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 mul can be sign-extended
396 /// without changing its value.
397 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
399 IntegerType::get(SE.getContext(),
400 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
401 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
404 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
405 /// and if the remainder is known to be zero, or null otherwise. If
406 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
407 /// to Y, ignoring that the multiplication may overflow, which is useful when
408 /// the result will be used in a context where the most significant bits are
410 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
412 bool IgnoreSignificantBits = false) {
413 // Handle the trivial case, which works for any SCEV type.
415 return SE.getConstant(LHS->getType(), 1);
417 // Handle a few RHS special cases.
418 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
420 const APInt &RA = RC->getValue()->getValue();
421 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
423 if (RA.isAllOnesValue())
424 return SE.getMulExpr(LHS, RC);
425 // Handle x /s 1 as x.
430 // Check for a division of a constant by a constant.
431 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
434 const APInt &LA = C->getValue()->getValue();
435 const APInt &RA = RC->getValue()->getValue();
436 if (LA.srem(RA) != 0)
438 return SE.getConstant(LA.sdiv(RA));
441 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
442 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
443 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
444 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
445 IgnoreSignificantBits);
446 if (!Start) return 0;
447 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
448 IgnoreSignificantBits);
450 return SE.getAddRecExpr(Start, Step, AR->getLoop());
455 // Distribute the sdiv over add operands, if the add doesn't overflow.
456 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
457 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
458 SmallVector<const SCEV *, 8> Ops;
459 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
461 const SCEV *Op = getExactSDiv(*I, RHS, SE,
462 IgnoreSignificantBits);
466 return SE.getAddExpr(Ops);
471 // Check for a multiply operand that we can pull RHS out of.
472 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
473 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
474 SmallVector<const SCEV *, 4> Ops;
476 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
480 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
481 IgnoreSignificantBits)) {
487 return Found ? SE.getMulExpr(Ops) : 0;
492 // Otherwise we don't know.
496 /// ExtractImmediate - If S involves the addition of a constant integer value,
497 /// return that integer value, and mutate S to point to a new SCEV with that
499 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
500 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
501 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
502 S = SE.getConstant(C->getType(), 0);
503 return C->getValue()->getSExtValue();
505 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
506 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
507 int64_t Result = ExtractImmediate(NewOps.front(), SE);
509 S = SE.getAddExpr(NewOps);
511 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
512 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
513 int64_t Result = ExtractImmediate(NewOps.front(), SE);
515 S = SE.getAddRecExpr(NewOps, AR->getLoop());
521 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
522 /// return that symbol, and mutate S to point to a new SCEV with that
524 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
525 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
526 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
527 S = SE.getConstant(GV->getType(), 0);
530 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
531 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
532 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
534 S = SE.getAddExpr(NewOps);
536 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
537 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
538 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
540 S = SE.getAddRecExpr(NewOps, AR->getLoop());
546 /// isAddressUse - Returns true if the specified instruction is using the
547 /// specified value as an address.
548 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
549 bool isAddress = isa<LoadInst>(Inst);
550 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
551 if (SI->getOperand(1) == OperandVal)
553 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
554 // Addressing modes can also be folded into prefetches and a variety
556 switch (II->getIntrinsicID()) {
558 case Intrinsic::prefetch:
559 case Intrinsic::x86_sse2_loadu_dq:
560 case Intrinsic::x86_sse2_loadu_pd:
561 case Intrinsic::x86_sse_loadu_ps:
562 case Intrinsic::x86_sse_storeu_ps:
563 case Intrinsic::x86_sse2_storeu_pd:
564 case Intrinsic::x86_sse2_storeu_dq:
565 case Intrinsic::x86_sse2_storel_dq:
566 if (II->getArgOperand(0) == OperandVal)
574 /// getAccessType - Return the type of the memory being accessed.
575 static const Type *getAccessType(const Instruction *Inst) {
576 const Type *AccessTy = Inst->getType();
577 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
578 AccessTy = SI->getOperand(0)->getType();
579 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
580 // Addressing modes can also be folded into prefetches and a variety
582 switch (II->getIntrinsicID()) {
584 case Intrinsic::x86_sse_storeu_ps:
585 case Intrinsic::x86_sse2_storeu_pd:
586 case Intrinsic::x86_sse2_storeu_dq:
587 case Intrinsic::x86_sse2_storel_dq:
588 AccessTy = II->getArgOperand(0)->getType();
593 // All pointers have the same requirements, so canonicalize them to an
594 // arbitrary pointer type to minimize variation.
595 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
596 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
597 PTy->getAddressSpace());
602 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
603 /// specified set are trivially dead, delete them and see if this makes any of
604 /// their operands subsequently dead.
606 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
607 bool Changed = false;
609 while (!DeadInsts.empty()) {
610 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
612 if (I == 0 || !isInstructionTriviallyDead(I))
615 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
616 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
619 DeadInsts.push_back(U);
622 I->eraseFromParent();
631 /// Cost - This class is used to measure and compare candidate formulae.
633 /// TODO: Some of these could be merged. Also, a lexical ordering
634 /// isn't always optimal.
638 unsigned NumBaseAdds;
644 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
647 unsigned getNumRegs() const { return NumRegs; }
649 bool operator<(const Cost &Other) const;
653 void RateFormula(const Formula &F,
654 SmallPtrSet<const SCEV *, 16> &Regs,
655 const DenseSet<const SCEV *> &VisitedRegs,
657 const SmallVectorImpl<int64_t> &Offsets,
658 ScalarEvolution &SE, DominatorTree &DT);
660 void print(raw_ostream &OS) const;
664 void RateRegister(const SCEV *Reg,
665 SmallPtrSet<const SCEV *, 16> &Regs,
667 ScalarEvolution &SE, DominatorTree &DT);
668 void RatePrimaryRegister(const SCEV *Reg,
669 SmallPtrSet<const SCEV *, 16> &Regs,
671 ScalarEvolution &SE, DominatorTree &DT);
676 /// RateRegister - Tally up interesting quantities from the given register.
677 void Cost::RateRegister(const SCEV *Reg,
678 SmallPtrSet<const SCEV *, 16> &Regs,
680 ScalarEvolution &SE, DominatorTree &DT) {
681 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
682 if (AR->getLoop() == L)
683 AddRecCost += 1; /// TODO: This should be a function of the stride.
685 // If this is an addrec for a loop that's already been visited by LSR,
686 // don't second-guess its addrec phi nodes. LSR isn't currently smart
687 // enough to reason about more than one loop at a time. Consider these
688 // registers free and leave them alone.
689 else if (L->contains(AR->getLoop()) ||
690 (!AR->getLoop()->contains(L) &&
691 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
692 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
693 PHINode *PN = dyn_cast<PHINode>(I); ++I)
694 if (SE.isSCEVable(PN->getType()) &&
695 (SE.getEffectiveSCEVType(PN->getType()) ==
696 SE.getEffectiveSCEVType(AR->getType())) &&
697 SE.getSCEV(PN) == AR)
700 // If this isn't one of the addrecs that the loop already has, it
701 // would require a costly new phi and add. TODO: This isn't
702 // precisely modeled right now.
704 if (!Regs.count(AR->getStart()))
705 RateRegister(AR->getStart(), Regs, L, SE, DT);
708 // Add the step value register, if it needs one.
709 // TODO: The non-affine case isn't precisely modeled here.
710 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
711 if (!Regs.count(AR->getStart()))
712 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
716 // Rough heuristic; favor registers which don't require extra setup
717 // instructions in the preheader.
718 if (!isa<SCEVUnknown>(Reg) &&
719 !isa<SCEVConstant>(Reg) &&
720 !(isa<SCEVAddRecExpr>(Reg) &&
721 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
722 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
726 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
728 void Cost::RatePrimaryRegister(const SCEV *Reg,
729 SmallPtrSet<const SCEV *, 16> &Regs,
731 ScalarEvolution &SE, DominatorTree &DT) {
732 if (Regs.insert(Reg))
733 RateRegister(Reg, Regs, L, SE, DT);
736 void Cost::RateFormula(const Formula &F,
737 SmallPtrSet<const SCEV *, 16> &Regs,
738 const DenseSet<const SCEV *> &VisitedRegs,
740 const SmallVectorImpl<int64_t> &Offsets,
741 ScalarEvolution &SE, DominatorTree &DT) {
742 // Tally up the registers.
743 if (const SCEV *ScaledReg = F.ScaledReg) {
744 if (VisitedRegs.count(ScaledReg)) {
748 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
750 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
751 E = F.BaseRegs.end(); I != E; ++I) {
752 const SCEV *BaseReg = *I;
753 if (VisitedRegs.count(BaseReg)) {
757 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
759 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
760 BaseReg->hasComputableLoopEvolution(L);
763 if (F.BaseRegs.size() > 1)
764 NumBaseAdds += F.BaseRegs.size() - 1;
766 // Tally up the non-zero immediates.
767 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
768 E = Offsets.end(); I != E; ++I) {
769 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
771 ImmCost += 64; // Handle symbolic values conservatively.
772 // TODO: This should probably be the pointer size.
773 else if (Offset != 0)
774 ImmCost += APInt(64, Offset, true).getMinSignedBits();
778 /// Loose - Set this cost to a loosing value.
788 /// operator< - Choose the lower cost.
789 bool Cost::operator<(const Cost &Other) const {
790 if (NumRegs != Other.NumRegs)
791 return NumRegs < Other.NumRegs;
792 if (AddRecCost != Other.AddRecCost)
793 return AddRecCost < Other.AddRecCost;
794 if (NumIVMuls != Other.NumIVMuls)
795 return NumIVMuls < Other.NumIVMuls;
796 if (NumBaseAdds != Other.NumBaseAdds)
797 return NumBaseAdds < Other.NumBaseAdds;
798 if (ImmCost != Other.ImmCost)
799 return ImmCost < Other.ImmCost;
800 if (SetupCost != Other.SetupCost)
801 return SetupCost < Other.SetupCost;
805 void Cost::print(raw_ostream &OS) const {
806 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
808 OS << ", with addrec cost " << AddRecCost;
810 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
811 if (NumBaseAdds != 0)
812 OS << ", plus " << NumBaseAdds << " base add"
813 << (NumBaseAdds == 1 ? "" : "s");
815 OS << ", plus " << ImmCost << " imm cost";
817 OS << ", plus " << SetupCost << " setup cost";
820 void Cost::dump() const {
821 print(errs()); errs() << '\n';
826 /// LSRFixup - An operand value in an instruction which is to be replaced
827 /// with some equivalent, possibly strength-reduced, replacement.
829 /// UserInst - The instruction which will be updated.
830 Instruction *UserInst;
832 /// OperandValToReplace - The operand of the instruction which will
833 /// be replaced. The operand may be used more than once; every instance
834 /// will be replaced.
835 Value *OperandValToReplace;
837 /// PostIncLoops - If this user is to use the post-incremented value of an
838 /// induction variable, this variable is non-null and holds the loop
839 /// associated with the induction variable.
