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 /// Formulae - A list of ways to build a value that can satisfy this user.
969 /// After the list is populated, one of these is selected heuristically and
970 /// used to formulate a replacement for OperandValToReplace in UserInst.
971 SmallVector<Formula, 12> Formulae;
973 /// Regs - The set of register candidates used by all formulae in this LSRUse.
974 SmallPtrSet<const SCEV *, 4> Regs;
976 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
977 MinOffset(INT64_MAX),
978 MaxOffset(INT64_MIN),
979 AllFixupsOutsideLoop(true) {}
981 bool HasFormulaWithSameRegs(const Formula &F) const;
982 bool InsertFormula(const Formula &F);
983 void DeleteFormula(Formula &F);
984 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
988 void print(raw_ostream &OS) const;
994 /// HasFormula - Test whether this use as a formula which has the same
995 /// registers as the given formula.
996 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
997 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
998 if (F.ScaledReg) Key.push_back(F.ScaledReg);
999 // Unstable sort by host order ok, because this is only used for uniquifying.
1000 std::sort(Key.begin(), Key.end());
1001 return Uniquifier.count(Key);
1004 /// InsertFormula - If the given formula has not yet been inserted, add it to
1005 /// the list, and return true. Return false otherwise.
1006 bool LSRUse::InsertFormula(const Formula &F) {
1007 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1008 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1009 // Unstable sort by host order ok, because this is only used for uniquifying.
1010 std::sort(Key.begin(), Key.end());
1012 if (!Uniquifier.insert(Key).second)
1015 // Using a register to hold the value of 0 is not profitable.
1016 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1017 "Zero allocated in a scaled register!");
1019 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1020 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1021 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1024 // Add the formula to the list.
1025 Formulae.push_back(F);
1027 // Record registers now being used by this use.
1028 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1029 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1034 /// DeleteFormula - Remove the given formula from this use's list.
1035 void LSRUse::DeleteFormula(Formula &F) {
1036 if (&F != &Formulae.back())
1037 std::swap(F, Formulae.back());
1038 Formulae.pop_back();
1039 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1042 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1043 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1044 // Now that we've filtered out some formulae, recompute the Regs set.
1045 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1047 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1048 E = Formulae.end(); I != E; ++I) {
1049 const Formula &F = *I;
1050 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1051 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1054 // Update the RegTracker.
1055 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1056 E = OldRegs.end(); I != E; ++I)
1057 if (!Regs.count(*I))
1058 RegUses.DropRegister(*I, LUIdx);
1061 void LSRUse::print(raw_ostream &OS) const {
1062 OS << "LSR Use: Kind=";
1064 case Basic: OS << "Basic"; break;
1065 case Special: OS << "Special"; break;
1066 case ICmpZero: OS << "ICmpZero"; break;
1068 OS << "Address of ";
1069 if (AccessTy->isPointerTy())
1070 OS << "pointer"; // the full pointer type could be really verbose
1075 OS << ", Offsets={";
1076 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1077 E = Offsets.end(); I != E; ++I) {
1084 if (AllFixupsOutsideLoop)
1085 OS << ", all-fixups-outside-loop";
1088 void LSRUse::dump() const {
1089 print(errs()); errs() << '\n';
1092 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1093 /// be completely folded into the user instruction at isel time. This includes
1094 /// address-mode folding and special icmp tricks.
1095 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1096 LSRUse::KindType Kind, const Type *AccessTy,
1097 const TargetLowering *TLI) {
1099 case LSRUse::Address:
1100 // If we have low-level target information, ask the target if it can
1101 // completely fold this address.
1102 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1104 // Otherwise, just guess that reg+reg addressing is legal.
1105 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1107 case LSRUse::ICmpZero:
1108 // There's not even a target hook for querying whether it would be legal to
1109 // fold a GV into an ICmp.
1113 // ICmp only has two operands; don't allow more than two non-trivial parts.
1114 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1117 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1118 // putting the scaled register in the other operand of the icmp.
1119 if (AM.Scale != 0 && AM.Scale != -1)
1122 // If we have low-level target information, ask the target if it can fold an
1123 // integer immediate on an icmp.
