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;
981 /// HasFormula - Test whether this use as a formula which has the same
982 /// registers as the given formula.
983 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
984 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
985 if (F.ScaledReg) Key.push_back(F.ScaledReg);
986 // Unstable sort by host order ok, because this is only used for uniquifying.
987 std::sort(Key.begin(), Key.end());
988 return Uniquifier.count(Key);
991 /// InsertFormula - If the given formula has not yet been inserted, add it to
992 /// the list, and return true. Return false otherwise.
993 bool LSRUse::InsertFormula(const Formula &F) {
994 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
995 if (F.ScaledReg) Key.push_back(F.ScaledReg);
996 // Unstable sort by host order ok, because this is only used for uniquifying.
997 std::sort(Key.begin(), Key.end());
999 if (!Uniquifier.insert(Key).second)
1002 // Using a register to hold the value of 0 is not profitable.
1003 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1004 "Zero allocated in a scaled register!");
1006 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1007 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1008 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1011 // Add the formula to the list.
1012 Formulae.push_back(F);
1014 // Record registers now being used by this use.
1015 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1016 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1021 /// DeleteFormula - Remove the given formula from this use's list.
1022 void LSRUse::DeleteFormula(Formula &F) {
1023 if (&F != &Formulae.back())
1024 std::swap(F, Formulae.back());
1025 Formulae.pop_back();
1026 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1029 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1030 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1031 // Now that we've filtered out some formulae, recompute the Regs set.
1032 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1034 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1035 E = Formulae.end(); I != E; ++I) {
1036 const Formula &F = *I;
1037 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1038 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1041 // Update the RegTracker.
1042 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1043 E = OldRegs.end(); I != E; ++I)
1044 if (!Regs.count(*I))
1045 RegUses.DropRegister(*I, LUIdx);
1048 void LSRUse::print(raw_ostream &OS) const {
1049 OS << "LSR Use: Kind=";
1051 case Basic: OS << "Basic"; break;
1052 case Special: OS << "Special"; break;
1053 case ICmpZero: OS << "ICmpZero"; break;
1055 OS << "Address of ";
1056 if (AccessTy->isPointerTy())
1057 OS << "pointer"; // the full pointer type could be really verbose
1062 OS << ", Offsets={";
1063 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1064 E = Offsets.end(); I != E; ++I) {
1071 if (AllFixupsOutsideLoop)
1072 OS << ", all-fixups-outside-loop";
1075 void LSRUse::dump() const {
1076 print(errs()); errs() << '\n';
1079 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1080 /// be completely folded into the user instruction at isel time. This includes
1081 /// address-mode folding and special icmp tricks.
1082 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1083 LSRUse::KindType Kind, const Type *AccessTy,
1084 const TargetLowering *TLI) {
1086 case LSRUse::Address:
1087 // If we have low-level target information, ask the target if it can
1088 // completely fold this address.
1089 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1091 // Otherwise, just guess that reg+reg addressing is legal.
1092 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1094 case LSRUse::ICmpZero:
1095 // There's not even a target hook for querying whether it would be legal to
1096 // fold a GV into an ICmp.
1100 // ICmp only has two operands; don't allow more than two non-trivial parts.
1101 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1104 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1105 // putting the scaled register in the other operand of the icmp.
1106 if (AM.Scale != 0 && AM.Scale != -1)
1109 // If we have low-level target information, ask the target if it can fold an
1110 // integer immediate on an icmp.
1111 if (AM.BaseOffs != 0) {
1112 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1119 // Only handle single-register values.
1120 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1122 case LSRUse::Special:
1123 // Only handle -1 scales, or no scale.
1124 return AM.Scale == 0 || AM.Scale == -1;
1130 static bool isLegalUse(TargetLowering::AddrMode AM,
1131 int64_t MinOffset, int64_t MaxOffset,
1132 LSRUse::KindType Kind, const Type *AccessTy,
1133 const TargetLowering *TLI) {
1134 // Check for overflow.
1135 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1138 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1139 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1140 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1141 // Check for overflow.
1142 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1145 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1146 return isLegalUse(AM, Kind, AccessTy, TLI);
1151 static bool isAlwaysFoldable(int64_t BaseOffs,
1152 GlobalValue *BaseGV,
1154 LSRUse::KindType Kind, const Type *AccessTy,
1155 const TargetLowering *TLI) {
1156 // Fast-path: zero is always foldable.
1157 if (BaseOffs == 0 && !BaseGV) return true;
1159 // Conservatively, create an address with an immediate and a
1160 // base and a scale.
1161 TargetLowering::AddrMode AM;
1162 AM.BaseOffs = BaseOffs;
1164 AM.HasBaseReg = HasBaseReg;
1165 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1167 // Canonicalize a scale of 1 to a base register if the formula doesn't
1168 // already have a base register.
1169 if (!AM.HasBaseReg && AM.Scale == 1) {
1171 AM.HasBaseReg = true;
1174 return isLegalUse(AM, Kind, AccessTy, TLI);
1177 static bool isAlwaysFoldable(const SCEV *S,
1178 int64_t MinOffset, int64_t MaxOffset,
1180 LSRUse::KindType Kind, const Type *AccessTy,
1181 const TargetLowering *TLI,
1182 ScalarEvolution &SE) {
1183 // Fast-path: zero is always foldable.
1184 if (S->isZero()) return true;
1186 // Conservatively, create an address with an immediate and a
1187 // base and a scale.
1188 int64_t BaseOffs = ExtractImmediate(S, SE);
1189 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1191 // If there's anything else involved, it's not foldable.
1192 if (!S->isZero()) return false;
1194 // Fast-path: zero is always foldable.
1195 if (BaseOffs == 0 && !BaseGV) return true;
1197 // Conservatively, create an address with an immediate and a
1198 // base and a scale.
1199 TargetLowering::AddrMode AM;
1200 AM.BaseOffs = BaseOffs;
1202 AM.HasBaseReg = HasBaseReg;
1203 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1205 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1210 /// FormulaSorter - This class implements an ordering for formulae which sorts
1211 /// the by their standalone cost.
1212 class FormulaSorter {
1213 /// These two sets are kept empty, so that we compute standalone costs.
1214 DenseSet<const SCEV *> VisitedRegs;
1215 SmallPtrSet<const SCEV *, 16> Regs;
1218 ScalarEvolution &SE;
1222 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1223 : L(l), LU(&lu), SE(se), DT(dt) {}
1225 bool operator()(const Formula &A, const Formula &B) {
1227 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1230 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1232 return CostA < CostB;
1236 /// LSRInstance - This class holds state for the main loop strength reduction
1240 ScalarEvolution &SE;
1243 const TargetLowering *const TLI;
1247 /// IVIncInsertPos - This is the insert position that the current loop's
1248 /// induction variable increment should be placed. In simple loops, this is
1249 /// the latch block's terminator. But in more complicated cases, this is a
1250 /// position which will dominate all the in-loop post-increment users.
1251 Instruction *IVIncInsertPos;
1253 /// Factors - Interesting factors between use strides.
1254 SmallSetVector<int64_t, 8> Factors;
1256 /// Types - Interesting use types, to facilitate truncation reuse.
1257 SmallSetVector<const Type *, 4> Types;
1259 /// Fixups - The list of operands which are to be replaced.
1260 SmallVector<LSRFixup, 16> Fixups;
1262 /// Uses - The list of interesting uses.
1263 SmallVector<LSRUse, 16> Uses;
1265 /// RegUses - Track which uses use which register candidates.
1266 RegUseTracker RegUses;
1268 void OptimizeShadowIV();
1269 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1270 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1271 void OptimizeLoopTermCond();
1273 void CollectInterestingTypesAndFactors();
1274 void CollectFixupsAndInitialFormulae();
1276 LSRFixup &getNewFixup() {
1277 Fixups.push_back(LSRFixup());
1278 return Fixups.back();
1281 // Support for sharing of LSRUses between LSRFixups.
1282 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1285 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1286 LSRUse::KindType Kind, const Type *AccessTy);
1288 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1289 LSRUse::KindType Kind,
1290 const Type *AccessTy);
1292 void DeleteUse(LSRUse &LU);
1294 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1297 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1298 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1299 void CountRegisters(const Formula &F, size_t LUIdx);
1300 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1302 void CollectLoopInvariantFixupsAndFormulae();
1304 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1305 unsigned Depth = 0);
1306 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1307 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1308 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1309 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1310 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1311 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1312 void GenerateCrossUseConstantOffsets();
1313 void GenerateAllReuseFormulae();
1315 void FilterOutUndesirableDedicatedRegisters();
1317 size_t EstimateSearchSpaceComplexity() const;
1318 void NarrowSearchSpaceUsingHeuristics();
1320 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1322 SmallVectorImpl<const Formula *> &Workspace,
1323 const Cost &CurCost,
1324 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1325 DenseSet<const SCEV *> &VisitedRegs) const;
1326 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1328 BasicBlock::iterator
1329 HoistInsertPosition(BasicBlock::iterator IP,
1330 const SmallVectorImpl<Instruction *> &Inputs) const;
1331 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1333 const LSRUse &LU) const;
1335 Value *Expand(const LSRFixup &LF,
1337 BasicBlock::iterator IP,
1338 SCEVExpander &Rewriter,
1339 SmallVectorImpl<WeakVH> &DeadInsts) const;
1340 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1342 SCEVExpander &Rewriter,
1343 SmallVectorImpl<WeakVH> &DeadInsts,
1345 void Rewrite(const LSRFixup &LF,
1347 SCEVExpander &Rewriter,
1348 SmallVectorImpl<WeakVH> &DeadInsts,
1350 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1353 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1355 bool getChanged() const { return Changed; }
1357 void print_factors_and_types(raw_ostream &OS) const;
1358 void print_fixups(raw_ostream &OS) const;
1359 void print_uses(raw_ostream &OS) const;
1360 void print(raw_ostream &OS) const;
1366 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1367 /// inside the loop then try to eliminate the cast operation.
