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 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1211 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1212 struct UseMapDenseMapInfo {
1213 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1214 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1217 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1218 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1222 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1223 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1224 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1228 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1229 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1234 /// FormulaSorter - This class implements an ordering for formulae which sorts
1235 /// the by their standalone cost.
1236 class FormulaSorter {
1237 /// These two sets are kept empty, so that we compute standalone costs.
1238 DenseSet<const SCEV *> VisitedRegs;
1239 SmallPtrSet<const SCEV *, 16> Regs;
1242 ScalarEvolution &SE;
1246 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1247 : L(l), LU(&lu), SE(se), DT(dt) {}
1249 bool operator()(const Formula &A, const Formula &B) {
1251 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1254 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1256 return CostA < CostB;
1260 /// LSRInstance - This class holds state for the main loop strength reduction
1264 ScalarEvolution &SE;
1267 const TargetLowering *const TLI;
1271 /// IVIncInsertPos - This is the insert position that the current loop's
1272 /// induction variable increment should be placed. In simple loops, this is
1273 /// the latch block's terminator. But in more complicated cases, this is a
1274 /// position which will dominate all the in-loop post-increment users.
1275 Instruction *IVIncInsertPos;
1277 /// Factors - Interesting factors between use strides.
1278 SmallSetVector<int64_t, 8> Factors;
1280 /// Types - Interesting use types, to facilitate truncation reuse.
1281 SmallSetVector<const Type *, 4> Types;
1283 /// Fixups - The list of operands which are to be replaced.
1284 SmallVector<LSRFixup, 16> Fixups;
1286 /// Uses - The list of interesting uses.
1287 SmallVector<LSRUse, 16> Uses;
1289 /// RegUses - Track which uses use which register candidates.
1290 RegUseTracker RegUses;
1292 void OptimizeShadowIV();
1293 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1294 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1295 void OptimizeLoopTermCond();
1297 void CollectInterestingTypesAndFactors();
1298 void CollectFixupsAndInitialFormulae();
1300 LSRFixup &getNewFixup() {
1301 Fixups.push_back(LSRFixup());
1302 return Fixups.back();
1305 // Support for sharing of LSRUses between LSRFixups.
1306 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1308 UseMapDenseMapInfo> UseMapTy;
1311 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1312 LSRUse::KindType Kind, const Type *AccessTy);
1314 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1315 LSRUse::KindType Kind,
1316 const Type *AccessTy);
1318 void DeleteUse(LSRUse &LU);
1320 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1323 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1324 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1325 void CountRegisters(const Formula &F, size_t LUIdx);
1326 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1328 void CollectLoopInvariantFixupsAndFormulae();
1330 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1331 unsigned Depth = 0);
1332 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1333 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1334 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1335 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1336 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1337 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1338 void GenerateCrossUseConstantOffsets();
1339 void GenerateAllReuseFormulae();
1341 void FilterOutUndesirableDedicatedRegisters();
1343 size_t EstimateSearchSpaceComplexity() const;
1344 void NarrowSearchSpaceUsingHeuristics();
1346 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1348 SmallVectorImpl<const Formula *> &Workspace,
1349 const Cost &CurCost,
1350 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1351 DenseSet<const SCEV *> &VisitedRegs) const;
1352 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1354 BasicBlock::iterator
1355 HoistInsertPosition(BasicBlock::iterator IP,
1356 const SmallVectorImpl<Instruction *> &Inputs) const;
1357 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1359 const LSRUse &LU) const;
1361 Value *Expand(const LSRFixup &LF,
1363 BasicBlock::iterator IP,
1364 SCEVExpander &Rewriter,
1365 SmallVectorImpl<WeakVH> &DeadInsts) const;
1366 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1368 SCEVExpander &Rewriter,
1369 SmallVectorImpl<WeakVH> &DeadInsts,
1371 void Rewrite(const LSRFixup &LF,
1373 SCEVExpander &Rewriter,
1374 SmallVectorImpl<WeakVH> &DeadInsts,
1376 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1379 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1381 bool getChanged() const { return Changed; }
1383 void print_factors_and_types(raw_ostream &OS) const;
1384 void print_fixups(raw_ostream &OS) const;
1385 void print_uses(raw_ostream &OS) const;
1386 void print(raw_ostream &OS) const;
1392 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1393 /// inside the loop then try to eliminate the cast operation.
1394 void LSRInstance::OptimizeShadowIV() {
1395 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1396 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1399 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1400 UI != E; /* empty */) {
1401 IVUsers::const_iterator CandidateUI = UI;
1403 Instruction *ShadowUse = CandidateUI->getUser();
1404 const Type *DestTy = NULL;
1406 /* If shadow use is a int->float cast then insert a second IV
1407 to eliminate this cast.
1409 for (unsigned i = 0; i < n; ++i)
1415 for (unsigned i = 0; i < n; ++i, ++d)
1418 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1419 DestTy = UCast->getDestTy();
1420 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1421 DestTy = SCast->getDestTy();
1422 if (!DestTy) continue;
1425 // If target does not support DestTy natively then do not apply
1426 // this transformation.
1427 EVT DVT = TLI->getValueType(DestTy);
1428 if (!TLI->isTypeLegal(DVT)) continue;
1431 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1433 if (PH->getNumIncomingValues() != 2) continue;
1435 const Type *SrcTy = PH->getType();
1436 int Mantissa = DestTy->getFPMantissaWidth();
1437 if (Mantissa == -1) continue;
1438 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1441 unsigned Entry, Latch;
1442 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1450 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1451 if (!Init) continue;
1452 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1454 BinaryOperator *Incr =
1455 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1456 if (!Incr) continue;
1457 if (Incr->getOpcode() != Instruction::Add
1458 && Incr->getOpcode() != Instruction::Sub)
1461 /* Initialize new IV, double d = 0.0 in above example. */
1462 ConstantInt *C = NULL;
1463 if (Incr->getOperand(0) == PH)
1464 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1465 else if (Incr->getOperand(1) == PH)
1466 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1472 // Ignore negative constants, as the code below doesn't handle them
1473 // correctly. TODO: Remove this restriction.
1474 if (!C->getValue().isStrictlyPositive()) continue;
1476 /* Add new PHINode. */
1477 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1479 /* create new increment. '++d' in above example. */
1480 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1481 BinaryOperator *NewIncr =
1482 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1483 Instruction::FAdd : Instruction::FSub,
1484 NewPH, CFP, "IV.S.next.", Incr);
1486 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1487 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1489 /* Remove cast operation */
1490 ShadowUse->replaceAllUsesWith(NewPH);
1491 ShadowUse->eraseFromParent();
1497 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1498 /// set the IV user and stride information and return true, otherwise return
1500 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1501 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1502 if (UI->getUser() == Cond) {
1503 // NOTE: we could handle setcc instructions with multiple uses here, but
1504 // InstCombine does it as well for simple uses, it's not clear that it
1505 // occurs enough in real life to handle.
1512 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1513 /// a max computation.
1515 /// This is a narrow solution to a specific, but acute, problem. For loops
1521 /// } while (++i < n);
1523 /// the trip count isn't just 'n', because 'n' might not be positive. And
1524 /// unfortunately this can come up even for loops where the user didn't use
1525 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1526 /// will commonly be lowered like this:
1532 /// } while (++i < n);
1535 /// and then it's possible for subsequent optimization to obscure the if
1536 /// test in such a way that indvars can't find it.
1538 /// When indvars can't find the if test in loops like this, it creates a
1539 /// max expression, which allows it to give the loop a canonical
1540 /// induction variable:
1543 /// max = n < 1 ? 1 : n;
1546 /// } while (++i != max);
1548 /// Canonical induction variables are necessary because the loop passes
1549 /// are designed around them. The most obvious example of this is the
1550 /// LoopInfo analysis, which doesn't remember trip count values. It
1551 /// expects to be able to rediscover the trip count each time it is
1552 /// needed, and it does this using a simple analysis that only succeeds if
1553 /// the loop has a canonical induction variable.
1555 /// However, when it comes time to generate code, the maximum operation
1556 /// can be quite costly, especially if it's inside of an outer loop.
1558 /// This function solves this problem by detecting this type of loop and
1559 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1560 /// the instructions for the maximum computation.
1562 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1563 // Check that the loop matches the pattern we're looking for.