840 PostIncLoopSet PostIncLoops;
842 /// LUIdx - The index of the LSRUse describing the expression which
843 /// this fixup needs, minus an offset (below).
846 /// Offset - A constant offset to be added to the LSRUse expression.
847 /// This allows multiple fixups to share the same LSRUse with different
848 /// offsets, for example in an unrolled loop.
851 bool isUseFullyOutsideLoop(const Loop *L) const;
855 void print(raw_ostream &OS) const;
862 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
864 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
865 /// value outside of the given loop.
866 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
867 // PHI nodes use their value in their incoming blocks.
868 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
869 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
870 if (PN->getIncomingValue(i) == OperandValToReplace &&
871 L->contains(PN->getIncomingBlock(i)))
876 return !L->contains(UserInst);
879 void LSRFixup::print(raw_ostream &OS) const {
881 // Store is common and interesting enough to be worth special-casing.
882 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
884 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
885 } else if (UserInst->getType()->isVoidTy())
886 OS << UserInst->getOpcodeName();
888 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
890 OS << ", OperandValToReplace=";
891 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
893 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
894 E = PostIncLoops.end(); I != E; ++I) {
895 OS << ", PostIncLoop=";
896 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
899 if (LUIdx != ~size_t(0))
900 OS << ", LUIdx=" << LUIdx;
903 OS << ", Offset=" << Offset;
906 void LSRFixup::dump() const {
907 print(errs()); errs() << '\n';
912 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
913 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
914 struct UniquifierDenseMapInfo {
915 static SmallVector<const SCEV *, 2> getEmptyKey() {
916 SmallVector<const SCEV *, 2> V;
917 V.push_back(reinterpret_cast<const SCEV *>(-1));
921 static SmallVector<const SCEV *, 2> getTombstoneKey() {
922 SmallVector<const SCEV *, 2> V;
923 V.push_back(reinterpret_cast<const SCEV *>(-2));
927 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
929 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
930 E = V.end(); I != E; ++I)
931 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
935 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
936 const SmallVector<const SCEV *, 2> &RHS) {
941 /// LSRUse - This class holds the state that LSR keeps for each use in
942 /// IVUsers, as well as uses invented by LSR itself. It includes information
943 /// about what kinds of things can be folded into the user, information about
944 /// the user itself, and information about how the use may be satisfied.
945 /// TODO: Represent multiple users of the same expression in common?
947 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
950 /// KindType - An enum for a kind of use, indicating what types of
951 /// scaled and immediate operands it might support.
953 Basic, ///< A normal use, with no folding.
954 Special, ///< A special case of basic, allowing -1 scales.
955 Address, ///< An address use; folding according to TargetLowering
956 ICmpZero ///< An equality icmp with both operands folded into one.
957 // TODO: Add a generic icmp too?
961 const Type *AccessTy;
963 SmallVector<int64_t, 8> Offsets;
967 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
968 /// LSRUse are outside of the loop, in which case some special-case heuristics
970 bool AllFixupsOutsideLoop;
972 /// WidestFixupType - This records the widest use type for any fixup using
973 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
974 /// max fixup widths to be equivalent, because the narrower one may be relying
975 /// on the implicit truncation to truncate away bogus bits.
976 const Type *WidestFixupType;
978 /// Formulae - A list of ways to build a value that can satisfy this user.
979 /// After the list is populated, one of these is selected heuristically and
980 /// used to formulate a replacement for OperandValToReplace in UserInst.
981 SmallVector<Formula, 12> Formulae;
983 /// Regs - The set of register candidates used by all formulae in this LSRUse.
984 SmallPtrSet<const SCEV *, 4> Regs;
986 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
987 MinOffset(INT64_MAX),
988 MaxOffset(INT64_MIN),
989 AllFixupsOutsideLoop(true),
990 WidestFixupType(0) {}
992 bool HasFormulaWithSameRegs(const Formula &F) const;
993 bool InsertFormula(const Formula &F);
994 void DeleteFormula(Formula &F);
995 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
999 void print(raw_ostream &OS) const;
1005 /// HasFormula - Test whether this use as a formula which has the same
1006 /// registers as the given formula.
1007 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1008 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1009 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1010 // Unstable sort by host order ok, because this is only used for uniquifying.
1011 std::sort(Key.begin(), Key.end());
1012 return Uniquifier.count(Key);
1015 /// InsertFormula - If the given formula has not yet been inserted, add it to
1016 /// the list, and return true. Return false otherwise.
1017 bool LSRUse::InsertFormula(const Formula &F) {
1018 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1019 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1020 // Unstable sort by host order ok, because this is only used for uniquifying.
1021 std::sort(Key.begin(), Key.end());
1023 if (!Uniquifier.insert(Key).second)
1026 // Using a register to hold the value of 0 is not profitable.
1027 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1028 "Zero allocated in a scaled register!");
1030 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1031 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1032 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1035 // Add the formula to the list.
1036 Formulae.push_back(F);
1038 // Record registers now being used by this use.
1039 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1040 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1045 /// DeleteFormula - Remove the given formula from this use's list.
1046 void LSRUse::DeleteFormula(Formula &F) {
1047 if (&F != &Formulae.back())
1048 std::swap(F, Formulae.back());
1049 Formulae.pop_back();
1050 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1053 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1054 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1055 // Now that we've filtered out some formulae, recompute the Regs set.
1056 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1058 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1059 E = Formulae.end(); I != E; ++I) {
1060 const Formula &F = *I;
1061 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1062 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1065 // Update the RegTracker.
1066 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1067 E = OldRegs.end(); I != E; ++I)
1068 if (!Regs.count(*I))
1069 RegUses.DropRegister(*I, LUIdx);
1072 void LSRUse::print(raw_ostream &OS) const {
1073 OS << "LSR Use: Kind=";
1075 case Basic: OS << "Basic"; break;
1076 case Special: OS << "Special"; break;
1077 case ICmpZero: OS << "ICmpZero"; break;
1079 OS << "Address of ";
1080 if (AccessTy->isPointerTy())
1081 OS << "pointer"; // the full pointer type could be really verbose
1086 OS << ", Offsets={";
1087 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1088 E = Offsets.end(); I != E; ++I) {
1090 if (llvm::next(I) != E)
1095 if (AllFixupsOutsideLoop)
1096 OS << ", all-fixups-outside-loop";
1098 if (WidestFixupType)
1099 OS << ", widest fixup type: " << *WidestFixupType;
1102 void LSRUse::dump() const {
1103 print(errs()); errs() << '\n';
1106 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1107 /// be completely folded into the user instruction at isel time. This includes
1108 /// address-mode folding and special icmp tricks.
1109 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1110 LSRUse::KindType Kind, const Type *AccessTy,
1111 const TargetLowering *TLI) {
1113 case LSRUse::Address:
1114 // If we have low-level target information, ask the target if it can
1115 // completely fold this address.
1116 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1118 // Otherwise, just guess that reg+reg addressing is legal.
1119 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1121 case LSRUse::ICmpZero:
1122 // There's not even a target hook for querying whether it would be legal to
1123 // fold a GV into an ICmp.
1127 // ICmp only has two operands; don't allow more than two non-trivial parts.
1128 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1131 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1132 // putting the scaled register in the other operand of the icmp.
1133 if (AM.Scale != 0 && AM.Scale != -1)
1136 // If we have low-level target information, ask the target if it can fold an
1137 // integer immediate on an icmp.
1138 if (AM.BaseOffs != 0) {
1139 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1146 // Only handle single-register values.
1147 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1149 case LSRUse::Special:
1150 // Only handle -1 scales, or no scale.
1151 return AM.Scale == 0 || AM.Scale == -1;
1157 static bool isLegalUse(TargetLowering::AddrMode AM,
1158 int64_t MinOffset, int64_t MaxOffset,
1159 LSRUse::KindType Kind, const Type *AccessTy,
1160 const TargetLowering *TLI) {
1161 // Check for overflow.
1162 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1165 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1166 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1167 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1168 // Check for overflow.
1169 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1172 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1173 return isLegalUse(AM, Kind, AccessTy, TLI);
1178 static bool isAlwaysFoldable(int64_t BaseOffs,
1179 GlobalValue *BaseGV,
1181 LSRUse::KindType Kind, const Type *AccessTy,
1182 const TargetLowering *TLI) {
1183 // Fast-path: zero is always foldable.
1184 if (BaseOffs == 0 && !BaseGV) return true;
1186 // Conservatively, create an address with an immediate and a
1187 // base and a scale.
1188 TargetLowering::AddrMode AM;
1189 AM.BaseOffs = BaseOffs;
1191 AM.HasBaseReg = HasBaseReg;
1192 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1194 // Canonicalize a scale of 1 to a base register if the formula doesn't
1195 // already have a base register.
1196 if (!AM.HasBaseReg && AM.Scale == 1) {
1198 AM.HasBaseReg = true;
1201 return isLegalUse(AM, Kind, AccessTy, TLI);
1204 static bool isAlwaysFoldable(const SCEV *S,
1205 int64_t MinOffset, int64_t MaxOffset,
1207 LSRUse::KindType Kind, const Type *AccessTy,
1208 const TargetLowering *TLI,
1209 ScalarEvolution &SE) {
1210 // Fast-path: zero is always foldable.
1211 if (S->isZero()) return true;
1213 // Conservatively, create an address with an immediate and a
1214 // base and a scale.
1215 int64_t BaseOffs = ExtractImmediate(S, SE);
1216 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1218 // If there's anything else involved, it's not foldable.
1219 if (!S->isZero()) return false;
1221 // Fast-path: zero is always foldable.
1222 if (BaseOffs == 0 && !BaseGV) return true;
1224 // Conservatively, create an address with an immediate and a
1225 // base and a scale.
1226 TargetLowering::AddrMode AM;
1227 AM.BaseOffs = BaseOffs;
1229 AM.HasBaseReg = HasBaseReg;
1230 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1232 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1237 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1238 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1239 struct UseMapDenseMapInfo {
1240 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1241 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1244 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1245 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1249 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1250 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1251 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1255 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1256 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1261 /// FormulaSorter - This class implements an ordering for formulae which sorts
1262 /// the by their standalone cost.
1263 class FormulaSorter {
1264 /// These two sets are kept empty, so that we compute standalone costs.
1265 DenseSet<const SCEV *> VisitedRegs;
1266 SmallPtrSet<const SCEV *, 16> Regs;
1269 ScalarEvolution &SE;
1273 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1274 : L(l), LU(&lu), SE(se), DT(dt) {}
1276 bool operator()(const Formula &A, const Formula &B) {
1278 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1281 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1283 return CostA < CostB;
1287 /// LSRInstance - This class holds state for the main loop strength reduction
1291 ScalarEvolution &SE;
1294 const TargetLowering *const TLI;
1298 /// IVIncInsertPos - This is the insert position that the current loop's
1299 /// induction variable increment should be placed. In simple loops, this is
1300 /// the latch block's terminator. But in more complicated cases, this is a
1301 /// position which will dominate all the in-loop post-increment users.
