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 if (!RegUsesMap.count(Reg)) return false;
165 const SmallBitVector &UsedByIndices =
166 RegUsesMap.find(Reg)->second.UsedByIndices;
167 int i = UsedByIndices.find_first();
168 if (i == -1) return false;
169 if ((size_t)i != LUIdx) return true;
170 return UsedByIndices.find_next(i) != -1;
173 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
174 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
175 assert(I != RegUsesMap.end() && "Unknown register!");
176 return I->second.UsedByIndices;
179 void RegUseTracker::clear() {
186 /// Formula - This class holds information that describes a formula for
187 /// computing satisfying a use. It may include broken-out immediates and scaled
190 /// AM - This is used to represent complex addressing, as well as other kinds
191 /// of interesting uses.
192 TargetLowering::AddrMode AM;
194 /// BaseRegs - The list of "base" registers for this use. When this is
195 /// non-empty, AM.HasBaseReg should be set to true.
196 SmallVector<const SCEV *, 2> BaseRegs;
198 /// ScaledReg - The 'scaled' register for this use. This should be non-null
199 /// when AM.Scale is not zero.
200 const SCEV *ScaledReg;
202 Formula() : ScaledReg(0) {}
204 void InitialMatch(const SCEV *S, Loop *L,
205 ScalarEvolution &SE, DominatorTree &DT);
207 unsigned getNumRegs() const;
208 const Type *getType() const;
210 void DeleteBaseReg(const SCEV *&S);
212 bool referencesReg(const SCEV *S) const;
213 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
214 const RegUseTracker &RegUses) const;
216 void print(raw_ostream &OS) const;
222 /// DoInitialMatch - Recursion helper for InitialMatch.
223 static void DoInitialMatch(const SCEV *S, Loop *L,
224 SmallVectorImpl<const SCEV *> &Good,
225 SmallVectorImpl<const SCEV *> &Bad,
226 ScalarEvolution &SE, DominatorTree &DT) {
227 // Collect expressions which properly dominate the loop header.
228 if (S->properlyDominates(L->getHeader(), &DT)) {
233 // Look at add operands.
234 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
235 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
237 DoInitialMatch(*I, L, Good, Bad, SE, DT);
241 // Look at addrec operands.
242 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
243 if (!AR->getStart()->isZero()) {
244 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
245 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
246 AR->getStepRecurrence(SE),
248 L, Good, Bad, SE, DT);
252 // Handle a multiplication by -1 (negation) if it didn't fold.
253 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
254 if (Mul->getOperand(0)->isAllOnesValue()) {
255 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
256 const SCEV *NewMul = SE.getMulExpr(Ops);
258 SmallVector<const SCEV *, 4> MyGood;
259 SmallVector<const SCEV *, 4> MyBad;
260 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
261 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
262 SE.getEffectiveSCEVType(NewMul->getType())));
263 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
264 E = MyGood.end(); I != E; ++I)
265 Good.push_back(SE.getMulExpr(NegOne, *I));
266 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
267 E = MyBad.end(); I != E; ++I)
268 Bad.push_back(SE.getMulExpr(NegOne, *I));
272 // Ok, we can't do anything interesting. Just stuff the whole thing into a
273 // register and hope for the best.
277 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
278 /// attempting to keep all loop-invariant and loop-computable values in a
279 /// single base register.
280 void Formula::InitialMatch(const SCEV *S, Loop *L,
281 ScalarEvolution &SE, DominatorTree &DT) {
282 SmallVector<const SCEV *, 4> Good;
283 SmallVector<const SCEV *, 4> Bad;
284 DoInitialMatch(S, L, Good, Bad, SE, DT);
286 const SCEV *Sum = SE.getAddExpr(Good);
288 BaseRegs.push_back(Sum);
289 AM.HasBaseReg = true;
292 const SCEV *Sum = SE.getAddExpr(Bad);
294 BaseRegs.push_back(Sum);
295 AM.HasBaseReg = true;
299 /// getNumRegs - Return the total number of register operands used by this
300 /// formula. This does not include register uses implied by non-constant
302 unsigned Formula::getNumRegs() const {
303 return !!ScaledReg + BaseRegs.size();
306 /// getType - Return the type of this formula, if it has one, or null
307 /// otherwise. This type is meaningless except for the bit size.
308 const Type *Formula::getType() const {
309 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
310 ScaledReg ? ScaledReg->getType() :
311 AM.BaseGV ? AM.BaseGV->getType() :
315 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
316 void Formula::DeleteBaseReg(const SCEV *&S) {
317 if (&S != &BaseRegs.back())
318 std::swap(S, BaseRegs.back());
322 /// referencesReg - Test if this formula references the given register.
323 bool Formula::referencesReg(const SCEV *S) const {
324 return S == ScaledReg ||
325 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
328 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
329 /// which are used by uses other than the use with the given index.
330 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
331 const RegUseTracker &RegUses) const {
333 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
335 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
336 E = BaseRegs.end(); I != E; ++I)
337 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
342 void Formula::print(raw_ostream &OS) const {
345 if (!First) OS << " + "; else First = false;
346 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
348 if (AM.BaseOffs != 0) {
349 if (!First) OS << " + "; else First = false;
352 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
353 E = BaseRegs.end(); I != E; ++I) {
354 if (!First) OS << " + "; else First = false;
355 OS << "reg(" << **I << ')';
357 if (AM.HasBaseReg && BaseRegs.empty()) {
358 if (!First) OS << " + "; else First = false;
359 OS << "**error: HasBaseReg**";
360 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
361 if (!First) OS << " + "; else First = false;
362 OS << "**error: !HasBaseReg**";
365 if (!First) OS << " + "; else First = false;
366 OS << AM.Scale << "*reg(";
375 void Formula::dump() const {
376 print(errs()); errs() << '\n';
379 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
380 /// without changing its value.
381 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
383 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
384 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
387 /// isAddSExtable - Return true if the given add can be sign-extended
388 /// without changing its value.
389 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
391 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
392 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
395 /// isMulSExtable - Return true if the given add can be sign-extended
396 /// without changing its value.
397 static bool isMulSExtable(const SCEVMulExpr *A, ScalarEvolution &SE) {
399 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
400 return isa<SCEVMulExpr>(SE.getSignExtendExpr(A, WideTy));
403 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
404 /// and if the remainder is known to be zero, or null otherwise. If
405 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
406 /// to Y, ignoring that the multiplication may overflow, which is useful when
407 /// the result will be used in a context where the most significant bits are
409 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
411 bool IgnoreSignificantBits = false) {
412 // Handle the trivial case, which works for any SCEV type.
414 return SE.getConstant(LHS->getType(), 1);
416 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
418 if (RHS->isAllOnesValue())
419 return SE.getMulExpr(LHS, RHS);
421 // Check for a division of a constant by a constant.
422 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
423 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
426 if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
428 return SE.getConstant(C->getValue()->getValue()
429 .sdiv(RC->getValue()->getValue()));
432 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
433 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
434 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
435 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
436 IgnoreSignificantBits);
437 if (!Start) return 0;
438 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
439 IgnoreSignificantBits);
441 return SE.getAddRecExpr(Start, Step, AR->getLoop());
445 // Distribute the sdiv over add operands, if the add doesn't overflow.
446 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
447 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
448 SmallVector<const SCEV *, 8> Ops;
449 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
451 const SCEV *Op = getExactSDiv(*I, RHS, SE,
452 IgnoreSignificantBits);
456 return SE.getAddExpr(Ops);
460 // Check for a multiply operand that we can pull RHS out of.
461 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
462 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
463 SmallVector<const SCEV *, 4> Ops;
465 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
469 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
470 IgnoreSignificantBits)) {
476 return Found ? SE.getMulExpr(Ops) : 0;
479 // Otherwise we don't know.
483 /// ExtractImmediate - If S involves the addition of a constant integer value,
484 /// return that integer value, and mutate S to point to a new SCEV with that
486 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
487 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
488 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
489 S = SE.getConstant(C->getType(), 0);
490 return C->getValue()->getSExtValue();
492 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
493 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
494 int64_t Result = ExtractImmediate(NewOps.front(), SE);
495 S = SE.getAddExpr(NewOps);
497 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
498 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
499 int64_t Result = ExtractImmediate(NewOps.front(), SE);
500 S = SE.getAddRecExpr(NewOps, AR->getLoop());
506 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
507 /// return that symbol, and mutate S to point to a new SCEV with that
509 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
510 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
511 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
512 S = SE.getConstant(GV->getType(), 0);
515 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
516 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
517 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
518 S = SE.getAddExpr(NewOps);
520 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
521 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
522 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
523 S = SE.getAddRecExpr(NewOps, AR->getLoop());
529 /// isAddressUse - Returns true if the specified instruction is using the
530 /// specified value as an address.
531 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
532 bool isAddress = isa<LoadInst>(Inst);
533 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
534 if (SI->getOperand(1) == OperandVal)
536 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
537 // Addressing modes can also be folded into prefetches and a variety
539 switch (II->getIntrinsicID()) {
541 case Intrinsic::prefetch:
542 case Intrinsic::x86_sse2_loadu_dq:
543 case Intrinsic::x86_sse2_loadu_pd:
544 case Intrinsic::x86_sse_loadu_ps:
545 case Intrinsic::x86_sse_storeu_ps:
546 case Intrinsic::x86_sse2_storeu_pd:
547 case Intrinsic::x86_sse2_storeu_dq:
548 case Intrinsic::x86_sse2_storel_dq:
549 if (II->getOperand(1) == OperandVal)
557 /// getAccessType - Return the type of the memory being accessed.
558 static const Type *getAccessType(const Instruction *Inst) {
559 const Type *AccessTy = Inst->getType();
560 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
561 AccessTy = SI->getOperand(0)->getType();
562 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
563 // Addressing modes can also be folded into prefetches and a variety
565 switch (II->getIntrinsicID()) {
567 case Intrinsic::x86_sse_storeu_ps:
568 case Intrinsic::x86_sse2_storeu_pd:
569 case Intrinsic::x86_sse2_storeu_dq:
570 case Intrinsic::x86_sse2_storel_dq:
571 AccessTy = II->getOperand(1)->getType();
576 // All pointers have the same requirements, so canonicalize them to an
577 // arbitrary pointer type to minimize variation.
578 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
579 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
580 PTy->getAddressSpace());
585 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
586 /// specified set are trivially dead, delete them and see if this makes any of
587 /// their operands subsequently dead.
589 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
590 bool Changed = false;
592 while (!DeadInsts.empty()) {
593 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
595 if (I == 0 || !isInstructionTriviallyDead(I))
598 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
599 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
602 DeadInsts.push_back(U);
605 I->eraseFromParent();
614 /// Cost - This class is used to measure and compare candidate formulae.
616 /// TODO: Some of these could be merged. Also, a lexical ordering
617 /// isn't always optimal.
621 unsigned NumBaseAdds;
627 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
630 unsigned getNumRegs() const { return NumRegs; }
632 bool operator<(const Cost &Other) const;
636 void RateFormula(const Formula &F,
637 SmallPtrSet<const SCEV *, 16> &Regs,
638 const DenseSet<const SCEV *> &VisitedRegs,
640 const SmallVectorImpl<int64_t> &Offsets,
641 ScalarEvolution &SE, DominatorTree &DT);
643 void print(raw_ostream &OS) const;
647 void RateRegister(const SCEV *Reg,
648 SmallPtrSet<const SCEV *, 16> &Regs,
650 ScalarEvolution &SE, DominatorTree &DT);
651 void RatePrimaryRegister(const SCEV *Reg,
652 SmallPtrSet<const SCEV *, 16> &Regs,
654 ScalarEvolution &SE, DominatorTree &DT);
659 /// RateRegister - Tally up interesting quantities from the given register.