1124 if (AM.BaseOffs != 0) {
1125 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1132 // Only handle single-register values.
1133 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1135 case LSRUse::Special:
1136 // Only handle -1 scales, or no scale.
1137 return AM.Scale == 0 || AM.Scale == -1;
1143 static bool isLegalUse(TargetLowering::AddrMode AM,
1144 int64_t MinOffset, int64_t MaxOffset,
1145 LSRUse::KindType Kind, const Type *AccessTy,
1146 const TargetLowering *TLI) {
1147 // Check for overflow.
1148 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1151 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1152 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1153 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1154 // Check for overflow.
1155 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1158 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1159 return isLegalUse(AM, Kind, AccessTy, TLI);
1164 static bool isAlwaysFoldable(int64_t BaseOffs,
1165 GlobalValue *BaseGV,
1167 LSRUse::KindType Kind, const Type *AccessTy,
1168 const TargetLowering *TLI) {
1169 // Fast-path: zero is always foldable.
1170 if (BaseOffs == 0 && !BaseGV) return true;
1172 // Conservatively, create an address with an immediate and a
1173 // base and a scale.
1174 TargetLowering::AddrMode AM;
1175 AM.BaseOffs = BaseOffs;
1177 AM.HasBaseReg = HasBaseReg;
1178 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1180 // Canonicalize a scale of 1 to a base register if the formula doesn't
1181 // already have a base register.
1182 if (!AM.HasBaseReg && AM.Scale == 1) {
1184 AM.HasBaseReg = true;
1187 return isLegalUse(AM, Kind, AccessTy, TLI);
1190 static bool isAlwaysFoldable(const SCEV *S,
1191 int64_t MinOffset, int64_t MaxOffset,
1193 LSRUse::KindType Kind, const Type *AccessTy,
1194 const TargetLowering *TLI,
1195 ScalarEvolution &SE) {
1196 // Fast-path: zero is always foldable.
1197 if (S->isZero()) return true;
1199 // Conservatively, create an address with an immediate and a
1200 // base and a scale.
1201 int64_t BaseOffs = ExtractImmediate(S, SE);
1202 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1204 // If there's anything else involved, it's not foldable.
1205 if (!S->isZero()) return false;
1207 // Fast-path: zero is always foldable.
1208 if (BaseOffs == 0 && !BaseGV) return true;
1210 // Conservatively, create an address with an immediate and a
1211 // base and a scale.
1212 TargetLowering::AddrMode AM;
1213 AM.BaseOffs = BaseOffs;
1215 AM.HasBaseReg = HasBaseReg;
1216 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1218 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1223 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1224 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1225 struct UseMapDenseMapInfo {
1226 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1227 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1230 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1231 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1235 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1236 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1237 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1241 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1242 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1247 /// FormulaSorter - This class implements an ordering for formulae which sorts
1248 /// the by their standalone cost.
1249 class FormulaSorter {
1250 /// These two sets are kept empty, so that we compute standalone costs.
1251 DenseSet<const SCEV *> VisitedRegs;
1252 SmallPtrSet<const SCEV *, 16> Regs;
1255 ScalarEvolution &SE;
1259 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1260 : L(l), LU(&lu), SE(se), DT(dt) {}
1262 bool operator()(const Formula &A, const Formula &B) {
1264 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1267 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1269 return CostA < CostB;
1273 /// LSRInstance - This class holds state for the main loop strength reduction
1277 ScalarEvolution &SE;
1280 const TargetLowering *const TLI;
1284 /// IVIncInsertPos - This is the insert position that the current loop's
1285 /// induction variable increment should be placed. In simple loops, this is
1286 /// the latch block's terminator. But in more complicated cases, this is a
1287 /// position which will dominate all the in-loop post-increment users.
1288 Instruction *IVIncInsertPos;
1290 /// Factors - Interesting factors between use strides.
1291 SmallSetVector<int64_t, 8> Factors;
1293 /// Types - Interesting use types, to facilitate truncation reuse.
