1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/ADT/SmallBitVector.h"
69 #include "llvm/ADT/SetVector.h"
70 #include "llvm/ADT/DenseSet.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/ValueHandle.h"
73 #include "llvm/Support/raw_ostream.h"
74 #include "llvm/Target/TargetLowering.h"
80 /// RegSortData - This class holds data which is used to order reuse candidates.
83 /// UsedByIndices - This represents the set of LSRUse indices which reference
84 /// a particular register.
85 SmallBitVector UsedByIndices;
89 void print(raw_ostream &OS) const;
95 void RegSortData::print(raw_ostream &OS) const {
96 OS << "[NumUses=" << UsedByIndices.count() << ']';
99 void RegSortData::dump() const {
100 print(errs()); errs() << '\n';
105 /// RegUseTracker - Map register candidates to information about how they are
107 class RegUseTracker {
108 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
110 RegUsesTy RegUsesMap;
111 SmallVector<const SCEV *, 16> RegSequence;
114 void CountRegister(const SCEV *Reg, size_t LUIdx);
115 void DropRegister(const SCEV *Reg, size_t LUIdx);
116 void DropUse(size_t LUIdx);
118 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
120 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
124 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
125 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
126 iterator begin() { return RegSequence.begin(); }
127 iterator end() { return RegSequence.end(); }
128 const_iterator begin() const { return RegSequence.begin(); }
129 const_iterator end() const { return RegSequence.end(); }
135 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
136 std::pair<RegUsesTy::iterator, bool> Pair =
137 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
138 RegSortData &RSD = Pair.first->second;
140 RegSequence.push_back(Reg);
141 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
142 RSD.UsedByIndices.set(LUIdx);
146 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
147 RegUsesTy::iterator It = RegUsesMap.find(Reg);
148 assert(It != RegUsesMap.end());
149 RegSortData &RSD = It->second;
150 assert(RSD.UsedByIndices.size() > LUIdx);
151 RSD.UsedByIndices.reset(LUIdx);
155 RegUseTracker::DropUse(size_t LUIdx) {
156 // Remove the use index from every register's use list.
157 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
159 I->second.UsedByIndices.reset(LUIdx);
163 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
164 if (!RegUsesMap.count(Reg)) return false;
165 const SmallBitVector &UsedByIndices =
166 RegUsesMap.find(Reg)->second.UsedByIndices;
167 int i = UsedByIndices.find_first();
168 if (i == -1) return false;
169 if ((size_t)i != LUIdx) return true;
170 return UsedByIndices.find_next(i) != -1;
173 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
174 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
175 assert(I != RegUsesMap.end() && "Unknown register!");
176 return I->second.UsedByIndices;
179 void RegUseTracker::clear() {
186 /// Formula - This class holds information that describes a formula for
187 /// computing satisfying a use. It may include broken-out immediates and scaled
190 /// AM - This is used to represent complex addressing, as well as other kinds
191 /// of interesting uses.
192 TargetLowering::AddrMode AM;
194 /// BaseRegs - The list of "base" registers for this use. When this is
195 /// non-empty, AM.HasBaseReg should be set to true.
196 SmallVector<const SCEV *, 2> BaseRegs;
198 /// ScaledReg - The 'scaled' register for this use. This should be non-null
199 /// when AM.Scale is not zero.
200 const SCEV *ScaledReg;
202 Formula() : ScaledReg(0) {}
204 void InitialMatch(const SCEV *S, Loop *L,
205 ScalarEvolution &SE, DominatorTree &DT);
207 unsigned getNumRegs() const;
208 const Type *getType() const;
210 void DeleteBaseReg(const SCEV *&S);
212 bool referencesReg(const SCEV *S) const;
213 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
214 const RegUseTracker &RegUses) const;
216 void print(raw_ostream &OS) const;
222 /// DoInitialMatch - Recursion helper for InitialMatch.
223 static void DoInitialMatch(const SCEV *S, Loop *L,
224 SmallVectorImpl<const SCEV *> &Good,
225 SmallVectorImpl<const SCEV *> &Bad,
226 ScalarEvolution &SE, DominatorTree &DT) {
227 // Collect expressions which properly dominate the loop header.
228 if (S->properlyDominates(L->getHeader(), &DT)) {
233 // Look at add operands.
234 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
235 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
237 DoInitialMatch(*I, L, Good, Bad, SE, DT);
241 // Look at addrec operands.
242 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
243 if (!AR->getStart()->isZero()) {
244 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
245 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
246 AR->getStepRecurrence(SE),
248 L, Good, Bad, SE, DT);
252 // Handle a multiplication by -1 (negation) if it didn't fold.
253 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
254 if (Mul->getOperand(0)->isAllOnesValue()) {
255 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
256 const SCEV *NewMul = SE.getMulExpr(Ops);
258 SmallVector<const SCEV *, 4> MyGood;
259 SmallVector<const SCEV *, 4> MyBad;
260 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
261 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
262 SE.getEffectiveSCEVType(NewMul->getType())));
263 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
264 E = MyGood.end(); I != E; ++I)
265 Good.push_back(SE.getMulExpr(NegOne, *I));
266 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
267 E = MyBad.end(); I != E; ++I)
268 Bad.push_back(SE.getMulExpr(NegOne, *I));
272 // Ok, we can't do anything interesting. Just stuff the whole thing into a
273 // register and hope for the best.
277 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
278 /// attempting to keep all loop-invariant and loop-computable values in a
279 /// single base register.
280 void Formula::InitialMatch(const SCEV *S, Loop *L,
281 ScalarEvolution &SE, DominatorTree &DT) {
282 SmallVector<const SCEV *, 4> Good;
283 SmallVector<const SCEV *, 4> Bad;
284 DoInitialMatch(S, L, Good, Bad, SE, DT);
286 const SCEV *Sum = SE.getAddExpr(Good);
288 BaseRegs.push_back(Sum);
289 AM.HasBaseReg = true;
292 const SCEV *Sum = SE.getAddExpr(Bad);
294 BaseRegs.push_back(Sum);
295 AM.HasBaseReg = true;
299 /// getNumRegs - Return the total number of register operands used by this
300 /// formula. This does not include register uses implied by non-constant
302 unsigned Formula::getNumRegs() const {
303 return !!ScaledReg + BaseRegs.size();
306 /// getType - Return the type of this formula, if it has one, or null
307 /// otherwise. This type is meaningless except for the bit size.
308 const Type *Formula::getType() const {
309 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
310 ScaledReg ? ScaledReg->getType() :
311 AM.BaseGV ? AM.BaseGV->getType() :
315 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
316 void Formula::DeleteBaseReg(const SCEV *&S) {
317 if (&S != &BaseRegs.back())
318 std::swap(S, BaseRegs.back());
322 /// referencesReg - Test if this formula references the given register.
323 bool Formula::referencesReg(const SCEV *S) const {
324 return S == ScaledReg ||
325 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
328 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
329 /// which are used by uses other than the use with the given index.
330 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
331 const RegUseTracker &RegUses) const {
333 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
335 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
336 E = BaseRegs.end(); I != E; ++I)
337 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
342 void Formula::print(raw_ostream &OS) const {
345 if (!First) OS << " + "; else First = false;
346 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
348 if (AM.BaseOffs != 0) {
349 if (!First) OS << " + "; else First = false;
352 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
353 E = BaseRegs.end(); I != E; ++I) {
354 if (!First) OS << " + "; else First = false;
355 OS << "reg(" << **I << ')';
357 if (AM.HasBaseReg && BaseRegs.empty()) {
358 if (!First) OS << " + "; else First = false;
359 OS << "**error: HasBaseReg**";
360 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
361 if (!First) OS << " + "; else First = false;
362 OS << "**error: !HasBaseReg**";
365 if (!First) OS << " + "; else First = false;
366 OS << AM.Scale << "*reg(";
375 void Formula::dump() const {
376 print(errs()); errs() << '\n';
379 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
380 /// without changing its value.
381 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
383 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
384 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
387 /// isAddSExtable - Return true if the given add can be sign-extended
388 /// without changing its value.
389 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
391 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
392 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
395 /// isMulSExtable - Return true if the given mul can be sign-extended
396 /// without changing its value.
397 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
399 IntegerType::get(SE.getContext(),
400 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
401 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
404 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
405 /// and if the remainder is known to be zero, or null otherwise. If
406 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
407 /// to Y, ignoring that the multiplication may overflow, which is useful when
408 /// the result will be used in a context where the most significant bits are
410 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
412 bool IgnoreSignificantBits = false) {
413 // Handle the trivial case, which works for any SCEV type.
415 return SE.getConstant(LHS->getType(), 1);
417 // Handle a few RHS special cases.
418 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
420 const APInt &RA = RC->getValue()->getValue();
421 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
423 if (RA.isAllOnesValue())
424 return SE.getMulExpr(LHS, RC);
425 // Handle x /s 1 as x.
430 // Check for a division of a constant by a constant.
431 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
434 const APInt &LA = C->getValue()->getValue();
435 const APInt &RA = RC->getValue()->getValue();
436 if (LA.srem(RA) != 0)
438 return SE.getConstant(LA.sdiv(RA));
441 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
442 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
443 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
444 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
445 IgnoreSignificantBits);
446 if (!Start) return 0;
447 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
448 IgnoreSignificantBits);
450 return SE.getAddRecExpr(Start, Step, AR->getLoop());
455 // Distribute the sdiv over add operands, if the add doesn't overflow.
456 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
457 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
458 SmallVector<const SCEV *, 8> Ops;
459 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
461 const SCEV *Op = getExactSDiv(*I, RHS, SE,
462 IgnoreSignificantBits);
466 return SE.getAddExpr(Ops);
471 // Check for a multiply operand that we can pull RHS out of.
472 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
473 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
474 SmallVector<const SCEV *, 4> Ops;
476 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
480 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
481 IgnoreSignificantBits)) {
487 return Found ? SE.getMulExpr(Ops) : 0;
492 // Otherwise we don't know.
496 /// ExtractImmediate - If S involves the addition of a constant integer value,
497 /// return that integer value, and mutate S to point to a new SCEV with that
499 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
500 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
501 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
502 S = SE.getConstant(C->getType(), 0);
503 return C->getValue()->getSExtValue();
505 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
506 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
507 int64_t Result = ExtractImmediate(NewOps.front(), SE);
508 S = SE.getAddExpr(NewOps);
510 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
511 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
512 int64_t Result = ExtractImmediate(NewOps.front(), SE);
513 S = SE.getAddRecExpr(NewOps, AR->getLoop());
519 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
520 /// return that symbol, and mutate S to point to a new SCEV with that
522 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
523 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
524 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
525 S = SE.getConstant(GV->getType(), 0);
528 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
529 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
530 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
531 S = SE.getAddExpr(NewOps);
533 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
534 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
535 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
536 S = SE.getAddRecExpr(NewOps, AR->getLoop());
542 /// isAddressUse - Returns true if the specified instruction is using the
543 /// specified value as an address.
544 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
545 bool isAddress = isa<LoadInst>(Inst);
546 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
547 if (SI->getOperand(1) == OperandVal)
549 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
550 // Addressing modes can also be folded into prefetches and a variety
552 switch (II->getIntrinsicID()) {
554 case Intrinsic::prefetch:
555 case Intrinsic::x86_sse2_loadu_dq:
556 case Intrinsic::x86_sse2_loadu_pd:
557 case Intrinsic::x86_sse_loadu_ps:
558 case Intrinsic::x86_sse_storeu_ps:
559 case Intrinsic::x86_sse2_storeu_pd:
560 case Intrinsic::x86_sse2_storeu_dq:
561 case Intrinsic::x86_sse2_storel_dq:
562 if (II->getArgOperand(0) == OperandVal)
570 /// getAccessType - Return the type of the memory being accessed.
