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, size_t NewLUIdx);
117 void DropUse(size_t LUIdx);
119 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
121 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
125 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
126 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
127 iterator begin() { return RegSequence.begin(); }
128 iterator end() { return RegSequence.end(); }
129 const_iterator begin() const { return RegSequence.begin(); }
130 const_iterator end() const { return RegSequence.end(); }
136 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
137 std::pair<RegUsesTy::iterator, bool> Pair =
138 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
139 RegSortData &RSD = Pair.first->second;
141 RegSequence.push_back(Reg);
142 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
143 RSD.UsedByIndices.set(LUIdx);
147 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
148 RegUsesTy::iterator It = RegUsesMap.find(Reg);
149 assert(It != RegUsesMap.end());
150 RegSortData &RSD = It->second;
151 assert(RSD.UsedByIndices.size() > LUIdx);
152 RSD.UsedByIndices.reset(LUIdx);
155 /// DropUse - Clear out reference by use LUIdx, and prepare for use NewLUIdx
156 /// to be swapped into LUIdx's position.
158 RegUseTracker::DropUse(size_t LUIdx, size_t NewLUIdx) {
159 // Remove the use index from every register's use list.
160 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
162 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
163 UsedByIndices.resize(std::max(UsedByIndices.size(), NewLUIdx + 1));
164 if (LUIdx < UsedByIndices.size()) {
165 UsedByIndices[LUIdx] = UsedByIndices[NewLUIdx];
166 UsedByIndices.reset(NewLUIdx);
168 UsedByIndices.reset(LUIdx);
172 /// DropUse - Clear out reference by use LUIdx.
174 RegUseTracker::DropUse(size_t LUIdx) {
175 // Remove the use index from every register's use list.
176 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
178 I->second.UsedByIndices.reset(LUIdx);
182 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
183 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
184 if (I == RegUsesMap.end())
186 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
187 int i = UsedByIndices.find_first();
188 if (i == -1) return false;
189 if ((size_t)i != LUIdx) return true;
190 return UsedByIndices.find_next(i) != -1;
193 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
194 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
195 assert(I != RegUsesMap.end() && "Unknown register!");
196 return I->second.UsedByIndices;
199 void RegUseTracker::clear() {
206 /// Formula - This class holds information that describes a formula for
207 /// computing satisfying a use. It may include broken-out immediates and scaled
210 /// AM - This is used to represent complex addressing, as well as other kinds
211 /// of interesting uses.
212 TargetLowering::AddrMode AM;
214 /// BaseRegs - The list of "base" registers for this use. When this is
215 /// non-empty, AM.HasBaseReg should be set to true.
216 SmallVector<const SCEV *, 2> BaseRegs;
218 /// ScaledReg - The 'scaled' register for this use. This should be non-null
219 /// when AM.Scale is not zero.
220 const SCEV *ScaledReg;
222 Formula() : ScaledReg(0) {}
224 void InitialMatch(const SCEV *S, Loop *L,
225 ScalarEvolution &SE, DominatorTree &DT);
227 unsigned getNumRegs() const;
228 const Type *getType() const;
230 void DeleteBaseReg(const SCEV *&S);
232 bool referencesReg(const SCEV *S) const;
233 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
234 const RegUseTracker &RegUses) const;
236 void print(raw_ostream &OS) const;
242 /// DoInitialMatch - Recursion helper for InitialMatch.
243 static void DoInitialMatch(const SCEV *S, Loop *L,
244 SmallVectorImpl<const SCEV *> &Good,
245 SmallVectorImpl<const SCEV *> &Bad,
246 ScalarEvolution &SE, DominatorTree &DT) {
247 // Collect expressions which properly dominate the loop header.
248 if (S->properlyDominates(L->getHeader(), &DT)) {
253 // Look at add operands.
254 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
255 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
257 DoInitialMatch(*I, L, Good, Bad, SE, DT);
261 // Look at addrec operands.
262 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
263 if (!AR->getStart()->isZero()) {
264 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
265 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
266 AR->getStepRecurrence(SE),
268 L, Good, Bad, SE, DT);
272 // Handle a multiplication by -1 (negation) if it didn't fold.
273 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
274 if (Mul->getOperand(0)->isAllOnesValue()) {
275 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
276 const SCEV *NewMul = SE.getMulExpr(Ops);
278 SmallVector<const SCEV *, 4> MyGood;
279 SmallVector<const SCEV *, 4> MyBad;
280 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
281 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
282 SE.getEffectiveSCEVType(NewMul->getType())));
283 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
284 E = MyGood.end(); I != E; ++I)
285 Good.push_back(SE.getMulExpr(NegOne, *I));
286 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
287 E = MyBad.end(); I != E; ++I)
288 Bad.push_back(SE.getMulExpr(NegOne, *I));
292 // Ok, we can't do anything interesting. Just stuff the whole thing into a
293 // register and hope for the best.
297 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
298 /// attempting to keep all loop-invariant and loop-computable values in a
299 /// single base register.
300 void Formula::InitialMatch(const SCEV *S, Loop *L,
301 ScalarEvolution &SE, DominatorTree &DT) {
302 SmallVector<const SCEV *, 4> Good;
303 SmallVector<const SCEV *, 4> Bad;
304 DoInitialMatch(S, L, Good, Bad, SE, DT);
306 const SCEV *Sum = SE.getAddExpr(Good);
308 BaseRegs.push_back(Sum);
309 AM.HasBaseReg = true;
312 const SCEV *Sum = SE.getAddExpr(Bad);
314 BaseRegs.push_back(Sum);
315 AM.HasBaseReg = true;
319 /// getNumRegs - Return the total number of register operands used by this
320 /// formula. This does not include register uses implied by non-constant
322 unsigned Formula::getNumRegs() const {
323 return !!ScaledReg + BaseRegs.size();
326 /// getType - Return the type of this formula, if it has one, or null
327 /// otherwise. This type is meaningless except for the bit size.
328 const Type *Formula::getType() const {
329 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
330 ScaledReg ? ScaledReg->getType() :
331 AM.BaseGV ? AM.BaseGV->getType() :
335 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
336 void Formula::DeleteBaseReg(const SCEV *&S) {
337 if (&S != &BaseRegs.back())
338 std::swap(S, BaseRegs.back());
342 /// referencesReg - Test if this formula references the given register.
343 bool Formula::referencesReg(const SCEV *S) const {
344 return S == ScaledReg ||
345 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
348 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
349 /// which are used by uses other than the use with the given index.
350 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
351 const RegUseTracker &RegUses) const {
353 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
355 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
356 E = BaseRegs.end(); I != E; ++I)
357 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
362 void Formula::print(raw_ostream &OS) const {
365 if (!First) OS << " + "; else First = false;
366 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
368 if (AM.BaseOffs != 0) {
369 if (!First) OS << " + "; else First = false;
372 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
373 E = BaseRegs.end(); I != E; ++I) {
374 if (!First) OS << " + "; else First = false;
375 OS << "reg(" << **I << ')';
377 if (AM.HasBaseReg && BaseRegs.empty()) {
378 if (!First) OS << " + "; else First = false;
379 OS << "**error: HasBaseReg**";
380 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
381 if (!First) OS << " + "; else First = false;
382 OS << "**error: !HasBaseReg**";
385 if (!First) OS << " + "; else First = false;
386 OS << AM.Scale << "*reg(";
395 void Formula::dump() const {
396 print(errs()); errs() << '\n';
399 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
400 /// without changing its value.
401 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
403 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
404 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
407 /// isAddSExtable - Return true if the given add can be sign-extended
408 /// without changing its value.
409 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
411 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
412 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
415 /// isMulSExtable - Return true if the given mul can be sign-extended
416 /// without changing its value.
417 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
419 IntegerType::get(SE.getContext(),
420 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
421 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
424 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
425 /// and if the remainder is known to be zero, or null otherwise. If
426 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
427 /// to Y, ignoring that the multiplication may overflow, which is useful when
428 /// the result will be used in a context where the most significant bits are
430 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
432 bool IgnoreSignificantBits = false) {
433 // Handle the trivial case, which works for any SCEV type.
435 return SE.getConstant(LHS->getType(), 1);
437 // Handle a few RHS special cases.
438 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
440 const APInt &RA = RC->getValue()->getValue();
441 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
443 if (RA.isAllOnesValue())
444 return SE.getMulExpr(LHS, RC);
445 // Handle x /s 1 as x.
450 // Check for a division of a constant by a constant.
451 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
454 const APInt &LA = C->getValue()->getValue();
455 const APInt &RA = RC->getValue()->getValue();
456 if (LA.srem(RA) != 0)
458 return SE.getConstant(LA.sdiv(RA));
461 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
462 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
463 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
464 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
465 IgnoreSignificantBits);
467 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
468 IgnoreSignificantBits);
469 if (!Start) return 0;
470 return SE.getAddRecExpr(Start, Step, AR->getLoop());
475 // Distribute the sdiv over add operands, if the add doesn't overflow.
476 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
477 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
478 SmallVector<const SCEV *, 8> Ops;
479 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
481 const SCEV *Op = getExactSDiv(*I, RHS, SE,
482 IgnoreSignificantBits);
486 return SE.getAddExpr(Ops);
491 // Check for a multiply operand that we can pull RHS out of.
492 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
493 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
494 SmallVector<const SCEV *, 4> Ops;
496 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
500 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
501 IgnoreSignificantBits)) {
507 return Found ? SE.getMulExpr(Ops) : 0;
512 // Otherwise we don't know.
516 /// ExtractImmediate - If S involves the addition of a constant integer value,
517 /// return that integer value, and mutate S to point to a new SCEV with that
519 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
520 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
521 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
522 S = SE.getConstant(C->getType(), 0);
523 return C->getValue()->getSExtValue();
525 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
526 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
527 int64_t Result = ExtractImmediate(NewOps.front(), SE);
529 S = SE.getAddExpr(NewOps);
531 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
532 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
533 int64_t Result = ExtractImmediate(NewOps.front(), SE);
535 S = SE.getAddRecExpr(NewOps, AR->getLoop());
541 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
542 /// return that symbol, and mutate S to point to a new SCEV with that
544 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
545 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
546 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
547 S = SE.getConstant(GV->getType(), 0);
550 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
551 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
552 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
554 S = SE.getAddExpr(NewOps);
556 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
557 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
558 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
560 S = SE.getAddRecExpr(NewOps, AR->getLoop());
566 /// isAddressUse - Returns true if the specified instruction is using the
567 /// specified value as an address.
568 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
569 bool isAddress = isa<LoadInst>(Inst);
570 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
571 if (SI->getOperand(1) == OperandVal)
573 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
574 // Addressing modes can also be folded into prefetches and a variety
576 switch (II->getIntrinsicID()) {
578 case Intrinsic::prefetch:
579 case Intrinsic::x86_sse2_loadu_dq:
580 case Intrinsic::x86_sse2_loadu_pd:
581 case Intrinsic::x86_sse_loadu_ps:
582 case Intrinsic::x86_sse_storeu_ps:
583 case Intrinsic::x86_sse2_storeu_pd:
584 case Intrinsic::x86_sse2_storeu_dq:
585 case Intrinsic::x86_sse2_storel_dq:
586 if (II->getArgOperand(0) == OperandVal)
594 /// getAccessType - Return the type of the memory being accessed.
595 static const Type *getAccessType(const Instruction *Inst) {
596 const Type *AccessTy = Inst->getType();
597 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
598 AccessTy = SI->getOperand(0)->getType();
599 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
600 // Addressing modes can also be folded into prefetches and a variety
602 switch (II->getIntrinsicID()) {
604 case Intrinsic::x86_sse_storeu_ps:
605 case Intrinsic::x86_sse2_storeu_pd:
606 case Intrinsic::x86_sse2_storeu_dq:
607 case Intrinsic::x86_sse2_storel_dq:
608 AccessTy = II->getArgOperand(0)->getType();
613 // All pointers have the same requirements, so canonicalize them to an
614 // arbitrary pointer type to minimize variation.
615 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
616 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
617 PTy->getAddressSpace());
622 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
623 /// specified set are trivially dead, delete them and see if this makes any of
624 /// their operands subsequently dead.
626 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
627 bool Changed = false;
629 while (!DeadInsts.empty()) {
630 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
632 if (I == 0 || !isInstructionTriviallyDead(I))
635 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
636 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
639 DeadInsts.push_back(U);
642 I->eraseFromParent();
651 /// Cost - This class is used to measure and compare candidate formulae.