1564 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1565 Cond->getPredicate() != CmpInst::ICMP_NE)
1568 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1569 if (!Sel || !Sel->hasOneUse()) return Cond;
1571 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1572 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1574 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1576 // Add one to the backedge-taken count to get the trip count.
1577 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1578 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1580 // Check for a max calculation that matches the pattern. There's no check
1581 // for ICMP_ULE here because the comparison would be with zero, which
1582 // isn't interesting.
1583 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1584 const SCEVNAryExpr *Max = 0;
1585 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1586 Pred = ICmpInst::ICMP_SLE;
1588 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1589 Pred = ICmpInst::ICMP_SLT;
1591 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1592 Pred = ICmpInst::ICMP_ULT;
1599 // To handle a max with more than two operands, this optimization would
1600 // require additional checking and setup.
1601 if (Max->getNumOperands() != 2)
1604 const SCEV *MaxLHS = Max->getOperand(0);
1605 const SCEV *MaxRHS = Max->getOperand(1);
1607 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1608 // for a comparison with 1. For <= and >=, a comparison with zero.
1610 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1613 // Check the relevant induction variable for conformance to
1615 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1616 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1617 if (!AR || !AR->isAffine() ||
1618 AR->getStart() != One ||
1619 AR->getStepRecurrence(SE) != One)
1622 assert(AR->getLoop() == L &&
1623 "Loop condition operand is an addrec in a different loop!");
1625 // Check the right operand of the select, and remember it, as it will
1626 // be used in the new comparison instruction.
1628 if (ICmpInst::isTrueWhenEqual(Pred)) {
1629 // Look for n+1, and grab n.
1630 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1631 if (isa<ConstantInt>(BO->getOperand(1)) &&
1632 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1633 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1634 NewRHS = BO->getOperand(0);
1635 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1636 if (isa<ConstantInt>(BO->getOperand(1)) &&
1637 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1638 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1639 NewRHS = BO->getOperand(0);
1642 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1643 NewRHS = Sel->getOperand(1);
1644 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1645 NewRHS = Sel->getOperand(2);
1646 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1647 NewRHS = SU->getValue();
1649 // Max doesn't match expected pattern.
1652 // Determine the new comparison opcode. It may be signed or unsigned,
1653 // and the original comparison may be either equality or inequality.
1654 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1655 Pred = CmpInst::getInversePredicate(Pred);
1657 // Ok, everything looks ok to change the condition into an SLT or SGE and
1658 // delete the max calculation.
1660 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1662 // Delete the max calculation instructions.
1663 Cond->replaceAllUsesWith(NewCond);
1664 CondUse->setUser(NewCond);
1665 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1666 Cond->eraseFromParent();
1667 Sel->eraseFromParent();
1668 if (Cmp->use_empty())
1669 Cmp->eraseFromParent();
1673 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1674 /// postinc iv when possible.
1676 LSRInstance::OptimizeLoopTermCond() {
1677 SmallPtrSet<Instruction *, 4> PostIncs;
1679 BasicBlock *LatchBlock = L->getLoopLatch();
1680 SmallVector<BasicBlock*, 8> ExitingBlocks;
1681 L->getExitingBlocks(ExitingBlocks);
1683 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1684 BasicBlock *ExitingBlock = ExitingBlocks[i];
1686 // Get the terminating condition for the loop if possible. If we
1687 // can, we want to change it to use a post-incremented version of its
1688 // induction variable, to allow coalescing the live ranges for the IV into
1689 // one register value.
1691 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1694 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1695 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1698 // Search IVUsesByStride to find Cond's IVUse if there is one.
1699 IVStrideUse *CondUse = 0;
1700 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1701 if (!FindIVUserForCond(Cond, CondUse))
1704 // If the trip count is computed in terms of a max (due to ScalarEvolution
1705 // being unable to find a sufficient guard, for example), change the loop
1706 // comparison to use SLT or ULT instead of NE.
1707 // One consequence of doing this now is that it disrupts the count-down
1708 // optimization. That's not always a bad thing though, because in such
1709 // cases it may still be worthwhile to avoid a max.
1710 Cond = OptimizeMax(Cond, CondUse);
1712 // If this exiting block dominates the latch block, it may also use
1713 // the post-inc value if it won't be shared with other uses.
1714 // Check for dominance.
1715 if (!DT.dominates(ExitingBlock, LatchBlock))
1718 // Conservatively avoid trying to use the post-inc value in non-latch
1719 // exits if there may be pre-inc users in intervening blocks.
1720 if (LatchBlock != ExitingBlock)
1721 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1722 // Test if the use is reachable from the exiting block. This dominator
1723 // query is a conservative approximation of reachability.
1724 if (&*UI != CondUse &&
1725 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1726 // Conservatively assume there may be reuse if the quotient of their
1727 // strides could be a legal scale.
1728 const SCEV *A = IU.getStride(*CondUse, L);
1729 const SCEV *B = IU.getStride(*UI, L);
1730 if (!A || !B) continue;
1731 if (SE.getTypeSizeInBits(A->getType()) !=
1732 SE.getTypeSizeInBits(B->getType())) {
1733 if (SE.getTypeSizeInBits(A->getType()) >
1734 SE.getTypeSizeInBits(B->getType()))
1735 B = SE.getSignExtendExpr(B, A->getType());
1737 A = SE.getSignExtendExpr(A, B->getType());
1739 if (const SCEVConstant *D =
1740 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1741 const ConstantInt *C = D->getValue();
1742 // Stride of one or negative one can have reuse with non-addresses.
1743 if (C->isOne() || C->isAllOnesValue())
1744 goto decline_post_inc;
1745 // Avoid weird situations.
1746 if (C->getValue().getMinSignedBits() >= 64 ||
1747 C->getValue().isMinSignedValue())
1748 goto decline_post_inc;
1749 // Without TLI, assume that any stride might be valid, and so any
1750 // use might be shared.
1752 goto decline_post_inc;
1753 // Check for possible scaled-address reuse.
1754 const Type *AccessTy = getAccessType(UI->getUser());
1755 TargetLowering::AddrMode AM;
1756 AM.Scale = C->getSExtValue();
1757 if (TLI->isLegalAddressingMode(AM, AccessTy))
1758 goto decline_post_inc;
1759 AM.Scale = -AM.Scale;
1760 if (TLI->isLegalAddressingMode(AM, AccessTy))
1761 goto decline_post_inc;
1765 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1768 // It's possible for the setcc instruction to be anywhere in the loop, and
1769 // possible for it to have multiple users. If it is not immediately before
1770 // the exiting block branch, move it.
1771 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1772 if (Cond->hasOneUse()) {
1773 Cond->moveBefore(TermBr);
1775 // Clone the terminating condition and insert into the loopend.
1776 ICmpInst *OldCond = Cond;
1777 Cond = cast<ICmpInst>(Cond->clone());
1778 Cond->setName(L->getHeader()->getName() + ".termcond");
1779 ExitingBlock->getInstList().insert(TermBr, Cond);
1781 // Clone the IVUse, as the old use still exists!
1782 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1783 TermBr->replaceUsesOfWith(OldCond, Cond);
1787 // If we get to here, we know that we can transform the setcc instruction to
1788 // use the post-incremented version of the IV, allowing us to coalesce the
1789 // live ranges for the IV correctly.
1790 CondUse->transformToPostInc(L);
1793 PostIncs.insert(Cond);
1797 // Determine an insertion point for the loop induction variable increment. It
1798 // must dominate all the post-inc comparisons we just set up, and it must
1799 // dominate the loop latch edge.
1800 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1801 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1802 E = PostIncs.end(); I != E; ++I) {
1804 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1806 if (BB == (*I)->getParent())
1807 IVIncInsertPos = *I;
1808 else if (BB != IVIncInsertPos->getParent())
1809 IVIncInsertPos = BB->getTerminator();
1813 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1814 /// at the given offset and other details. If so, update the use and
1817 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1818 LSRUse::KindType Kind, const Type *AccessTy) {
1819 int64_t NewMinOffset = LU.MinOffset;
1820 int64_t NewMaxOffset = LU.MaxOffset;
1821 const Type *NewAccessTy = AccessTy;
1823 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1824 // something conservative, however this can pessimize in the case that one of
1825 // the uses will have all its uses outside the loop, for example.
