) return true;
1185 // Conservatively, create an address with an immediate and a
1186 // base and a scale.
1187 TargetLowering::AddrMode AM;
1188 AM.BaseOffs = BaseOffs;
1190 AM.HasBaseReg = HasBaseReg;
1191 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1193 // Canonicalize a scale of 1 to a base register if the formula doesn't
1194 // already have a base register.
1195 if (!AM.HasBaseReg && AM.Scale == 1) {
1197 AM.HasBaseReg = true;
1200 return isLegalUse(AM, Kind, AccessTy, TLI);
1203 static bool isAlwaysFoldable(const SCEV *S,
1204 int64_t MinOffset, int64_t MaxOffset,
1206 LSRUse::KindType Kind, const Type *AccessTy,
1207 const TargetLowering *TLI,
1208 ScalarEvolution &SE) {
1209 // Fast-path: zero is always foldable.
1210 if (S->isZero()) return true;
1212 // Conservatively, create an address with an immediate and a
1213 // base and a scale.
1214 int64_t BaseOffs = ExtractImmediate(S, SE);
1215 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1217 // If there's anything else involved, it's not foldable.
1218 if (!S->isZero()) return false;
1220 // Fast-path: zero is always foldable.
1221 if (BaseOffs == 0 && !BaseGV) return true;
1223 // Conservatively, create an address with an immediate and a
1224 // base and a scale.
1225 TargetLowering::AddrMode AM;
1226 AM.BaseOffs = BaseOffs;
1228 AM.HasBaseReg = HasBaseReg;
1229 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1231 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1236 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1237 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1238 struct UseMapDenseMapInfo {
1239 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1240 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1243 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1244 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1248 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1249 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1250 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1254 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1255 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1260 /// FormulaSorter - This class implements an ordering for formulae which sorts
1261 /// the by their standalone cost.
1262 class FormulaSorter {
1263 /// These two sets are kept empty, so that we compute standalone costs.
1264 DenseSet<const SCEV *> VisitedRegs;
1265 SmallPtrSet<const SCEV *, 16> Regs;
1268 ScalarEvolution &SE;
1272 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1273 : L(l), LU(&lu), SE(se), DT(dt) {}
1275 bool operator()(const Formula &A, const Formula &B) {
1277 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1280 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1282 return CostA < CostB;
1286 /// LSRInstance - This class holds state for the main loop strength reduction
1290 ScalarEvolution &SE;
1293 const TargetLowering *const TLI;
1297 /// IVIncInsertPos - This is the insert position that the current loop's
1298 /// induction variable increment should be placed. In simple loops, this is
1299 /// the latch block's terminator. But in more complicated cases, this is a
1300 /// position which will dominate all the in-loop post-increment users.
1301 Instruction *IVIncInsertPos;
1303 /// Factors - Interesting factors between use strides.
1304 SmallSetVector<int64_t, 8> Factors;
1306 /// Types - Interesting use types, to facilitate truncation reuse.
1307 SmallSetVector<const Type *, 4> Types;
1309 /// Fixups - The list of operands which are to be replaced.
1310 SmallVector<LSRFixup, 16> Fixups;
1312 /// Uses - The list of interesting uses.
1313 SmallVector<LSRUse, 16> Uses;
1315 /// RegUses - Track which uses use which register candidates.
1316 RegUseTracker RegUses;
1318 void OptimizeShadowIV();
1319 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1320 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1321 void OptimizeLoopTermCond();
1323 void CollectInterestingTypesAndFactors();
1324 void CollectFixupsAndInitialFormulae();
1326 LSRFixup &getNewFixup() {
1327 Fixups.push_back(LSRFixup());
1328 return Fixups.back();
1331 // Support for sharing of LSRUses between LSRFixups.
1332 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1334 UseMapDenseMapInfo> UseMapTy;
1337 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1338 LSRUse::KindType Kind, const Type *AccessTy);
1340 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1341 LSRUse::KindType Kind,
1342 const Type *AccessTy);
1344 void DeleteUse(LSRUse &LU);
1346 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1349 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1350 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1351 void CountRegisters(const Formula &F, size_t LUIdx);
1352 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1354 void CollectLoopInvariantFixupsAndFormulae();
1356 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1357 unsigned Depth = 0);
1358 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1359 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1360 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1361 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1362 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1363 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1364 void GenerateCrossUseConstantOffsets();
1365 void GenerateAllReuseFormulae();
1367 void FilterOutUndesirableDedicatedRegisters();
1369 size_t EstimateSearchSpaceComplexity() const;
1370 void NarrowSearchSpaceByDetectingSupersets();
1371 void NarrowSearchSpaceByCollapsingUnrolledCode();
1372 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1373 void NarrowSearchSpaceByPickingWinnerRegs();
1374 void NarrowSearchSpaceUsingHeuristics();
1376 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1378 SmallVectorImpl<const Formula *> &Workspace,
1379 const Cost &CurCost,
1380 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1381 DenseSet<const SCEV *> &VisitedRegs) const;
1382 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1384 BasicBlock::iterator
1385 HoistInsertPosition(BasicBlock::iterator IP,
1386 const SmallVectorImpl<Instruction *> &Inputs) const;
1387 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1389 const LSRUse &LU) const;
1391 Value *Expand(const LSRFixup &LF,
1393 BasicBlock::iterator IP,
1394 SCEVExpander &Rewriter,
1395 SmallVectorImpl<WeakVH> &DeadInsts) const;
1396 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1398 SCEVExpander &Rewriter,
1399 SmallVectorImpl<WeakVH> &DeadInsts,
1401 void Rewrite(const LSRFixup &LF,
1403 SCEVExpander &Rewriter,
1404 SmallVectorImpl<WeakVH> &DeadInsts,
1406 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1409 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1411 bool getChanged() const { return Changed; }
1413 void print_factors_and_types(raw_ostream &OS) const;
1414 void print_fixups(raw_ostream &OS) const;
1415 void print_uses(raw_ostream &OS) const;
1416 void print(raw_ostream &OS) const;
1422 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1423 /// inside the loop then try to eliminate the cast operation.
1424 void LSRInstance::OptimizeShadowIV() {
1425 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1426 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1429 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1430 UI != E; /* empty */) {
1431 IVUsers::const_iterator CandidateUI = UI;
1433 Instruction *ShadowUse = CandidateUI->getUser();
1434 const Type *DestTy = NULL;
1436 /* If shadow use is a int->float cast then insert a second IV
1437 to eliminate this cast.
1439 for (unsigned i = 0; i < n; ++i)
1445 for (unsigned i = 0; i < n; ++i, ++d)
1448 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1449 DestTy = UCast->getDestTy();
1450 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1451 DestTy = SCast->getDestTy();
1452 if (!DestTy) continue;
1455 // If target does not support DestTy natively then do not apply
1456 // this transformation.
1457 EVT DVT = TLI->getValueType(DestTy);
1458 if (!TLI->isTypeLegal(DVT)) continue;
1461 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1463 if (PH->getNumIncomingValues() != 2) continue;
1465 const Type *SrcTy = PH->getType();
1466 int Mantissa = DestTy->getFPMantissaWidth();
1467 if (Mantissa == -1) continue;
1468 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1471 unsigned Entry, Latch;
1472 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1480 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1481 if (!Init) continue;
1482 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1484 BinaryOperator *Incr =
1485 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1486 if (!Incr) continue;
1487 if (Incr->getOpcode() != Instruction::Add
1488 && Incr->getOpcode() != Instruction::Sub)
1491 /* Initialize new IV, double d = 0.0 in above example. */
1492 ConstantInt *C = NULL;
1493 if (Incr->getOperand(0) == PH)
1494 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1495 else if (Incr->getOperand(1) == PH)
1496 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1502 // Ignore negative constants, as the code below doesn't handle them
1503 // correctly. TODO: Remove this restriction.