653 /// TODO: Some of these could be merged. Also, a lexical ordering
654 /// isn't always optimal.
658 unsigned NumBaseAdds;
664 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
667 unsigned getNumRegs() const { return NumRegs; }
669 bool operator<(const Cost &Other) const;
673 void RateFormula(const Formula &F,
674 SmallPtrSet<const SCEV *, 16> &Regs,
675 const DenseSet<const SCEV *> &VisitedRegs,
677 const SmallVectorImpl<int64_t> &Offsets,
678 ScalarEvolution &SE, DominatorTree &DT);
680 void print(raw_ostream &OS) const;
684 void RateRegister(const SCEV *Reg,
685 SmallPtrSet<const SCEV *, 16> &Regs,
687 ScalarEvolution &SE, DominatorTree &DT);
688 void RatePrimaryRegister(const SCEV *Reg,
689 SmallPtrSet<const SCEV *, 16> &Regs,
691 ScalarEvolution &SE, DominatorTree &DT);
696 /// RateRegister - Tally up interesting quantities from the given register.
697 void Cost::RateRegister(const SCEV *Reg,
698 SmallPtrSet<const SCEV *, 16> &Regs,
700 ScalarEvolution &SE, DominatorTree &DT) {
701 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
702 if (AR->getLoop() == L)
703 AddRecCost += 1; /// TODO: This should be a function of the stride.
705 // If this is an addrec for a loop that's already been visited by LSR,
706 // don't second-guess its addrec phi nodes. LSR isn't currently smart
707 // enough to reason about more than one loop at a time. Consider these
708 // registers free and leave them alone.
709 else if (L->contains(AR->getLoop()) ||
710 (!AR->getLoop()->contains(L) &&
711 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
712 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
713 PHINode *PN = dyn_cast<PHINode>(I); ++I)
714 if (SE.isSCEVable(PN->getType()) &&
715 (SE.getEffectiveSCEVType(PN->getType()) ==
716 SE.getEffectiveSCEVType(AR->getType())) &&
717 SE.getSCEV(PN) == AR)
720 // If this isn't one of the addrecs that the loop already has, it
721 // would require a costly new phi and add. TODO: This isn't
722 // precisely modeled right now.
724 if (!Regs.count(AR->getStart()))
725 RateRegister(AR->getStart(), Regs, L, SE, DT);
728 // Add the step value register, if it needs one.
729 // TODO: The non-affine case isn't precisely modeled here.
730 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
731 if (!Regs.count(AR->getStart()))
732 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
736 // Rough heuristic; favor registers which don't require extra setup
737 // instructions in the preheader.
738 if (!isa<SCEVUnknown>(Reg) &&
739 !isa<SCEVConstant>(Reg) &&
740 !(isa<SCEVAddRecExpr>(Reg) &&
741 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
742 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
746 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
748 void Cost::RatePrimaryRegister(const SCEV *Reg,
749 SmallPtrSet<const SCEV *, 16> &Regs,
751 ScalarEvolution &SE, DominatorTree &DT) {
752 if (Regs.insert(Reg))
753 RateRegister(Reg, Regs, L, SE, DT);
756 void Cost::RateFormula(const Formula &F,
757 SmallPtrSet<const SCEV *, 16> &Regs,
758 const DenseSet<const SCEV *> &VisitedRegs,
760 const SmallVectorImpl<int64_t> &Offsets,
761 ScalarEvolution &SE, DominatorTree &DT) {
762 // Tally up the registers.
763 if (const SCEV *ScaledReg = F.ScaledReg) {
764 if (VisitedRegs.count(ScaledReg)) {
768 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
770 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
771 E = F.BaseRegs.end(); I != E; ++I) {
772 const SCEV *BaseReg = *I;
773 if (VisitedRegs.count(BaseReg)) {
777 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
779 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
780 BaseReg->hasComputableLoopEvolution(L);
783 if (F.BaseRegs.size() > 1)
784 NumBaseAdds += F.BaseRegs.size() - 1;
786 // Tally up the non-zero immediates.
787 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
788 E = Offsets.end(); I != E; ++I) {
789 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
791 ImmCost += 64; // Handle symbolic values conservatively.
792 // TODO: This should probably be the pointer size.
793 else if (Offset != 0)
794 ImmCost += APInt(64, Offset, true).getMinSignedBits();
798 /// Loose - Set this cost to a loosing value.
808 /// operator< - Choose the lower cost.
809 bool Cost::operator<(const Cost &Other) const {
810 if (NumRegs != Other.NumRegs)
811 return NumRegs < Other.NumRegs;
812 if (AddRecCost != Other.AddRecCost)
813 return AddRecCost < Other.AddRecCost;
814 if (NumIVMuls != Other.NumIVMuls)
815 return NumIVMuls < Other.NumIVMuls;
816 if (NumBaseAdds != Other.NumBaseAdds)
817 return NumBaseAdds < Other.NumBaseAdds;
818 if (ImmCost != Other.ImmCost)
819 return ImmCost < Other.ImmCost;
820 if (SetupCost != Other.SetupCost)
821 return SetupCost < Other.SetupCost;
825 void Cost::print(raw_ostream &OS) const {
826 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
828 OS << ", with addrec cost " << AddRecCost;
830 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
831 if (NumBaseAdds != 0)
832 OS << ", plus " << NumBaseAdds << " base add"
833 << (NumBaseAdds == 1 ? "" : "s");
835 OS << ", plus " << ImmCost << " imm cost";
837 OS << ", plus " << SetupCost << " setup cost";
840 void Cost::dump() const {
841 print(errs()); errs() << '\n';
846 /// LSRFixup - An operand value in an instruction which is to be replaced
847 /// with some equivalent, possibly strength-reduced, replacement.
849 /// UserInst - The instruction which will be updated.
850 Instruction *UserInst;
852 /// OperandValToReplace - The operand of the instruction which will
853 /// be replaced. The operand may be used more than once; every instance
854 /// will be replaced.
855 Value *OperandValToReplace;
857 /// PostIncLoops - If this user is to use the post-incremented value of an
858 /// induction variable, this variable is non-null and holds the loop
859 /// associated with the induction variable.
860 PostIncLoopSet PostIncLoops;
862 /// LUIdx - The index of the LSRUse describing the expression which
863 /// this fixup needs, minus an offset (below).
866 /// Offset - A constant offset to be added to the LSRUse expression.
867 /// This allows multiple fixups to share the same LSRUse with different
868 /// offsets, for example in an unrolled loop.
871 bool isUseFullyOutsideLoop(const Loop *L) const;
875 void print(raw_ostream &OS) const;
882 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
884 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
885 /// value outside of the given loop.
886 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
887 // PHI nodes use their value in their incoming blocks.
888 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
889 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
890 if (PN->getIncomingValue(i) == OperandValToReplace &&
891 L->contains(PN->getIncomingBlock(i)))
896 return !L->contains(UserInst);
899 void LSRFixup::print(raw_ostream &OS) const {
901 // Store is common and interesting enough to be worth special-casing.
902 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
904 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
905 } else if (UserInst->getType()->isVoidTy())
906 OS << UserInst->getOpcodeName();
908 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
910 OS << ", OperandValToReplace=";
911 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
913 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
914 E = PostIncLoops.end(); I != E; ++I) {
915 OS << ", PostIncLoop=";
916 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
919 if (LUIdx != ~size_t(0))
920 OS << ", LUIdx=" << LUIdx;
923 OS << ", Offset=" << Offset;
926 void LSRFixup::dump() const {
927 print(errs()); errs() << '\n';
932 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
933 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
934 struct UniquifierDenseMapInfo {
935 static SmallVector<const SCEV *, 2> getEmptyKey() {
936 SmallVector<const SCEV *, 2> V;
937 V.push_back(reinterpret_cast<const SCEV *>(-1));
941 static SmallVector<const SCEV *, 2> getTombstoneKey() {
942 SmallVector<const SCEV *, 2> V;
943 V.push_back(reinterpret_cast<const SCEV *>(-2));
947 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
949 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
950 E = V.end(); I != E; ++I)
951 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
955 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
956 const SmallVector<const SCEV *, 2> &RHS) {
961 /// LSRUse - This class holds the state that LSR keeps for each use in
962 /// IVUsers, as well as uses invented by LSR itself. It includes information
963 /// about what kinds of things can be folded into the user, information about
964 /// the user itself, and information about how the use may be satisfied.
965 /// TODO: Represent multiple users of the same expression in common?
967 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
970 /// KindType - An enum for a kind of use, indicating what types of
971 /// scaled and immediate operands it might support.
973 Basic, ///< A normal use, with no folding.
974 Special, ///< A special case of basic, allowing -1 scales.
975 Address, ///< An address use; folding according to TargetLowering
976 ICmpZero ///< An equality icmp with both operands folded into one.
977 // TODO: Add a generic icmp too?
981 const Type *AccessTy;
983 SmallVector<int64_t, 8> Offsets;
987 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
988 /// LSRUse are outside of the loop, in which case some special-case heuristics
990 bool AllFixupsOutsideLoop;
992 /// WidestFixupType - This records the widest use type for any fixup using
993 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
994 /// max fixup widths to be equivalent, because the narrower one may be relying
995 /// on the implicit truncation to truncate away bogus bits.
996 const Type *WidestFixupType;
998 /// Formulae - A list of ways to build a value that can satisfy this user.
999 /// After the list is populated, one of these is selected heuristically and
1000 /// used to formulate a replacement for OperandValToReplace in UserInst.
1001 SmallVector<Formula, 12> Formulae;
1003 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1004 SmallPtrSet<const SCEV *, 4> Regs;
1006 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
1007 MinOffset(INT64_MAX),
1008 MaxOffset(INT64_MIN),
1009 AllFixupsOutsideLoop(true),
1010 WidestFixupType(0) {}
1012 bool HasFormulaWithSameRegs(const Formula &F) const;
1013 bool InsertFormula(const Formula &F);
1014 void DeleteFormula(Formula &F);
1015 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1017 void print(raw_ostream &OS) const;
1023 /// HasFormula - Test whether this use as a formula which has the same
1024 /// registers as the given formula.
1025 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1026 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1027 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1028 // Unstable sort by host order ok, because this is only used for uniquifying.
1029 std::sort(Key.begin(), Key.end());
1030 return Uniquifier.count(Key);
1033 /// InsertFormula - If the given formula has not yet been inserted, add it to
1034 /// the list, and return true. Return false otherwise.
1035 bool LSRUse::InsertFormula(const Formula &F) {
1036 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1037 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1038 // Unstable sort by host order ok, because this is only used for uniquifying.
1039 std::sort(Key.begin(), Key.end());
1041 if (!Uniquifier.insert(Key).second)
1044 // Using a register to hold the value of 0 is not profitable.
1045 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1046 "Zero allocated in a scaled register!");
1048 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1049 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1050 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1053 // Add the formula to the list.
1054 Formulae.push_back(F);
1056 // Record registers now being used by this use.
1057 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1058 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1063 /// DeleteFormula - Remove the given formula from this use's list.
1064 void LSRUse::DeleteFormula(Formula &F) {
1065 if (&F != &Formulae.back())
1066 std::swap(F, Formulae.back());
1067 Formulae.pop_back();
1068 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1071 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1072 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1073 // Now that we've filtered out some formulae, recompute the Regs set.
1074 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1076 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1077 E = Formulae.end(); I != E; ++I) {
1078 const Formula &F = *I;
1079 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1080 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1083 // Update the RegTracker.
1084 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1085 E = OldRegs.end(); I != E; ++I)
1086 if (!Regs.count(*I))
1087 RegUses.DropRegister(*I, LUIdx);
1090 void LSRUse::print(raw_ostream &OS) const {
1091 OS << "LSR Use: Kind=";
1093 case Basic: OS << "Basic"; break;
1094 case Special: OS << "Special"; break;
1095 case ICmpZero: OS << "ICmpZero"; break;
1097 OS << "Address of ";
1098 if (AccessTy->isPointerTy())
1099 OS << "pointer"; // the full pointer type could be really verbose
1104 OS << ", Offsets={";
1105 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1106 E = Offsets.end(); I != E; ++I) {
1108 if (llvm::next(I) != E)
1113 if (AllFixupsOutsideLoop)
1114 OS << ", all-fixups-outside-loop";
1116 if (WidestFixupType)
1117 OS << ", widest fixup type: " << *WidestFixupType;
1120 void LSRUse::dump() const {
1121 print(errs()); errs() << '\n';
1124 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1125 /// be completely folded into the user instruction at isel time. This includes
1126 /// address-mode folding and special icmp tricks.