571 static const Type *getAccessType(const Instruction *Inst) {
572 const Type *AccessTy = Inst->getType();
573 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
574 AccessTy = SI->getOperand(0)->getType();
575 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
576 // Addressing modes can also be folded into prefetches and a variety
578 switch (II->getIntrinsicID()) {
580 case Intrinsic::x86_sse_storeu_ps:
581 case Intrinsic::x86_sse2_storeu_pd:
582 case Intrinsic::x86_sse2_storeu_dq:
583 case Intrinsic::x86_sse2_storel_dq:
584 AccessTy = II->getArgOperand(0)->getType();
589 // All pointers have the same requirements, so canonicalize them to an
590 // arbitrary pointer type to minimize variation.
591 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
592 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
593 PTy->getAddressSpace());
598 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
599 /// specified set are trivially dead, delete them and see if this makes any of
600 /// their operands subsequently dead.
602 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
603 bool Changed = false;
605 while (!DeadInsts.empty()) {
606 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
608 if (I == 0 || !isInstructionTriviallyDead(I))
611 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
612 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
615 DeadInsts.push_back(U);
618 I->eraseFromParent();
627 /// Cost - This class is used to measure and compare candidate formulae.
629 /// TODO: Some of these could be merged. Also, a lexical ordering
630 /// isn't always optimal.
634 unsigned NumBaseAdds;
640 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
643 unsigned getNumRegs() const { return NumRegs; }
645 bool operator<(const Cost &Other) const;
649 void RateFormula(const Formula &F,
650 SmallPtrSet<const SCEV *, 16> &Regs,
651 const DenseSet<const SCEV *> &VisitedRegs,
653 const SmallVectorImpl<int64_t> &Offsets,
654 ScalarEvolution &SE, DominatorTree &DT);
656 void print(raw_ostream &OS) const;
660 void RateRegister(const SCEV *Reg,
661 SmallPtrSet<const SCEV *, 16> &Regs,
663 ScalarEvolution &SE, DominatorTree &DT);
664 void RatePrimaryRegister(const SCEV *Reg,
665 SmallPtrSet<const SCEV *, 16> &Regs,
667 ScalarEvolution &SE, DominatorTree &DT);
672 /// RateRegister - Tally up interesting quantities from the given register.
673 void Cost::RateRegister(const SCEV *Reg,
674 SmallPtrSet<const SCEV *, 16> &Regs,
676 ScalarEvolution &SE, DominatorTree &DT) {
677 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
678 if (AR->getLoop() == L)
679 AddRecCost += 1; /// TODO: This should be a function of the stride.
681 // If this is an addrec for a loop that's already been visited by LSR,
682 // don't second-guess its addrec phi nodes. LSR isn't currently smart
683 // enough to reason about more than one loop at a time. Consider these
684 // registers free and leave them alone.
685 else if (L->contains(AR->getLoop()) ||
686 (!AR->getLoop()->contains(L) &&
687 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
688 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
689 PHINode *PN = dyn_cast<PHINode>(I); ++I)
690 if (SE.isSCEVable(PN->getType()) &&
691 (SE.getEffectiveSCEVType(PN->getType()) ==
692 SE.getEffectiveSCEVType(AR->getType())) &&
693 SE.getSCEV(PN) == AR)
696 // If this isn't one of the addrecs that the loop already has, it
697 // would require a costly new phi and add. TODO: This isn't
698 // precisely modeled right now.
700 if (!Regs.count(AR->getStart()))
701 RateRegister(AR->getStart(), Regs, L, SE, DT);
704 // Add the step value register, if it needs one.
705 // TODO: The non-affine case isn't precisely modeled here.
706 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
707 if (!Regs.count(AR->getStart()))
708 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
712 // Rough heuristic; favor registers which don't require extra setup
713 // instructions in the preheader.
714 if (!isa<SCEVUnknown>(Reg) &&
715 !isa<SCEVConstant>(Reg) &&
716 !(isa<SCEVAddRecExpr>(Reg) &&
717 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
718 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
722 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
724 void Cost::RatePrimaryRegister(const SCEV *Reg,
725 SmallPtrSet<const SCEV *, 16> &Regs,
727 ScalarEvolution &SE, DominatorTree &DT) {
728 if (Regs.insert(Reg))
729 RateRegister(Reg, Regs, L, SE, DT);
732 void Cost::RateFormula(const Formula &F,
733 SmallPtrSet<const SCEV *, 16> &Regs,
734 const DenseSet<const SCEV *> &VisitedRegs,
736 const SmallVectorImpl<int64_t> &Offsets,
737 ScalarEvolution &SE, DominatorTree &DT) {
738 // Tally up the registers.
739 if (const SCEV *ScaledReg = F.ScaledReg) {
740 if (VisitedRegs.count(ScaledReg)) {
744 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
746 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
747 E = F.BaseRegs.end(); I != E; ++I) {
748 const SCEV *BaseReg = *I;
749 if (VisitedRegs.count(BaseReg)) {
753 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
755 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
756 BaseReg->hasComputableLoopEvolution(L);
759 if (F.BaseRegs.size() > 1)
760 NumBaseAdds += F.BaseRegs.size() - 1;
762 // Tally up the non-zero immediates.
763 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
764 E = Offsets.end(); I != E; ++I) {
765 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
767 ImmCost += 64; // Handle symbolic values conservatively.
768 // TODO: This should probably be the pointer size.
769 else if (Offset != 0)
770 ImmCost += APInt(64, Offset, true).getMinSignedBits();
774 /// Loose - Set this cost to a loosing value.
784 /// operator< - Choose the lower cost.
785 bool Cost::operator<(const Cost &Other) const {
786 if (NumRegs != Other.NumRegs)
787 return NumRegs < Other.NumRegs;
788 if (AddRecCost != Other.AddRecCost)
789 return AddRecCost < Other.AddRecCost;
790 if (NumIVMuls != Other.NumIVMuls)
791 return NumIVMuls < Other.NumIVMuls;
792 if (NumBaseAdds != Other.NumBaseAdds)
793 return NumBaseAdds < Other.NumBaseAdds;
794 if (ImmCost != Other.ImmCost)
795 return ImmCost < Other.ImmCost;
796 if (SetupCost != Other.SetupCost)
797 return SetupCost < Other.SetupCost;
801 void Cost::print(raw_ostream &OS) const {
802 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
804 OS << ", with addrec cost " << AddRecCost;
806 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
807 if (NumBaseAdds != 0)
808 OS << ", plus " << NumBaseAdds << " base add"
809 << (NumBaseAdds == 1 ? "" : "s");
811 OS << ", plus " << ImmCost << " imm cost";
813 OS << ", plus " << SetupCost << " setup cost";
816 void Cost::dump() const {
817 print(errs()); errs() << '\n';
822 /// LSRFixup - An operand value in an instruction which is to be replaced
823 /// with some equivalent, possibly strength-reduced, replacement.
825 /// UserInst - The instruction which will be updated.
826 Instruction *UserInst;
828 /// OperandValToReplace - The operand of the instruction which will
829 /// be replaced. The operand may be used more than once; every instance
830 /// will be replaced.
831 Value *OperandValToReplace;
833 /// PostIncLoops - If this user is to use the post-incremented value of an
834 /// induction variable, this variable is non-null and holds the loop
835 /// associated with the induction variable.
836 PostIncLoopSet PostIncLoops;
838 /// LUIdx - The index of the LSRUse describing the expression which
839 /// this fixup needs, minus an offset (below).
842 /// Offset - A constant offset to be added to the LSRUse expression.
843 /// This allows multiple fixups to share the same LSRUse with different
844 /// offsets, for example in an unrolled loop.
847 bool isUseFullyOutsideLoop(const Loop *L) const;
851 void print(raw_ostream &OS) const;
858 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
860 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
861 /// value outside of the given loop.
862 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
863 // PHI nodes use their value in their incoming blocks.
864 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
865 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
866 if (PN->getIncomingValue(i) == OperandValToReplace &&
867 L->contains(PN->getIncomingBlock(i)))
872 return !L->contains(UserInst);
875 void LSRFixup::print(raw_ostream &OS) const {
877 // Store is common and interesting enough to be worth special-casing.
878 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
880 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
881 } else if (UserInst->getType()->isVoidTy())
882 OS << UserInst->getOpcodeName();
884 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
886 OS << ", OperandValToReplace=";
887 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
889 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
890 E = PostIncLoops.end(); I != E; ++I) {
891 OS << ", PostIncLoop=";
892 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
895 if (LUIdx != ~size_t(0))
896 OS << ", LUIdx=" << LUIdx;
899 OS << ", Offset=" << Offset;
902 void LSRFixup::dump() const {
903 print(errs()); errs() << '\n';
908 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
909 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
910 struct UniquifierDenseMapInfo {
911 static SmallVector<const SCEV *, 2> getEmptyKey() {
912 SmallVector<const SCEV *, 2> V;
913 V.push_back(reinterpret_cast<const SCEV *>(-1));
917 static SmallVector<const SCEV *, 2> getTombstoneKey() {
918 SmallVector<const SCEV *, 2> V;
919 V.push_back(reinterpret_cast<const SCEV *>(-2));
923 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
925 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
926 E = V.end(); I != E; ++I)
927 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
931 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
932 const SmallVector<const SCEV *, 2> &RHS) {
937 /// LSRUse - This class holds the state that LSR keeps for each use in
938 /// IVUsers, as well as uses invented by LSR itself. It includes information
939 /// about what kinds of things can be folded into the user, information about
940 /// the user itself, and information about how the use may be satisfied.
941 /// TODO: Represent multiple users of the same expression in common?
943 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
946 /// KindType - An enum for a kind of use, indicating what types of
947 /// scaled and immediate operands it might support.
949 Basic, ///< A normal use, with no folding.
950 Special, ///< A special case of basic, allowing -1 scales.
951 Address, ///< An address use; folding according to TargetLowering
952 ICmpZero ///< An equality icmp with both operands folded into one.
953 // TODO: Add a generic icmp too?
957 const Type *AccessTy;
959 SmallVector<int64_t, 8> Offsets;
963 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
964 /// LSRUse are outside of the loop, in which case some special-case heuristics
966 bool AllFixupsOutsideLoop;
968 /// WidestFixupType - This records the widest use type for any fixup using
969 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
970 /// max fixup widths to be equivalent, because the narrower one may be relying
971 /// on the implicit truncation to truncate away bogus bits.
972 const Type *WidestFixupType;
974 /// Formulae - A list of ways to build a value that can satisfy this user.
975 /// After the list is populated, one of these is selected heuristically and
976 /// used to formulate a replacement for OperandValToReplace in UserInst.
977 SmallVector<Formula, 12> Formulae;
979 /// Regs - The set of register candidates used by all formulae in this LSRUse.
980 SmallPtrSet<const SCEV *, 4> Regs;
982 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
983 MinOffset(INT64_MAX),
984 MaxOffset(INT64_MIN),
985 AllFixupsOutsideLoop(true),
986 WidestFixupType(0) {}
988 bool HasFormulaWithSameRegs(const Formula &F) const;
989 bool InsertFormula(const Formula &F);
990 void DeleteFormula(Formula &F);
991 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
995 void print(raw_ostream &OS) const;
1001 /// HasFormula - Test whether this use as a formula which has the same
1002 /// registers as the given formula.
1003 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1004 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1005 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1006 // Unstable sort by host order ok, because this is only used for uniquifying.
1007 std::sort(Key.begin(), Key.end());
1008 return Uniquifier.count(Key);
1011 /// InsertFormula - If the given formula has not yet been inserted, add it to
1012 /// the list, and return true. Return false otherwise.
1013 bool LSRUse::InsertFormula(const Formula &F) {
1014 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1015 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1016 // Unstable sort by host order ok, because this is only used for uniquifying.
1017 std::sort(Key.begin(), Key.end());
1019 if (!Uniquifier.insert(Key).second)
1022 // Using a register to hold the value of 0 is not profitable.
1023 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1024 "Zero allocated in a scaled register!");
1026 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1027 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1028 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1031 // Add the formula to the list.
1032 Formulae.push_back(F);
1034 // Record registers now being used by this use.
1035 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1036 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1041 /// DeleteFormula - Remove the given formula from this use's list.
