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 bool operator<(const Cost &Other) const;
652 void RateFormula(const Formula &F,
653 SmallPtrSet<const SCEV *, 16> &Regs,
654 const DenseSet<const SCEV *> &VisitedRegs,
656 const SmallVectorImpl<int64_t> &Offsets,
657 ScalarEvolution &SE, DominatorTree &DT);
659 void print(raw_ostream &OS) const;
663 void RateRegister(const SCEV *Reg,
664 SmallPtrSet<const SCEV *, 16> &Regs,
666 ScalarEvolution &SE, DominatorTree &DT);
667 void RatePrimaryRegister(const SCEV *Reg,
668 SmallPtrSet<const SCEV *, 16> &Regs,
670 ScalarEvolution &SE, DominatorTree &DT);
675 /// RateRegister - Tally up interesting quantities from the given register.
676 void Cost::RateRegister(const SCEV *Reg,
677 SmallPtrSet<const SCEV *, 16> &Regs,
679 ScalarEvolution &SE, DominatorTree &DT) {
680 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
681 if (AR->getLoop() == L)
682 AddRecCost += 1; /// TODO: This should be a function of the stride.
684 // If this is an addrec for a loop that's already been visited by LSR,
685 // don't second-guess its addrec phi nodes. LSR isn't currently smart
686 // enough to reason about more than one loop at a time. Consider these
687 // registers free and leave them alone.
688 else if (L->contains(AR->getLoop()) ||
689 (!AR->getLoop()->contains(L) &&
690 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
691 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
692 PHINode *PN = dyn_cast<PHINode>(I); ++I)
693 if (SE.isSCEVable(PN->getType()) &&
694 (SE.getEffectiveSCEVType(PN->getType()) ==
695 SE.getEffectiveSCEVType(AR->getType())) &&
696 SE.getSCEV(PN) == AR)
699 // If this isn't one of the addrecs that the loop already has, it
700 // would require a costly new phi and add. TODO: This isn't
701 // precisely modeled right now.
703 if (!Regs.count(AR->getStart()))
704 RateRegister(AR->getStart(), Regs, L, SE, DT);
707 // Add the step value register, if it needs one.
708 // TODO: The non-affine case isn't precisely modeled here.
709 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
710 if (!Regs.count(AR->getStart()))
711 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
715 // Rough heuristic; favor registers which don't require extra setup
716 // instructions in the preheader.
717 if (!isa<SCEVUnknown>(Reg) &&
718 !isa<SCEVConstant>(Reg) &&
719 !(isa<SCEVAddRecExpr>(Reg) &&
720 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
721 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
725 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
727 void Cost::RatePrimaryRegister(const SCEV *Reg,
728 SmallPtrSet<const SCEV *, 16> &Regs,
730 ScalarEvolution &SE, DominatorTree &DT) {
731 if (Regs.insert(Reg))
732 RateRegister(Reg, Regs, L, SE, DT);
735 void Cost::RateFormula(const Formula &F,
736 SmallPtrSet<const SCEV *, 16> &Regs,
737 const DenseSet<const SCEV *> &VisitedRegs,
739 const SmallVectorImpl<int64_t> &Offsets,
740 ScalarEvolution &SE, DominatorTree &DT) {
741 // Tally up the registers.
742 if (const SCEV *ScaledReg = F.ScaledReg) {
743 if (VisitedRegs.count(ScaledReg)) {
747 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
749 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
750 E = F.BaseRegs.end(); I != E; ++I) {
751 const SCEV *BaseReg = *I;
752 if (VisitedRegs.count(BaseReg)) {
756 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
758 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
759 BaseReg->hasComputableLoopEvolution(L);
762 if (F.BaseRegs.size() > 1)
763 NumBaseAdds += F.BaseRegs.size() - 1;
765 // Tally up the non-zero immediates.
766 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
767 E = Offsets.end(); I != E; ++I) {
768 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
770 ImmCost += 64; // Handle symbolic values conservatively.
771 // TODO: This should probably be the pointer size.
772 else if (Offset != 0)
773 ImmCost += APInt(64, Offset, true).getMinSignedBits();
777 /// Loose - Set this cost to a loosing value.
787 /// operator< - Choose the lower cost.
788 bool Cost::operator<(const Cost &Other) const {
789 if (NumRegs != Other.NumRegs)
790 return NumRegs < Other.NumRegs;
791 if (AddRecCost != Other.AddRecCost)
792 return AddRecCost < Other.AddRecCost;
793 if (NumIVMuls != Other.NumIVMuls)
794 return NumIVMuls < Other.NumIVMuls;
795 if (NumBaseAdds != Other.NumBaseAdds)
796 return NumBaseAdds < Other.NumBaseAdds;
797 if (ImmCost != Other.ImmCost)
798 return ImmCost < Other.ImmCost;
799 if (SetupCost != Other.SetupCost)
800 return SetupCost < Other.SetupCost;
804 void Cost::print(raw_ostream &OS) const {
805 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
807 OS << ", with addrec cost " << AddRecCost;
809 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
810 if (NumBaseAdds != 0)
811 OS << ", plus " << NumBaseAdds << " base add"
812 << (NumBaseAdds == 1 ? "" : "s");
814 OS << ", plus " << ImmCost << " imm cost";
816 OS << ", plus " << SetupCost << " setup cost";
819 void Cost::dump() const {
820 print(errs()); errs() << '\n';
825 /// LSRFixup - An operand value in an instruction which is to be replaced
826 /// with some equivalent, possibly strength-reduced, replacement.
828 /// UserInst - The instruction which will be updated.
829 Instruction *UserInst;
831 /// OperandValToReplace - The operand of the instruction which will
832 /// be replaced. The operand may be used more than once; every instance
833 /// will be replaced.
834 Value *OperandValToReplace;
836 /// PostIncLoops - If this user is to use the post-incremented value of an
837 /// induction variable, this variable is non-null and holds the loop
838 /// associated with the induction variable.
839 PostIncLoopSet PostIncLoops;
841 /// LUIdx - The index of the LSRUse describing the expression which
842 /// this fixup needs, minus an offset (below).
845 /// Offset - A constant offset to be added to the LSRUse expression.
846 /// This allows multiple fixups to share the same LSRUse with different
847 /// offsets, for example in an unrolled loop.
850 bool isUseFullyOutsideLoop(const Loop *L) const;
854 void print(raw_ostream &OS) const;
861 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
863 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
864 /// value outside of the given loop.
865 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
866 // PHI nodes use their value in their incoming blocks.
867 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
868 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
869 if (PN->getIncomingValue(i) == OperandValToReplace &&
870 L->contains(PN->getIncomingBlock(i)))
875 return !L->contains(UserInst);
878 void LSRFixup::print(raw_ostream &OS) const {
880 // Store is common and interesting enough to be worth special-casing.
881 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
883 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
884 } else if (UserInst->getType()->isVoidTy())
885 OS << UserInst->getOpcodeName();
887 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
889 OS << ", OperandValToReplace=";
890 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
892 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
893 E = PostIncLoops.end(); I != E; ++I) {
894 OS << ", PostIncLoop=";
895 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
898 if (LUIdx != ~size_t(0))
899 OS << ", LUIdx=" << LUIdx;
902 OS << ", Offset=" << Offset;
905 void LSRFixup::dump() const {
906 print(errs()); errs() << '\n';
911 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
912 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
913 struct UniquifierDenseMapInfo {
914 static SmallVector<const SCEV *, 2> getEmptyKey() {
915 SmallVector<const SCEV *, 2> V;
916 V.push_back(reinterpret_cast<const SCEV *>(-1));
920 static SmallVector<const SCEV *, 2> getTombstoneKey() {
921 SmallVector<const SCEV *, 2> V;
922 V.push_back(reinterpret_cast<const SCEV *>(-2));
926 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
928 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
929 E = V.end(); I != E; ++I)
930 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
934 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
935 const SmallVector<const SCEV *, 2> &RHS) {
940 /// LSRUse - This class holds the state that LSR keeps for each use in
941 /// IVUsers, as well as uses invented by LSR itself. It includes information
942 /// about what kinds of things can be folded into the user, information about
943 /// the user itself, and information about how the use may be satisfied.
944 /// TODO: Represent multiple users of the same expression in common?
946 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
949 /// KindType - An enum for a kind of use, indicating what types of
950 /// scaled and immediate operands it might support.
952 Basic, ///< A normal use, with no folding.
953 Special, ///< A special case of basic, allowing -1 scales.
954 Address, ///< An address use; folding according to TargetLowering
955 ICmpZero ///< An equality icmp with both operands folded into one.
956 // TODO: Add a generic icmp too?
960 const Type *AccessTy;
962 SmallVector<int64_t, 8> Offsets;
966 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
967 /// LSRUse are outside of the loop, in which case some special-case heuristics
969 bool AllFixupsOutsideLoop;
971 /// WidestFixupType - This records the widest use type for any fixup using
972 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
973 /// max fixup widths to be equivalent, because the narrower one may be relying
974 /// on the implicit truncation to truncate away bogus bits.
975 const Type *WidestFixupType;
977 /// Formulae - A list of ways to build a value that can satisfy this user.
978 /// After the list is populated, one of these is selected heuristically and
979 /// used to formulate a replacement for OperandValToReplace in UserInst.
980 SmallVector<Formula, 12> Formulae;
982 /// Regs - The set of register candidates used by all formulae in this LSRUse.
983 SmallPtrSet<const SCEV *, 4> Regs;
985 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
986 MinOffset(INT64_MAX),
987 MaxOffset(INT64_MIN),
988 AllFixupsOutsideLoop(true),
989 WidestFixupType(0) {}
991 bool HasFormulaWithSameRegs(const Formula &F) const;
992 bool InsertFormula(const Formula &F);
993 void DeleteFormula(Formula &F);
994 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
996 void print(raw_ostream &OS) const;
1002 /// HasFormula - Test whether this use as a formula which has the same
1003 /// registers as the given formula.
1004 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1005 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1006 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1007 // Unstable sort by host order ok, because this is only used for uniquifying.
1008 std::sort(Key.begin(), Key.end());
1009 return Uniquifier.count(Key);
1012 /// InsertFormula - If the given formula has not yet been inserted, add it to
1013 /// the list, and return true. Return false otherwise.
1014 bool LSRUse::InsertFormula(const Formula &F) {
1015 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1016 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1017 // Unstable sort by host order ok, because this is only used for uniquifying.
1018 std::sort(Key.begin(), Key.end());
1020 if (!Uniquifier.insert(Key).second)
1023 // Using a register to hold the value of 0 is not profitable.