1127 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1128 LSRUse::KindType Kind, const Type *AccessTy,
1129 const TargetLowering *TLI) {
1131 case LSRUse::Address:
1132 // If we have low-level target information, ask the target if it can
1133 // completely fold this address.
1134 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1136 // Otherwise, just guess that reg+reg addressing is legal.
1137 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1139 case LSRUse::ICmpZero:
1140 // There's not even a target hook for querying whether it would be legal to
1141 // fold a GV into an ICmp.
1145 // ICmp only has two operands; don't allow more than two non-trivial parts.
1146 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1149 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1150 // putting the scaled register in the other operand of the icmp.
1151 if (AM.Scale != 0 && AM.Scale != -1)
1154 // If we have low-level target information, ask the target if it can fold an
1155 // integer immediate on an icmp.
1156 if (AM.BaseOffs != 0) {
1157 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1164 // Only handle single-register values.
1165 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1167 case LSRUse::Special:
1168 // Only handle -1 scales, or no scale.
1169 return AM.Scale == 0 || AM.Scale == -1;
1175 static bool isLegalUse(TargetLowering::AddrMode AM,
1176 int64_t MinOffset, int64_t MaxOffset,
1177 LSRUse::KindType Kind, const Type *AccessTy,
1178 const TargetLowering *TLI) {
1179 // Check for overflow.
1180 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1183 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1184 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1185 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1186 // Check for overflow.
1187 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1190 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1191 return isLegalUse(AM, Kind, AccessTy, TLI);
1196 static bool isAlwaysFoldable(int64_t BaseOffs,
1197 GlobalValue *BaseGV,
1199 LSRUse::KindType Kind, const Type *AccessTy,
1200 const TargetLowering *TLI) {
1201 // Fast-path: zero is always foldable.
1202 if (BaseOffs == 0 && !BaseGV) return true;
1204 // Conservatively, create an address with an immediate and a
1205 // base and a scale.
1206 TargetLowering::AddrMode AM;
1207 AM.BaseOffs = BaseOffs;
1209 AM.HasBaseReg = HasBaseReg;
1210 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1212 // Canonicalize a scale of 1 to a base register if the formula doesn't
1213 // already have a base register.
1214 if (!AM.HasBaseReg && AM.Scale == 1) {
1216 AM.HasBaseReg = true;
1219 return isLegalUse(AM, Kind, AccessTy, TLI);
1222 static bool isAlwaysFoldable(const SCEV *S,
1223 int64_t MinOffset, int64_t MaxOffset,
1225 LSRUse::KindType Kind, const Type *AccessTy,
1226 const TargetLowering *TLI,
1227 ScalarEvolution &SE) {
1228 // Fast-path: zero is always foldable.
1229 if (S->isZero()) return true;
1231 // Conservatively, create an address with an immediate and a
1232 // base and a scale.
1233 int64_t BaseOffs = ExtractImmediate(S, SE);
1234 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1236 // If there's anything else involved, it's not foldable.
1237 if (!S->isZero()) return false;
1239 // Fast-path: zero is always foldable.
1240 if (BaseOffs == 0 && !BaseGV) return true;
1242 // Conservatively, create an address with an immediate and a
1243 // base and a scale.
1244 TargetLowering::AddrMode AM;
1245 AM.BaseOffs = BaseOffs;
1247 AM.HasBaseReg = HasBaseReg;
1248 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1250 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1255 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1256 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1257 struct UseMapDenseMapInfo {
1258 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1259 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1262 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1263 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1267 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1268 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1269 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1273 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1274 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1279 /// FormulaSorter - This class implements an ordering for formulae which sorts
1280 /// the by their standalone cost.
1281 class FormulaSorter {
1282 /// These two sets are kept empty, so that we compute standalone costs.
1283 DenseSet<const SCEV *> VisitedRegs;
1284 SmallPtrSet<const SCEV *, 16> Regs;
1287 ScalarEvolution &SE;
1291 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1292 : L(l), LU(&lu), SE(se), DT(dt) {}
1294 bool operator()(const Formula &A, const Formula &B) {
1296 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1299 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1301 return CostA < CostB;
1305 /// LSRInstance - This class holds state for the main loop strength reduction
1309 ScalarEvolution &SE;
1312 const TargetLowering *const TLI;
1316 /// IVIncInsertPos - This is the insert position that the current loop's
1317 /// induction variable increment should be placed. In simple loops, this is
1318 /// the latch block's terminator. But in more complicated cases, this is a
1319 /// position which will dominate all the in-loop post-increment users.
1320 Instruction *IVIncInsertPos;
1322 /// Factors - Interesting factors between use strides.
1323 SmallSetVector<int64_t, 8> Factors;
1325 /// Types - Interesting use types, to facilitate truncation reuse.
1326 SmallSetVector<const Type *, 4> Types;
1328 /// Fixups - The list of operands which are to be replaced.
1329 SmallVector<LSRFixup, 16> Fixups;
1331 /// Uses - The list of interesting uses.
1332 SmallVector<LSRUse, 16> Uses;
1334 /// RegUses - Track which uses use which register candidates.
1335 RegUseTracker RegUses;
1337 void OptimizeShadowIV();
1338 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1339 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1340 void OptimizeLoopTermCond();
1342 void CollectInterestingTypesAndFactors();
1343 void CollectFixupsAndInitialFormulae();
1345 LSRFixup &getNewFixup() {
1346 Fixups.push_back(LSRFixup());
1347 return Fixups.back();
1350 // Support for sharing of LSRUses between LSRFixups.
1351 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1353 UseMapDenseMapInfo> UseMapTy;
1356 bool reconcileNewOffset(LSRUse &LU,
1357 int64_t NewMinOffset, int64_t NewMaxOffset,
1359 LSRUse::KindType Kind, const Type *AccessTy);
1361 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1362 LSRUse::KindType Kind,
1363 const Type *AccessTy);
1365 void DeleteUse(LSRUse &LU);
1367 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU,
1368 int64_t &NewBaseOffs);
1371 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1372 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1373 void CountRegisters(const Formula &F, size_t LUIdx);
1374 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1376 void CollectLoopInvariantFixupsAndFormulae();
1378 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1379 unsigned Depth = 0);
1380 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1381 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1382 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1383 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1384 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1385 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1386 void GenerateCrossUseConstantOffsets();
1387 void GenerateAllReuseFormulae();
1389 void FilterOutUndesirableDedicatedRegisters();
1391 size_t EstimateSearchSpaceComplexity() const;
1392 void NarrowSearchSpaceByDetectingSupersets();
1393 void NarrowSearchSpaceByCollapsingUnrolledCode();
1394 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1395 void NarrowSearchSpaceByPickingWinnerRegs();
1396 void NarrowSearchSpaceUsingHeuristics();
1398 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1400 SmallVectorImpl<const Formula *> &Workspace,
1401 const Cost &CurCost,
1402 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1403 DenseSet<const SCEV *> &VisitedRegs) const;
1404 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1406 BasicBlock::iterator
1407 HoistInsertPosition(BasicBlock::iterator IP,
1408 const SmallVectorImpl<Instruction *> &Inputs) const;
1409 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1411 const LSRUse &LU) const;
1413 Value *Expand(const LSRFixup &LF,
1415 BasicBlock::iterator IP,
1416 SCEVExpander &Rewriter,
1417 SmallVectorImpl<WeakVH> &DeadInsts) const;
1418 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1420 SCEVExpander &Rewriter,
1421 SmallVectorImpl<WeakVH> &DeadInsts,
1423 void Rewrite(const LSRFixup &LF,
1425 SCEVExpander &Rewriter,
1426 SmallVectorImpl<WeakVH> &DeadInsts,
1428 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1431 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1433 bool getChanged() const { return Changed; }
1435 void print_factors_and_types(raw_ostream &OS) const;
1436 void print_fixups(raw_ostream &OS) const;
1437 void print_uses(raw_ostream &OS) const;
1438 void print(raw_ostream &OS) const;
1444 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1445 /// inside the loop then try to eliminate the cast operation.
1446 void LSRInstance::OptimizeShadowIV() {
1447 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1448 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1451 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1452 UI != E; /* empty */) {
1453 IVUsers::const_iterator CandidateUI = UI;
1455 Instruction *ShadowUse = CandidateUI->getUser();
1456 const Type *DestTy = NULL;
1458 /* If shadow use is a int->float cast then insert a second IV
1459 to eliminate this cast.
1461 for (unsigned i = 0; i < n; ++i)
1467 for (unsigned i = 0; i < n; ++i, ++d)
1470 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1471 DestTy = UCast->getDestTy();
1472 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1473 DestTy = SCast->getDestTy();
1474 if (!DestTy) continue;
1477 // If target does not support DestTy natively then do not apply
1478 // this transformation.
1479 EVT DVT = TLI->getValueType(DestTy);
1480 if (!TLI->isTypeLegal(DVT)) continue;
1483 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1485 if (PH->getNumIncomingValues() != 2) continue;
1487 const Type *SrcTy = PH->getType();
1488 int Mantissa = DestTy->getFPMantissaWidth();
1489 if (Mantissa == -1) continue;
1490 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1493 unsigned Entry, Latch;
1494 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1502 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1503 if (!Init) continue;
1504 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1506 BinaryOperator *Incr =
1507 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1508 if (!Incr) continue;
1509 if (Incr->getOpcode() != Instruction::Add
1510 && Incr->getOpcode() != Instruction::Sub)
1513 /* Initialize new IV, double d = 0.0 in above example. */
1514 ConstantInt *C = NULL;
1515 if (Incr->getOperand(0) == PH)
1516 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1517 else if (Incr->getOperand(1) == PH)
1518 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1524 // Ignore negative constants, as the code below doesn't handle them
1525 // correctly. TODO: Remove this restriction.
1526 if (!C->getValue().isStrictlyPositive()) continue;
1528 /* Add new PHINode. */
1529 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1531 /* create new increment. '++d' in above example. */
1532 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1533 BinaryOperator *NewIncr =
1534 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1535 Instruction::FAdd : Instruction::FSub,
1536 NewPH, CFP, "IV.S.next.", Incr);
1538 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1539 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1541 /* Remove cast operation */
1542 ShadowUse->replaceAllUsesWith(NewPH);
1543 ShadowUse->eraseFromParent();
1549 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1550 /// set the IV user and stride information and return true, otherwise return
1552 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1553 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1554 if (UI->getUser() == Cond) {
1555 // NOTE: we could handle setcc instructions with multiple uses here, but
1556 // InstCombine does it as well for simple uses, it's not clear that it
1557 // occurs enough in real life to handle.
1564 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1565 /// a max computation.
1567 /// This is a narrow solution to a specific, but acute, problem. For loops
1573 /// } while (++i < n);
1575 /// the trip count isn't just 'n', because 'n' might not be positive. And
1576 /// unfortunately this can come up even for loops where the user didn't use
1577 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1578 /// will commonly be lowered like this:
1584 /// } while (++i < n);
1587 /// and then it's possible for subsequent optimization to obscure the if
1588 /// test in such a way that indvars can't find it.
1590 /// When indvars can't find the if test in loops like this, it creates a
1591 /// max expression, which allows it to give the loop a canonical
1592 /// induction variable:
1595 /// max = n < 1 ? 1 : n;
1598 /// } while (++i != max);
1600 /// Canonical induction variables are necessary because the loop passes
1601 /// are designed around them. The most obvious example of this is the
1602 /// LoopInfo analysis, which doesn't remember trip count values. It
1603 /// expects to be able to rediscover the trip count each time it is
1604 /// needed, and it does this using a simple analysis that only succeeds if
1605 /// the loop has a canonical induction variable.
1607 /// However, when it comes time to generate code, the maximum operation
1608 /// can be quite costly, especially if it's inside of an outer loop.
1610 /// This function solves this problem by detecting this type of loop and
1611 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1612 /// the instructions for the maximum computation.
1614 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1615 // Check that the loop matches the pattern we're looking for.
1616 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1617 Cond->getPredicate() != CmpInst::ICMP_NE)
1620 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1621 if (!Sel || !Sel->hasOneUse()) return Cond;
1623 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1624 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1626 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1628 // Add one to the backedge-taken count to get the trip count.