1042 void LSRUse::DeleteFormula(Formula &F) {
1043 if (&F != &Formulae.back())
1044 std::swap(F, Formulae.back());
1045 Formulae.pop_back();
1046 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1049 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1050 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1051 // Now that we've filtered out some formulae, recompute the Regs set.
1052 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1054 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1055 E = Formulae.end(); I != E; ++I) {
1056 const Formula &F = *I;
1057 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1058 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1061 // Update the RegTracker.
1062 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1063 E = OldRegs.end(); I != E; ++I)
1064 if (!Regs.count(*I))
1065 RegUses.DropRegister(*I, LUIdx);
1068 void LSRUse::print(raw_ostream &OS) const {
1069 OS << "LSR Use: Kind=";
1071 case Basic: OS << "Basic"; break;
1072 case Special: OS << "Special"; break;
1073 case ICmpZero: OS << "ICmpZero"; break;
1075 OS << "Address of ";
1076 if (AccessTy->isPointerTy())
1077 OS << "pointer"; // the full pointer type could be really verbose
1082 OS << ", Offsets={";
1083 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1084 E = Offsets.end(); I != E; ++I) {
1086 if (llvm::next(I) != E)
1091 if (AllFixupsOutsideLoop)
1092 OS << ", all-fixups-outside-loop";
1094 if (WidestFixupType)
1095 OS << ", widest fixup type: " << *WidestFixupType;
1098 void LSRUse::dump() const {
1099 print(errs()); errs() << '\n';
1102 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1103 /// be completely folded into the user instruction at isel time. This includes
1104 /// address-mode folding and special icmp tricks.
1105 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1106 LSRUse::KindType Kind, const Type *AccessTy,
1107 const TargetLowering *TLI) {
1109 case LSRUse::Address:
1110 // If we have low-level target information, ask the target if it can
1111 // completely fold this address.
1112 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1114 // Otherwise, just guess that reg+reg addressing is legal.
1115 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1117 case LSRUse::ICmpZero:
1118 // There's not even a target hook for querying whether it would be legal to
1119 // fold a GV into an ICmp.
1123 // ICmp only has two operands; don't allow more than two non-trivial parts.
1124 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1127 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1128 // putting the scaled register in the other operand of the icmp.
1129 if (AM.Scale != 0 && AM.Scale != -1)
1132 // If we have low-level target information, ask the target if it can fold an
1133 // integer immediate on an icmp.
1134 if (AM.BaseOffs != 0) {
1135 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1142 // Only handle single-register values.
1143 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1145 case LSRUse::Special:
1146 // Only handle -1 scales, or no scale.
1147 return AM.Scale == 0 || AM.Scale == -1;
1153 static bool isLegalUse(TargetLowering::AddrMode AM,
1154 int64_t MinOffset, int64_t MaxOffset,
1155 LSRUse::KindType Kind, const Type *AccessTy,
1156 const TargetLowering *TLI) {
1157 // Check for overflow.
1158 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1161 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1162 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1163 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1164 // Check for overflow.
1165 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1168 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1169 return isLegalUse(AM, Kind, AccessTy, TLI);
1174 static bool isAlwaysFoldable(int64_t BaseOffs,
1175 GlobalValue *BaseGV,
1177 LSRUse::KindType Kind, const Type *AccessTy,
1178 const TargetLowering *TLI) {
1179 // Fast-path: zero is always foldable.
1180 if (BaseOffs == 0 && !BaseGV) return true;
1182 // Conservatively, create an address with an immediate and a
1183 // base and a scale.
1184 TargetLowering::AddrMode AM;
1185 AM.BaseOffs = BaseOffs;
1187 AM.HasBaseReg = HasBaseReg;
1188 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1190 // Canonicalize a scale of 1 to a base register if the formula doesn't
1191 // already have a base register.
1192 if (!AM.HasBaseReg && AM.Scale == 1) {
1194 AM.HasBaseReg = true;
1197 return isLegalUse(AM, Kind, AccessTy, TLI);
1200 static bool isAlwaysFoldable(const SCEV *S,
1201 int64_t MinOffset, int64_t MaxOffset,
1203 LSRUse::KindType Kind, const Type *AccessTy,
1204 const TargetLowering *TLI,
1205 ScalarEvolution &SE) {
1206 // Fast-path: zero is always foldable.
1207 if (S->isZero()) return true;
1209 // Conservatively, create an address with an immediate and a
1210 // base and a scale.
1211 int64_t BaseOffs = ExtractImmediate(S, SE);
1212 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1214 // If there's anything else involved, it's not foldable.
1215 if (!S->isZero()) return false;
1217 // Fast-path: zero is always foldable.
1218 if (BaseOffs == 0 && !BaseGV) return true;
1220 // Conservatively, create an address with an immediate and a
1221 // base and a scale.
1222 TargetLowering::AddrMode AM;
1223 AM.BaseOffs = BaseOffs;
1225 AM.HasBaseReg = HasBaseReg;
1226 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1228 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1233 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1234 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1235 struct UseMapDenseMapInfo {
1236 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1237 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1240 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1241 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1245 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1246 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1247 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1251 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1252 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1257 /// FormulaSorter - This class implements an ordering for formulae which sorts
1258 /// the by their standalone cost.
1259 class FormulaSorter {
1260 /// These two sets are kept empty, so that we compute standalone costs.
1261 DenseSet<const SCEV *> VisitedRegs;
1262 SmallPtrSet<const SCEV *, 16> Regs;
1265 ScalarEvolution &SE;
1269 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1270 : L(l), LU(&lu), SE(se), DT(dt) {}
1272 bool operator()(const Formula &A, const Formula &B) {
1274 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1277 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1279 return CostA < CostB;
1283 /// LSRInstance - This class holds state for the main loop strength reduction
1287 ScalarEvolution &SE;
1290 const TargetLowering *const TLI;
1294 /// IVIncInsertPos - This is the insert position that the current loop's
1295 /// induction variable increment should be placed. In simple loops, this is
1296 /// the latch block's terminator. But in more complicated cases, this is a
1297 /// position which will dominate all the in-loop post-increment users.
1298 Instruction *IVIncInsertPos;
1300 /// Factors - Interesting factors between use strides.
1301 SmallSetVector<int64_t, 8> Factors;
1303 /// Types - Interesting use types, to facilitate truncation reuse.
1304 SmallSetVector<const Type *, 4> Types;
1306 /// Fixups - The list of operands which are to be replaced.
1307 SmallVector<LSRFixup, 16> Fixups;
1309 /// Uses - The list of interesting uses.
1310 SmallVector<LSRUse, 16> Uses;
1312 /// RegUses - Track which uses use which register candidates.
1313 RegUseTracker RegUses;
1315 void OptimizeShadowIV();
1316 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1317 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1318 void OptimizeLoopTermCond();
1320 void CollectInterestingTypesAndFactors();
1321 void CollectFixupsAndInitialFormulae();
1323 LSRFixup &getNewFixup() {
1324 Fixups.push_back(LSRFixup());
1325 return Fixups.back();
1328 // Support for sharing of LSRUses between LSRFixups.
1329 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1331 UseMapDenseMapInfo> UseMapTy;
1334 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1335 LSRUse::KindType Kind, const Type *AccessTy);
1337 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1338 LSRUse::KindType Kind,
1339 const Type *AccessTy);
1341 void DeleteUse(LSRUse &LU);
1343 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1346 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1347 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1348 void CountRegisters(const Formula &F, size_t LUIdx);
1349 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1351 void CollectLoopInvariantFixupsAndFormulae();
1353 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1354 unsigned Depth = 0);
1355 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1356 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1357 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1358 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1359 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1360 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1361 void GenerateCrossUseConstantOffsets();
1362 void GenerateAllReuseFormulae();
1364 void FilterOutUndesirableDedicatedRegisters();
1366 size_t EstimateSearchSpaceComplexity() const;
1367 void NarrowSearchSpaceUsingHeuristics();
1369 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1371 SmallVectorImpl<const Formula *> &Workspace,
1372 const Cost &CurCost,
1373 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1374 DenseSet<const SCEV *> &VisitedRegs) const;
1375 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1377 BasicBlock::iterator
1378 HoistInsertPosition(BasicBlock::iterator IP,
1379 const SmallVectorImpl<Instruction *> &Inputs) const;
1380 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1382 const LSRUse &LU) const;
1384 Value *Expand(const LSRFixup &LF,
1386 BasicBlock::iterator IP,
1387 SCEVExpander &Rewriter,
1388 SmallVectorImpl<WeakVH> &DeadInsts) const;
1389 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1391 SCEVExpander &Rewriter,
1392 SmallVectorImpl<WeakVH> &DeadInsts,
1394 void Rewrite(const LSRFixup &LF,
1396 SCEVExpander &Rewriter,
1397 SmallVectorImpl<WeakVH> &DeadInsts,
1399 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1402 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1404 bool getChanged() const { return Changed; }
1406 void print_factors_and_types(raw_ostream &OS) const;
1407 void print_fixups(raw_ostream &OS) const;
1408 void print_uses(raw_ostream &OS) const;
1409 void print(raw_ostream &OS) const;
1415 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1416 /// inside the loop then try to eliminate the cast operation.
1417 void LSRInstance::OptimizeShadowIV() {
1418 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1419 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1422 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1423 UI != E; /* empty */) {
1424 IVUsers::const_iterator CandidateUI = UI;
1426 Instruction *ShadowUse = CandidateUI->getUser();
1427 const Type *DestTy = NULL;
1429 /* If shadow use is a int->float cast then insert a second IV
1430 to eliminate this cast.
1432 for (unsigned i = 0; i < n; ++i)
1438 for (unsigned i = 0; i < n; ++i, ++d)
1441 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1442 DestTy = UCast->getDestTy();
1443 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1444 DestTy = SCast->getDestTy();
1445 if (!DestTy) continue;
1448 // If target does not support DestTy natively then do not apply
1449 // this transformation.
1450 EVT DVT = TLI->getValueType(DestTy);
1451 if (!TLI->isTypeLegal(DVT)) continue;
1454 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1456 if (PH->getNumIncomingValues() != 2) continue;
1458 const Type *SrcTy = PH->getType();
1459 int Mantissa = DestTy->getFPMantissaWidth();
1460 if (Mantissa == -1) continue;
1461 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1464 unsigned Entry, Latch;
1465 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1473 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1474 if (!Init) continue;
1475 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1477 BinaryOperator *Incr =
1478 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1479 if (!Incr) continue;
1480 if (Incr->getOpcode() != Instruction::Add
1481 && Incr->getOpcode() != Instruction::Sub)
1484 /* Initialize new IV, double d = 0.0 in above example. */
1485 ConstantInt *C = NULL;
1486 if (Incr->getOperand(0) == PH)
1487 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1488 else if (Incr->getOperand(1) == PH)
1489 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1495 // Ignore negative constants, as the code below doesn't handle them
1496 // correctly. TODO: Remove this restriction.
1497 if (!C->getValue().isStrictlyPositive()) continue;
1499 /* Add new PHINode. */
1500 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1502 /* create new increment. '++d' in above example. */
1503 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1504 BinaryOperator *NewIncr =
1505 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1506 Instruction::FAdd : Instruction::FSub,
1507 NewPH, CFP, "IV.S.next.", Incr);
1509 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1510 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1512 /* Remove cast operation */
1513 ShadowUse->replaceAllUsesWith(NewPH);
1514 ShadowUse->eraseFromParent();
1520 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1521 /// set the IV user and stride information and return true, otherwise return
1523 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1524 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1525 if (UI->getUser() == Cond) {
1526 // NOTE: we could handle setcc instructions with multiple uses here, but
1527 // InstCombine does it as well for simple uses, it's not clear that it
1528 // occurs enough in real life to handle.
1535 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1536 /// a max computation.
1538 /// This is a narrow solution to a specific, but acute, problem. For loops
1544 /// } while (++i < n);
1546 /// the trip count isn't just 'n', because 'n' might not be positive. And
1547 /// unfortunately this can come up even for loops where the user didn't use
1548 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1549 /// will commonly be lowered like this:
1555 /// } while (++i < n);
1558 /// and then it's possible for subsequent optimization to obscure the if
1559 /// test in such a way that indvars can't find it.