1302 Instruction *IVIncInsertPos;
1304 /// Factors - Interesting factors between use strides.
1305 SmallSetVector<int64_t, 8> Factors;
1307 /// Types - Interesting use types, to facilitate truncation reuse.
1308 SmallSetVector<const Type *, 4> Types;
1310 /// Fixups - The list of operands which are to be replaced.
1311 SmallVector<LSRFixup, 16> Fixups;
1313 /// Uses - The list of interesting uses.
1314 SmallVector<LSRUse, 16> Uses;
1316 /// RegUses - Track which uses use which register candidates.
1317 RegUseTracker RegUses;
1319 void OptimizeShadowIV();
1320 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1321 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1322 void OptimizeLoopTermCond();
1324 void CollectInterestingTypesAndFactors();
1325 void CollectFixupsAndInitialFormulae();
1327 LSRFixup &getNewFixup() {
1328 Fixups.push_back(LSRFixup());
1329 return Fixups.back();
1332 // Support for sharing of LSRUses between LSRFixups.
1333 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1335 UseMapDenseMapInfo> UseMapTy;
1338 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1339 LSRUse::KindType Kind, const Type *AccessTy);
1341 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1342 LSRUse::KindType Kind,
1343 const Type *AccessTy);
1345 void DeleteUse(LSRUse &LU);
1347 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1350 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1351 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1352 void CountRegisters(const Formula &F, size_t LUIdx);
1353 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1355 void CollectLoopInvariantFixupsAndFormulae();
1357 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1358 unsigned Depth = 0);
1359 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1360 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1361 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1362 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1363 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1364 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1365 void GenerateCrossUseConstantOffsets();
1366 void GenerateAllReuseFormulae();
1368 void FilterOutUndesirableDedicatedRegisters();
1370 size_t EstimateSearchSpaceComplexity() const;
1371 void NarrowSearchSpaceUsingHeuristics();
1373 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1375 SmallVectorImpl<const Formula *> &Workspace,
1376 const Cost &CurCost,
1377 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1378 DenseSet<const SCEV *> &VisitedRegs) const;
1379 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1381 BasicBlock::iterator
1382 HoistInsertPosition(BasicBlock::iterator IP,
1383 const SmallVectorImpl<Instruction *> &Inputs) const;
1384 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1386 const LSRUse &LU) const;
1388 Value *Expand(const LSRFixup &LF,
1390 BasicBlock::iterator IP,
1391 SCEVExpander &Rewriter,
1392 SmallVectorImpl<WeakVH> &DeadInsts) const;
1393 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1395 SCEVExpander &Rewriter,
1396 SmallVectorImpl<WeakVH> &DeadInsts,
1398 void Rewrite(const LSRFixup &LF,
1400 SCEVExpander &Rewriter,
1401 SmallVectorImpl<WeakVH> &DeadInsts,
1403 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1406 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1408 bool getChanged() const { return Changed; }
1410 void print_factors_and_types(raw_ostream &OS) const;
1411 void print_fixups(raw_ostream &OS) const;
1412 void print_uses(raw_ostream &OS) const;
1413 void print(raw_ostream &OS) const;
1419 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1420 /// inside the loop then try to eliminate the cast operation.
1421 void LSRInstance::OptimizeShadowIV() {
1422 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1423 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1426 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1427 UI != E; /* empty */) {
1428 IVUsers::const_iterator CandidateUI = UI;
1430 Instruction *ShadowUse = CandidateUI->getUser();
1431 const Type *DestTy = NULL;
1433 /* If shadow use is a int->float cast then insert a second IV
1434 to eliminate this cast.
1436 for (unsigned i = 0; i < n; ++i)
1442 for (unsigned i = 0; i < n; ++i, ++d)
1445 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1446 DestTy = UCast->getDestTy();
1447 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1448 DestTy = SCast->getDestTy();
1449 if (!DestTy) continue;
1452 // If target does not support DestTy natively then do not apply
1453 // this transformation.
1454 EVT DVT = TLI->getValueType(DestTy);
1455 if (!TLI->isTypeLegal(DVT)) continue;
1458 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1460 if (PH->getNumIncomingValues() != 2) continue;
1462 const Type *SrcTy = PH->getType();
1463 int Mantissa = DestTy->getFPMantissaWidth();
1464 if (Mantissa == -1) continue;
1465 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1468 unsigned Entry, Latch;
1469 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1477 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1478 if (!Init) continue;
1479 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1481 BinaryOperator *Incr =
1482 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1483 if (!Incr) continue;
1484 if (Incr->getOpcode() != Instruction::Add
1485 && Incr->getOpcode() != Instruction::Sub)
1488 /* Initialize new IV, double d = 0.0 in above example. */
1489 ConstantInt *C = NULL;
1490 if (Incr->getOperand(0) == PH)
1491 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1492 else if (Incr->getOperand(1) == PH)
1493 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1499 // Ignore negative constants, as the code below doesn't handle them
1500 // correctly. TODO: Remove this restriction.
1501 if (!C->getValue().isStrictlyPositive()) continue;
1503 /* Add new PHINode. */
1504 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1506 /* create new increment. '++d' in above example. */
1507 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1508 BinaryOperator *NewIncr =
1509 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1510 Instruction::FAdd : Instruction::FSub,
1511 NewPH, CFP, "IV.S.next.", Incr);
1513 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1514 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1516 /* Remove cast operation */
1517 ShadowUse->replaceAllUsesWith(NewPH);
1518 ShadowUse->eraseFromParent();
1524 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1525 /// set the IV user and stride information and return true, otherwise return
1527 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1528 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1529 if (UI->getUser() == Cond) {
1530 // NOTE: we could handle setcc instructions with multiple uses here, but
1531 // InstCombine does it as well for simple uses, it's not clear that it
1532 // occurs enough in real life to handle.
1539 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1540 /// a max computation.
1542 /// This is a narrow solution to a specific, but acute, problem. For loops
1548 /// } while (++i < n);
1550 /// the trip count isn't just 'n', because 'n' might not be positive. And
1551 /// unfortunately this can come up even for loops where the user didn't use
1552 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1553 /// will commonly be lowered like this:
1559 /// } while (++i < n);
1562 /// and then it's possible for subsequent optimization to obscure the if
1563 /// test in such a way that indvars can't find it.
1565 /// When indvars can't find the if test in loops like this, it creates a
1566 /// max expression, which allows it to give the loop a canonical
1567 /// induction variable:
1570 /// max = n < 1 ? 1 : n;
1573 /// } while (++i != max);
1575 /// Canonical induction variables are necessary because the loop passes
1576 /// are designed around them. The most obvious example of this is the
1577 /// LoopInfo analysis, which doesn't remember trip count values. It
1578 /// expects to be able to rediscover the trip count each time it is
1579 /// needed, and it does this using a simple analysis that only succeeds if
1580 /// the loop has a canonical induction variable.
1582 /// However, when it comes time to generate code, the maximum operation
1583 /// can be quite costly, especially if it's inside of an outer loop.
1585 /// This function solves this problem by detecting this type of loop and
1586 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1587 /// the instructions for the maximum computation.
1589 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1590 // Check that the loop matches the pattern we're looking for.
1591 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1592 Cond->getPredicate() != CmpInst::ICMP_NE)
1595 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1596 if (!Sel || !Sel->hasOneUse()) return Cond;
1598 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1599 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1601 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1603 // Add one to the backedge-taken count to get the trip count.
1604 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1605 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1607 // Check for a max calculation that matches the pattern. There's no check
1608 // for ICMP_ULE here because the comparison would be with zero, which
1609 // isn't interesting.
1610 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1611 const SCEVNAryExpr *Max = 0;
1612 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1613 Pred = ICmpInst::ICMP_SLE;
1615 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1616 Pred = ICmpInst::ICMP_SLT;
1618 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1619 Pred = ICmpInst::ICMP_ULT;
1626 // To handle a max with more than two operands, this optimization would
1627 // require additional checking and setup.
1628 if (Max->getNumOperands() != 2)
1631 const SCEV *MaxLHS = Max->getOperand(0);
1632 const SCEV *MaxRHS = Max->getOperand(1);
1634 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1635 // for a comparison with 1. For <= and >=, a comparison with zero.
1637 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1640 // Check the relevant induction variable for conformance to
1642 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1643 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1644 if (!AR || !AR->isAffine() ||
1645 AR->getStart() != One ||
1646 AR->getStepRecurrence(SE) != One)
1649 assert(AR->getLoop() == L &&
1650 "Loop condition operand is an addrec in a different loop!");
1652 // Check the right operand of the select, and remember it, as it will
1653 // be used in the new comparison instruction.
1655 if (ICmpInst::isTrueWhenEqual(Pred)) {
1656 // Look for n+1, and grab n.
1657 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1658 if (isa<ConstantInt>(BO->getOperand(1)) &&
1659 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1660 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1661 NewRHS = BO->getOperand(0);
1662 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1663 if (isa<ConstantInt>(BO->getOperand(1)) &&
1664 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1665 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1666 NewRHS = BO->getOperand(0);
1669 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1670 NewRHS = Sel->getOperand(1);
1671 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1672 NewRHS = Sel->getOperand(2);
1673 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1674 NewRHS = SU->getValue();
1676 // Max doesn't match expected pattern.
1679 // Determine the new comparison opcode. It may be signed or unsigned,
1680 // and the original comparison may be either equality or inequality.
1681 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1682 Pred = CmpInst::getInversePredicate(Pred);
1684 // Ok, everything looks ok to change the condition into an SLT or SGE and
1685 // delete the max calculation.
1687 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1689 // Delete the max calculation instructions.
1690 Cond->replaceAllUsesWith(NewCond);
1691 CondUse->setUser(NewCond);
1692 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1693 Cond->eraseFromParent();
1694 Sel->eraseFromParent();
1695 if (Cmp->use_empty())
1696 Cmp->eraseFromParent();
1700 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1701 /// postinc iv when possible.
1703 LSRInstance::OptimizeLoopTermCond() {
1704 SmallPtrSet<Instruction *, 4> PostIncs;
1706 BasicBlock *LatchBlock = L->getLoopLatch();
1707 SmallVector<BasicBlock*, 8> ExitingBlocks;
1708 L->getExitingBlocks(ExitingBlocks);
1710 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1711 BasicBlock *ExitingBlock = ExitingBlocks[i];
1713 // Get the terminating condition for the loop if possible. If we
1714 // can, we want to change it to use a post-incremented version of its
1715 // induction variable, to allow coalescing the live ranges for the IV into
1716 // one register value.
1718 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1721 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1722 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1725 // Search IVUsesByStride to find Cond's IVUse if there is one.
1726 IVStrideUse *CondUse = 0;
1727 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1728 if (!FindIVUserForCond(Cond, CondUse))
1731 // If the trip count is computed in terms of a max (due to ScalarEvolution
1732 // being unable to find a sufficient guard, for example), change the loop
1733 // comparison to use SLT or ULT instead of NE.