660 void Cost::RateRegister(const SCEV *Reg,
661 SmallPtrSet<const SCEV *, 16> &Regs,
663 ScalarEvolution &SE, DominatorTree &DT) {
664 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
665 if (AR->getLoop() == L)
666 AddRecCost += 1; /// TODO: This should be a function of the stride.
668 // If this is an addrec for a loop that's already been visited by LSR,
669 // don't second-guess its addrec phi nodes. LSR isn't currently smart
670 // enough to reason about more than one loop at a time. Consider these
671 // registers free and leave them alone.
672 else if (L->contains(AR->getLoop()) ||
673 (!AR->getLoop()->contains(L) &&
674 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
675 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
676 PHINode *PN = dyn_cast<PHINode>(I); ++I)
677 if (SE.isSCEVable(PN->getType()) &&
678 (SE.getEffectiveSCEVType(PN->getType()) ==
679 SE.getEffectiveSCEVType(AR->getType())) &&
680 SE.getSCEV(PN) == AR)
683 // If this isn't one of the addrecs that the loop already has, it
684 // would require a costly new phi and add. TODO: This isn't
685 // precisely modeled right now.
687 if (!Regs.count(AR->getStart()))
688 RateRegister(AR->getStart(), Regs, L, SE, DT);
691 // Add the step value register, if it needs one.
692 // TODO: The non-affine case isn't precisely modeled here.
693 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
694 if (!Regs.count(AR->getStart()))
695 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
699 // Rough heuristic; favor registers which don't require extra setup
700 // instructions in the preheader.
701 if (!isa<SCEVUnknown>(Reg) &&
702 !isa<SCEVConstant>(Reg) &&
703 !(isa<SCEVAddRecExpr>(Reg) &&
704 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
705 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
709 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
711 void Cost::RatePrimaryRegister(const SCEV *Reg,
712 SmallPtrSet<const SCEV *, 16> &Regs,
714 ScalarEvolution &SE, DominatorTree &DT) {
715 if (Regs.insert(Reg))
716 RateRegister(Reg, Regs, L, SE, DT);
719 void Cost::RateFormula(const Formula &F,
720 SmallPtrSet<const SCEV *, 16> &Regs,
721 const DenseSet<const SCEV *> &VisitedRegs,
723 const SmallVectorImpl<int64_t> &Offsets,
724 ScalarEvolution &SE, DominatorTree &DT) {
725 // Tally up the registers.
726 if (const SCEV *ScaledReg = F.ScaledReg) {
727 if (VisitedRegs.count(ScaledReg)) {
731 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
733 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
734 E = F.BaseRegs.end(); I != E; ++I) {
735 const SCEV *BaseReg = *I;
736 if (VisitedRegs.count(BaseReg)) {
740 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
742 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
743 BaseReg->hasComputableLoopEvolution(L);
746 if (F.BaseRegs.size() > 1)
747 NumBaseAdds += F.BaseRegs.size() - 1;
749 // Tally up the non-zero immediates.
750 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
751 E = Offsets.end(); I != E; ++I) {
752 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
754 ImmCost += 64; // Handle symbolic values conservatively.
755 // TODO: This should probably be the pointer size.
756 else if (Offset != 0)
757 ImmCost += APInt(64, Offset, true).getMinSignedBits();
761 /// Loose - Set this cost to a loosing value.
771 /// operator< - Choose the lower cost.
772 bool Cost::operator<(const Cost &Other) const {
773 if (NumRegs != Other.NumRegs)
774 return NumRegs < Other.NumRegs;
775 if (AddRecCost != Other.AddRecCost)
776 return AddRecCost < Other.AddRecCost;
777 if (NumIVMuls != Other.NumIVMuls)
778 return NumIVMuls < Other.NumIVMuls;
779 if (NumBaseAdds != Other.NumBaseAdds)
780 return NumBaseAdds < Other.NumBaseAdds;
781 if (ImmCost != Other.ImmCost)
782 return ImmCost < Other.ImmCost;
783 if (SetupCost != Other.SetupCost)
784 return SetupCost < Other.SetupCost;
788 void Cost::print(raw_ostream &OS) const {
789 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
791 OS << ", with addrec cost " << AddRecCost;
793 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
794 if (NumBaseAdds != 0)
795 OS << ", plus " << NumBaseAdds << " base add"
796 << (NumBaseAdds == 1 ? "" : "s");
798 OS << ", plus " << ImmCost << " imm cost";
800 OS << ", plus " << SetupCost << " setup cost";
803 void Cost::dump() const {
804 print(errs()); errs() << '\n';
809 /// LSRFixup - An operand value in an instruction which is to be replaced
810 /// with some equivalent, possibly strength-reduced, replacement.
812 /// UserInst - The instruction which will be updated.
813 Instruction *UserInst;
815 /// OperandValToReplace - The operand of the instruction which will
816 /// be replaced. The operand may be used more than once; every instance
817 /// will be replaced.
818 Value *OperandValToReplace;
820 /// PostIncLoops - If this user is to use the post-incremented value of an
821 /// induction variable, this variable is non-null and holds the loop
822 /// associated with the induction variable.
823 PostIncLoopSet PostIncLoops;
825 /// LUIdx - The index of the LSRUse describing the expression which
826 /// this fixup needs, minus an offset (below).
829 /// Offset - A constant offset to be added to the LSRUse expression.
830 /// This allows multiple fixups to share the same LSRUse with different
831 /// offsets, for example in an unrolled loop.
834 bool isUseFullyOutsideLoop(const Loop *L) const;
838 void print(raw_ostream &OS) const;
845 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
847 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
848 /// value outside of the given loop.
849 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
850 // PHI nodes use their value in their incoming blocks.
851 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
852 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
853 if (PN->getIncomingValue(i) == OperandValToReplace &&
854 L->contains(PN->getIncomingBlock(i)))
859 return !L->contains(UserInst);
862 void LSRFixup::print(raw_ostream &OS) const {
864 // Store is common and interesting enough to be worth special-casing.
865 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
867 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
868 } else if (UserInst->getType()->isVoidTy())
869 OS << UserInst->getOpcodeName();
871 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
873 OS << ", OperandValToReplace=";
874 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
876 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
877 E = PostIncLoops.end(); I != E; ++I) {
878 OS << ", PostIncLoop=";
879 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
882 if (LUIdx != ~size_t(0))
883 OS << ", LUIdx=" << LUIdx;
886 OS << ", Offset=" << Offset;
889 void LSRFixup::dump() const {
890 print(errs()); errs() << '\n';
895 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
896 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
897 struct UniquifierDenseMapInfo {
898 static SmallVector<const SCEV *, 2> getEmptyKey() {
899 SmallVector<const SCEV *, 2> V;
900 V.push_back(reinterpret_cast<const SCEV *>(-1));
904 static SmallVector<const SCEV *, 2> getTombstoneKey() {
905 SmallVector<const SCEV *, 2> V;
906 V.push_back(reinterpret_cast<const SCEV *>(-2));
910 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
912 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
913 E = V.end(); I != E; ++I)
914 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
918 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
919 const SmallVector<const SCEV *, 2> &RHS) {
924 /// LSRUse - This class holds the state that LSR keeps for each use in
925 /// IVUsers, as well as uses invented by LSR itself. It includes information
926 /// about what kinds of things can be folded into the user, information about
927 /// the user itself, and information about how the use may be satisfied.
928 /// TODO: Represent multiple users of the same expression in common?
930 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
933 /// KindType - An enum for a kind of use, indicating what types of
934 /// scaled and immediate operands it might support.
936 Basic, ///< A normal use, with no folding.
937 Special, ///< A special case of basic, allowing -1 scales.
938 Address, ///< An address use; folding according to TargetLowering
939 ICmpZero ///< An equality icmp with both operands folded into one.
940 // TODO: Add a generic icmp too?
944 const Type *AccessTy;
946 SmallVector<int64_t, 8> Offsets;
950 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
951 /// LSRUse are outside of the loop, in which case some special-case heuristics
953 bool AllFixupsOutsideLoop;
955 /// Formulae - A list of ways to build a value that can satisfy this user.
956 /// After the list is populated, one of these is selected heuristically and
957 /// used to formulate a replacement for OperandValToReplace in UserInst.
958 SmallVector<Formula, 12> Formulae;
960 /// Regs - The set of register candidates used by all formulae in this LSRUse.
961 SmallPtrSet<const SCEV *, 4> Regs;
963 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
964 MinOffset(INT64_MAX),
965 MaxOffset(INT64_MIN),
966 AllFixupsOutsideLoop(true) {}
968 bool HasFormulaWithSameRegs(const Formula &F) const;
969 bool InsertFormula(const Formula &F);
970 void DeleteFormula(Formula &F);
971 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
975 void print(raw_ostream &OS) const;
979 /// HasFormula - Test whether this use as a formula which has the same
980 /// registers as the given formula.
981 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
982 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
983 if (F.ScaledReg) Key.push_back(F.ScaledReg);
984 // Unstable sort by host order ok, because this is only used for uniquifying.
985 std::sort(Key.begin(), Key.end());
986 return Uniquifier.count(Key);
989 /// InsertFormula - If the given formula has not yet been inserted, add it to
990 /// the list, and return true. Return false otherwise.
991 bool LSRUse::InsertFormula(const Formula &F) {
992 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
993 if (F.ScaledReg) Key.push_back(F.ScaledReg);
994 // Unstable sort by host order ok, because this is only used for uniquifying.
995 std::sort(Key.begin(), Key.end());
997 if (!Uniquifier.insert(Key).second)
1000 // Using a register to hold the value of 0 is not profitable.
1001 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1002 "Zero allocated in a scaled register!");
1004 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1005 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1006 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1009 // Add the formula to the list.
1010 Formulae.push_back(F);
1012 // Record registers now being used by this use.
1013 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1014 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1019 /// DeleteFormula - Remove the given formula from this use's list.
1020 void LSRUse::DeleteFormula(Formula &F) {
1021 if (&F != &Formulae.back())
1022 std::swap(F, Formulae.back());
1023 Formulae.pop_back();
1024 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1027 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1028 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1029 // Now that we've filtered out some formulae, recompute the Regs set.
1030 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1032 for (size_t FIdx = 0, NumForms = Formulae.size(); FIdx != NumForms; ++FIdx) {
1033 Formula &F = Formulae[FIdx];
1034 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1035 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1038 // Update the RegTracker.
1039 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1040 E = OldRegs.end(); I != E; ++I)
1041 if (!Regs.count(*I))
1042 RegUses.DropRegister(*I, LUIdx);
1045 void LSRUse::print(raw_ostream &OS) const {
1046 OS << "LSR Use: Kind=";
1048 case Basic: OS << "Basic"; break;
1049 case Special: OS << "Special"; break;
1050 case ICmpZero: OS << "ICmpZero"; break;
1052 OS << "Address of ";
1053 if (AccessTy->isPointerTy())
1054 OS << "pointer"; // the full pointer type could be really verbose
1059 OS << ", Offsets={";
1060 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1061 E = Offsets.end(); I != E; ++I) {
1068 if (AllFixupsOutsideLoop)
1069 OS << ", all-fixups-outside-loop";
1072 void LSRUse::dump() const {
1073 print(errs()); errs() << '\n';
1076 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1077 /// be completely folded into the user instruction at isel time. This includes
1078 /// address-mode folding and special icmp tricks.