1561 /// When indvars can't find the if test in loops like this, it creates a
1562 /// max expression, which allows it to give the loop a canonical
1563 /// induction variable:
1566 /// max = n < 1 ? 1 : n;
1569 /// } while (++i != max);
1571 /// Canonical induction variables are necessary because the loop passes
1572 /// are designed around them. The most obvious example of this is the
1573 /// LoopInfo analysis, which doesn't remember trip count values. It
1574 /// expects to be able to rediscover the trip count each time it is
1575 /// needed, and it does this using a simple analysis that only succeeds if
1576 /// the loop has a canonical induction variable.
1578 /// However, when it comes time to generate code, the maximum operation
1579 /// can be quite costly, especially if it's inside of an outer loop.
1581 /// This function solves this problem by detecting this type of loop and
1582 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1583 /// the instructions for the maximum computation.
1585 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1586 // Check that the loop matches the pattern we're looking for.
1587 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1588 Cond->getPredicate() != CmpInst::ICMP_NE)
1591 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1592 if (!Sel || !Sel->hasOneUse()) return Cond;
1594 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1595 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1597 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1599 // Add one to the backedge-taken count to get the trip count.
1600 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1601 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1603 // Check for a max calculation that matches the pattern. There's no check
1604 // for ICMP_ULE here because the comparison would be with zero, which
1605 // isn't interesting.
1606 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1607 const SCEVNAryExpr *Max = 0;
1608 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1609 Pred = ICmpInst::ICMP_SLE;
1611 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1612 Pred = ICmpInst::ICMP_SLT;
1614 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1615 Pred = ICmpInst::ICMP_ULT;
1622 // To handle a max with more than two operands, this optimization would
1623 // require additional checking and setup.
1624 if (Max->getNumOperands() != 2)
1627 const SCEV *MaxLHS = Max->getOperand(0);
1628 const SCEV *MaxRHS = Max->getOperand(1);
1630 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1631 // for a comparison with 1. For <= and >=, a comparison with zero.
1633 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1636 // Check the relevant induction variable for conformance to
1638 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1639 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1640 if (!AR || !AR->isAffine() ||
1641 AR->getStart() != One ||
1642 AR->getStepRecurrence(SE) != One)
1645 assert(AR->getLoop() == L &&
1646 "Loop condition operand is an addrec in a different loop!");
1648 // Check the right operand of the select, and remember it, as it will
1649 // be used in the new comparison instruction.
1651 if (ICmpInst::isTrueWhenEqual(Pred)) {
1652 // Look for n+1, and grab n.
1653 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1654 if (isa<ConstantInt>(BO->getOperand(1)) &&
1655 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1656 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1657 NewRHS = BO->getOperand(0);
1658 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1659 if (isa<ConstantInt>(BO->getOperand(1)) &&
1660 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1661 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1662 NewRHS = BO->getOperand(0);
1665 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1666 NewRHS = Sel->getOperand(1);
1667 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1668 NewRHS = Sel->getOperand(2);
1669 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1670 NewRHS = SU->getValue();
1672 // Max doesn't match expected pattern.
1675 // Determine the new comparison opcode. It may be signed or unsigned,
1676 // and the original comparison may be either equality or inequality.
1677 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1678 Pred = CmpInst::getInversePredicate(Pred);
1680 // Ok, everything looks ok to change the condition into an SLT or SGE and
1681 // delete the max calculation.
1683 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1685 // Delete the max calculation instructions.
1686 Cond->replaceAllUsesWith(NewCond);
1687 CondUse->setUser(NewCond);
1688 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1689 Cond->eraseFromParent();
1690 Sel->eraseFromParent();
1691 if (Cmp->use_empty())
1692 Cmp->eraseFromParent();
1696 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1697 /// postinc iv when possible.
1699 LSRInstance::OptimizeLoopTermCond() {
1700 SmallPtrSet<Instruction *, 4> PostIncs;
1702 BasicBlock *LatchBlock = L->getLoopLatch();
1703 SmallVector<BasicBlock*, 8> ExitingBlocks;
1704 L->getExitingBlocks(ExitingBlocks);
1706 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1707 BasicBlock *ExitingBlock = ExitingBlocks[i];
1709 // Get the terminating condition for the loop if possible. If we
1710 // can, we want to change it to use a post-incremented version of its
1711 // induction variable, to allow coalescing the live ranges for the IV into
1712 // one register value.
1714 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1717 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1718 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1721 // Search IVUsesByStride to find Cond's IVUse if there is one.
1722 IVStrideUse *CondUse = 0;
1723 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1724 if (!FindIVUserForCond(Cond, CondUse))
1727 // If the trip count is computed in terms of a max (due to ScalarEvolution
1728 // being unable to find a sufficient guard, for example), change the loop
1729 // comparison to use SLT or ULT instead of NE.
1730 // One consequence of doing this now is that it disrupts the count-down
1731 // optimization. That's not always a bad thing though, because in such
1732 // cases it may still be worthwhile to avoid a max.
1733 Cond = OptimizeMax(Cond, CondUse);
1735 // If this exiting block dominates the latch block, it may also use
1736 // the post-inc value if it won't be shared with other uses.
1737 // Check for dominance.
1738 if (!DT.dominates(ExitingBlock, LatchBlock))
1741 // Conservatively avoid trying to use the post-inc value in non-latch
1742 // exits if there may be pre-inc users in intervening blocks.
1743 if (LatchBlock != ExitingBlock)
1744 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1745 // Test if the use is reachable from the exiting block. This dominator
1746 // query is a conservative approximation of reachability.
1747 if (&*UI != CondUse &&
1748 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1749 // Conservatively assume there may be reuse if the quotient of their
1750 // strides could be a legal scale.
1751 const SCEV *A = IU.getStride(*CondUse, L);
1752 const SCEV *B = IU.getStride(*UI, L);
1753 if (!A || !B) continue;
1754 if (SE.getTypeSizeInBits(A->getType()) !=
1755 SE.getTypeSizeInBits(B->getType())) {
1756 if (SE.getTypeSizeInBits(A->getType()) >
1757 SE.getTypeSizeInBits(B->getType()))
1758 B = SE.getSignExtendExpr(B, A->getType());
1760 A = SE.getSignExtendExpr(A, B->getType());
1762 if (const SCEVConstant *D =
1763 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1764 const ConstantInt *C = D->getValue();
1765 // Stride of one or negative one can have reuse with non-addresses.
1766 if (C->isOne() || C->isAllOnesValue())
1767 goto decline_post_inc;
1768 // Avoid weird situations.
1769 if (C->getValue().getMinSignedBits() >= 64 ||
1770 C->getValue().isMinSignedValue())
1771 goto decline_post_inc;
1772 // Without TLI, assume that any stride might be valid, and so any
1773 // use might be shared.
1775 goto decline_post_inc;
1776 // Check for possible scaled-address reuse.
1777 const Type *AccessTy = getAccessType(UI->getUser());
1778 TargetLowering::AddrMode AM;
1779 AM.Scale = C->getSExtValue();
1780 if (TLI->isLegalAddressingMode(AM, AccessTy))
1781 goto decline_post_inc;
1782 AM.Scale = -AM.Scale;
1783 if (TLI->isLegalAddressingMode(AM, AccessTy))
1784 goto decline_post_inc;
1788 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1791 // It's possible for the setcc instruction to be anywhere in the loop, and
1792 // possible for it to have multiple users. If it is not immediately before
1793 // the exiting block branch, move it.
1794 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1795 if (Cond->hasOneUse()) {
1796 Cond->moveBefore(TermBr);
1798 // Clone the terminating condition and insert into the loopend.
1799 ICmpInst *OldCond = Cond;
1800 Cond = cast<ICmpInst>(Cond->clone());
1801 Cond->setName(L->getHeader()->getName() + ".termcond");
1802 ExitingBlock->getInstList().insert(TermBr, Cond);
1804 // Clone the IVUse, as the old use still exists!
1805 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1806 TermBr->replaceUsesOfWith(OldCond, Cond);
1810 // If we get to here, we know that we can transform the setcc instruction to
1811 // use the post-incremented version of the IV, allowing us to coalesce the
1812 // live ranges for the IV correctly.
1813 CondUse->transformToPostInc(L);
1816 PostIncs.insert(Cond);
1820 // Determine an insertion point for the loop induction variable increment. It
1821 // must dominate all the post-inc comparisons we just set up, and it must
1822 // dominate the loop latch edge.
1823 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1824 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1825 E = PostIncs.end(); I != E; ++I) {
1827 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1829 if (BB == (*I)->getParent())
1830 IVIncInsertPos = *I;
1831 else if (BB != IVIncInsertPos->getParent())
1832 IVIncInsertPos = BB->getTerminator();
1836 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1837 /// at the given offset and other details. If so, update the use and
1840 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1841 LSRUse::KindType Kind, const Type *AccessTy) {
1842 int64_t NewMinOffset = LU.MinOffset;
1843 int64_t NewMaxOffset = LU.MaxOffset;
1844 const Type *NewAccessTy = AccessTy;
1846 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1847 // something conservative, however this can pessimize in the case that one of
1848 // the uses will have all its uses outside the loop, for example.
1849 if (LU.Kind != Kind)
1851 // Conservatively assume HasBaseReg is true for now.
1852 if (NewOffset < LU.MinOffset) {
1853 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1854 Kind, AccessTy, TLI))
1856 NewMinOffset = NewOffset;
1857 } else if (NewOffset > LU.MaxOffset) {
1858 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1859 Kind, AccessTy, TLI))
1861 NewMaxOffset = NewOffset;
1863 // Check for a mismatched access type, and fall back conservatively as needed.
1864 // TODO: Be less conservative when the type is similar and can use the same
1865 // addressing modes.
1866 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1867 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1870 LU.MinOffset = NewMinOffset;
1871 LU.MaxOffset = NewMaxOffset;
1872 LU.AccessTy = NewAccessTy;
1873 if (NewOffset != LU.Offsets.back())
1874 LU.Offsets.push_back(NewOffset);
1878 /// getUse - Return an LSRUse index and an offset value for a fixup which
1879 /// needs the given expression, with the given kind and optional access type.
1880 /// Either reuse an existing use or create a new one, as needed.
1881 std::pair<size_t, int64_t>
1882 LSRInstance::getUse(const SCEV *&Expr,
1883 LSRUse::KindType Kind, const Type *AccessTy) {
1884 const SCEV *Copy = Expr;
1885 int64_t Offset = ExtractImmediate(Expr, SE);
1887 // Basic uses can't accept any offset, for example.
1888 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1893 std::pair<UseMapTy::iterator, bool> P =
1894 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1896 // A use already existed with this base.
1897 size_t LUIdx = P.first->second;
1898 LSRUse &LU = Uses[LUIdx];
1899 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1901 return std::make_pair(LUIdx, Offset);
1904 // Create a new use.
1905 size_t LUIdx = Uses.size();
1906 P.first->second = LUIdx;
1907 Uses.push_back(LSRUse(Kind, AccessTy));
1908 LSRUse &LU = Uses[LUIdx];
1910 // We don't need to track redundant offsets, but we don't need to go out
1911 // of our way here to avoid them.
1912 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1913 LU.Offsets.push_back(Offset);
1915 LU.MinOffset = Offset;
1916 LU.MaxOffset = Offset;
1917 return std::make_pair(LUIdx, Offset);
1920 /// DeleteUse - Delete the given use from the Uses list.
1921 void LSRInstance::DeleteUse(LSRUse &LU) {
1922 if (&LU != &Uses.back())
1923 std::swap(LU, Uses.back());
1927 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1928 /// a formula that has the same registers as the given formula.