1826 if (LU.Kind != Kind)
1828 // Conservatively assume HasBaseReg is true for now.
1829 if (NewOffset < LU.MinOffset) {
1830 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1831 Kind, AccessTy, TLI))
1833 NewMinOffset = NewOffset;
1834 } else if (NewOffset > LU.MaxOffset) {
1835 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1836 Kind, AccessTy, TLI))
1838 NewMaxOffset = NewOffset;
1840 // Check for a mismatched access type, and fall back conservatively as needed.
1841 // TODO: Be less conservative when the type is similar and can use the same
1842 // addressing modes.
1843 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1844 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1847 LU.MinOffset = NewMinOffset;
1848 LU.MaxOffset = NewMaxOffset;
1849 LU.AccessTy = NewAccessTy;
1850 if (NewOffset != LU.Offsets.back())
1851 LU.Offsets.push_back(NewOffset);
1855 /// getUse - Return an LSRUse index and an offset value for a fixup which
1856 /// needs the given expression, with the given kind and optional access type.
1857 /// Either reuse an existing use or create a new one, as needed.
1858 std::pair<size_t, int64_t>
1859 LSRInstance::getUse(const SCEV *&Expr,
1860 LSRUse::KindType Kind, const Type *AccessTy) {
1861 const SCEV *Copy = Expr;
1862 int64_t Offset = ExtractImmediate(Expr, SE);
1864 // Basic uses can't accept any offset, for example.
1865 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1870 std::pair<UseMapTy::iterator, bool> P =
1871 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1873 // A use already existed with this base.
1874 size_t LUIdx = P.first->second;
1875 LSRUse &LU = Uses[LUIdx];
1876 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1878 return std::make_pair(LUIdx, Offset);
1881 // Create a new use.
1882 size_t LUIdx = Uses.size();
1883 P.first->second = LUIdx;
1884 Uses.push_back(LSRUse(Kind, AccessTy));
1885 LSRUse &LU = Uses[LUIdx];
1887 // We don't need to track redundant offsets, but we don't need to go out
1888 // of our way here to avoid them.
1889 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1890 LU.Offsets.push_back(Offset);
1892 LU.MinOffset = Offset;
1893 LU.MaxOffset = Offset;
1894 return std::make_pair(LUIdx, Offset);
1897 /// DeleteUse - Delete the given use from the Uses list.
1898 void LSRInstance::DeleteUse(LSRUse &LU) {
1899 if (&LU != &Uses.back())
1900 std::swap(LU, Uses.back());
1904 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1905 /// a formula that has the same registers as the given formula.
1907 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1908 const LSRUse &OrigLU) {
1909 // Search all uses for the formula. This could be more clever. Ignore
1910 // ICmpZero uses because they may contain formulae generated by
1911 // GenerateICmpZeroScales, in which case adding fixup offsets may
1913 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1914 LSRUse &LU = Uses[LUIdx];
1915 if (&LU != &OrigLU &&
1916 LU.Kind != LSRUse::ICmpZero &&
1917 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1918 LU.HasFormulaWithSameRegs(OrigF)) {
1919 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1920 E = LU.Formulae.end(); I != E; ++I) {
1921 const Formula &F = *I;
1922 if (F.BaseRegs == OrigF.BaseRegs &&
1923 F.ScaledReg == OrigF.ScaledReg &&
1924 F.AM.BaseGV == OrigF.AM.BaseGV &&
1925 F.AM.Scale == OrigF.AM.Scale &&
1927 if (F.AM.BaseOffs == 0)
1938 void LSRInstance::CollectInterestingTypesAndFactors() {
1939 SmallSetVector<const SCEV *, 4> Strides;
1941 // Collect interesting types and strides.
1942 SmallVector<const SCEV *, 4> Worklist;
1943 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1944 const SCEV *Expr = IU.getExpr(*UI);
1946 // Collect interesting types.
1947 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1949 // Add strides for mentioned loops.
1950 Worklist.push_back(Expr);
1952 const SCEV *S = Worklist.pop_back_val();
1953 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1954 Strides.insert(AR->getStepRecurrence(SE));
1955 Worklist.push_back(AR->getStart());
1956 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1957 Worklist.append(Add->op_begin(), Add->op_end());
1959 } while (!Worklist.empty());
1962 // Compute interesting factors from the set of interesting strides.
1963 for (SmallSetVector<const SCEV *, 4>::const_iterator
1964 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1965 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1966 next(I); NewStrideIter != E; ++NewStrideIter) {
1967 const SCEV *OldStride = *I;
1968 const SCEV *NewStride = *NewStrideIter;
1970 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1971 SE.getTypeSizeInBits(NewStride->getType())) {
1972 if (SE.getTypeSizeInBits(OldStride->getType()) >
1973 SE.getTypeSizeInBits(NewStride->getType()))
1974 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1976 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1978 if (const SCEVConstant *Factor =
1979 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1981 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1982 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1983 } else if (const SCEVConstant *Factor =
1984 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1987 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1988 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1992 // If all uses use the same type, don't bother looking for truncation-based
1994 if (Types.size() == 1)
1997 DEBUG(print_factors_and_types(dbgs()));
2000 void LSRInstance::CollectFixupsAndInitialFormulae() {
2001 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2003 LSRFixup &LF = getNewFixup();
2004 LF.UserInst = UI->getUser();
2005 LF.OperandValToReplace = UI->getOperandValToReplace();
2006 LF.PostIncLoops = UI->getPostIncLoops();
2008 LSRUse::KindType Kind = LSRUse::Basic;
2009 const Type *AccessTy = 0;
2010 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2011 Kind = LSRUse::Address;
2012 AccessTy = getAccessType(LF.UserInst);
2015 const SCEV *S = IU.getExpr(*UI);
2017 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2018 // (N - i == 0), and this allows (N - i) to be the expression that we work
2019 // with rather than just N or i, so we can consider the register
2020 // requirements for both N and i at the same time. Limiting this code to
2021 // equality icmps is not a problem because all interesting loops use
2022 // equality icmps, thanks to IndVarSimplify.
2023 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2024 if (CI->isEquality()) {
2025 // Swap the operands if needed to put the OperandValToReplace on the
2026 // left, for consistency.
2027 Value *NV = CI->getOperand(1);
2028 if (NV == LF.OperandValToReplace) {
2029 CI->setOperand(1, CI->getOperand(0));
2030 CI->setOperand(0, NV);
2031 NV = CI->getOperand(1);
2035 // x == y --> x - y == 0
2036 const SCEV *N = SE.getSCEV(NV);
2037 if (N->isLoopInvariant(L)) {
2038 Kind = LSRUse::ICmpZero;
2039 S = SE.getMinusSCEV(N, S);
2042 // -1 and the negations of all interesting strides (except the negation
2043 // of -1) are now also interesting.
2044 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2045 if (Factors[i] != -1)
2046 Factors.insert(-(uint64_t)Factors[i]);
2050 // Set up the initial formula for this use.
2051 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2053 LF.Offset = P.second;
2054 LSRUse &LU = Uses[LF.LUIdx];
2055 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2057 // If this is the first use of this LSRUse, give it a formula.
2058 if (LU.Formulae.empty()) {
2059 InsertInitialFormula(S, LU, LF.LUIdx);
2060 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2064 DEBUG(print_fixups(dbgs()));
2067 /// InsertInitialFormula - Insert a formula for the given expression into
2068 /// the given use, separating out loop-variant portions from loop-invariant
2069 /// and loop-computable portions.
2071 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2073 F.InitialMatch(S, L, SE, DT);
2074 bool Inserted = InsertFormula(LU, LUIdx, F);
2075 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2078 /// InsertSupplementalFormula - Insert a simple single-register formula for
2079 /// the given expression into the given use.
2081 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2082 LSRUse &LU, size_t LUIdx) {
2084 F.BaseRegs.push_back(S);
2085 F.AM.HasBaseReg = true;
2086 bool Inserted = InsertFormula(LU, LUIdx, F);
2087 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2090 /// CountRegisters - Note which registers are used by the given formula,
2091 /// updating RegUses.
2092 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2094 RegUses.CountRegister(F.ScaledReg, LUIdx);
2095 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2096 E = F.BaseRegs.end(); I != E; ++I)
2097 RegUses.CountRegister(*I, LUIdx);
2100 /// InsertFormula - If the given formula has not yet been inserted, add it to
2101 /// the list, and return true. Return false otherwise.