1368 void LSRInstance::OptimizeShadowIV() {
1369 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1370 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1373 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1374 UI != E; /* empty */) {
1375 IVUsers::const_iterator CandidateUI = UI;
1377 Instruction *ShadowUse = CandidateUI->getUser();
1378 const Type *DestTy = NULL;
1380 /* If shadow use is a int->float cast then insert a second IV
1381 to eliminate this cast.
1383 for (unsigned i = 0; i < n; ++i)
1389 for (unsigned i = 0; i < n; ++i, ++d)
1392 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1393 DestTy = UCast->getDestTy();
1394 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1395 DestTy = SCast->getDestTy();
1396 if (!DestTy) continue;
1399 // If target does not support DestTy natively then do not apply
1400 // this transformation.
1401 EVT DVT = TLI->getValueType(DestTy);
1402 if (!TLI->isTypeLegal(DVT)) continue;
1405 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1407 if (PH->getNumIncomingValues() != 2) continue;
1409 const Type *SrcTy = PH->getType();
1410 int Mantissa = DestTy->getFPMantissaWidth();
1411 if (Mantissa == -1) continue;
1412 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1415 unsigned Entry, Latch;
1416 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1424 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1425 if (!Init) continue;
1426 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1428 BinaryOperator *Incr =
1429 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1430 if (!Incr) continue;
1431 if (Incr->getOpcode() != Instruction::Add
1432 && Incr->getOpcode() != Instruction::Sub)
1435 /* Initialize new IV, double d = 0.0 in above example. */
1436 ConstantInt *C = NULL;
1437 if (Incr->getOperand(0) == PH)
1438 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1439 else if (Incr->getOperand(1) == PH)
1440 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1446 // Ignore negative constants, as the code below doesn't handle them
1447 // correctly. TODO: Remove this restriction.
1448 if (!C->getValue().isStrictlyPositive()) continue;
1450 /* Add new PHINode. */
1451 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1453 /* create new increment. '++d' in above example. */
1454 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1455 BinaryOperator *NewIncr =
1456 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1457 Instruction::FAdd : Instruction::FSub,
1458 NewPH, CFP, "IV.S.next.", Incr);
1460 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1461 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1463 /* Remove cast operation */
1464 ShadowUse->replaceAllUsesWith(NewPH);
1465 ShadowUse->eraseFromParent();
1471 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1472 /// set the IV user and stride information and return true, otherwise return
1474 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1475 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1476 if (UI->getUser() == Cond) {
1477 // NOTE: we could handle setcc instructions with multiple uses here, but
1478 // InstCombine does it as well for simple uses, it's not clear that it
1479 // occurs enough in real life to handle.
1486 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1487 /// a max computation.
1489 /// This is a narrow solution to a specific, but acute, problem. For loops
1495 /// } while (++i < n);
1497 /// the trip count isn't just 'n', because 'n' might not be positive. And
1498 /// unfortunately this can come up even for loops where the user didn't use
1499 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1500 /// will commonly be lowered like this:
1506 /// } while (++i < n);
1509 /// and then it's possible for subsequent optimization to obscure the if
1510 /// test in such a way that indvars can't find it.
1512 /// When indvars can't find the if test in loops like this, it creates a
1513 /// max expression, which allows it to give the loop a canonical
1514 /// induction variable:
1517 /// max = n < 1 ? 1 : n;
1520 /// } while (++i != max);
1522 /// Canonical induction variables are necessary because the loop passes
1523 /// are designed around them. The most obvious example of this is the
1524 /// LoopInfo analysis, which doesn't remember trip count values. It
1525 /// expects to be able to rediscover the trip count each time it is
1526 /// needed, and it does this using a simple analysis that only succeeds if
1527 /// the loop has a canonical induction variable.
1529 /// However, when it comes time to generate code, the maximum operation
1530 /// can be quite costly, especially if it's inside of an outer loop.
1532 /// This function solves this problem by detecting this type of loop and
1533 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1534 /// the instructions for the maximum computation.
1536 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1537 // Check that the loop matches the pattern we're looking for.
1538 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1539 Cond->getPredicate() != CmpInst::ICMP_NE)
1542 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1543 if (!Sel || !Sel->hasOneUse()) return Cond;
1545 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1546 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1548 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1550 // Add one to the backedge-taken count to get the trip count.
1551 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1552 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1554 // Check for a max calculation that matches the pattern. There's no check
1555 // for ICMP_ULE here because the comparison would be with zero, which
1556 // isn't interesting.
1557 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1558 const SCEVNAryExpr *Max = 0;
1559 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1560 Pred = ICmpInst::ICMP_SLE;
1562 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1563 Pred = ICmpInst::ICMP_SLT;
1565 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1566 Pred = ICmpInst::ICMP_ULT;
1573 // To handle a max with more than two operands, this optimization would
1574 // require additional checking and setup.
1575 if (Max->getNumOperands() != 2)
1578 const SCEV *MaxLHS = Max->getOperand(0);
1579 const SCEV *MaxRHS = Max->getOperand(1);
1581 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1582 // for a comparison with 1. For <= and >=, a comparison with zero.
1584 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1587 // Check the relevant induction variable for conformance to
1589 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1590 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1591 if (!AR || !AR->isAffine() ||
1592 AR->getStart() != One ||
1593 AR->getStepRecurrence(SE) != One)
1596 assert(AR->getLoop() == L &&
1597 "Loop condition operand is an addrec in a different loop!");
1599 // Check the right operand of the select, and remember it, as it will
1600 // be used in the new comparison instruction.
1602 if (ICmpInst::isTrueWhenEqual(Pred)) {
1603 // Look for n+1, and grab n.
1604 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
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);
1609 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1610 if (isa<ConstantInt>(BO->getOperand(1)) &&
1611 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1612 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1613 NewRHS = BO->getOperand(0);
1616 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1617 NewRHS = Sel->getOperand(1);
1618 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1619 NewRHS = Sel->getOperand(2);
1621 llvm_unreachable("Max doesn't match expected pattern!");
1623 // Determine the new comparison opcode. It may be signed or unsigned,
1624 // and the original comparison may be either equality or inequality.
1625 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1626 Pred = CmpInst::getInversePredicate(Pred);
1628 // Ok, everything looks ok to change the condition into an SLT or SGE and
1629 // delete the max calculation.
1631 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1633 // Delete the max calculation instructions.
1634 Cond->replaceAllUsesWith(NewCond);
1635 CondUse->setUser(NewCond);
1636 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1637 Cond->eraseFromParent();
1638 Sel->eraseFromParent();
1639 if (Cmp->use_empty())
1640 Cmp->eraseFromParent();
1644 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1645 /// postinc iv when possible.
1647 LSRInstance::OptimizeLoopTermCond() {
1648 SmallPtrSet<Instruction *, 4> PostIncs;
1650 BasicBlock *LatchBlock = L->getLoopLatch();
1651 SmallVector<BasicBlock*, 8> ExitingBlocks;
1652 L->getExitingBlocks(ExitingBlocks);
1654 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1655 BasicBlock *ExitingBlock = ExitingBlocks[i];
1657 // Get the terminating condition for the loop if possible. If we
1658 // can, we want to change it to use a post-incremented version of its
1659 // induction variable, to allow coalescing the live ranges for the IV into
1660 // one register value.
1662 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1665 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1666 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1669 // Search IVUsesByStride to find Cond's IVUse if there is one.
1670 IVStrideUse *CondUse = 0;
1671 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1672 if (!FindIVUserForCond(Cond, CondUse))
1675 // If the trip count is computed in terms of a max (due to ScalarEvolution
1676 // being unable to find a sufficient guard, for example), change the loop
1677 // comparison to use SLT or ULT instead of NE.
1678 // One consequence of doing this now is that it disrupts the count-down
1679 // optimization. That's not always a bad thing though, because in such
1680 // cases it may still be worthwhile to avoid a max.
1681 Cond = OptimizeMax(Cond, CondUse);
1683 // If this exiting block dominates the latch block, it may also use
1684 // the post-inc value if it won't be shared with other uses.