1024 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1025 "Zero allocated in a scaled register!");
1027 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1028 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1029 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1032 // Add the formula to the list.
1033 Formulae.push_back(F);
1035 // Record registers now being used by this use.
1036 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1037 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1042 /// DeleteFormula - Remove the given formula from this use's list.
1043 void LSRUse::DeleteFormula(Formula &F) {
1044 if (&F != &Formulae.back())
1045 std::swap(F, Formulae.back());
1046 Formulae.pop_back();
1047 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1050 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1051 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1052 // Now that we've filtered out some formulae, recompute the Regs set.
1053 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1055 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1056 E = Formulae.end(); I != E; ++I) {
1057 const Formula &F = *I;
1058 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1059 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1062 // Update the RegTracker.
1063 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1064 E = OldRegs.end(); I != E; ++I)
1065 if (!Regs.count(*I))
1066 RegUses.DropRegister(*I, LUIdx);
1069 void LSRUse::print(raw_ostream &OS) const {
1070 OS << "LSR Use: Kind=";
1072 case Basic: OS << "Basic"; break;
1073 case Special: OS << "Special"; break;
1074 case ICmpZero: OS << "ICmpZero"; break;
1076 OS << "Address of ";
1077 if (AccessTy->isPointerTy())
1078 OS << "pointer"; // the full pointer type could be really verbose
1083 OS << ", Offsets={";
1084 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1085 E = Offsets.end(); I != E; ++I) {
1087 if (llvm::next(I) != E)
1092 if (AllFixupsOutsideLoop)
1093 OS << ", all-fixups-outside-loop";
1095 if (WidestFixupType)
1096 OS << ", widest fixup type: " << *WidestFixupType;
1099 void LSRUse::dump() const {
1100 print(errs()); errs() << '\n';
1103 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1104 /// be completely folded into the user instruction at isel time. This includes
1105 /// address-mode folding and special icmp tricks.
1106 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1107 LSRUse::KindType Kind, const Type *AccessTy,
1108 const TargetLowering *TLI) {
1110 case LSRUse::Address:
1111 // If we have low-level target information, ask the target if it can
1112 // completely fold this address.
1113 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1115 // Otherwise, just guess that reg+reg addressing is legal.
1116 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1118 case LSRUse::ICmpZero:
1119 // There's not even a target hook for querying whether it would be legal to
1120 // fold a GV into an ICmp.
1124 // ICmp only has two operands; don't allow more than two non-trivial parts.
1125 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1128 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1129 // putting the scaled register in the other operand of the icmp.
1130 if (AM.Scale != 0 && AM.Scale != -1)
1133 // If we have low-level target information, ask the target if it can fold an
1134 // integer immediate on an icmp.
1135 if (AM.BaseOffs != 0) {
1136 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1143 // Only handle single-register values.
1144 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1146 case LSRUse::Special:
1147 // Only handle -1 scales, or no scale.
1148 return AM.Scale == 0 || AM.Scale == -1;
1154 static bool isLegalUse(TargetLowering::AddrMode AM,
1155 int64_t MinOffset, int64_t MaxOffset,
1156 LSRUse::KindType Kind, const Type *AccessTy,
1157 const TargetLowering *TLI) {
1158 // Check for overflow.
1159 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1162 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1163 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1164 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1165 // Check for overflow.
1166 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1169 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1170 return isLegalUse(AM, Kind, AccessTy, TLI);
1175 static bool isAlwaysFoldable(int64_t BaseOffs,
1176 GlobalValue *BaseGV,
1178 LSRUse::KindType Kind, const Type *AccessTy,
1179 const TargetLowering *TLI) {
1180 // Fast-path: zero is always foldable.
1181 if (BaseOffs == 0 && !BaseGV) return true;
1183 // Conservatively, create an address with an immediate and a
1184 // base and a scale.
1185 TargetLowering::AddrMode AM;
1186 AM.BaseOffs = BaseOffs;
1188 AM.HasBaseReg = HasBaseReg;
1189 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1191 // Canonicalize a scale of 1 to a base register if the formula doesn't
1192 // already have a base register.
1193 if (!AM.HasBaseReg && AM.Scale == 1) {
1195 AM.HasBaseReg = true;
1198 return isLegalUse(AM, Kind, AccessTy, TLI);
1201 static bool isAlwaysFoldable(const SCEV *S,
1202 int64_t MinOffset, int64_t MaxOffset,
1204 LSRUse::KindType Kind, const Type *AccessTy,
1205 const TargetLowering *TLI,
1206 ScalarEvolution &SE) {
1207 // Fast-path: zero is always foldable.
1208 if (S->isZero()) return true;
1210 // Conservatively, create an address with an immediate and a
1211 // base and a scale.
1212 int64_t BaseOffs = ExtractImmediate(S, SE);
1213 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1215 // If there's anything else involved, it's not foldable.
1216 if (!S->isZero()) return false;
1218 // Fast-path: zero is always foldable.
1219 if (BaseOffs == 0 && !BaseGV) return true;
1221 // Conservatively, create an address with an immediate and a
1222 // base and a scale.
1223 TargetLowering::AddrMode AM;
1224 AM.BaseOffs = BaseOffs;
1226 AM.HasBaseReg = HasBaseReg;
1227 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1229 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1234 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1235 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1236 struct UseMapDenseMapInfo {
1237 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1238 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1241 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1242 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1246 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1247 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1248 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1252 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1253 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1258 /// FormulaSorter - This class implements an ordering for formulae which sorts
1259 /// the by their standalone cost.
1260 class FormulaSorter {
1261 /// These two sets are kept empty, so that we compute standalone costs.
1262 DenseSet<const SCEV *> VisitedRegs;
1263 SmallPtrSet<const SCEV *, 16> Regs;
1266 ScalarEvolution &SE;
1270 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1271 : L(l), LU(&lu), SE(se), DT(dt) {}
1273 bool operator()(const Formula &A, const Formula &B) {
1275 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1278 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1280 return CostA < CostB;
1284 /// LSRInstance - This class holds state for the main loop strength reduction
1288 ScalarEvolution &SE;
1291 const TargetLowering *const TLI;
1295 /// IVIncInsertPos - This is the insert position that the current loop's
1296 /// induction variable increment should be placed. In simple loops, this is
1297 /// the latch block's terminator. But in more complicated cases, this is a
1298 /// position which will dominate all the in-loop post-increment users.
1299 Instruction *IVIncInsertPos;
1301 /// Factors - Interesting factors between use strides.
1302 SmallSetVector<int64_t, 8> Factors;
1304 /// Types - Interesting use types, to facilitate truncation reuse.
1305 SmallSetVector<const Type *, 4> Types;
1307 /// Fixups - The list of operands which are to be replaced.
1308 SmallVector<LSRFixup, 16> Fixups;
1310 /// Uses - The list of interesting uses.
1311 SmallVector<LSRUse, 16> Uses;
1313 /// RegUses - Track which uses use which register candidates.
1314 RegUseTracker RegUses;
1316 void OptimizeShadowIV();
1317 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1318 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1319 void OptimizeLoopTermCond();
1321 void CollectInterestingTypesAndFactors();
1322 void CollectFixupsAndInitialFormulae();
1324 LSRFixup &getNewFixup() {
1325 Fixups.push_back(LSRFixup());
1326 return Fixups.back();
1329 // Support for sharing of LSRUses between LSRFixups.
1330 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1332 UseMapDenseMapInfo> UseMapTy;
1335 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1336 LSRUse::KindType Kind, const Type *AccessTy);
1338 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1339 LSRUse::KindType Kind,
1340 const Type *AccessTy);
1342 void DeleteUse(LSRUse &LU);
1344 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1347 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1348 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1349 void CountRegisters(const Formula &F, size_t LUIdx);
1350 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1352 void CollectLoopInvariantFixupsAndFormulae();
1354 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1355 unsigned Depth = 0);
1356 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1357 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1358 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1359 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1360 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1361 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1362 void GenerateCrossUseConstantOffsets();
1363 void GenerateAllReuseFormulae();
1365 void FilterOutUndesirableDedicatedRegisters();
1367 size_t EstimateSearchSpaceComplexity() const;
1368 void NarrowSearchSpaceByDetectingSupersets();
1369 void NarrowSearchSpaceByCollapsingUnrolledCode();
1370 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1371 void NarrowSearchSpaceByPickingWinnerRegs();
1372 void NarrowSearchSpaceUsingHeuristics();
1374 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1376 SmallVectorImpl<const Formula *> &Workspace,
1377 const Cost &CurCost,
1378 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1379 DenseSet<const SCEV *> &VisitedRegs) const;
1380 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1382 BasicBlock::iterator
1383 HoistInsertPosition(BasicBlock::iterator IP,
1384 const SmallVectorImpl<Instruction *> &Inputs) const;
1385 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1387 const LSRUse &LU) const;
1389 Value *Expand(const LSRFixup &LF,
1391 BasicBlock::iterator IP,
1392 SCEVExpander &Rewriter,
1393 SmallVectorImpl<WeakVH> &DeadInsts) const;
1394 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1396 SCEVExpander &Rewriter,
1397 SmallVectorImpl<WeakVH> &DeadInsts,
1399 void Rewrite(const LSRFixup &LF,
1401 SCEVExpander &Rewriter,
1402 SmallVectorImpl<WeakVH> &DeadInsts,
1404 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1407 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1409 bool getChanged() const { return Changed; }
1411 void print_factors_and_types(raw_ostream &OS) const;
1412 void print_fixups(raw_ostream &OS) const;
1413 void print_uses(raw_ostream &OS) const;
1414 void print(raw_ostream &OS) const;
1420 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1421 /// inside the loop then try to eliminate the cast operation.
1422 void LSRInstance::OptimizeShadowIV() {
1423 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1424 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1427 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1428 UI != E; /* empty */) {
1429 IVUsers::const_iterator CandidateUI = UI;
1431 Instruction *ShadowUse = CandidateUI->getUser();
1432 const Type *DestTy = NULL;
1434 /* If shadow use is a int->float cast then insert a second IV
1435 to eliminate this cast.
1437 for (unsigned i = 0; i < n; ++i)
1443 for (unsigned i = 0; i < n; ++i, ++d)
1446 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1447 DestTy = UCast->getDestTy();
1448 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1449 DestTy = SCast->getDestTy();
1450 if (!DestTy) continue;
1453 // If target does not support DestTy natively then do not apply
1454 // this transformation.