1504 if (!C->getValue().isStrictlyPositive()) continue;
1506 /* Add new PHINode. */
1507 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1509 /* create new increment. '++d' in above example. */
1510 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1511 BinaryOperator *NewIncr =
1512 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1513 Instruction::FAdd : Instruction::FSub,
1514 NewPH, CFP, "IV.S.next.", Incr);
1516 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1517 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1519 /* Remove cast operation */
1520 ShadowUse->replaceAllUsesWith(NewPH);
1521 ShadowUse->eraseFromParent();
1527 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1528 /// set the IV user and stride information and return true, otherwise return
1530 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1531 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1532 if (UI->getUser() == Cond) {
1533 // NOTE: we could handle setcc instructions with multiple uses here, but
1534 // InstCombine does it as well for simple uses, it's not clear that it
1535 // occurs enough in real life to handle.
1542 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1543 /// a max computation.
1545 /// This is a narrow solution to a specific, but acute, problem. For loops
1551 /// } while (++i < n);
1553 /// the trip count isn't just 'n', because 'n' might not be positive. And
1554 /// unfortunately this can come up even for loops where the user didn't use
1555 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1556 /// will commonly be lowered like this:
1562 /// } while (++i < n);
1565 /// and then it's possible for subsequent optimization to obscure the if
1566 /// test in such a way that indvars can't find it.
1568 /// When indvars can't find the if test in loops like this, it creates a
1569 /// max expression, which allows it to give the loop a canonical
1570 /// induction variable:
1573 /// max = n < 1 ? 1 : n;
1576 /// } while (++i != max);
1578 /// Canonical induction variables are necessary because the loop passes
1579 /// are designed around them. The most obvious example of this is the
1580 /// LoopInfo analysis, which doesn't remember trip count values. It
1581 /// expects to be able to rediscover the trip count each time it is
1582 /// needed, and it does this using a simple analysis that only succeeds if
1583 /// the loop has a canonical induction variable.
1585 /// However, when it comes time to generate code, the maximum operation
1586 /// can be quite costly, especially if it's inside of an outer loop.
1588 /// This function solves this problem by detecting this type of loop and
1589 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1590 /// the instructions for the maximum computation.
1592 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1593 // Check that the loop matches the pattern we're looking for.
1594 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1595 Cond->getPredicate() != CmpInst::ICMP_NE)
1598 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1599 if (!Sel || !Sel->hasOneUse()) return Cond;
1601 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1602 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1604 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1606 // Add one to the backedge-taken count to get the trip count.
1607 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1608 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1610 // Check for a max calculation that matches the pattern. There's no check
1611 // for ICMP_ULE here because the comparison would be with zero, which
1612 // isn't interesting.
1613 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1614 const SCEVNAryExpr *Max = 0;
1615 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1616 Pred = ICmpInst::ICMP_SLE;
1618 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1619 Pred = ICmpInst::ICMP_SLT;
1621 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1622 Pred = ICmpInst::ICMP_ULT;
1629 // To handle a max with more than two operands, this optimization would
1630 // require additional checking and setup.
1631 if (Max->getNumOperands() != 2)
1634 const SCEV *MaxLHS = Max->getOperand(0);
1635 const SCEV *MaxRHS = Max->getOperand(1);
1637 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1638 // for a comparison with 1. For <= and >=, a comparison with zero.
1640 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1643 // Check the relevant induction variable for conformance to
1645 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1646 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1647 if (!AR || !AR->isAffine() ||
1648 AR->getStart() != One ||
1649 AR->getStepRecurrence(SE) != One)
1652 assert(AR->getLoop() == L &&
1653 "Loop condition operand is an addrec in a different loop!");
1655 // Check the right operand of the select, and remember it, as it will
1656 // be used in the new comparison instruction.
1658 if (ICmpInst::isTrueWhenEqual(Pred)) {
1659 // Look for n+1, and grab n.
1660 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1661 if (isa<ConstantInt>(BO->getOperand(1)) &&
1662 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1663 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1664 NewRHS = BO->getOperand(0);
1665 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1666 if (isa<ConstantInt>(BO->getOperand(1)) &&
1667 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1668 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1669 NewRHS = BO->getOperand(0);
1672 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1673 NewRHS = Sel->getOperand(1);
1674 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1675 NewRHS = Sel->getOperand(2);
1676 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1677 NewRHS = SU->getValue();
1679 // Max doesn't match expected pattern.
1682 // Determine the new comparison opcode. It may be signed or unsigned,
1683 // and the original comparison may be either equality or inequality.
1684 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1685 Pred = CmpInst::getInversePredicate(Pred);
1687 // Ok, everything looks ok to change the condition into an SLT or SGE and
1688 // delete the max calculation.
1690 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1692 // Delete the max calculation instructions.
1693 Cond->replaceAllUsesWith(NewCond);
1694 CondUse->setUser(NewCond);
1695 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1696 Cond->eraseFromParent();
1697 Sel->eraseFromParent();
1698 if (Cmp->use_empty())
1699 Cmp->eraseFromParent();
1703 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1704 /// postinc iv when possible.
1706 LSRInstance::OptimizeLoopTermCond() {
1707 SmallPtrSet<Instruction *, 4> PostIncs;
1709 BasicBlock *LatchBlock = L->getLoopLatch();
1710 SmallVector<BasicBlock*, 8> ExitingBlocks;
1711 L->getExitingBlocks(ExitingBlocks);
1713 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1714 BasicBlock *ExitingBlock = ExitingBlocks[i];
1716 // Get the terminating condition for the loop if possible. If we
1717 // can, we want to change it to use a post-incremented version of its
1718 // induction variable, to allow coalescing the live ranges for the IV into
1719 // one register value.
1721 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1724 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1725 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1728 // Search IVUsesByStride to find Cond's IVUse if there is one.
1729 IVStrideUse *CondUse = 0;
1730 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1731 if (!FindIVUserForCond(Cond, CondUse))
1734 // If the trip count is computed in terms of a max (due to ScalarEvolution
1735 // being unable to find a sufficient guard, for example), change the loop
1736 // comparison to use SLT or ULT instead of NE.
1737 // One consequence of doing this now is that it disrupts the count-down
1738 // optimization. That's not always a bad thing though, because in such
1739 // cases it may still be worthwhile to avoid a max.
1740 Cond = OptimizeMax(Cond, CondUse);
1742 // If this exiting block dominates the latch block, it may also use
1743 // the post-inc value if it won't be shared with other uses.
1744 // Check for dominance.
1745 if (!DT.dominates(ExitingBlock, LatchBlock))
1748 // Conservatively avoid trying to use the post-inc value in non-latch
1749 // exits if there may be pre-inc users in intervening blocks.
1750 if (LatchBlock != ExitingBlock)
1751 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1752 // Test if the use is reachable from the exiting block. This dominator
1753 // query is a conservative approximation of reachability.
1754 if (&*UI != CondUse &&
1755 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1756 // Conservatively assume there may be reuse if the quotient of their
1757 // strides could be a legal scale.
1758 const SCEV *A = IU.getStride(*CondUse, L);
1759 const SCEV *B = IU.getStride(*UI, L);
1760 if (!A || !B) continue;
1761 if (SE.getTypeSizeInBits(A->getType()) !=
1762 SE.getTypeSizeInBits(B->getType())) {
1763 if (SE.getTypeSizeInBits(A->getType()) >
1764 SE.getTypeSizeInBits(B->getType()))
1765 B = SE.getSignExtendExpr(B, A->getType());
1767 A = SE.getSignExtendExpr(A, B->getType());
1769 if (const SCEVConstant *D =
1770 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1771 const ConstantInt *C = D->getValue();
1772 // Stride of one or negative one can have reuse with non-addresses.
1773 if (C->isOne() || C->isAllOnesValue())
1774 goto decline_post_inc;
1775 // Avoid weird situations.
1776 if (C->getValue().getMinSignedBits() >= 64 ||
1777 C->getValue().isMinSignedValue())
1778 goto decline_post_inc;
1779 // Without TLI, assume that any stride might be valid, and so any
1780 // use might be shared.
1782 goto decline_post_inc;
1783 // Check for possible scaled-address reuse.