1079 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1080 LSRUse::KindType Kind, const Type *AccessTy,
1081 const TargetLowering *TLI) {
1083 case LSRUse::Address:
1084 // If we have low-level target information, ask the target if it can
1085 // completely fold this address.
1086 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1088 // Otherwise, just guess that reg+reg addressing is legal.
1089 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1091 case LSRUse::ICmpZero:
1092 // There's not even a target hook for querying whether it would be legal to
1093 // fold a GV into an ICmp.
1097 // ICmp only has two operands; don't allow more than two non-trivial parts.
1098 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1101 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1102 // putting the scaled register in the other operand of the icmp.
1103 if (AM.Scale != 0 && AM.Scale != -1)
1106 // If we have low-level target information, ask the target if it can fold an
1107 // integer immediate on an icmp.
1108 if (AM.BaseOffs != 0) {
1109 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1116 // Only handle single-register values.
1117 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1119 case LSRUse::Special:
1120 // Only handle -1 scales, or no scale.
1121 return AM.Scale == 0 || AM.Scale == -1;
1127 static bool isLegalUse(TargetLowering::AddrMode AM,
1128 int64_t MinOffset, int64_t MaxOffset,
1129 LSRUse::KindType Kind, const Type *AccessTy,
1130 const TargetLowering *TLI) {
1131 // Check for overflow.
1132 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1135 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1136 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1137 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1138 // Check for overflow.
1139 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1142 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1143 return isLegalUse(AM, Kind, AccessTy, TLI);
1148 static bool isAlwaysFoldable(int64_t BaseOffs,
1149 GlobalValue *BaseGV,
1151 LSRUse::KindType Kind, const Type *AccessTy,
1152 const TargetLowering *TLI) {
1153 // Fast-path: zero is always foldable.
1154 if (BaseOffs == 0 && !BaseGV) return true;
1156 // Conservatively, create an address with an immediate and a
1157 // base and a scale.
1158 TargetLowering::AddrMode AM;
1159 AM.BaseOffs = BaseOffs;
1161 AM.HasBaseReg = HasBaseReg;
1162 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1164 // Canonicalize a scale of 1 to a base register if the formula doesn't
1165 // already have a base register.
1166 if (!AM.HasBaseReg && AM.Scale == 1) {
1168 AM.HasBaseReg = true;
1171 return isLegalUse(AM, Kind, AccessTy, TLI);
1174 static bool isAlwaysFoldable(const SCEV *S,
1175 int64_t MinOffset, int64_t MaxOffset,
1177 LSRUse::KindType Kind, const Type *AccessTy,
1178 const TargetLowering *TLI,
1179 ScalarEvolution &SE) {
1180 // Fast-path: zero is always foldable.
1181 if (S->isZero()) return true;
1183 // Conservatively, create an address with an immediate and a
1184 // base and a scale.
1185 int64_t BaseOffs = ExtractImmediate(S, SE);
1186 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1188 // If there's anything else involved, it's not foldable.
1189 if (!S->isZero()) return false;
1191 // Fast-path: zero is always foldable.
1192 if (BaseOffs == 0 && !BaseGV) return true;
1194 // Conservatively, create an address with an immediate and a
1195 // base and a scale.
1196 TargetLowering::AddrMode AM;
1197 AM.BaseOffs = BaseOffs;
1199 AM.HasBaseReg = HasBaseReg;
1200 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1202 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1205 /// FormulaSorter - This class implements an ordering for formulae which sorts
1206 /// the by their standalone cost.
1207 class FormulaSorter {
1208 /// These two sets are kept empty, so that we compute standalone costs.
1209 DenseSet<const SCEV *> VisitedRegs;
1210 SmallPtrSet<const SCEV *, 16> Regs;
1213 ScalarEvolution &SE;
1217 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1218 : L(l), LU(&lu), SE(se), DT(dt) {}
1220 bool operator()(const Formula &A, const Formula &B) {
1222 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1225 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1227 return CostA < CostB;
1231 /// LSRInstance - This class holds state for the main loop strength reduction
1235 ScalarEvolution &SE;
1238 const TargetLowering *const TLI;
1242 /// IVIncInsertPos - This is the insert position that the current loop's
1243 /// induction variable increment should be placed. In simple loops, this is
1244 /// the latch block's terminator. But in more complicated cases, this is a
1245 /// position which will dominate all the in-loop post-increment users.
1246 Instruction *IVIncInsertPos;
1248 /// Factors - Interesting factors between use strides.
1249 SmallSetVector<int64_t, 8> Factors;
1251 /// Types - Interesting use types, to facilitate truncation reuse.
1252 SmallSetVector<const Type *, 4> Types;
1254 /// Fixups - The list of operands which are to be replaced.
1255 SmallVector<LSRFixup, 16> Fixups;
1257 /// Uses - The list of interesting uses.
1258 SmallVector<LSRUse, 16> Uses;
1260 /// RegUses - Track which uses use which register candidates.
1261 RegUseTracker RegUses;
1263 void OptimizeShadowIV();
1264 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1265 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1266 void OptimizeLoopTermCond();
1268 void CollectInterestingTypesAndFactors();
1269 void CollectFixupsAndInitialFormulae();
1271 LSRFixup &getNewFixup() {
1272 Fixups.push_back(LSRFixup());
1273 return Fixups.back();
1276 // Support for sharing of LSRUses between LSRFixups.
1277 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1280 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1281 LSRUse::KindType Kind, const Type *AccessTy);
1283 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1284 LSRUse::KindType Kind,
1285 const Type *AccessTy);
1287 void DeleteUse(LSRUse &LU);
1289 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1292 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1293 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1294 void CountRegisters(const Formula &F, size_t LUIdx);
1295 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1297 void CollectLoopInvariantFixupsAndFormulae();
1299 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1300 unsigned Depth = 0);
1301 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1302 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1303 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1304 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1305 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1306 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1307 void GenerateCrossUseConstantOffsets();
1308 void GenerateAllReuseFormulae();
1310 void FilterOutUndesirableDedicatedRegisters();
1312 size_t EstimateSearchSpaceComplexity() const;
1313 void NarrowSearchSpaceUsingHeuristics();
1315 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1317 SmallVectorImpl<const Formula *> &Workspace,
1318 const Cost &CurCost,
1319 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1320 DenseSet<const SCEV *> &VisitedRegs) const;
1321 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1323 BasicBlock::iterator
1324 HoistInsertPosition(BasicBlock::iterator IP,
1325 const SmallVectorImpl<Instruction *> &Inputs) const;
1326 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1328 const LSRUse &LU) const;
1330 Value *Expand(const LSRFixup &LF,
1332 BasicBlock::iterator IP,
1333 SCEVExpander &Rewriter,
1334 SmallVectorImpl<WeakVH> &DeadInsts) const;
1335 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1337 SCEVExpander &Rewriter,
1338 SmallVectorImpl<WeakVH> &DeadInsts,
1340 void Rewrite(const LSRFixup &LF,
1342 SCEVExpander &Rewriter,
1343 SmallVectorImpl<WeakVH> &DeadInsts,
1345 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1348 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1350 bool getChanged() const { return Changed; }
1352 void print_factors_and_types(raw_ostream &OS) const;
1353 void print_fixups(raw_ostream &OS) const;
1354 void print_uses(raw_ostream &OS) const;
1355 void print(raw_ostream &OS) const;
1361 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1362 /// inside the loop then try to eliminate the cast operation.
1363 void LSRInstance::OptimizeShadowIV() {
1364 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1365 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1368 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1369 UI != E; /* empty */) {
1370 IVUsers::const_iterator CandidateUI = UI;
1372 Instruction *ShadowUse = CandidateUI->getUser();
1373 const Type *DestTy = NULL;
1375 /* If shadow use is a int->float cast then insert a second IV
1376 to eliminate this cast.
1378 for (unsigned i = 0; i < n; ++i)
1384 for (unsigned i = 0; i < n; ++i, ++d)
1387 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1388 DestTy = UCast->getDestTy();
1389 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1390 DestTy = SCast->getDestTy();
1391 if (!DestTy) continue;
1394 // If target does not support DestTy natively then do not apply
1395 // this transformation.
1396 EVT DVT = TLI->getValueType(DestTy);
1397 if (!TLI->isTypeLegal(DVT)) continue;
1400 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1402 if (PH->getNumIncomingValues() != 2) continue;
1404 const Type *SrcTy = PH->getType();
1405 int Mantissa = DestTy->getFPMantissaWidth();
1406 if (Mantissa == -1) continue;
1407 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1410 unsigned Entry, Latch;
1411 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1419 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1420 if (!Init) continue;
1421 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1423 BinaryOperator *Incr =
1424 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1425 if (!Incr) continue;
1426 if (Incr->getOpcode() != Instruction::Add
1427 && Incr->getOpcode() != Instruction::Sub)
1430 /* Initialize new IV, double d = 0.0 in above example. */
1431 ConstantInt *C = NULL;
1432 if (Incr->getOperand(0) == PH)
1433 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1434 else if (Incr->getOperand(1) == PH)
1435 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1441 // Ignore negative constants, as the code below doesn't handle them
1442 // correctly. TODO: Remove this restriction.
1443 if (!C->getValue().isStrictlyPositive()) continue;
1445 /* Add new PHINode. */
1446 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1448 /* create new increment. '++d' in above example. */
1449 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1450 BinaryOperator *NewIncr =
1451 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1452 Instruction::FAdd : Instruction::FSub,
1453 NewPH, CFP, "IV.S.next.", Incr);
1455 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1456 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1458 /* Remove cast operation */
1459 ShadowUse->replaceAllUsesWith(NewPH);
1460 ShadowUse->eraseFromParent();
1466 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1467 /// set the IV user and stride information and return true, otherwise return
1469 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1470 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1471 if (UI->getUser() == Cond) {
1472 // NOTE: we could handle setcc instructions with multiple uses here, but
1473 // InstCombine does it as well for simple uses, it's not clear that it
1474 // occurs enough in real life to handle.
1481 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1482 /// a max computation.
1484 /// This is a narrow solution to a specific, but acute, problem. For loops
1490 /// } while (++i < n);
1492 /// the trip count isn't just 'n', because 'n' might not be positive. And
1493 /// unfortunately this can come up even for loops where the user didn't use
1494 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1495 /// will commonly be lowered like this:
1501 /// } while (++i < n);
1504 /// and then it's possible for subsequent optimization to obscure the if
1505 /// test in such a way that indvars can't find it.
1507 /// When indvars can't find the if test in loops like this, it creates a
1508 /// max expression, which allows it to give the loop a canonical
1509 /// induction variable:
1512 /// max = n < 1 ? 1 : n;
1515 /// } while (++i != max);
1517 /// Canonical induction variables are necessary because the loop passes
1518 /// are designed around them. The most obvious example of this is the
1519 /// LoopInfo analysis, which doesn't remember trip count values. It
1520 /// expects to be able to rediscover the trip count each time it is
1521 /// needed, and it does this using a simple analysis that only succeeds if
1522 /// the loop has a canonical induction variable.
1524 /// However, when it comes time to generate code, the maximum operation
1525 /// can be quite costly, especially if it's inside of an outer loop.