1685 // Check for dominance.
1686 if (!DT.dominates(ExitingBlock, LatchBlock))
1689 // Conservatively avoid trying to use the post-inc value in non-latch
1690 // exits if there may be pre-inc users in intervening blocks.
1691 if (LatchBlock != ExitingBlock)
1692 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1693 // Test if the use is reachable from the exiting block. This dominator
1694 // query is a conservative approximation of reachability.
1695 if (&*UI != CondUse &&
1696 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1697 // Conservatively assume there may be reuse if the quotient of their
1698 // strides could be a legal scale.
1699 const SCEV *A = IU.getStride(*CondUse, L);
1700 const SCEV *B = IU.getStride(*UI, L);
1701 if (!A || !B) continue;
1702 if (SE.getTypeSizeInBits(A->getType()) !=
1703 SE.getTypeSizeInBits(B->getType())) {
1704 if (SE.getTypeSizeInBits(A->getType()) >
1705 SE.getTypeSizeInBits(B->getType()))
1706 B = SE.getSignExtendExpr(B, A->getType());
1708 A = SE.getSignExtendExpr(A, B->getType());
1710 if (const SCEVConstant *D =
1711 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1712 const ConstantInt *C = D->getValue();
1713 // Stride of one or negative one can have reuse with non-addresses.
1714 if (C->isOne() || C->isAllOnesValue())
1715 goto decline_post_inc;
1716 // Avoid weird situations.
1717 if (C->getValue().getMinSignedBits() >= 64 ||
1718 C->getValue().isMinSignedValue())
1719 goto decline_post_inc;
1720 // Without TLI, assume that any stride might be valid, and so any
1721 // use might be shared.
1723 goto decline_post_inc;
1724 // Check for possible scaled-address reuse.
1725 const Type *AccessTy = getAccessType(UI->getUser());
1726 TargetLowering::AddrMode AM;
1727 AM.Scale = C->getSExtValue();
1728 if (TLI->isLegalAddressingMode(AM, AccessTy))
1729 goto decline_post_inc;
1730 AM.Scale = -AM.Scale;
1731 if (TLI->isLegalAddressingMode(AM, AccessTy))
1732 goto decline_post_inc;
1736 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1739 // It's possible for the setcc instruction to be anywhere in the loop, and
1740 // possible for it to have multiple users. If it is not immediately before
1741 // the exiting block branch, move it.
1742 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1743 if (Cond->hasOneUse()) {
1744 Cond->moveBefore(TermBr);
1746 // Clone the terminating condition and insert into the loopend.
1747 ICmpInst *OldCond = Cond;
1748 Cond = cast<ICmpInst>(Cond->clone());
1749 Cond->setName(L->getHeader()->getName() + ".termcond");
1750 ExitingBlock->getInstList().insert(TermBr, Cond);
1752 // Clone the IVUse, as the old use still exists!
1753 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1754 TermBr->replaceUsesOfWith(OldCond, Cond);
1758 // If we get to here, we know that we can transform the setcc instruction to
1759 // use the post-incremented version of the IV, allowing us to coalesce the
1760 // live ranges for the IV correctly.
1761 CondUse->transformToPostInc(L);
1764 PostIncs.insert(Cond);
1768 // Determine an insertion point for the loop induction variable increment. It
1769 // must dominate all the post-inc comparisons we just set up, and it must
1770 // dominate the loop latch edge.
1771 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1772 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1773 E = PostIncs.end(); I != E; ++I) {
1775 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1777 if (BB == (*I)->getParent())
1778 IVIncInsertPos = *I;
1779 else if (BB != IVIncInsertPos->getParent())
1780 IVIncInsertPos = BB->getTerminator();
1784 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1785 /// at the given offset and other details. If so, update the use and
1788 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1789 LSRUse::KindType Kind, const Type *AccessTy) {
1790 int64_t NewMinOffset = LU.MinOffset;
1791 int64_t NewMaxOffset = LU.MaxOffset;
1792 const Type *NewAccessTy = AccessTy;
1794 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1795 // something conservative, however this can pessimize in the case that one of
1796 // the uses will have all its uses outside the loop, for example.
1797 if (LU.Kind != Kind)
1799 // Conservatively assume HasBaseReg is true for now.
1800 if (NewOffset < LU.MinOffset) {
1801 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1802 Kind, AccessTy, TLI))
1804 NewMinOffset = NewOffset;
1805 } else if (NewOffset > LU.MaxOffset) {
1806 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1807 Kind, AccessTy, TLI))
1809 NewMaxOffset = NewOffset;
1811 // Check for a mismatched access type, and fall back conservatively as needed.
1812 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1813 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1816 LU.MinOffset = NewMinOffset;
1817 LU.MaxOffset = NewMaxOffset;
1818 LU.AccessTy = NewAccessTy;
1819 if (NewOffset != LU.Offsets.back())
1820 LU.Offsets.push_back(NewOffset);
1824 /// getUse - Return an LSRUse index and an offset value for a fixup which
1825 /// needs the given expression, with the given kind and optional access type.
1826 /// Either reuse an existing use or create a new one, as needed.
1827 std::pair<size_t, int64_t>
1828 LSRInstance::getUse(const SCEV *&Expr,
1829 LSRUse::KindType Kind, const Type *AccessTy) {
1830 const SCEV *Copy = Expr;
1831 int64_t Offset = ExtractImmediate(Expr, SE);
1833 // Basic uses can't accept any offset, for example.
1834 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1839 std::pair<UseMapTy::iterator, bool> P =
1840 UseMap.insert(std::make_pair(Expr, 0));
1842 // A use already existed with this base.
1843 size_t LUIdx = P.first->second;
1844 LSRUse &LU = Uses[LUIdx];
1845 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1847 return std::make_pair(LUIdx, Offset);
1850 // Create a new use.
1851 size_t LUIdx = Uses.size();
1852 P.first->second = LUIdx;
1853 Uses.push_back(LSRUse(Kind, AccessTy));
1854 LSRUse &LU = Uses[LUIdx];
1856 // We don't need to track redundant offsets, but we don't need to go out
1857 // of our way here to avoid them.
1858 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1859 LU.Offsets.push_back(Offset);
1861 LU.MinOffset = Offset;
1862 LU.MaxOffset = Offset;
1863 return std::make_pair(LUIdx, Offset);
1866 /// DeleteUse - Delete the given use from the Uses list.
1867 void LSRInstance::DeleteUse(LSRUse &LU) {
1868 if (&LU != &Uses.back())
1869 std::swap(LU, Uses.back());
1873 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1874 /// a formula that has the same registers as the given formula.
1876 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1877 const LSRUse &OrigLU) {
1878 // Search all uses for the formula. This could be more clever. Ignore
1879 // ICmpZero uses because they may contain formulae generated by
1880 // GenerateICmpZeroScales, in which case adding fixup offsets may
1882 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1883 LSRUse &LU = Uses[LUIdx];
1884 if (&LU != &OrigLU &&
1885 LU.Kind != LSRUse::ICmpZero &&
1886 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1887 LU.HasFormulaWithSameRegs(OrigF)) {
1888 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1889 E = LU.Formulae.end(); I != E; ++I) {
1890 const Formula &F = *I;
1891 if (F.BaseRegs == OrigF.BaseRegs &&
1892 F.ScaledReg == OrigF.ScaledReg &&
1893 F.AM.BaseGV == OrigF.AM.BaseGV &&
1894 F.AM.Scale == OrigF.AM.Scale &&
1896 if (F.AM.BaseOffs == 0)
1907 void LSRInstance::CollectInterestingTypesAndFactors() {
1908 SmallSetVector<const SCEV *, 4> Strides;
1910 // Collect interesting types and strides.
1911 SmallVector<const SCEV *, 4> Worklist;
1912 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1913 const SCEV *Expr = IU.getExpr(*UI);
1915 // Collect interesting types.
1916 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1918 // Add strides for mentioned loops.
1919 Worklist.push_back(Expr);
1921 const SCEV *S = Worklist.pop_back_val();
1922 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1923 Strides.insert(AR->getStepRecurrence(SE));
1924 Worklist.push_back(AR->getStart());
1925 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1926 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1928 } while (!Worklist.empty());
1931 // Compute interesting factors from the set of interesting strides.