1784 const Type *AccessTy = getAccessType(UI->getUser());
1785 TargetLowering::AddrMode AM;
1786 AM.Scale = C->getSExtValue();
1787 if (TLI->isLegalAddressingMode(AM, AccessTy))
1788 goto decline_post_inc;
1789 AM.Scale = -AM.Scale;
1790 if (TLI->isLegalAddressingMode(AM, AccessTy))
1791 goto decline_post_inc;
1795 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1798 // It's possible for the setcc instruction to be anywhere in the loop, and
1799 // possible for it to have multiple users. If it is not immediately before
1800 // the exiting block branch, move it.
1801 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1802 if (Cond->hasOneUse()) {
1803 Cond->moveBefore(TermBr);
1805 // Clone the terminating condition and insert into the loopend.
1806 ICmpInst *OldCond = Cond;
1807 Cond = cast<ICmpInst>(Cond->clone());
1808 Cond->setName(L->getHeader()->getName() + ".termcond");
1809 ExitingBlock->getInstList().insert(TermBr, Cond);
1811 // Clone the IVUse, as the old use still exists!
1812 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1813 TermBr->replaceUsesOfWith(OldCond, Cond);
1817 // If we get to here, we know that we can transform the setcc instruction to
1818 // use the post-incremented version of the IV, allowing us to coalesce the
1819 // live ranges for the IV correctly.
1820 CondUse->transformToPostInc(L);
1823 PostIncs.insert(Cond);
1827 // Determine an insertion point for the loop induction variable increment. It
1828 // must dominate all the post-inc comparisons we just set up, and it must
1829 // dominate the loop latch edge.
1830 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1831 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1832 E = PostIncs.end(); I != E; ++I) {
1834 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1836 if (BB == (*I)->getParent())
1837 IVIncInsertPos = *I;
1838 else if (BB != IVIncInsertPos->getParent())
1839 IVIncInsertPos = BB->getTerminator();
1843 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1844 /// at the given offset and other details. If so, update the use and
1847 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1848 LSRUse::KindType Kind, const Type *AccessTy) {
1849 int64_t NewMinOffset = LU.MinOffset;
1850 int64_t NewMaxOffset = LU.MaxOffset;
1851 const Type *NewAccessTy = AccessTy;
1853 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1854 // something conservative, however this can pessimize in the case that one of
1855 // the uses will have all its uses outside the loop, for example.
1856 if (LU.Kind != Kind)
1858 // Conservatively assume HasBaseReg is true for now.
1859 if (NewOffset < LU.MinOffset) {
1860 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1861 Kind, AccessTy, TLI))
1863 NewMinOffset = NewOffset;
1864 } else if (NewOffset > LU.MaxOffset) {
1865 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1866 Kind, AccessTy, TLI))
1868 NewMaxOffset = NewOffset;
1870 // Check for a mismatched access type, and fall back conservatively as needed.
1871 // TODO: Be less conservative when the type is similar and can use the same
1872 // addressing modes.
1873 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1874 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1877 LU.MinOffset = NewMinOffset;
1878 LU.MaxOffset = NewMaxOffset;
1879 LU.AccessTy = NewAccessTy;
1880 if (NewOffset != LU.Offsets.back())
1881 LU.Offsets.push_back(NewOffset);
1885 /// getUse - Return an LSRUse index and an offset value for a fixup which
1886 /// needs the given expression, with the given kind and optional access type.
1887 /// Either reuse an existing use or create a new one, as needed.
1888 std::pair<size_t, int64_t>
1889 LSRInstance::getUse(const SCEV *&Expr,
1890 LSRUse::KindType Kind, const Type *AccessTy) {
1891 const SCEV *Copy = Expr;
1892 int64_t Offset = ExtractImmediate(Expr, SE);
1894 // Basic uses can't accept any offset, for example.
1895 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1900 std::pair<UseMapTy::iterator, bool> P =
1901 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1903 // A use already existed with this base.
1904 size_t LUIdx = P.first->second;
1905 LSRUse &LU = Uses[LUIdx];
1906 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1908 return std::make_pair(LUIdx, Offset);
1911 // Create a new use.
1912 size_t LUIdx = Uses.size();
1913 P.first->second = LUIdx;
1914 Uses.push_back(LSRUse(Kind, AccessTy));
1915 LSRUse &LU = Uses[LUIdx];
1917 // We don't need to track redundant offsets, but we don't need to go out
1918 // of our way here to avoid them.
1919 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1920 LU.Offsets.push_back(Offset);
1922 LU.MinOffset = Offset;
1923 LU.MaxOffset = Offset;
1924 return std::make_pair(LUIdx, Offset);
1927 /// DeleteUse - Delete the given use from the Uses list.
1928 void LSRInstance::DeleteUse(LSRUse &LU) {
1929 if (&LU != &Uses.back())
1930 std::swap(LU, Uses.back());
1934 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1935 /// a formula that has the same registers as the given formula.
1937 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1938 const LSRUse &OrigLU) {
1939 // Search all uses for the formula. This could be more clever.
1940 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1941 LSRUse &LU = Uses[LUIdx];
1942 // Check whether this use is close enough to OrigLU, to see whether it's
1943 // worthwhile looking through its formulae.
1944 // Ignore ICmpZero uses because they may contain formulae generated by
1945 // GenerateICmpZeroScales, in which case adding fixup offsets may
1947 if (&LU != &OrigLU &&
1948 LU.Kind != LSRUse::ICmpZero &&
1949 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1950 LU.WidestFixupType == OrigLU.WidestFixupType &&
1951 LU.HasFormulaWithSameRegs(OrigF)) {
1952 // Scan through this use's formulae.
1953 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1954 E = LU.Formulae.end(); I != E; ++I) {
1955 const Formula &F = *I;
1956 // Check to see if this formula has the same registers and symbols
1958 if (F.BaseRegs == OrigF.BaseRegs &&
1959 F.ScaledReg == OrigF.ScaledReg &&
1960 F.AM.BaseGV == OrigF.AM.BaseGV &&
1961 F.AM.Scale == OrigF.AM.Scale) {
1962 if (F.AM.BaseOffs == 0)
1964 // This is the formula where all the registers and symbols matched;
1965 // there aren't going to be any others. Since we declined it, we
1966 // can skip the rest of the formulae and procede to the next LSRUse.
1973 // Nothing looked good.
1977 void LSRInstance::CollectInterestingTypesAndFactors() {
1978 SmallSetVector<const SCEV *, 4> Strides;
1980 // Collect interesting types and strides.
1981 SmallVector<const SCEV *, 4> Worklist;
1982 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1983 const SCEV *Expr = IU.getExpr(*UI);
1985 // Collect interesting types.
1986 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1988 // Add strides for mentioned loops.
1989 Worklist.push_back(Expr);
1991 const SCEV *S = Worklist.pop_back_val();
1992 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1993 Strides.insert(AR->getStepRecurrence(SE));
1994 Worklist.push_back(AR->getStart());
1995 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1996 Worklist.append(Add->op_begin(), Add->op_end());
1998 } while (!Worklist.empty());
2001 // Compute interesting factors from the set of interesting strides.
2002 for (SmallSetVector<const SCEV *, 4>::const_iterator
2003 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2004 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2005 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2006 const SCEV *OldStride = *I;
2007 const SCEV *NewStride = *NewStrideIter;
2009 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2010 SE.getTypeSizeInBits(NewStride->getType())) {
2011 if (SE.getTypeSizeInBits(OldStride->getType()) >
2012 SE.getTypeSizeInBits(NewStride->getType()))
2013 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2015 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2017 if (const SCEVConstant *Factor =
2018 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2020 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2021 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2022 } else if (const SCEVConstant *Factor =
2023 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2026 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2027 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2031 // If all uses use the same type, don't bother looking for truncation-based
2033 if (Types.size() == 1)
2036 DEBUG(print_factors_and_types(dbgs()));
2039 void LSRInstance::CollectFixupsAndInitialFormulae() {
2040 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2042 LSRFixup &LF = getNewFixup();
2043 LF.UserInst = UI->getUser();
2044 LF.OperandValToReplace = UI->getOperandValToReplace();
2045 LF.PostIncLoops = UI->getPostIncLoops();
2047 LSRUse::KindType Kind = LSRUse::Basic;
2048 const Type *AccessTy = 0;
2049 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2050 Kind = LSRUse::Address;
2051 AccessTy = getAccessType(LF.UserInst);
2054 const SCEV *S = IU.getExpr(*UI);
2056 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2057 // (N - i == 0), and this allows (N - i) to be the expression that we work
2058 // with rather than just N or i, so we can consider the register
2059 // requirements for both N and i at the same time. Limiting this code to
2060 // equality icmps is not a problem because all interesting loops use
2061 // equality icmps, thanks to IndVarSimplify.