1455 EVT DVT = TLI->getValueType(DestTy);
1456 if (!TLI->isTypeLegal(DVT)) continue;
1459 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1461 if (PH->getNumIncomingValues() != 2) continue;
1463 const Type *SrcTy = PH->getType();
1464 int Mantissa = DestTy->getFPMantissaWidth();
1465 if (Mantissa == -1) continue;
1466 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1469 unsigned Entry, Latch;
1470 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1478 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1479 if (!Init) continue;
1480 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1482 BinaryOperator *Incr =
1483 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1484 if (!Incr) continue;
1485 if (Incr->getOpcode() != Instruction::Add
1486 && Incr->getOpcode() != Instruction::Sub)
1489 /* Initialize new IV, double d = 0.0 in above example. */
1490 ConstantInt *C = NULL;
1491 if (Incr->getOperand(0) == PH)
1492 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1493 else if (Incr->getOperand(1) == PH)
1494 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1500 // Ignore negative constants, as the code below doesn't handle them
1501 // correctly. TODO: Remove this restriction.
1502 if (!C->getValue().isStrictlyPositive()) continue;
1504 /* Add new PHINode. */
1505 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1507 /* create new increment. '++d' in above example. */
1508 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1509 BinaryOperator *NewIncr =
1510 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1511 Instruction::FAdd : Instruction::FSub,
1512 NewPH, CFP, "IV.S.next.", Incr);
1514 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1515 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1517 /* Remove cast operation */
1518 ShadowUse->replaceAllUsesWith(NewPH);
1519 ShadowUse->eraseFromParent();
1525 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1526 /// set the IV user and stride information and return true, otherwise return
1528 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1529 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1530 if (UI->getUser() == Cond) {
1531 // NOTE: we could handle setcc instructions with multiple uses here, but
1532 // InstCombine does it as well for simple uses, it's not clear that it
1533 // occurs enough in real life to handle.
1540 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1541 /// a max computation.
1543 /// This is a narrow solution to a specific, but acute, problem. For loops
1549 /// } while (++i < n);
1551 /// the trip count isn't just 'n', because 'n' might not be positive. And
1552 /// unfortunately this can come up even for loops where the user didn't use
1553 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1554 /// will commonly be lowered like this:
1560 /// } while (++i < n);
1563 /// and then it's possible for subsequent optimization to obscure the if
1564 /// test in such a way that indvars can't find it.
1566 /// When indvars can't find the if test in loops like this, it creates a
1567 /// max expression, which allows it to give the loop a canonical
1568 /// induction variable:
1571 /// max = n < 1 ? 1 : n;
1574 /// } while (++i != max);
1576 /// Canonical induction variables are necessary because the loop passes
1577 /// are designed around them. The most obvious example of this is the
1578 /// LoopInfo analysis, which doesn't remember trip count values. It
1579 /// expects to be able to rediscover the trip count each time it is
1580 /// needed, and it does this using a simple analysis that only succeeds if
1581 /// the loop has a canonical induction variable.
1583 /// However, when it comes time to generate code, the maximum operation
1584 /// can be quite costly, especially if it's inside of an outer loop.
1586 /// This function solves this problem by detecting this type of loop and
1587 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1588 /// the instructions for the maximum computation.
1590 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1591 // Check that the loop matches the pattern we're looking for.
1592 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1593 Cond->getPredicate() != CmpInst::ICMP_NE)
1596 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1597 if (!Sel || !Sel->hasOneUse()) return Cond;
1599 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1600 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1602 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1604 // Add one to the backedge-taken count to get the trip count.
1605 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1606 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1608 // Check for a max calculation that matches the pattern. There's no check
1609 // for ICMP_ULE here because the comparison would be with zero, which
1610 // isn't interesting.
1611 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1612 const SCEVNAryExpr *Max = 0;
1613 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1614 Pred = ICmpInst::ICMP_SLE;
1616 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1617 Pred = ICmpInst::ICMP_SLT;
1619 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1620 Pred = ICmpInst::ICMP_ULT;
1627 // To handle a max with more than two operands, this optimization would
1628 // require additional checking and setup.
1629 if (Max->getNumOperands() != 2)
1632 const SCEV *MaxLHS = Max->getOperand(0);
1633 const SCEV *MaxRHS = Max->getOperand(1);
1635 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1636 // for a comparison with 1. For <= and >=, a comparison with zero.
1638 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1641 // Check the relevant induction variable for conformance to
1643 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1644 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1645 if (!AR || !AR->isAffine() ||
1646 AR->getStart() != One ||
1647 AR->getStepRecurrence(SE) != One)
1650 assert(AR->getLoop() == L &&
1651 "Loop condition operand is an addrec in a different loop!");
1653 // Check the right operand of the select, and remember it, as it will
1654 // be used in the new comparison instruction.
1656 if (ICmpInst::isTrueWhenEqual(Pred)) {
1657 // Look for n+1, and grab n.
1658 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1659 if (isa<ConstantInt>(BO->getOperand(1)) &&
1660 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1661 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1662 NewRHS = BO->getOperand(0);
1663 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1664 if (isa<ConstantInt>(BO->getOperand(1)) &&
1665 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1666 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1667 NewRHS = BO->getOperand(0);
1670 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1671 NewRHS = Sel->getOperand(1);
1672 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1673 NewRHS = Sel->getOperand(2);
1674 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1675 NewRHS = SU->getValue();
1677 // Max doesn't match expected pattern.
1680 // Determine the new comparison opcode. It may be signed or unsigned,
1681 // and the original comparison may be either equality or inequality.
1682 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1683 Pred = CmpInst::getInversePredicate(Pred);
1685 // Ok, everything looks ok to change the condition into an SLT or SGE and
1686 // delete the max calculation.
1688 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1690 // Delete the max calculation instructions.
1691 Cond->replaceAllUsesWith(NewCond);
1692 CondUse->setUser(NewCond);
1693 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1694 Cond->eraseFromParent();
1695 Sel->eraseFromParent();
1696 if (Cmp->use_empty())
1697 Cmp->eraseFromParent();
1701 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1702 /// postinc iv when possible.
1704 LSRInstance::OptimizeLoopTermCond() {
1705 SmallPtrSet<Instruction *, 4> PostIncs;
1707 BasicBlock *LatchBlock = L->getLoopLatch();
1708 SmallVector<BasicBlock*, 8> ExitingBlocks;
1709 L->getExitingBlocks(ExitingBlocks);
1711 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1712 BasicBlock *ExitingBlock = ExitingBlocks[i];
1714 // Get the terminating condition for the loop if possible. If we
1715 // can, we want to change it to use a post-incremented version of its
1716 // induction variable, to allow coalescing the live ranges for the IV into
1717 // one register value.
1719 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1722 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1723 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1726 // Search IVUsesByStride to find Cond's IVUse if there is one.
1727 IVStrideUse *CondUse = 0;
1728 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1729 if (!FindIVUserForCond(Cond, CondUse))
1732 // If the trip count is computed in terms of a max (due to ScalarEvolution
1733 // being unable to find a sufficient guard, for example), change the loop
1734 // comparison to use SLT or ULT instead of NE.
1735 // One consequence of doing this now is that it disrupts the count-down
1736 // optimization. That's not always a bad thing though, because in such
1737 // cases it may still be worthwhile to avoid a max.
1738 Cond = OptimizeMax(Cond, CondUse);
1740 // If this exiting block dominates the latch block, it may also use
1741 // the post-inc value if it won't be shared with other uses.
1742 // Check for dominance.
1743 if (!DT.dominates(ExitingBlock, LatchBlock))
1746 // Conservatively avoid trying to use the post-inc value in non-latch
1747 // exits if there may be pre-inc users in intervening blocks.
1748 if (LatchBlock != ExitingBlock)
1749 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1750 // Test if the use is reachable from the exiting block. This dominator
1751 // query is a conservative approximation of reachability.
1752 if (&*UI != CondUse &&
1753 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1754 // Conservatively assume there may be reuse if the quotient of their
1755 // strides could be a legal scale.
1756 const SCEV *A = IU.getStride(*CondUse, L);
1757 const SCEV *B = IU.getStride(*UI, L);
1758 if (!A || !B) continue;
1759 if (SE.getTypeSizeInBits(A->getType()) !=
1760 SE.getTypeSizeInBits(B->getType())) {
1761 if (SE.getTypeSizeInBits(A->getType()) >
1762 SE.getTypeSizeInBits(B->getType()))
1763 B = SE.getSignExtendExpr(B, A->getType());
1765 A = SE.getSignExtendExpr(A, B->getType());
1767 if (const SCEVConstant *D =
1768 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1769 const ConstantInt *C = D->getValue();
1770 // Stride of one or negative one can have reuse with non-addresses.
1771 if (C->isOne() || C->isAllOnesValue())
1772 goto decline_post_inc;
1773 // Avoid weird situations.
1774 if (C->getValue().getMinSignedBits() >= 64 ||
1775 C->getValue().isMinSignedValue())
1776 goto decline_post_inc;
1777 // Without TLI, assume that any stride might be valid, and so any
1778 // use might be shared.
1780 goto decline_post_inc;
1781 // Check for possible scaled-address reuse.
1782 const Type *AccessTy = getAccessType(UI->getUser());
1783 TargetLowering::AddrMode AM;
1784 AM.Scale = C->getSExtValue();
1785 if (TLI->isLegalAddressingMode(AM, AccessTy))
1786 goto decline_post_inc;
1787 AM.Scale = -AM.Scale;
1788 if (TLI->isLegalAddressingMode(AM, AccessTy))
1789 goto decline_post_inc;
1793 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1796 // It's possible for the setcc instruction to be anywhere in the loop, and
1797 // possible for it to have multiple users. If it is not immediately before
1798 // the exiting block branch, move it.
1799 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1800 if (Cond->hasOneUse()) {
1801 Cond->moveBefore(TermBr);
1803 // Clone the terminating condition and insert into the loopend.
1804 ICmpInst *OldCond = Cond;
1805 Cond = cast<ICmpInst>(Cond->clone());
1806 Cond->setName(L->getHeader()->getName() + ".termcond");
1807 ExitingBlock->getInstList().insert(TermBr, Cond);
1809 // Clone the IVUse, as the old use still exists!
1810 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1811 TermBr->replaceUsesOfWith(OldCond, Cond);
1815 // If we get to here, we know that we can transform the setcc instruction to
1816 // use the post-incremented version of the IV, allowing us to coalesce the
1817 // live ranges for the IV correctly.
1818 CondUse->transformToPostInc(L);
1821 PostIncs.insert(Cond);
1825 // Determine an insertion point for the loop induction variable increment. It
1826 // must dominate all the post-inc comparisons we just set up, and it must
1827 // dominate the loop latch edge.