1930 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1931 const LSRUse &OrigLU) {
1932 // Search all uses for the formula. This could be more clever. Ignore
1933 // ICmpZero uses because they may contain formulae generated by
1934 // GenerateICmpZeroScales, in which case adding fixup offsets may
1936 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1937 LSRUse &LU = Uses[LUIdx];
1938 if (&LU != &OrigLU &&
1939 LU.Kind != LSRUse::ICmpZero &&
1940 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1941 LU.WidestFixupType == OrigLU.WidestFixupType &&
1942 LU.HasFormulaWithSameRegs(OrigF)) {
1943 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1944 E = LU.Formulae.end(); I != E; ++I) {
1945 const Formula &F = *I;
1946 if (F.BaseRegs == OrigF.BaseRegs &&
1947 F.ScaledReg == OrigF.ScaledReg &&
1948 F.AM.BaseGV == OrigF.AM.BaseGV &&
1949 F.AM.Scale == OrigF.AM.Scale &&
1951 if (F.AM.BaseOffs == 0)
1962 void LSRInstance::CollectInterestingTypesAndFactors() {
1963 SmallSetVector<const SCEV *, 4> Strides;
1965 // Collect interesting types and strides.
1966 SmallVector<const SCEV *, 4> Worklist;
1967 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1968 const SCEV *Expr = IU.getExpr(*UI);
1970 // Collect interesting types.
1971 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1973 // Add strides for mentioned loops.
1974 Worklist.push_back(Expr);
1976 const SCEV *S = Worklist.pop_back_val();
1977 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1978 Strides.insert(AR->getStepRecurrence(SE));
1979 Worklist.push_back(AR->getStart());
1980 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1981 Worklist.append(Add->op_begin(), Add->op_end());
1983 } while (!Worklist.empty());
1986 // Compute interesting factors from the set of interesting strides.
1987 for (SmallSetVector<const SCEV *, 4>::const_iterator
1988 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1989 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1990 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
1991 const SCEV *OldStride = *I;
1992 const SCEV *NewStride = *NewStrideIter;
1994 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1995 SE.getTypeSizeInBits(NewStride->getType())) {
1996 if (SE.getTypeSizeInBits(OldStride->getType()) >
1997 SE.getTypeSizeInBits(NewStride->getType()))
1998 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2000 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2002 if (const SCEVConstant *Factor =
2003 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2005 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2006 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2007 } else if (const SCEVConstant *Factor =
2008 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2011 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2012 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2016 // If all uses use the same type, don't bother looking for truncation-based
2018 if (Types.size() == 1)
2021 DEBUG(print_factors_and_types(dbgs()));
2024 void LSRInstance::CollectFixupsAndInitialFormulae() {
2025 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2027 LSRFixup &LF = getNewFixup();
2028 LF.UserInst = UI->getUser();
2029 LF.OperandValToReplace = UI->getOperandValToReplace();
2030 LF.PostIncLoops = UI->getPostIncLoops();
2032 LSRUse::KindType Kind = LSRUse::Basic;
2033 const Type *AccessTy = 0;
2034 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2035 Kind = LSRUse::Address;
2036 AccessTy = getAccessType(LF.UserInst);
2039 const SCEV *S = IU.getExpr(*UI);
2041 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2042 // (N - i == 0), and this allows (N - i) to be the expression that we work
2043 // with rather than just N or i, so we can consider the register
2044 // requirements for both N and i at the same time. Limiting this code to
2045 // equality icmps is not a problem because all interesting loops use
2046 // equality icmps, thanks to IndVarSimplify.
2047 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2048 if (CI->isEquality()) {
2049 // Swap the operands if needed to put the OperandValToReplace on the
2050 // left, for consistency.
2051 Value *NV = CI->getOperand(1);
2052 if (NV == LF.OperandValToReplace) {
2053 CI->setOperand(1, CI->getOperand(0));
2054 CI->setOperand(0, NV);
2055 NV = CI->getOperand(1);
2059 // x == y --> x - y == 0
2060 const SCEV *N = SE.getSCEV(NV);
2061 if (N->isLoopInvariant(L)) {
2062 Kind = LSRUse::ICmpZero;
2063 S = SE.getMinusSCEV(N, S);
2066 // -1 and the negations of all interesting strides (except the negation
2067 // of -1) are now also interesting.
2068 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2069 if (Factors[i] != -1)
2070 Factors.insert(-(uint64_t)Factors[i]);
2074 // Set up the initial formula for this use.
2075 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2077 LF.Offset = P.second;
2078 LSRUse &LU = Uses[LF.LUIdx];
2079 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2080 if (!LU.WidestFixupType ||
2081 SE.getTypeSizeInBits(LU.WidestFixupType) <
2082 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2083 LU.WidestFixupType = LF.OperandValToReplace->getType();
2085 // If this is the first use of this LSRUse, give it a formula.
2086 if (LU.Formulae.empty()) {
2087 InsertInitialFormula(S, LU, LF.LUIdx);
2088 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2092 DEBUG(print_fixups(dbgs()));
2095 /// InsertInitialFormula - Insert a formula for the given expression into
2096 /// the given use, separating out loop-variant portions from loop-invariant
2097 /// and loop-computable portions.
2099 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2101 F.InitialMatch(S, L, SE, DT);
2102 bool Inserted = InsertFormula(LU, LUIdx, F);
2103 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2106 /// InsertSupplementalFormula - Insert a simple single-register formula for
2107 /// the given expression into the given use.
2109 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2110 LSRUse &LU, size_t LUIdx) {
2112 F.BaseRegs.push_back(S);
2113 F.AM.HasBaseReg = true;
2114 bool Inserted = InsertFormula(LU, LUIdx, F);
2115 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2118 /// CountRegisters - Note which registers are used by the given formula,
2119 /// updating RegUses.
2120 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2122 RegUses.CountRegister(F.ScaledReg, LUIdx);
2123 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2124 E = F.BaseRegs.end(); I != E; ++I)
2125 RegUses.CountRegister(*I, LUIdx);
2128 /// InsertFormula - If the given formula has not yet been inserted, add it to
2129 /// the list, and return true. Return false otherwise.
2130 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2131 if (!LU.InsertFormula(F))
2134 CountRegisters(F, LUIdx);
2138 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2139 /// loop-invariant values which we're tracking. These other uses will pin these
2140 /// values in registers, making them less profitable for elimination.
2141 /// TODO: This currently misses non-constant addrec step registers.
2142 /// TODO: Should this give more weight to users inside the loop?
2144 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2145 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2146 SmallPtrSet<const SCEV *, 8> Inserted;
2148 while (!Worklist.empty()) {
2149 const SCEV *S = Worklist.pop_back_val();
2151 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2152 Worklist.append(N->op_begin(), N->op_end());
2153 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2154 Worklist.push_back(C->getOperand());
2155 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2156 Worklist.push_back(D->getLHS());
2157 Worklist.push_back(D->getRHS());
2158 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2159 if (!Inserted.insert(U)) continue;
2160 const Value *V = U->getValue();
2161 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2162 // Look for instructions defined outside the loop.
2163 if (L->contains(Inst)) continue;
2164 } else if (isa<UndefValue>(V))
2165 // Undef doesn't have a live range, so it doesn't matter.
2167 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2169 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2170 // Ignore non-instructions.
2173 // Ignore instructions in other functions (as can happen with
2175 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2177 // Ignore instructions not dominated by the loop.
2178 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2179 UserInst->getParent() :
2180 cast<PHINode>(UserInst)->getIncomingBlock(
2181 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2182 if (!DT.dominates(L->getHeader(), UseBB))
2184 // Ignore uses which are part of other SCEV expressions, to avoid
2185 // analyzing them multiple times.
2186 if (SE.isSCEVable(UserInst->getType())) {
2187 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2188 // If the user is a no-op, look through to its uses.
2189 if (!isa<SCEVUnknown>(UserS))
2193 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2197 // Ignore icmp instructions which are already being analyzed.
2198 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2199 unsigned OtherIdx = !UI.getOperandNo();
2200 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2201 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2205 LSRFixup &LF = getNewFixup();
2206 LF.UserInst = const_cast<Instruction *>(UserInst);
2207 LF.OperandValToReplace = UI.getUse();
2208 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2210 LF.Offset = P.second;
2211 LSRUse &LU = Uses[LF.LUIdx];
2212 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2213 if (!LU.WidestFixupType ||
2214 SE.getTypeSizeInBits(LU.WidestFixupType) <
2215 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2216 LU.WidestFixupType = LF.OperandValToReplace->getType();
2217 InsertSupplementalFormula(U, LU, LF.LUIdx);
2218 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2225 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2226 /// separate registers. If C is non-null, multiply each subexpression by C.
2227 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2228 SmallVectorImpl<const SCEV *> &Ops,
2229 SmallVectorImpl<const SCEV *> &UninterestingOps,
2231 ScalarEvolution &SE) {
2232 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2233 // Break out add operands.
2234 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2236 CollectSubexprs(*I, C, Ops, UninterestingOps, L, SE);
2238 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2239 // Split a non-zero base out of an addrec.
2240 if (!AR->getStart()->isZero()) {
2241 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2242 AR->getStepRecurrence(SE),
2244 C, Ops, UninterestingOps, L, SE);
2245 CollectSubexprs(AR->getStart(), C, Ops, UninterestingOps, L, SE);
2248 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2249 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2250 if (Mul->getNumOperands() == 2)
2251 if (const SCEVConstant *Op0 =
2252 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2253 CollectSubexprs(Mul->getOperand(1),
2254 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2255 Ops, UninterestingOps, L, SE);
2260 // Otherwise use the value itself. Loop-variant "unknown" values are
2261 // uninteresting; we won't be able to do anything meaningful with them.
2262 if (!C && isa<SCEVUnknown>(S) && !S->isLoopInvariant(L))
2263 UninterestingOps.push_back(S);
2265 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2268 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2270 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2273 // Arbitrarily cap recursion to protect compile time.
2274 if (Depth >= 3) return;
2276 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2277 const SCEV *BaseReg = Base.BaseRegs[i];
2279 SmallVector<const SCEV *, 8> AddOps, UninterestingAddOps;
2280 CollectSubexprs(BaseReg, 0, AddOps, UninterestingAddOps, L, SE);
2282 // Add any uninteresting values as one register, as we won't be able to
2283 // form any interesting reassociation opportunities with them. They'll
2284 // just have to be added inside the loop no matter what we do.
2285 if (!UninterestingAddOps.empty())
2286 AddOps.push_back(SE.getAddExpr(UninterestingAddOps));
2288 if (AddOps.size() == 1) continue;
2290 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2291 JE = AddOps.end(); J != JE; ++J) {
2292 // Don't pull a constant into a register if the constant could be folded
2293 // into an immediate field.
2294 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2295 Base.getNumRegs() > 1,
2296 LU.Kind, LU.AccessTy, TLI, SE))
2299 // Collect all operands except *J.
2300 SmallVector<const SCEV *, 8> InnerAddOps
2301 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2303 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2305 // Don't leave just a constant behind in a register if the constant could
2306 // be folded into an immediate field.
2307 if (InnerAddOps.size() == 1 &&
2308 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2309 Base.getNumRegs() > 1,
2310 LU.Kind, LU.AccessTy, TLI, SE))
2313 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2314 if (InnerSum->isZero())
2317 F.BaseRegs[i] = InnerSum;
2318 F.BaseRegs.push_back(*J);
2319 if (InsertFormula(LU, LUIdx, F))
2320 // If that formula hadn't been seen before, recurse to find more like
2322 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2327 /// GenerateCombinations - Generate a formula consisting of all of the
2328 /// loop-dominating registers added into a single register.
2329 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2331 // This method is only interesting on a plurality of registers.
2332 if (Base.BaseRegs.size() <= 1) return;
2336 SmallVector<const SCEV *, 4> Ops;
2337 for (SmallVectorImpl<const SCEV *>::const_iterator
2338 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2339 const SCEV *BaseReg = *I;
2340 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2341 !BaseReg->hasComputableLoopEvolution(L))
2342 Ops.push_back(BaseReg);
2344 F.BaseRegs.push_back(BaseReg);
2346 if (Ops.size() > 1) {
2347 const SCEV *Sum = SE.getAddExpr(Ops);
2348 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2349 // opportunity to fold something. For now, just ignore such cases
2350 // rather than proceed with zero in a register.