1734 // One consequence of doing this now is that it disrupts the count-down
1735 // optimization. That's not always a bad thing though, because in such
1736 // cases it may still be worthwhile to avoid a max.
1737 Cond = OptimizeMax(Cond, CondUse);
1739 // If this exiting block dominates the latch block, it may also use
1740 // the post-inc value if it won't be shared with other uses.
1741 // Check for dominance.
1742 if (!DT.dominates(ExitingBlock, LatchBlock))
1745 // Conservatively avoid trying to use the post-inc value in non-latch
1746 // exits if there may be pre-inc users in intervening blocks.
1747 if (LatchBlock != ExitingBlock)
1748 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1749 // Test if the use is reachable from the exiting block. This dominator
1750 // query is a conservative approximation of reachability.
1751 if (&*UI != CondUse &&
1752 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1753 // Conservatively assume there may be reuse if the quotient of their
1754 // strides could be a legal scale.
1755 const SCEV *A = IU.getStride(*CondUse, L);
1756 const SCEV *B = IU.getStride(*UI, L);
1757 if (!A || !B) continue;
1758 if (SE.getTypeSizeInBits(A->getType()) !=
1759 SE.getTypeSizeInBits(B->getType())) {
1760 if (SE.getTypeSizeInBits(A->getType()) >
1761 SE.getTypeSizeInBits(B->getType()))
1762 B = SE.getSignExtendExpr(B, A->getType());
1764 A = SE.getSignExtendExpr(A, B->getType());
1766 if (const SCEVConstant *D =
1767 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1768 const ConstantInt *C = D->getValue();
1769 // Stride of one or negative one can have reuse with non-addresses.
1770 if (C->isOne() || C->isAllOnesValue())
1771 goto decline_post_inc;
1772 // Avoid weird situations.
1773 if (C->getValue().getMinSignedBits() >= 64 ||
1774 C->getValue().isMinSignedValue())
1775 goto decline_post_inc;
1776 // Without TLI, assume that any stride might be valid, and so any
1777 // use might be shared.
1779 goto decline_post_inc;
1780 // Check for possible scaled-address reuse.
1781 const Type *AccessTy = getAccessType(UI->getUser());
1782 TargetLowering::AddrMode AM;
1783 AM.Scale = C->getSExtValue();
1784 if (TLI->isLegalAddressingMode(AM, AccessTy))
1785 goto decline_post_inc;
1786 AM.Scale = -AM.Scale;
1787 if (TLI->isLegalAddressingMode(AM, AccessTy))
1788 goto decline_post_inc;
1792 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1795 // It's possible for the setcc instruction to be anywhere in the loop, and
1796 // possible for it to have multiple users. If it is not immediately before
1797 // the exiting block branch, move it.
1798 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1799 if (Cond->hasOneUse()) {
1800 Cond->moveBefore(TermBr);
1802 // Clone the terminating condition and insert into the loopend.
1803 ICmpInst *OldCond = Cond;
1804 Cond = cast<ICmpInst>(Cond->clone());
1805 Cond->setName(L->getHeader()->getName() + ".termcond");
1806 ExitingBlock->getInstList().insert(TermBr, Cond);
1808 // Clone the IVUse, as the old use still exists!
1809 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1810 TermBr->replaceUsesOfWith(OldCond, Cond);
1814 // If we get to here, we know that we can transform the setcc instruction to
1815 // use the post-incremented version of the IV, allowing us to coalesce the
1816 // live ranges for the IV correctly.
1817 CondUse->transformToPostInc(L);
1820 PostIncs.insert(Cond);
1824 // Determine an insertion point for the loop induction variable increment. It
1825 // must dominate all the post-inc comparisons we just set up, and it must
1826 // dominate the loop latch edge.
1827 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1828 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1829 E = PostIncs.end(); I != E; ++I) {
1831 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1833 if (BB == (*I)->getParent())
1834 IVIncInsertPos = *I;
1835 else if (BB != IVIncInsertPos->getParent())
1836 IVIncInsertPos = BB->getTerminator();
1840 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1841 /// at the given offset and other details. If so, update the use and
1844 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1845 LSRUse::KindType Kind, const Type *AccessTy) {
1846 int64_t NewMinOffset = LU.MinOffset;
1847 int64_t NewMaxOffset = LU.MaxOffset;
1848 const Type *NewAccessTy = AccessTy;
1850 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1851 // something conservative, however this can pessimize in the case that one of
1852 // the uses will have all its uses outside the loop, for example.
1853 if (LU.Kind != Kind)
1855 // Conservatively assume HasBaseReg is true for now.
1856 if (NewOffset < LU.MinOffset) {
1857 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1858 Kind, AccessTy, TLI))
1860 NewMinOffset = NewOffset;
1861 } else if (NewOffset > LU.MaxOffset) {
1862 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1863 Kind, AccessTy, TLI))
1865 NewMaxOffset = NewOffset;
1867 // Check for a mismatched access type, and fall back conservatively as needed.
1868 // TODO: Be less conservative when the type is similar and can use the same
1869 // addressing modes.
1870 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1871 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1874 LU.MinOffset = NewMinOffset;
1875 LU.MaxOffset = NewMaxOffset;
1876 LU.AccessTy = NewAccessTy;
1877 if (NewOffset != LU.Offsets.back())
1878 LU.Offsets.push_back(NewOffset);
1882 /// getUse - Return an LSRUse index and an offset value for a fixup which
1883 /// needs the given expression, with the given kind and optional access type.
1884 /// Either reuse an existing use or create a new one, as needed.
1885 std::pair<size_t, int64_t>
1886 LSRInstance::getUse(const SCEV *&Expr,
1887 LSRUse::KindType Kind, const Type *AccessTy) {
1888 const SCEV *Copy = Expr;
1889 int64_t Offset = ExtractImmediate(Expr, SE);
1891 // Basic uses can't accept any offset, for example.
1892 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1897 std::pair<UseMapTy::iterator, bool> P =
1898 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1900 // A use already existed with this base.
1901 size_t LUIdx = P.first->second;
1902 LSRUse &LU = Uses[LUIdx];
1903 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1905 return std::make_pair(LUIdx, Offset);
1908 // Create a new use.
1909 size_t LUIdx = Uses.size();
1910 P.first->second = LUIdx;
1911 Uses.push_back(LSRUse(Kind, AccessTy));
1912 LSRUse &LU = Uses[LUIdx];
1914 // We don't need to track redundant offsets, but we don't need to go out
1915 // of our way here to avoid them.
1916 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1917 LU.Offsets.push_back(Offset);
1919 LU.MinOffset = Offset;
1920 LU.MaxOffset = Offset;
1921 return std::make_pair(LUIdx, Offset);
1924 /// DeleteUse - Delete the given use from the Uses list.
1925 void LSRInstance::DeleteUse(LSRUse &LU) {
1926 if (&LU != &Uses.back())
1927 std::swap(LU, Uses.back());
1931 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1932 /// a formula that has the same registers as the given formula.
1934 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1935 const LSRUse &OrigLU) {
1936 // Search all uses for the formula. This could be more clever. Ignore
1937 // ICmpZero uses because they may contain formulae generated by
1938 // GenerateICmpZeroScales, in which case adding fixup offsets may
1940 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1941 LSRUse &LU = Uses[LUIdx];
1942 if (&LU != &OrigLU &&
1943 LU.Kind != LSRUse::ICmpZero &&
1944 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1945 LU.WidestFixupType == OrigLU.WidestFixupType &&
1946 LU.HasFormulaWithSameRegs(OrigF)) {
1947 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1948 E = LU.Formulae.end(); I != E; ++I) {
1949 const Formula &F = *I;
1950 if (F.BaseRegs == OrigF.BaseRegs &&
1951 F.ScaledReg == OrigF.ScaledReg &&
1952 F.AM.BaseGV == OrigF.AM.BaseGV &&
1953 F.AM.Scale == OrigF.AM.Scale &&
1955 if (F.AM.BaseOffs == 0)
1966 void LSRInstance::CollectInterestingTypesAndFactors() {
1967 SmallSetVector<const SCEV *, 4> Strides;
1969 // Collect interesting types and strides.
1970 SmallVector<const SCEV *, 4> Worklist;
1971 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1972 const SCEV *Expr = IU.getExpr(*UI);
1974 // Collect interesting types.
1975 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1977 // Add strides for mentioned loops.
1978 Worklist.push_back(Expr);
1980 const SCEV *S = Worklist.pop_back_val();
1981 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1982 Strides.insert(AR->getStepRecurrence(SE));
1983 Worklist.push_back(AR->getStart());
1984 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1985 Worklist.append(Add->op_begin(), Add->op_end());
1987 } while (!Worklist.empty());
1990 // Compute interesting factors from the set of interesting strides.
1991 for (SmallSetVector<const SCEV *, 4>::const_iterator
1992 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1993 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1994 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
1995 const SCEV *OldStride = *I;
1996 const SCEV *NewStride = *NewStrideIter;
1998 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1999 SE.getTypeSizeInBits(NewStride->getType())) {
2000 if (SE.getTypeSizeInBits(OldStride->getType()) >
2001 SE.getTypeSizeInBits(NewStride->getType()))
2002 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2004 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2006 if (const SCEVConstant *Factor =
2007 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2009 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2010 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2011 } else if (const SCEVConstant *Factor =
2012 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2015 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2016 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2020 // If all uses use the same type, don't bother looking for truncation-based
2022 if (Types.size() == 1)
2025 DEBUG(print_factors_and_types(dbgs()));
2028 void LSRInstance::CollectFixupsAndInitialFormulae() {
2029 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2031 LSRFixup &LF = getNewFixup();
2032 LF.UserInst = UI->getUser();
2033 LF.OperandValToReplace = UI->getOperandValToReplace();
2034 LF.PostIncLoops = UI->getPostIncLoops();
2036 LSRUse::KindType Kind = LSRUse::Basic;
2037 const Type *AccessTy = 0;
2038 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2039 Kind = LSRUse::Address;
2040 AccessTy = getAccessType(LF.UserInst);
2043 const SCEV *S = IU.getExpr(*UI);
2045 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2046 // (N - i == 0), and this allows (N - i) to be the expression that we work
2047 // with rather than just N or i, so we can consider the register
2048 // requirements for both N and i at the same time. Limiting this code to
2049 // equality icmps is not a problem because all interesting loops use
2050 // equality icmps, thanks to IndVarSimplify.
2051 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2052 if (CI->isEquality()) {
2053 // Swap the operands if needed to put the OperandValToReplace on the
2054 // left, for consistency.
2055 Value *NV = CI->getOperand(1);
2056 if (NV == LF.OperandValToReplace) {
2057 CI->setOperand(1, CI->getOperand(0));
2058 CI->setOperand(0, NV);
2059 NV = CI->getOperand(1);
2063 // x == y --> x - y == 0
2064 const SCEV *N = SE.getSCEV(NV);
2065 if (N->isLoopInvariant(L)) {
2066 Kind = LSRUse::ICmpZero;
2067 S = SE.getMinusSCEV(N, S);
2070 // -1 and the negations of all interesting strides (except the negation
2071 // of -1) are now also interesting.