1828 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1829 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1830 E = PostIncs.end(); I != E; ++I) {
1832 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1834 if (BB == (*I)->getParent())
1835 IVIncInsertPos = *I;
1836 else if (BB != IVIncInsertPos->getParent())
1837 IVIncInsertPos = BB->getTerminator();
1841 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1842 /// at the given offset and other details. If so, update the use and
1845 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1846 LSRUse::KindType Kind, const Type *AccessTy) {
1847 int64_t NewMinOffset = LU.MinOffset;
1848 int64_t NewMaxOffset = LU.MaxOffset;
1849 const Type *NewAccessTy = AccessTy;
1851 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1852 // something conservative, however this can pessimize in the case that one of
1853 // the uses will have all its uses outside the loop, for example.
1854 if (LU.Kind != Kind)
1856 // Conservatively assume HasBaseReg is true for now.
1857 if (NewOffset < LU.MinOffset) {
1858 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1859 Kind, AccessTy, TLI))
1861 NewMinOffset = NewOffset;
1862 } else if (NewOffset > LU.MaxOffset) {
1863 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1864 Kind, AccessTy, TLI))
1866 NewMaxOffset = NewOffset;
1868 // Check for a mismatched access type, and fall back conservatively as needed.
1869 // TODO: Be less conservative when the type is similar and can use the same
1870 // addressing modes.
1871 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1872 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1875 LU.MinOffset = NewMinOffset;
1876 LU.MaxOffset = NewMaxOffset;
1877 LU.AccessTy = NewAccessTy;
1878 if (NewOffset != LU.Offsets.back())
1879 LU.Offsets.push_back(NewOffset);
1883 /// getUse - Return an LSRUse index and an offset value for a fixup which
1884 /// needs the given expression, with the given kind and optional access type.
1885 /// Either reuse an existing use or create a new one, as needed.
1886 std::pair<size_t, int64_t>
1887 LSRInstance::getUse(const SCEV *&Expr,
1888 LSRUse::KindType Kind, const Type *AccessTy) {
1889 const SCEV *Copy = Expr;
1890 int64_t Offset = ExtractImmediate(Expr, SE);
1892 // Basic uses can't accept any offset, for example.
1893 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1898 std::pair<UseMapTy::iterator, bool> P =
1899 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1901 // A use already existed with this base.
1902 size_t LUIdx = P.first->second;
1903 LSRUse &LU = Uses[LUIdx];
1904 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1906 return std::make_pair(LUIdx, Offset);
1909 // Create a new use.
1910 size_t LUIdx = Uses.size();
1911 P.first->second = LUIdx;
1912 Uses.push_back(LSRUse(Kind, AccessTy));
1913 LSRUse &LU = Uses[LUIdx];
1915 // We don't need to track redundant offsets, but we don't need to go out
1916 // of our way here to avoid them.
1917 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1918 LU.Offsets.push_back(Offset);
1920 LU.MinOffset = Offset;
1921 LU.MaxOffset = Offset;
1922 return std::make_pair(LUIdx, Offset);
1925 /// DeleteUse - Delete the given use from the Uses list.
1926 void LSRInstance::DeleteUse(LSRUse &LU) {
1927 if (&LU != &Uses.back())
1928 std::swap(LU, Uses.back());
1932 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1933 /// a formula that has the same registers as the given formula.
1935 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1936 const LSRUse &OrigLU) {
1937 // Search all uses for the formula. This could be more clever.
1938 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1939 LSRUse &LU = Uses[LUIdx];
1940 // Check whether this use is close enough to OrigLU, to see whether it's
1941 // worthwhile looking through its formulae.
1942 // Ignore ICmpZero uses because they may contain formulae generated by
1943 // GenerateICmpZeroScales, in which case adding fixup offsets may
1945 if (&LU != &OrigLU &&
1946 LU.Kind != LSRUse::ICmpZero &&
1947 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1948 LU.WidestFixupType == OrigLU.WidestFixupType &&
1949 LU.HasFormulaWithSameRegs(OrigF)) {
1950 // Scan through this use's formulae.
1951 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1952 E = LU.Formulae.end(); I != E; ++I) {
1953 const Formula &F = *I;
1954 // Check to see if this formula has the same registers and symbols
1956 if (F.BaseRegs == OrigF.BaseRegs &&
1957 F.ScaledReg == OrigF.ScaledReg &&
1958 F.AM.BaseGV == OrigF.AM.BaseGV &&
1959 F.AM.Scale == OrigF.AM.Scale) {
1960 if (F.AM.BaseOffs == 0)
1962 // This is the formula where all the registers and symbols matched;
1963 // there aren't going to be any others. Since we declined it, we
1964 // can skip the rest of the formulae and procede to the next LSRUse.
1971 // Nothing looked good.
1975 void LSRInstance::CollectInterestingTypesAndFactors() {
1976 SmallSetVector<const SCEV *, 4> Strides;
1978 // Collect interesting types and strides.
1979 SmallVector<const SCEV *, 4> Worklist;
1980 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1981 const SCEV *Expr = IU.getExpr(*UI);
1983 // Collect interesting types.
1984 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1986 // Add strides for mentioned loops.
1987 Worklist.push_back(Expr);
1989 const SCEV *S = Worklist.pop_back_val();
1990 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1991 Strides.insert(AR->getStepRecurrence(SE));
1992 Worklist.push_back(AR->getStart());
1993 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1994 Worklist.append(Add->op_begin(), Add->op_end());
1996 } while (!Worklist.empty());
1999 // Compute interesting factors from the set of interesting strides.
2000 for (SmallSetVector<const SCEV *, 4>::const_iterator
2001 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2002 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2003 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2004 const SCEV *OldStride = *I;
2005 const SCEV *NewStride = *NewStrideIter;
2007 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2008 SE.getTypeSizeInBits(NewStride->getType())) {
2009 if (SE.getTypeSizeInBits(OldStride->getType()) >
2010 SE.getTypeSizeInBits(NewStride->getType()))
2011 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2013 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2015 if (const SCEVConstant *Factor =
2016 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2018 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2019 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2020 } else if (const SCEVConstant *Factor =
2021 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2024 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2025 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2029 // If all uses use the same type, don't bother looking for truncation-based
2031 if (Types.size() == 1)
2034 DEBUG(print_factors_and_types(dbgs()));
2037 void LSRInstance::CollectFixupsAndInitialFormulae() {
2038 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2040 LSRFixup &LF = getNewFixup();
2041 LF.UserInst = UI->getUser();
2042 LF.OperandValToReplace = UI->getOperandValToReplace();
2043 LF.PostIncLoops = UI->getPostIncLoops();
2045 LSRUse::KindType Kind = LSRUse::Basic;
2046 const Type *AccessTy = 0;
2047 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2048 Kind = LSRUse::Address;
2049 AccessTy = getAccessType(LF.UserInst);
2052 const SCEV *S = IU.getExpr(*UI);
2054 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2055 // (N - i == 0), and this allows (N - i) to be the expression that we work
2056 // with rather than just N or i, so we can consider the register
2057 // requirements for both N and i at the same time. Limiting this code to
2058 // equality icmps is not a problem because all interesting loops use
2059 // equality icmps, thanks to IndVarSimplify.
2060 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2061 if (CI->isEquality()) {
2062 // Swap the operands if needed to put the OperandValToReplace on the
2063 // left, for consistency.
2064 Value *NV = CI->getOperand(1);
2065 if (NV == LF.OperandValToReplace) {
2066 CI->setOperand(1, CI->getOperand(0));
2067 CI->setOperand(0, NV);
2068 NV = CI->getOperand(1);
2072 // x == y --> x - y == 0
2073 const SCEV *N = SE.getSCEV(NV);
2074 if (N->isLoopInvariant(L)) {
2075 Kind = LSRUse::ICmpZero;
2076 S = SE.getMinusSCEV(N, S);
2079 // -1 and the negations of all interesting strides (except the negation
2080 // of -1) are now also interesting.
2081 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2082 if (Factors[i] != -1)
2083 Factors.insert(-(uint64_t)Factors[i]);
2087 // Set up the initial formula for this use.
2088 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2090 LF.Offset = P.second;
2091 LSRUse &LU = Uses[LF.LUIdx];
2092 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2093 if (!LU.WidestFixupType ||
2094 SE.getTypeSizeInBits(LU.WidestFixupType) <
2095 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2096 LU.WidestFixupType = LF.OperandValToReplace->getType();
2098 // If this is the first use of this LSRUse, give it a formula.
2099 if (LU.Formulae.empty()) {
2100 InsertInitialFormula(S, LU, LF.LUIdx);
2101 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2105 DEBUG(print_fixups(dbgs()));
2108 /// InsertInitialFormula - Insert a formula for the given expression into
2109 /// the given use, separating out loop-variant portions from loop-invariant
2110 /// and loop-computable portions.
2112 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2114 F.InitialMatch(S, L, SE, DT);
2115 bool Inserted = InsertFormula(LU, LUIdx, F);
2116 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2119 /// InsertSupplementalFormula - Insert a simple single-register formula for
2120 /// the given expression into the given use.
2122 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2123 LSRUse &LU, size_t LUIdx) {
2125 F.BaseRegs.push_back(S);
2126 F.AM.HasBaseReg = true;
2127 bool Inserted = InsertFormula(LU, LUIdx, F);
2128 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2131 /// CountRegisters - Note which registers are used by the given formula,
2132 /// updating RegUses.
2133 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2135 RegUses.CountRegister(F.ScaledReg, LUIdx);
2136 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2137 E = F.BaseRegs.end(); I != E; ++I)
2138 RegUses.CountRegister(*I, LUIdx);
2141 /// InsertFormula - If the given formula has not yet been inserted, add it to
2142 /// the list, and return true. Return false otherwise.
2143 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2144 if (!LU.InsertFormula(F))
2147 CountRegisters(F, LUIdx);
2151 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2152 /// loop-invariant values which we're tracking. These other uses will pin these
2153 /// values in registers, making them less profitable for elimination.
2154 /// TODO: This currently misses non-constant addrec step registers.
2155 /// TODO: Should this give more weight to users inside the loop?
2157 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2158 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2159 SmallPtrSet<const SCEV *, 8> Inserted;
2161 while (!Worklist.empty()) {
2162 const SCEV *S = Worklist.pop_back_val();
2164 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2165 Worklist.append(N->op_begin(), N->op_end());
2166 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2167 Worklist.push_back(C->getOperand());
2168 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2169 Worklist.push_back(D->getLHS());
2170 Worklist.push_back(D->getRHS());
2171 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2172 if (!Inserted.insert(U)) continue;
2173 const Value *V = U->getValue();
2174 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2175 // Look for instructions defined outside the loop.