1932 for (SmallSetVector<const SCEV *, 4>::const_iterator
1933 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1934 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1935 next(I); NewStrideIter != E; ++NewStrideIter) {
1936 const SCEV *OldStride = *I;
1937 const SCEV *NewStride = *NewStrideIter;
1939 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1940 SE.getTypeSizeInBits(NewStride->getType())) {
1941 if (SE.getTypeSizeInBits(OldStride->getType()) >
1942 SE.getTypeSizeInBits(NewStride->getType()))
1943 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1945 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1947 if (const SCEVConstant *Factor =
1948 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1950 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1951 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1952 } else if (const SCEVConstant *Factor =
1953 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1956 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1957 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1961 // If all uses use the same type, don't bother looking for truncation-based
1963 if (Types.size() == 1)
1966 DEBUG(print_factors_and_types(dbgs()));
1969 void LSRInstance::CollectFixupsAndInitialFormulae() {
1970 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1972 LSRFixup &LF = getNewFixup();
1973 LF.UserInst = UI->getUser();
1974 LF.OperandValToReplace = UI->getOperandValToReplace();
1975 LF.PostIncLoops = UI->getPostIncLoops();
1977 LSRUse::KindType Kind = LSRUse::Basic;
1978 const Type *AccessTy = 0;
1979 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1980 Kind = LSRUse::Address;
1981 AccessTy = getAccessType(LF.UserInst);
1984 const SCEV *S = IU.getExpr(*UI);
1986 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1987 // (N - i == 0), and this allows (N - i) to be the expression that we work
1988 // with rather than just N or i, so we can consider the register
1989 // requirements for both N and i at the same time. Limiting this code to
1990 // equality icmps is not a problem because all interesting loops use
1991 // equality icmps, thanks to IndVarSimplify.
1992 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1993 if (CI->isEquality()) {
1994 // Swap the operands if needed to put the OperandValToReplace on the
1995 // left, for consistency.
1996 Value *NV = CI->getOperand(1);
1997 if (NV == LF.OperandValToReplace) {
1998 CI->setOperand(1, CI->getOperand(0));
1999 CI->setOperand(0, NV);
2000 NV = CI->getOperand(1);
2004 // x == y --> x - y == 0
2005 const SCEV *N = SE.getSCEV(NV);
2006 if (N->isLoopInvariant(L)) {
2007 Kind = LSRUse::ICmpZero;
2008 S = SE.getMinusSCEV(N, S);
2011 // -1 and the negations of all interesting strides (except the negation
2012 // of -1) are now also interesting.
2013 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2014 if (Factors[i] != -1)
2015 Factors.insert(-(uint64_t)Factors[i]);
2019 // Set up the initial formula for this use.
2020 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2022 LF.Offset = P.second;
2023 LSRUse &LU = Uses[LF.LUIdx];
2024 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2026 // If this is the first use of this LSRUse, give it a formula.
2027 if (LU.Formulae.empty()) {
2028 InsertInitialFormula(S, LU, LF.LUIdx);
2029 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2033 DEBUG(print_fixups(dbgs()));
2036 /// InsertInitialFormula - Insert a formula for the given expression into
2037 /// the given use, separating out loop-variant portions from loop-invariant
2038 /// and loop-computable portions.
2040 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2042 F.InitialMatch(S, L, SE, DT);
2043 bool Inserted = InsertFormula(LU, LUIdx, F);
2044 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2047 /// InsertSupplementalFormula - Insert a simple single-register formula for
2048 /// the given expression into the given use.
2050 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2051 LSRUse &LU, size_t LUIdx) {
2053 F.BaseRegs.push_back(S);
2054 F.AM.HasBaseReg = true;
2055 bool Inserted = InsertFormula(LU, LUIdx, F);
2056 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2059 /// CountRegisters - Note which registers are used by the given formula,
2060 /// updating RegUses.
2061 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2063 RegUses.CountRegister(F.ScaledReg, LUIdx);
2064 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2065 E = F.BaseRegs.end(); I != E; ++I)
2066 RegUses.CountRegister(*I, LUIdx);
2069 /// InsertFormula - If the given formula has not yet been inserted, add it to
2070 /// the list, and return true. Return false otherwise.
2071 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2072 if (!LU.InsertFormula(F))
2075 CountRegisters(F, LUIdx);
2079 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2080 /// loop-invariant values which we're tracking. These other uses will pin these
2081 /// values in registers, making them less profitable for elimination.
2082 /// TODO: This currently misses non-constant addrec step registers.
2083 /// TODO: Should this give more weight to users inside the loop?
2085 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2086 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2087 SmallPtrSet<const SCEV *, 8> Inserted;
2089 while (!Worklist.empty()) {
2090 const SCEV *S = Worklist.pop_back_val();
2092 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2093 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
2094 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2095 Worklist.push_back(C->getOperand());
2096 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2097 Worklist.push_back(D->getLHS());
2098 Worklist.push_back(D->getRHS());
2099 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2100 if (!Inserted.insert(U)) continue;
2101 const Value *V = U->getValue();
2102 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2103 // Look for instructions defined outside the loop.
2104 if (L->contains(Inst)) continue;
2105 } else if (isa<UndefValue>(V))
2106 // Undef doesn't have a live range, so it doesn't matter.
2108 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2110 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2111 // Ignore non-instructions.
2114 // Ignore instructions in other functions (as can happen with
2116 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2118 // Ignore instructions not dominated by the loop.
2119 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2120 UserInst->getParent() :
2121 cast<PHINode>(UserInst)->getIncomingBlock(
2122 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2123 if (!DT.dominates(L->getHeader(), UseBB))
2125 // Ignore uses which are part of other SCEV expressions, to avoid
2126 // analyzing them multiple times.
2127 if (SE.isSCEVable(UserInst->getType())) {
2128 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2129 // If the user is a no-op, look through to its uses.
2130 if (!isa<SCEVUnknown>(UserS))
2134 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2138 // Ignore icmp instructions which are already being analyzed.
2139 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2140 unsigned OtherIdx = !UI.getOperandNo();
2141 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2142 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2146 LSRFixup &LF = getNewFixup();
2147 LF.UserInst = const_cast<Instruction *>(UserInst);
2148 LF.OperandValToReplace = UI.getUse();
2149 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2151 LF.Offset = P.second;
2152 LSRUse &LU = Uses[LF.LUIdx];
2153 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2154 InsertSupplementalFormula(U, LU, LF.LUIdx);
2155 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2162 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2163 /// separate registers. If C is non-null, multiply each subexpression by C.
2164 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2165 SmallVectorImpl<const SCEV *> &Ops,
2166 ScalarEvolution &SE) {
2167 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2168 // Break out add operands.
2169 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2171 CollectSubexprs(*I, C, Ops, SE);
2173 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2174 // Split a non-zero base out of an addrec.
2175 if (!AR->getStart()->isZero()) {
2176 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2177 AR->getStepRecurrence(SE),
2178 AR->getLoop()), C, Ops, SE);
2179 CollectSubexprs(AR->getStart(), C, Ops, SE);
2182 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2183 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2184 if (Mul->getNumOperands() == 2)
2185 if (const SCEVConstant *Op0 =
2186 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2187 CollectSubexprs(Mul->getOperand(1),
2188 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2194 // Otherwise use the value itself.
2195 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2198 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2200 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2203 // Arbitrarily cap recursion to protect compile time.
2204 if (Depth >= 3) return;
2206 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2207 const SCEV *BaseReg = Base.BaseRegs[i];
2209 SmallVector<const SCEV *, 8> AddOps;
2210 CollectSubexprs(BaseReg, 0, AddOps, SE);
2211 if (AddOps.size() == 1) continue;
2213 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2214 JE = AddOps.end(); J != JE; ++J) {
2215 // Don't pull a constant into a register if the constant could be folded
2216 // into an immediate field.
2217 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2218 Base.getNumRegs() > 1,
2219 LU.Kind, LU.AccessTy, TLI, SE))
2222 // Collect all operands except *J.
2223 SmallVector<const SCEV *, 8> InnerAddOps;
2224 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2225 KE = AddOps.end(); K != KE; ++K)
2227 InnerAddOps.push_back(*K);
2229 // Don't leave just a constant behind in a register if the constant could
2230 // be folded into an immediate field.
2231 if (InnerAddOps.size() == 1 &&
2232 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2233 Base.getNumRegs() > 1,
2234 LU.Kind, LU.AccessTy, TLI, SE))
2237 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2238 if (InnerSum->isZero())
2241 F.BaseRegs[i] = InnerSum;
2242 F.BaseRegs.push_back(*J);
2243 if (InsertFormula(LU, LUIdx, F))
2244 // If that formula hadn't been seen before, recurse to find more like
2246 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2251 /// GenerateCombinations - Generate a formula consisting of all of the
2252 /// loop-dominating registers added into a single register.
2253 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2255 // This method is only interesting on a plurality of registers.
2256 if (Base.BaseRegs.size() <= 1) return;
2260 SmallVector<const SCEV *, 4> Ops;
2261 for (SmallVectorImpl<const SCEV *>::const_iterator
2262 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2263 const SCEV *BaseReg = *I;
2264 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2265 !BaseReg->hasComputableLoopEvolution(L))
2266 Ops.push_back(BaseReg);
2268 F.BaseRegs.push_back(BaseReg);
2270 if (Ops.size() > 1) {
2271 const SCEV *Sum = SE.getAddExpr(Ops);
2272 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2273 // opportunity to fold something. For now, just ignore such cases
2274 // rather than proceed with zero in a register.
2275 if (!Sum->isZero()) {
2276 F.BaseRegs.push_back(Sum);
2277 (void)InsertFormula(LU, LUIdx, F);
2282 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2283 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2285 // We can't add a symbolic offset if the address already contains one.