1294 SmallSetVector<const Type *, 4> Types;
1296 /// Fixups - The list of operands which are to be replaced.
1297 SmallVector<LSRFixup, 16> Fixups;
1299 /// Uses - The list of interesting uses.
1300 SmallVector<LSRUse, 16> Uses;
1302 /// RegUses - Track which uses use which register candidates.
1303 RegUseTracker RegUses;
1305 void OptimizeShadowIV();
1306 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1307 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1308 void OptimizeLoopTermCond();
1310 void CollectInterestingTypesAndFactors();
1311 void CollectFixupsAndInitialFormulae();
1313 LSRFixup &getNewFixup() {
1314 Fixups.push_back(LSRFixup());
1315 return Fixups.back();
1318 // Support for sharing of LSRUses between LSRFixups.
1319 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1321 UseMapDenseMapInfo> UseMapTy;
1324 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1325 LSRUse::KindType Kind, const Type *AccessTy);
1327 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1328 LSRUse::KindType Kind,
1329 const Type *AccessTy);
1331 void DeleteUse(LSRUse &LU);
1333 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1336 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1337 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1338 void CountRegisters(const Formula &F, size_t LUIdx);
1339 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1341 void CollectLoopInvariantFixupsAndFormulae();
1343 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1344 unsigned Depth = 0);
1345 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1346 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1347 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1348 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1349 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1350 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1351 void GenerateCrossUseConstantOffsets();
1352 void GenerateAllReuseFormulae();
1354 void FilterOutUndesirableDedicatedRegisters();
1356 size_t EstimateSearchSpaceComplexity() const;
1357 void NarrowSearchSpaceUsingHeuristics();
1359 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1361 SmallVectorImpl<const Formula *> &Workspace,
1362 const Cost &CurCost,
1363 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1364 DenseSet<const SCEV *> &VisitedRegs) const;
1365 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1367 BasicBlock::iterator
1368 HoistInsertPosition(BasicBlock::iterator IP,
1369 const SmallVectorImpl<Instruction *> &Inputs) const;
1370 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1372 const LSRUse &LU) const;
1374 Value *Expand(const LSRFixup &LF,
1376 BasicBlock::iterator IP,
1377 SCEVExpander &Rewriter,
1378 SmallVectorImpl<WeakVH> &DeadInsts) const;
1379 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1381 SCEVExpander &Rewriter,
1382 SmallVectorImpl<WeakVH> &DeadInsts,
1384 void Rewrite(const LSRFixup &LF,
1386 SCEVExpander &Rewriter,
1387 SmallVectorImpl<WeakVH> &DeadInsts,
1389 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1392 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1394 bool getChanged() const { return Changed; }
1396 void print_factors_and_types(raw_ostream &OS) const;
1397 void print_fixups(raw_ostream &OS) const;
1398 void print_uses(raw_ostream &OS) const;
1399 void print(raw_ostream &OS) const;
1405 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1406 /// inside the loop then try to eliminate the cast operation.
1407 void LSRInstance::OptimizeShadowIV() {
1408 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1409 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1412 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1413 UI != E; /* empty */) {
1414 IVUsers::const_iterator CandidateUI = UI;
1416 Instruction *ShadowUse = CandidateUI->getUser();
1417 const Type *DestTy = NULL;
1419 /* If shadow use is a int->float cast then insert a second IV
1420 to eliminate this cast.
1422 for (unsigned i = 0; i < n; ++i)
1428 for (unsigned i = 0; i < n; ++i, ++d)
1431 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1432 DestTy = UCast->getDestTy();
1433 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1434 DestTy = SCast->getDestTy();
1435 if (!DestTy) continue;
1438 // If target does not support DestTy natively then do not apply
1439 // this transformation.