1629 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1630 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1632 // Check for a max calculation that matches the pattern. There's no check
1633 // for ICMP_ULE here because the comparison would be with zero, which
1634 // isn't interesting.
1635 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1636 const SCEVNAryExpr *Max = 0;
1637 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1638 Pred = ICmpInst::ICMP_SLE;
1640 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1641 Pred = ICmpInst::ICMP_SLT;
1643 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1644 Pred = ICmpInst::ICMP_ULT;
1651 // To handle a max with more than two operands, this optimization would
1652 // require additional checking and setup.
1653 if (Max->getNumOperands() != 2)
1656 const SCEV *MaxLHS = Max->getOperand(0);
1657 const SCEV *MaxRHS = Max->getOperand(1);
1659 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1660 // for a comparison with 1. For <= and >=, a comparison with zero.
1662 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1665 // Check the relevant induction variable for conformance to
1667 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1668 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1669 if (!AR || !AR->isAffine() ||
1670 AR->getStart() != One ||
1671 AR->getStepRecurrence(SE) != One)
1674 assert(AR->getLoop() == L &&
1675 "Loop condition operand is an addrec in a different loop!");
1677 // Check the right operand of the select, and remember it, as it will
1678 // be used in the new comparison instruction.
1680 if (ICmpInst::isTrueWhenEqual(Pred)) {
1681 // Look for n+1, and grab n.
1682 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1683 if (isa<ConstantInt>(BO->getOperand(1)) &&
1684 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1685 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1686 NewRHS = BO->getOperand(0);
1687 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1688 if (isa<ConstantInt>(BO->getOperand(1)) &&
1689 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1690 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1691 NewRHS = BO->getOperand(0);
1694 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1695 NewRHS = Sel->getOperand(1);
1696 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1697 NewRHS = Sel->getOperand(2);
1698 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1699 NewRHS = SU->getValue();
1701 // Max doesn't match expected pattern.
1704 // Determine the new comparison opcode. It may be signed or unsigned,
1705 // and the original comparison may be either equality or inequality.
1706 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1707 Pred = CmpInst::getInversePredicate(Pred);
1709 // Ok, everything looks ok to change the condition into an SLT or SGE and
1710 // delete the max calculation.
1712 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1714 // Delete the max calculation instructions.
1715 Cond->replaceAllUsesWith(NewCond);
1716 CondUse->setUser(NewCond);
1717 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1718 Cond->eraseFromParent();
1719 Sel->eraseFromParent();
1720 if (Cmp->use_empty())
1721 Cmp->eraseFromParent();
1725 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1726 /// postinc iv when possible.
1728 LSRInstance::OptimizeLoopTermCond() {
1729 SmallPtrSet<Instruction *, 4> PostIncs;
1731 BasicBlock *LatchBlock = L->getLoopLatch();
1732 SmallVector<BasicBlock*, 8> ExitingBlocks;
1733 L->getExitingBlocks(ExitingBlocks);
1735 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1736 BasicBlock *ExitingBlock = ExitingBlocks[i];
1738 // Get the terminating condition for the loop if possible. If we
1739 // can, we want to change it to use a post-incremented version of its
1740 // induction variable, to allow coalescing the live ranges for the IV into
1741 // one register value.
1743 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1746 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1747 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1750 // Search IVUsesByStride to find Cond's IVUse if there is one.
1751 IVStrideUse *CondUse = 0;
1752 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1753 if (!FindIVUserForCond(Cond, CondUse))
1756 // If the trip count is computed in terms of a max (due to ScalarEvolution
1757 // being unable to find a sufficient guard, for example), change the loop
1758 // comparison to use SLT or ULT instead of NE.
1759 // One consequence of doing this now is that it disrupts the count-down
1760 // optimization. That's not always a bad thing though, because in such
1761 // cases it may still be worthwhile to avoid a max.
1762 Cond = OptimizeMax(Cond, CondUse);
1764 // If this exiting block dominates the latch block, it may also use
1765 // the post-inc value if it won't be shared with other uses.
1766 // Check for dominance.
1767 if (!DT.dominates(ExitingBlock, LatchBlock))
1770 // Conservatively avoid trying to use the post-inc value in non-latch
1771 // exits if there may be pre-inc users in intervening blocks.
1772 if (LatchBlock != ExitingBlock)
1773 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1774 // Test if the use is reachable from the exiting block. This dominator
1775 // query is a conservative approximation of reachability.
1776 if (&*UI != CondUse &&
1777 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1778 // Conservatively assume there may be reuse if the quotient of their
1779 // strides could be a legal scale.
1780 const SCEV *A = IU.getStride(*CondUse, L);
1781 const SCEV *B = IU.getStride(*UI, L);
1782 if (!A || !B) continue;
1783 if (SE.getTypeSizeInBits(A->getType()) !=
1784 SE.getTypeSizeInBits(B->getType())) {
1785 if (SE.getTypeSizeInBits(A->getType()) >
1786 SE.getTypeSizeInBits(B->getType()))
1787 B = SE.getSignExtendExpr(B, A->getType());
1789 A = SE.getSignExtendExpr(A, B->getType());
1791 if (const SCEVConstant *D =
1792 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1793 const ConstantInt *C = D->getValue();
1794 // Stride of one or negative one can have reuse with non-addresses.
1795 if (C->isOne() || C->isAllOnesValue())
1796 goto decline_post_inc;
1797 // Avoid weird situations.
1798 if (C->getValue().getMinSignedBits() >= 64 ||
1799 C->getValue().isMinSignedValue())
1800 goto decline_post_inc;
1801 // Without TLI, assume that any stride might be valid, and so any
1802 // use might be shared.
1804 goto decline_post_inc;
1805 // Check for possible scaled-address reuse.
1806 const Type *AccessTy = getAccessType(UI->getUser());
1807 TargetLowering::AddrMode AM;
1808 AM.Scale = C->getSExtValue();
1809 if (TLI->isLegalAddressingMode(AM, AccessTy))
1810 goto decline_post_inc;
1811 AM.Scale = -AM.Scale;
1812 if (TLI->isLegalAddressingMode(AM, AccessTy))
1813 goto decline_post_inc;
1817 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1820 // It's possible for the setcc instruction to be anywhere in the loop, and
1821 // possible for it to have multiple users. If it is not immediately before
1822 // the exiting block branch, move it.
1823 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1824 if (Cond->hasOneUse()) {
1825 Cond->moveBefore(TermBr);
1827 // Clone the terminating condition and insert into the loopend.
1828 ICmpInst *OldCond = Cond;
1829 Cond = cast<ICmpInst>(Cond->clone());
1830 Cond->setName(L->getHeader()->getName() + ".termcond");
1831 ExitingBlock->getInstList().insert(TermBr, Cond);
1833 // Clone the IVUse, as the old use still exists!
1834 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1835 TermBr->replaceUsesOfWith(OldCond, Cond);
1839 // If we get to here, we know that we can transform the setcc instruction to
1840 // use the post-incremented version of the IV, allowing us to coalesce the
1841 // live ranges for the IV correctly.
1842 CondUse->transformToPostInc(L);
1845 PostIncs.insert(Cond);
1849 // Determine an insertion point for the loop induction variable increment. It
1850 // must dominate all the post-inc comparisons we just set up, and it must
1851 // dominate the loop latch edge.
1852 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1853 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1854 E = PostIncs.end(); I != E; ++I) {
1856 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1858 if (BB == (*I)->getParent())
1859 IVIncInsertPos = *I;
1860 else if (BB != IVIncInsertPos->getParent())
1861 IVIncInsertPos = BB->getTerminator();
1865 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1866 /// at the given offset and other details. If so, update the use and
1869 LSRInstance::reconcileNewOffset(LSRUse &LU,
1870 int64_t NewMinOffset, int64_t NewMaxOffset,
1872 LSRUse::KindType Kind, const Type *AccessTy) {
1873 int64_t ResultMinOffset = LU.MinOffset;
1874 int64_t ResultMaxOffset = LU.MaxOffset;
1875 const Type *ResultAccessTy = AccessTy;
1877 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1878 // something conservative, however this can pessimize in the case that one of
1879 // the uses will have all its uses outside the loop, for example.
1880 if (LU.Kind != Kind)
1882 // Conservatively assume HasBaseReg is true for now.
1883 if (NewMinOffset < LU.MinOffset) {
1884 if (!isAlwaysFoldable(LU.MaxOffset - NewMinOffset, 0, HasBaseReg,
1885 Kind, AccessTy, TLI))
1887 ResultMinOffset = NewMinOffset;
1888 } else if (NewMaxOffset > LU.MaxOffset) {
1889 if (!isAlwaysFoldable(NewMaxOffset - LU.MinOffset, 0, HasBaseReg,
1890 Kind, AccessTy, TLI))
1892 ResultMaxOffset = NewMaxOffset;
1894 // Check for a mismatched access type, and fall back conservatively as needed.
1895 // TODO: Be less conservative when the type is similar and can use the same
1896 // addressing modes.
1897 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1898 ResultAccessTy = Type::getVoidTy(AccessTy->getContext());
1901 LU.MinOffset = ResultMinOffset;
1902 LU.MaxOffset = ResultMaxOffset;
1903 LU.AccessTy = ResultAccessTy;
1907 /// getUse - Return an LSRUse index and an offset value for a fixup which
1908 /// needs the given expression, with the given kind and optional access type.
1909 /// Either reuse an existing use or create a new one, as needed.
1910 std::pair<size_t, int64_t>
1911 LSRInstance::getUse(const SCEV *&Expr,
1912 LSRUse::KindType Kind, const Type *AccessTy) {
1913 const SCEV *Copy = Expr;
1914 int64_t Offset = ExtractImmediate(Expr, SE);
1916 // Basic uses can't accept any offset, for example.
1917 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1922 std::pair<UseMapTy::iterator, bool> P =
1923 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1925 // A use already existed with this base.
1926 size_t LUIdx = P.first->second;
1927 LSRUse &LU = Uses[LUIdx];
1928 if (reconcileNewOffset(LU, Offset, Offset,
1929 /*HasBaseReg=*/true, Kind, AccessTy)) {
1930 LU.Offsets.push_back(Offset);
1932 return std::make_pair(LUIdx, Offset);
1936 // Create a new use.
1937 size_t LUIdx = Uses.size();
1938 P.first->second = LUIdx;
1939 Uses.push_back(LSRUse(Kind, AccessTy));
1940 LSRUse &LU = Uses[LUIdx];
1942 LU.Offsets.push_back(Offset);
1943 LU.MinOffset = Offset;
1944 LU.MaxOffset = Offset;
1945 return std::make_pair(LUIdx, Offset);
1948 /// DeleteUse - Delete the given use from the Uses list.
1949 void LSRInstance::DeleteUse(LSRUse &LU) {
1950 if (&LU != &Uses.back()) {
1951 std::swap(LU, Uses.back());
1952 RegUses.DropUse(&LU - Uses.begin(), Uses.size() - 1);
1954 RegUses.DropUse(&LU - Uses.begin());
1959 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1960 /// a formula that has the same registers as the given formula.
1962 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1963 const LSRUse &OrigLU,
1964 int64_t &NewBaseOffs) {
1965 // Search all uses for a formula similar to OrigF. This could be more clever.
1966 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1967 LSRUse &LU = Uses[LUIdx];
1968 // Check whether this use is close enough to OrigLU, to see whether it's
1969 // worthwhile looking through its formulae.
1970 // Ignore ICmpZero uses because they may contain formulae generated by
1971 // GenerateICmpZeroScales, in which case adding fixup offsets may
1973 if (&LU != &OrigLU &&
1974 LU.Kind != LSRUse::ICmpZero &&
1975 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1976 LU.WidestFixupType == OrigLU.WidestFixupType &&
1977 LU.HasFormulaWithSameRegs(OrigF)) {
1978 // Scan through this use's formulae.
1979 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1980 E = LU.Formulae.end(); I != E; ++I) {
1981 const Formula &F = *I;
1982 // Check to see if this formula has the same registers and symbols
1984 if (F.BaseRegs == OrigF.BaseRegs &&
1985 F.ScaledReg == OrigF.ScaledReg &&
1986 F.AM.BaseGV == OrigF.AM.BaseGV &&
1987 F.AM.Scale == OrigF.AM.Scale) {
1988 // Ok, all the registers and symbols matched. Check to see if the
1989 // immediate looks nicer than our old one.