2102 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2103 if (!LU.InsertFormula(F))
2106 CountRegisters(F, LUIdx);
2110 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2111 /// loop-invariant values which we're tracking. These other uses will pin these
2112 /// values in registers, making them less profitable for elimination.
2113 /// TODO: This currently misses non-constant addrec step registers.
2114 /// TODO: Should this give more weight to users inside the loop?
2116 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2117 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2118 SmallPtrSet<const SCEV *, 8> Inserted;
2120 while (!Worklist.empty()) {
2121 const SCEV *S = Worklist.pop_back_val();
2123 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2124 Worklist.append(N->op_begin(), N->op_end());
2125 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2126 Worklist.push_back(C->getOperand());
2127 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2128 Worklist.push_back(D->getLHS());
2129 Worklist.push_back(D->getRHS());
2130 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2131 if (!Inserted.insert(U)) continue;
2132 const Value *V = U->getValue();
2133 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2134 // Look for instructions defined outside the loop.
2135 if (L->contains(Inst)) continue;
2136 } else if (isa<UndefValue>(V))
2137 // Undef doesn't have a live range, so it doesn't matter.
2139 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2141 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2142 // Ignore non-instructions.
2145 // Ignore instructions in other functions (as can happen with
2147 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2149 // Ignore instructions not dominated by the loop.
2150 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2151 UserInst->getParent() :
2152 cast<PHINode>(UserInst)->getIncomingBlock(
2153 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2154 if (!DT.dominates(L->getHeader(), UseBB))
2156 // Ignore uses which are part of other SCEV expressions, to avoid
2157 // analyzing them multiple times.
2158 if (SE.isSCEVable(UserInst->getType())) {
2159 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2160 // If the user is a no-op, look through to its uses.
2161 if (!isa<SCEVUnknown>(UserS))
2165 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2169 // Ignore icmp instructions which are already being analyzed.
2170 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2171 unsigned OtherIdx = !UI.getOperandNo();
2172 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2173 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2177 LSRFixup &LF = getNewFixup();
2178 LF.UserInst = const_cast<Instruction *>(UserInst);
2179 LF.OperandValToReplace = UI.getUse();
2180 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2182 LF.Offset = P.second;
2183 LSRUse &LU = Uses[LF.LUIdx];
2184 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2185 InsertSupplementalFormula(U, LU, LF.LUIdx);
2186 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2193 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2194 /// separate registers. If C is non-null, multiply each subexpression by C.
2195 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2196 SmallVectorImpl<const SCEV *> &Ops,
2197 ScalarEvolution &SE) {
2198 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2199 // Break out add operands.
2200 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2202 CollectSubexprs(*I, C, Ops, SE);
2204 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2205 // Split a non-zero base out of an addrec.
2206 if (!AR->getStart()->isZero()) {
2207 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2208 AR->getStepRecurrence(SE),
2209 AR->getLoop()), C, Ops, SE);
2210 CollectSubexprs(AR->getStart(), C, Ops, SE);
2213 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2214 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2215 if (Mul->getNumOperands() == 2)
2216 if (const SCEVConstant *Op0 =
2217 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2218 CollectSubexprs(Mul->getOperand(1),
2219 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2225 // Otherwise use the value itself.
2226 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2229 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2231 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2234 // Arbitrarily cap recursion to protect compile time.
2235 if (Depth >= 3) return;
2237 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2238 const SCEV *BaseReg = Base.BaseRegs[i];
2240 SmallVector<const SCEV *, 8> AddOps;
2241 CollectSubexprs(BaseReg, 0, AddOps, SE);
2242 if (AddOps.size() == 1) continue;
2244 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2245 JE = AddOps.end(); J != JE; ++J) {
2246 // Don't pull a constant into a register if the constant could be folded
2247 // into an immediate field.
2248 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2249 Base.getNumRegs() > 1,
2250 LU.Kind, LU.AccessTy, TLI, SE))
2253 // Collect all operands except *J.
2254 SmallVector<const SCEV *, 8> InnerAddOps
2255 ( ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2257 (next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2259 // Don't leave just a constant behind in a register if the constant could
2260 // be folded into an immediate field.
2261 if (InnerAddOps.size() == 1 &&
2262 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2263 Base.getNumRegs() > 1,
2264 LU.Kind, LU.AccessTy, TLI, SE))
2267 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2268 if (InnerSum->isZero())
2271 F.BaseRegs[i] = InnerSum;
2272 F.BaseRegs.push_back(*J);
2273 if (InsertFormula(LU, LUIdx, F))
2274 // If that formula hadn't been seen before, recurse to find more like
2276 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2281 /// GenerateCombinations - Generate a formula consisting of all of the
2282 /// loop-dominating registers added into a single register.
2283 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2285 // This method is only interesting on a plurality of registers.
2286 if (Base.BaseRegs.size() <= 1) return;
2290 SmallVector<const SCEV *, 4> Ops;
2291 for (SmallVectorImpl<const SCEV *>::const_iterator
2292 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2293 const SCEV *BaseReg = *I;
2294 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2295 !BaseReg->hasComputableLoopEvolution(L))
2296 Ops.push_back(BaseReg);
2298 F.BaseRegs.push_back(BaseReg);
2300 if (Ops.size() > 1) {
2301 const SCEV *Sum = SE.getAddExpr(Ops);
2302 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2303 // opportunity to fold something. For now, just ignore such cases
2304 // rather than proceed with zero in a register.
2305 if (!Sum->isZero()) {
2306 F.BaseRegs.push_back(Sum);
2307 (void)InsertFormula(LU, LUIdx, F);
2312 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2313 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2315 // We can't add a symbolic offset if the address already contains one.
2316 if (Base.AM.BaseGV) return;
2318 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2319 const SCEV *G = Base.BaseRegs[i];
2320 GlobalValue *GV = ExtractSymbol(G, SE);
2321 if (G->isZero() || !GV)
2325 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2326 LU.Kind, LU.AccessTy, TLI))
2329 (void)InsertFormula(LU, LUIdx, F);
2333 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2334 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2336 // TODO: For now, just add the min and max offset, because it usually isn't
2337 // worthwhile looking at everything inbetween.
2338 SmallVector<int64_t, 4> Worklist;
2339 Worklist.push_back(LU.MinOffset);
2340 if (LU.MaxOffset != LU.MinOffset)
2341 Worklist.push_back(LU.MaxOffset);
2343 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2344 const SCEV *G = Base.BaseRegs[i];
2346 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2347 E = Worklist.end(); I != E; ++I) {
2349 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2350 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2351 LU.Kind, LU.AccessTy, TLI)) {
2352 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2354 (void)InsertFormula(LU, LUIdx, F);
2358 int64_t Imm = ExtractImmediate(G, SE);
2359 if (G->isZero() || Imm == 0)
2362 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2363 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2364 LU.Kind, LU.AccessTy, TLI))
2367 (void)InsertFormula(LU, LUIdx, F);
2371 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2372 /// the comparison. For example, x == y -> x*c == y*c.
2373 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2375 if (LU.Kind != LSRUse::ICmpZero) return;
2377 // Determine the integer type for the base formula.
2378 const Type *IntTy = Base.getType();
2380 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2382 // Don't do this if there is more than one offset.
2383 if (LU.MinOffset != LU.MaxOffset) return;
2385 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2387 // Check each interesting stride.
2388 for (SmallSetVector<int64_t, 8>::const_iterator
2389 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2390 int64_t Factor = *I;
2393 // Check that the multiplication doesn't overflow.
2394 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2396 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2397 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2400 // Check that multiplying with the use offset doesn't overflow.
2401 int64_t Offset = LU.MinOffset;
2402 if (Offset == INT64_MIN && Factor == -1)
2404 Offset = (uint64_t)Offset * Factor;
2405 if (Offset / Factor != LU.MinOffset)
2408 // Check that this scale is legal.
2409 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2412 // Compensate for the use having MinOffset built into it.
2413 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2415 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2417 // Check that multiplying with each base register doesn't overflow.
2418 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2419 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2420 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2424 // Check that multiplying with the scaled register doesn't overflow.
2426 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2427 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2431 // If we make it here and it's legal, add it.