1440 EVT DVT = TLI->getValueType(DestTy);
1441 if (!TLI->isTypeLegal(DVT)) continue;
1444 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1446 if (PH->getNumIncomingValues() != 2) continue;
1448 const Type *SrcTy = PH->getType();
1449 int Mantissa = DestTy->getFPMantissaWidth();
1450 if (Mantissa == -1) continue;
1451 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1454 unsigned Entry, Latch;
1455 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1463 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1464 if (!Init) continue;
1465 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1467 BinaryOperator *Incr =
1468 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1469 if (!Incr) continue;
1470 if (Incr->getOpcode() != Instruction::Add
1471 && Incr->getOpcode() != Instruction::Sub)
1474 /* Initialize new IV, double d = 0.0 in above example. */
1475 ConstantInt *C = NULL;
1476 if (Incr->getOperand(0) == PH)
1477 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1478 else if (Incr->getOperand(1) == PH)
1479 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1485 // Ignore negative constants, as the code below doesn't handle them
1486 // correctly. TODO: Remove this restriction.
1487 if (!C->getValue().isStrictlyPositive()) continue;
1489 /* Add new PHINode. */
1490 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1492 /* create new increment. '++d' in above example. */
1493 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1494 BinaryOperator *NewIncr =
1495 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1496 Instruction::FAdd : Instruction::FSub,
1497 NewPH, CFP, "IV.S.next.", Incr);
1499 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1500 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1502 /* Remove cast operation */
1503 ShadowUse->replaceAllUsesWith(NewPH);
1504 ShadowUse->eraseFromParent();
1510 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1511 /// set the IV user and stride information and return true, otherwise return
1513 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1514 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1515 if (UI->getUser() == Cond) {
1516 // NOTE: we could handle setcc instructions with multiple uses here, but
1517 // InstCombine does it as well for simple uses, it's not clear that it
1518 // occurs enough in real life to handle.
1525 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1526 /// a max computation.
1528 /// This is a narrow solution to a specific, but acute, problem. For loops
1534 /// } while (++i < n);
1536 /// the trip count isn't just 'n', because 'n' might not be positive. And
1537 /// unfortunately this can come up even for loops where the user didn't use
1538 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1539 /// will commonly be lowered like this:
1545 /// } while (++i < n);
1548 /// and then it's possible for subsequent optimization to obscure the if
1549 /// test in such a way that indvars can't find it.
1551 /// When indvars can't find the if test in loops like this, it creates a
1552 /// max expression, which allows it to give the loop a canonical
1553 /// induction variable:
1556 /// max = n < 1 ? 1 : n;
1559 /// } while (++i != max);
1561 /// Canonical induction variables are necessary because the loop passes
1562 /// are designed around them. The most obvious example of this is the
1563 /// LoopInfo analysis, which doesn't remember trip count values. It
1564 /// expects to be able to rediscover the trip count each time it is
1565 /// needed, and it does this using a simple analysis that only succeeds if
1566 /// the loop has a canonical induction variable.
1568 /// However, when it comes time to generate code, the maximum operation
1569 /// can be quite costly, especially if it's inside of an outer loop.
1571 /// This function solves this problem by detecting this type of loop and
1572 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1573 /// the instructions for the maximum computation.
1575 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1576 // Check that the loop matches the pattern we're looking for.
1577 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1578 Cond->getPredicate() != CmpInst::ICMP_NE)
1581 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1582 if (!Sel || !Sel->hasOneUse()) return Cond;
1584 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1585 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1587 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1589 // Add one to the backedge-taken count to get the trip count.
1590 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1591 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1593 // Check for a max calculation that matches the pattern. There's no check
1594 // for ICMP_ULE here because the comparison would be with zero, which
1595 // isn't interesting.
1596 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1597 const SCEVNAryExpr *Max = 0;
1598 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1599 Pred = ICmpInst::ICMP_SLE;
1601 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1602 Pred = ICmpInst::ICMP_SLT;
1604 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1605 Pred = ICmpInst::ICMP_ULT;
1612 // To handle a max with more than two operands, this optimization would
1613 // require additional checking and setup.
1614 if (Max->getNumOperands() != 2)
1617 const SCEV *MaxLHS = Max->getOperand(0);
1618 const SCEV *MaxRHS = Max->getOperand(1);
1620 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1621 // for a comparison with 1. For <= and >=, a comparison with zero.