2176 if (L->contains(Inst)) continue;
2177 } else if (isa<UndefValue>(V))
2178 // Undef doesn't have a live range, so it doesn't matter.
2180 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2182 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2183 // Ignore non-instructions.
2186 // Ignore instructions in other functions (as can happen with
2188 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2190 // Ignore instructions not dominated by the loop.
2191 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2192 UserInst->getParent() :
2193 cast<PHINode>(UserInst)->getIncomingBlock(
2194 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2195 if (!DT.dominates(L->getHeader(), UseBB))
2197 // Ignore uses which are part of other SCEV expressions, to avoid
2198 // analyzing them multiple times.
2199 if (SE.isSCEVable(UserInst->getType())) {
2200 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2201 // If the user is a no-op, look through to its uses.
2202 if (!isa<SCEVUnknown>(UserS))
2206 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2210 // Ignore icmp instructions which are already being analyzed.
2211 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2212 unsigned OtherIdx = !UI.getOperandNo();
2213 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2214 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2218 LSRFixup &LF = getNewFixup();
2219 LF.UserInst = const_cast<Instruction *>(UserInst);
2220 LF.OperandValToReplace = UI.getUse();
2221 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2223 LF.Offset = P.second;
2224 LSRUse &LU = Uses[LF.LUIdx];
2225 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2226 if (!LU.WidestFixupType ||
2227 SE.getTypeSizeInBits(LU.WidestFixupType) <
2228 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2229 LU.WidestFixupType = LF.OperandValToReplace->getType();
2230 InsertSupplementalFormula(U, LU, LF.LUIdx);
2231 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2238 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2239 /// separate registers. If C is non-null, multiply each subexpression by C.
2240 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2241 SmallVectorImpl<const SCEV *> &Ops,
2243 ScalarEvolution &SE) {
2244 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2245 // Break out add operands.
2246 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2248 CollectSubexprs(*I, C, Ops, L, SE);
2250 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2251 // Split a non-zero base out of an addrec.
2252 if (!AR->getStart()->isZero()) {
2253 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2254 AR->getStepRecurrence(SE),
2257 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2260 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2261 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2262 if (Mul->getNumOperands() == 2)
2263 if (const SCEVConstant *Op0 =
2264 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2265 CollectSubexprs(Mul->getOperand(1),
2266 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2272 // Otherwise use the value itself, optionally with a scale applied.
2273 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2276 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2278 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2281 // Arbitrarily cap recursion to protect compile time.
2282 if (Depth >= 3) return;
2284 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2285 const SCEV *BaseReg = Base.BaseRegs[i];
2287 SmallVector<const SCEV *, 8> AddOps;
2288 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2290 if (AddOps.size() == 1) continue;
2292 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2293 JE = AddOps.end(); J != JE; ++J) {
2295 // Loop-variant "unknown" values are uninteresting; we won't be able to
2296 // do anything meaningful with them.
2297 if (isa<SCEVUnknown>(*J) && !(*J)->isLoopInvariant(L))
2300 // Don't pull a constant into a register if the constant could be folded
2301 // into an immediate field.
2302 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2303 Base.getNumRegs() > 1,
2304 LU.Kind, LU.AccessTy, TLI, SE))
2307 // Collect all operands except *J.
2308 SmallVector<const SCEV *, 8> InnerAddOps
2309 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2311 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2313 // Don't leave just a constant behind in a register if the constant could
2314 // be folded into an immediate field.
2315 if (InnerAddOps.size() == 1 &&
2316 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2317 Base.getNumRegs() > 1,
2318 LU.Kind, LU.AccessTy, TLI, SE))
2321 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2322 if (InnerSum->isZero())
2325 F.BaseRegs[i] = InnerSum;
2326 F.BaseRegs.push_back(*J);
2327 if (InsertFormula(LU, LUIdx, F))
2328 // If that formula hadn't been seen before, recurse to find more like
2330 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2335 /// GenerateCombinations - Generate a formula consisting of all of the
2336 /// loop-dominating registers added into a single register.
2337 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2339 // This method is only interesting on a plurality of registers.
2340 if (Base.BaseRegs.size() <= 1) return;
2344 SmallVector<const SCEV *, 4> Ops;
2345 for (SmallVectorImpl<const SCEV *>::const_iterator
2346 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2347 const SCEV *BaseReg = *I;
2348 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2349 !BaseReg->hasComputableLoopEvolution(L))
2350 Ops.push_back(BaseReg);
2352 F.BaseRegs.push_back(BaseReg);
2354 if (Ops.size() > 1) {
2355 const SCEV *Sum = SE.getAddExpr(Ops);
2356 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2357 // opportunity to fold something. For now, just ignore such cases
2358 // rather than proceed with zero in a register.
2359 if (!Sum->isZero()) {
2360 F.BaseRegs.push_back(Sum);
2361 (void)InsertFormula(LU, LUIdx, F);
2366 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2367 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2369 // We can't add a symbolic offset if the address already contains one.
2370 if (Base.AM.BaseGV) return;
2372 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2373 const SCEV *G = Base.BaseRegs[i];
2374 GlobalValue *GV = ExtractSymbol(G, SE);
2375 if (G->isZero() || !GV)
2379 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2380 LU.Kind, LU.AccessTy, TLI))
2383 (void)InsertFormula(LU, LUIdx, F);
2387 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2388 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2390 // TODO: For now, just add the min and max offset, because it usually isn't
2391 // worthwhile looking at everything inbetween.
2392 SmallVector<int64_t, 2> Worklist;
2393 Worklist.push_back(LU.MinOffset);
2394 if (LU.MaxOffset != LU.MinOffset)
2395 Worklist.push_back(LU.MaxOffset);
2397 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2398 const SCEV *G = Base.BaseRegs[i];
2400 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2401 E = Worklist.end(); I != E; ++I) {
2403 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2404 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2405 LU.Kind, LU.AccessTy, TLI)) {
2406 // Add the offset to the base register.
2407 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2408 // If it cancelled out, drop the base register, otherwise update it.
2409 if (NewG->isZero()) {
2410 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2411 F.BaseRegs.pop_back();
2413 F.BaseRegs[i] = NewG;
2415 (void)InsertFormula(LU, LUIdx, F);
2419 int64_t Imm = ExtractImmediate(G, SE);
2420 if (G->isZero() || Imm == 0)
2423 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2424 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2425 LU.Kind, LU.AccessTy, TLI))
2428 (void)InsertFormula(LU, LUIdx, F);
2432 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2433 /// the comparison. For example, x == y -> x*c == y*c.
2434 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2436 if (LU.Kind != LSRUse::ICmpZero) return;
2438 // Determine the integer type for the base formula.
2439 const Type *IntTy = Base.getType();
2441 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2443 // Don't do this if there is more than one offset.
2444 if (LU.MinOffset != LU.MaxOffset) return;
2446 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2448 // Check each interesting stride.
2449 for (SmallSetVector<int64_t, 8>::const_iterator
2450 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2451 int64_t Factor = *I;
2453 // Check that the multiplication doesn't overflow.
2454 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2456 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2457 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2460 // Check that multiplying with the use offset doesn't overflow.
2461 int64_t Offset = LU.MinOffset;
2462 if (Offset == INT64_MIN && Factor == -1)
2464 Offset = (uint64_t)Offset * Factor;
2465 if (Offset / Factor != LU.MinOffset)
2469 F.AM.BaseOffs = NewBaseOffs;
2471 // Check that this scale is legal.
2472 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2475 // Compensate for the use having MinOffset built into it.
2476 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2478 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2480 // Check that multiplying with each base register doesn't overflow.
2481 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2482 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2483 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2487 // Check that multiplying with the scaled register doesn't overflow.
2489 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2490 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2494 // If we make it here and it's legal, add it.
2495 (void)InsertFormula(LU, LUIdx, F);
2500 /// GenerateScales - Generate stride factor reuse formulae by making use of
2501 /// scaled-offset address modes, for example.
2502 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2503 // Determine the integer type for the base formula.
2504 const Type *IntTy = Base.getType();
2507 // If this Formula already has a scaled register, we can't add another one.
2508 if (Base.AM.Scale != 0) return;
2510 // Check each interesting stride.
2511 for (SmallSetVector<int64_t, 8>::const_iterator
2512 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2513 int64_t Factor = *I;
2515 Base.AM.Scale = Factor;
2516 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2517 // Check whether this scale is going to be legal.
2518 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2519 LU.Kind, LU.AccessTy, TLI)) {
2520 // As a special-case, handle special out-of-loop Basic users specially.
2521 // TODO: Reconsider this special case.
2522 if (LU.Kind == LSRUse::Basic &&
2523 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2524 LSRUse::Special, LU.AccessTy, TLI) &&
2525 LU.AllFixupsOutsideLoop)
2526 LU.Kind = LSRUse::Special;
2530 // For an ICmpZero, negating a solitary base register won't lead to
2532 if (LU.Kind == LSRUse::ICmpZero &&
2533 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2535 // For each addrec base reg, apply the scale, if possible.
2536 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2537 if (const SCEVAddRecExpr *AR =
2538 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2539 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2540 if (FactorS->isZero())
2542 // Divide out the factor, ignoring high bits, since we'll be
2543 // scaling the value back up in the end.
2544 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2545 // TODO: This could be optimized to avoid all the copying.
2547 F.ScaledReg = Quotient;
2548 F.DeleteBaseReg(F.BaseRegs[i]);
2549 (void)InsertFormula(LU, LUIdx, F);
2555 /// GenerateTruncates - Generate reuse formulae from different IV types.
2556 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2557 // This requires TargetLowering to tell us which truncates are free.
2560 // Don't bother truncating symbolic values.
2561 if (Base.AM.BaseGV) return;
2563 // Determine the integer type for the base formula.
2564 const Type *DstTy = Base.getType();
2566 DstTy = SE.getEffectiveSCEVType(DstTy);
2568 for (SmallSetVector<const Type *, 4>::const_iterator
2569 I = Types.begin(), E = Types.end(); I != E; ++I) {
2570 const Type *SrcTy = *I;
2571 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2574 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2575 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2576 JE = F.BaseRegs.end(); J != JE; ++J)
2577 *J = SE.getAnyExtendExpr(*J, SrcTy);
2579 // TODO: This assumes we've done basic processing on all uses and
2580 // have an idea what the register usage is.