1990 if (OrigF.AM.BaseOffs == INT64_MIN ||
1991 (F.AM.BaseOffs != INT64_MIN &&
1992 abs64(F.AM.BaseOffs) < abs64(OrigF.AM.BaseOffs))) {
1993 // Looks good. Take it.
1994 NewBaseOffs = F.AM.BaseOffs;
1997 // This is the formula where all the registers and symbols matched;
1998 // there aren't going to be any others. Since we declined it, we
1999 // can skip the rest of the formulae and procede to the next LSRUse.
2006 // Nothing looked good.
2010 void LSRInstance::CollectInterestingTypesAndFactors() {
2011 SmallSetVector<const SCEV *, 4> Strides;
2013 // Collect interesting types and strides.
2014 SmallVector<const SCEV *, 4> Worklist;
2015 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2016 const SCEV *Expr = IU.getExpr(*UI);
2018 // Collect interesting types.
2019 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2021 // Add strides for mentioned loops.
2022 Worklist.push_back(Expr);
2024 const SCEV *S = Worklist.pop_back_val();
2025 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2026 Strides.insert(AR->getStepRecurrence(SE));
2027 Worklist.push_back(AR->getStart());
2028 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2029 Worklist.append(Add->op_begin(), Add->op_end());
2031 } while (!Worklist.empty());
2034 // Compute interesting factors from the set of interesting strides.
2035 for (SmallSetVector<const SCEV *, 4>::const_iterator
2036 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2037 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2038 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2039 const SCEV *OldStride = *I;
2040 const SCEV *NewStride = *NewStrideIter;
2042 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2043 SE.getTypeSizeInBits(NewStride->getType())) {
2044 if (SE.getTypeSizeInBits(OldStride->getType()) >
2045 SE.getTypeSizeInBits(NewStride->getType()))
2046 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2048 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2050 if (const SCEVConstant *Factor =
2051 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2053 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2054 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2055 } else if (const SCEVConstant *Factor =
2056 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2059 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2060 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2064 // If all uses use the same type, don't bother looking for truncation-based
2066 if (Types.size() == 1)
2069 DEBUG(print_factors_and_types(dbgs()));
2072 void LSRInstance::CollectFixupsAndInitialFormulae() {
2073 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2075 LSRFixup &LF = getNewFixup();
2076 LF.UserInst = UI->getUser();
2077 LF.OperandValToReplace = UI->getOperandValToReplace();
2078 LF.PostIncLoops = UI->getPostIncLoops();
2080 LSRUse::KindType Kind = LSRUse::Basic;
2081 const Type *AccessTy = 0;
2082 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2083 Kind = LSRUse::Address;
2084 AccessTy = getAccessType(LF.UserInst);
2087 const SCEV *S = IU.getExpr(*UI);
2089 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2090 // (N - i == 0), and this allows (N - i) to be the expression that we work
2091 // with rather than just N or i, so we can consider the register
2092 // requirements for both N and i at the same time. Limiting this code to
2093 // equality icmps is not a problem because all interesting loops use
2094 // equality icmps, thanks to IndVarSimplify.
2095 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2096 if (CI->isEquality()) {
2097 // Swap the operands if needed to put the OperandValToReplace on the
2098 // left, for consistency.
2099 Value *NV = CI->getOperand(1);
2100 if (NV == LF.OperandValToReplace) {
2101 CI->setOperand(1, CI->getOperand(0));
2102 CI->setOperand(0, NV);
2103 NV = CI->getOperand(1);
2107 // x == y --> x - y == 0
2108 const SCEV *N = SE.getSCEV(NV);
2109 if (N->isLoopInvariant(L)) {
2110 Kind = LSRUse::ICmpZero;
2111 S = SE.getMinusSCEV(N, S);
2114 // -1 and the negations of all interesting strides (except the negation
2115 // of -1) are now also interesting.
2116 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2117 if (Factors[i] != -1)
2118 Factors.insert(-(uint64_t)Factors[i]);
2122 // Set up the initial formula for this use.
2123 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2125 LF.Offset = P.second;
2126 LSRUse &LU = Uses[LF.LUIdx];
2127 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2128 if (!LU.WidestFixupType ||
2129 SE.getTypeSizeInBits(LU.WidestFixupType) <
2130 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2131 LU.WidestFixupType = LF.OperandValToReplace->getType();
2133 // If this is the first use of this LSRUse, give it a formula.
2134 if (LU.Formulae.empty()) {
2135 InsertInitialFormula(S, LU, LF.LUIdx);
2136 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2140 DEBUG(print_fixups(dbgs()));
2143 /// InsertInitialFormula - Insert a formula for the given expression into
2144 /// the given use, separating out loop-variant portions from loop-invariant
2145 /// and loop-computable portions.
2147 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2149 F.InitialMatch(S, L, SE, DT);
2150 bool Inserted = InsertFormula(LU, LUIdx, F);
2151 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2154 /// InsertSupplementalFormula - Insert a simple single-register formula for
2155 /// the given expression into the given use.
2157 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2158 LSRUse &LU, size_t LUIdx) {
2160 F.BaseRegs.push_back(S);
2161 F.AM.HasBaseReg = true;
2162 bool Inserted = InsertFormula(LU, LUIdx, F);
2163 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2166 /// CountRegisters - Note which registers are used by the given formula,
2167 /// updating RegUses.
2168 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2170 RegUses.CountRegister(F.ScaledReg, LUIdx);
2171 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2172 E = F.BaseRegs.end(); I != E; ++I)
2173 RegUses.CountRegister(*I, LUIdx);
2176 /// InsertFormula - If the given formula has not yet been inserted, add it to
2177 /// the list, and return true. Return false otherwise.
2178 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2179 if (!LU.InsertFormula(F))
2182 CountRegisters(F, LUIdx);
2186 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2187 /// loop-invariant values which we're tracking. These other uses will pin these
2188 /// values in registers, making them less profitable for elimination.
2189 /// TODO: This currently misses non-constant addrec step registers.
2190 /// TODO: Should this give more weight to users inside the loop?
2192 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2193 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2194 SmallPtrSet<const SCEV *, 8> Inserted;
2196 while (!Worklist.empty()) {
2197 const SCEV *S = Worklist.pop_back_val();
2199 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2200 Worklist.append(N->op_begin(), N->op_end());
2201 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2202 Worklist.push_back(C->getOperand());
2203 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2204 Worklist.push_back(D->getLHS());
2205 Worklist.push_back(D->getRHS());
2206 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2207 if (!Inserted.insert(U)) continue;
2208 const Value *V = U->getValue();
2209 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2210 // Look for instructions defined outside the loop.
2211 if (L->contains(Inst)) continue;
2212 } else if (isa<UndefValue>(V))
2213 // Undef doesn't have a live range, so it doesn't matter.
2215 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2217 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2218 // Ignore non-instructions.
2221 // Ignore instructions in other functions (as can happen with
2223 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2225 // Ignore instructions not dominated by the loop.
2226 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2227 UserInst->getParent() :
2228 cast<PHINode>(UserInst)->getIncomingBlock(
2229 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2230 if (!DT.dominates(L->getHeader(), UseBB))
2232 // Ignore uses which are part of other SCEV expressions, to avoid
2233 // analyzing them multiple times.
2234 if (SE.isSCEVable(UserInst->getType())) {
2235 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2236 // If the user is a no-op, look through to its uses.
2237 if (!isa<SCEVUnknown>(UserS))
2241 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2245 // Ignore icmp instructions which are already being analyzed.
2246 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2247 unsigned OtherIdx = !UI.getOperandNo();
2248 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2249 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2253 LSRFixup &LF = getNewFixup();
2254 LF.UserInst = const_cast<Instruction *>(UserInst);
2255 LF.OperandValToReplace = UI.getUse();
2256 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2258 LF.Offset = P.second;
2259 LSRUse &LU = Uses[LF.LUIdx];
2260 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2261 if (!LU.WidestFixupType ||
2262 SE.getTypeSizeInBits(LU.WidestFixupType) <
2263 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2264 LU.WidestFixupType = LF.OperandValToReplace->getType();
2265 InsertSupplementalFormula(U, LU, LF.LUIdx);
2266 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2273 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2274 /// separate registers. If C is non-null, multiply each subexpression by C.
2275 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2276 SmallVectorImpl<const SCEV *> &Ops,
2278 ScalarEvolution &SE) {
2279 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2280 // Break out add operands.
2281 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2283 CollectSubexprs(*I, C, Ops, L, SE);
2285 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2286 // Split a non-zero base out of an addrec.
2287 if (!AR->getStart()->isZero()) {
2288 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2289 AR->getStepRecurrence(SE),
2292 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2295 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2296 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2297 if (Mul->getNumOperands() == 2)
2298 if (const SCEVConstant *Op0 =
2299 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2300 CollectSubexprs(Mul->getOperand(1),
2301 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2307 // Otherwise use the value itself, optionally with a scale applied.
2308 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2311 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2313 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2316 // Arbitrarily cap recursion to protect compile time.
2317 if (Depth >= 3) return;
2319 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2320 const SCEV *BaseReg = Base.BaseRegs[i];
2322 SmallVector<const SCEV *, 8> AddOps;
2323 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2325 if (AddOps.size() == 1) continue;
2327 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2328 JE = AddOps.end(); J != JE; ++J) {
2330 // Loop-variant "unknown" values are uninteresting; we won't be able to
2331 // do anything meaningful with them.
2332 if (isa<SCEVUnknown>(*J) && !(*J)->isLoopInvariant(L))
2335 // Don't pull a constant into a register if the constant could be folded
2336 // into an immediate field.
2337 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2338 Base.getNumRegs() > 1,
2339 LU.Kind, LU.AccessTy, TLI, SE))
2342 // Collect all operands except *J.
2343 SmallVector<const SCEV *, 8> InnerAddOps
2344 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2346 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2348 // Don't leave just a constant behind in a register if the constant could
2349 // be folded into an immediate field.
2350 if (InnerAddOps.size() == 1 &&
2351 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2352 Base.getNumRegs() > 1,
2353 LU.Kind, LU.AccessTy, TLI, SE))
2356 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2357 if (InnerSum->isZero())
2360 F.BaseRegs[i] = InnerSum;
2361 F.BaseRegs.push_back(*J);
2362 if (InsertFormula(LU, LUIdx, F))
2363 // If that formula hadn't been seen before, recurse to find more like
2365 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2370 /// GenerateCombinations - Generate a formula consisting of all of the
2371 /// loop-dominating registers added into a single register.
2372 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2374 // This method is only interesting on a plurality of registers.
2375 if (Base.BaseRegs.size() <= 1) return;
2379 SmallVector<const SCEV *, 4> Ops;
2380 for (SmallVectorImpl<const SCEV *>::const_iterator
2381 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2382 const SCEV *BaseReg = *I;
2383 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2384 !BaseReg->hasComputableLoopEvolution(L))
2385 Ops.push_back(BaseReg);
2387 F.BaseRegs.push_back(BaseReg);
2389 if (Ops.size() > 1) {
2390 const SCEV *Sum = SE.getAddExpr(Ops);
2391 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2392 // opportunity to fold something. For now, just ignore such cases
2393 // rather than proceed with zero in a register.
2394 if (!Sum->isZero()) {
2395 F.BaseRegs.push_back(Sum);
2396 (void)InsertFormula(LU, LUIdx, F);
2401 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2402 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2404 // We can't add a symbolic offset if the address already contains one.
2405 if (Base.AM.BaseGV) return;
2407 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2408 const SCEV *G = Base.BaseRegs[i];
2409 GlobalValue *GV = ExtractSymbol(G, SE);
2410 if (G->isZero() || !GV)
2414 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2415 LU.Kind, LU.AccessTy, TLI))
2418 (void)InsertFormula(LU, LUIdx, F);
2422 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2423 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2425 // TODO: For now, just add the min and max offset, because it usually isn't
2426 // worthwhile looking at everything inbetween.
2427 SmallVector<int64_t, 2> Worklist;
2428 Worklist.push_back(LU.MinOffset);
2429 if (LU.MaxOffset != LU.MinOffset)
2430 Worklist.push_back(LU.MaxOffset);
2432 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2433 const SCEV *G = Base.BaseRegs[i];
2435 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2436 E = Worklist.end(); I != E; ++I) {
2438 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2439 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2440 LU.Kind, LU.AccessTy, TLI)) {
2441 // Add the offset to the base register.