1623 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1626 // Check the relevant induction variable for conformance to
1628 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1629 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1630 if (!AR || !AR->isAffine() ||
1631 AR->getStart() != One ||
1632 AR->getStepRecurrence(SE) != One)
1635 assert(AR->getLoop() == L &&
1636 "Loop condition operand is an addrec in a different loop!");
1638 // Check the right operand of the select, and remember it, as it will
1639 // be used in the new comparison instruction.
1641 if (ICmpInst::isTrueWhenEqual(Pred)) {
1642 // Look for n+1, and grab n.
1643 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1644 if (isa<ConstantInt>(BO->getOperand(1)) &&
1645 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1646 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1647 NewRHS = BO->getOperand(0);
1648 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1649 if (isa<ConstantInt>(BO->getOperand(1)) &&
1650 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1651 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1652 NewRHS = BO->getOperand(0);
1655 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1656 NewRHS = Sel->getOperand(1);
1657 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1658 NewRHS = Sel->getOperand(2);
1659 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1660 NewRHS = SU->getValue();
1662 // Max doesn't match expected pattern.
1665 // Determine the new comparison opcode. It may be signed or unsigned,
1666 // and the original comparison may be either equality or inequality.
1667 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1668 Pred = CmpInst::getInversePredicate(Pred);
1670 // Ok, everything looks ok to change the condition into an SLT or SGE and
1671 // delete the max calculation.
1673 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1675 // Delete the max calculation instructions.
1676 Cond->replaceAllUsesWith(NewCond);
1677 CondUse->setUser(NewCond);
1678 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1679 Cond->eraseFromParent();
1680 Sel->eraseFromParent();
1681 if (Cmp->use_empty())
1682 Cmp->eraseFromParent();
1686 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1687 /// postinc iv when possible.
1689 LSRInstance::OptimizeLoopTermCond() {
1690 SmallPtrSet<Instruction *, 4> PostIncs;
1692 BasicBlock *LatchBlock = L->getLoopLatch();
1693 SmallVector<BasicBlock*, 8> ExitingBlocks;
1694 L->getExitingBlocks(ExitingBlocks);
1696 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1697 BasicBlock *ExitingBlock = ExitingBlocks[i];
1699 // Get the terminating condition for the loop if possible. If we
1700 // can, we want to change it to use a post-incremented version of its
1701 // induction variable, to allow coalescing the live ranges for the IV into
1702 // one register value.
1704 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1707 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1708 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1711 // Search IVUsesByStride to find Cond's IVUse if there is one.
1712 IVStrideUse *CondUse = 0;
1713 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1714 if (!FindIVUserForCond(Cond, CondUse))
1717 // If the trip count is computed in terms of a max (due to ScalarEvolution
1718 // being unable to find a sufficient guard, for example), change the loop
1719 // comparison to use SLT or ULT instead of NE.
1720 // One consequence of doing this now is that it disrupts the count-down
1721 // optimization. That's not always a bad thing though, because in such
1722 // cases it may still be worthwhile to avoid a max.
1723 Cond = OptimizeMax(Cond, CondUse);
1725 // If this exiting block dominates the latch block, it may also use
1726 // the post-inc value if it won't be shared with other uses.
1727 // Check for dominance.
1728 if (!DT.dominates(ExitingBlock, LatchBlock))
1731 // Conservatively avoid trying to use the post-inc value in non-latch
1732 // exits if there may be pre-inc users in intervening blocks.
1733 if (LatchBlock != ExitingBlock)
1734 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1735 // Test if the use is reachable from the exiting block. This dominator
1736 // query is a conservative approximation of reachability.
1737 if (&*UI != CondUse &&
1738 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1739 // Conservatively assume there may be reuse if the quotient of their
1740 // strides could be a legal scale.