2432 (void)InsertFormula(LU, LUIdx, F);
2437 /// GenerateScales - Generate stride factor reuse formulae by making use of
2438 /// scaled-offset address modes, for example.
2439 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2440 // Determine the integer type for the base formula.
2441 const Type *IntTy = Base.getType();
2444 // If this Formula already has a scaled register, we can't add another one.
2445 if (Base.AM.Scale != 0) return;
2447 // Check each interesting stride.
2448 for (SmallSetVector<int64_t, 8>::const_iterator
2449 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2450 int64_t Factor = *I;
2452 Base.AM.Scale = Factor;
2453 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2454 // Check whether this scale is going to be legal.
2455 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2456 LU.Kind, LU.AccessTy, TLI)) {
2457 // As a special-case, handle special out-of-loop Basic users specially.
2458 // TODO: Reconsider this special case.
2459 if (LU.Kind == LSRUse::Basic &&
2460 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2461 LSRUse::Special, LU.AccessTy, TLI) &&
2462 LU.AllFixupsOutsideLoop)
2463 LU.Kind = LSRUse::Special;
2467 // For an ICmpZero, negating a solitary base register won't lead to
2469 if (LU.Kind == LSRUse::ICmpZero &&
2470 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2472 // For each addrec base reg, apply the scale, if possible.
2473 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2474 if (const SCEVAddRecExpr *AR =
2475 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2476 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2477 if (FactorS->isZero())
2479 // Divide out the factor, ignoring high bits, since we'll be
2480 // scaling the value back up in the end.
2481 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2482 // TODO: This could be optimized to avoid all the copying.
2484 F.ScaledReg = Quotient;
2485 F.DeleteBaseReg(F.BaseRegs[i]);
2486 (void)InsertFormula(LU, LUIdx, F);
2492 /// GenerateTruncates - Generate reuse formulae from different IV types.
2493 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2494 // This requires TargetLowering to tell us which truncates are free.
2497 // Don't bother truncating symbolic values.
2498 if (Base.AM.BaseGV) return;
2500 // Determine the integer type for the base formula.
2501 const Type *DstTy = Base.getType();
2503 DstTy = SE.getEffectiveSCEVType(DstTy);
2505 for (SmallSetVector<const Type *, 4>::const_iterator
2506 I = Types.begin(), E = Types.end(); I != E; ++I) {
2507 const Type *SrcTy = *I;
2508 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2511 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2512 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2513 JE = F.BaseRegs.end(); J != JE; ++J)
2514 *J = SE.getAnyExtendExpr(*J, SrcTy);
2516 // TODO: This assumes we've done basic processing on all uses and
2517 // have an idea what the register usage is.
2518 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2521 (void)InsertFormula(LU, LUIdx, F);
2528 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2529 /// defer modifications so that the search phase doesn't have to worry about
2530 /// the data structures moving underneath it.
2534 const SCEV *OrigReg;
2536 WorkItem(size_t LI, int64_t I, const SCEV *R)
2537 : LUIdx(LI), Imm(I), OrigReg(R) {}
2539 void print(raw_ostream &OS) const;
2545 void WorkItem::print(raw_ostream &OS) const {
2546 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2547 << " , add offset " << Imm;
2550 void WorkItem::dump() const {
2551 print(errs()); errs() << '\n';
2554 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2555 /// distance apart and try to form reuse opportunities between them.
2556 void LSRInstance::GenerateCrossUseConstantOffsets() {
2557 // Group the registers by their value without any added constant offset.
2558 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2559 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2561 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2562 SmallVector<const SCEV *, 8> Sequence;
2563 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2565 const SCEV *Reg = *I;
2566 int64_t Imm = ExtractImmediate(Reg, SE);
2567 std::pair<RegMapTy::iterator, bool> Pair =
2568 Map.insert(std::make_pair(Reg, ImmMapTy()));
2570 Sequence.push_back(Reg);
2571 Pair.first->second.insert(std::make_pair(Imm, *I));
2572 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2575 // Now examine each set of registers with the same base value. Build up
2576 // a list of work to do and do the work in a separate step so that we're
2577 // not adding formulae and register counts while we're searching.
2578 SmallVector<WorkItem, 32> WorkItems;
2579 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2580 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2581 E = Sequence.end(); I != E; ++I) {
2582 const SCEV *Reg = *I;
2583 const ImmMapTy &Imms = Map.find(Reg)->second;
2585 // It's not worthwhile looking for reuse if there's only one offset.
2586 if (Imms.size() == 1)
2589 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2590 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2592 dbgs() << ' ' << J->first;
2595 // Examine each offset.
2596 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2598 const SCEV *OrigReg = J->second;
2600 int64_t JImm = J->first;
2601 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2603 if (!isa<SCEVConstant>(OrigReg) &&
2604 UsedByIndicesMap[Reg].count() == 1) {
2605 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2609 // Conservatively examine offsets between this orig reg a few selected
2611 ImmMapTy::const_iterator OtherImms[] = {
2612 Imms.begin(), prior(Imms.end()),
2613 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2615 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2616 ImmMapTy::const_iterator M = OtherImms[i];
2617 if (M == J || M == JE) continue;
2619 // Compute the difference between the two.
2620 int64_t Imm = (uint64_t)JImm - M->first;
2621 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2622 LUIdx = UsedByIndices.find_next(LUIdx))
2623 // Make a memo of this use, offset, and register tuple.
2624 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2625 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2632 UsedByIndicesMap.clear();
2633 UniqueItems.clear();
2635 // Now iterate through the worklist and add new formulae.
2636 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2637 E = WorkItems.end(); I != E; ++I) {
2638 const WorkItem &WI = *I;
2639 size_t LUIdx = WI.LUIdx;
2640 LSRUse &LU = Uses[LUIdx];
2641 int64_t Imm = WI.Imm;
2642 const SCEV *OrigReg = WI.OrigReg;
2644 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2645 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2646 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2648 // TODO: Use a more targeted data structure.
2649 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2650 const Formula &F = LU.Formulae[L];
2651 // Use the immediate in the scaled register.
2652 if (F.ScaledReg == OrigReg) {
2653 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2654 Imm * (uint64_t)F.AM.Scale;
2655 // Don't create 50 + reg(-50).
2656 if (F.referencesReg(SE.getSCEV(
2657 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2660 NewF.AM.BaseOffs = Offs;
2661 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2662 LU.Kind, LU.AccessTy, TLI))
2664 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2666 // If the new scale is a constant in a register, and adding the constant
2667 // value to the immediate would produce a value closer to zero than the
2668 // immediate itself, then the formula isn't worthwhile.
2669 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2670 if (C->getValue()->getValue().isNegative() !=
2671 (NewF.AM.BaseOffs < 0) &&
2672 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2673 .ule(abs64(NewF.AM.BaseOffs)))
2677 (void)InsertFormula(LU, LUIdx, NewF);
2679 // Use the immediate in a base register.
2680 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2681 const SCEV *BaseReg = F.BaseRegs[N];
2682 if (BaseReg != OrigReg)
2685 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2686 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2687 LU.Kind, LU.AccessTy, TLI))
2689 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2691 // If the new formula has a constant in a register, and adding the
2692 // constant value to the immediate would produce a value closer to
2693 // zero than the immediate itself, then the formula isn't worthwhile.
2694 for (SmallVectorImpl<const SCEV *>::const_iterator
2695 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2697 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2698 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2699 abs64(NewF.AM.BaseOffs)) &&
2700 (C->getValue()->getValue() +
2701 NewF.AM.BaseOffs).countTrailingZeros() >=
2702 CountTrailingZeros_64(NewF.AM.BaseOffs))
2706 (void)InsertFormula(LU, LUIdx, NewF);
2715 /// GenerateAllReuseFormulae - Generate formulae for each use.
2717 LSRInstance::GenerateAllReuseFormulae() {
2718 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2719 // queries are more precise.
2720 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2721 LSRUse &LU = Uses[LUIdx];
2722 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2723 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2724 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2725 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2727 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2728 LSRUse &LU = Uses[LUIdx];
2729 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2730 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2731 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2732 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2733 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2734 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2735 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2736 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2738 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2739 LSRUse &LU = Uses[LUIdx];
2740 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2741 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2744 GenerateCrossUseConstantOffsets();
2747 /// If their are multiple formulae with the same set of registers used
2748 /// by other uses, pick the best one and delete the others.