2351 if (!Sum->isZero()) {
2352 F.BaseRegs.push_back(Sum);
2353 (void)InsertFormula(LU, LUIdx, F);
2358 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2359 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2361 // We can't add a symbolic offset if the address already contains one.
2362 if (Base.AM.BaseGV) return;
2364 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2365 const SCEV *G = Base.BaseRegs[i];
2366 GlobalValue *GV = ExtractSymbol(G, SE);
2367 if (G->isZero() || !GV)
2371 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2372 LU.Kind, LU.AccessTy, TLI))
2375 (void)InsertFormula(LU, LUIdx, F);
2379 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2380 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2382 // TODO: For now, just add the min and max offset, because it usually isn't
2383 // worthwhile looking at everything inbetween.
2384 SmallVector<int64_t, 2> Worklist;
2385 Worklist.push_back(LU.MinOffset);
2386 if (LU.MaxOffset != LU.MinOffset)
2387 Worklist.push_back(LU.MaxOffset);
2389 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2390 const SCEV *G = Base.BaseRegs[i];
2392 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2393 E = Worklist.end(); I != E; ++I) {
2395 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2396 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2397 LU.Kind, LU.AccessTy, TLI)) {
2398 // Add the offset to the base register.
2399 const SCEV *NewG = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2400 // If it cancelled out, drop the base register, otherwise update it.
2401 if (NewG->isZero()) {
2402 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2403 F.BaseRegs.pop_back();
2405 F.BaseRegs[i] = NewG;
2407 (void)InsertFormula(LU, LUIdx, F);
2411 int64_t Imm = ExtractImmediate(G, SE);
2412 if (G->isZero() || Imm == 0)
2415 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2416 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2417 LU.Kind, LU.AccessTy, TLI))
2420 (void)InsertFormula(LU, LUIdx, F);
2424 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2425 /// the comparison. For example, x == y -> x*c == y*c.
2426 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2428 if (LU.Kind != LSRUse::ICmpZero) return;
2430 // Determine the integer type for the base formula.
2431 const Type *IntTy = Base.getType();
2433 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2435 // Don't do this if there is more than one offset.
2436 if (LU.MinOffset != LU.MaxOffset) return;
2438 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2440 // Check each interesting stride.
2441 for (SmallSetVector<int64_t, 8>::const_iterator
2442 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2443 int64_t Factor = *I;
2445 // Check that the multiplication doesn't overflow.
2446 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2448 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2449 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2452 // Check that multiplying with the use offset doesn't overflow.
2453 int64_t Offset = LU.MinOffset;
2454 if (Offset == INT64_MIN && Factor == -1)
2456 Offset = (uint64_t)Offset * Factor;
2457 if (Offset / Factor != LU.MinOffset)
2461 F.AM.BaseOffs = NewBaseOffs;
2463 // Check that this scale is legal.
2464 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2467 // Compensate for the use having MinOffset built into it.
2468 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2470 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2472 // Check that multiplying with each base register doesn't overflow.
2473 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2474 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2475 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2479 // Check that multiplying with the scaled register doesn't overflow.
2481 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2482 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2486 // If we make it here and it's legal, add it.
2487 (void)InsertFormula(LU, LUIdx, F);
2492 /// GenerateScales - Generate stride factor reuse formulae by making use of
2493 /// scaled-offset address modes, for example.
2494 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2495 // Determine the integer type for the base formula.
2496 const Type *IntTy = Base.getType();
2499 // If this Formula already has a scaled register, we can't add another one.
2500 if (Base.AM.Scale != 0) return;
2502 // Check each interesting stride.
2503 for (SmallSetVector<int64_t, 8>::const_iterator
2504 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2505 int64_t Factor = *I;
2507 Base.AM.Scale = Factor;
2508 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2509 // Check whether this scale is going to be legal.
2510 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2511 LU.Kind, LU.AccessTy, TLI)) {
2512 // As a special-case, handle special out-of-loop Basic users specially.
2513 // TODO: Reconsider this special case.
2514 if (LU.Kind == LSRUse::Basic &&
2515 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2516 LSRUse::Special, LU.AccessTy, TLI) &&
2517 LU.AllFixupsOutsideLoop)
2518 LU.Kind = LSRUse::Special;
2522 // For an ICmpZero, negating a solitary base register won't lead to
2524 if (LU.Kind == LSRUse::ICmpZero &&
2525 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2527 // For each addrec base reg, apply the scale, if possible.
2528 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2529 if (const SCEVAddRecExpr *AR =
2530 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2531 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2532 if (FactorS->isZero())
2534 // Divide out the factor, ignoring high bits, since we'll be
2535 // scaling the value back up in the end.
2536 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2537 // TODO: This could be optimized to avoid all the copying.
2539 F.ScaledReg = Quotient;
2540 F.DeleteBaseReg(F.BaseRegs[i]);
2541 (void)InsertFormula(LU, LUIdx, F);
2547 /// GenerateTruncates - Generate reuse formulae from different IV types.
2548 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2549 // This requires TargetLowering to tell us which truncates are free.
2552 // Don't bother truncating symbolic values.
2553 if (Base.AM.BaseGV) return;
2555 // Determine the integer type for the base formula.
2556 const Type *DstTy = Base.getType();
2558 DstTy = SE.getEffectiveSCEVType(DstTy);
2560 for (SmallSetVector<const Type *, 4>::const_iterator
2561 I = Types.begin(), E = Types.end(); I != E; ++I) {
2562 const Type *SrcTy = *I;
2563 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2566 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2567 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2568 JE = F.BaseRegs.end(); J != JE; ++J)
2569 *J = SE.getAnyExtendExpr(*J, SrcTy);
2571 // TODO: This assumes we've done basic processing on all uses and
2572 // have an idea what the register usage is.
2573 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2576 (void)InsertFormula(LU, LUIdx, F);
2583 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2584 /// defer modifications so that the search phase doesn't have to worry about
2585 /// the data structures moving underneath it.
2589 const SCEV *OrigReg;
2591 WorkItem(size_t LI, int64_t I, const SCEV *R)
2592 : LUIdx(LI), Imm(I), OrigReg(R) {}
2594 void print(raw_ostream &OS) const;
2600 void WorkItem::print(raw_ostream &OS) const {
2601 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2602 << " , add offset " << Imm;
2605 void WorkItem::dump() const {
2606 print(errs()); errs() << '\n';
2609 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2610 /// distance apart and try to form reuse opportunities between them.
2611 void LSRInstance::GenerateCrossUseConstantOffsets() {
2612 // Group the registers by their value without any added constant offset.
2613 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2614 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2616 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2617 SmallVector<const SCEV *, 8> Sequence;
2618 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2620 const SCEV *Reg = *I;
2621 int64_t Imm = ExtractImmediate(Reg, SE);
2622 std::pair<RegMapTy::iterator, bool> Pair =
2623 Map.insert(std::make_pair(Reg, ImmMapTy()));
2625 Sequence.push_back(Reg);
2626 Pair.first->second.insert(std::make_pair(Imm, *I));
2627 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2630 // Now examine each set of registers with the same base value. Build up
2631 // a list of work to do and do the work in a separate step so that we're
2632 // not adding formulae and register counts while we're searching.
2633 SmallVector<WorkItem, 32> WorkItems;
2634 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2635 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2636 E = Sequence.end(); I != E; ++I) {
2637 const SCEV *Reg = *I;
2638 const ImmMapTy &Imms = Map.find(Reg)->second;
2640 // It's not worthwhile looking for reuse if there's only one offset.
2641 if (Imms.size() == 1)
2644 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2645 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2647 dbgs() << ' ' << J->first;
2650 // Examine each offset.
2651 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2653 const SCEV *OrigReg = J->second;
2655 int64_t JImm = J->first;
2656 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2658 if (!isa<SCEVConstant>(OrigReg) &&
2659 UsedByIndicesMap[Reg].count() == 1) {
2660 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2664 // Conservatively examine offsets between this orig reg a few selected
2666 ImmMapTy::const_iterator OtherImms[] = {
2667 Imms.begin(), prior(Imms.end()),
2668 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2670 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2671 ImmMapTy::const_iterator M = OtherImms[i];
2672 if (M == J || M == JE) continue;
2674 // Compute the difference between the two.
2675 int64_t Imm = (uint64_t)JImm - M->first;
2676 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2677 LUIdx = UsedByIndices.find_next(LUIdx))
2678 // Make a memo of this use, offset, and register tuple.
2679 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2680 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2687 UsedByIndicesMap.clear();
2688 UniqueItems.clear();
2690 // Now iterate through the worklist and add new formulae.
2691 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2692 E = WorkItems.end(); I != E; ++I) {
2693 const WorkItem &WI = *I;
2694 size_t LUIdx = WI.LUIdx;
2695 LSRUse &LU = Uses[LUIdx];
2696 int64_t Imm = WI.Imm;
2697 const SCEV *OrigReg = WI.OrigReg;
2699 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2700 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2701 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2703 // TODO: Use a more targeted data structure.
2704 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2705 const Formula &F = LU.Formulae[L];
2706 // Use the immediate in the scaled register.
2707 if (F.ScaledReg == OrigReg) {
2708 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2709 Imm * (uint64_t)F.AM.Scale;
2710 // Don't create 50 + reg(-50).
2711 if (F.referencesReg(SE.getSCEV(
2712 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2715 NewF.AM.BaseOffs = Offs;
2716 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2717 LU.Kind, LU.AccessTy, TLI))
2719 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2721 // If the new scale is a constant in a register, and adding the constant
2722 // value to the immediate would produce a value closer to zero than the
2723 // immediate itself, then the formula isn't worthwhile.
2724 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2725 if (C->getValue()->getValue().isNegative() !=
2726 (NewF.AM.BaseOffs < 0) &&
2727 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2728 .ule(abs64(NewF.AM.BaseOffs)))
2732 (void)InsertFormula(LU, LUIdx, NewF);
2734 // Use the immediate in a base register.
2735 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2736 const SCEV *BaseReg = F.BaseRegs[N];
2737 if (BaseReg != OrigReg)
2740 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2741 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2742 LU.Kind, LU.AccessTy, TLI))
2744 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2746 // If the new formula has a constant in a register, and adding the
2747 // constant value to the immediate would produce a value closer to
2748 // zero than the immediate itself, then the formula isn't worthwhile.
2749 for (SmallVectorImpl<const SCEV *>::const_iterator
2750 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2752 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2753 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2754 abs64(NewF.AM.BaseOffs)) &&
2755 (C->getValue()->getValue() +
2756 NewF.AM.BaseOffs).countTrailingZeros() >=
2757 CountTrailingZeros_64(NewF.AM.BaseOffs))
2761 (void)InsertFormula(LU, LUIdx, NewF);
2770 /// GenerateAllReuseFormulae - Generate formulae for each use.
2772 LSRInstance::GenerateAllReuseFormulae() {
2773 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2774 // queries are more precise.
2775 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2776 LSRUse &LU = Uses[LUIdx];
2777 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2778 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2779 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2780 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2782 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2783 LSRUse &LU = Uses[LUIdx];
2784 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2785 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2786 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2787 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2788 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2789 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2790 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2791 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2793 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2794 LSRUse &LU = Uses[LUIdx];
2795 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2796 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2799 GenerateCrossUseConstantOffsets();
2802 /// If their are multiple formulae with the same set of registers used
2803 /// by other uses, pick the best one and delete the others.
2804 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2806 bool ChangedFormulae = false;
2809 // Collect the best formula for each unique set of shared registers. This
2810 // is reset for each use.