1741 const SCEV *A = IU.getStride(*CondUse, L);
1742 const SCEV *B = IU.getStride(*UI, L);
1743 if (!A || !B) continue;
1744 if (SE.getTypeSizeInBits(A->getType()) !=
1745 SE.getTypeSizeInBits(B->getType())) {
1746 if (SE.getTypeSizeInBits(A->getType()) >
1747 SE.getTypeSizeInBits(B->getType()))
1748 B = SE.getSignExtendExpr(B, A->getType());
1750 A = SE.getSignExtendExpr(A, B->getType());
1752 if (const SCEVConstant *D =
1753 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1754 const ConstantInt *C = D->getValue();
1755 // Stride of one or negative one can have reuse with non-addresses.
1756 if (C->isOne() || C->isAllOnesValue())
1757 goto decline_post_inc;
1758 // Avoid weird situations.
1759 if (C->getValue().getMinSignedBits() >= 64 ||
1760 C->getValue().isMinSignedValue())
1761 goto decline_post_inc;
1762 // Without TLI, assume that any stride might be valid, and so any
1763 // use might be shared.
1765 goto decline_post_inc;
1766 // Check for possible scaled-address reuse.
1767 const Type *AccessTy = getAccessType(UI->getUser());
1768 TargetLowering::AddrMode AM;
1769 AM.Scale = C->getSExtValue();
1770 if (TLI->isLegalAddressingMode(AM, AccessTy))
1771 goto decline_post_inc;
1772 AM.Scale = -AM.Scale;
1773 if (TLI->isLegalAddressingMode(AM, AccessTy))
1774 goto decline_post_inc;
1778 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1781 // It's possible for the setcc instruction to be anywhere in the loop, and
1782 // possible for it to have multiple users. If it is not immediately before
1783 // the exiting block branch, move it.
1784 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1785 if (Cond->hasOneUse()) {
1786 Cond->moveBefore(TermBr);
1788 // Clone the terminating condition and insert into the loopend.
1789 ICmpInst *OldCond = Cond;
1790 Cond = cast<ICmpInst>(Cond->clone());
1791 Cond->setName(L->getHeader()->getName() + ".termcond");
1792 ExitingBlock->getInstList().insert(TermBr, Cond);
1794 // Clone the IVUse, as the old use still exists!
1795 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1796 TermBr->replaceUsesOfWith(OldCond, Cond);
1800 // If we get to here, we know that we can transform the setcc instruction to
1801 // use the post-incremented version of the IV, allowing us to coalesce the
1802 // live ranges for the IV correctly.
1803 CondUse->transformToPostInc(L);
1806 PostIncs.insert(Cond);
1810 // Determine an insertion point for the loop induction variable increment. It
1811 // must dominate all the post-inc comparisons we just set up, and it must
1812 // dominate the loop latch edge.
1813 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1814 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1815 E = PostIncs.end(); I != E; ++I) {
1817 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1819 if (BB == (*I)->getParent())
1820 IVIncInsertPos = *I;
1821 else if (BB != IVIncInsertPos->getParent())
1822 IVIncInsertPos = BB->getTerminator();
1826 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1827 /// at the given offset and other details. If so, update the use and
1830 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1831 LSRUse::KindType Kind, const Type *AccessTy) {
1832 int64_t NewMinOffset = LU.MinOffset;
1833 int64_t NewMaxOffset = LU.MaxOffset;
1834 const Type *NewAccessTy = AccessTy;
1836 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1837 // something conservative, however this can pessimize in the case that one of
1838 // the uses will have all its uses outside the loop, for example.
1839 if (LU.Kind != Kind)
1841 // Conservatively assume HasBaseReg is true for now.
1842 if (NewOffset < LU.MinOffset) {
1843 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1844 Kind, AccessTy, TLI))
1846 NewMinOffset = NewOffset;
1847 } else if (NewOffset > LU.MaxOffset) {
1848 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1849 Kind, AccessTy, TLI))
1851 NewMaxOffset = NewOffset;
1853 // Check for a mismatched access type, and fall back conservatively as needed.
1854 // TODO: Be less conservative when the type is similar and can use the same
1855 // addressing modes.