1527 /// This function solves this problem by detecting this type of loop and
1528 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1529 /// the instructions for the maximum computation.
1531 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1532 // Check that the loop matches the pattern we're looking for.
1533 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1534 Cond->getPredicate() != CmpInst::ICMP_NE)
1537 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1538 if (!Sel || !Sel->hasOneUse()) return Cond;
1540 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1541 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1543 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1545 // Add one to the backedge-taken count to get the trip count.
1546 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1547 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1549 // Check for a max calculation that matches the pattern. There's no check
1550 // for ICMP_ULE here because the comparison would be with zero, which
1551 // isn't interesting.
1552 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1553 const SCEVNAryExpr *Max = 0;
1554 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1555 Pred = ICmpInst::ICMP_SLE;
1557 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1558 Pred = ICmpInst::ICMP_SLT;
1560 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1561 Pred = ICmpInst::ICMP_ULT;
1568 // To handle a max with more than two operands, this optimization would
1569 // require additional checking and setup.
1570 if (Max->getNumOperands() != 2)
1573 const SCEV *MaxLHS = Max->getOperand(0);
1574 const SCEV *MaxRHS = Max->getOperand(1);
1576 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1577 // for a comparison with 1. For <= and >=, a comparison with zero.
1579 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1582 // Check the relevant induction variable for conformance to
1584 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1585 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1586 if (!AR || !AR->isAffine() ||
1587 AR->getStart() != One ||
1588 AR->getStepRecurrence(SE) != One)
1591 assert(AR->getLoop() == L &&
1592 "Loop condition operand is an addrec in a different loop!");
1594 // Check the right operand of the select, and remember it, as it will
1595 // be used in the new comparison instruction.
1597 if (ICmpInst::isTrueWhenEqual(Pred)) {
1598 // Look for n+1, and grab n.
1599 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1600 if (isa<ConstantInt>(BO->getOperand(1)) &&
1601 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1602 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1603 NewRHS = BO->getOperand(0);
1604 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1605 if (isa<ConstantInt>(BO->getOperand(1)) &&
1606 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1607 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1608 NewRHS = BO->getOperand(0);
1611 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1612 NewRHS = Sel->getOperand(1);
1613 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1614 NewRHS = Sel->getOperand(2);
1616 llvm_unreachable("Max doesn't match expected pattern!");
1618 // Determine the new comparison opcode. It may be signed or unsigned,
1619 // and the original comparison may be either equality or inequality.
1620 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1621 Pred = CmpInst::getInversePredicate(Pred);
1623 // Ok, everything looks ok to change the condition into an SLT or SGE and
1624 // delete the max calculation.
1626 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1628 // Delete the max calculation instructions.
1629 Cond->replaceAllUsesWith(NewCond);
1630 CondUse->setUser(NewCond);
1631 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1632 Cond->eraseFromParent();
1633 Sel->eraseFromParent();
1634 if (Cmp->use_empty())
1635 Cmp->eraseFromParent();
1639 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1640 /// postinc iv when possible.
1642 LSRInstance::OptimizeLoopTermCond() {
1643 SmallPtrSet<Instruction *, 4> PostIncs;
1645 BasicBlock *LatchBlock = L->getLoopLatch();
1646 SmallVector<BasicBlock*, 8> ExitingBlocks;
1647 L->getExitingBlocks(ExitingBlocks);
1649 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1650 BasicBlock *ExitingBlock = ExitingBlocks[i];
1652 // Get the terminating condition for the loop if possible. If we
1653 // can, we want to change it to use a post-incremented version of its
1654 // induction variable, to allow coalescing the live ranges for the IV into
1655 // one register value.
1657 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1660 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1661 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1664 // Search IVUsesByStride to find Cond's IVUse if there is one.
1665 IVStrideUse *CondUse = 0;
1666 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1667 if (!FindIVUserForCond(Cond, CondUse))
1670 // If the trip count is computed in terms of a max (due to ScalarEvolution
1671 // being unable to find a sufficient guard, for example), change the loop
1672 // comparison to use SLT or ULT instead of NE.
1673 // One consequence of doing this now is that it disrupts the count-down
1674 // optimization. That's not always a bad thing though, because in such
1675 // cases it may still be worthwhile to avoid a max.
1676 Cond = OptimizeMax(Cond, CondUse);
1678 // If this exiting block dominates the latch block, it may also use
1679 // the post-inc value if it won't be shared with other uses.
1680 // Check for dominance.
1681 if (!DT.dominates(ExitingBlock, LatchBlock))
1684 // Conservatively avoid trying to use the post-inc value in non-latch
1685 // exits if there may be pre-inc users in intervening blocks.
1686 if (LatchBlock != ExitingBlock)
1687 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1688 // Test if the use is reachable from the exiting block. This dominator
1689 // query is a conservative approximation of reachability.
1690 if (&*UI != CondUse &&
1691 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1692 // Conservatively assume there may be reuse if the quotient of their
1693 // strides could be a legal scale.
1694 const SCEV *A = IU.getStride(*CondUse, L);
1695 const SCEV *B = IU.getStride(*UI, L);
1696 if (!A || !B) continue;
1697 if (SE.getTypeSizeInBits(A->getType()) !=
1698 SE.getTypeSizeInBits(B->getType())) {
1699 if (SE.getTypeSizeInBits(A->getType()) >
1700 SE.getTypeSizeInBits(B->getType()))
1701 B = SE.getSignExtendExpr(B, A->getType());
1703 A = SE.getSignExtendExpr(A, B->getType());
1705 if (const SCEVConstant *D =
1706 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1707 // Stride of one or negative one can have reuse with non-addresses.
1708 if (D->getValue()->isOne() ||
1709 D->getValue()->isAllOnesValue())
1710 goto decline_post_inc;
1711 // Avoid weird situations.
1712 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1713 D->getValue()->getValue().isMinSignedValue())
1714 goto decline_post_inc;
1715 // Without TLI, assume that any stride might be valid, and so any
1716 // use might be shared.
1718 goto decline_post_inc;
1719 // Check for possible scaled-address reuse.
1720 const Type *AccessTy = getAccessType(UI->getUser());
1721 TargetLowering::AddrMode AM;
1722 AM.Scale = D->getValue()->getSExtValue();
1723 if (TLI->isLegalAddressingMode(AM, AccessTy))
1724 goto decline_post_inc;
1725 AM.Scale = -AM.Scale;
1726 if (TLI->isLegalAddressingMode(AM, AccessTy))
1727 goto decline_post_inc;
1731 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1734 // It's possible for the setcc instruction to be anywhere in the loop, and
1735 // possible for it to have multiple users. If it is not immediately before
1736 // the exiting block branch, move it.
1737 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1738 if (Cond->hasOneUse()) {
1739 Cond->moveBefore(TermBr);
1741 // Clone the terminating condition and insert into the loopend.
1742 ICmpInst *OldCond = Cond;
1743 Cond = cast<ICmpInst>(Cond->clone());
1744 Cond->setName(L->getHeader()->getName() + ".termcond");
1745 ExitingBlock->getInstList().insert(TermBr, Cond);
1747 // Clone the IVUse, as the old use still exists!
1748 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1749 TermBr->replaceUsesOfWith(OldCond, Cond);
1753 // If we get to here, we know that we can transform the setcc instruction to
1754 // use the post-incremented version of the IV, allowing us to coalesce the
1755 // live ranges for the IV correctly.
1756 CondUse->transformToPostInc(L);
1759 PostIncs.insert(Cond);
1763 // Determine an insertion point for the loop induction variable increment. It
1764 // must dominate all the post-inc comparisons we just set up, and it must
1765 // dominate the loop latch edge.
1766 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1767 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1768 E = PostIncs.end(); I != E; ++I) {
1770 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1772 if (BB == (*I)->getParent())
1773 IVIncInsertPos = *I;
1774 else if (BB != IVIncInsertPos->getParent())
1775 IVIncInsertPos = BB->getTerminator();
1780 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1781 LSRUse::KindType Kind, const Type *AccessTy) {
1782 int64_t NewMinOffset = LU.MinOffset;
1783 int64_t NewMaxOffset = LU.MaxOffset;
1784 const Type *NewAccessTy = AccessTy;
1786 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1787 // something conservative, however this can pessimize in the case that one of
1788 // the uses will have all its uses outside the loop, for example.
1789 if (LU.Kind != Kind)
1791 // Conservatively assume HasBaseReg is true for now.
1792 if (NewOffset < LU.MinOffset) {
1793 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1794 Kind, AccessTy, TLI))
1796 NewMinOffset = NewOffset;
1797 } else if (NewOffset > LU.MaxOffset) {
1798 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1799 Kind, AccessTy, TLI))
1801 NewMaxOffset = NewOffset;
1803 // Check for a mismatched access type, and fall back conservatively as needed.
1804 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1805 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1808 LU.MinOffset = NewMinOffset;
1809 LU.MaxOffset = NewMaxOffset;
1810 LU.AccessTy = NewAccessTy;
1811 if (NewOffset != LU.Offsets.back())
1812 LU.Offsets.push_back(NewOffset);
1816 /// getUse - Return an LSRUse index and an offset value for a fixup which
1817 /// needs the given expression, with the given kind and optional access type.
1818 /// Either reuse an existing use or create a new one, as needed.
1819 std::pair<size_t, int64_t>
1820 LSRInstance::getUse(const SCEV *&Expr,
1821 LSRUse::KindType Kind, const Type *AccessTy) {
1822 const SCEV *Copy = Expr;
1823 int64_t Offset = ExtractImmediate(Expr, SE);
1825 // Basic uses can't accept any offset, for example.
1826 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1831 std::pair<UseMapTy::iterator, bool> P =
1832 UseMap.insert(std::make_pair(Expr, 0));
1834 // A use already existed with this base.
1835 size_t LUIdx = P.first->second;
1836 LSRUse &LU = Uses[LUIdx];
1837 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1839 return std::make_pair(LUIdx, Offset);
1842 // Create a new use.
1843 size_t LUIdx = Uses.size();
1844 P.first->second = LUIdx;
1845 Uses.push_back(LSRUse(Kind, AccessTy));
1846 LSRUse &LU = Uses[LUIdx];
1848 // We don't need to track redundant offsets, but we don't need to go out
1849 // of our way here to avoid them.
1850 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1851 LU.Offsets.push_back(Offset);
1853 LU.MinOffset = Offset;
1854 LU.MaxOffset = Offset;
1855 return std::make_pair(LUIdx, Offset);
1858 /// DeleteUse - Delete the given use from the Uses list.
1859 void LSRInstance::DeleteUse(LSRUse &LU) {
1860 if (&LU != &Uses.back())
1861 std::swap(LU, Uses.back());
1865 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1866 /// a formula that has the same registers as the given formula.
1868 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1869 const LSRUse &OrigLU) {
1870 // Search all uses for the formula. This could be more clever. Ignore
1871 // ICmpZero uses because they may contain formulae generated by
1872 // GenerateICmpZeroScales, in which case adding fixup offsets may
1874 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1875 LSRUse &LU = Uses[LUIdx];
1876 if (&LU != &OrigLU &&
1877 LU.Kind != LSRUse::ICmpZero &&
1878 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1879 LU.HasFormulaWithSameRegs(OrigF)) {
1880 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
1881 FIdx != NumForms; ++FIdx) {
1882 Formula &F = LU.Formulae[FIdx];
1883 if (F.BaseRegs == OrigF.BaseRegs &&
1884 F.ScaledReg == OrigF.ScaledReg &&
1885 F.AM.BaseGV == OrigF.AM.BaseGV &&
1886 F.AM.Scale == OrigF.AM.Scale &&
1888 if (F.AM.BaseOffs == 0)
1899 void LSRInstance::CollectInterestingTypesAndFactors() {
1900 SmallSetVector<const SCEV *, 4> Strides;
1902 // Collect interesting types and strides.