2286 if (Base.AM.BaseGV) return;
2288 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2289 const SCEV *G = Base.BaseRegs[i];
2290 GlobalValue *GV = ExtractSymbol(G, SE);
2291 if (G->isZero() || !GV)
2295 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2296 LU.Kind, LU.AccessTy, TLI))
2299 (void)InsertFormula(LU, LUIdx, F);
2303 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2304 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2306 // TODO: For now, just add the min and max offset, because it usually isn't
2307 // worthwhile looking at everything inbetween.
2308 SmallVector<int64_t, 4> Worklist;
2309 Worklist.push_back(LU.MinOffset);
2310 if (LU.MaxOffset != LU.MinOffset)
2311 Worklist.push_back(LU.MaxOffset);
2313 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2314 const SCEV *G = Base.BaseRegs[i];
2316 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2317 E = Worklist.end(); I != E; ++I) {
2319 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2320 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2321 LU.Kind, LU.AccessTy, TLI)) {
2322 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2324 (void)InsertFormula(LU, LUIdx, F);
2328 int64_t Imm = ExtractImmediate(G, SE);
2329 if (G->isZero() || Imm == 0)
2332 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2333 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2334 LU.Kind, LU.AccessTy, TLI))
2337 (void)InsertFormula(LU, LUIdx, F);
2341 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2342 /// the comparison. For example, x == y -> x*c == y*c.
2343 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2345 if (LU.Kind != LSRUse::ICmpZero) return;
2347 // Determine the integer type for the base formula.
2348 const Type *IntTy = Base.getType();
2350 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2352 // Don't do this if there is more than one offset.
2353 if (LU.MinOffset != LU.MaxOffset) return;
2355 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2357 // Check each interesting stride.
2358 for (SmallSetVector<int64_t, 8>::const_iterator
2359 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2360 int64_t Factor = *I;
2363 // Check that the multiplication doesn't overflow.
2364 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2366 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2367 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2370 // Check that multiplying with the use offset doesn't overflow.
2371 int64_t Offset = LU.MinOffset;
2372 if (Offset == INT64_MIN && Factor == -1)
2374 Offset = (uint64_t)Offset * Factor;
2375 if (Offset / Factor != LU.MinOffset)
2378 // Check that this scale is legal.
2379 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2382 // Compensate for the use having MinOffset built into it.
2383 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2385 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2387 // Check that multiplying with each base register doesn't overflow.
2388 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2389 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2390 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2394 // Check that multiplying with the scaled register doesn't overflow.
2396 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2397 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2401 // If we make it here and it's legal, add it.
2402 (void)InsertFormula(LU, LUIdx, F);
2407 /// GenerateScales - Generate stride factor reuse formulae by making use of
2408 /// scaled-offset address modes, for example.
2409 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2410 // Determine the integer type for the base formula.
2411 const Type *IntTy = Base.getType();
2414 // If this Formula already has a scaled register, we can't add another one.
2415 if (Base.AM.Scale != 0) return;
2417 // Check each interesting stride.
2418 for (SmallSetVector<int64_t, 8>::const_iterator
2419 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2420 int64_t Factor = *I;
2422 Base.AM.Scale = Factor;
2423 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2424 // Check whether this scale is going to be legal.
2425 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2426 LU.Kind, LU.AccessTy, TLI)) {
2427 // As a special-case, handle special out-of-loop Basic users specially.
2428 // TODO: Reconsider this special case.
2429 if (LU.Kind == LSRUse::Basic &&
2430 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2431 LSRUse::Special, LU.AccessTy, TLI) &&
2432 LU.AllFixupsOutsideLoop)
2433 LU.Kind = LSRUse::Special;
2437 // For an ICmpZero, negating a solitary base register won't lead to
2439 if (LU.Kind == LSRUse::ICmpZero &&
2440 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2442 // For each addrec base reg, apply the scale, if possible.
2443 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2444 if (const SCEVAddRecExpr *AR =
2445 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2446 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2447 if (FactorS->isZero())
2449 // Divide out the factor, ignoring high bits, since we'll be
2450 // scaling the value back up in the end.
2451 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2452 // TODO: This could be optimized to avoid all the copying.
2454 F.ScaledReg = Quotient;
2455 F.DeleteBaseReg(F.BaseRegs[i]);
2456 (void)InsertFormula(LU, LUIdx, F);
2462 /// GenerateTruncates - Generate reuse formulae from different IV types.
2463 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2464 // This requires TargetLowering to tell us which truncates are free.
2467 // Don't bother truncating symbolic values.
2468 if (Base.AM.BaseGV) return;
2470 // Determine the integer type for the base formula.
2471 const Type *DstTy = Base.getType();
2473 DstTy = SE.getEffectiveSCEVType(DstTy);
2475 for (SmallSetVector<const Type *, 4>::const_iterator
2476 I = Types.begin(), E = Types.end(); I != E; ++I) {
2477 const Type *SrcTy = *I;
2478 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2481 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2482 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2483 JE = F.BaseRegs.end(); J != JE; ++J)
2484 *J = SE.getAnyExtendExpr(*J, SrcTy);
2486 // TODO: This assumes we've done basic processing on all uses and
2487 // have an idea what the register usage is.
2488 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2491 (void)InsertFormula(LU, LUIdx, F);
2498 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2499 /// defer modifications so that the search phase doesn't have to worry about
2500 /// the data structures moving underneath it.
2504 const SCEV *OrigReg;
2506 WorkItem(size_t LI, int64_t I, const SCEV *R)
2507 : LUIdx(LI), Imm(I), OrigReg(R) {}
2509 void print(raw_ostream &OS) const;
2515 void WorkItem::print(raw_ostream &OS) const {
2516 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2517 << " , add offset " << Imm;
2520 void WorkItem::dump() const {
2521 print(errs()); errs() << '\n';
2524 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2525 /// distance apart and try to form reuse opportunities between them.
2526 void LSRInstance::GenerateCrossUseConstantOffsets() {
2527 // Group the registers by their value without any added constant offset.
2528 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2529 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2531 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2532 SmallVector<const SCEV *, 8> Sequence;
2533 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2535 const SCEV *Reg = *I;
2536 int64_t Imm = ExtractImmediate(Reg, SE);
2537 std::pair<RegMapTy::iterator, bool> Pair =
2538 Map.insert(std::make_pair(Reg, ImmMapTy()));
2540 Sequence.push_back(Reg);
2541 Pair.first->second.insert(std::make_pair(Imm, *I));
2542 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2545 // Now examine each set of registers with the same base value. Build up
2546 // a list of work to do and do the work in a separate step so that we're
2547 // not adding formulae and register counts while we're searching.
2548 SmallVector<WorkItem, 32> WorkItems;
2549 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2550 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2551 E = Sequence.end(); I != E; ++I) {
2552 const SCEV *Reg = *I;
2553 const ImmMapTy &Imms = Map.find(Reg)->second;
2555 // It's not worthwhile looking for reuse if there's only one offset.
2556 if (Imms.size() == 1)
2559 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2560 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2562 dbgs() << ' ' << J->first;
2565 // Examine each offset.
2566 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2568 const SCEV *OrigReg = J->second;
2570 int64_t JImm = J->first;
2571 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2573 if (!isa<SCEVConstant>(OrigReg) &&
2574 UsedByIndicesMap[Reg].count() == 1) {
2575 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2579 // Conservatively examine offsets between this orig reg a few selected
2581 ImmMapTy::const_iterator OtherImms[] = {
2582 Imms.begin(), prior(Imms.end()),
2583 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2585 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2586 ImmMapTy::const_iterator M = OtherImms[i];
2587 if (M == J || M == JE) continue;
2589 // Compute the difference between the two.
2590 int64_t Imm = (uint64_t)JImm - M->first;
2591 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2592 LUIdx = UsedByIndices.find_next(LUIdx))
2593 // Make a memo of this use, offset, and register tuple.
2594 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2595 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2602 UsedByIndicesMap.clear();
2603 UniqueItems.clear();
2605 // Now iterate through the worklist and add new formulae.
2606 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2607 E = WorkItems.end(); I != E; ++I) {
2608 const WorkItem &WI = *I;
2609 size_t LUIdx = WI.LUIdx;
2610 LSRUse &LU = Uses[LUIdx];
2611 int64_t Imm = WI.Imm;
2612 const SCEV *OrigReg = WI.OrigReg;
2614 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2615 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2616 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2618 // TODO: Use a more targeted data structure.
2619 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2620 const Formula &F = LU.Formulae[L];
2621 // Use the immediate in the scaled register.
2622 if (F.ScaledReg == OrigReg) {
2623 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2624 Imm * (uint64_t)F.AM.Scale;
2625 // Don't create 50 + reg(-50).