2581 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2584 (void)InsertFormula(LU, LUIdx, F);
2591 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2592 /// defer modifications so that the search phase doesn't have to worry about
2593 /// the data structures moving underneath it.
2597 const SCEV *OrigReg;
2599 WorkItem(size_t LI, int64_t I, const SCEV *R)
2600 : LUIdx(LI), Imm(I), OrigReg(R) {}
2602 void print(raw_ostream &OS) const;
2608 void WorkItem::print(raw_ostream &OS) const {
2609 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2610 << " , add offset " << Imm;
2613 void WorkItem::dump() const {
2614 print(errs()); errs() << '\n';
2617 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2618 /// distance apart and try to form reuse opportunities between them.
2619 void LSRInstance::GenerateCrossUseConstantOffsets() {
2620 // Group the registers by their value without any added constant offset.
2621 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2622 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2624 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2625 SmallVector<const SCEV *, 8> Sequence;
2626 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2628 const SCEV *Reg = *I;
2629 int64_t Imm = ExtractImmediate(Reg, SE);
2630 std::pair<RegMapTy::iterator, bool> Pair =
2631 Map.insert(std::make_pair(Reg, ImmMapTy()));
2633 Sequence.push_back(Reg);
2634 Pair.first->second.insert(std::make_pair(Imm, *I));
2635 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2638 // Now examine each set of registers with the same base value. Build up
2639 // a list of work to do and do the work in a separate step so that we're
2640 // not adding formulae and register counts while we're searching.
2641 SmallVector<WorkItem, 32> WorkItems;
2642 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2643 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2644 E = Sequence.end(); I != E; ++I) {
2645 const SCEV *Reg = *I;
2646 const ImmMapTy &Imms = Map.find(Reg)->second;
2648 // It's not worthwhile looking for reuse if there's only one offset.
2649 if (Imms.size() == 1)
2652 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2653 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2655 dbgs() << ' ' << J->first;
2658 // Examine each offset.
2659 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2661 const SCEV *OrigReg = J->second;
2663 int64_t JImm = J->first;
2664 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2666 if (!isa<SCEVConstant>(OrigReg) &&
2667 UsedByIndicesMap[Reg].count() == 1) {
2668 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2672 // Conservatively examine offsets between this orig reg a few selected
2674 ImmMapTy::const_iterator OtherImms[] = {
2675 Imms.begin(), prior(Imms.end()),
2676 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2678 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2679 ImmMapTy::const_iterator M = OtherImms[i];
2680 if (M == J || M == JE) continue;
2682 // Compute the difference between the two.
2683 int64_t Imm = (uint64_t)JImm - M->first;
2684 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2685 LUIdx = UsedByIndices.find_next(LUIdx))
2686 // Make a memo of this use, offset, and register tuple.
2687 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2688 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2695 UsedByIndicesMap.clear();
2696 UniqueItems.clear();
2698 // Now iterate through the worklist and add new formulae.
2699 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2700 E = WorkItems.end(); I != E; ++I) {
2701 const WorkItem &WI = *I;
2702 size_t LUIdx = WI.LUIdx;
2703 LSRUse &LU = Uses[LUIdx];
2704 int64_t Imm = WI.Imm;
2705 const SCEV *OrigReg = WI.OrigReg;
2707 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2708 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2709 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2711 // TODO: Use a more targeted data structure.
2712 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2713 const Formula &F = LU.Formulae[L];
2714 // Use the immediate in the scaled register.
2715 if (F.ScaledReg == OrigReg) {
2716 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2717 Imm * (uint64_t)F.AM.Scale;
2718 // Don't create 50 + reg(-50).
2719 if (F.referencesReg(SE.getSCEV(
2720 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2723 NewF.AM.BaseOffs = Offs;
2724 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2725 LU.Kind, LU.AccessTy, TLI))
2727 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2729 // If the new scale is a constant in a register, and adding the constant
2730 // value to the immediate would produce a value closer to zero than the
2731 // immediate itself, then the formula isn't worthwhile.
2732 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2733 if (C->getValue()->getValue().isNegative() !=
2734 (NewF.AM.BaseOffs < 0) &&
2735 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2736 .ule(abs64(NewF.AM.BaseOffs)))
2740 (void)InsertFormula(LU, LUIdx, NewF);
2742 // Use the immediate in a base register.
2743 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2744 const SCEV *BaseReg = F.BaseRegs[N];
2745 if (BaseReg != OrigReg)
2748 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2749 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2750 LU.Kind, LU.AccessTy, TLI))
2752 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2754 // If the new formula has a constant in a register, and adding the
2755 // constant value to the immediate would produce a value closer to
2756 // zero than the immediate itself, then the formula isn't worthwhile.
2757 for (SmallVectorImpl<const SCEV *>::const_iterator
2758 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2760 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2761 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2762 abs64(NewF.AM.BaseOffs)) &&
2763 (C->getValue()->getValue() +
2764 NewF.AM.BaseOffs).countTrailingZeros() >=
2765 CountTrailingZeros_64(NewF.AM.BaseOffs))
2769 (void)InsertFormula(LU, LUIdx, NewF);
2778 /// GenerateAllReuseFormulae - Generate formulae for each use.
2780 LSRInstance::GenerateAllReuseFormulae() {
2781 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2782 // queries are more precise.
2783 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2784 LSRUse &LU = Uses[LUIdx];
2785 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2786 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2787 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2788 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2790 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2791 LSRUse &LU = Uses[LUIdx];
2792 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2793 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2794 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2795 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2796 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2797 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2798 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2799 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2801 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2802 LSRUse &LU = Uses[LUIdx];
2803 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2804 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2807 GenerateCrossUseConstantOffsets();
2809 DEBUG(dbgs() << "\n"
2810 "After generating reuse formulae:\n";
2811 print_uses(dbgs()));
2814 /// If their are multiple formulae with the same set of registers used
2815 /// by other uses, pick the best one and delete the others.
2816 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2818 bool ChangedFormulae = false;
2821 // Collect the best formula for each unique set of shared registers. This
2822 // is reset for each use.
2823 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2825 BestFormulaeTy BestFormulae;
2827 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2828 LSRUse &LU = Uses[LUIdx];
2829 FormulaSorter Sorter(L, LU, SE, DT);
2830 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2833 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2834 FIdx != NumForms; ++FIdx) {
2835 Formula &F = LU.Formulae[FIdx];
2837 SmallVector<const SCEV *, 2> Key;
2838 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2839 JE = F.BaseRegs.end(); J != JE; ++J) {
2840 const SCEV *Reg = *J;
2841 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2845 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2846 Key.push_back(F.ScaledReg);
2847 // Unstable sort by host order ok, because this is only used for
2849 std::sort(Key.begin(), Key.end());
2851 std::pair<BestFormulaeTy::const_iterator, bool> P =
2852 BestFormulae.insert(std::make_pair(Key, FIdx));
2854 Formula &Best = LU.Formulae[P.first->second];
2855 if (Sorter.operator()(F, Best))
2857 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2859 " in favor of formula "; Best.print(dbgs());
2862 ChangedFormulae = true;
2864 LU.DeleteFormula(F);
2872 // Now that we've filtered out some formulae, recompute the Regs set.
2874 LU.RecomputeRegs(LUIdx, RegUses);
2876 // Reset this to prepare for the next use.
2877 BestFormulae.clear();
2880 DEBUG(if (ChangedFormulae) {
2882 "After filtering out undesirable candidates:\n";
2887 // This is a rough guess that seems to work fairly well.
2888 static const size_t ComplexityLimit = UINT16_MAX;
2890 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2891 /// solutions the solver might have to consider. It almost never considers
2892 /// this many solutions because it prune the search space, but the pruning
2893 /// isn't always sufficient.
2894 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2896 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2897 E = Uses.end(); I != E; ++I) {
2898 size_t FSize = I->Formulae.size();
2899 if (FSize >= ComplexityLimit) {
2900 Power = ComplexityLimit;
2904 if (Power >= ComplexityLimit)
2910 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2911 /// of the registers of another formula, it won't help reduce register
2912 /// pressure (though it may not necessarily hurt register pressure); remove
2913 /// it to simplify the system.
2914 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2915 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2916 DEBUG(dbgs() << "The search space is too complex.\n");
2918 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2919 "which use a superset of registers used by other "
2922 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2923 LSRUse &LU = Uses[LUIdx];
2925 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2926 Formula &F = LU.Formulae[i];
2927 // Look for a formula with a constant or GV in a register. If the use
2928 // also has a formula with that same value in an immediate field,
2929 // delete the one that uses a register.
2930 for (SmallVectorImpl<const SCEV *>::const_iterator
2931 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2932 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2934 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2935 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2936 (I - F.BaseRegs.begin()));
2937 if (LU.HasFormulaWithSameRegs(NewF)) {
2938 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2939 LU.DeleteFormula(F);
2945 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2946 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2949 NewF.AM.BaseGV = GV;
2950 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2951 (I - F.BaseRegs.begin()));
2952 if (LU.HasFormulaWithSameRegs(NewF)) {
2953 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2955 LU.DeleteFormula(F);
2966 LU.RecomputeRegs(LUIdx, RegUses);
2969 DEBUG(dbgs() << "After pre-selection:\n";
2970 print_uses(dbgs()));
2974 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
2975 /// for expressions like A, A+1, A+2, etc., allocate a single register for
2977 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
2978 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2979 DEBUG(dbgs() << "The search space is too complex.\n");
2981 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2982 "separated by a constant offset will use the same "
2985 // This is especially useful for unrolled loops.
2987 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2988 LSRUse &LU = Uses[LUIdx];
2989 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2990 E = LU.Formulae.end(); I != E; ++I) {
2991 const Formula &F = *I;
2992 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2993 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2994 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2995 /*HasBaseReg=*/false,
2996 LU.Kind, LU.AccessTy)) {
2997 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3000 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3002 // Delete formulae from the new use which are no longer legal.
3004 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3005 Formula &F = LUThatHas->Formulae[i];
3006 if (!isLegalUse(F.AM,
3007 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3008 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3009 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3011 LUThatHas->DeleteFormula(F);
3018 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3020 // Update the relocs to reference the new use.
3021 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3022 E = Fixups.end(); I != E; ++I) {
3023 LSRFixup &Fixup = *I;
3024 if (Fixup.LUIdx == LUIdx) {
3025 Fixup.LUIdx = LUThatHas - &Uses.front();
3026 Fixup.Offset += F.AM.BaseOffs;
3027 DEBUG(dbgs() << "New fixup has offset "
3028 << Fixup.Offset << '\n');
3030 if (Fixup.LUIdx == NumUses-1)
3031 Fixup.LUIdx = LUIdx;
3034 // Delete the old use.