1903 SmallVector<const SCEV *, 4> Worklist;
1904 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1905 const SCEV *Expr = IU.getExpr(*UI);
1907 // Collect interesting types.
1908 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1910 // Add strides for mentioned loops.
1911 Worklist.push_back(Expr);
1913 const SCEV *S = Worklist.pop_back_val();
1914 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1915 Strides.insert(AR->getStepRecurrence(SE));
1916 Worklist.push_back(AR->getStart());
1917 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1918 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1920 } while (!Worklist.empty());
1923 // Compute interesting factors from the set of interesting strides.
1924 for (SmallSetVector<const SCEV *, 4>::const_iterator
1925 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1926 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1927 next(I); NewStrideIter != E; ++NewStrideIter) {
1928 const SCEV *OldStride = *I;
1929 const SCEV *NewStride = *NewStrideIter;
1931 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1932 SE.getTypeSizeInBits(NewStride->getType())) {
1933 if (SE.getTypeSizeInBits(OldStride->getType()) >
1934 SE.getTypeSizeInBits(NewStride->getType()))
1935 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1937 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1939 if (const SCEVConstant *Factor =
1940 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1942 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1943 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1944 } else if (const SCEVConstant *Factor =
1945 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1948 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1949 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1953 // If all uses use the same type, don't bother looking for truncation-based
1955 if (Types.size() == 1)
1958 DEBUG(print_factors_and_types(dbgs()));
1961 void LSRInstance::CollectFixupsAndInitialFormulae() {
1962 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1964 LSRFixup &LF = getNewFixup();
1965 LF.UserInst = UI->getUser();
1966 LF.OperandValToReplace = UI->getOperandValToReplace();
1967 LF.PostIncLoops = UI->getPostIncLoops();
1969 LSRUse::KindType Kind = LSRUse::Basic;
1970 const Type *AccessTy = 0;
1971 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1972 Kind = LSRUse::Address;
1973 AccessTy = getAccessType(LF.UserInst);
1976 const SCEV *S = IU.getExpr(*UI);
1978 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1979 // (N - i == 0), and this allows (N - i) to be the expression that we work
1980 // with rather than just N or i, so we can consider the register
1981 // requirements for both N and i at the same time. Limiting this code to
1982 // equality icmps is not a problem because all interesting loops use
1983 // equality icmps, thanks to IndVarSimplify.
1984 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1985 if (CI->isEquality()) {
1986 // Swap the operands if needed to put the OperandValToReplace on the
1987 // left, for consistency.
1988 Value *NV = CI->getOperand(1);
1989 if (NV == LF.OperandValToReplace) {
1990 CI->setOperand(1, CI->getOperand(0));
1991 CI->setOperand(0, NV);
1992 NV = CI->getOperand(1);
1996 // x == y --> x - y == 0
1997 const SCEV *N = SE.getSCEV(NV);
1998 if (N->isLoopInvariant(L)) {
1999 Kind = LSRUse::ICmpZero;
2000 S = SE.getMinusSCEV(N, S);
2003 // -1 and the negations of all interesting strides (except the negation
2004 // of -1) are now also interesting.
2005 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2006 if (Factors[i] != -1)
2007 Factors.insert(-(uint64_t)Factors[i]);
2011 // Set up the initial formula for this use.
2012 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2014 LF.Offset = P.second;
2015 LSRUse &LU = Uses[LF.LUIdx];
2016 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2018 // If this is the first use of this LSRUse, give it a formula.
2019 if (LU.Formulae.empty()) {
2020 InsertInitialFormula(S, LU, LF.LUIdx);
2021 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2025 DEBUG(print_fixups(dbgs()));
2029 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2031 F.InitialMatch(S, L, SE, DT);
2032 bool Inserted = InsertFormula(LU, LUIdx, F);
2033 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2037 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2038 LSRUse &LU, size_t LUIdx) {
2040 F.BaseRegs.push_back(S);
2041 F.AM.HasBaseReg = true;
2042 bool Inserted = InsertFormula(LU, LUIdx, F);
2043 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2046 /// CountRegisters - Note which registers are used by the given formula,
2047 /// updating RegUses.
2048 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2050 RegUses.CountRegister(F.ScaledReg, LUIdx);
2051 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2052 E = F.BaseRegs.end(); I != E; ++I)
2053 RegUses.CountRegister(*I, LUIdx);
2056 /// InsertFormula - If the given formula has not yet been inserted, add it to
2057 /// the list, and return true. Return false otherwise.
2058 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2059 if (!LU.InsertFormula(F))
2062 CountRegisters(F, LUIdx);
2066 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2067 /// loop-invariant values which we're tracking. These other uses will pin these
2068 /// values in registers, making them less profitable for elimination.
2069 /// TODO: This currently misses non-constant addrec step registers.
2070 /// TODO: Should this give more weight to users inside the loop?
2072 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2073 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2074 SmallPtrSet<const SCEV *, 8> Inserted;
2076 while (!Worklist.empty()) {
2077 const SCEV *S = Worklist.pop_back_val();
2079 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2080 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
2081 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2082 Worklist.push_back(C->getOperand());
2083 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2084 Worklist.push_back(D->getLHS());
2085 Worklist.push_back(D->getRHS());
2086 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2087 if (!Inserted.insert(U)) continue;
2088 const Value *V = U->getValue();
2089 if (const Instruction *Inst = dyn_cast<Instruction>(V))
2090 if (L->contains(Inst)) continue;
2091 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2093 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2094 // Ignore non-instructions.
2097 // Ignore instructions in other functions (as can happen with
2099 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2101 // Ignore instructions not dominated by the loop.
2102 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2103 UserInst->getParent() :
2104 cast<PHINode>(UserInst)->getIncomingBlock(
2105 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2106 if (!DT.dominates(L->getHeader(), UseBB))
2108 // Ignore uses which are part of other SCEV expressions, to avoid
2109 // analyzing them multiple times.
2110 if (SE.isSCEVable(UserInst->getType())) {
2111 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2112 // If the user is a no-op, look through to its uses.
2113 if (!isa<SCEVUnknown>(UserS))
2117 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2121 // Ignore icmp instructions which are already being analyzed.
2122 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2123 unsigned OtherIdx = !UI.getOperandNo();
2124 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2125 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2129 LSRFixup &LF = getNewFixup();
2130 LF.UserInst = const_cast<Instruction *>(UserInst);
2131 LF.OperandValToReplace = UI.getUse();
2132 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2134 LF.Offset = P.second;
2135 LSRUse &LU = Uses[LF.LUIdx];
2136 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2137 InsertSupplementalFormula(U, LU, LF.LUIdx);
2138 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2145 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2146 /// separate registers. If C is non-null, multiply each subexpression by C.
2147 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2148 SmallVectorImpl<const SCEV *> &Ops,
2149 ScalarEvolution &SE) {
2150 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2151 // Break out add operands.
2152 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2154 CollectSubexprs(*I, C, Ops, SE);
2156 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2157 // Split a non-zero base out of an addrec.
2158 if (!AR->getStart()->isZero()) {
2159 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2160 AR->getStepRecurrence(SE),
2161 AR->getLoop()), C, Ops, SE);
2162 CollectSubexprs(AR->getStart(), C, Ops, SE);
2165 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2166 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2167 if (Mul->getNumOperands() == 2)
2168 if (const SCEVConstant *Op0 =
2169 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2170 CollectSubexprs(Mul->getOperand(1),
2171 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2177 // Otherwise use the value itself.
2178 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2181 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2183 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2186 // Arbitrarily cap recursion to protect compile time.
2187 if (Depth >= 3) return;
2189 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2190 const SCEV *BaseReg = Base.BaseRegs[i];
2192 SmallVector<const SCEV *, 8> AddOps;
2193 CollectSubexprs(BaseReg, 0, AddOps, SE);
2194 if (AddOps.size() == 1) continue;
2196 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2197 JE = AddOps.end(); J != JE; ++J) {
2198 // Don't pull a constant into a register if the constant could be folded
2199 // into an immediate field.
2200 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2201 Base.getNumRegs() > 1,
2202 LU.Kind, LU.AccessTy, TLI, SE))
2205 // Collect all operands except *J.
2206 SmallVector<const SCEV *, 8> InnerAddOps;
2207 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2208 KE = AddOps.end(); K != KE; ++K)
2210 InnerAddOps.push_back(*K);
2212 // Don't leave just a constant behind in a register if the constant could
2213 // be folded into an immediate field.
2214 if (InnerAddOps.size() == 1 &&
2215 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2216 Base.getNumRegs() > 1,
2217 LU.Kind, LU.AccessTy, TLI, SE))
2220 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2221 if (InnerSum->isZero())
2224 F.BaseRegs[i] = InnerSum;
2225 F.BaseRegs.push_back(*J);
2226 if (InsertFormula(LU, LUIdx, F))
2227 // If that formula hadn't been seen before, recurse to find more like
2229 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2234 /// GenerateCombinations - Generate a formula consisting of all of the
2235 /// loop-dominating registers added into a single register.
2236 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2238 // This method is only interesting on a plurality of registers.
2239 if (Base.BaseRegs.size() <= 1) return;
2243 SmallVector<const SCEV *, 4> Ops;
2244 for (SmallVectorImpl<const SCEV *>::const_iterator
2245 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2246 const SCEV *BaseReg = *I;
2247 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2248 !BaseReg->hasComputableLoopEvolution(L))
2249 Ops.push_back(BaseReg);
2251 F.BaseRegs.push_back(BaseReg);
2253 if (Ops.size() > 1) {
2254 const SCEV *Sum = SE.getAddExpr(Ops);
2255 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2256 // opportunity to fold something. For now, just ignore such cases
2257 // rather than proceed with zero in a register.
2258 if (!Sum->isZero()) {
2259 F.BaseRegs.push_back(Sum);
2260 (void)InsertFormula(LU, LUIdx, F);
2265 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2266 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2268 // We can't add a symbolic offset if the address already contains one.
2269 if (Base.AM.BaseGV) return;
2271 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2272 const SCEV *G = Base.BaseRegs[i];
2273 GlobalValue *GV = ExtractSymbol(G, SE);
2274 if (G->isZero() || !GV)
2278 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2279 LU.Kind, LU.AccessTy, TLI))
2282 (void)InsertFormula(LU, LUIdx, F);
2286 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2287 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2289 // TODO: For now, just add the min and max offset, because it usually isn't
2290 // worthwhile looking at everything inbetween.
2291 SmallVector<int64_t, 4> Worklist;
2292 Worklist.push_back(LU.MinOffset);
2293 if (LU.MaxOffset != LU.MinOffset)
2294 Worklist.push_back(LU.MaxOffset);
2296 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2297 const SCEV *G = Base.BaseRegs[i];
2299 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2300 E = Worklist.end(); I != E; ++I) {
2302 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2303 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2304 LU.Kind, LU.AccessTy, TLI)) {
2305 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2307 (void)InsertFormula(LU, LUIdx, F);
2311 int64_t Imm = ExtractImmediate(G, SE);
2312 if (G->isZero() || Imm == 0)
2315 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2316 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2317 LU.Kind, LU.AccessTy, TLI))
2320 (void)InsertFormula(LU, LUIdx, F);
2324 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2325 /// the comparison. For example, x == y -> x*c == y*c.