2062 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2063 if (CI->isEquality()) {
2064 // Swap the operands if needed to put the OperandValToReplace on the
2065 // left, for consistency.
2066 Value *NV = CI->getOperand(1);
2067 if (NV == LF.OperandValToReplace) {
2068 CI->setOperand(1, CI->getOperand(0));
2069 CI->setOperand(0, NV);
2070 NV = CI->getOperand(1);
2074 // x == y --> x - y == 0
2075 const SCEV *N = SE.getSCEV(NV);
2076 if (N->isLoopInvariant(L)) {
2077 Kind = LSRUse::ICmpZero;
2078 S = SE.getMinusSCEV(N, S);
2081 // -1 and the negations of all interesting strides (except the negation
2082 // of -1) are now also interesting.
2083 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2084 if (Factors[i] != -1)
2085 Factors.insert(-(uint64_t)Factors[i]);
2089 // Set up the initial formula for this use.
2090 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2092 LF.Offset = P.second;
2093 LSRUse &LU = Uses[LF.LUIdx];
2094 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2095 if (!LU.WidestFixupType ||
2096 SE.getTypeSizeInBits(LU.WidestFixupType) <
2097 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2098 LU.WidestFixupType = LF.OperandValToReplace->getType();
2100 // If this is the first use of this LSRUse, give it a formula.
2101 if (LU.Formulae.empty()) {
2102 InsertInitialFormula(S, LU, LF.LUIdx);
2103 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2107 DEBUG(print_fixups(dbgs()));
2110 /// InsertInitialFormula - Insert a formula for the given expression into
2111 /// the given use, separating out loop-variant portions from loop-invariant
2112 /// and loop-computable portions.
2114 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2116 F.InitialMatch(S, L, SE, DT);
2117 bool Inserted = InsertFormula(LU, LUIdx, F);
2118 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2121 /// InsertSupplementalFormula - Insert a simple single-register formula for
2122 /// the given expression into the given use.
2124 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2125 LSRUse &LU, size_t LUIdx) {
2127 F.BaseRegs.push_back(S);
2128 F.AM.HasBaseReg = true;
2129 bool Inserted = InsertFormula(LU, LUIdx, F);
2130 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2133 /// CountRegisters - Note which registers are used by the given formula,
2134 /// updating RegUses.
2135 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2137 RegUses.CountRegister(F.ScaledReg, LUIdx);
2138 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2139 E = F.BaseRegs.end(); I != E; ++I)
2140 RegUses.CountRegister(*I, LUIdx);
2143 /// InsertFormula - If the given formula has not yet been inserted, add it to
2144 /// the list, and return true. Return false otherwise.
2145 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2146 if (!LU.InsertFormula(F))
2149 CountRegisters(F, LUIdx);
2153 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2154 /// loop-invariant values which we're tracking. These other uses will pin these
2155 /// values in registers, making them less profitable for elimination.
2156 /// TODO: This currently misses non-constant addrec step registers.
2157 /// TODO: Should this give more weight to users inside the loop?
2159 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2160 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2161 SmallPtrSet<const SCEV *, 8> Inserted;
2163 while (!Worklist.empty()) {
2164 const SCEV *S = Worklist.pop_back_val();
2166 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2167 Worklist.append(N->op_begin(), N->op_end());
2168 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2169 Worklist.push_back(C->getOperand());
2170 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2171 Worklist.push_back(D->getLHS());
2172 Worklist.push_back(D->getRHS());
2173 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2174 if (!Inserted.insert(U)) continue;
2175 const Value *V = U->getValue();
2176 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2177 // Look for instructions defined outside the loop.
2178 if (L->contains(Inst)) continue;
2179 } else if (isa<UndefValue>(V))
2180 // Undef doesn't have a live range, so it doesn't matter.
2182 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2184 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2185 // Ignore non-instructions.
2188 // Ignore instructions in other functions (as can happen with
2190 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2192 // Ignore instructions not dominated by the loop.
2193 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2194 UserInst->getParent() :
2195 cast<PHINode>(UserInst)->getIncomingBlock(
2196 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2197 if (!DT.dominates(L->getHeader(), UseBB))
2199 // Ignore uses which are part of other SCEV expressions, to avoid
2200 // analyzing them multiple times.
2201 if (SE.isSCEVable(UserInst->getType())) {
2202 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2203 // If the user is a no-op, look through to its uses.
2204 if (!isa<SCEVUnknown>(UserS))
2208 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2212 // Ignore icmp instructions which are already being analyzed.
2213 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2214 unsigned OtherIdx = !UI.getOperandNo();
2215 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2216 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2220 LSRFixup &LF = getNewFixup();
2221 LF.UserInst = const_cast<Instruction *>(UserInst);
2222 LF.OperandValToReplace = UI.getUse();
2223 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2225 LF.Offset = P.second;
2226 LSRUse &LU = Uses[LF.LUIdx];
2227 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2228 if (!LU.WidestFixupType ||
2229 SE.getTypeSizeInBits(LU.WidestFixupType) <
2230 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2231 LU.WidestFixupType = LF.OperandValToReplace->getType();
2232 InsertSupplementalFormula(U, LU, LF.LUIdx);
2233 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2240 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2241 /// separate registers. If C is non-null, multiply each subexpression by C.
2242 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2243 SmallVectorImpl<const SCEV *> &Ops,
2245 ScalarEvolution &SE) {
2246 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2247 // Break out add operands.
2248 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2250 CollectSubexprs(*I, C, Ops, L, SE);
2252 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2253 // Split a non-zero base out of an addrec.
2254 if (!AR->getStart()->isZero()) {
2255 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2256 AR->getStepRecurrence(SE),
2259 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2262 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2263 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2264 if (Mul->getNumOperands() == 2)
2265 if (const SCEVConstant *Op0 =
2266 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2267 CollectSubexprs(Mul->getOperand(1),
2268 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2274 // Otherwise use the value itself, optionally with a scale applied.
2275 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2278 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2280 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2283 // Arbitrarily cap recursion to protect compile time.
2284 if (Depth >= 3) return;
2286 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2287 const SCEV *BaseReg = Base.BaseRegs[i];
2289 SmallVector<const SCEV *, 8> AddOps;
2290 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2292 if (AddOps.size() == 1) continue;
2294 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2295 JE = AddOps.end(); J != JE; ++J) {
2297 // Loop-variant "unknown" values are uninteresting; we won't be able to
2298 // do anything meaningful with them.
2299 if (isa<SCEVUnknown>(*J) && !(*J)->isLoopInvariant(L))
2302 // Don't pull a constant into a register if the constant could be folded
2303 // into an immediate field.
2304 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2305 Base.getNumRegs() > 1,
2306 LU.Kind, LU.AccessTy, TLI, SE))
2309 // Collect all operands except *J.
2310 SmallVector<const SCEV *, 8> InnerAddOps
2311 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2313 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2315 // Don't leave just a constant behind in a register if the constant could
2316 // be folded into an immediate field.
2317 if (InnerAddOps.size() == 1 &&
2318 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2319 Base.getNumRegs() > 1,
2320 LU.Kind, LU.AccessTy, TLI, SE))
2323 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2324 if (InnerSum->isZero())
2327 F.BaseRegs[i] = InnerSum;
2328 F.BaseRegs.push_back(*J);
2329 if (InsertFormula(LU, LUIdx, F))
2330 // If that formula hadn't been seen before, recurse to find more like
2332 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2337 /// GenerateCombinations - Generate a formula consisting of all of the
2338 /// loop-dominating registers added into a single register.
2339 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2341 // This method is only interesting on a plurality of registers.