2626 if (F.referencesReg(SE.getSCEV(
2627 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2630 NewF.AM.BaseOffs = Offs;
2631 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2632 LU.Kind, LU.AccessTy, TLI))
2634 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2636 // If the new scale is a constant in a register, and adding the constant
2637 // value to the immediate would produce a value closer to zero than the
2638 // immediate itself, then the formula isn't worthwhile.
2639 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2640 if (C->getValue()->getValue().isNegative() !=
2641 (NewF.AM.BaseOffs < 0) &&
2642 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2643 .ule(abs64(NewF.AM.BaseOffs)))
2647 (void)InsertFormula(LU, LUIdx, NewF);
2649 // Use the immediate in a base register.
2650 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2651 const SCEV *BaseReg = F.BaseRegs[N];
2652 if (BaseReg != OrigReg)
2655 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2656 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2657 LU.Kind, LU.AccessTy, TLI))
2659 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2661 // If the new formula has a constant in a register, and adding the
2662 // constant value to the immediate would produce a value closer to
2663 // zero than the immediate itself, then the formula isn't worthwhile.
2664 for (SmallVectorImpl<const SCEV *>::const_iterator
2665 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2667 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2668 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2669 abs64(NewF.AM.BaseOffs)) &&
2670 (C->getValue()->getValue() +
2671 NewF.AM.BaseOffs).countTrailingZeros() >=
2672 CountTrailingZeros_64(NewF.AM.BaseOffs))
2676 (void)InsertFormula(LU, LUIdx, NewF);
2685 /// GenerateAllReuseFormulae - Generate formulae for each use.
2687 LSRInstance::GenerateAllReuseFormulae() {
2688 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2689 // queries are more precise.
2690 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2691 LSRUse &LU = Uses[LUIdx];
2692 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2693 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2694 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2695 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2697 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2698 LSRUse &LU = Uses[LUIdx];
2699 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2700 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2701 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2702 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2703 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2704 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2705 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2706 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2708 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2709 LSRUse &LU = Uses[LUIdx];
2710 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2711 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2714 GenerateCrossUseConstantOffsets();
2717 /// If their are multiple formulae with the same set of registers used
2718 /// by other uses, pick the best one and delete the others.
2719 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2721 bool ChangedFormulae = false;
2724 // Collect the best formula for each unique set of shared registers. This
2725 // is reset for each use.
2726 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2728 BestFormulaeTy BestFormulae;
2730 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2731 LSRUse &LU = Uses[LUIdx];
2732 FormulaSorter Sorter(L, LU, SE, DT);
2733 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2736 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2737 FIdx != NumForms; ++FIdx) {
2738 Formula &F = LU.Formulae[FIdx];
2740 SmallVector<const SCEV *, 2> Key;
2741 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2742 JE = F.BaseRegs.end(); J != JE; ++J) {
2743 const SCEV *Reg = *J;
2744 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2748 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2749 Key.push_back(F.ScaledReg);
2750 // Unstable sort by host order ok, because this is only used for
2752 std::sort(Key.begin(), Key.end());
2754 std::pair<BestFormulaeTy::const_iterator, bool> P =
2755 BestFormulae.insert(std::make_pair(Key, FIdx));
2757 Formula &Best = LU.Formulae[P.first->second];
2758 if (Sorter.operator()(F, Best))
2760 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2762 " in favor of formula "; Best.print(dbgs());
2765 ChangedFormulae = true;
2767 LU.DeleteFormula(F);
2775 // Now that we've filtered out some formulae, recompute the Regs set.
2777 LU.RecomputeRegs(LUIdx, RegUses);
2779 // Reset this to prepare for the next use.
2780 BestFormulae.clear();
2783 DEBUG(if (ChangedFormulae) {
2785 "After filtering out undesirable candidates:\n";
2790 // This is a rough guess that seems to work fairly well.
2791 static const size_t ComplexityLimit = UINT16_MAX;
2793 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2794 /// solutions the solver might have to consider. It almost never considers
2795 /// this many solutions because it prune the search space, but the pruning
2796 /// isn't always sufficient.
2797 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2799 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2800 E = Uses.end(); I != E; ++I) {
2801 size_t FSize = I->Formulae.size();
2802 if (FSize >= ComplexityLimit) {
2803 Power = ComplexityLimit;
2807 if (Power >= ComplexityLimit)
2813 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2814 /// formulae to choose from, use some rough heuristics to prune down the number
2815 /// of formulae. This keeps the main solver from taking an extraordinary amount
2816 /// of time in some worst-case scenarios.
2817 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2818 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2819 DEBUG(dbgs() << "The search space is too complex.\n");
2821 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2822 "which use a superset of registers used by other "
2825 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2826 LSRUse &LU = Uses[LUIdx];
2828 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2829 Formula &F = LU.Formulae[i];
2830 // Look for a formula with a constant or GV in a register. If the use
2831 // also has a formula with that same value in an immediate field,
2832 // delete the one that uses a register.
2833 for (SmallVectorImpl<const SCEV *>::const_iterator
2834 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2835 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2837 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2838 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2839 (I - F.BaseRegs.begin()));
2840 if (LU.HasFormulaWithSameRegs(NewF)) {
2841 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2842 LU.DeleteFormula(F);
2848 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2849 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2852 NewF.AM.BaseGV = GV;
2853 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2854 (I - F.BaseRegs.begin()));
2855 if (LU.HasFormulaWithSameRegs(NewF)) {
2856 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2858 LU.DeleteFormula(F);
2869 LU.RecomputeRegs(LUIdx, RegUses);
2872 DEBUG(dbgs() << "After pre-selection:\n";
2873 print_uses(dbgs()));
2876 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2877 DEBUG(dbgs() << "The search space is too complex.\n");
2879 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2880 "separated by a constant offset will use the same "
2883 // This is especially useful for unrolled loops.
2885 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2886 LSRUse &LU = Uses[LUIdx];
2887 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2888 E = LU.Formulae.end(); I != E; ++I) {
2889 const Formula &F = *I;
2890 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2891 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2892 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2893 /*HasBaseReg=*/false,
2894 LU.Kind, LU.AccessTy)) {
2895 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2898 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2900 // Delete formulae from the new use which are no longer legal.
2902 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2903 Formula &F = LUThatHas->Formulae[i];
2904 if (!isLegalUse(F.AM,
2905 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2906 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2907 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2909 LUThatHas->DeleteFormula(F);
2916 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
2918 // Update the relocs to reference the new use.
2919 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
2920 E = Fixups.end(); I != E; ++I) {
2921 LSRFixup &Fixup = *I;
2922 if (Fixup.LUIdx == LUIdx) {
2923 Fixup.LUIdx = LUThatHas - &Uses.front();
2924 Fixup.Offset += F.AM.BaseOffs;
2925 DEBUG(errs() << "New fixup has offset "
2926 << Fixup.Offset << '\n');
2928 if (Fixup.LUIdx == NumUses-1)
2929 Fixup.LUIdx = LUIdx;
2932 // Delete the old use.
2943 DEBUG(dbgs() << "After pre-selection:\n";
2944 print_uses(dbgs()));
2947 // With all other options exhausted, loop until the system is simple
2948 // enough to handle.
2949 SmallPtrSet<const SCEV *, 4> Taken;
2950 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2951 // Ok, we have too many of formulae on our hands to conveniently handle.
2952 // Use a rough heuristic to thin out the list.
2953 DEBUG(dbgs() << "The search space is too complex.\n");
2955 // Pick the register which is used by the most LSRUses, which is likely
2956 // to be a good reuse register candidate.
2957 const SCEV *Best = 0;
2958 unsigned BestNum = 0;
2959 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2961 const SCEV *Reg = *I;
2962 if (Taken.count(Reg))
2967 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2968 if (Count > BestNum) {
2975 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2976 << " will yield profitable reuse.\n");
2979 // In any use with formulae which references this register, delete formulae
2980 // which don't reference it.
2981 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2982 LSRUse &LU = Uses[LUIdx];
2983 if (!LU.Regs.count(Best)) continue;
2986 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2987 Formula &F = LU.Formulae[i];
2988 if (!F.referencesReg(Best)) {
2989 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2990 LU.DeleteFormula(F);
2994 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3000 LU.RecomputeRegs(LUIdx, RegUses);
3003 DEBUG(dbgs() << "After pre-selection:\n";
3004 print_uses(dbgs()));
3008 /// SolveRecurse - This is the recursive solver.
3009 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3011 SmallVectorImpl<const Formula *> &Workspace,
3012 const Cost &CurCost,
3013 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3014 DenseSet<const SCEV *> &VisitedRegs) const {
3017 // - use more aggressive filtering
3018 // - sort the formula so that the most profitable solutions are found first
3019 // - sort the uses too
3021 // - don't compute a cost, and then compare. compare while computing a cost
3023 // - track register sets with SmallBitVector
3025 const LSRUse &LU = Uses[Workspace.size()];
3027 // If this use references any register that's already a part of the
3028 // in-progress solution, consider it a requirement that a formula must
3029 // reference that register in order to be considered. This prunes out
3030 // unprofitable searching.