2072 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2073 if (Factors[i] != -1)
2074 Factors.insert(-(uint64_t)Factors[i]);
2078 // Set up the initial formula for this use.
2079 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2081 LF.Offset = P.second;
2082 LSRUse &LU = Uses[LF.LUIdx];
2083 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2084 if (!LU.WidestFixupType ||
2085 SE.getTypeSizeInBits(LU.WidestFixupType) <
2086 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2087 LU.WidestFixupType = LF.OperandValToReplace->getType();
2089 // If this is the first use of this LSRUse, give it a formula.
2090 if (LU.Formulae.empty()) {
2091 InsertInitialFormula(S, LU, LF.LUIdx);
2092 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2096 DEBUG(print_fixups(dbgs()));
2099 /// InsertInitialFormula - Insert a formula for the given expression into
2100 /// the given use, separating out loop-variant portions from loop-invariant
2101 /// and loop-computable portions.
2103 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2105 F.InitialMatch(S, L, SE, DT);
2106 bool Inserted = InsertFormula(LU, LUIdx, F);
2107 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2110 /// InsertSupplementalFormula - Insert a simple single-register formula for
2111 /// the given expression into the given use.
2113 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2114 LSRUse &LU, size_t LUIdx) {
2116 F.BaseRegs.push_back(S);
2117 F.AM.HasBaseReg = true;
2118 bool Inserted = InsertFormula(LU, LUIdx, F);
2119 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2122 /// CountRegisters - Note which registers are used by the given formula,
2123 /// updating RegUses.
2124 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2126 RegUses.CountRegister(F.ScaledReg, LUIdx);
2127 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2128 E = F.BaseRegs.end(); I != E; ++I)
2129 RegUses.CountRegister(*I, LUIdx);
2132 /// InsertFormula - If the given formula has not yet been inserted, add it to
2133 /// the list, and return true. Return false otherwise.
2134 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2135 if (!LU.InsertFormula(F))
2138 CountRegisters(F, LUIdx);
2142 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2143 /// loop-invariant values which we're tracking. These other uses will pin these
2144 /// values in registers, making them less profitable for elimination.
2145 /// TODO: This currently misses non-constant addrec step registers.
2146 /// TODO: Should this give more weight to users inside the loop?
2148 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2149 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2150 SmallPtrSet<const SCEV *, 8> Inserted;
2152 while (!Worklist.empty()) {
2153 const SCEV *S = Worklist.pop_back_val();
2155 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2156 Worklist.append(N->op_begin(), N->op_end());
2157 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2158 Worklist.push_back(C->getOperand());
2159 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2160 Worklist.push_back(D->getLHS());
2161 Worklist.push_back(D->getRHS());
2162 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2163 if (!Inserted.insert(U)) continue;
2164 const Value *V = U->getValue();
2165 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2166 // Look for instructions defined outside the loop.
2167 if (L->contains(Inst)) continue;
2168 } else if (isa<UndefValue>(V))
2169 // Undef doesn't have a live range, so it doesn't matter.
2171 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2173 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2174 // Ignore non-instructions.
2177 // Ignore instructions in other functions (as can happen with
2179 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2181 // Ignore instructions not dominated by the loop.
2182 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2183 UserInst->getParent() :
2184 cast<PHINode>(UserInst)->getIncomingBlock(
2185 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2186 if (!DT.dominates(L->getHeader(), UseBB))
2188 // Ignore uses which are part of other SCEV expressions, to avoid
2189 // analyzing them multiple times.
2190 if (SE.isSCEVable(UserInst->getType())) {
2191 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2192 // If the user is a no-op, look through to its uses.
2193 if (!isa<SCEVUnknown>(UserS))
2197 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2201 // Ignore icmp instructions which are already being analyzed.
2202 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2203 unsigned OtherIdx = !UI.getOperandNo();
2204 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2205 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2209 LSRFixup &LF = getNewFixup();
2210 LF.UserInst = const_cast<Instruction *>(UserInst);
2211 LF.OperandValToReplace = UI.getUse();
2212 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2214 LF.Offset = P.second;
2215 LSRUse &LU = Uses[LF.LUIdx];
2216 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2217 if (!LU.WidestFixupType ||
2218 SE.getTypeSizeInBits(LU.WidestFixupType) <
2219 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2220 LU.WidestFixupType = LF.OperandValToReplace->getType();
2221 InsertSupplementalFormula(U, LU, LF.LUIdx);
2222 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2229 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2230 /// separate registers. If C is non-null, multiply each subexpression by C.
2231 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2232 SmallVectorImpl<const SCEV *> &Ops,
2234 ScalarEvolution &SE) {
2235 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2236 // Break out add operands.
2237 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2239 CollectSubexprs(*I, C, Ops, L, SE);
2241 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2242 // Split a non-zero base out of an addrec.
2243 if (!AR->getStart()->isZero()) {
2244 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2245 AR->getStepRecurrence(SE),
2248 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2251 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2252 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2253 if (Mul->getNumOperands() == 2)
2254 if (const SCEVConstant *Op0 =
2255 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2256 CollectSubexprs(Mul->getOperand(1),
2257 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2263 // Otherwise use the value itself, optionally with a scale applied.
2264 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2267 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2269 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2272 // Arbitrarily cap recursion to protect compile time.
2273 if (Depth >= 3) return;
2275 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2276 const SCEV *BaseReg = Base.BaseRegs[i];
2278 SmallVector<const SCEV *, 8> AddOps;
2279 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2281 if (AddOps.size() == 1) continue;
2283 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2284 JE = AddOps.end(); J != JE; ++J) {
2286 // Loop-variant "unknown" values are uninteresting; we won't be able to
2287 // do anything meaningful with them.
2288 if (isa<SCEVUnknown>(*J) && !(*J)->isLoopInvariant(L))
2291 // Don't pull a constant into a register if the constant could be folded
2292 // into an immediate field.
2293 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2294 Base.getNumRegs() > 1,
2295 LU.Kind, LU.AccessTy, TLI, SE))
2298 // Collect all operands except *J.
2299 SmallVector<const SCEV *, 8> InnerAddOps
2300 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2302 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2304 // Don't leave just a constant behind in a register if the constant could
2305 // be folded into an immediate field.
2306 if (InnerAddOps.size() == 1 &&
2307 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2308 Base.getNumRegs() > 1,
2309 LU.Kind, LU.AccessTy, TLI, SE))
2312 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2313 if (InnerSum->isZero())
2316 F.BaseRegs[i] = InnerSum;
2317 F.BaseRegs.push_back(*J);
2318 if (InsertFormula(LU, LUIdx, F))
2319 // If that formula hadn't been seen before, recurse to find more like
2321 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2326 /// GenerateCombinations - Generate a formula consisting of all of the
2327 /// loop-dominating registers added into a single register.
2328 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2330 // This method is only interesting on a plurality of registers.
2331 if (Base.BaseRegs.size() <= 1) return;
2335 SmallVector<const SCEV *, 4> Ops;
2336 for (SmallVectorImpl<const SCEV *>::const_iterator
2337 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2338 const SCEV *BaseReg = *I;
2339 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2340 !BaseReg->hasComputableLoopEvolution(L))
2341 Ops.push_back(BaseReg);
2343 F.BaseRegs.push_back(BaseReg);
2345 if (Ops.size() > 1) {
2346 const SCEV *Sum = SE.getAddExpr(Ops);
2347 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2348 // opportunity to fold something. For now, just ignore such cases
2349 // rather than proceed with zero in a register.
2350 if (!Sum->isZero()) {
2351 F.BaseRegs.push_back(Sum);
2352 (void)InsertFormula(LU, LUIdx, F);
2357 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2358 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2360 // We can't add a symbolic offset if the address already contains one.
2361 if (Base.AM.BaseGV) return;
2363 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2364 const SCEV *G = Base.BaseRegs[i];
2365 GlobalValue *GV = ExtractSymbol(G, SE);
2366 if (G->isZero() || !GV)
2370 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2371 LU.Kind, LU.AccessTy, TLI))
2374 (void)InsertFormula(LU, LUIdx, F);
2378 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2379 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2381 // TODO: For now, just add the min and max offset, because it usually isn't
2382 // worthwhile looking at everything inbetween.
2383 SmallVector<int64_t, 2> Worklist;
2384 Worklist.push_back(LU.MinOffset);
2385 if (LU.MaxOffset != LU.MinOffset)
2386 Worklist.push_back(LU.MaxOffset);
2388 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2389 const SCEV *G = Base.BaseRegs[i];
2391 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2392 E = Worklist.end(); I != E; ++I) {
2394 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2395 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2396 LU.Kind, LU.AccessTy, TLI)) {
2397 // Add the offset to the base register.
2398 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2399 // If it cancelled out, drop the base register, otherwise update it.
2400 if (NewG->isZero()) {
2401 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2402 F.BaseRegs.pop_back();
2404 F.BaseRegs[i] = NewG;
2406 (void)InsertFormula(LU, LUIdx, F);
2410 int64_t Imm = ExtractImmediate(G, SE);
2411 if (G->isZero() || Imm == 0)
2414 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2415 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2416 LU.Kind, LU.AccessTy, TLI))
2419 (void)InsertFormula(LU, LUIdx, F);
2423 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2424 /// the comparison. For example, x == y -> x*c == y*c.
2425 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2427 if (LU.Kind != LSRUse::ICmpZero) return;
2429 // Determine the integer type for the base formula.
2430 const Type *IntTy = Base.getType();
2432 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2434 // Don't do this if there is more than one offset.
2435 if (LU.MinOffset != LU.MaxOffset) return;
2437 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2439 // Check each interesting stride.
2440 for (SmallSetVector<int64_t, 8>::const_iterator
2441 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2442 int64_t Factor = *I;
2444 // Check that the multiplication doesn't overflow.
2445 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2447 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2448 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2451 // Check that multiplying with the use offset doesn't overflow.
2452 int64_t Offset = LU.MinOffset;
2453 if (Offset == INT64_MIN && Factor == -1)
2455 Offset = (uint64_t)Offset * Factor;
2456 if (Offset / Factor != LU.MinOffset)
2460 F.AM.BaseOffs = NewBaseOffs;
2462 // Check that this scale is legal.
2463 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2466 // Compensate for the use having MinOffset built into it.
2467 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2469 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2471 // Check that multiplying with each base register doesn't overflow.
2472 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2473 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2474 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2478 // Check that multiplying with the scaled register doesn't overflow.
2480 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2481 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2485 // If we make it here and it's legal, add it.
2486 (void)InsertFormula(LU, LUIdx, F);
2491 /// GenerateScales - Generate stride factor reuse formulae by making use of
2492 /// scaled-offset address modes, for example.
2493 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2494 // Determine the integer type for the base formula.