2342 if (Base.BaseRegs.size() <= 1) return;
2346 SmallVector<const SCEV *, 4> Ops;
2347 for (SmallVectorImpl<const SCEV *>::const_iterator
2348 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2349 const SCEV *BaseReg = *I;
2350 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2351 !BaseReg->hasComputableLoopEvolution(L))
2352 Ops.push_back(BaseReg);
2354 F.BaseRegs.push_back(BaseReg);
2356 if (Ops.size() > 1) {
2357 const SCEV *Sum = SE.getAddExpr(Ops);
2358 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2359 // opportunity to fold something. For now, just ignore such cases
2360 // rather than proceed with zero in a register.
2361 if (!Sum->isZero()) {
2362 F.BaseRegs.push_back(Sum);
2363 (void)InsertFormula(LU, LUIdx, F);
2368 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2369 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2371 // We can't add a symbolic offset if the address already contains one.
2372 if (Base.AM.BaseGV) return;
2374 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2375 const SCEV *G = Base.BaseRegs[i];
2376 GlobalValue *GV = ExtractSymbol(G, SE);
2377 if (G->isZero() || !GV)
2381 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2382 LU.Kind, LU.AccessTy, TLI))
2385 (void)InsertFormula(LU, LUIdx, F);
2389 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2390 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2392 // TODO: For now, just add the min and max offset, because it usually isn't
2393 // worthwhile looking at everything inbetween.
2394 SmallVector<int64_t, 2> Worklist;
2395 Worklist.push_back(LU.MinOffset);
2396 if (LU.MaxOffset != LU.MinOffset)
2397 Worklist.push_back(LU.MaxOffset);
2399 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2400 const SCEV *G = Base.BaseRegs[i];
2402 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2403 E = Worklist.end(); I != E; ++I) {
2405 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2406 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2407 LU.Kind, LU.AccessTy, TLI)) {
2408 // Add the offset to the base register.
2409 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2410 // If it cancelled out, drop the base register, otherwise update it.
2411 if (NewG->isZero()) {
2412 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2413 F.BaseRegs.pop_back();
2415 F.BaseRegs[i] = NewG;
2417 (void)InsertFormula(LU, LUIdx, F);
2421 int64_t Imm = ExtractImmediate(G, SE);
2422 if (G->isZero() || Imm == 0)
2425 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2426 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2427 LU.Kind, LU.AccessTy, TLI))
2430 (void)InsertFormula(LU, LUIdx, F);
2434 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2435 /// the comparison. For example, x == y -> x*c == y*c.
2436 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2438 if (LU.Kind != LSRUse::ICmpZero) return;
2440 // Determine the integer type for the base formula.
2441 const Type *IntTy = Base.getType();
2443 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2445 // Don't do this if there is more than one offset.
2446 if (LU.MinOffset != LU.MaxOffset) return;
2448 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2450 // Check each interesting stride.
2451 for (SmallSetVector<int64_t, 8>::const_iterator
2452 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2453 int64_t Factor = *I;
2455 // Check that the multiplication doesn't overflow.
2456 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2458 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2459 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2462 // Check that multiplying with the use offset doesn't overflow.
2463 int64_t Offset = LU.MinOffset;
2464 if (Offset == INT64_MIN && Factor == -1)
2466 Offset = (uint64_t)Offset * Factor;
2467 if (Offset / Factor != LU.MinOffset)
2471 F.AM.BaseOffs = NewBaseOffs;
2473 // Check that this scale is legal.
2474 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2477 // Compensate for the use having MinOffset built into it.
2478 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2480 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2482 // Check that multiplying with each base register doesn't overflow.
2483 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2484 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2485 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2489 // Check that multiplying with the scaled register doesn't overflow.
2491 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2492 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2496 // If we make it here and it's legal, add it.
2497 (void)InsertFormula(LU, LUIdx, F);
2502 /// GenerateScales - Generate stride factor reuse formulae by making use of
2503 /// scaled-offset address modes, for example.
2504 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2505 // Determine the integer type for the base formula.
2506 const Type *IntTy = Base.getType();
2509 // If this Formula already has a scaled register, we can't add another one.
2510 if (Base.AM.Scale != 0) return;
2512 // Check each interesting stride.
2513 for (SmallSetVector<int64_t, 8>::const_iterator
2514 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2515 int64_t Factor = *I;
2517 Base.AM.Scale = Factor;
2518 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2519 // Check whether this scale is going to be legal.
2520 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2521 LU.Kind, LU.AccessTy, TLI)) {
2522 // As a special-case, handle special out-of-loop Basic users specially.
2523 // TODO: Reconsider this special case.
2524 if (LU.Kind == LSRUse::Basic &&
2525 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2526 LSRUse::Special, LU.AccessTy, TLI) &&
2527 LU.AllFixupsOutsideLoop)
2528 LU.Kind = LSRUse::Special;
2532 // For an ICmpZero, negating a solitary base register won't lead to
2534 if (LU.Kind == LSRUse::ICmpZero &&
2535 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2537 // For each addrec base reg, apply the scale, if possible.
2538 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2539 if (const SCEVAddRecExpr *AR =
2540 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2541 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2542 if (FactorS->isZero())
2544 // Divide out the factor, ignoring high bits, since we'll be
2545 // scaling the value back up in the end.
2546 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2547 // TODO: This could be optimized to avoid all the copying.
2549 F.ScaledReg = Quotient;
2550 F.DeleteBaseReg(F.BaseRegs[i]);
2551 (void)InsertFormula(LU, LUIdx, F);
2557 /// GenerateTruncates - Generate reuse formulae from different IV types.
2558 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2559 // This requires TargetLowering to tell us which truncates are free.
2562 // Don't bother truncating symbolic values.
2563 if (Base.AM.BaseGV) return;
2565 // Determine the integer type for the base formula.
2566 const Type *DstTy = Base.getType();
2568 DstTy = SE.getEffectiveSCEVType(DstTy);
2570 for (SmallSetVector<const Type *, 4>::const_iterator
2571 I = Types.begin(), E = Types.end(); I != E; ++I) {
2572 const Type *SrcTy = *I;
2573 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2576 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2577 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2578 JE = F.BaseRegs.end(); J != JE; ++J)
2579 *J = SE.getAnyExtendExpr(*J, SrcTy);
2581 // TODO: This assumes we've done basic processing on all uses and
2582 // have an idea what the register usage is.
2583 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2586 (void)InsertFormula(LU, LUIdx, F);
2593 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2594 /// defer modifications so that the search phase doesn't have to worry about
2595 /// the data structures moving underneath it.
2599 const SCEV *OrigReg;
2601 WorkItem(size_t LI, int64_t I, const SCEV *R)
2602 : LUIdx(LI), Imm(I), OrigReg(R) {}
2604 void print(raw_ostream &OS) const;
2610 void WorkItem::print(raw_ostream &OS) const {
2611 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2612 << " , add offset " << Imm;
2615 void WorkItem::dump() const {
2616 print(errs()); errs() << '\n';
2619 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2620 /// distance apart and try to form reuse opportunities between them.
2621 void LSRInstance::GenerateCrossUseConstantOffsets() {
2622 // Group the registers by their value without any added constant offset.
2623 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2624 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2626 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2627 SmallVector<const SCEV *, 8> Sequence;
2628 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2630 const SCEV *Reg = *I;
2631 int64_t Imm = ExtractImmediate(Reg, SE);
2632 std::pair<RegMapTy::iterator, bool> Pair =
2633 Map.insert(std::make_pair(Reg, ImmMapTy()));
2635 Sequence.push_back(Reg);
2636 Pair.first->second.insert(std::make_pair(Imm, *I));
2637 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2640 // Now examine each set of registers with the same base value. Build up
2641 // a list of work to do and do the work in a separate step so that we're
2642 // not adding formulae and register counts while we're searching.
2643 SmallVector<WorkItem, 32> WorkItems;
2644 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2645 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2646 E = Sequence.end(); I != E; ++I) {
2647 const SCEV *Reg = *I;
2648 const ImmMapTy &Imms = Map.find(Reg)->second;
2650 // It's not worthwhile looking for reuse if there's only one offset.