3031 SmallSetVector<const SCEV *, 4> ReqRegs;
3032 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3033 E = CurRegs.end(); I != E; ++I)
3034 if (LU.Regs.count(*I))
3037 bool AnySatisfiedReqRegs = false;
3038 SmallPtrSet<const SCEV *, 16> NewRegs;
3041 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3042 E = LU.Formulae.end(); I != E; ++I) {
3043 const Formula &F = *I;
3045 // Ignore formulae which do not use any of the required registers.
3046 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3047 JE = ReqRegs.end(); J != JE; ++J) {
3048 const SCEV *Reg = *J;
3049 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3050 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3054 AnySatisfiedReqRegs = true;
3056 // Evaluate the cost of the current formula. If it's already worse than
3057 // the current best, prune the search at that point.
3060 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3061 if (NewCost < SolutionCost) {
3062 Workspace.push_back(&F);
3063 if (Workspace.size() != Uses.size()) {
3064 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3065 NewRegs, VisitedRegs);
3066 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3067 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3069 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3070 dbgs() << ". Regs:";
3071 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3072 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3073 dbgs() << ' ' << **I;
3076 SolutionCost = NewCost;
3077 Solution = Workspace;
3079 Workspace.pop_back();
3084 // If none of the formulae had all of the required registers, relax the
3085 // constraint so that we don't exclude all formulae.
3086 if (!AnySatisfiedReqRegs) {
3087 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3093 /// Solve - Choose one formula from each use. Return the results in the given
3094 /// Solution vector.
3095 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3096 SmallVector<const Formula *, 8> Workspace;
3098 SolutionCost.Loose();
3100 SmallPtrSet<const SCEV *, 16> CurRegs;
3101 DenseSet<const SCEV *> VisitedRegs;
3102 Workspace.reserve(Uses.size());
3104 // SolveRecurse does all the work.
3105 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3106 CurRegs, VisitedRegs);
3108 // Ok, we've now made all our decisions.
3109 DEBUG(dbgs() << "\n"
3110 "The chosen solution requires "; SolutionCost.print(dbgs());
3112 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3114 Uses[i].print(dbgs());
3117 Solution[i]->print(dbgs());
3121 assert(Solution.size() == Uses.size() && "Malformed solution!");
3124 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3125 /// the dominator tree far as we can go while still being dominated by the
3126 /// input positions. This helps canonicalize the insert position, which
3127 /// encourages sharing.
3128 BasicBlock::iterator
3129 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3130 const SmallVectorImpl<Instruction *> &Inputs)
3133 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3134 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3137 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3138 if (!Rung) return IP;
3139 Rung = Rung->getIDom();
3140 if (!Rung) return IP;
3141 IDom = Rung->getBlock();
3143 // Don't climb into a loop though.
3144 const Loop *IDomLoop = LI.getLoopFor(IDom);
3145 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3146 if (IDomDepth <= IPLoopDepth &&
3147 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3151 bool AllDominate = true;
3152 Instruction *BetterPos = 0;
3153 Instruction *Tentative = IDom->getTerminator();
3154 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3155 E = Inputs.end(); I != E; ++I) {
3156 Instruction *Inst = *I;
3157 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3158 AllDominate = false;
3161 // Attempt to find an insert position in the middle of the block,
3162 // instead of at the end, so that it can be used for other expansions.
3163 if (IDom == Inst->getParent() &&
3164 (!BetterPos || DT.dominates(BetterPos, Inst)))
3165 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3178 /// AdjustInsertPositionForExpand - Determine an input position which will be
3179 /// dominated by the operands and which will dominate the result.
3180 BasicBlock::iterator
3181 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3183 const LSRUse &LU) const {
3184 // Collect some instructions which must be dominated by the
3185 // expanding replacement. These must be dominated by any operands that
3186 // will be required in the expansion.
3187 SmallVector<Instruction *, 4> Inputs;
3188 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3189 Inputs.push_back(I);
3190 if (LU.Kind == LSRUse::ICmpZero)
3191 if (Instruction *I =
3192 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3193 Inputs.push_back(I);
3194 if (LF.PostIncLoops.count(L)) {
3195 if (LF.isUseFullyOutsideLoop(L))
3196 Inputs.push_back(L->getLoopLatch()->getTerminator());
3198 Inputs.push_back(IVIncInsertPos);
3200 // The expansion must also be dominated by the increment positions of any
3201 // loops it for which it is using post-inc mode.
3202 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3203 E = LF.PostIncLoops.end(); I != E; ++I) {
3204 const Loop *PIL = *I;
3205 if (PIL == L) continue;
3207 // Be dominated by the loop exit.
3208 SmallVector<BasicBlock *, 4> ExitingBlocks;
3209 PIL->getExitingBlocks(ExitingBlocks);
3210 if (!ExitingBlocks.empty()) {
3211 BasicBlock *BB = ExitingBlocks[0];
3212 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3213 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3214 Inputs.push_back(BB->getTerminator());
3218 // Then, climb up the immediate dominator tree as far as we can go while
3219 // still being dominated by the input positions.
3220 IP = HoistInsertPosition(IP, Inputs);
3222 // Don't insert instructions before PHI nodes.
3223 while (isa<PHINode>(IP)) ++IP;
3225 // Ignore debug intrinsics.
3226 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3231 /// Expand - Emit instructions for the leading candidate expression for this
3232 /// LSRUse (this is called "expanding").
3233 Value *LSRInstance::Expand(const LSRFixup &LF,
3235 BasicBlock::iterator IP,
3236 SCEVExpander &Rewriter,
3237 SmallVectorImpl<WeakVH> &DeadInsts) const {
3238 const LSRUse &LU = Uses[LF.LUIdx];
3240 // Determine an input position which will be dominated by the operands and
3241 // which will dominate the result.
3242 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3244 // Inform the Rewriter if we have a post-increment use, so that it can
3245 // perform an advantageous expansion.
3246 Rewriter.setPostInc(LF.PostIncLoops);
3248 // This is the type that the user actually needs.
3249 const Type *OpTy = LF.OperandValToReplace->getType();
3250 // This will be the type that we'll initially expand to.
3251 const Type *Ty = F.getType();
3253 // No type known; just expand directly to the ultimate type.
3255 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3256 // Expand directly to the ultimate type if it's the right size.
3258 // This is the type to do integer arithmetic in.
3259 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3261 // Build up a list of operands to add together to form the full base.
3262 SmallVector<const SCEV *, 8> Ops;
3264 // Expand the BaseRegs portion.
3265 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3266 E = F.BaseRegs.end(); I != E; ++I) {
3267 const SCEV *Reg = *I;
3268 assert(!Reg->isZero() && "Zero allocated in a base register!");
3270 // If we're expanding for a post-inc user, make the post-inc adjustment.
3271 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3272 Reg = TransformForPostIncUse(Denormalize, Reg,
3273 LF.UserInst, LF.OperandValToReplace,
3276 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3279 // Flush the operand list to suppress SCEVExpander hoisting.
3281 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3283 Ops.push_back(SE.getUnknown(FullV));
3286 // Expand the ScaledReg portion.
3287 Value *ICmpScaledV = 0;
3288 if (F.AM.Scale != 0) {
3289 const SCEV *ScaledS = F.ScaledReg;
3291 // If we're expanding for a post-inc user, make the post-inc adjustment.
3292 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3293 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3294 LF.UserInst, LF.OperandValToReplace,
3297 if (LU.Kind == LSRUse::ICmpZero) {
3298 // An interesting way of "folding" with an icmp is to use a negated
3299 // scale, which we'll implement by inserting it into the other operand
3301 assert(F.AM.Scale == -1 &&
3302 "The only scale supported by ICmpZero uses is -1!");
3303 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3305 // Otherwise just expand the scaled register and an explicit scale,
3306 // which is expected to be matched as part of the address.
3307 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3308 ScaledS = SE.getMulExpr(ScaledS,
3309 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3310 Ops.push_back(ScaledS);
3312 // Flush the operand list to suppress SCEVExpander hoisting.
3313 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3315 Ops.push_back(SE.getUnknown(FullV));
3319 // Expand the GV portion.
3321 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3323 // Flush the operand list to suppress SCEVExpander hoisting.
3324 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3326 Ops.push_back(SE.getUnknown(FullV));
3329 // Expand the immediate portion.
3330 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3332 if (LU.Kind == LSRUse::ICmpZero) {
3333 // The other interesting way of "folding" with an ICmpZero is to use a
3334 // negated immediate.
3336 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3338 Ops.push_back(SE.getUnknown(ICmpScaledV));
3339 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3342 // Just add the immediate values. These again are expected to be matched
3343 // as part of the address.
3344 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3348 // Emit instructions summing all the operands.
3349 const SCEV *FullS = Ops.empty() ?
3350 SE.getConstant(IntTy, 0) :
3352 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3354 // We're done expanding now, so reset the rewriter.
3355 Rewriter.clearPostInc();
3357 // An ICmpZero Formula represents an ICmp which we're handling as a
3358 // comparison against zero. Now that we've expanded an expression for that
3359 // form, update the ICmp's other operand.