2442 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2443 // If it cancelled out, drop the base register, otherwise update it.
2444 if (NewG->isZero()) {
2445 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2446 F.BaseRegs.pop_back();
2448 F.BaseRegs[i] = NewG;
2450 (void)InsertFormula(LU, LUIdx, F);
2454 int64_t Imm = ExtractImmediate(G, SE);
2455 if (G->isZero() || Imm == 0)
2458 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2459 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2460 LU.Kind, LU.AccessTy, TLI))
2463 (void)InsertFormula(LU, LUIdx, F);
2467 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2468 /// the comparison. For example, x == y -> x*c == y*c.
2469 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2471 if (LU.Kind != LSRUse::ICmpZero) return;
2473 // Determine the integer type for the base formula.
2474 const Type *IntTy = Base.getType();
2476 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2478 // Don't do this if there is more than one offset.
2479 if (LU.MinOffset != LU.MaxOffset) return;
2481 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2483 // Check each interesting stride.
2484 for (SmallSetVector<int64_t, 8>::const_iterator
2485 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2486 int64_t Factor = *I;
2488 // Check that the multiplication doesn't overflow.
2489 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2491 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2492 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2495 // Check that multiplying with the use offset doesn't overflow.
2496 int64_t Offset = LU.MinOffset;
2497 if (Offset == INT64_MIN && Factor == -1)
2499 Offset = (uint64_t)Offset * Factor;
2500 if (Offset / Factor != LU.MinOffset)
2504 F.AM.BaseOffs = NewBaseOffs;
2506 // Check that this scale is legal.
2507 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2510 // Compensate for the use having MinOffset built into it.
2511 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2513 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2515 // Check that multiplying with each base register doesn't overflow.
2516 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2517 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2518 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2522 // Check that multiplying with the scaled register doesn't overflow.
2524 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2525 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2529 // If we make it here and it's legal, add it.
2530 (void)InsertFormula(LU, LUIdx, F);
2535 /// GenerateScales - Generate stride factor reuse formulae by making use of
2536 /// scaled-offset address modes, for example.
2537 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2538 // Determine the integer type for the base formula.
2539 const Type *IntTy = Base.getType();
2542 // If this Formula already has a scaled register, we can't add another one.
2543 if (Base.AM.Scale != 0) return;
2545 // Check each interesting stride.
2546 for (SmallSetVector<int64_t, 8>::const_iterator
2547 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2548 int64_t Factor = *I;
2550 Base.AM.Scale = Factor;
2551 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2552 // Check whether this scale is going to be legal.
2553 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2554 LU.Kind, LU.AccessTy, TLI)) {
2555 // As a special-case, handle special out-of-loop Basic users specially.
2556 // TODO: Reconsider this special case.
2557 if (LU.Kind == LSRUse::Basic &&
2558 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2559 LSRUse::Special, LU.AccessTy, TLI) &&
2560 LU.AllFixupsOutsideLoop)
2561 LU.Kind = LSRUse::Special;
2565 // For an ICmpZero, negating a solitary base register won't lead to
2567 if (LU.Kind == LSRUse::ICmpZero &&
2568 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2570 // For each addrec base reg, apply the scale, if possible.
2571 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2572 if (const SCEVAddRecExpr *AR =
2573 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2574 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2575 if (FactorS->isZero())
2577 // Divide out the factor, ignoring high bits, since we'll be
2578 // scaling the value back up in the end.
2579 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2580 // TODO: This could be optimized to avoid all the copying.
2582 F.ScaledReg = Quotient;
2583 F.DeleteBaseReg(F.BaseRegs[i]);
2584 (void)InsertFormula(LU, LUIdx, F);
2590 /// GenerateTruncates - Generate reuse formulae from different IV types.
2591 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2592 // This requires TargetLowering to tell us which truncates are free.
2595 // Don't bother truncating symbolic values.
2596 if (Base.AM.BaseGV) return;
2598 // Determine the integer type for the base formula.
2599 const Type *DstTy = Base.getType();
2601 DstTy = SE.getEffectiveSCEVType(DstTy);
2603 for (SmallSetVector<const Type *, 4>::const_iterator
2604 I = Types.begin(), E = Types.end(); I != E; ++I) {
2605 const Type *SrcTy = *I;
2606 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2609 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2610 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2611 JE = F.BaseRegs.end(); J != JE; ++J)
2612 *J = SE.getAnyExtendExpr(*J, SrcTy);
2614 // TODO: This assumes we've done basic processing on all uses and
2615 // have an idea what the register usage is.
2616 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2619 (void)InsertFormula(LU, LUIdx, F);
2626 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2627 /// defer modifications so that the search phase doesn't have to worry about
2628 /// the data structures moving underneath it.
2632 const SCEV *OrigReg;
2634 WorkItem(size_t LI, int64_t I, const SCEV *R)
2635 : LUIdx(LI), Imm(I), OrigReg(R) {}
2637 bool operator==(const WorkItem &that) const {
2638 return LUIdx == that.LUIdx && Imm == that.Imm && OrigReg == that.OrigReg;
2640 bool operator<(const WorkItem &that) const {
2641 if (LUIdx != that.LUIdx)
2642 return LUIdx < that.LUIdx;
2643 if (Imm != that.Imm)
2644 return Imm < that.Imm;
2645 return OrigReg < that.OrigReg;
2648 void print(raw_ostream &OS) const;
2654 void WorkItem::print(raw_ostream &OS) const {
2655 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2656 << " , add offset " << Imm;
2659 void WorkItem::dump() const {
2660 print(errs()); errs() << '\n';
2663 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2664 /// distance apart and try to form reuse opportunities between them.
2665 void LSRInstance::GenerateCrossUseConstantOffsets() {
2666 // Group the registers by their value without any added constant offset.
2667 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2668 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2670 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2671 SmallVector<const SCEV *, 8> Sequence;
2672 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2674 const SCEV *Reg = *I;
2675 int64_t Imm = ExtractImmediate(Reg, SE);
2676 std::pair<RegMapTy::iterator, bool> Pair =
2677 Map.insert(std::make_pair(Reg, ImmMapTy()));
2679 Sequence.push_back(Reg);
2680 Pair.first->second.insert(std::make_pair(Imm, *I));
2681 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2684 // Now examine each set of registers with the same base value. Build up
2685 // a list of work to do and do the work in a separate step so that we're
2686 // not adding formulae and register counts while we're searching.
2687 SmallSetVector<WorkItem, 32> WorkItems;
2688 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2689 E = Sequence.end(); I != E; ++I) {
2690 const SCEV *Reg = *I;
2691 const ImmMapTy &Imms = Map.find(Reg)->second;
2693 // It's not worthwhile looking for reuse if there's only one offset.
2694 if (Imms.size() == 1)
2697 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2698 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2700 dbgs() << ' ' << J->first;
2703 // Examine each offset.
2704 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2706 const SCEV *OrigReg = J->second;
2708 int64_t JImm = J->first;
2709 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2711 if (!isa<SCEVConstant>(OrigReg) &&
2712 UsedByIndicesMap[Reg].count() == 1) {
2713 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2717 // Conservatively examine offsets between this orig reg a few selected
2719 ImmMapTy::const_iterator OtherImms[] = {
2720 Imms.begin(), prior(Imms.end()),
2721 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2723 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2724 ImmMapTy::const_iterator M = OtherImms[i];
2725 if (M == J || M == JE) continue;
2727 // Compute the difference between the two.
2728 int64_t Imm = (uint64_t)JImm - M->first;
2729 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2730 LUIdx = UsedByIndices.find_next(LUIdx)) {
2731 // Make a memo of this use, offset, and register tuple.
2732 WorkItems.insert(WorkItem(LUIdx, Imm, OrigReg));
2740 UsedByIndicesMap.clear();
2742 // Now iterate through the worklist and add new formulae.
2743 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2744 E = WorkItems.end(); I != E; ++I) {
2745 const WorkItem &WI = *I;
2746 size_t LUIdx = WI.LUIdx;
2747 LSRUse &LU = Uses[LUIdx];
2748 int64_t Imm = WI.Imm;
2749 const SCEV *OrigReg = WI.OrigReg;
2751 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2752 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2753 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2755 // TODO: Use a more targeted data structure.
2756 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2757 const Formula &F = LU.Formulae[L];
2758 // Use the immediate in the scaled register.
2759 if (F.ScaledReg == OrigReg) {
2760 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2761 Imm * (uint64_t)F.AM.Scale;
2762 // Don't create 50 + reg(-50).
2763 if (F.referencesReg(SE.getSCEV(
2764 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2767 NewF.AM.BaseOffs = Offs;
2768 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2769 LU.Kind, LU.AccessTy, TLI))
2771 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2773 // If the new scale is a constant in a register, and adding the constant
2774 // value to the immediate would produce a value closer to zero than the
2775 // immediate itself, then the formula isn't worthwhile.
2776 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2777 if (C->getValue()->getValue().isNegative() !=
2778 (NewF.AM.BaseOffs < 0) &&
2779 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2780 .ule(abs64(NewF.AM.BaseOffs)))
2784 (void)InsertFormula(LU, LUIdx, NewF);
2786 // Use the immediate in a base register.
2787 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2788 const SCEV *BaseReg = F.BaseRegs[N];
2789 if (BaseReg != OrigReg)
2792 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2793 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2794 LU.Kind, LU.AccessTy, TLI))
2796 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2798 // If the new formula has a constant in a register, and adding the
2799 // constant value to the immediate would produce a value closer to
2800 // zero than the immediate itself, then the formula isn't worthwhile.
2801 for (SmallVectorImpl<const SCEV *>::const_iterator
2802 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2804 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2805 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2806 abs64(NewF.AM.BaseOffs)) &&
2807 (C->getValue()->getValue() +
2808 NewF.AM.BaseOffs).countTrailingZeros() >=
2809 CountTrailingZeros_64(NewF.AM.BaseOffs))
2813 (void)InsertFormula(LU, LUIdx, NewF);
2822 /// GenerateAllReuseFormulae - Generate formulae for each use.
2824 LSRInstance::GenerateAllReuseFormulae() {
2825 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2826 // queries are more precise.
2827 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2828 LSRUse &LU = Uses[LUIdx];
2829 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2830 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2831 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2832 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2834 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2835 LSRUse &LU = Uses[LUIdx];
2836 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2837 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2838 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2839 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2840 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2841 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2842 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2843 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2845 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2846 LSRUse &LU = Uses[LUIdx];
2847 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2848 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2851 GenerateCrossUseConstantOffsets();
2853 DEBUG(dbgs() << "\n"
2854 "After generating reuse formulae:\n";
2855 print_uses(dbgs()));
2858 /// If their are multiple formulae with the same set of registers used
2859 /// by other uses, pick the best one and delete the others.
2860 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2862 bool ChangedFormulae = false;
2865 // Collect the best formula for each unique set of shared registers. This
2866 // is reset for each use.
2867 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2869 BestFormulaeTy BestFormulae;
2871 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2872 LSRUse &LU = Uses[LUIdx];
2873 FormulaSorter Sorter(L, LU, SE, DT);
2874 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2877 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2878 FIdx != NumForms; ++FIdx) {
2879 Formula &F = LU.Formulae[FIdx];
2881 SmallVector<const SCEV *, 2> Key;
2882 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2883 JE = F.BaseRegs.end(); J != JE; ++J) {
2884 const SCEV *Reg = *J;
2885 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2889 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2890 Key.push_back(F.ScaledReg);
2891 // Unstable sort by host order ok, because this is only used for
2893 std::sort(Key.begin(), Key.end());
2895 std::pair<BestFormulaeTy::const_iterator, bool> P =
2896 BestFormulae.insert(std::make_pair(Key, FIdx));
2898 Formula &Best = LU.Formulae[P.first->second];
2899 if (Sorter.operator()(F, Best))
2901 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2903 " in favor of formula "; Best.print(dbgs());
2906 ChangedFormulae = true;
2908 LU.DeleteFormula(F);
2916 // Now that we've filtered out some formulae, recompute the Regs set.
2918 LU.RecomputeRegs(LUIdx, RegUses);
2920 // Reset this to prepare for the next use.
2921 BestFormulae.clear();
2924 DEBUG(if (ChangedFormulae) {
2926 "After filtering out undesirable candidates:\n";
2931 // This is a rough guess that seems to work fairly well.