3045 DEBUG(dbgs() << "After pre-selection:\n";
3046 print_uses(dbgs()));
3050 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3051 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3052 /// we've done more filtering, as it may be able to find more formulae to
3054 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3055 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3056 DEBUG(dbgs() << "The search space is too complex.\n");
3058 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3059 "undesirable dedicated registers.\n");
3061 FilterOutUndesirableDedicatedRegisters();
3063 DEBUG(dbgs() << "After pre-selection:\n";
3064 print_uses(dbgs()));
3068 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3069 /// to be profitable, and then in any use which has any reference to that
3070 /// register, delete all formulae which do not reference that register.
3071 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3072 // With all other options exhausted, loop until the system is simple
3073 // enough to handle.
3074 SmallPtrSet<const SCEV *, 4> Taken;
3075 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3076 // Ok, we have too many of formulae on our hands to conveniently handle.
3077 // Use a rough heuristic to thin out the list.
3078 DEBUG(dbgs() << "The search space is too complex.\n");
3080 // Pick the register which is used by the most LSRUses, which is likely
3081 // to be a good reuse register candidate.
3082 const SCEV *Best = 0;
3083 unsigned BestNum = 0;
3084 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3086 const SCEV *Reg = *I;
3087 if (Taken.count(Reg))
3092 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3093 if (Count > BestNum) {
3100 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3101 << " will yield profitable reuse.\n");
3104 // In any use with formulae which references this register, delete formulae
3105 // which don't reference it.
3106 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3107 LSRUse &LU = Uses[LUIdx];
3108 if (!LU.Regs.count(Best)) continue;
3111 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3112 Formula &F = LU.Formulae[i];
3113 if (!F.referencesReg(Best)) {
3114 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3115 LU.DeleteFormula(F);
3119 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3125 LU.RecomputeRegs(LUIdx, RegUses);
3128 DEBUG(dbgs() << "After pre-selection:\n";
3129 print_uses(dbgs()));
3133 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3134 /// formulae to choose from, use some rough heuristics to prune down the number
3135 /// of formulae. This keeps the main solver from taking an extraordinary amount
3136 /// of time in some worst-case scenarios.
3137 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3138 NarrowSearchSpaceByDetectingSupersets();
3139 NarrowSearchSpaceByCollapsingUnrolledCode();
3140 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3141 NarrowSearchSpaceByPickingWinnerRegs();
3144 /// SolveRecurse - This is the recursive solver.
3145 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3147 SmallVectorImpl<const Formula *> &Workspace,
3148 const Cost &CurCost,
3149 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3150 DenseSet<const SCEV *> &VisitedRegs) const {
3153 // - use more aggressive filtering
3154 // - sort the formula so that the most profitable solutions are found first
3155 // - sort the uses too
3157 // - don't compute a cost, and then compare. compare while computing a cost
3159 // - track register sets with SmallBitVector
3161 const LSRUse &LU = Uses[Workspace.size()];
3163 // If this use references any register that's already a part of the
3164 // in-progress solution, consider it a requirement that a formula must
3165 // reference that register in order to be considered. This prunes out
3166 // unprofitable searching.
3167 SmallSetVector<const SCEV *, 4> ReqRegs;
3168 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3169 E = CurRegs.end(); I != E; ++I)
3170 if (LU.Regs.count(*I))
3173 bool AnySatisfiedReqRegs = false;
3174 SmallPtrSet<const SCEV *, 16> NewRegs;
3177 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3178 E = LU.Formulae.end(); I != E; ++I) {
3179 const Formula &F = *I;
3181 // Ignore formulae which do not use any of the required registers.
3182 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3183 JE = ReqRegs.end(); J != JE; ++J) {
3184 const SCEV *Reg = *J;
3185 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3186 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3190 AnySatisfiedReqRegs = true;
3192 // Evaluate the cost of the current formula. If it's already worse than
3193 // the current best, prune the search at that point.
3196 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3197 if (NewCost < SolutionCost) {
3198 Workspace.push_back(&F);
3199 if (Workspace.size() != Uses.size()) {
3200 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3201 NewRegs, VisitedRegs);
3202 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3203 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3205 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3206 dbgs() << ". Regs:";
3207 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3208 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3209 dbgs() << ' ' << **I;
3212 SolutionCost = NewCost;
3213 Solution = Workspace;
3215 Workspace.pop_back();
3220 // If none of the formulae had all of the required registers, relax the
3221 // constraint so that we don't exclude all formulae.
3222 if (!AnySatisfiedReqRegs) {
3223 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3229 /// Solve - Choose one formula from each use. Return the results in the given
3230 /// Solution vector.
3231 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3232 SmallVector<const Formula *, 8> Workspace;
3234 SolutionCost.Loose();
3236 SmallPtrSet<const SCEV *, 16> CurRegs;
3237 DenseSet<const SCEV *> VisitedRegs;
3238 Workspace.reserve(Uses.size());
3240 // SolveRecurse does all the work.
3241 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3242 CurRegs, VisitedRegs);
3244 // Ok, we've now made all our decisions.
3245 DEBUG(dbgs() << "\n"
3246 "The chosen solution requires "; SolutionCost.print(dbgs());
3248 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3250 Uses[i].print(dbgs());
3253 Solution[i]->print(dbgs());
3257 assert(Solution.size() == Uses.size() && "Malformed solution!");
3260 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3261 /// the dominator tree far as we can go while still being dominated by the
3262 /// input positions. This helps canonicalize the insert position, which
3263 /// encourages sharing.
3264 BasicBlock::iterator
3265 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3266 const SmallVectorImpl<Instruction *> &Inputs)
3269 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3270 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3273 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3274 if (!Rung) return IP;
3275 Rung = Rung->getIDom();
3276 if (!Rung) return IP;
3277 IDom = Rung->getBlock();
3279 // Don't climb into a loop though.
3280 const Loop *IDomLoop = LI.getLoopFor(IDom);
3281 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3282 if (IDomDepth <= IPLoopDepth &&
3283 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3287 bool AllDominate = true;
3288 Instruction *BetterPos = 0;
3289 Instruction *Tentative = IDom->getTerminator();
3290 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3291 E = Inputs.end(); I != E; ++I) {
3292 Instruction *Inst = *I;
3293 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3294 AllDominate = false;
3297 // Attempt to find an insert position in the middle of the block,
3298 // instead of at the end, so that it can be used for other expansions.
3299 if (IDom == Inst->getParent() &&
3300 (!BetterPos || DT.dominates(BetterPos, Inst)))
3301 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3314 /// AdjustInsertPositionForExpand - Determine an input position which will be
3315 /// dominated by the operands and which will dominate the result.
3316 BasicBlock::iterator
3317 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3319 const LSRUse &LU) const {
3320 // Collect some instructions which must be dominated by the
3321 // expanding replacement. These must be dominated by any operands that
3322 // will be required in the expansion.
3323 SmallVector<Instruction *, 4> Inputs;
3324 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3325 Inputs.push_back(I);
3326 if (LU.Kind == LSRUse::ICmpZero)
3327 if (Instruction *I =
3328 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3329 Inputs.push_back(I);
3330 if (LF.PostIncLoops.count(L)) {
3331 if (LF.isUseFullyOutsideLoop(L))
3332 Inputs.push_back(L->getLoopLatch()->getTerminator());
3334 Inputs.push_back(IVIncInsertPos);
3336 // The expansion must also be dominated by the increment positions of any
3337 // loops it for which it is using post-inc mode.
3338 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3339 E = LF.PostIncLoops.end(); I != E; ++I) {
3340 const Loop *PIL = *I;
3341 if (PIL == L) continue;
3343 // Be dominated by the loop exit.
3344 SmallVector<BasicBlock *, 4> ExitingBlocks;
3345 PIL->getExitingBlocks(ExitingBlocks);
3346 if (!ExitingBlocks.empty()) {
3347 BasicBlock *BB = ExitingBlocks[0];
3348 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3349 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3350 Inputs.push_back(BB->getTerminator());
3354 // Then, climb up the immediate dominator tree as far as we can go while
3355 // still being dominated by the input positions.
3356 IP = HoistInsertPosition(IP, Inputs);
3358 // Don't insert instructions before PHI nodes.
3359 while (isa<PHINode>(IP)) ++IP;
3361 // Ignore debug intrinsics.
3362 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3367 /// Expand - Emit instructions for the leading candidate expression for this
3368 /// LSRUse (this is called "expanding").
3369 Value *LSRInstance::Expand(const LSRFixup &LF,
3371 BasicBlock::iterator IP,
3372 SCEVExpander &Rewriter,
3373 SmallVectorImpl<WeakVH> &DeadInsts) const {
3374 const LSRUse &LU = Uses[LF.LUIdx];
3376 // Determine an input position which will be dominated by the operands and
3377 // which will dominate the result.
3378 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3380 // Inform the Rewriter if we have a post-increment use, so that it can
3381 // perform an advantageous expansion.
3382 Rewriter.setPostInc(LF.PostIncLoops);
3384 // This is the type that the user actually needs.
3385 const Type *OpTy = LF.OperandValToReplace->getType();
3386 // This will be the type that we'll initially expand to.
3387 const Type *Ty = F.getType();
3389 // No type known; just expand directly to the ultimate type.
3391 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3392 // Expand directly to the ultimate type if it's the right size.
3394 // This is the type to do integer arithmetic in.
3395 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3397 // Build up a list of operands to add together to form the full base.
3398 SmallVector<const SCEV *, 8> Ops;
3400 // Expand the BaseRegs portion.
3401 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3402 E = F.BaseRegs.end(); I != E; ++I) {
3403 const SCEV *Reg = *I;
3404 assert(!Reg->isZero() && "Zero allocated in a base register!");
3406 // If we're expanding for a post-inc user, make the post-inc adjustment.
3407 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3408 Reg = TransformForPostIncUse(Denormalize, Reg,
3409 LF.UserInst, LF.OperandValToReplace,
3412 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3415 // Flush the operand list to suppress SCEVExpander hoisting.
3417 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3419 Ops.push_back(SE.getUnknown(FullV));
3422 // Expand the ScaledReg portion.
3423 Value *ICmpScaledV = 0;
3424 if (F.AM.Scale != 0) {
3425 const SCEV *ScaledS = F.ScaledReg;
3427 // If we're expanding for a post-inc user, make the post-inc adjustment.