3360 if (LU.Kind == LSRUse::ICmpZero) {
3361 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3362 DeadInsts.push_back(CI->getOperand(1));
3363 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3364 "a scale at the same time!");
3365 if (F.AM.Scale == -1) {
3366 if (ICmpScaledV->getType() != OpTy) {
3368 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3370 ICmpScaledV, OpTy, "tmp", CI);
3373 CI->setOperand(1, ICmpScaledV);
3375 assert(F.AM.Scale == 0 &&
3376 "ICmp does not support folding a global value and "
3377 "a scale at the same time!");
3378 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3380 if (C->getType() != OpTy)
3381 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3385 CI->setOperand(1, C);
3392 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3393 /// of their operands effectively happens in their predecessor blocks, so the
3394 /// expression may need to be expanded in multiple places.
3395 void LSRInstance::RewriteForPHI(PHINode *PN,
3398 SCEVExpander &Rewriter,
3399 SmallVectorImpl<WeakVH> &DeadInsts,
3401 DenseMap<BasicBlock *, Value *> Inserted;
3402 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3403 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3404 BasicBlock *BB = PN->getIncomingBlock(i);
3406 // If this is a critical edge, split the edge so that we do not insert
3407 // the code on all predecessor/successor paths. We do this unless this
3408 // is the canonical backedge for this loop, which complicates post-inc
3410 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3411 !isa<IndirectBrInst>(BB->getTerminator()) &&
3412 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3413 // Split the critical edge.
3414 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3416 // If PN is outside of the loop and BB is in the loop, we want to
3417 // move the block to be immediately before the PHI block, not
3418 // immediately after BB.
3419 if (L->contains(BB) && !L->contains(PN))
3420 NewBB->moveBefore(PN->getParent());
3422 // Splitting the edge can reduce the number of PHI entries we have.
3423 e = PN->getNumIncomingValues();
3425 i = PN->getBasicBlockIndex(BB);
3428 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3429 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3431 PN->setIncomingValue(i, Pair.first->second);
3433 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3435 // If this is reuse-by-noop-cast, insert the noop cast.
3436 const Type *OpTy = LF.OperandValToReplace->getType();
3437 if (FullV->getType() != OpTy)
3439 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3441 FullV, LF.OperandValToReplace->getType(),
3442 "tmp", BB->getTerminator());
3444 PN->setIncomingValue(i, FullV);
3445 Pair.first->second = FullV;
3450 /// Rewrite - Emit instructions for the leading candidate expression for this
3451 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3452 /// the newly expanded value.
3453 void LSRInstance::Rewrite(const LSRFixup &LF,
3455 SCEVExpander &Rewriter,
3456 SmallVectorImpl<WeakVH> &DeadInsts,
3458 // First, find an insertion point that dominates UserInst. For PHI nodes,
3459 // find the nearest block which dominates all the relevant uses.
3460 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3461 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3463 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3465 // If this is reuse-by-noop-cast, insert the noop cast.
3466 const Type *OpTy = LF.OperandValToReplace->getType();
3467 if (FullV->getType() != OpTy) {
3469 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3470 FullV, OpTy, "tmp", LF.UserInst);
3474 // Update the user. ICmpZero is handled specially here (for now) because
3475 // Expand may have updated one of the operands of the icmp already, and
3476 // its new value may happen to be equal to LF.OperandValToReplace, in
3477 // which case doing replaceUsesOfWith leads to replacing both operands
3478 // with the same value. TODO: Reorganize this.
3479 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3480 LF.UserInst->setOperand(0, FullV);
3482 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3485 DeadInsts.push_back(LF.OperandValToReplace);
3488 /// ImplementSolution - Rewrite all the fixup locations with new values,
3489 /// following the chosen solution.
3491 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3493 // Keep track of instructions we may have made dead, so that
3494 // we can remove them after we are done working.
3495 SmallVector<WeakVH, 16> DeadInsts;
3497 SCEVExpander Rewriter(SE);
3498 Rewriter.disableCanonicalMode();
3499 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3501 // Expand the new value definitions and update the users.
3502 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3503 E = Fixups.end(); I != E; ++I) {
3504 const LSRFixup &Fixup = *I;
3506 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3511 // Clean up after ourselves. This must be done before deleting any
3515 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3518 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3519 : IU(P->getAnalysis<IVUsers>()),
3520 SE(P->getAnalysis<ScalarEvolution>()),
3521 DT(P->getAnalysis<DominatorTree>()),
3522 LI(P->getAnalysis<LoopInfo>()),
3523 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3525 // If LoopSimplify form is not available, stay out of trouble.
3526 if (!L->isLoopSimplifyForm()) return;
3528 // If there's no interesting work to be done, bail early.
3529 if (IU.empty()) return;
3531 DEBUG(dbgs() << "\nLSR on loop ";
3532 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3535 // First, perform some low-level loop optimizations.
3537 OptimizeLoopTermCond();
3539 // Start collecting data and preparing for the solver.
3540 CollectInterestingTypesAndFactors();
3541 CollectFixupsAndInitialFormulae();
3542 CollectLoopInvariantFixupsAndFormulae();
3544 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3545 print_uses(dbgs()));
3547 // Now use the reuse data to generate a bunch of interesting ways
3548 // to formulate the values needed for the uses.
3549 GenerateAllReuseFormulae();
3551 DEBUG(dbgs() << "\n"
3552 "After generating reuse formulae:\n";
3553 print_uses(dbgs()));
3555 FilterOutUndesirableDedicatedRegisters();
3556 NarrowSearchSpaceUsingHeuristics();
3558 SmallVector<const Formula *, 8> Solution;
3561 // Release memory that is no longer needed.
3567 // Formulae should be legal.
3568 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3569 E = Uses.end(); I != E; ++I) {
3570 const LSRUse &LU = *I;
3571 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3572 JE = LU.Formulae.end(); J != JE; ++J)
3573 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3574 LU.Kind, LU.AccessTy, TLI) &&
3575 "Illegal formula generated!");
3579 // Now that we've decided what we want, make it so.
3580 ImplementSolution(Solution, P);
3583 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3584 if (Factors.empty() && Types.empty()) return;
3586 OS << "LSR has identified the following interesting factors and types: ";
3589 for (SmallSetVector<int64_t, 8>::const_iterator
3590 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3591 if (!First) OS << ", ";
3596 for (SmallSetVector<const Type *, 4>::const_iterator
3597 I = Types.begin(), E = Types.end(); I != E; ++I) {
3598 if (!First) OS << ", ";
3600 OS << '(' << **I << ')';
3605 void LSRInstance::print_fixups(raw_ostream &OS) const {
3606 OS << "LSR is examining the following fixup sites:\n";
3607 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3608 E = Fixups.end(); I != E; ++I) {
3615 void LSRInstance::print_uses(raw_ostream &OS) const {
3616 OS << "LSR is examining the following uses:\n";
3617 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3618 E = Uses.end(); I != E; ++I) {
3619 const LSRUse &LU = *I;
3623 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3624 JE = LU.Formulae.end(); J != JE; ++J) {
3632 void LSRInstance::print(raw_ostream &OS) const {
3633 print_factors_and_types(OS);
3638 void LSRInstance::dump() const {
3639 print(errs()); errs() << '\n';
3644 class LoopStrengthReduce : public LoopPass {
3645 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3646 /// transformation profitability.
3647 const TargetLowering *const TLI;
3650 static char ID; // Pass ID, replacement for typeid
3651 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3654 bool runOnLoop(Loop *L, LPPassManager &LPM);
3655 void getAnalysisUsage(AnalysisUsage &AU) const;
3660 char LoopStrengthReduce::ID = 0;
3661 static RegisterPass<LoopStrengthReduce>
3662 X("loop-reduce", "Loop Strength Reduction");
3664 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3665 return new LoopStrengthReduce(TLI);
3668 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3669 : LoopPass(&ID), TLI(tli) {}
3671 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3672 // We split critical edges, so we change the CFG. However, we do update
3673 // many analyses if they are around.
3674 AU.addPreservedID(LoopSimplifyID);
3675 AU.addPreserved("domfrontier");
3677 AU.addRequired<LoopInfo>();
3678 AU.addPreserved<LoopInfo>();
3679 AU.addRequiredID(LoopSimplifyID);
3680 AU.addRequired<DominatorTree>();
3681 AU.addPreserved<DominatorTree>();
3682 AU.addRequired<ScalarEvolution>();
3683 AU.addPreserved<ScalarEvolution>();
3684 AU.addRequired<IVUsers>();
3685 AU.addPreserved<IVUsers>();
3688 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3689 bool Changed = false;
3691 // Run the main LSR transformation.
3692 Changed |= LSRInstance(TLI, L, this).getChanged();
3694 // At this point, it is worth checking to see if any recurrence PHIs are also
3695 // dead, so that we can remove them as well.
3696 Changed |= DeleteDeadPHIs(L->getHeader());