2326 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2328 if (LU.Kind != LSRUse::ICmpZero) return;
2330 // Determine the integer type for the base formula.
2331 const Type *IntTy = Base.getType();
2333 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2335 // Don't do this if there is more than one offset.
2336 if (LU.MinOffset != LU.MaxOffset) return;
2338 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2340 // Check each interesting stride.
2341 for (SmallSetVector<int64_t, 8>::const_iterator
2342 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2343 int64_t Factor = *I;
2346 // Check that the multiplication doesn't overflow.
2347 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2349 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2350 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2353 // Check that multiplying with the use offset doesn't overflow.
2354 int64_t Offset = LU.MinOffset;
2355 if (Offset == INT64_MIN && Factor == -1)
2357 Offset = (uint64_t)Offset * Factor;
2358 if (Offset / Factor != LU.MinOffset)
2361 // Check that this scale is legal.
2362 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2365 // Compensate for the use having MinOffset built into it.
2366 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2368 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2370 // Check that multiplying with each base register doesn't overflow.
2371 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2372 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2373 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2377 // Check that multiplying with the scaled register doesn't overflow.
2379 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2380 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2384 // If we make it here and it's legal, add it.
2385 (void)InsertFormula(LU, LUIdx, F);
2390 /// GenerateScales - Generate stride factor reuse formulae by making use of
2391 /// scaled-offset address modes, for example.
2392 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2393 // Determine the integer type for the base formula.
2394 const Type *IntTy = Base.getType();
2397 // If this Formula already has a scaled register, we can't add another one.
2398 if (Base.AM.Scale != 0) return;
2400 // Check each interesting stride.
2401 for (SmallSetVector<int64_t, 8>::const_iterator
2402 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2403 int64_t Factor = *I;
2405 Base.AM.Scale = Factor;
2406 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2407 // Check whether this scale is going to be legal.
2408 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2409 LU.Kind, LU.AccessTy, TLI)) {
2410 // As a special-case, handle special out-of-loop Basic users specially.
2411 // TODO: Reconsider this special case.
2412 if (LU.Kind == LSRUse::Basic &&
2413 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2414 LSRUse::Special, LU.AccessTy, TLI) &&
2415 LU.AllFixupsOutsideLoop)
2416 LU.Kind = LSRUse::Special;
2420 // For an ICmpZero, negating a solitary base register won't lead to
2422 if (LU.Kind == LSRUse::ICmpZero &&
2423 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2425 // For each addrec base reg, apply the scale, if possible.
2426 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2427 if (const SCEVAddRecExpr *AR =
2428 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2429 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2430 if (FactorS->isZero())
2432 // Divide out the factor, ignoring high bits, since we'll be
2433 // scaling the value back up in the end.
2434 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2435 // TODO: This could be optimized to avoid all the copying.
2437 F.ScaledReg = Quotient;
2438 F.DeleteBaseReg(F.BaseRegs[i]);
2439 (void)InsertFormula(LU, LUIdx, F);
2445 /// GenerateTruncates - Generate reuse formulae from different IV types.
2446 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2447 // This requires TargetLowering to tell us which truncates are free.
2450 // Don't bother truncating symbolic values.
2451 if (Base.AM.BaseGV) return;
2453 // Determine the integer type for the base formula.
2454 const Type *DstTy = Base.getType();
2456 DstTy = SE.getEffectiveSCEVType(DstTy);
2458 for (SmallSetVector<const Type *, 4>::const_iterator
2459 I = Types.begin(), E = Types.end(); I != E; ++I) {
2460 const Type *SrcTy = *I;
2461 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2464 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2465 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2466 JE = F.BaseRegs.end(); J != JE; ++J)
2467 *J = SE.getAnyExtendExpr(*J, SrcTy);
2469 // TODO: This assumes we've done basic processing on all uses and
2470 // have an idea what the register usage is.
2471 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2474 (void)InsertFormula(LU, LUIdx, F);
2481 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2482 /// defer modifications so that the search phase doesn't have to worry about
2483 /// the data structures moving underneath it.
2487 const SCEV *OrigReg;
2489 WorkItem(size_t LI, int64_t I, const SCEV *R)
2490 : LUIdx(LI), Imm(I), OrigReg(R) {}
2492 void print(raw_ostream &OS) const;
2498 void WorkItem::print(raw_ostream &OS) const {
2499 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2500 << " , add offset " << Imm;
2503 void WorkItem::dump() const {
2504 print(errs()); errs() << '\n';
2507 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2508 /// distance apart and try to form reuse opportunities between them.
2509 void LSRInstance::GenerateCrossUseConstantOffsets() {
2510 // Group the registers by their value without any added constant offset.
2511 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2512 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2514 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2515 SmallVector<const SCEV *, 8> Sequence;
2516 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2518 const SCEV *Reg = *I;
2519 int64_t Imm = ExtractImmediate(Reg, SE);
2520 std::pair<RegMapTy::iterator, bool> Pair =
2521 Map.insert(std::make_pair(Reg, ImmMapTy()));
2523 Sequence.push_back(Reg);
2524 Pair.first->second.insert(std::make_pair(Imm, *I));
2525 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2528 // Now examine each set of registers with the same base value. Build up
2529 // a list of work to do and do the work in a separate step so that we're
2530 // not adding formulae and register counts while we're searching.
2531 SmallVector<WorkItem, 32> WorkItems;
2532 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2533 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2534 E = Sequence.end(); I != E; ++I) {
2535 const SCEV *Reg = *I;
2536 const ImmMapTy &Imms = Map.find(Reg)->second;
2538 // It's not worthwhile looking for reuse if there's only one offset.
2539 if (Imms.size() == 1)
2542 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2543 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2545 dbgs() << ' ' << J->first;
2548 // Examine each offset.
2549 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2551 const SCEV *OrigReg = J->second;
2553 int64_t JImm = J->first;
2554 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2556 if (!isa<SCEVConstant>(OrigReg) &&
2557 UsedByIndicesMap[Reg].count() == 1) {
2558 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2562 // Conservatively examine offsets between this orig reg a few selected
2564 ImmMapTy::const_iterator OtherImms[] = {
2565 Imms.begin(), prior(Imms.end()),
2566 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2568 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2569 ImmMapTy::const_iterator M = OtherImms[i];
2570 if (M == J || M == JE) continue;
2572 // Compute the difference between the two.
2573 int64_t Imm = (uint64_t)JImm - M->first;
2574 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2575 LUIdx = UsedByIndices.find_next(LUIdx))
2576 // Make a memo of this use, offset, and register tuple.
2577 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2578 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2585 UsedByIndicesMap.clear();
2586 UniqueItems.clear();
2588 // Now iterate through the worklist and add new formulae.
2589 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2590 E = WorkItems.end(); I != E; ++I) {
2591 const WorkItem &WI = *I;
2592 size_t LUIdx = WI.LUIdx;
2593 LSRUse &LU = Uses[LUIdx];
2594 int64_t Imm = WI.Imm;
2595 const SCEV *OrigReg = WI.OrigReg;
2597 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2598 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2599 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2601 // TODO: Use a more targeted data structure.
2602 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2603 Formula F = LU.Formulae[L];
2604 // Use the immediate in the scaled register.
2605 if (F.ScaledReg == OrigReg) {
2606 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2607 Imm * (uint64_t)F.AM.Scale;
2608 // Don't create 50 + reg(-50).
2609 if (F.referencesReg(SE.getSCEV(
2610 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2613 NewF.AM.BaseOffs = Offs;
2614 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2615 LU.Kind, LU.AccessTy, TLI))
2617 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2619 // If the new scale is a constant in a register, and adding the constant
2620 // value to the immediate would produce a value closer to zero than the
2621 // immediate itself, then the formula isn't worthwhile.
2622 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2623 if (C->getValue()->getValue().isNegative() !=
2624 (NewF.AM.BaseOffs < 0) &&
2625 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2626 .ule(abs64(NewF.AM.BaseOffs)))
2630 (void)InsertFormula(LU, LUIdx, NewF);
2632 // Use the immediate in a base register.
2633 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2634 const SCEV *BaseReg = F.BaseRegs[N];
2635 if (BaseReg != OrigReg)
2638 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2639 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2640 LU.Kind, LU.AccessTy, TLI))
2642 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2644 // If the new formula has a constant in a register, and adding the
2645 // constant value to the immediate would produce a value closer to
2646 // zero than the immediate itself, then the formula isn't worthwhile.
2647 for (SmallVectorImpl<const SCEV *>::const_iterator
2648 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2650 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2651 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2652 abs64(NewF.AM.BaseOffs)) &&
2653 (C->getValue()->getValue() +
2654 NewF.AM.BaseOffs).countTrailingZeros() >=
2655 CountTrailingZeros_64(NewF.AM.BaseOffs))
2659 (void)InsertFormula(LU, LUIdx, NewF);
2668 /// GenerateAllReuseFormulae - Generate formulae for each use.
2670 LSRInstance::GenerateAllReuseFormulae() {
2671 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2672 // queries are more precise.
2673 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2674 LSRUse &LU = Uses[LUIdx];
2675 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2676 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2677 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2678 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2680 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2681 LSRUse &LU = Uses[LUIdx];
2682 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2683 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2684 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2685 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2686 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2687 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2688 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2689 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2691 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2692 LSRUse &LU = Uses[LUIdx];
2693 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2694 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2697 GenerateCrossUseConstantOffsets();
2700 /// If their are multiple formulae with the same set of registers used
2701 /// by other uses, pick the best one and delete the others.
2702 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2704 bool ChangedFormulae = false;
2707 // Collect the best formula for each unique set of shared registers. This
2708 // is reset for each use.
2709 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2711 BestFormulaeTy BestFormulae;
2713 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2714 LSRUse &LU = Uses[LUIdx];
2715 FormulaSorter Sorter(L, LU, SE, DT);
2716 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2719 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2720 FIdx != NumForms; ++FIdx) {
2721 Formula &F = LU.Formulae[FIdx];
2723 SmallVector<const SCEV *, 2> Key;
2724 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2725 JE = F.BaseRegs.end(); J != JE; ++J) {
2726 const SCEV *Reg = *J;
2727 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2731 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2732 Key.push_back(F.ScaledReg);
2733 // Unstable sort by host order ok, because this is only used for
2735 std::sort(Key.begin(), Key.end());
2737 std::pair<BestFormulaeTy::const_iterator, bool> P =
2738 BestFormulae.insert(std::make_pair(Key, FIdx));
2740 Formula &Best = LU.Formulae[P.first->second];
2741 if (Sorter.operator()(F, Best))
2743 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2745 " in favor of formula "; Best.print(dbgs());
2748 ChangedFormulae = true;
2750 LU.DeleteFormula(F);
2758 // Now that we've filtered out some formulae, recompute the Regs set.
2760 LU.RecomputeRegs(LUIdx, RegUses);
2762 // Reset this to prepare for the next use.
2763 BestFormulae.clear();
2766 DEBUG(if (ChangedFormulae) {
2768 "After filtering out undesirable candidates:\n";
2773 // This is a rough guess that seems to work fairly well.