2811 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2813 BestFormulaeTy BestFormulae;
2815 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2816 LSRUse &LU = Uses[LUIdx];
2817 FormulaSorter Sorter(L, LU, SE, DT);
2818 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2821 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2822 FIdx != NumForms; ++FIdx) {
2823 Formula &F = LU.Formulae[FIdx];
2825 SmallVector<const SCEV *, 2> Key;
2826 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2827 JE = F.BaseRegs.end(); J != JE; ++J) {
2828 const SCEV *Reg = *J;
2829 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2833 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2834 Key.push_back(F.ScaledReg);
2835 // Unstable sort by host order ok, because this is only used for
2837 std::sort(Key.begin(), Key.end());
2839 std::pair<BestFormulaeTy::const_iterator, bool> P =
2840 BestFormulae.insert(std::make_pair(Key, FIdx));
2842 Formula &Best = LU.Formulae[P.first->second];
2843 if (Sorter.operator()(F, Best))
2845 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2847 " in favor of formula "; Best.print(dbgs());
2850 ChangedFormulae = true;
2852 LU.DeleteFormula(F);
2860 // Now that we've filtered out some formulae, recompute the Regs set.
2862 LU.RecomputeRegs(LUIdx, RegUses);
2864 // Reset this to prepare for the next use.
2865 BestFormulae.clear();
2868 DEBUG(if (ChangedFormulae) {
2870 "After filtering out undesirable candidates:\n";
2875 // This is a rough guess that seems to work fairly well.
2876 static const size_t ComplexityLimit = UINT16_MAX;
2878 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2879 /// solutions the solver might have to consider. It almost never considers
2880 /// this many solutions because it prune the search space, but the pruning
2881 /// isn't always sufficient.
2882 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2884 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2885 E = Uses.end(); I != E; ++I) {
2886 size_t FSize = I->Formulae.size();
2887 if (FSize >= ComplexityLimit) {
2888 Power = ComplexityLimit;
2892 if (Power >= ComplexityLimit)
2898 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2899 /// formulae to choose from, use some rough heuristics to prune down the number
2900 /// of formulae. This keeps the main solver from taking an extraordinary amount
2901 /// of time in some worst-case scenarios.
2902 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2903 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2904 DEBUG(dbgs() << "The search space is too complex.\n");
2906 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2907 "which use a superset of registers used by other "
2910 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2911 LSRUse &LU = Uses[LUIdx];
2913 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2914 Formula &F = LU.Formulae[i];
2915 // Look for a formula with a constant or GV in a register. If the use
2916 // also has a formula with that same value in an immediate field,
2917 // delete the one that uses a register.
2918 for (SmallVectorImpl<const SCEV *>::const_iterator
2919 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2920 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2922 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2923 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2924 (I - F.BaseRegs.begin()));
2925 if (LU.HasFormulaWithSameRegs(NewF)) {
2926 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2927 LU.DeleteFormula(F);
2933 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2934 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2937 NewF.AM.BaseGV = GV;
2938 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2939 (I - F.BaseRegs.begin()));
2940 if (LU.HasFormulaWithSameRegs(NewF)) {
2941 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2943 LU.DeleteFormula(F);
2954 LU.RecomputeRegs(LUIdx, RegUses);
2957 DEBUG(dbgs() << "After pre-selection:\n";
2958 print_uses(dbgs()));
2961 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2962 DEBUG(dbgs() << "The search space is too complex.\n");
2964 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2965 "separated by a constant offset will use the same "
2968 // This is especially useful for unrolled loops.
2970 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2971 LSRUse &LU = Uses[LUIdx];
2972 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2973 E = LU.Formulae.end(); I != E; ++I) {
2974 const Formula &F = *I;
2975 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2976 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2977 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2978 /*HasBaseReg=*/false,
2979 LU.Kind, LU.AccessTy)) {
2980 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2983 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2985 // Delete formulae from the new use which are no longer legal.
2987 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2988 Formula &F = LUThatHas->Formulae[i];
2989 if (!isLegalUse(F.AM,
2990 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2991 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2992 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2994 LUThatHas->DeleteFormula(F);
3001 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3003 // Update the relocs to reference the new use.
3004 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3005 E = Fixups.end(); I != E; ++I) {
3006 LSRFixup &Fixup = *I;
3007 if (Fixup.LUIdx == LUIdx) {
3008 Fixup.LUIdx = LUThatHas - &Uses.front();
3009 Fixup.Offset += F.AM.BaseOffs;
3010 DEBUG(dbgs() << "New fixup has offset "
3011 << Fixup.Offset << '\n');
3013 if (Fixup.LUIdx == NumUses-1)
3014 Fixup.LUIdx = LUIdx;
3017 // Delete the old use.
3028 DEBUG(dbgs() << "After pre-selection:\n";
3029 print_uses(dbgs()));
3032 // With all other options exhausted, loop until the system is simple
3033 // enough to handle.
3034 SmallPtrSet<const SCEV *, 4> Taken;
3035 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3036 // Ok, we have too many of formulae on our hands to conveniently handle.
3037 // Use a rough heuristic to thin out the list.
3038 DEBUG(dbgs() << "The search space is too complex.\n");
3040 // Pick the register which is used by the most LSRUses, which is likely
3041 // to be a good reuse register candidate.
3042 const SCEV *Best = 0;
3043 unsigned BestNum = 0;
3044 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3046 const SCEV *Reg = *I;
3047 if (Taken.count(Reg))
3052 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3053 if (Count > BestNum) {
3060 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3061 << " will yield profitable reuse.\n");
3064 // In any use with formulae which references this register, delete formulae
3065 // which don't reference it.
3066 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3067 LSRUse &LU = Uses[LUIdx];
3068 if (!LU.Regs.count(Best)) continue;
3071 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3072 Formula &F = LU.Formulae[i];
3073 if (!F.referencesReg(Best)) {
3074 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3075 LU.DeleteFormula(F);
3079 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3085 LU.RecomputeRegs(LUIdx, RegUses);
3088 DEBUG(dbgs() << "After pre-selection:\n";
3089 print_uses(dbgs()));
3093 /// SolveRecurse - This is the recursive solver.
3094 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3096 SmallVectorImpl<const Formula *> &Workspace,
3097 const Cost &CurCost,
3098 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3099 DenseSet<const SCEV *> &VisitedRegs) const {
3102 // - use more aggressive filtering
3103 // - sort the formula so that the most profitable solutions are found first
3104 // - sort the uses too
3106 // - don't compute a cost, and then compare. compare while computing a cost
3108 // - track register sets with SmallBitVector
3110 const LSRUse &LU = Uses[Workspace.size()];
3112 // If this use references any register that's already a part of the
3113 // in-progress solution, consider it a requirement that a formula must
3114 // reference that register in order to be considered. This prunes out
3115 // unprofitable searching.
3116 SmallSetVector<const SCEV *, 4> ReqRegs;
3117 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3118 E = CurRegs.end(); I != E; ++I)
3119 if (LU.Regs.count(*I))
3122 bool AnySatisfiedReqRegs = false;
3123 SmallPtrSet<const SCEV *, 16> NewRegs;
3126 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3127 E = LU.Formulae.end(); I != E; ++I) {
3128 const Formula &F = *I;
3130 // Ignore formulae which do not use any of the required registers.
3131 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3132 JE = ReqRegs.end(); J != JE; ++J) {
3133 const SCEV *Reg = *J;
3134 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3135 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3139 AnySatisfiedReqRegs = true;
3141 // Evaluate the cost of the current formula. If it's already worse than
3142 // the current best, prune the search at that point.
3145 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3146 if (NewCost < SolutionCost) {
3147 Workspace.push_back(&F);
3148 if (Workspace.size() != Uses.size()) {
3149 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3150 NewRegs, VisitedRegs);
3151 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3152 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3154 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3155 dbgs() << ". Regs:";
3156 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3157 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3158 dbgs() << ' ' << **I;
3161 SolutionCost = NewCost;
3162 Solution = Workspace;
3164 Workspace.pop_back();
3169 // If none of the formulae had all of the required registers, relax the
3170 // constraint so that we don't exclude all formulae.
3171 if (!AnySatisfiedReqRegs) {
3172 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3178 /// Solve - Choose one formula from each use. Return the results in the given
3179 /// Solution vector.
3180 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3181 SmallVector<const Formula *, 8> Workspace;
3183 SolutionCost.Loose();
3185 SmallPtrSet<const SCEV *, 16> CurRegs;
3186 DenseSet<const SCEV *> VisitedRegs;
3187 Workspace.reserve(Uses.size());
3189 // SolveRecurse does all the work.
3190 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3191 CurRegs, VisitedRegs);
3193 // Ok, we've now made all our decisions.
3194 DEBUG(dbgs() << "\n"
3195 "The chosen solution requires "; SolutionCost.print(dbgs());
3197 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3199 Uses[i].print(dbgs());
3202 Solution[i]->print(dbgs());
3206 assert(Solution.size() == Uses.size() && "Malformed solution!");
3209 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3210 /// the dominator tree far as we can go while still being dominated by the
3211 /// input positions. This helps canonicalize the insert position, which
3212 /// encourages sharing.
3213 BasicBlock::iterator
3214 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3215 const SmallVectorImpl<Instruction *> &Inputs)
3218 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3219 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3222 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3223 if (!Rung) return IP;
3224 Rung = Rung->getIDom();
3225 if (!Rung) return IP;
3226 IDom = Rung->getBlock();
3228 // Don't climb into a loop though.
3229 const Loop *IDomLoop = LI.getLoopFor(IDom);
3230 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3231 if (IDomDepth <= IPLoopDepth &&
3232 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3236 bool AllDominate = true;
3237 Instruction *BetterPos = 0;
3238 Instruction *Tentative = IDom->getTerminator();
3239 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3240 E = Inputs.end(); I != E; ++I) {
3241 Instruction *Inst = *I;
3242 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3243 AllDominate = false;
3246 // Attempt to find an insert position in the middle of the block,
3247 // instead of at the end, so that it can be used for other expansions.
3248 if (IDom == Inst->getParent() &&
3249 (!BetterPos || DT.dominates(BetterPos, Inst)))
3250 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3263 /// AdjustInsertPositionForExpand - Determine an input position which will be
3264 /// dominated by the operands and which will dominate the result.
3265 BasicBlock::iterator
3266 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3268 const LSRUse &LU) const {
3269 // Collect some instructions which must be dominated by the
3270 // expanding replacement. These must be dominated by any operands that
3271 // will be required in the expansion.
3272 SmallVector<Instruction *, 4> Inputs;
3273 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3274 Inputs.push_back(I);
3275 if (LU.Kind == LSRUse::ICmpZero)
3276 if (Instruction *I =
3277 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3278 Inputs.push_back(I);
3279 if (LF.PostIncLoops.count(L)) {
3280 if (LF.isUseFullyOutsideLoop(L))
3281 Inputs.push_back(L->getLoopLatch()->getTerminator());
3283 Inputs.push_back(IVIncInsertPos);
3285 // The expansion must also be dominated by the increment positions of any
3286 // loops it for which it is using post-inc mode.
3287 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3288 E = LF.PostIncLoops.end(); I != E; ++I) {
3289 const Loop *PIL = *I;
3290 if (PIL == L) continue;
3292 // Be dominated by the loop exit.
3293 SmallVector<BasicBlock *, 4> ExitingBlocks;
3294 PIL->getExitingBlocks(ExitingBlocks);
3295 if (!ExitingBlocks.empty()) {
3296 BasicBlock *BB = ExitingBlocks[0];
3297 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3298 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3299 Inputs.push_back(BB->getTerminator());
3303 // Then, climb up the immediate dominator tree as far as we can go while
3304 // still being dominated by the input positions.
3305 IP = HoistInsertPosition(IP, Inputs);
3307 // Don't insert instructions before PHI nodes.
3308 while (isa<PHINode>(IP)) ++IP;
3310 // Ignore debug intrinsics.
3311 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3316 /// Expand - Emit instructions for the leading candidate expression for this
3317 /// LSRUse (this is called "expanding").
3318 Value *LSRInstance::Expand(const LSRFixup &LF,
3320 BasicBlock::iterator IP,
3321 SCEVExpander &Rewriter,
3322 SmallVectorImpl<WeakVH> &DeadInsts) const {
3323 const LSRUse &LU = Uses[LF.LUIdx];
3325 // Determine an input position which will be dominated by the operands and
3326 // which will dominate the result.
3327 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3329 // Inform the Rewriter if we have a post-increment use, so that it can
3330 // perform an advantageous expansion.