2651 if (Imms.size() == 1)
2654 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2655 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2657 dbgs() << ' ' << J->first;
2660 // Examine each offset.
2661 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2663 const SCEV *OrigReg = J->second;
2665 int64_t JImm = J->first;
2666 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2668 if (!isa<SCEVConstant>(OrigReg) &&
2669 UsedByIndicesMap[Reg].count() == 1) {
2670 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2674 // Conservatively examine offsets between this orig reg a few selected
2676 ImmMapTy::const_iterator OtherImms[] = {
2677 Imms.begin(), prior(Imms.end()),
2678 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2680 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2681 ImmMapTy::const_iterator M = OtherImms[i];
2682 if (M == J || M == JE) continue;
2684 // Compute the difference between the two.
2685 int64_t Imm = (uint64_t)JImm - M->first;
2686 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2687 LUIdx = UsedByIndices.find_next(LUIdx))
2688 // Make a memo of this use, offset, and register tuple.
2689 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2690 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2697 UsedByIndicesMap.clear();
2698 UniqueItems.clear();
2700 // Now iterate through the worklist and add new formulae.
2701 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2702 E = WorkItems.end(); I != E; ++I) {
2703 const WorkItem &WI = *I;
2704 size_t LUIdx = WI.LUIdx;
2705 LSRUse &LU = Uses[LUIdx];
2706 int64_t Imm = WI.Imm;
2707 const SCEV *OrigReg = WI.OrigReg;
2709 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2710 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2711 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2713 // TODO: Use a more targeted data structure.
2714 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2715 const Formula &F = LU.Formulae[L];
2716 // Use the immediate in the scaled register.
2717 if (F.ScaledReg == OrigReg) {
2718 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2719 Imm * (uint64_t)F.AM.Scale;
2720 // Don't create 50 + reg(-50).
2721 if (F.referencesReg(SE.getSCEV(
2722 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2725 NewF.AM.BaseOffs = Offs;
2726 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2727 LU.Kind, LU.AccessTy, TLI))
2729 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2731 // If the new scale is a constant in a register, and adding the constant
2732 // value to the immediate would produce a value closer to zero than the
2733 // immediate itself, then the formula isn't worthwhile.
2734 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2735 if (C->getValue()->getValue().isNegative() !=
2736 (NewF.AM.BaseOffs < 0) &&
2737 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2738 .ule(abs64(NewF.AM.BaseOffs)))
2742 (void)InsertFormula(LU, LUIdx, NewF);
2744 // Use the immediate in a base register.
2745 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2746 const SCEV *BaseReg = F.BaseRegs[N];
2747 if (BaseReg != OrigReg)
2750 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2751 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2752 LU.Kind, LU.AccessTy, TLI))
2754 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2756 // If the new formula has a constant in a register, and adding the
2757 // constant value to the immediate would produce a value closer to
2758 // zero than the immediate itself, then the formula isn't worthwhile.
2759 for (SmallVectorImpl<const SCEV *>::const_iterator
2760 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2762 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2763 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2764 abs64(NewF.AM.BaseOffs)) &&
2765 (C->getValue()->getValue() +
2766 NewF.AM.BaseOffs).countTrailingZeros() >=
2767 CountTrailingZeros_64(NewF.AM.BaseOffs))
2771 (void)InsertFormula(LU, LUIdx, NewF);
2780 /// GenerateAllReuseFormulae - Generate formulae for each use.
2782 LSRInstance::GenerateAllReuseFormulae() {
2783 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2784 // queries are more precise.
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 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2789 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2790 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2792 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2793 LSRUse &LU = Uses[LUIdx];
2794 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2795 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2796 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2797 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2798 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2799 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2800 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2801 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2803 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2804 LSRUse &LU = Uses[LUIdx];
2805 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2806 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2809 GenerateCrossUseConstantOffsets();
2811 DEBUG(dbgs() << "\n"
2812 "After generating reuse formulae:\n";
2813 print_uses(dbgs()));
2816 /// If their are multiple formulae with the same set of registers used
2817 /// by other uses, pick the best one and delete the others.
2818 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2820 bool ChangedFormulae = false;
2823 // Collect the best formula for each unique set of shared registers. This
2824 // is reset for each use.
2825 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2827 BestFormulaeTy BestFormulae;
2829 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2830 LSRUse &LU = Uses[LUIdx];
2831 FormulaSorter Sorter(L, LU, SE, DT);
2832 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2835 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2836 FIdx != NumForms; ++FIdx) {
2837 Formula &F = LU.Formulae[FIdx];
2839 SmallVector<const SCEV *, 2> Key;
2840 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2841 JE = F.BaseRegs.end(); J != JE; ++J) {
2842 const SCEV *Reg = *J;
2843 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2847 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2848 Key.push_back(F.ScaledReg);
2849 // Unstable sort by host order ok, because this is only used for
2851 std::sort(Key.begin(), Key.end());
2853 std::pair<BestFormulaeTy::const_iterator, bool> P =
2854 BestFormulae.insert(std::make_pair(Key, FIdx));
2856 Formula &Best = LU.Formulae[P.first->second];
2857 if (Sorter.operator()(F, Best))
2859 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2861 " in favor of formula "; Best.print(dbgs());
2864 ChangedFormulae = true;
2866 LU.DeleteFormula(F);
2874 // Now that we've filtered out some formulae, recompute the Regs set.
2876 LU.RecomputeRegs(LUIdx, RegUses);
2878 // Reset this to prepare for the next use.
2879 BestFormulae.clear();
2882 DEBUG(if (ChangedFormulae) {
2884 "After filtering out undesirable candidates:\n";
2889 // This is a rough guess that seems to work fairly well.
2890 static const size_t ComplexityLimit = UINT16_MAX;
2892 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2893 /// solutions the solver might have to consider. It almost never considers
2894 /// this many solutions because it prune the search space, but the pruning
2895 /// isn't always sufficient.
2896 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2898 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2899 E = Uses.end(); I != E; ++I) {
2900 size_t FSize = I->Formulae.size();
2901 if (FSize >= ComplexityLimit) {
2902 Power = ComplexityLimit;
2906 if (Power >= ComplexityLimit)
2912 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2913 /// of the registers of another formula, it won't help reduce register
2914 /// pressure (though it may not necessarily hurt register pressure); remove
2915 /// it to simplify the system.
2916 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2917 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2918 DEBUG(dbgs() << "The search space is too complex.\n");
2920 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2921 "which use a superset of registers used by other "
2924 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2925 LSRUse &LU = Uses[LUIdx];
2927 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2928 Formula &F = LU.Formulae[i];
2929 // Look for a formula with a constant or GV in a register. If the use
2930 // also has a formula with that same value in an immediate field,
2931 // delete the one that uses a register.
2932 for (SmallVectorImpl<const SCEV *>::const_iterator
2933 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2934 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2936 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2937 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2938 (I - F.BaseRegs.begin()));
2939 if (LU.HasFormulaWithSameRegs(NewF)) {
2940 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2941 LU.DeleteFormula(F);
2947 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2948 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2951 NewF.AM.BaseGV = GV;
2952 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2953 (I - F.BaseRegs.begin()));
2954 if (LU.HasFormulaWithSameRegs(NewF)) {
2955 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2957 LU.DeleteFormula(F);
2968 LU.RecomputeRegs(LUIdx, RegUses);
2971 DEBUG(dbgs() << "After pre-selection:\n";
2972 print_uses(dbgs()));
2976 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
2977 /// for expressions like A, A+1, A+2, etc., allocate a single register for
2979 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
2980 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2981 DEBUG(dbgs() << "The search space is too complex.\n");
2983 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2984 "separated by a constant offset will use the same "
2987 // This is especially useful for unrolled loops.
2989 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2990 LSRUse &LU = Uses[LUIdx];
2991 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2992 E = LU.Formulae.end(); I != E; ++I) {
2993 const Formula &F = *I;
2994 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2995 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2996 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2997 /*HasBaseReg=*/false,
2998 LU.Kind, LU.AccessTy)) {
2999 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3002 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3004 // Delete formulae from the new use which are no longer legal.