1856 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1857 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1860 LU.MinOffset = NewMinOffset;
1861 LU.MaxOffset = NewMaxOffset;
1862 LU.AccessTy = NewAccessTy;
1863 if (NewOffset != LU.Offsets.back())
1864 LU.Offsets.push_back(NewOffset);
1868 /// getUse - Return an LSRUse index and an offset value for a fixup which
1869 /// needs the given expression, with the given kind and optional access type.
1870 /// Either reuse an existing use or create a new one, as needed.
1871 std::pair<size_t, int64_t>
1872 LSRInstance::getUse(const SCEV *&Expr,
1873 LSRUse::KindType Kind, const Type *AccessTy) {
1874 const SCEV *Copy = Expr;
1875 int64_t Offset = ExtractImmediate(Expr, SE);
1877 // Basic uses can't accept any offset, for example.
1878 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1883 std::pair<UseMapTy::iterator, bool> P =
1884 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1886 // A use already existed with this base.
1887 size_t LUIdx = P.first->second;
1888 LSRUse &LU = Uses[LUIdx];
1889 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1891 return std::make_pair(LUIdx, Offset);
1894 // Create a new use.
1895 size_t LUIdx = Uses.size();
1896 P.first->second = LUIdx;
1897 Uses.push_back(LSRUse(Kind, AccessTy));
1898 LSRUse &LU = Uses[LUIdx];
1900 // We don't need to track redundant offsets, but we don't need to go out
1901 // of our way here to avoid them.
1902 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1903 LU.Offsets.push_back(Offset);
1905 LU.MinOffset = Offset;
1906 LU.MaxOffset = Offset;
1907 return std::make_pair(LUIdx, Offset);
1910 /// DeleteUse - Delete the given use from the Uses list.
1911 void LSRInstance::DeleteUse(LSRUse &LU) {
1912 if (&LU != &Uses.back())
1913 std::swap(LU, Uses.back());
1917 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1918 /// a formula that has the same registers as the given formula.
1920 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1921 const LSRUse &OrigLU) {
1922 // Search all uses for the formula. This could be more clever. Ignore
1923 // ICmpZero uses because they may contain formulae generated by
1924 // GenerateICmpZeroScales, in which case adding fixup offsets may
1926 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1927 LSRUse &LU = Uses[LUIdx];
1928 if (&LU != &OrigLU &&
1929 LU.Kind != LSRUse::ICmpZero &&
1930 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1931 LU.HasFormulaWithSameRegs(OrigF)) {
1932 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1933 E = LU.Formulae.end(); I != E; ++I) {
1934 const Formula &F = *I;
1935 if (F.BaseRegs == OrigF.BaseRegs &&
1936 F.ScaledReg == OrigF.ScaledReg &&
1937 F.AM.BaseGV == OrigF.AM.BaseGV &&
1938 F.AM.Scale == OrigF.AM.Scale &&
1940 if (F.AM.BaseOffs == 0)
1951 void LSRInstance::CollectInterestingTypesAndFactors() {
1952 SmallSetVector<const SCEV *, 4> Strides;
1954 // Collect interesting types and strides.
1955 SmallVector<const SCEV *, 4> Worklist;
1956 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1957 const SCEV *Expr = IU.getExpr(*UI);
1959 // Collect interesting types.
1960 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1962 // Add strides for mentioned loops.
1963 Worklist.push_back(Expr);
1965 const SCEV *S = Worklist.pop_back_val();
1966 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1967 Strides.insert(AR->getStepRecurrence(SE));
1968 Worklist.push_back(AR->getStart());
1969 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1970 Worklist.append(Add->op_begin(), Add->op_end());
1972 } while (!Worklist.empty());
1975 // Compute interesting factors from the set of interesting strides.