3331 Rewriter.setPostInc(LF.PostIncLoops);
3333 // This is the type that the user actually needs.
3334 const Type *OpTy = LF.OperandValToReplace->getType();
3335 // This will be the type that we'll initially expand to.
3336 const Type *Ty = F.getType();
3338 // No type known; just expand directly to the ultimate type.
3340 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3341 // Expand directly to the ultimate type if it's the right size.
3343 // This is the type to do integer arithmetic in.
3344 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3346 // Build up a list of operands to add together to form the full base.
3347 SmallVector<const SCEV *, 8> Ops;
3349 // Expand the BaseRegs portion.
3350 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3351 E = F.BaseRegs.end(); I != E; ++I) {
3352 const SCEV *Reg = *I;
3353 assert(!Reg->isZero() && "Zero allocated in a base register!");
3355 // If we're expanding for a post-inc user, make the post-inc adjustment.
3356 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3357 Reg = TransformForPostIncUse(Denormalize, Reg,
3358 LF.UserInst, LF.OperandValToReplace,
3361 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3364 // Flush the operand list to suppress SCEVExpander hoisting.
3366 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3368 Ops.push_back(SE.getUnknown(FullV));
3371 // Expand the ScaledReg portion.
3372 Value *ICmpScaledV = 0;
3373 if (F.AM.Scale != 0) {
3374 const SCEV *ScaledS = F.ScaledReg;
3376 // If we're expanding for a post-inc user, make the post-inc adjustment.
3377 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3378 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3379 LF.UserInst, LF.OperandValToReplace,
3382 if (LU.Kind == LSRUse::ICmpZero) {
3383 // An interesting way of "folding" with an icmp is to use a negated
3384 // scale, which we'll implement by inserting it into the other operand
3386 assert(F.AM.Scale == -1 &&
3387 "The only scale supported by ICmpZero uses is -1!");
3388 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3390 // Otherwise just expand the scaled register and an explicit scale,
3391 // which is expected to be matched as part of the address.
3392 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3393 ScaledS = SE.getMulExpr(ScaledS,
3394 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3395 Ops.push_back(ScaledS);
3397 // Flush the operand list to suppress SCEVExpander hoisting.
3398 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3400 Ops.push_back(SE.getUnknown(FullV));
3404 // Expand the GV portion.
3406 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3408 // Flush the operand list to suppress SCEVExpander hoisting.
3409 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3411 Ops.push_back(SE.getUnknown(FullV));
3414 // Expand the immediate portion.
3415 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3417 if (LU.Kind == LSRUse::ICmpZero) {
3418 // The other interesting way of "folding" with an ICmpZero is to use a
3419 // negated immediate.
3421 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3423 Ops.push_back(SE.getUnknown(ICmpScaledV));
3424 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3427 // Just add the immediate values. These again are expected to be matched
3428 // as part of the address.
3429 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3433 // Emit instructions summing all the operands.
3434 const SCEV *FullS = Ops.empty() ?
3435 SE.getConstant(IntTy, 0) :
3437 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3439 // We're done expanding now, so reset the rewriter.
3440 Rewriter.clearPostInc();
3442 // An ICmpZero Formula represents an ICmp which we're handling as a
3443 // comparison against zero. Now that we've expanded an expression for that
3444 // form, update the ICmp's other operand.
3445 if (LU.Kind == LSRUse::ICmpZero) {
3446 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3447 DeadInsts.push_back(CI->getOperand(1));
3448 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3449 "a scale at the same time!");
3450 if (F.AM.Scale == -1) {
3451 if (ICmpScaledV->getType() != OpTy) {
3453 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3455 ICmpScaledV, OpTy, "tmp", CI);
3458 CI->setOperand(1, ICmpScaledV);
3460 assert(F.AM.Scale == 0 &&
3461 "ICmp does not support folding a global value and "
3462 "a scale at the same time!");
3463 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3465 if (C->getType() != OpTy)
3466 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3470 CI->setOperand(1, C);
3477 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3478 /// of their operands effectively happens in their predecessor blocks, so the
3479 /// expression may need to be expanded in multiple places.
3480 void LSRInstance::RewriteForPHI(PHINode *PN,
3483 SCEVExpander &Rewriter,
3484 SmallVectorImpl<WeakVH> &DeadInsts,
3486 DenseMap<BasicBlock *, Value *> Inserted;
3487 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3488 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3489 BasicBlock *BB = PN->getIncomingBlock(i);
3491 // If this is a critical edge, split the edge so that we do not insert
3492 // the code on all predecessor/successor paths. We do this unless this
3493 // is the canonical backedge for this loop, which complicates post-inc
3495 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3496 !isa<IndirectBrInst>(BB->getTerminator()) &&
3497 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3498 // Split the critical edge.
3499 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3501 // If PN is outside of the loop and BB is in the loop, we want to
3502 // move the block to be immediately before the PHI block, not
3503 // immediately after BB.
3504 if (L->contains(BB) && !L->contains(PN))
3505 NewBB->moveBefore(PN->getParent());
3507 // Splitting the edge can reduce the number of PHI entries we have.
3508 e = PN->getNumIncomingValues();
3510 i = PN->getBasicBlockIndex(BB);
3513 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3514 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3516 PN->setIncomingValue(i, Pair.first->second);
3518 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3520 // If this is reuse-by-noop-cast, insert the noop cast.
3521 const Type *OpTy = LF.OperandValToReplace->getType();
3522 if (FullV->getType() != OpTy)
3524 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3526 FullV, LF.OperandValToReplace->getType(),
3527 "tmp", BB->getTerminator());
3529 PN->setIncomingValue(i, FullV);
3530 Pair.first->second = FullV;
3535 /// Rewrite - Emit instructions for the leading candidate expression for this
3536 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3537 /// the newly expanded value.
3538 void LSRInstance::Rewrite(const LSRFixup &LF,
3540 SCEVExpander &Rewriter,
3541 SmallVectorImpl<WeakVH> &DeadInsts,
3543 // First, find an insertion point that dominates UserInst. For PHI nodes,
3544 // find the nearest block which dominates all the relevant uses.
3545 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3546 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3548 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3550 // If this is reuse-by-noop-cast, insert the noop cast.
3551 const Type *OpTy = LF.OperandValToReplace->getType();
3552 if (FullV->getType() != OpTy) {
3554 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3555 FullV, OpTy, "tmp", LF.UserInst);
3559 // Update the user. ICmpZero is handled specially here (for now) because
3560 // Expand may have updated one of the operands of the icmp already, and
3561 // its new value may happen to be equal to LF.OperandValToReplace, in
3562 // which case doing replaceUsesOfWith leads to replacing both operands
3563 // with the same value. TODO: Reorganize this.
3564 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3565 LF.UserInst->setOperand(0, FullV);
3567 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3570 DeadInsts.push_back(LF.OperandValToReplace);
3573 /// ImplementSolution - Rewrite all the fixup locations with new values,
3574 /// following the chosen solution.
3576 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3578 // Keep track of instructions we may have made dead, so that
3579 // we can remove them after we are done working.
3580 SmallVector<WeakVH, 16> DeadInsts;
3582 SCEVExpander Rewriter(SE);
3583 Rewriter.disableCanonicalMode();
3584 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3586 // Expand the new value definitions and update the users.
3587 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3588 E = Fixups.end(); I != E; ++I) {
3589 const LSRFixup &Fixup = *I;
3591 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3596 // Clean up after ourselves. This must be done before deleting any
3600 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3603 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3604 : IU(P->getAnalysis<IVUsers>()),
3605 SE(P->getAnalysis<ScalarEvolution>()),
3606 DT(P->getAnalysis<DominatorTree>()),
3607 LI(P->getAnalysis<LoopInfo>()),
3608 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3610 // If LoopSimplify form is not available, stay out of trouble.
3611 if (!L->isLoopSimplifyForm()) return;
3613 // If there's no interesting work to be done, bail early.
3614 if (IU.empty()) return;
3616 DEBUG(dbgs() << "\nLSR on loop ";
3617 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3620 // First, perform some low-level loop optimizations.
3622 OptimizeLoopTermCond();
3624 // Start collecting data and preparing for the solver.
3625 CollectInterestingTypesAndFactors();
3626 CollectFixupsAndInitialFormulae();
3627 CollectLoopInvariantFixupsAndFormulae();
3629 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3630 print_uses(dbgs()));
3632 // Now use the reuse data to generate a bunch of interesting ways
3633 // to formulate the values needed for the uses.
3634 GenerateAllReuseFormulae();
3636 DEBUG(dbgs() << "\n"
3637 "After generating reuse formulae:\n";
3638 print_uses(dbgs()));
3640 FilterOutUndesirableDedicatedRegisters();
3641 NarrowSearchSpaceUsingHeuristics();
3643 SmallVector<const Formula *, 8> Solution;
3646 // Release memory that is no longer needed.
3652 // Formulae should be legal.
3653 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3654 E = Uses.end(); I != E; ++I) {
3655 const LSRUse &LU = *I;
3656 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3657 JE = LU.Formulae.end(); J != JE; ++J)
3658 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3659 LU.Kind, LU.AccessTy, TLI) &&
3660 "Illegal formula generated!");
3664 // Now that we've decided what we want, make it so.
3665 ImplementSolution(Solution, P);
3668 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3669 if (Factors.empty() && Types.empty()) return;
3671 OS << "LSR has identified the following interesting factors and types: ";
3674 for (SmallSetVector<int64_t, 8>::const_iterator
3675 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3676 if (!First) OS << ", ";
3681 for (SmallSetVector<const Type *, 4>::const_iterator
3682 I = Types.begin(), E = Types.end(); I != E; ++I) {
3683 if (!First) OS << ", ";
3685 OS << '(' << **I << ')';
3690 void LSRInstance::print_fixups(raw_ostream &OS) const {
3691 OS << "LSR is examining the following fixup sites:\n";
3692 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3693 E = Fixups.end(); I != E; ++I) {
3700 void LSRInstance::print_uses(raw_ostream &OS) const {
3701 OS << "LSR is examining the following uses:\n";
3702 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3703 E = Uses.end(); I != E; ++I) {
3704 const LSRUse &LU = *I;
3708 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3709 JE = LU.Formulae.end(); J != JE; ++J) {
3717 void LSRInstance::print(raw_ostream &OS) const {
3718 print_factors_and_types(OS);
3723 void LSRInstance::dump() const {
3724 print(errs()); errs() << '\n';
3729 class LoopStrengthReduce : public LoopPass {
3730 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3731 /// transformation profitability.
3732 const TargetLowering *const TLI;
3735 static char ID; // Pass ID, replacement for typeid
3736 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3739 bool runOnLoop(Loop *L, LPPassManager &LPM);
3740 void getAnalysisUsage(AnalysisUsage &AU) const;
3745 char LoopStrengthReduce::ID = 0;
3746 INITIALIZE_PASS(LoopStrengthReduce, "loop-reduce",
3747 "Loop Strength Reduction", false, false);
3749 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3750 return new LoopStrengthReduce(TLI);
3753 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3754 : LoopPass(&ID), TLI(tli) {}
3756 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3757 // We split critical edges, so we change the CFG. However, we do update
3758 // many analyses if they are around.
3759 AU.addPreservedID(LoopSimplifyID);
3760 AU.addPreserved("domfrontier");
3762 AU.addRequired<LoopInfo>();
3763 AU.addPreserved<LoopInfo>();
3764 AU.addRequiredID(LoopSimplifyID);
3765 AU.addRequired<DominatorTree>();
3766 AU.addPreserved<DominatorTree>();
3767 AU.addRequired<ScalarEvolution>();
3768 AU.addPreserved<ScalarEvolution>();
3769 AU.addRequired<IVUsers>();
3770 AU.addPreserved<IVUsers>();
3773 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3774 bool Changed = false;
3776 // Run the main LSR transformation.
3777 Changed |= LSRInstance(TLI, L, this).getChanged();
3779 // At this point, it is worth checking to see if any recurrence PHIs are also
3780 // dead, so that we can remove them as well.
3781 Changed |= DeleteDeadPHIs(L->getHeader());