3006 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3007 Formula &F = LUThatHas->Formulae[i];
3008 if (!isLegalUse(F.AM,
3009 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3010 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3011 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3013 LUThatHas->DeleteFormula(F);
3020 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3022 // Update the relocs to reference the new use.
3023 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3024 E = Fixups.end(); I != E; ++I) {
3025 LSRFixup &Fixup = *I;
3026 if (Fixup.LUIdx == LUIdx) {
3027 Fixup.LUIdx = LUThatHas - &Uses.front();
3028 Fixup.Offset += F.AM.BaseOffs;
3029 DEBUG(dbgs() << "New fixup has offset "
3030 << Fixup.Offset << '\n');
3032 if (Fixup.LUIdx == NumUses-1)
3033 Fixup.LUIdx = LUIdx;
3036 // Delete the old use.
3047 DEBUG(dbgs() << "After pre-selection:\n";
3048 print_uses(dbgs()));
3052 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3053 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3054 /// we've done more filtering, as it may be able to find more formulae to
3056 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3057 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3058 DEBUG(dbgs() << "The search space is too complex.\n");
3060 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3061 "undesirable dedicated registers.\n");
3063 FilterOutUndesirableDedicatedRegisters();
3065 DEBUG(dbgs() << "After pre-selection:\n";
3066 print_uses(dbgs()));
3070 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3071 /// to be profitable, and then in any use which has any reference to that
3072 /// register, delete all formulae which do not reference that register.
3073 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3074 // With all other options exhausted, loop until the system is simple
3075 // enough to handle.
3076 SmallPtrSet<const SCEV *, 4> Taken;
3077 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3078 // Ok, we have too many of formulae on our hands to conveniently handle.
3079 // Use a rough heuristic to thin out the list.
3080 DEBUG(dbgs() << "The search space is too complex.\n");
3082 // Pick the register which is used by the most LSRUses, which is likely
3083 // to be a good reuse register candidate.
3084 const SCEV *Best = 0;
3085 unsigned BestNum = 0;
3086 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3088 const SCEV *Reg = *I;
3089 if (Taken.count(Reg))
3094 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3095 if (Count > BestNum) {
3102 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3103 << " will yield profitable reuse.\n");
3106 // In any use with formulae which references this register, delete formulae
3107 // which don't reference it.
3108 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3109 LSRUse &LU = Uses[LUIdx];
3110 if (!LU.Regs.count(Best)) continue;
3113 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3114 Formula &F = LU.Formulae[i];
3115 if (!F.referencesReg(Best)) {
3116 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3117 LU.DeleteFormula(F);
3121 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3127 LU.RecomputeRegs(LUIdx, RegUses);
3130 DEBUG(dbgs() << "After pre-selection:\n";
3131 print_uses(dbgs()));
3135 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3136 /// formulae to choose from, use some rough heuristics to prune down the number
3137 /// of formulae. This keeps the main solver from taking an extraordinary amount
3138 /// of time in some worst-case scenarios.
3139 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3140 NarrowSearchSpaceByDetectingSupersets();
3141 NarrowSearchSpaceByCollapsingUnrolledCode();
3142 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3143 NarrowSearchSpaceByPickingWinnerRegs();
3146 /// SolveRecurse - This is the recursive solver.
3147 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3149 SmallVectorImpl<const Formula *> &Workspace,
3150 const Cost &CurCost,
3151 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3152 DenseSet<const SCEV *> &VisitedRegs) const {
3155 // - use more aggressive filtering
3156 // - sort the formula so that the most profitable solutions are found first
3157 // - sort the uses too
3159 // - don't compute a cost, and then compare. compare while computing a cost
3161 // - track register sets with SmallBitVector
3163 const LSRUse &LU = Uses[Workspace.size()];
3165 // If this use references any register that's already a part of the
3166 // in-progress solution, consider it a requirement that a formula must
3167 // reference that register in order to be considered. This prunes out
3168 // unprofitable searching.
3169 SmallSetVector<const SCEV *, 4> ReqRegs;
3170 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3171 E = CurRegs.end(); I != E; ++I)
3172 if (LU.Regs.count(*I))
3175 bool AnySatisfiedReqRegs = false;
3176 SmallPtrSet<const SCEV *, 16> NewRegs;
3179 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3180 E = LU.Formulae.end(); I != E; ++I) {
3181 const Formula &F = *I;
3183 // Ignore formulae which do not use any of the required registers.
3184 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3185 JE = ReqRegs.end(); J != JE; ++J) {
3186 const SCEV *Reg = *J;
3187 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3188 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3192 AnySatisfiedReqRegs = true;
3194 // Evaluate the cost of the current formula. If it's already worse than
3195 // the current best, prune the search at that point.
3198 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3199 if (NewCost < SolutionCost) {
3200 Workspace.push_back(&F);
3201 if (Workspace.size() != Uses.size()) {
3202 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3203 NewRegs, VisitedRegs);
3204 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3205 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3207 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3208 dbgs() << ". Regs:";
3209 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3210 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3211 dbgs() << ' ' << **I;
3214 SolutionCost = NewCost;
3215 Solution = Workspace;
3217 Workspace.pop_back();
3222 // If none of the formulae had all of the required registers, relax the
3223 // constraint so that we don't exclude all formulae.
3224 if (!AnySatisfiedReqRegs) {
3225 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3231 /// Solve - Choose one formula from each use. Return the results in the given
3232 /// Solution vector.
3233 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3234 SmallVector<const Formula *, 8> Workspace;
3236 SolutionCost.Loose();
3238 SmallPtrSet<const SCEV *, 16> CurRegs;
3239 DenseSet<const SCEV *> VisitedRegs;
3240 Workspace.reserve(Uses.size());
3242 // SolveRecurse does all the work.
3243 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3244 CurRegs, VisitedRegs);
3246 // Ok, we've now made all our decisions.
3247 DEBUG(dbgs() << "\n"
3248 "The chosen solution requires "; SolutionCost.print(dbgs());
3250 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3252 Uses[i].print(dbgs());
3255 Solution[i]->print(dbgs());
3259 assert(Solution.size() == Uses.size() && "Malformed solution!");
3262 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3263 /// the dominator tree far as we can go while still being dominated by the
3264 /// input positions. This helps canonicalize the insert position, which
3265 /// encourages sharing.
3266 BasicBlock::iterator
3267 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3268 const SmallVectorImpl<Instruction *> &Inputs)
3271 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3272 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3275 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3276 if (!Rung) return IP;
3277 Rung = Rung->getIDom();
3278 if (!Rung) return IP;
3279 IDom = Rung->getBlock();
3281 // Don't climb into a loop though.
3282 const Loop *IDomLoop = LI.getLoopFor(IDom);
3283 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3284 if (IDomDepth <= IPLoopDepth &&
3285 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3289 bool AllDominate = true;
3290 Instruction *BetterPos = 0;
3291 Instruction *Tentative = IDom->getTerminator();
3292 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3293 E = Inputs.end(); I != E; ++I) {
3294 Instruction *Inst = *I;
3295 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3296 AllDominate = false;
3299 // Attempt to find an insert position in the middle of the block,
3300 // instead of at the end, so that it can be used for other expansions.
3301 if (IDom == Inst->getParent() &&
3302 (!BetterPos || DT.dominates(BetterPos, Inst)))
3303 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3316 /// AdjustInsertPositionForExpand - Determine an input position which will be
3317 /// dominated by the operands and which will dominate the result.
3318 BasicBlock::iterator
3319 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3321 const LSRUse &LU) const {
3322 // Collect some instructions which must be dominated by the
3323 // expanding replacement. These must be dominated by any operands that
3324 // will be required in the expansion.
3325 SmallVector<Instruction *, 4> Inputs;
3326 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3327 Inputs.push_back(I);
3328 if (LU.Kind == LSRUse::ICmpZero)
3329 if (Instruction *I =
3330 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3331 Inputs.push_back(I);
3332 if (LF.PostIncLoops.count(L)) {
3333 if (LF.isUseFullyOutsideLoop(L))
3334 Inputs.push_back(L->getLoopLatch()->getTerminator());
3336 Inputs.push_back(IVIncInsertPos);
3338 // The expansion must also be dominated by the increment positions of any
3339 // loops it for which it is using post-inc mode.