2932 static const size_t ComplexityLimit = UINT16_MAX;
2934 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2935 /// solutions the solver might have to consider. It almost never considers
2936 /// this many solutions because it prune the search space, but the pruning
2937 /// isn't always sufficient.
2938 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2940 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2941 E = Uses.end(); I != E; ++I) {
2942 size_t FSize = I->Formulae.size();
2943 if (FSize >= ComplexityLimit) {
2944 Power = ComplexityLimit;
2948 if (Power >= ComplexityLimit)
2954 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2955 /// of the registers of another formula, it won't help reduce register
2956 /// pressure (though it may not necessarily hurt register pressure); remove
2957 /// it to simplify the system.
2958 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2959 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2960 DEBUG(dbgs() << "The search space is too complex.\n");
2962 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2963 "which use a superset of registers used by other "
2966 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2967 LSRUse &LU = Uses[LUIdx];
2969 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2970 Formula &F = LU.Formulae[i];
2971 // Look for a formula with a constant or GV in a register. If the use
2972 // also has a formula with that same value in an immediate field,
2973 // delete the one that uses a register.
2974 for (SmallVectorImpl<const SCEV *>::const_iterator
2975 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2976 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2978 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2979 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2980 (I - F.BaseRegs.begin()));
2981 if (LU.HasFormulaWithSameRegs(NewF)) {
2982 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2983 LU.DeleteFormula(F);
2989 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2990 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2993 NewF.AM.BaseGV = GV;
2994 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2995 (I - F.BaseRegs.begin()));
2996 if (LU.HasFormulaWithSameRegs(NewF)) {
2997 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2999 LU.DeleteFormula(F);
3010 LU.RecomputeRegs(LUIdx, RegUses);
3013 DEBUG(dbgs() << "After pre-selection:\n";
3014 print_uses(dbgs()));
3018 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3019 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3021 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3022 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3023 DEBUG(dbgs() << "The search space is too complex.\n");
3025 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3026 "separated by a constant offset will use the same "
3029 // This is especially useful for unrolled loops.
3031 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3032 LSRUse &LU = Uses[LUIdx];
3033 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3034 E = LU.Formulae.end(); I != E; ++I) {
3035 const Formula &F = *I;
3036 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3037 int64_t NewBaseOffs;
3038 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU,
3040 if (reconcileNewOffset(*LUThatHas,
3041 F.AM.BaseOffs + LU.MinOffset - NewBaseOffs,
3042 F.AM.BaseOffs + LU.MaxOffset - NewBaseOffs,
3043 /*HasBaseReg=*/false,
3044 LU.Kind, LU.AccessTy)) {
3045 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3048 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3050 // Update the relocs to reference the new use.
3051 // Do this first so that MinOffset and MaxOffset are updated
3052 // before we begin to determine which formulae to delete.
3053 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3054 E = Fixups.end(); I != E; ++I) {
3055 LSRFixup &Fixup = *I;
3056 if (Fixup.LUIdx == LUIdx) {
3057 Fixup.LUIdx = LUThatHas - &Uses.front();
3058 Fixup.Offset += F.AM.BaseOffs - NewBaseOffs;
3059 DEBUG(dbgs() << "New fixup has offset "
3060 << Fixup.Offset << '\n');
3061 LUThatHas->Offsets.push_back(Fixup.Offset);
3062 if (Fixup.Offset > LUThatHas->MaxOffset)
3063 LUThatHas->MaxOffset = Fixup.Offset;
3064 if (Fixup.Offset < LUThatHas->MinOffset)
3065 LUThatHas->MinOffset = Fixup.Offset;
3067 // DeleteUse will do a swap+pop_back, so if this fixup is
3068 // now pointing to the last LSRUse, update it to point to the
3069 // position it'll be swapped to.
3070 if (Fixup.LUIdx == NumUses-1)
3071 Fixup.LUIdx = LUIdx;
3074 // Delete formulae from the new use which are no longer legal.
3076 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3077 Formula &F = LUThatHas->Formulae[i];
3078 if (!isLegalUse(F.AM,
3079 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3080 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3081 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3083 LUThatHas->DeleteFormula(F);
3090 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3092 // Delete the old use.
3103 DEBUG(dbgs() << "After pre-selection:\n";
3104 print_uses(dbgs()));
3108 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3109 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3110 /// we've done more filtering, as it may be able to find more formulae to
3112 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3113 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3114 DEBUG(dbgs() << "The search space is too complex.\n");
3116 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3117 "undesirable dedicated registers.\n");
3119 FilterOutUndesirableDedicatedRegisters();
3121 DEBUG(dbgs() << "After pre-selection:\n";
3122 print_uses(dbgs()));
3126 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3127 /// to be profitable, and then in any use which has any reference to that
3128 /// register, delete all formulae which do not reference that register.
3129 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3130 // With all other options exhausted, loop until the system is simple
3131 // enough to handle.
3132 SmallPtrSet<const SCEV *, 4> Taken;
3133 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3134 // Ok, we have too many of formulae on our hands to conveniently handle.
3135 // Use a rough heuristic to thin out the list.
3136 DEBUG(dbgs() << "The search space is too complex.\n");
3138 // Pick the register which is used by the most LSRUses, which is likely
3139 // to be a good reuse register candidate.
3140 const SCEV *Best = 0;
3141 unsigned BestNum = 0;
3142 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3144 const SCEV *Reg = *I;
3145 if (Taken.count(Reg))
3150 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3151 if (Count > BestNum) {
3158 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3159 << " will yield profitable reuse.\n");
3162 // In any use with formulae which references this register, delete formulae
3163 // which don't reference it.
3164 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3165 LSRUse &LU = Uses[LUIdx];
3166 if (!LU.Regs.count(Best)) continue;
3169 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3170 Formula &F = LU.Formulae[i];
3171 if (!F.referencesReg(Best)) {
3172 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3173 LU.DeleteFormula(F);
3177 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3183 LU.RecomputeRegs(LUIdx, RegUses);
3186 DEBUG(dbgs() << "After pre-selection:\n";
3187 print_uses(dbgs()));
3191 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3192 /// formulae to choose from, use some rough heuristics to prune down the number
3193 /// of formulae. This keeps the main solver from taking an extraordinary amount
3194 /// of time in some worst-case scenarios.
3195 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3196 NarrowSearchSpaceByDetectingSupersets();
3197 NarrowSearchSpaceByCollapsingUnrolledCode();
3198 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3199 NarrowSearchSpaceByPickingWinnerRegs();
3202 /// SolveRecurse - This is the recursive solver.
3203 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3205 SmallVectorImpl<const Formula *> &Workspace,
3206 const Cost &CurCost,
3207 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3208 DenseSet<const SCEV *> &VisitedRegs) const {
3211 // - use more aggressive filtering
3212 // - sort the formula so that the most profitable solutions are found first
3213 // - sort the uses too
3215 // - don't compute a cost, and then compare. compare while computing a cost
3217 // - track register sets with SmallBitVector
3219 const LSRUse &LU = Uses[Workspace.size()];
3221 // If this use references any register that's already a part of the
3222 // in-progress solution, consider it a requirement that a formula must
3223 // reference that register in order to be considered. This prunes out
3224 // unprofitable searching.
3225 SmallSetVector<const SCEV *, 4> ReqRegs;
3226 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3227 E = CurRegs.end(); I != E; ++I)
3228 if (LU.Regs.count(*I))
3231 bool AnySatisfiedReqRegs = false;
3232 SmallPtrSet<const SCEV *, 16> NewRegs;
3235 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3236 E = LU.Formulae.end(); I != E; ++I) {
3237 const Formula &F = *I;
3239 // Ignore formulae which do not use any of the required registers.
3240 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3241 JE = ReqRegs.end(); J != JE; ++J) {
3242 const SCEV *Reg = *J;
3243 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3244 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3248 AnySatisfiedReqRegs = true;
3250 // Evaluate the cost of the current formula. If it's already worse than
3251 // the current best, prune the search at that point.
3254 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3255 if (NewCost < SolutionCost) {
3256 Workspace.push_back(&F);
3257 if (Workspace.size() != Uses.size()) {
3258 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3259 NewRegs, VisitedRegs);
3260 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3261 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3263 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3264 dbgs() << ". Regs:";
3265 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3266 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3267 dbgs() << ' ' << **I;
3270 SolutionCost = NewCost;
3271 Solution = Workspace;
3273 Workspace.pop_back();
3278 // If none of the formulae had all of the required registers, relax the
3279 // constraint so that we don't exclude all formulae.
3280 if (!AnySatisfiedReqRegs) {
3281 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3287 /// Solve - Choose one formula from each use. Return the results in the given
3288 /// Solution vector.
3289 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3290 SmallVector<const Formula *, 8> Workspace;
3292 SolutionCost.Loose();
3294 SmallPtrSet<const SCEV *, 16> CurRegs;
3295 DenseSet<const SCEV *> VisitedRegs;
3296 Workspace.reserve(Uses.size());
3298 // SolveRecurse does all the work.
3299 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3300 CurRegs, VisitedRegs);
3302 // Ok, we've now made all our decisions.
3303 DEBUG(dbgs() << "\n"
3304 "The chosen solution requires "; SolutionCost.print(dbgs());
3306 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3308 Uses[i].print(dbgs());
3311 Solution[i]->print(dbgs());
3315 assert(Solution.size() == Uses.size() && "Malformed solution!");
3318 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3319 /// the dominator tree far as we can go while still being dominated by the
3320 /// input positions. This helps canonicalize the insert position, which
3321 /// encourages sharing.
3322 BasicBlock::iterator
3323 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3324 const SmallVectorImpl<Instruction *> &Inputs)
3327 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3328 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3331 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3332 if (!Rung) return IP;
3333 Rung = Rung->getIDom();
3334 if (!Rung) return IP;
3335 IDom = Rung->getBlock();
3337 // Don't climb into a loop though.
3338 const Loop *IDomLoop = LI.getLoopFor(IDom);
3339 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3340 if (IDomDepth <= IPLoopDepth &&
3341 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3345 bool AllDominate = true;
3346 Instruction *BetterPos = 0;
3347 Instruction *Tentative = IDom->getTerminator();
3348 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3349 E = Inputs.end(); I != E; ++I) {
3350 Instruction *Inst = *I;
3351 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3352 AllDominate = false;
3355 // Attempt to find an insert position in the middle of the block,
3356 // instead of at the end, so that it can be used for other expansions.
3357 if (IDom == Inst->getParent() &&
3358 (!BetterPos || DT.dominates(BetterPos, Inst)))
3359 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3372 /// AdjustInsertPositionForExpand - Determine an input position which will be
3373 /// dominated by the operands and which will dominate the result.
3374 BasicBlock::iterator
3375 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3377 const LSRUse &LU) const {
3378 // Collect some instructions which must be dominated by the
3379 // expanding replacement. These must be dominated by any operands that
3380 // will be required in the expansion.
3381 SmallVector<Instruction *, 4> Inputs;
3382 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3383 Inputs.push_back(I);
3384 if (LU.Kind == LSRUse::ICmpZero)
3385 if (Instruction *I =
3386 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3387 Inputs.push_back(I);
3388 if (LF.PostIncLoops.count(L)) {
3389 if (LF.isUseFullyOutsideLoop(L))
3390 Inputs.push_back(L->getLoopLatch()->getTerminator());
3392 Inputs.push_back(IVIncInsertPos);
3394 // The expansion must also be dominated by the increment positions of any
3395 // loops it for which it is using post-inc mode.
3396 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3397 E = LF.PostIncLoops.end(); I != E; ++I) {
3398 const Loop *PIL = *I;
3399 if (PIL == L) continue;
3401 // Be dominated by the loop exit.
3402 SmallVector<BasicBlock *, 4> ExitingBlocks;
3403 PIL->getExitingBlocks(ExitingBlocks);
3404 if (!ExitingBlocks.empty()) {
3405 BasicBlock *BB = ExitingBlocks[0];
3406 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3407 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3408 Inputs.push_back(BB->getTerminator());
3412 // Then, climb up the immediate dominator tree as far as we can go while
3413 // still being dominated by the input positions.
3414 IP = HoistInsertPosition(IP, Inputs);
3416 // Don't insert instructions before PHI nodes.
3417 while (isa<PHINode>(IP)) ++IP;
3419 // Ignore debug intrinsics.
3420 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3425 /// Expand - Emit instructions for the leading candidate expression for this
3426 /// LSRUse (this is called "expanding").