2749 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2751 bool ChangedFormulae = false;
2754 // Collect the best formula for each unique set of shared registers. This
2755 // is reset for each use.
2756 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2758 BestFormulaeTy BestFormulae;
2760 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2761 LSRUse &LU = Uses[LUIdx];
2762 FormulaSorter Sorter(L, LU, SE, DT);
2763 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2766 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2767 FIdx != NumForms; ++FIdx) {
2768 Formula &F = LU.Formulae[FIdx];
2770 SmallVector<const SCEV *, 2> Key;
2771 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2772 JE = F.BaseRegs.end(); J != JE; ++J) {
2773 const SCEV *Reg = *J;
2774 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2778 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2779 Key.push_back(F.ScaledReg);
2780 // Unstable sort by host order ok, because this is only used for
2782 std::sort(Key.begin(), Key.end());
2784 std::pair<BestFormulaeTy::const_iterator, bool> P =
2785 BestFormulae.insert(std::make_pair(Key, FIdx));
2787 Formula &Best = LU.Formulae[P.first->second];
2788 if (Sorter.operator()(F, Best))
2790 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2792 " in favor of formula "; Best.print(dbgs());
2795 ChangedFormulae = true;
2797 LU.DeleteFormula(F);
2805 // Now that we've filtered out some formulae, recompute the Regs set.
2807 LU.RecomputeRegs(LUIdx, RegUses);
2809 // Reset this to prepare for the next use.
2810 BestFormulae.clear();
2813 DEBUG(if (ChangedFormulae) {
2815 "After filtering out undesirable candidates:\n";
2820 // This is a rough guess that seems to work fairly well.
2821 static const size_t ComplexityLimit = UINT16_MAX;
2823 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2824 /// solutions the solver might have to consider. It almost never considers
2825 /// this many solutions because it prune the search space, but the pruning
2826 /// isn't always sufficient.
2827 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2829 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2830 E = Uses.end(); I != E; ++I) {
2831 size_t FSize = I->Formulae.size();
2832 if (FSize >= ComplexityLimit) {
2833 Power = ComplexityLimit;
2837 if (Power >= ComplexityLimit)
2843 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2844 /// formulae to choose from, use some rough heuristics to prune down the number
2845 /// of formulae. This keeps the main solver from taking an extraordinary amount
2846 /// of time in some worst-case scenarios.
2847 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2848 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2849 DEBUG(dbgs() << "The search space is too complex.\n");
2851 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2852 "which use a superset of registers used by other "
2855 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2856 LSRUse &LU = Uses[LUIdx];
2858 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2859 Formula &F = LU.Formulae[i];
2860 // Look for a formula with a constant or GV in a register. If the use
2861 // also has a formula with that same value in an immediate field,
2862 // delete the one that uses a register.
2863 for (SmallVectorImpl<const SCEV *>::const_iterator
2864 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2865 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2867 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2868 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2869 (I - F.BaseRegs.begin()));
2870 if (LU.HasFormulaWithSameRegs(NewF)) {
2871 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2872 LU.DeleteFormula(F);
2878 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2879 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2882 NewF.AM.BaseGV = GV;
2883 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2884 (I - F.BaseRegs.begin()));
2885 if (LU.HasFormulaWithSameRegs(NewF)) {
2886 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2888 LU.DeleteFormula(F);
2899 LU.RecomputeRegs(LUIdx, RegUses);
2902 DEBUG(dbgs() << "After pre-selection:\n";
2903 print_uses(dbgs()));
2906 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2907 DEBUG(dbgs() << "The search space is too complex.\n");
2909 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2910 "separated by a constant offset will use the same "
2913 // This is especially useful for unrolled loops.
2915 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2916 LSRUse &LU = Uses[LUIdx];
2917 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2918 E = LU.Formulae.end(); I != E; ++I) {
2919 const Formula &F = *I;
2920 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2921 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2922 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2923 /*HasBaseReg=*/false,
2924 LU.Kind, LU.AccessTy)) {
2925 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2928 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2930 // Delete formulae from the new use which are no longer legal.
2932 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2933 Formula &F = LUThatHas->Formulae[i];
2934 if (!isLegalUse(F.AM,
2935 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2936 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2937 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2939 LUThatHas->DeleteFormula(F);
2946 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
2948 // Update the relocs to reference the new use.
2949 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
2950 E = Fixups.end(); I != E; ++I) {
2951 LSRFixup &Fixup = *I;
2952 if (Fixup.LUIdx == LUIdx) {
2953 Fixup.LUIdx = LUThatHas - &Uses.front();
2954 Fixup.Offset += F.AM.BaseOffs;
2955 DEBUG(errs() << "New fixup has offset "
2956 << Fixup.Offset << '\n');
2958 if (Fixup.LUIdx == NumUses-1)
2959 Fixup.LUIdx = LUIdx;
2962 // Delete the old use.
2973 DEBUG(dbgs() << "After pre-selection:\n";
2974 print_uses(dbgs()));
2977 // With all other options exhausted, loop until the system is simple
2978 // enough to handle.
2979 SmallPtrSet<const SCEV *, 4> Taken;
2980 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2981 // Ok, we have too many of formulae on our hands to conveniently handle.
2982 // Use a rough heuristic to thin out the list.
2983 DEBUG(dbgs() << "The search space is too complex.\n");
2985 // Pick the register which is used by the most LSRUses, which is likely
2986 // to be a good reuse register candidate.
2987 const SCEV *Best = 0;
2988 unsigned BestNum = 0;
2989 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2991 const SCEV *Reg = *I;
2992 if (Taken.count(Reg))
2997 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2998 if (Count > BestNum) {
3005 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3006 << " will yield profitable reuse.\n");
3009 // In any use with formulae which references this register, delete formulae
3010 // which don't reference it.
3011 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3012 LSRUse &LU = Uses[LUIdx];
3013 if (!LU.Regs.count(Best)) continue;
3016 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3017 Formula &F = LU.Formulae[i];
3018 if (!F.referencesReg(Best)) {
3019 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3020 LU.DeleteFormula(F);
3024 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3030 LU.RecomputeRegs(LUIdx, RegUses);
3033 DEBUG(dbgs() << "After pre-selection:\n";
3034 print_uses(dbgs()));
3038 /// SolveRecurse - This is the recursive solver.
3039 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3041 SmallVectorImpl<const Formula *> &Workspace,
3042 const Cost &CurCost,
3043 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3044 DenseSet<const SCEV *> &VisitedRegs) const {
3047 // - use more aggressive filtering
3048 // - sort the formula so that the most profitable solutions are found first
3049 // - sort the uses too
3051 // - don't compute a cost, and then compare. compare while computing a cost
3053 // - track register sets with SmallBitVector
3055 const LSRUse &LU = Uses[Workspace.size()];
3057 // If this use references any register that's already a part of the
3058 // in-progress solution, consider it a requirement that a formula must
3059 // reference that register in order to be considered. This prunes out
3060 // unprofitable searching.
3061 SmallSetVector<const SCEV *, 4> ReqRegs;
3062 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3063 E = CurRegs.end(); I != E; ++I)
3064 if (LU.Regs.count(*I))
3067 bool AnySatisfiedReqRegs = false;
3068 SmallPtrSet<const SCEV *, 16> NewRegs;
3071 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3072 E = LU.Formulae.end(); I != E; ++I) {
3073 const Formula &F = *I;
3075 // Ignore formulae which do not use any of the required registers.
3076 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3077 JE = ReqRegs.end(); J != JE; ++J) {
3078 const SCEV *Reg = *J;
3079 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3080 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3084 AnySatisfiedReqRegs = true;
3086 // Evaluate the cost of the current formula. If it's already worse than
3087 // the current best, prune the search at that point.
3090 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3091 if (NewCost < SolutionCost) {
3092 Workspace.push_back(&F);
3093 if (Workspace.size() != Uses.size()) {
3094 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3095 NewRegs, VisitedRegs);
3096 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3097 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3099 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3100 dbgs() << ". Regs:";
3101 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3102 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3103 dbgs() << ' ' << **I;
3106 SolutionCost = NewCost;
3107 Solution = Workspace;
3109 Workspace.pop_back();
3114 // If none of the formulae had all of the required registers, relax the
3115 // constraint so that we don't exclude all formulae.