2495 const Type *IntTy = Base.getType();
2498 // If this Formula already has a scaled register, we can't add another one.
2499 if (Base.AM.Scale != 0) return;
2501 // Check each interesting stride.
2502 for (SmallSetVector<int64_t, 8>::const_iterator
2503 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2504 int64_t Factor = *I;
2506 Base.AM.Scale = Factor;
2507 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2508 // Check whether this scale is going to be legal.
2509 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2510 LU.Kind, LU.AccessTy, TLI)) {
2511 // As a special-case, handle special out-of-loop Basic users specially.
2512 // TODO: Reconsider this special case.
2513 if (LU.Kind == LSRUse::Basic &&
2514 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2515 LSRUse::Special, LU.AccessTy, TLI) &&
2516 LU.AllFixupsOutsideLoop)
2517 LU.Kind = LSRUse::Special;
2521 // For an ICmpZero, negating a solitary base register won't lead to
2523 if (LU.Kind == LSRUse::ICmpZero &&
2524 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2526 // For each addrec base reg, apply the scale, if possible.
2527 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2528 if (const SCEVAddRecExpr *AR =
2529 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2530 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2531 if (FactorS->isZero())
2533 // Divide out the factor, ignoring high bits, since we'll be
2534 // scaling the value back up in the end.
2535 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2536 // TODO: This could be optimized to avoid all the copying.
2538 F.ScaledReg = Quotient;
2539 F.DeleteBaseReg(F.BaseRegs[i]);
2540 (void)InsertFormula(LU, LUIdx, F);
2546 /// GenerateTruncates - Generate reuse formulae from different IV types.
2547 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2548 // This requires TargetLowering to tell us which truncates are free.
2551 // Don't bother truncating symbolic values.
2552 if (Base.AM.BaseGV) return;
2554 // Determine the integer type for the base formula.
2555 const Type *DstTy = Base.getType();
2557 DstTy = SE.getEffectiveSCEVType(DstTy);
2559 for (SmallSetVector<const Type *, 4>::const_iterator
2560 I = Types.begin(), E = Types.end(); I != E; ++I) {
2561 const Type *SrcTy = *I;
2562 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2565 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2566 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2567 JE = F.BaseRegs.end(); J != JE; ++J)
2568 *J = SE.getAnyExtendExpr(*J, SrcTy);
2570 // TODO: This assumes we've done basic processing on all uses and
2571 // have an idea what the register usage is.
2572 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2575 (void)InsertFormula(LU, LUIdx, F);
2582 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2583 /// defer modifications so that the search phase doesn't have to worry about
2584 /// the data structures moving underneath it.
2588 const SCEV *OrigReg;
2590 WorkItem(size_t LI, int64_t I, const SCEV *R)
2591 : LUIdx(LI), Imm(I), OrigReg(R) {}
2593 void print(raw_ostream &OS) const;
2599 void WorkItem::print(raw_ostream &OS) const {
2600 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2601 << " , add offset " << Imm;
2604 void WorkItem::dump() const {
2605 print(errs()); errs() << '\n';
2608 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2609 /// distance apart and try to form reuse opportunities between them.
2610 void LSRInstance::GenerateCrossUseConstantOffsets() {
2611 // Group the registers by their value without any added constant offset.
2612 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2613 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2615 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2616 SmallVector<const SCEV *, 8> Sequence;
2617 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2619 const SCEV *Reg = *I;
2620 int64_t Imm = ExtractImmediate(Reg, SE);
2621 std::pair<RegMapTy::iterator, bool> Pair =
2622 Map.insert(std::make_pair(Reg, ImmMapTy()));
2624 Sequence.push_back(Reg);
2625 Pair.first->second.insert(std::make_pair(Imm, *I));
2626 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2629 // Now examine each set of registers with the same base value. Build up
2630 // a list of work to do and do the work in a separate step so that we're
2631 // not adding formulae and register counts while we're searching.
2632 SmallVector<WorkItem, 32> WorkItems;
2633 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2634 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2635 E = Sequence.end(); I != E; ++I) {
2636 const SCEV *Reg = *I;
2637 const ImmMapTy &Imms = Map.find(Reg)->second;
2639 // It's not worthwhile looking for reuse if there's only one offset.
2640 if (Imms.size() == 1)
2643 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2644 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2646 dbgs() << ' ' << J->first;
2649 // Examine each offset.
2650 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2652 const SCEV *OrigReg = J->second;
2654 int64_t JImm = J->first;
2655 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2657 if (!isa<SCEVConstant>(OrigReg) &&
2658 UsedByIndicesMap[Reg].count() == 1) {
2659 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2663 // Conservatively examine offsets between this orig reg a few selected
2665 ImmMapTy::const_iterator OtherImms[] = {
2666 Imms.begin(), prior(Imms.end()),
2667 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2669 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2670 ImmMapTy::const_iterator M = OtherImms[i];
2671 if (M == J || M == JE) continue;
2673 // Compute the difference between the two.
2674 int64_t Imm = (uint64_t)JImm - M->first;
2675 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2676 LUIdx = UsedByIndices.find_next(LUIdx))
2677 // Make a memo of this use, offset, and register tuple.
2678 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2679 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2686 UsedByIndicesMap.clear();
2687 UniqueItems.clear();
2689 // Now iterate through the worklist and add new formulae.
2690 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2691 E = WorkItems.end(); I != E; ++I) {
2692 const WorkItem &WI = *I;
2693 size_t LUIdx = WI.LUIdx;
2694 LSRUse &LU = Uses[LUIdx];
2695 int64_t Imm = WI.Imm;
2696 const SCEV *OrigReg = WI.OrigReg;
2698 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2699 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2700 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2702 // TODO: Use a more targeted data structure.
2703 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2704 const Formula &F = LU.Formulae[L];
2705 // Use the immediate in the scaled register.
2706 if (F.ScaledReg == OrigReg) {
2707 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2708 Imm * (uint64_t)F.AM.Scale;
2709 // Don't create 50 + reg(-50).
2710 if (F.referencesReg(SE.getSCEV(
2711 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2714 NewF.AM.BaseOffs = Offs;
2715 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2716 LU.Kind, LU.AccessTy, TLI))
2718 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2720 // If the new scale is a constant in a register, and adding the constant
2721 // value to the immediate would produce a value closer to zero than the
2722 // immediate itself, then the formula isn't worthwhile.
2723 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2724 if (C->getValue()->getValue().isNegative() !=
2725 (NewF.AM.BaseOffs < 0) &&
2726 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2727 .ule(abs64(NewF.AM.BaseOffs)))
2731 (void)InsertFormula(LU, LUIdx, NewF);
2733 // Use the immediate in a base register.
2734 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2735 const SCEV *BaseReg = F.BaseRegs[N];
2736 if (BaseReg != OrigReg)
2739 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2740 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2741 LU.Kind, LU.AccessTy, TLI))
2743 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2745 // If the new formula has a constant in a register, and adding the
2746 // constant value to the immediate would produce a value closer to
2747 // zero than the immediate itself, then the formula isn't worthwhile.
2748 for (SmallVectorImpl<const SCEV *>::const_iterator
2749 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2751 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2752 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2753 abs64(NewF.AM.BaseOffs)) &&
2754 (C->getValue()->getValue() +
2755 NewF.AM.BaseOffs).countTrailingZeros() >=
2756 CountTrailingZeros_64(NewF.AM.BaseOffs))
2760 (void)InsertFormula(LU, LUIdx, NewF);
2769 /// GenerateAllReuseFormulae - Generate formulae for each use.
2771 LSRInstance::GenerateAllReuseFormulae() {
2772 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2773 // queries are more precise.
2774 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2775 LSRUse &LU = Uses[LUIdx];
2776 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2777 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2778 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2779 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2781 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2782 LSRUse &LU = Uses[LUIdx];
2783 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2784 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2785 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2786 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2787 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2788 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2789 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2790 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2792 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2793 LSRUse &LU = Uses[LUIdx];
2794 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2795 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2798 GenerateCrossUseConstantOffsets();
2801 /// If their are multiple formulae with the same set of registers used
2802 /// by other uses, pick the best one and delete the others.
2803 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2805 bool ChangedFormulae = false;
2808 // Collect the best formula for each unique set of shared registers. This
2809 // is reset for each use.
2810 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2812 BestFormulaeTy BestFormulae;
2814 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2815 LSRUse &LU = Uses[LUIdx];
2816 FormulaSorter Sorter(L, LU, SE, DT);
2817 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2820 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2821 FIdx != NumForms; ++FIdx) {
2822 Formula &F = LU.Formulae[FIdx];
2824 SmallVector<const SCEV *, 2> Key;
2825 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2826 JE = F.BaseRegs.end(); J != JE; ++J) {
2827 const SCEV *Reg = *J;
2828 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2832 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2833 Key.push_back(F.ScaledReg);
2834 // Unstable sort by host order ok, because this is only used for
2836 std::sort(Key.begin(), Key.end());
2838 std::pair<BestFormulaeTy::const_iterator, bool> P =
2839 BestFormulae.insert(std::make_pair(Key, FIdx));
2841 Formula &Best = LU.Formulae[P.first->second];
2842 if (Sorter.operator()(F, Best))
2844 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2846 " in favor of formula "; Best.print(dbgs());
2849 ChangedFormulae = true;
2851 LU.DeleteFormula(F);
2859 // Now that we've filtered out some formulae, recompute the Regs set.
2861 LU.RecomputeRegs(LUIdx, RegUses);
2863 // Reset this to prepare for the next use.
2864 BestFormulae.clear();
2867 DEBUG(if (ChangedFormulae) {
2869 "After filtering out undesirable candidates:\n";
2874 // This is a rough guess that seems to work fairly well.
2875 static const size_t ComplexityLimit = UINT16_MAX;
2877 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2878 /// solutions the solver might have to consider. It almost never considers
2879 /// this many solutions because it prune the search space, but the pruning
2880 /// isn't always sufficient.
2881 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2883 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2884 E = Uses.end(); I != E; ++I) {
2885 size_t FSize = I->Formulae.size();
2886 if (FSize >= ComplexityLimit) {
2887 Power = ComplexityLimit;
2891 if (Power >= ComplexityLimit)
2897 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2898 /// formulae to choose from, use some rough heuristics to prune down the number
2899 /// of formulae. This keeps the main solver from taking an extraordinary amount
2900 /// of time in some worst-case scenarios.
2901 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2902 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2903 DEBUG(dbgs() << "The search space is too complex.\n");
2905 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2906 "which use a superset of registers used by other "
2909 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2910 LSRUse &LU = Uses[LUIdx];
2912 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2913 Formula &F = LU.Formulae[i];
2914 // Look for a formula with a constant or GV in a register. If the use
2915 // also has a formula with that same value in an immediate field,
2916 // delete the one that uses a register.