3427 Value *LSRInstance::Expand(const LSRFixup &LF,
3429 BasicBlock::iterator IP,
3430 SCEVExpander &Rewriter,
3431 SmallVectorImpl<WeakVH> &DeadInsts) const {
3432 const LSRUse &LU = Uses[LF.LUIdx];
3434 // Determine an input position which will be dominated by the operands and
3435 // which will dominate the result.
3436 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3438 // Inform the Rewriter if we have a post-increment use, so that it can
3439 // perform an advantageous expansion.
3440 Rewriter.setPostInc(LF.PostIncLoops);
3442 // This is the type that the user actually needs.
3443 const Type *OpTy = LF.OperandValToReplace->getType();
3444 // This will be the type that we'll initially expand to.
3445 const Type *Ty = F.getType();
3447 // No type known; just expand directly to the ultimate type.
3449 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3450 // Expand directly to the ultimate type if it's the right size.
3452 // This is the type to do integer arithmetic in.
3453 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3455 // Build up a list of operands to add together to form the full base.
3456 SmallVector<const SCEV *, 8> Ops;
3458 // Expand the BaseRegs portion.
3459 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3460 E = F.BaseRegs.end(); I != E; ++I) {
3461 const SCEV *Reg = *I;
3462 assert(!Reg->isZero() && "Zero allocated in a base register!");
3464 // If we're expanding for a post-inc user, make the post-inc adjustment.
3465 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3466 Reg = TransformForPostIncUse(Denormalize, Reg,
3467 LF.UserInst, LF.OperandValToReplace,
3470 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3473 // Flush the operand list to suppress SCEVExpander hoisting.
3475 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3477 Ops.push_back(SE.getUnknown(FullV));
3480 // Expand the ScaledReg portion.
3481 Value *ICmpScaledV = 0;
3482 if (F.AM.Scale != 0) {
3483 const SCEV *ScaledS = F.ScaledReg;
3485 // If we're expanding for a post-inc user, make the post-inc adjustment.
3486 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3487 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3488 LF.UserInst, LF.OperandValToReplace,
3491 if (LU.Kind == LSRUse::ICmpZero) {
3492 // An interesting way of "folding" with an icmp is to use a negated
3493 // scale, which we'll implement by inserting it into the other operand
3495 assert(F.AM.Scale == -1 &&
3496 "The only scale supported by ICmpZero uses is -1!");
3497 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3499 // Otherwise just expand the scaled register and an explicit scale,
3500 // which is expected to be matched as part of the address.
3501 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3502 ScaledS = SE.getMulExpr(ScaledS,
3503 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3504 Ops.push_back(ScaledS);
3506 // Flush the operand list to suppress SCEVExpander hoisting.
3507 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3509 Ops.push_back(SE.getUnknown(FullV));
3513 // Expand the GV portion.
3515 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3517 // Flush the operand list to suppress SCEVExpander hoisting.
3518 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3520 Ops.push_back(SE.getUnknown(FullV));
3523 // Expand the immediate portion.
3524 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3526 if (LU.Kind == LSRUse::ICmpZero) {
3527 // The other interesting way of "folding" with an ICmpZero is to use a
3528 // negated immediate.
3530 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3532 Ops.push_back(SE.getUnknown(ICmpScaledV));
3533 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3536 // Just add the immediate values. These again are expected to be matched
3537 // as part of the address.
3538 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3542 // Emit instructions summing all the operands.
3543 const SCEV *FullS = Ops.empty() ?
3544 SE.getConstant(IntTy, 0) :
3546 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3548 // We're done expanding now, so reset the rewriter.
3549 Rewriter.clearPostInc();
3551 // An ICmpZero Formula represents an ICmp which we're handling as a
3552 // comparison against zero. Now that we've expanded an expression for that
3553 // form, update the ICmp's other operand.
3554 if (LU.Kind == LSRUse::ICmpZero) {
3555 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3556 DeadInsts.push_back(CI->getOperand(1));
3557 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3558 "a scale at the same time!");
3559 if (F.AM.Scale == -1) {
3560 if (ICmpScaledV->getType() != OpTy) {
3562 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3564 ICmpScaledV, OpTy, "tmp", CI);
3567 CI->setOperand(1, ICmpScaledV);
3569 assert(F.AM.Scale == 0 &&
3570 "ICmp does not support folding a global value and "
3571 "a scale at the same time!");
3572 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3574 if (C->getType() != OpTy)
3575 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3579 CI->setOperand(1, C);
3586 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3587 /// of their operands effectively happens in their predecessor blocks, so the
3588 /// expression may need to be expanded in multiple places.
3589 void LSRInstance::RewriteForPHI(PHINode *PN,
3592 SCEVExpander &Rewriter,
3593 SmallVectorImpl<WeakVH> &DeadInsts,
3595 DenseMap<BasicBlock *, Value *> Inserted;
3596 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3597 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3598 BasicBlock *BB = PN->getIncomingBlock(i);
3600 // If this is a critical edge, split the edge so that we do not insert
3601 // the code on all predecessor/successor paths. We do this unless this
3602 // is the canonical backedge for this loop, which complicates post-inc
3604 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3605 !isa<IndirectBrInst>(BB->getTerminator()) &&
3606 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3607 // Split the critical edge.
3608 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3610 // If PN is outside of the loop and BB is in the loop, we want to
3611 // move the block to be immediately before the PHI block, not
3612 // immediately after BB.
3613 if (L->contains(BB) && !L->contains(PN))
3614 NewBB->moveBefore(PN->getParent());
3616 // Splitting the edge can reduce the number of PHI entries we have.
3617 e = PN->getNumIncomingValues();
3619 i = PN->getBasicBlockIndex(BB);
3622 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3623 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3625 PN->setIncomingValue(i, Pair.first->second);
3627 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3629 // If this is reuse-by-noop-cast, insert the noop cast.
3630 const Type *OpTy = LF.OperandValToReplace->getType();
3631 if (FullV->getType() != OpTy)
3633 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3635 FullV, LF.OperandValToReplace->getType(),
3636 "tmp", BB->getTerminator());
3638 PN->setIncomingValue(i, FullV);
3639 Pair.first->second = FullV;
3644 /// Rewrite - Emit instructions for the leading candidate expression for this
3645 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3646 /// the newly expanded value.
3647 void LSRInstance::Rewrite(const LSRFixup &LF,
3649 SCEVExpander &Rewriter,
3650 SmallVectorImpl<WeakVH> &DeadInsts,
3652 // First, find an insertion point that dominates UserInst. For PHI nodes,
3653 // find the nearest block which dominates all the relevant uses.
3654 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3655 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3657 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3659 // If this is reuse-by-noop-cast, insert the noop cast.
3660 const Type *OpTy = LF.OperandValToReplace->getType();
3661 if (FullV->getType() != OpTy) {
3663 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3664 FullV, OpTy, "tmp", LF.UserInst);
3668 // Update the user. ICmpZero is handled specially here (for now) because
3669 // Expand may have updated one of the operands of the icmp already, and
3670 // its new value may happen to be equal to LF.OperandValToReplace, in
3671 // which case doing replaceUsesOfWith leads to replacing both operands
3672 // with the same value. TODO: Reorganize this.
3673 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3674 LF.UserInst->setOperand(0, FullV);
3676 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3679 DeadInsts.push_back(LF.OperandValToReplace);
3682 /// ImplementSolution - Rewrite all the fixup locations with new values,
3683 /// following the chosen solution.
3685 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3687 // Keep track of instructions we may have made dead, so that
3688 // we can remove them after we are done working.
3689 SmallVector<WeakVH, 16> DeadInsts;
3691 SCEVExpander Rewriter(SE);
3692 Rewriter.disableCanonicalMode();
3693 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3695 // Expand the new value definitions and update the users.
3696 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3697 E = Fixups.end(); I != E; ++I) {
3698 const LSRFixup &Fixup = *I;
3700 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3705 // Clean up after ourselves. This must be done before deleting any
3709 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3712 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3713 : IU(P->getAnalysis<IVUsers>()),
3714 SE(P->getAnalysis<ScalarEvolution>()),
3715 DT(P->getAnalysis<DominatorTree>()),
3716 LI(P->getAnalysis<LoopInfo>()),
3717 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3719 // If LoopSimplify form is not available, stay out of trouble.
3720 if (!L->isLoopSimplifyForm()) return;
3722 // If there's no interesting work to be done, bail early.
3723 if (IU.empty()) return;
3725 DEBUG(dbgs() << "\nLSR on loop ";
3726 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3729 // First, perform some low-level loop optimizations.
3731 OptimizeLoopTermCond();
3733 // Start collecting data and preparing for the solver.
3734 CollectInterestingTypesAndFactors();
3735 CollectFixupsAndInitialFormulae();
3736 CollectLoopInvariantFixupsAndFormulae();
3738 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3739 print_uses(dbgs()));
3741 // Now use the reuse data to generate a bunch of interesting ways
3742 // to formulate the values needed for the uses.
3743 GenerateAllReuseFormulae();
3745 FilterOutUndesirableDedicatedRegisters();
3746 NarrowSearchSpaceUsingHeuristics();
3748 SmallVector<const Formula *, 8> Solution;
3751 // Release memory that is no longer needed.
3757 // Formulae should be legal.
3758 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3759 E = Uses.end(); I != E; ++I) {
3760 const LSRUse &LU = *I;
3761 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3762 JE = LU.Formulae.end(); J != JE; ++J)
3763 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3764 LU.Kind, LU.AccessTy, TLI) &&
3765 "Illegal formula generated!");
3769 // Now that we've decided what we want, make it so.
3770 ImplementSolution(Solution, P);
3773 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3774 if (Factors.empty() && Types.empty()) return;
3776 OS << "LSR has identified the following interesting factors and types: ";
3779 for (SmallSetVector<int64_t, 8>::const_iterator
3780 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3781 if (!First) OS << ", ";
3786 for (SmallSetVector<const Type *, 4>::const_iterator
3787 I = Types.begin(), E = Types.end(); I != E; ++I) {
3788 if (!First) OS << ", ";
3790 OS << '(' << **I << ')';
3795 void LSRInstance::print_fixups(raw_ostream &OS) const {
3796 OS << "LSR is examining the following fixup sites:\n";
3797 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3798 E = Fixups.end(); I != E; ++I) {
3805 void LSRInstance::print_uses(raw_ostream &OS) const {
3806 OS << "LSR is examining the following uses:\n";
3807 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3808 E = Uses.end(); I != E; ++I) {
3809 const LSRUse &LU = *I;
3813 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3814 JE = LU.Formulae.end(); J != JE; ++J) {
3822 void LSRInstance::print(raw_ostream &OS) const {
3823 print_factors_and_types(OS);
3828 void LSRInstance::dump() const {
3829 print(errs()); errs() << '\n';
3834 class LoopStrengthReduce : public LoopPass {
3835 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3836 /// transformation profitability.
3837 const TargetLowering *const TLI;
3840 static char ID; // Pass ID, replacement for typeid
3841 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3844 bool runOnLoop(Loop *L, LPPassManager &LPM);
3845 void getAnalysisUsage(AnalysisUsage &AU) const;
3850 char LoopStrengthReduce::ID = 0;
3851 INITIALIZE_PASS(LoopStrengthReduce, "loop-reduce",
3852 "Loop Strength Reduction", false, false);
3854 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3855 return new LoopStrengthReduce(TLI);
3858 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3859 : LoopPass(ID), TLI(tli) {}
3861 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3862 // We split critical edges, so we change the CFG. However, we do update
3863 // many analyses if they are around.
3864 AU.addPreservedID(LoopSimplifyID);
3865 AU.addPreserved("domfrontier");
3867 AU.addRequired<LoopInfo>();
3868 AU.addPreserved<LoopInfo>();
3869 AU.addRequiredID(LoopSimplifyID);
3870 AU.addRequired<DominatorTree>();
3871 AU.addPreserved<DominatorTree>();
3872 AU.addRequired<ScalarEvolution>();
3873 AU.addPreserved<ScalarEvolution>();
3874 AU.addRequired<IVUsers>();
3875 AU.addPreserved<IVUsers>();
3878 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3879 bool Changed = false;
3881 // Run the main LSR transformation.
3882 Changed |= LSRInstance(TLI, L, this).getChanged();
3884 // At this point, it is worth checking to see if any recurrence PHIs are also
3885 // dead, so that we can remove them as well.
3886 Changed |= DeleteDeadPHIs(L->getHeader());