3116 if (!AnySatisfiedReqRegs) {
3117 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3123 /// Solve - Choose one formula from each use. Return the results in the given
3124 /// Solution vector.
3125 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3126 SmallVector<const Formula *, 8> Workspace;
3128 SolutionCost.Loose();
3130 SmallPtrSet<const SCEV *, 16> CurRegs;
3131 DenseSet<const SCEV *> VisitedRegs;
3132 Workspace.reserve(Uses.size());
3134 // SolveRecurse does all the work.
3135 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3136 CurRegs, VisitedRegs);
3138 // Ok, we've now made all our decisions.
3139 DEBUG(dbgs() << "\n"
3140 "The chosen solution requires "; SolutionCost.print(dbgs());
3142 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3144 Uses[i].print(dbgs());
3147 Solution[i]->print(dbgs());
3151 assert(Solution.size() == Uses.size() && "Malformed solution!");
3154 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3155 /// the dominator tree far as we can go while still being dominated by the
3156 /// input positions. This helps canonicalize the insert position, which
3157 /// encourages sharing.
3158 BasicBlock::iterator
3159 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3160 const SmallVectorImpl<Instruction *> &Inputs)
3163 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3164 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3167 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3168 if (!Rung) return IP;
3169 Rung = Rung->getIDom();
3170 if (!Rung) return IP;
3171 IDom = Rung->getBlock();
3173 // Don't climb into a loop though.
3174 const Loop *IDomLoop = LI.getLoopFor(IDom);
3175 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3176 if (IDomDepth <= IPLoopDepth &&
3177 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3181 bool AllDominate = true;
3182 Instruction *BetterPos = 0;
3183 Instruction *Tentative = IDom->getTerminator();
3184 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3185 E = Inputs.end(); I != E; ++I) {
3186 Instruction *Inst = *I;
3187 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3188 AllDominate = false;
3191 // Attempt to find an insert position in the middle of the block,
3192 // instead of at the end, so that it can be used for other expansions.
3193 if (IDom == Inst->getParent() &&
3194 (!BetterPos || DT.dominates(BetterPos, Inst)))
3195 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3208 /// AdjustInsertPositionForExpand - Determine an input position which will be
3209 /// dominated by the operands and which will dominate the result.
3210 BasicBlock::iterator
3211 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3213 const LSRUse &LU) const {
3214 // Collect some instructions which must be dominated by the
3215 // expanding replacement. These must be dominated by any operands that
3216 // will be required in the expansion.
3217 SmallVector<Instruction *, 4> Inputs;
3218 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3219 Inputs.push_back(I);
3220 if (LU.Kind == LSRUse::ICmpZero)
3221 if (Instruction *I =
3222 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3223 Inputs.push_back(I);
3224 if (LF.PostIncLoops.count(L)) {
3225 if (LF.isUseFullyOutsideLoop(L))
3226 Inputs.push_back(L->getLoopLatch()->getTerminator());
3228 Inputs.push_back(IVIncInsertPos);
3230 // The expansion must also be dominated by the increment positions of any
3231 // loops it for which it is using post-inc mode.
3232 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3233 E = LF.PostIncLoops.end(); I != E; ++I) {
3234 const Loop *PIL = *I;
3235 if (PIL == L) continue;
3237 // Be dominated by the loop exit.
3238 SmallVector<BasicBlock *, 4> ExitingBlocks;
3239 PIL->getExitingBlocks(ExitingBlocks);
3240 if (!ExitingBlocks.empty()) {
3241 BasicBlock *BB = ExitingBlocks[0];
3242 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3243 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3244 Inputs.push_back(BB->getTerminator());
3248 // Then, climb up the immediate dominator tree as far as we can go while
3249 // still being dominated by the input positions.
3250 IP = HoistInsertPosition(IP, Inputs);
3252 // Don't insert instructions before PHI nodes.
3253 while (isa<PHINode>(IP)) ++IP;
3255 // Ignore debug intrinsics.
3256 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3261 /// Expand - Emit instructions for the leading candidate expression for this
3262 /// LSRUse (this is called "expanding").
3263 Value *LSRInstance::Expand(const LSRFixup &LF,
3265 BasicBlock::iterator IP,
3266 SCEVExpander &Rewriter,
3267 SmallVectorImpl<WeakVH> &DeadInsts) const {
3268 const LSRUse &LU = Uses[LF.LUIdx];
3270 // Determine an input position which will be dominated by the operands and
3271 // which will dominate the result.
3272 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3274 // Inform the Rewriter if we have a post-increment use, so that it can
3275 // perform an advantageous expansion.
3276 Rewriter.setPostInc(LF.PostIncLoops);
3278 // This is the type that the user actually needs.
3279 const Type *OpTy = LF.OperandValToReplace->getType();
3280 // This will be the type that we'll initially expand to.
3281 const Type *Ty = F.getType();
3283 // No type known; just expand directly to the ultimate type.
3285 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3286 // Expand directly to the ultimate type if it's the right size.
3288 // This is the type to do integer arithmetic in.
3289 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3291 // Build up a list of operands to add together to form the full base.
3292 SmallVector<const SCEV *, 8> Ops;
3294 // Expand the BaseRegs portion.
3295 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3296 E = F.BaseRegs.end(); I != E; ++I) {
3297 const SCEV *Reg = *I;
3298 assert(!Reg->isZero() && "Zero allocated in a base register!");
3300 // If we're expanding for a post-inc user, make the post-inc adjustment.
3301 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3302 Reg = TransformForPostIncUse(Denormalize, Reg,
3303 LF.UserInst, LF.OperandValToReplace,
3306 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3309 // Flush the operand list to suppress SCEVExpander hoisting.
3311 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3313 Ops.push_back(SE.getUnknown(FullV));
3316 // Expand the ScaledReg portion.
3317 Value *ICmpScaledV = 0;
3318 if (F.AM.Scale != 0) {
3319 const SCEV *ScaledS = F.ScaledReg;
3321 // If we're expanding for a post-inc user, make the post-inc adjustment.
3322 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3323 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3324 LF.UserInst, LF.OperandValToReplace,
3327 if (LU.Kind == LSRUse::ICmpZero) {
3328 // An interesting way of "folding" with an icmp is to use a negated
3329 // scale, which we'll implement by inserting it into the other operand
3331 assert(F.AM.Scale == -1 &&
3332 "The only scale supported by ICmpZero uses is -1!");
3333 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3335 // Otherwise just expand the scaled register and an explicit scale,
3336 // which is expected to be matched as part of the address.
3337 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3338 ScaledS = SE.getMulExpr(ScaledS,
3339 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3340 Ops.push_back(ScaledS);
3342 // Flush the operand list to suppress SCEVExpander hoisting.
3343 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3345 Ops.push_back(SE.getUnknown(FullV));
3349 // Expand the GV portion.
3351 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3353 // Flush the operand list to suppress SCEVExpander hoisting.
3354 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3356 Ops.push_back(SE.getUnknown(FullV));
3359 // Expand the immediate portion.
3360 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3362 if (LU.Kind == LSRUse::ICmpZero) {
3363 // The other interesting way of "folding" with an ICmpZero is to use a
3364 // negated immediate.
3366 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3368 Ops.push_back(SE.getUnknown(ICmpScaledV));
3369 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3372 // Just add the immediate values. These again are expected to be matched
3373 // as part of the address.
3374 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3378 // Emit instructions summing all the operands.
3379 const SCEV *FullS = Ops.empty() ?
3380 SE.getConstant(IntTy, 0) :
3382 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3384 // We're done expanding now, so reset the rewriter.
3385 Rewriter.clearPostInc();
3387 // An ICmpZero Formula represents an ICmp which we're handling as a
3388 // comparison against zero. Now that we've expanded an expression for that
3389 // form, update the ICmp's other operand.
3390 if (LU.Kind == LSRUse::ICmpZero) {
3391 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3392 DeadInsts.push_back(CI->getOperand(1));
3393 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3394 "a scale at the same time!");
3395 if (F.AM.Scale == -1) {
3396 if (ICmpScaledV->getType() != OpTy) {
3398 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3400 ICmpScaledV, OpTy, "tmp", CI);
3403 CI->setOperand(1, ICmpScaledV);
3405 assert(F.AM.Scale == 0 &&
3406 "ICmp does not support folding a global value and "
3407 "a scale at the same time!");
3408 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3410 if (C->getType() != OpTy)
3411 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3415 CI->setOperand(1, C);
3422 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3423 /// of their operands effectively happens in their predecessor blocks, so the
3424 /// expression may need to be expanded in multiple places.