2917 for (SmallVectorImpl<const SCEV *>::const_iterator
2918 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2919 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2921 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2922 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2923 (I - F.BaseRegs.begin()));
2924 if (LU.HasFormulaWithSameRegs(NewF)) {
2925 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2926 LU.DeleteFormula(F);
2932 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2933 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2936 NewF.AM.BaseGV = GV;
2937 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2938 (I - F.BaseRegs.begin()));
2939 if (LU.HasFormulaWithSameRegs(NewF)) {
2940 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2942 LU.DeleteFormula(F);
2953 LU.RecomputeRegs(LUIdx, RegUses);
2956 DEBUG(dbgs() << "After pre-selection:\n";
2957 print_uses(dbgs()));
2960 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2961 DEBUG(dbgs() << "The search space is too complex.\n");
2963 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2964 "separated by a constant offset will use the same "
2967 // This is especially useful for unrolled loops.
2969 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2970 LSRUse &LU = Uses[LUIdx];
2971 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2972 E = LU.Formulae.end(); I != E; ++I) {
2973 const Formula &F = *I;
2974 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2975 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2976 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2977 /*HasBaseReg=*/false,
2978 LU.Kind, LU.AccessTy)) {
2979 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2982 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2984 // Delete formulae from the new use which are no longer legal.
2986 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2987 Formula &F = LUThatHas->Formulae[i];
2988 if (!isLegalUse(F.AM,
2989 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2990 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2991 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2993 LUThatHas->DeleteFormula(F);
3000 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3002 // Update the relocs to reference the new use.
3003 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3004 E = Fixups.end(); I != E; ++I) {
3005 LSRFixup &Fixup = *I;
3006 if (Fixup.LUIdx == LUIdx) {
3007 Fixup.LUIdx = LUThatHas - &Uses.front();
3008 Fixup.Offset += F.AM.BaseOffs;
3009 DEBUG(dbgs() << "New fixup has offset "
3010 << Fixup.Offset << '\n');
3012 if (Fixup.LUIdx == NumUses-1)
3013 Fixup.LUIdx = LUIdx;
3016 // Delete the old use.
3027 DEBUG(dbgs() << "After pre-selection:\n";
3028 print_uses(dbgs()));
3031 // With all other options exhausted, loop until the system is simple
3032 // enough to handle.
3033 SmallPtrSet<const SCEV *, 4> Taken;
3034 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3035 // Ok, we have too many of formulae on our hands to conveniently handle.
3036 // Use a rough heuristic to thin out the list.
3037 DEBUG(dbgs() << "The search space is too complex.\n");
3039 // Pick the register which is used by the most LSRUses, which is likely
3040 // to be a good reuse register candidate.
3041 const SCEV *Best = 0;
3042 unsigned BestNum = 0;
3043 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3045 const SCEV *Reg = *I;
3046 if (Taken.count(Reg))
3051 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3052 if (Count > BestNum) {
3059 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3060 << " will yield profitable reuse.\n");
3063 // In any use with formulae which references this register, delete formulae
3064 // which don't reference it.
3065 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3066 LSRUse &LU = Uses[LUIdx];
3067 if (!LU.Regs.count(Best)) continue;
3070 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3071 Formula &F = LU.Formulae[i];
3072 if (!F.referencesReg(Best)) {
3073 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3074 LU.DeleteFormula(F);
3078 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3084 LU.RecomputeRegs(LUIdx, RegUses);
3087 DEBUG(dbgs() << "After pre-selection:\n";
3088 print_uses(dbgs()));
3092 /// SolveRecurse - This is the recursive solver.
3093 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3095 SmallVectorImpl<const Formula *> &Workspace,
3096 const Cost &CurCost,
3097 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3098 DenseSet<const SCEV *> &VisitedRegs) const {
3101 // - use more aggressive filtering
3102 // - sort the formula so that the most profitable solutions are found first
3103 // - sort the uses too
3105 // - don't compute a cost, and then compare. compare while computing a cost
3107 // - track register sets with SmallBitVector
3109 const LSRUse &LU = Uses[Workspace.size()];
3111 // If this use references any register that's already a part of the
3112 // in-progress solution, consider it a requirement that a formula must
3113 // reference that register in order to be considered. This prunes out
3114 // unprofitable searching.
3115 SmallSetVector<const SCEV *, 4> ReqRegs;
3116 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3117 E = CurRegs.end(); I != E; ++I)
3118 if (LU.Regs.count(*I))
3121 bool AnySatisfiedReqRegs = false;
3122 SmallPtrSet<const SCEV *, 16> NewRegs;
3125 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3126 E = LU.Formulae.end(); I != E; ++I) {
3127 const Formula &F = *I;
3129 // Ignore formulae which do not use any of the required registers.
3130 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3131 JE = ReqRegs.end(); J != JE; ++J) {
3132 const SCEV *Reg = *J;
3133 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3134 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3138 AnySatisfiedReqRegs = true;
3140 // Evaluate the cost of the current formula. If it's already worse than
3141 // the current best, prune the search at that point.
3144 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3145 if (NewCost < SolutionCost) {
3146 Workspace.push_back(&F);
3147 if (Workspace.size() != Uses.size()) {
3148 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3149 NewRegs, VisitedRegs);
3150 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3151 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3153 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3154 dbgs() << ". Regs:";
3155 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3156 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3157 dbgs() << ' ' << **I;
3160 SolutionCost = NewCost;
3161 Solution = Workspace;
3163 Workspace.pop_back();
3168 // If none of the formulae had all of the required registers, relax the
3169 // constraint so that we don't exclude all formulae.
3170 if (!AnySatisfiedReqRegs) {
3171 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3177 /// Solve - Choose one formula from each use. Return the results in the given
3178 /// Solution vector.
3179 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3180 SmallVector<const Formula *, 8> Workspace;
3182 SolutionCost.Loose();
3184 SmallPtrSet<const SCEV *, 16> CurRegs;
3185 DenseSet<const SCEV *> VisitedRegs;
3186 Workspace.reserve(Uses.size());
3188 // SolveRecurse does all the work.
3189 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3190 CurRegs, VisitedRegs);
3192 // Ok, we've now made all our decisions.
3193 DEBUG(dbgs() << "\n"
3194 "The chosen solution requires "; SolutionCost.print(dbgs());
3196 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3198 Uses[i].print(dbgs());
3201 Solution[i]->print(dbgs());
3205 assert(Solution.size() == Uses.size() && "Malformed solution!");
3208 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3209 /// the dominator tree far as we can go while still being dominated by the
3210 /// input positions. This helps canonicalize the insert position, which
3211 /// encourages sharing.
3212 BasicBlock::iterator
3213 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3214 const SmallVectorImpl<Instruction *> &Inputs)
3217 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3218 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3221 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3222 if (!Rung) return IP;
3223 Rung = Rung->getIDom();
3224 if (!Rung) return IP;
3225 IDom = Rung->getBlock();
3227 // Don't climb into a loop though.
3228 const Loop *IDomLoop = LI.getLoopFor(IDom);
3229 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3230 if (IDomDepth <= IPLoopDepth &&
3231 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3235 bool AllDominate = true;
3236 Instruction *BetterPos = 0;
3237 Instruction *Tentative = IDom->getTerminator();
3238 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3239 E = Inputs.end(); I != E; ++I) {
3240 Instruction *Inst = *I;
3241 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3242 AllDominate = false;
3245 // Attempt to find an insert position in the middle of the block,
3246 // instead of at the end, so that it can be used for other expansions.
3247 if (IDom == Inst->getParent() &&
3248 (!BetterPos || DT.dominates(BetterPos, Inst)))
3249 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3262 /// AdjustInsertPositionForExpand - Determine an input position which will be
3263 /// dominated by the operands and which will dominate the result.
3264 BasicBlock::iterator
3265 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3267 const LSRUse &LU) const {
3268 // Collect some instructions which must be dominated by the
3269 // expanding replacement. These must be dominated by any operands that
3270 // will be required in the expansion.
3271 SmallVector<Instruction *, 4> Inputs;
3272 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3273 Inputs.push_back(I);
3274 if (LU.Kind == LSRUse::ICmpZero)
3275 if (Instruction *I =
3276 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3277 Inputs.push_back(I);
3278 if (LF.PostIncLoops.count(L)) {
3279 if (LF.isUseFullyOutsideLoop(L))
3280 Inputs.push_back(L->getLoopLatch()->getTerminator());
3282 Inputs.push_back(IVIncInsertPos);
3284 // The expansion must also be dominated by the increment positions of any
3285 // loops it for which it is using post-inc mode.
3286 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3287 E = LF.PostIncLoops.end(); I != E; ++I) {
3288 const Loop *PIL = *I;
3289 if (PIL == L) continue;
3291 // Be dominated by the loop exit.
3292 SmallVector<BasicBlock *, 4> ExitingBlocks;
3293 PIL->getExitingBlocks(ExitingBlocks);
3294 if (!ExitingBlocks.empty()) {
3295 BasicBlock *BB = ExitingBlocks[0];
3296 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3297 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3298 Inputs.push_back(BB->getTerminator());
3302 // Then, climb up the immediate dominator tree as far as we can go while
3303 // still being dominated by the input positions.
3304 IP = HoistInsertPosition(IP, Inputs);
3306 // Don't insert instructions before PHI nodes.
3307 while (isa<PHINode>(IP)) ++IP;
3309 // Ignore debug intrinsics.
3310 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3315 /// Expand - Emit instructions for the leading candidate expression for this
3316 /// LSRUse (this is called "expanding").
3317 Value *LSRInstance::Expand(const LSRFixup &LF,
3319 BasicBlock::iterator IP,
3320 SCEVExpander &Rewriter,
3321 SmallVectorImpl<WeakVH> &DeadInsts) const {
3322 const LSRUse &LU = Uses[LF.LUIdx];
3324 // Determine an input position which will be dominated by the operands and
3325 // which will dominate the result.
3326 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3328 // Inform the Rewriter if we have a post-increment use, so that it can
3329 // perform an advantageous expansion.
3330 Rewriter.setPostInc(LF.PostIncLoops);
3332 // This is the type that the user actually needs.
3333 const Type *OpTy = LF.OperandValToReplace->getType();
3334 // This will be the type that we'll initially expand to.
3335 const Type *Ty = F.getType();
3337 // No type known; just expand directly to the ultimate type.
3339 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3340 // Expand directly to the ultimate type if it's the right size.
3342 // This is the type to do integer arithmetic in.
3343 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3345 // Build up a list of operands to add together to form the full base.
3346 SmallVector<const SCEV *, 8> Ops;
3348 // Expand the BaseRegs portion.
3349 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3350 E = F.BaseRegs.end(); I != E; ++I) {
3351 const SCEV *Reg = *I;
3352 assert(!Reg->isZero() && "Zero allocated in a base register!");
3354 // If we're expanding for a post-inc user, make the post-inc adjustment.
3355 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3356 Reg = TransformForPostIncUse(Denormalize, Reg,
3357 LF.UserInst, LF.OperandValToReplace,
3360 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3363 // Flush the operand list to suppress SCEVExpander hoisting.