2774 static const size_t ComplexityLimit = UINT16_MAX;
2776 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2777 /// solutions the solver might have to consider. It almost never considers
2778 /// this many solutions because it prune the search space, but the pruning
2779 /// isn't always sufficient.
2780 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2782 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2783 E = Uses.end(); I != E; ++I) {
2784 size_t FSize = I->Formulae.size();
2785 if (FSize >= ComplexityLimit) {
2786 Power = ComplexityLimit;
2790 if (Power >= ComplexityLimit)
2796 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2797 /// formulae to choose from, use some rough heuristics to prune down the number
2798 /// of formulae. This keeps the main solver from taking an extraordinary amount
2799 /// of time in some worst-case scenarios.
2800 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2801 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2802 DEBUG(dbgs() << "The search space is too complex.\n");
2804 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2805 "which use a superset of registers used by other "
2808 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2809 LSRUse &LU = Uses[LUIdx];
2811 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2812 Formula &F = LU.Formulae[i];
2813 // Look for a formula with a constant or GV in a register. If the use
2814 // also has a formula with that same value in an immediate field,
2815 // delete the one that uses a register.
2816 for (SmallVectorImpl<const SCEV *>::const_iterator
2817 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2818 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2820 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2821 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2822 (I - F.BaseRegs.begin()));
2823 if (LU.HasFormulaWithSameRegs(NewF)) {
2824 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2825 LU.DeleteFormula(F);
2831 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2832 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2835 NewF.AM.BaseGV = GV;
2836 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2837 (I - F.BaseRegs.begin()));
2838 if (LU.HasFormulaWithSameRegs(NewF)) {
2839 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2841 LU.DeleteFormula(F);
2852 LU.RecomputeRegs(LUIdx, RegUses);
2855 DEBUG(dbgs() << "After pre-selection:\n";
2856 print_uses(dbgs()));
2859 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2860 DEBUG(dbgs() << "The search space is too complex.\n");
2862 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2863 "separated by a constant offset will use the same "
2866 // This is especially useful for unrolled loops.
2868 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2869 LSRUse &LU = Uses[LUIdx];
2870 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2871 FIdx != NumForms; ++FIdx) {
2872 Formula &F = LU.Formulae[FIdx];
2873 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2874 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2875 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2876 /*HasBaseReg=*/false,
2877 LU.Kind, LU.AccessTy)) {
2878 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2881 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2883 // Delete formulae from the new use which are no longer legal.
2885 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2886 Formula &F = LUThatHas->Formulae[i];
2887 if (!isLegalUse(F.AM,
2888 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2889 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2890 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2892 LUThatHas->DeleteFormula(F);
2899 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
2901 // Update the relocs to reference the new use.
2902 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
2903 if (Fixups[i].LUIdx == LUIdx) {
2904 Fixups[i].LUIdx = LUThatHas - &Uses.front();
2905 Fixups[i].Offset += F.AM.BaseOffs;
2906 DEBUG(errs() << "New fixup has offset "
2907 << Fixups[i].Offset << '\n');
2909 if (Fixups[i].LUIdx == NumUses-1)
2910 Fixups[i].LUIdx = LUIdx;
2913 // Delete the old use.
2924 DEBUG(dbgs() << "After pre-selection:\n";
2925 print_uses(dbgs()));
2928 SmallPtrSet<const SCEV *, 4> Taken;
2929 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2930 // Ok, we have too many of formulae on our hands to conveniently handle.
2931 // Use a rough heuristic to thin out the list.
2932 DEBUG(dbgs() << "The search space is too complex.\n");
2934 // Pick the register which is used by the most LSRUses, which is likely
2935 // to be a good reuse register candidate.
2936 const SCEV *Best = 0;
2937 unsigned BestNum = 0;
2938 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2940 const SCEV *Reg = *I;
2941 if (Taken.count(Reg))
2946 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2947 if (Count > BestNum) {
2954 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2955 << " will yield profitable reuse.\n");
2958 // In any use with formulae which references this register, delete formulae
2959 // which don't reference it.
2960 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2961 LSRUse &LU = Uses[LUIdx];
2962 if (!LU.Regs.count(Best)) continue;
2965 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2966 Formula &F = LU.Formulae[i];
2967 if (!F.referencesReg(Best)) {
2968 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2969 LU.DeleteFormula(F);
2973 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
2979 LU.RecomputeRegs(LUIdx, RegUses);
2982 DEBUG(dbgs() << "After pre-selection:\n";
2983 print_uses(dbgs()));
2987 /// SolveRecurse - This is the recursive solver.
2988 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2990 SmallVectorImpl<const Formula *> &Workspace,
2991 const Cost &CurCost,
2992 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2993 DenseSet<const SCEV *> &VisitedRegs) const {
2996 // - use more aggressive filtering
2997 // - sort the formula so that the most profitable solutions are found first
2998 // - sort the uses too
3000 // - don't compute a cost, and then compare. compare while computing a cost
3002 // - track register sets with SmallBitVector
3004 const LSRUse &LU = Uses[Workspace.size()];
3006 // If this use references any register that's already a part of the
3007 // in-progress solution, consider it a requirement that a formula must
3008 // reference that register in order to be considered. This prunes out
3009 // unprofitable searching.
3010 SmallSetVector<const SCEV *, 4> ReqRegs;
3011 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3012 E = CurRegs.end(); I != E; ++I)
3013 if (LU.Regs.count(*I))
3016 bool AnySatisfiedReqRegs = false;
3017 SmallPtrSet<const SCEV *, 16> NewRegs;
3020 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3021 E = LU.Formulae.end(); I != E; ++I) {
3022 const Formula &F = *I;
3024 // Ignore formulae which do not use any of the required registers.
3025 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3026 JE = ReqRegs.end(); J != JE; ++J) {
3027 const SCEV *Reg = *J;
3028 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3029 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3033 AnySatisfiedReqRegs = true;
3035 // Evaluate the cost of the current formula. If it's already worse than
3036 // the current best, prune the search at that point.
3039 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3040 if (NewCost < SolutionCost) {
3041 Workspace.push_back(&F);
3042 if (Workspace.size() != Uses.size()) {
3043 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3044 NewRegs, VisitedRegs);
3045 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3046 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3048 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3049 dbgs() << ". Regs:";
3050 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3051 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3052 dbgs() << ' ' << **I;
3055 SolutionCost = NewCost;
3056 Solution = Workspace;
3058 Workspace.pop_back();
3063 // If none of the formulae had all of the required registers, relax the
3064 // constraint so that we don't exclude all formulae.
3065 if (!AnySatisfiedReqRegs) {
3066 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3072 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3073 SmallVector<const Formula *, 8> Workspace;
3075 SolutionCost.Loose();
3077 SmallPtrSet<const SCEV *, 16> CurRegs;
3078 DenseSet<const SCEV *> VisitedRegs;
3079 Workspace.reserve(Uses.size());
3081 // SolveRecurse does all the work.
3082 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3083 CurRegs, VisitedRegs);
3085 // Ok, we've now made all our decisions.
3086 DEBUG(dbgs() << "\n"
3087 "The chosen solution requires "; SolutionCost.print(dbgs());
3089 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3091 Uses[i].print(dbgs());
3094 Solution[i]->print(dbgs());
3099 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3100 /// the dominator tree far as we can go while still being dominated by the
3101 /// input positions. This helps canonicalize the insert position, which
3102 /// encourages sharing.
3103 BasicBlock::iterator
3104 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3105 const SmallVectorImpl<Instruction *> &Inputs)
3108 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3109 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3112 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3113 assert(Rung && "Block has no DomTreeNode!");
3114 Rung = Rung->getIDom();
3115 if (!Rung) return IP;
3116 IDom = Rung->getBlock();
3118 // Don't climb into a loop though.
3119 const Loop *IDomLoop = LI.getLoopFor(IDom);
3120 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3121 if (IDomDepth <= IPLoopDepth &&
3122 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3126 bool AllDominate = true;
3127 Instruction *BetterPos = 0;
3128 Instruction *Tentative = IDom->getTerminator();
3129 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3130 E = Inputs.end(); I != E; ++I) {
3131 Instruction *Inst = *I;
3132 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3133 AllDominate = false;
3136 // Attempt to find an insert position in the middle of the block,
3137 // instead of at the end, so that it can be used for other expansions.
3138 if (IDom == Inst->getParent() &&
3139 (!BetterPos || DT.dominates(BetterPos, Inst)))
3140 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3153 /// AdjustInsertPositionForExpand - Determine an input position which will be
3154 /// dominated by the operands and which will dominate the result.
3155 BasicBlock::iterator
3156 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3158 const LSRUse &LU) const {
3159 // Collect some instructions which must be dominated by the
3160 // expanding replacement. These must be dominated by any operands that
3161 // will be required in the expansion.
3162 SmallVector<Instruction *, 4> Inputs;
3163 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3164 Inputs.push_back(I);
3165 if (LU.Kind == LSRUse::ICmpZero)
3166 if (Instruction *I =
3167 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3168 Inputs.push_back(I);
3169 if (LF.PostIncLoops.count(L)) {
3170 if (LF.isUseFullyOutsideLoop(L))
3171 Inputs.push_back(L->getLoopLatch()->getTerminator());
3173 Inputs.push_back(IVIncInsertPos);
3175 // The expansion must also be dominated by the increment positions of any
3176 // loops it for which it is using post-inc mode.
3177 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3178 E = LF.PostIncLoops.end(); I != E; ++I) {
3179 const Loop *PIL = *I;
3180 if (PIL == L) continue;
3182 // Be dominated by the loop exit.
3183 SmallVector<BasicBlock *, 4> ExitingBlocks;
3184 PIL->getExitingBlocks(ExitingBlocks);
3185 if (!ExitingBlocks.empty()) {
3186 BasicBlock *BB = ExitingBlocks[0];
3187 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3188 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3189 Inputs.push_back(BB->getTerminator());
3193 // Then, climb up the immediate dominator tree as far as we can go while
3194 // still being dominated by the input positions.
3195 IP = HoistInsertPosition(IP, Inputs);
3197 // Don't insert instructions before PHI nodes.
3198 while (isa<PHINode>(IP)) ++IP;
3200 // Ignore debug intrinsics.
3201 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3206 Value *LSRInstance::Expand(const LSRFixup &LF,
3208 BasicBlock::iterator IP,
3209 SCEVExpander &Rewriter,
3210 SmallVectorImpl<WeakVH> &DeadInsts) const {
3211 const LSRUse &LU = Uses[LF.LUIdx];
3213 // Determine an input position which will be dominated by the operands and
3214 // which will dominate the result.
3215 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3217 // Inform the Rewriter if we have a post-increment use, so that it can
3218 // perform an advantageous expansion.
3219 Rewriter.setPostInc(LF.PostIncLoops);
3221 // This is the type that the user actually needs.
3222 const Type *OpTy = LF.OperandValToReplace->getType();
3223 // This will be the type that we'll initially expand to.
3224 const Type *Ty = F.getType();
3226 // No type known; just expand directly to the ultimate type.
3228 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3229 // Expand directly to the ultimate type if it's the right size.
3231 // This is the type to do integer arithmetic in.
3232 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3234 // Build up a list of operands to add together to form the full base.
3235 SmallVector<const SCEV *, 8> Ops;
3237 // Expand the BaseRegs portion.