3425 void LSRInstance::RewriteForPHI(PHINode *PN,
3428 SCEVExpander &Rewriter,
3429 SmallVectorImpl<WeakVH> &DeadInsts,
3431 DenseMap<BasicBlock *, Value *> Inserted;
3432 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3433 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3434 BasicBlock *BB = PN->getIncomingBlock(i);
3436 // If this is a critical edge, split the edge so that we do not insert
3437 // the code on all predecessor/successor paths. We do this unless this
3438 // is the canonical backedge for this loop, which complicates post-inc
3440 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3441 !isa<IndirectBrInst>(BB->getTerminator()) &&
3442 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3443 // Split the critical edge.
3444 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3446 // If PN is outside of the loop and BB is in the loop, we want to
3447 // move the block to be immediately before the PHI block, not
3448 // immediately after BB.
3449 if (L->contains(BB) && !L->contains(PN))
3450 NewBB->moveBefore(PN->getParent());
3452 // Splitting the edge can reduce the number of PHI entries we have.
3453 e = PN->getNumIncomingValues();
3455 i = PN->getBasicBlockIndex(BB);
3458 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3459 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3461 PN->setIncomingValue(i, Pair.first->second);
3463 Value *FullV = Expand(LF, F, BB->getTerminator(), 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,
3471 FullV, LF.OperandValToReplace->getType(),
3472 "tmp", BB->getTerminator());
3474 PN->setIncomingValue(i, FullV);
3475 Pair.first->second = FullV;
3480 /// Rewrite - Emit instructions for the leading candidate expression for this
3481 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3482 /// the newly expanded value.
3483 void LSRInstance::Rewrite(const LSRFixup &LF,
3485 SCEVExpander &Rewriter,
3486 SmallVectorImpl<WeakVH> &DeadInsts,
3488 // First, find an insertion point that dominates UserInst. For PHI nodes,
3489 // find the nearest block which dominates all the relevant uses.
3490 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3491 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3493 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3495 // If this is reuse-by-noop-cast, insert the noop cast.
3496 const Type *OpTy = LF.OperandValToReplace->getType();
3497 if (FullV->getType() != OpTy) {
3499 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3500 FullV, OpTy, "tmp", LF.UserInst);
3504 // Update the user. ICmpZero is handled specially here (for now) because
3505 // Expand may have updated one of the operands of the icmp already, and
3506 // its new value may happen to be equal to LF.OperandValToReplace, in
3507 // which case doing replaceUsesOfWith leads to replacing both operands
3508 // with the same value. TODO: Reorganize this.
3509 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3510 LF.UserInst->setOperand(0, FullV);
3512 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3515 DeadInsts.push_back(LF.OperandValToReplace);
3518 /// ImplementSolution - Rewrite all the fixup locations with new values,
3519 /// following the chosen solution.
3521 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3523 // Keep track of instructions we may have made dead, so that
3524 // we can remove them after we are done working.
3525 SmallVector<WeakVH, 16> DeadInsts;
3527 SCEVExpander Rewriter(SE);
3528 Rewriter.disableCanonicalMode();
3529 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3531 // Expand the new value definitions and update the users.
3532 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3533 E = Fixups.end(); I != E; ++I) {
3534 const LSRFixup &Fixup = *I;
3536 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3541 // Clean up after ourselves. This must be done before deleting any
3545 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3548 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3549 : IU(P->getAnalysis<IVUsers>()),
3550 SE(P->getAnalysis<ScalarEvolution>()),
3551 DT(P->getAnalysis<DominatorTree>()),
3552 LI(P->getAnalysis<LoopInfo>()),
3553 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3555 // If LoopSimplify form is not available, stay out of trouble.
3556 if (!L->isLoopSimplifyForm()) return;
3558 // If there's no interesting work to be done, bail early.
3559 if (IU.empty()) return;
3561 DEBUG(dbgs() << "\nLSR on loop ";
3562 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3565 // First, perform some low-level loop optimizations.
3567 OptimizeLoopTermCond();
3569 // Start collecting data and preparing for the solver.
3570 CollectInterestingTypesAndFactors();
3571 CollectFixupsAndInitialFormulae();
3572 CollectLoopInvariantFixupsAndFormulae();
3574 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3575 print_uses(dbgs()));
3577 // Now use the reuse data to generate a bunch of interesting ways
3578 // to formulate the values needed for the uses.
3579 GenerateAllReuseFormulae();
3581 DEBUG(dbgs() << "\n"
3582 "After generating reuse formulae:\n";
3583 print_uses(dbgs()));
3585 FilterOutUndesirableDedicatedRegisters();
3586 NarrowSearchSpaceUsingHeuristics();
3588 SmallVector<const Formula *, 8> Solution;
3591 // Release memory that is no longer needed.
3597 // Formulae should be legal.
3598 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3599 E = Uses.end(); I != E; ++I) {
3600 const LSRUse &LU = *I;
3601 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3602 JE = LU.Formulae.end(); J != JE; ++J)
3603 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3604 LU.Kind, LU.AccessTy, TLI) &&
3605 "Illegal formula generated!");
3609 // Now that we've decided what we want, make it so.
3610 ImplementSolution(Solution, P);
3613 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3614 if (Factors.empty() && Types.empty()) return;
3616 OS << "LSR has identified the following interesting factors and types: ";
3619 for (SmallSetVector<int64_t, 8>::const_iterator
3620 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3621 if (!First) OS << ", ";
3626 for (SmallSetVector<const Type *, 4>::const_iterator
3627 I = Types.begin(), E = Types.end(); I != E; ++I) {
3628 if (!First) OS << ", ";
3630 OS << '(' << **I << ')';
3635 void LSRInstance::print_fixups(raw_ostream &OS) const {
3636 OS << "LSR is examining the following fixup sites:\n";
3637 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3638 E = Fixups.end(); I != E; ++I) {
3645 void LSRInstance::print_uses(raw_ostream &OS) const {
3646 OS << "LSR is examining the following uses:\n";
3647 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3648 E = Uses.end(); I != E; ++I) {
3649 const LSRUse &LU = *I;
3653 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3654 JE = LU.Formulae.end(); J != JE; ++J) {
3662 void LSRInstance::print(raw_ostream &OS) const {
3663 print_factors_and_types(OS);
3668 void LSRInstance::dump() const {
3669 print(errs()); errs() << '\n';
3674 class LoopStrengthReduce : public LoopPass {
3675 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3676 /// transformation profitability.
3677 const TargetLowering *const TLI;
3680 static char ID; // Pass ID, replacement for typeid
3681 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3684 bool runOnLoop(Loop *L, LPPassManager &LPM);
3685 void getAnalysisUsage(AnalysisUsage &AU) const;
3690 char LoopStrengthReduce::ID = 0;
3691 static RegisterPass<LoopStrengthReduce>
3692 X("loop-reduce", "Loop Strength Reduction");
3694 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3695 return new LoopStrengthReduce(TLI);
3698 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3699 : LoopPass(&ID), TLI(tli) {}
3701 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3702 // We split critical edges, so we change the CFG. However, we do update
3703 // many analyses if they are around.
3704 AU.addPreservedID(LoopSimplifyID);
3705 AU.addPreserved("domfrontier");
3707 AU.addRequired<LoopInfo>();
3708 AU.addPreserved<LoopInfo>();
3709 AU.addRequiredID(LoopSimplifyID);
3710 AU.addRequired<DominatorTree>();
3711 AU.addPreserved<DominatorTree>();
3712 AU.addRequired<ScalarEvolution>();
3713 AU.addPreserved<ScalarEvolution>();
3714 AU.addRequired<IVUsers>();
3715 AU.addPreserved<IVUsers>();
3718 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3719 bool Changed = false;
3721 // Run the main LSR transformation.
3722 Changed |= LSRInstance(TLI, L, this).getChanged();
3724 // At this point, it is worth checking to see if any recurrence PHIs are also
3725 // dead, so that we can remove them as well.
3726 Changed |= DeleteDeadPHIs(L->getHeader());