3428 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3429 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3430 LF.UserInst, LF.OperandValToReplace,
3433 if (LU.Kind == LSRUse::ICmpZero) {
3434 // An interesting way of "folding" with an icmp is to use a negated
3435 // scale, which we'll implement by inserting it into the other operand
3437 assert(F.AM.Scale == -1 &&
3438 "The only scale supported by ICmpZero uses is -1!");
3439 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3441 // Otherwise just expand the scaled register and an explicit scale,
3442 // which is expected to be matched as part of the address.
3443 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3444 ScaledS = SE.getMulExpr(ScaledS,
3445 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3446 Ops.push_back(ScaledS);
3448 // Flush the operand list to suppress SCEVExpander hoisting.
3449 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3451 Ops.push_back(SE.getUnknown(FullV));
3455 // Expand the GV portion.
3457 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3459 // Flush the operand list to suppress SCEVExpander hoisting.
3460 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3462 Ops.push_back(SE.getUnknown(FullV));
3465 // Expand the immediate portion.
3466 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3468 if (LU.Kind == LSRUse::ICmpZero) {
3469 // The other interesting way of "folding" with an ICmpZero is to use a
3470 // negated immediate.
3472 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3474 Ops.push_back(SE.getUnknown(ICmpScaledV));
3475 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3478 // Just add the immediate values. These again are expected to be matched
3479 // as part of the address.
3480 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3484 // Emit instructions summing all the operands.
3485 const SCEV *FullS = Ops.empty() ?
3486 SE.getConstant(IntTy, 0) :
3488 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3490 // We're done expanding now, so reset the rewriter.
3491 Rewriter.clearPostInc();
3493 // An ICmpZero Formula represents an ICmp which we're handling as a
3494 // comparison against zero. Now that we've expanded an expression for that
3495 // form, update the ICmp's other operand.
3496 if (LU.Kind == LSRUse::ICmpZero) {
3497 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3498 DeadInsts.push_back(CI->getOperand(1));
3499 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3500 "a scale at the same time!");
3501 if (F.AM.Scale == -1) {
3502 if (ICmpScaledV->getType() != OpTy) {
3504 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3506 ICmpScaledV, OpTy, "tmp", CI);
3509 CI->setOperand(1, ICmpScaledV);
3511 assert(F.AM.Scale == 0 &&
3512 "ICmp does not support folding a global value and "
3513 "a scale at the same time!");
3514 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3516 if (C->getType() != OpTy)
3517 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3521 CI->setOperand(1, C);
3528 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3529 /// of their operands effectively happens in their predecessor blocks, so the
3530 /// expression may need to be expanded in multiple places.
3531 void LSRInstance::RewriteForPHI(PHINode *PN,
3534 SCEVExpander &Rewriter,
3535 SmallVectorImpl<WeakVH> &DeadInsts,
3537 DenseMap<BasicBlock *, Value *> Inserted;
3538 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3539 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3540 BasicBlock *BB = PN->getIncomingBlock(i);
3542 // If this is a critical edge, split the edge so that we do not insert
3543 // the code on all predecessor/successor paths. We do this unless this
3544 // is the canonical backedge for this loop, which complicates post-inc
3546 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3547 !isa<IndirectBrInst>(BB->getTerminator()) &&
3548 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3549 // Split the critical edge.
3550 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3552 // If PN is outside of the loop and BB is in the loop, we want to
3553 // move the block to be immediately before the PHI block, not
3554 // immediately after BB.
3555 if (L->contains(BB) && !L->contains(PN))
3556 NewBB->moveBefore(PN->getParent());
3558 // Splitting the edge can reduce the number of PHI entries we have.
3559 e = PN->getNumIncomingValues();
3561 i = PN->getBasicBlockIndex(BB);
3564 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3565 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3567 PN->setIncomingValue(i, Pair.first->second);
3569 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3571 // If this is reuse-by-noop-cast, insert the noop cast.
3572 const Type *OpTy = LF.OperandValToReplace->getType();
3573 if (FullV->getType() != OpTy)
3575 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3577 FullV, LF.OperandValToReplace->getType(),
3578 "tmp", BB->getTerminator());
3580 PN->setIncomingValue(i, FullV);
3581 Pair.first->second = FullV;
3586 /// Rewrite - Emit instructions for the leading candidate expression for this
3587 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3588 /// the newly expanded value.
3589 void LSRInstance::Rewrite(const LSRFixup &LF,
3591 SCEVExpander &Rewriter,
3592 SmallVectorImpl<WeakVH> &DeadInsts,
3594 // First, find an insertion point that dominates UserInst. For PHI nodes,
3595 // find the nearest block which dominates all the relevant uses.
3596 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3597 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3599 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3601 // If this is reuse-by-noop-cast, insert the noop cast.
3602 const Type *OpTy = LF.OperandValToReplace->getType();
3603 if (FullV->getType() != OpTy) {
3605 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3606 FullV, OpTy, "tmp", LF.UserInst);
3610 // Update the user. ICmpZero is handled specially here (for now) because
3611 // Expand may have updated one of the operands of the icmp already, and
3612 // its new value may happen to be equal to LF.OperandValToReplace, in
3613 // which case doing replaceUsesOfWith leads to replacing both operands
3614 // with the same value. TODO: Reorganize this.
3615 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3616 LF.UserInst->setOperand(0, FullV);
3618 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3621 DeadInsts.push_back(LF.OperandValToReplace);
3624 /// ImplementSolution - Rewrite all the fixup locations with new values,
3625 /// following the chosen solution.
3627 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3629 // Keep track of instructions we may have made dead, so that
3630 // we can remove them after we are done working.
3631 SmallVector<WeakVH, 16> DeadInsts;
3633 SCEVExpander Rewriter(SE);
3634 Rewriter.disableCanonicalMode();
3635 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3637 // Expand the new value definitions and update the users.
3638 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3639 E = Fixups.end(); I != E; ++I) {
3640 const LSRFixup &Fixup = *I;
3642 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3647 // Clean up after ourselves. This must be done before deleting any
3651 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3654 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3655 : IU(P->getAnalysis<IVUsers>()),
3656 SE(P->getAnalysis<ScalarEvolution>()),
3657 DT(P->getAnalysis<DominatorTree>()),
3658 LI(P->getAnalysis<LoopInfo>()),
3659 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3661 // If LoopSimplify form is not available, stay out of trouble.
3662 if (!L->isLoopSimplifyForm()) return;
3664 // If there's no interesting work to be done, bail early.
3665 if (IU.empty()) return;
3667 DEBUG(dbgs() << "\nLSR on loop ";
3668 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3671 // First, perform some low-level loop optimizations.
3673 OptimizeLoopTermCond();
3675 // Start collecting data and preparing for the solver.
3676 CollectInterestingTypesAndFactors();
3677 CollectFixupsAndInitialFormulae();
3678 CollectLoopInvariantFixupsAndFormulae();
3680 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3681 print_uses(dbgs()));
3683 // Now use the reuse data to generate a bunch of interesting ways
3684 // to formulate the values needed for the uses.
3685 GenerateAllReuseFormulae();
3687 FilterOutUndesirableDedicatedRegisters();
3688 NarrowSearchSpaceUsingHeuristics();
3690 SmallVector<const Formula *, 8> Solution;
3693 // Release memory that is no longer needed.
3699 // Formulae should be legal.
3700 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3701 E = Uses.end(); I != E; ++I) {
3702 const LSRUse &LU = *I;
3703 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3704 JE = LU.Formulae.end(); J != JE; ++J)
3705 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3706 LU.Kind, LU.AccessTy, TLI) &&
3707 "Illegal formula generated!");
3711 // Now that we've decided what we want, make it so.
3712 ImplementSolution(Solution, P);
3715 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3716 if (Factors.empty() && Types.empty()) return;
3718 OS << "LSR has identified the following interesting factors and types: ";
3721 for (SmallSetVector<int64_t, 8>::const_iterator
3722 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3723 if (!First) OS << ", ";
3728 for (SmallSetVector<const Type *, 4>::const_iterator
3729 I = Types.begin(), E = Types.end(); I != E; ++I) {
3730 if (!First) OS << ", ";
3732 OS << '(' << **I << ')';
3737 void LSRInstance::print_fixups(raw_ostream &OS) const {
3738 OS << "LSR is examining the following fixup sites:\n";
3739 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3740 E = Fixups.end(); I != E; ++I) {
3747 void LSRInstance::print_uses(raw_ostream &OS) const {
3748 OS << "LSR is examining the following uses:\n";
3749 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3750 E = Uses.end(); I != E; ++I) {
3751 const LSRUse &LU = *I;
3755 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3756 JE = LU.Formulae.end(); J != JE; ++J) {
3764 void LSRInstance::print(raw_ostream &OS) const {
3765 print_factors_and_types(OS);
3770 void LSRInstance::dump() const {
3771 print(errs()); errs() << '\n';
3776 class LoopStrengthReduce : public LoopPass {
3777 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3778 /// transformation profitability.
3779 const TargetLowering *const TLI;
3782 static char ID; // Pass ID, replacement for typeid
3783 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3786 bool runOnLoop(Loop *L, LPPassManager &LPM);
3787 void getAnalysisUsage(AnalysisUsage &AU) const;
3792 char LoopStrengthReduce::ID = 0;
3793 INITIALIZE_PASS(LoopStrengthReduce, "loop-reduce",
3794 "Loop Strength Reduction", false, false);
3796 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3797 return new LoopStrengthReduce(TLI);
3800 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3801 : LoopPass(ID), TLI(tli) {}
3803 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3804 // We split critical edges, so we change the CFG. However, we do update
3805 // many analyses if they are around.
3806 AU.addPreservedID(LoopSimplifyID);
3807 AU.addPreserved("domfrontier");
3809 AU.addRequired<LoopInfo>();
3810 AU.addPreserved<LoopInfo>();
3811 AU.addRequiredID(LoopSimplifyID);
3812 AU.addRequired<DominatorTree>();
3813 AU.addPreserved<DominatorTree>();
3814 AU.addRequired<ScalarEvolution>();
3815 AU.addPreserved<ScalarEvolution>();
3816 AU.addRequired<IVUsers>();
3817 AU.addPreserved<IVUsers>();
3820 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3821 bool Changed = false;
3823 // Run the main LSR transformation.
3824 Changed |= LSRInstance(TLI, L, this).getChanged();
3826 // At this point, it is worth checking to see if any recurrence PHIs are also
3827 // dead, so that we can remove them as well.
3828 Changed |= DeleteDeadPHIs(L->getHeader());