3238 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3239 E = F.BaseRegs.end(); I != E; ++I) {
3240 const SCEV *Reg = *I;
3241 assert(!Reg->isZero() && "Zero allocated in a base register!");
3243 // If we're expanding for a post-inc user, make the post-inc adjustment.
3244 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3245 Reg = TransformForPostIncUse(Denormalize, Reg,
3246 LF.UserInst, LF.OperandValToReplace,
3249 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3252 // Flush the operand list to suppress SCEVExpander hoisting.
3254 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3256 Ops.push_back(SE.getUnknown(FullV));
3259 // Expand the ScaledReg portion.
3260 Value *ICmpScaledV = 0;
3261 if (F.AM.Scale != 0) {
3262 const SCEV *ScaledS = F.ScaledReg;
3264 // If we're expanding for a post-inc user, make the post-inc adjustment.
3265 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3266 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3267 LF.UserInst, LF.OperandValToReplace,
3270 if (LU.Kind == LSRUse::ICmpZero) {
3271 // An interesting way of "folding" with an icmp is to use a negated
3272 // scale, which we'll implement by inserting it into the other operand
3274 assert(F.AM.Scale == -1 &&
3275 "The only scale supported by ICmpZero uses is -1!");
3276 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3278 // Otherwise just expand the scaled register and an explicit scale,
3279 // which is expected to be matched as part of the address.
3280 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3281 ScaledS = SE.getMulExpr(ScaledS,
3282 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3283 Ops.push_back(ScaledS);
3285 // Flush the operand list to suppress SCEVExpander hoisting.
3286 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3288 Ops.push_back(SE.getUnknown(FullV));
3292 // Expand the GV portion.
3294 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3296 // Flush the operand list to suppress SCEVExpander hoisting.
3297 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3299 Ops.push_back(SE.getUnknown(FullV));
3302 // Expand the immediate portion.
3303 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3305 if (LU.Kind == LSRUse::ICmpZero) {
3306 // The other interesting way of "folding" with an ICmpZero is to use a
3307 // negated immediate.
3309 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3311 Ops.push_back(SE.getUnknown(ICmpScaledV));
3312 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3315 // Just add the immediate values. These again are expected to be matched
3316 // as part of the address.
3317 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3321 // Emit instructions summing all the operands.
3322 const SCEV *FullS = Ops.empty() ?
3323 SE.getConstant(IntTy, 0) :
3325 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3327 // We're done expanding now, so reset the rewriter.
3328 Rewriter.clearPostInc();
3330 // An ICmpZero Formula represents an ICmp which we're handling as a
3331 // comparison against zero. Now that we've expanded an expression for that
3332 // form, update the ICmp's other operand.
3333 if (LU.Kind == LSRUse::ICmpZero) {
3334 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3335 DeadInsts.push_back(CI->getOperand(1));
3336 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3337 "a scale at the same time!");
3338 if (F.AM.Scale == -1) {
3339 if (ICmpScaledV->getType() != OpTy) {
3341 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3343 ICmpScaledV, OpTy, "tmp", CI);
3346 CI->setOperand(1, ICmpScaledV);
3348 assert(F.AM.Scale == 0 &&
3349 "ICmp does not support folding a global value and "
3350 "a scale at the same time!");
3351 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3353 if (C->getType() != OpTy)
3354 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3358 CI->setOperand(1, C);
3365 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3366 /// of their operands effectively happens in their predecessor blocks, so the
3367 /// expression may need to be expanded in multiple places.
3368 void LSRInstance::RewriteForPHI(PHINode *PN,
3371 SCEVExpander &Rewriter,
3372 SmallVectorImpl<WeakVH> &DeadInsts,
3374 DenseMap<BasicBlock *, Value *> Inserted;
3375 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3376 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3377 BasicBlock *BB = PN->getIncomingBlock(i);
3379 // If this is a critical edge, split the edge so that we do not insert
3380 // the code on all predecessor/successor paths. We do this unless this
3381 // is the canonical backedge for this loop, which complicates post-inc
3383 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3384 !isa<IndirectBrInst>(BB->getTerminator()) &&
3385 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3386 // Split the critical edge.
3387 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3389 // If PN is outside of the loop and BB is in the loop, we want to
3390 // move the block to be immediately before the PHI block, not
3391 // immediately after BB.
3392 if (L->contains(BB) && !L->contains(PN))
3393 NewBB->moveBefore(PN->getParent());
3395 // Splitting the edge can reduce the number of PHI entries we have.
3396 e = PN->getNumIncomingValues();
3398 i = PN->getBasicBlockIndex(BB);
3401 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3402 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3404 PN->setIncomingValue(i, Pair.first->second);
3406 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3408 // If this is reuse-by-noop-cast, insert the noop cast.
3409 const Type *OpTy = LF.OperandValToReplace->getType();
3410 if (FullV->getType() != OpTy)
3412 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3414 FullV, LF.OperandValToReplace->getType(),
3415 "tmp", BB->getTerminator());
3417 PN->setIncomingValue(i, FullV);
3418 Pair.first->second = FullV;
3423 /// Rewrite - Emit instructions for the leading candidate expression for this
3424 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3425 /// the newly expanded value.
3426 void LSRInstance::Rewrite(const LSRFixup &LF,
3428 SCEVExpander &Rewriter,
3429 SmallVectorImpl<WeakVH> &DeadInsts,
3431 // First, find an insertion point that dominates UserInst. For PHI nodes,
3432 // find the nearest block which dominates all the relevant uses.
3433 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3434 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3436 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3438 // If this is reuse-by-noop-cast, insert the noop cast.
3439 const Type *OpTy = LF.OperandValToReplace->getType();
3440 if (FullV->getType() != OpTy) {
3442 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3443 FullV, OpTy, "tmp", LF.UserInst);
3447 // Update the user. ICmpZero is handled specially here (for now) because
3448 // Expand may have updated one of the operands of the icmp already, and
3449 // its new value may happen to be equal to LF.OperandValToReplace, in
3450 // which case doing replaceUsesOfWith leads to replacing both operands
3451 // with the same value. TODO: Reorganize this.
3452 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3453 LF.UserInst->setOperand(0, FullV);
3455 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3458 DeadInsts.push_back(LF.OperandValToReplace);
3462 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3464 // Keep track of instructions we may have made dead, so that
3465 // we can remove them after we are done working.
3466 SmallVector<WeakVH, 16> DeadInsts;
3468 SCEVExpander Rewriter(SE);
3469 Rewriter.disableCanonicalMode();
3470 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3472 // Expand the new value definitions and update the users.
3473 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3474 size_t LUIdx = Fixups[i].LUIdx;
3476 Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
3481 // Clean up after ourselves. This must be done before deleting any
3485 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3488 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3489 : IU(P->getAnalysis<IVUsers>()),
3490 SE(P->getAnalysis<ScalarEvolution>()),
3491 DT(P->getAnalysis<DominatorTree>()),
3492 LI(P->getAnalysis<LoopInfo>()),
3493 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3495 // If LoopSimplify form is not available, stay out of trouble.
3496 if (!L->isLoopSimplifyForm()) return;
3498 // If there's no interesting work to be done, bail early.
3499 if (IU.empty()) return;
3501 DEBUG(dbgs() << "\nLSR on loop ";
3502 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3505 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3506 /// inside the loop then try to eliminate the cast operation.
3509 // Change loop terminating condition to use the postinc iv when possible.
3510 OptimizeLoopTermCond();
3512 CollectInterestingTypesAndFactors();
3513 CollectFixupsAndInitialFormulae();
3514 CollectLoopInvariantFixupsAndFormulae();
3516 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3517 print_uses(dbgs()));
3519 // Now use the reuse data to generate a bunch of interesting ways
3520 // to formulate the values needed for the uses.
3521 GenerateAllReuseFormulae();
3523 DEBUG(dbgs() << "\n"
3524 "After generating reuse formulae:\n";
3525 print_uses(dbgs()));
3527 FilterOutUndesirableDedicatedRegisters();
3528 NarrowSearchSpaceUsingHeuristics();
3530 SmallVector<const Formula *, 8> Solution;
3532 assert(Solution.size() == Uses.size() && "Malformed solution!");
3534 // Release memory that is no longer needed.
3540 // Formulae should be legal.
3541 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3542 E = Uses.end(); I != E; ++I) {
3543 const LSRUse &LU = *I;
3544 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3545 JE = LU.Formulae.end(); J != JE; ++J)
3546 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3547 LU.Kind, LU.AccessTy, TLI) &&
3548 "Illegal formula generated!");
3552 // Now that we've decided what we want, make it so.
3553 ImplementSolution(Solution, P);
3556 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3557 if (Factors.empty() && Types.empty()) return;
3559 OS << "LSR has identified the following interesting factors and types: ";
3562 for (SmallSetVector<int64_t, 8>::const_iterator
3563 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3564 if (!First) OS << ", ";
3569 for (SmallSetVector<const Type *, 4>::const_iterator
3570 I = Types.begin(), E = Types.end(); I != E; ++I) {
3571 if (!First) OS << ", ";
3573 OS << '(' << **I << ')';
3578 void LSRInstance::print_fixups(raw_ostream &OS) const {
3579 OS << "LSR is examining the following fixup sites:\n";
3580 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3581 E = Fixups.end(); I != E; ++I) {
3582 const LSRFixup &LF = *I;
3589 void LSRInstance::print_uses(raw_ostream &OS) const {
3590 OS << "LSR is examining the following uses:\n";
3591 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3592 E = Uses.end(); I != E; ++I) {
3593 const LSRUse &LU = *I;
3597 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3598 JE = LU.Formulae.end(); J != JE; ++J) {
3606 void LSRInstance::print(raw_ostream &OS) const {
3607 print_factors_and_types(OS);
3612 void LSRInstance::dump() const {
3613 print(errs()); errs() << '\n';
3618 class LoopStrengthReduce : public LoopPass {
3619 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3620 /// transformation profitability.
3621 const TargetLowering *const TLI;
3624 static char ID; // Pass ID, replacement for typeid
3625 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3628 bool runOnLoop(Loop *L, LPPassManager &LPM);
3629 void getAnalysisUsage(AnalysisUsage &AU) const;
3634 char LoopStrengthReduce::ID = 0;
3635 static RegisterPass<LoopStrengthReduce>
3636 X("loop-reduce", "Loop Strength Reduction");
3638 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3639 return new LoopStrengthReduce(TLI);
3642 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3643 : LoopPass(&ID), TLI(tli) {}
3645 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3646 // We split critical edges, so we change the CFG. However, we do update
3647 // many analyses if they are around.
3648 AU.addPreservedID(LoopSimplifyID);
3649 AU.addPreserved("domfrontier");
3651 AU.addRequired<LoopInfo>();
3652 AU.addPreserved<LoopInfo>();
3653 AU.addRequiredID(LoopSimplifyID);
3654 AU.addRequired<DominatorTree>();
3655 AU.addPreserved<DominatorTree>();
3656 AU.addRequired<ScalarEvolution>();
3657 AU.addPreserved<ScalarEvolution>();
3658 AU.addRequired<IVUsers>();
3659 AU.addPreserved<IVUsers>();
3662 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3663 bool Changed = false;
3665 // Run the main LSR transformation.
3666 Changed |= LSRInstance(TLI, L, this).getChanged();
3668 // At this point, it is worth checking to see if any recurrence PHIs are also
3669 // dead, so that we can remove them as well.
3670 Changed |= DeleteDeadPHIs(L->getHeader());