3365 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3367 Ops.push_back(SE.getUnknown(FullV));
3370 // Expand the ScaledReg portion.
3371 Value *ICmpScaledV = 0;
3372 if (F.AM.Scale != 0) {
3373 const SCEV *ScaledS = F.ScaledReg;
3375 // If we're expanding for a post-inc user, make the post-inc adjustment.
3376 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3377 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3378 LF.UserInst, LF.OperandValToReplace,
3381 if (LU.Kind == LSRUse::ICmpZero) {
3382 // An interesting way of "folding" with an icmp is to use a negated
3383 // scale, which we'll implement by inserting it into the other operand
3385 assert(F.AM.Scale == -1 &&
3386 "The only scale supported by ICmpZero uses is -1!");
3387 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3389 // Otherwise just expand the scaled register and an explicit scale,
3390 // which is expected to be matched as part of the address.
3391 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3392 ScaledS = SE.getMulExpr(ScaledS,
3393 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3394 Ops.push_back(ScaledS);
3396 // Flush the operand list to suppress SCEVExpander hoisting.
3397 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3399 Ops.push_back(SE.getUnknown(FullV));
3403 // Expand the GV portion.
3405 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3407 // Flush the operand list to suppress SCEVExpander hoisting.
3408 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3410 Ops.push_back(SE.getUnknown(FullV));
3413 // Expand the immediate portion.
3414 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3416 if (LU.Kind == LSRUse::ICmpZero) {
3417 // The other interesting way of "folding" with an ICmpZero is to use a
3418 // negated immediate.
3420 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3422 Ops.push_back(SE.getUnknown(ICmpScaledV));
3423 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3426 // Just add the immediate values. These again are expected to be matched
3427 // as part of the address.
3428 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3432 // Emit instructions summing all the operands.
3433 const SCEV *FullS = Ops.empty() ?
3434 SE.getConstant(IntTy, 0) :
3436 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3438 // We're done expanding now, so reset the rewriter.
3439 Rewriter.clearPostInc();
3441 // An ICmpZero Formula represents an ICmp which we're handling as a
3442 // comparison against zero. Now that we've expanded an expression for that
3443 // form, update the ICmp's other operand.
3444 if (LU.Kind == LSRUse::ICmpZero) {
3445 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3446 DeadInsts.push_back(CI->getOperand(1));
3447 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3448 "a scale at the same time!");
3449 if (F.AM.Scale == -1) {
3450 if (ICmpScaledV->getType() != OpTy) {
3452 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3454 ICmpScaledV, OpTy, "tmp", CI);
3457 CI->setOperand(1, ICmpScaledV);
3459 assert(F.AM.Scale == 0 &&
3460 "ICmp does not support folding a global value and "
3461 "a scale at the same time!");
3462 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3464 if (C->getType() != OpTy)
3465 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3469 CI->setOperand(1, C);
3476 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3477 /// of their operands effectively happens in their predecessor blocks, so the
3478 /// expression may need to be expanded in multiple places.
3479 void LSRInstance::RewriteForPHI(PHINode *PN,
3482 SCEVExpander &Rewriter,
3483 SmallVectorImpl<WeakVH> &DeadInsts,
3485 DenseMap<BasicBlock *, Value *> Inserted;
3486 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3487 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3488 BasicBlock *BB = PN->getIncomingBlock(i);
3490 // If this is a critical edge, split the edge so that we do not insert
3491 // the code on all predecessor/successor paths. We do this unless this
3492 // is the canonical backedge for this loop, which complicates post-inc
3494 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3495 !isa<IndirectBrInst>(BB->getTerminator()) &&
3496 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3497 // Split the critical edge.
3498 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3500 // If PN is outside of the loop and BB is in the loop, we want to
3501 // move the block to be immediately before the PHI block, not
3502 // immediately after BB.
3503 if (L->contains(BB) && !L->contains(PN))
3504 NewBB->moveBefore(PN->getParent());
3506 // Splitting the edge can reduce the number of PHI entries we have.
3507 e = PN->getNumIncomingValues();
3509 i = PN->getBasicBlockIndex(BB);
3512 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3513 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3515 PN->setIncomingValue(i, Pair.first->second);
3517 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3519 // If this is reuse-by-noop-cast, insert the noop cast.
3520 const Type *OpTy = LF.OperandValToReplace->getType();
3521 if (FullV->getType() != OpTy)
3523 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3525 FullV, LF.OperandValToReplace->getType(),
3526 "tmp", BB->getTerminator());
3528 PN->setIncomingValue(i, FullV);
3529 Pair.first->second = FullV;
3534 /// Rewrite - Emit instructions for the leading candidate expression for this
3535 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3536 /// the newly expanded value.
3537 void LSRInstance::Rewrite(const LSRFixup &LF,
3539 SCEVExpander &Rewriter,
3540 SmallVectorImpl<WeakVH> &DeadInsts,
3542 // First, find an insertion point that dominates UserInst. For PHI nodes,
3543 // find the nearest block which dominates all the relevant uses.
3544 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3545 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3547 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3549 // If this is reuse-by-noop-cast, insert the noop cast.
3550 const Type *OpTy = LF.OperandValToReplace->getType();
3551 if (FullV->getType() != OpTy) {
3553 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3554 FullV, OpTy, "tmp", LF.UserInst);
3558 // Update the user. ICmpZero is handled specially here (for now) because
3559 // Expand may have updated one of the operands of the icmp already, and
3560 // its new value may happen to be equal to LF.OperandValToReplace, in
3561 // which case doing replaceUsesOfWith leads to replacing both operands
3562 // with the same value. TODO: Reorganize this.
3563 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3564 LF.UserInst->setOperand(0, FullV);
3566 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3569 DeadInsts.push_back(LF.OperandValToReplace);
3572 /// ImplementSolution - Rewrite all the fixup locations with new values,
3573 /// following the chosen solution.
3575 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3577 // Keep track of instructions we may have made dead, so that
3578 // we can remove them after we are done working.
3579 SmallVector<WeakVH, 16> DeadInsts;
3581 SCEVExpander Rewriter(SE);
3582 Rewriter.disableCanonicalMode();
3583 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3585 // Expand the new value definitions and update the users.
3586 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3587 E = Fixups.end(); I != E; ++I) {
3588 const LSRFixup &Fixup = *I;
3590 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3595 // Clean up after ourselves. This must be done before deleting any
3599 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3602 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3603 : IU(P->getAnalysis<IVUsers>()),
3604 SE(P->getAnalysis<ScalarEvolution>()),
3605 DT(P->getAnalysis<DominatorTree>()),
3606 LI(P->getAnalysis<LoopInfo>()),
3607 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3609 // If LoopSimplify form is not available, stay out of trouble.
3610 if (!L->isLoopSimplifyForm()) return;
3612 // If there's no interesting work to be done, bail early.
3613 if (IU.empty()) return;
3615 DEBUG(dbgs() << "\nLSR on loop ";
3616 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3619 // First, perform some low-level loop optimizations.
3621 OptimizeLoopTermCond();
3623 // Start collecting data and preparing for the solver.
3624 CollectInterestingTypesAndFactors();
3625 CollectFixupsAndInitialFormulae();
3626 CollectLoopInvariantFixupsAndFormulae();
3628 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3629 print_uses(dbgs()));
3631 // Now use the reuse data to generate a bunch of interesting ways
3632 // to formulate the values needed for the uses.
3633 GenerateAllReuseFormulae();
3635 DEBUG(dbgs() << "\n"
3636 "After generating reuse formulae:\n";
3637 print_uses(dbgs()));
3639 FilterOutUndesirableDedicatedRegisters();
3640 NarrowSearchSpaceUsingHeuristics();
3642 SmallVector<const Formula *, 8> Solution;
3645 // Release memory that is no longer needed.
3651 // Formulae should be legal.
3652 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3653 E = Uses.end(); I != E; ++I) {
3654 const LSRUse &LU = *I;
3655 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3656 JE = LU.Formulae.end(); J != JE; ++J)
3657 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3658 LU.Kind, LU.AccessTy, TLI) &&
3659 "Illegal formula generated!");
3663 // Now that we've decided what we want, make it so.
3664 ImplementSolution(Solution, P);
3667 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3668 if (Factors.empty() && Types.empty()) return;
3670 OS << "LSR has identified the following interesting factors and types: ";
3673 for (SmallSetVector<int64_t, 8>::const_iterator
3674 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3675 if (!First) OS << ", ";
3680 for (SmallSetVector<const Type *, 4>::const_iterator
3681 I = Types.begin(), E = Types.end(); I != E; ++I) {
3682 if (!First) OS << ", ";
3684 OS << '(' << **I << ')';
3689 void LSRInstance::print_fixups(raw_ostream &OS) const {
3690 OS << "LSR is examining the following fixup sites:\n";
3691 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3692 E = Fixups.end(); I != E; ++I) {
3699 void LSRInstance::print_uses(raw_ostream &OS) const {
3700 OS << "LSR is examining the following uses:\n";
3701 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3702 E = Uses.end(); I != E; ++I) {
3703 const LSRUse &LU = *I;
3707 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3708 JE = LU.Formulae.end(); J != JE; ++J) {
3716 void LSRInstance::print(raw_ostream &OS) const {
3717 print_factors_and_types(OS);
3722 void LSRInstance::dump() const {
3723 print(errs()); errs() << '\n';
3728 class LoopStrengthReduce : public LoopPass {
3729 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3730 /// transformation profitability.
3731 const TargetLowering *const TLI;
3734 static char ID; // Pass ID, replacement for typeid
3735 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3738 bool runOnLoop(Loop *L, LPPassManager &LPM);
3739 void getAnalysisUsage(AnalysisUsage &AU) const;
3744 char LoopStrengthReduce::ID = 0;
3745 INITIALIZE_PASS(LoopStrengthReduce, "loop-reduce",
3746 "Loop Strength Reduction", false, false);
3748 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3749 return new LoopStrengthReduce(TLI);
3752 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3753 : LoopPass(ID), TLI(tli) {}
3755 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3756 // We split critical edges, so we change the CFG. However, we do update
3757 // many analyses if they are around.
3758 AU.addPreservedID(LoopSimplifyID);
3759 AU.addPreserved("domfrontier");
3761 AU.addRequired<LoopInfo>();
3762 AU.addPreserved<LoopInfo>();
3763 AU.addRequiredID(LoopSimplifyID);
3764 AU.addRequired<DominatorTree>();
3765 AU.addPreserved<DominatorTree>();
3766 AU.addRequired<ScalarEvolution>();
3767 AU.addPreserved<ScalarEvolution>();
3768 AU.addRequired<IVUsers>();
3769 AU.addPreserved<IVUsers>();
3772 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3773 bool Changed = false;
3775 // Run the main LSR transformation.
3776 Changed |= LSRInstance(TLI, L, this).getChanged();
3778 // At this point, it is worth checking to see if any recurrence PHIs are also
3779 // dead, so that we can remove them as well.
3780 Changed |= DeleteDeadPHIs(L->getHeader());