3340 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3341 E = LF.PostIncLoops.end(); I != E; ++I) {
3342 const Loop *PIL = *I;
3343 if (PIL == L) continue;
3345 // Be dominated by the loop exit.
3346 SmallVector<BasicBlock *, 4> ExitingBlocks;
3347 PIL->getExitingBlocks(ExitingBlocks);
3348 if (!ExitingBlocks.empty()) {
3349 BasicBlock *BB = ExitingBlocks[0];
3350 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3351 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3352 Inputs.push_back(BB->getTerminator());
3356 // Then, climb up the immediate dominator tree as far as we can go while
3357 // still being dominated by the input positions.
3358 IP = HoistInsertPosition(IP, Inputs);
3360 // Don't insert instructions before PHI nodes.
3361 while (isa<PHINode>(IP)) ++IP;
3363 // Ignore debug intrinsics.
3364 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3369 /// Expand - Emit instructions for the leading candidate expression for this
3370 /// LSRUse (this is called "expanding").
3371 Value *LSRInstance::Expand(const LSRFixup &LF,
3373 BasicBlock::iterator IP,
3374 SCEVExpander &Rewriter,
3375 SmallVectorImpl<WeakVH> &DeadInsts) const {
3376 const LSRUse &LU = Uses[LF.LUIdx];
3378 // Determine an input position which will be dominated by the operands and
3379 // which will dominate the result.
3380 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3382 // Inform the Rewriter if we have a post-increment use, so that it can
3383 // perform an advantageous expansion.
3384 Rewriter.setPostInc(LF.PostIncLoops);
3386 // This is the type that the user actually needs.
3387 const Type *OpTy = LF.OperandValToReplace->getType();
3388 // This will be the type that we'll initially expand to.
3389 const Type *Ty = F.getType();
3391 // No type known; just expand directly to the ultimate type.
3393 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3394 // Expand directly to the ultimate type if it's the right size.
3396 // This is the type to do integer arithmetic in.
3397 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3399 // Build up a list of operands to add together to form the full base.
3400 SmallVector<const SCEV *, 8> Ops;
3402 // Expand the BaseRegs portion.
3403 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3404 E = F.BaseRegs.end(); I != E; ++I) {
3405 const SCEV *Reg = *I;
3406 assert(!Reg->isZero() && "Zero allocated in a base register!");
3408 // If we're expanding for a post-inc user, make the post-inc adjustment.
3409 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3410 Reg = TransformForPostIncUse(Denormalize, Reg,
3411 LF.UserInst, LF.OperandValToReplace,
3414 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3417 // Flush the operand list to suppress SCEVExpander hoisting.
3419 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3421 Ops.push_back(SE.getUnknown(FullV));
3424 // Expand the ScaledReg portion.
3425 Value *ICmpScaledV = 0;
3426 if (F.AM.Scale != 0) {
3427 const SCEV *ScaledS = F.ScaledReg;
3429 // If we're expanding for a post-inc user, make the post-inc adjustment.
3430 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3431 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3432 LF.UserInst, LF.OperandValToReplace,
3435 if (LU.Kind == LSRUse::ICmpZero) {
3436 // An interesting way of "folding" with an icmp is to use a negated
3437 // scale, which we'll implement by inserting it into the other operand
3439 assert(F.AM.Scale == -1 &&
3440 "The only scale supported by ICmpZero uses is -1!");
3441 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3443 // Otherwise just expand the scaled register and an explicit scale,
3444 // which is expected to be matched as part of the address.
3445 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3446 ScaledS = SE.getMulExpr(ScaledS,
3447 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3448 Ops.push_back(ScaledS);
3450 // Flush the operand list to suppress SCEVExpander hoisting.
3451 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3453 Ops.push_back(SE.getUnknown(FullV));
3457 // Expand the GV portion.
3459 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3461 // Flush the operand list to suppress SCEVExpander hoisting.
3462 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3464 Ops.push_back(SE.getUnknown(FullV));
3467 // Expand the immediate portion.
3468 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3470 if (LU.Kind == LSRUse::ICmpZero) {
3471 // The other interesting way of "folding" with an ICmpZero is to use a
3472 // negated immediate.
3474 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3476 Ops.push_back(SE.getUnknown(ICmpScaledV));
3477 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3480 // Just add the immediate values. These again are expected to be matched
3481 // as part of the address.
3482 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3486 // Emit instructions summing all the operands.
3487 const SCEV *FullS = Ops.empty() ?
3488 SE.getConstant(IntTy, 0) :
3490 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3492 // We're done expanding now, so reset the rewriter.
3493 Rewriter.clearPostInc();
3495 // An ICmpZero Formula represents an ICmp which we're handling as a
3496 // comparison against zero. Now that we've expanded an expression for that
3497 // form, update the ICmp's other operand.
3498 if (LU.Kind == LSRUse::ICmpZero) {
3499 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3500 DeadInsts.push_back(CI->getOperand(1));
3501 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3502 "a scale at the same time!");
3503 if (F.AM.Scale == -1) {
3504 if (ICmpScaledV->getType() != OpTy) {
3506 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3508 ICmpScaledV, OpTy, "tmp", CI);
3511 CI->setOperand(1, ICmpScaledV);
3513 assert(F.AM.Scale == 0 &&
3514 "ICmp does not support folding a global value and "
3515 "a scale at the same time!");
3516 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3518 if (C->getType() != OpTy)
3519 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3523 CI->setOperand(1, C);
3530 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3531 /// of their operands effectively happens in their predecessor blocks, so the
3532 /// expression may need to be expanded in multiple places.
3533 void LSRInstance::RewriteForPHI(PHINode *PN,
3536 SCEVExpander &Rewriter,
3537 SmallVectorImpl<WeakVH> &DeadInsts,
3539 DenseMap<BasicBlock *, Value *> Inserted;
3540 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3541 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3542 BasicBlock *BB = PN->getIncomingBlock(i);
3544 // If this is a critical edge, split the edge so that we do not insert
3545 // the code on all predecessor/successor paths. We do this unless this
3546 // is the canonical backedge for this loop, which complicates post-inc
3548 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3549 !isa<IndirectBrInst>(BB->getTerminator()) &&
3550 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3551 // Split the critical edge.
3552 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3554 // If PN is outside of the loop and BB is in the loop, we want to
3555 // move the block to be immediately before the PHI block, not
3556 // immediately after BB.
3557 if (L->contains(BB) && !L->contains(PN))
3558 NewBB->moveBefore(PN->getParent());
3560 // Splitting the edge can reduce the number of PHI entries we have.
3561 e = PN->getNumIncomingValues();
3563 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(LoopStrengthReduce, "loop-reduce",
3796 "Loop Strength Reduction", false, false);
3798 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3799 return new LoopStrengthReduce(TLI);
3802 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3803 : LoopPass(ID), TLI(tli) {}
3805 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3806 // We split critical edges, so we change the CFG. However, we do update
3807 // many analyses if they are around.
3808 AU.addPreservedID(LoopSimplifyID);
3809 AU.addPreserved("domfrontier");
3811 AU.addRequired<LoopInfo>();
3812 AU.addPreserved<LoopInfo>();
3813 AU.addRequiredID(LoopSimplifyID);
3814 AU.addRequired<DominatorTree>();
3815 AU.addPreserved<DominatorTree>();
3816 AU.addRequired<ScalarEvolution>();
3817 AU.addPreserved<ScalarEvolution>();
3818 AU.addRequired<IVUsers>();
3819 AU.addPreserved<IVUsers>();
3822 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3823 bool Changed = false;
3825 // Run the main LSR transformation.
3826 Changed |= LSRInstance(TLI, L, this).getChanged();
3828 // At this point, it is worth checking to see if any recurrence PHIs are also
3829 // dead, so that we can remove them as well.
3830 Changed |= DeleteDeadPHIs(L->getHeader());