1976 for (SmallSetVector<const SCEV *, 4>::const_iterator
1977 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1978 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1979 next(I); NewStrideIter != E; ++NewStrideIter) {
1980 const SCEV *OldStride = *I;
1981 const SCEV *NewStride = *NewStrideIter;
1983 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1984 SE.getTypeSizeInBits(NewStride->getType())) {
1985 if (SE.getTypeSizeInBits(OldStride->getType()) >
1986 SE.getTypeSizeInBits(NewStride->getType()))
1987 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1989 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1991 if (const SCEVConstant *Factor =
1992 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1994 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1995 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1996 } else if (const SCEVConstant *Factor =
1997 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2000 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2001 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2005 // If all uses use the same type, don't bother looking for truncation-based
2007 if (Types.size() == 1)
2010 DEBUG(print_factors_and_types(dbgs()));
2013 void LSRInstance::CollectFixupsAndInitialFormulae() {
2014 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2016 LSRFixup &LF = getNewFixup();
2017 LF.UserInst = UI->getUser();
2018 LF.OperandValToReplace = UI->getOperandValToReplace();
2019 LF.PostIncLoops = UI->getPostIncLoops();
2021 LSRUse::KindType Kind = LSRUse::Basic;
2022 const Type *AccessTy = 0;
2023 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2024 Kind = LSRUse::Address;
2025 AccessTy = getAccessType(LF.UserInst);
2028 const SCEV *S = IU.getExpr(*UI);
2030 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2031 // (N - i == 0), and this allows (N - i) to be the expression that we work
2032 // with rather than just N or i, so we can consider the register
2033 // requirements for both N and i at the same time. Limiting this code to
2034 // equality icmps is not a problem because all interesting loops use
2035 // equality icmps, thanks to IndVarSimplify.
2036 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2037 if (CI->isEquality()) {
2038 // Swap the operands if needed to put the OperandValToReplace on the
2039 // left, for consistency.
2040 Value *NV = CI->getOperand(1);
2041 if (NV == LF.OperandValToReplace) {
2042 CI->setOperand(1, CI->getOperand(0));
2043 CI->setOperand(0, NV);
2044 NV = CI->getOperand(1);
2048 // x == y --> x - y == 0
2049 const SCEV *N = SE.getSCEV(NV);
2050 if (N->isLoopInvariant(L)) {
2051 Kind = LSRUse::ICmpZero;
2052 S = SE.getMinusSCEV(N, S);
2055 // -1 and the negations of all interesting strides (except the negation
2056 // of -1) are now also interesting.
2057 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2058 if (Factors[i] != -1)
2059 Factors.insert(-(uint64_t)Factors[i]);
2063 // Set up the initial formula for this use.
2064 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2066 LF.Offset = P.second;
2067 LSRUse &LU = Uses[LF.LUIdx];
2068 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2070 // If this is the first use of this LSRUse, give it a formula.
2071 if (LU.Formulae.empty()) {
2072 InsertInitialFormula(S, LU, LF.LUIdx);
2073 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2077 DEBUG(print_fixups(dbgs()));
2080 /// InsertInitialFormula - Insert a formula for the given expression into
2081 /// the given use, separating out loop-variant portions from loop-invariant
2082 /// and loop-computable portions.
2084 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2086 F.InitialMatch(S, L, SE, DT);
2087 bool Inserted = InsertFormula(LU, LUIdx, F);
2088 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2091 /// InsertSupplementalFormula - Insert a simple single-register formula for
2092 /// the given expression into the given use.
2094 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2095 LSRUse &LU, size_t LUIdx) {
2097 F.BaseRegs.push_back(S);
2098 F.AM.HasBaseReg = true;
2099 bool Inserted = InsertFormula(LU, LUIdx, F);
2100 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2103 /// CountRegisters - Note which registers are used by the given formula,
2104 /// updating RegUses.
2105 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2107 RegUses.CountRegister(F.ScaledReg, LUIdx);
2108 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2109 E = F.BaseRegs.end(); I != E; ++I)
2110 RegUses.CountRegister(*I, LUIdx);
2113 /// InsertFormula - If the given formula has not yet been inserted, add it to
2114 /// the list, and return true. Return false otherwise.
2115 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2116 if (!LU.InsertFormula(F))
2119 CountRegisters(F, LUIdx);
2123 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2124 /// loop-invariant values which we're tracking. These other uses will pin these
2125 /// values in registers, making them less profitable for elimination.
2126 /// TODO: This currently misses non-constant addrec step registers.
2127 /// TODO: Should this give more weight to users inside the loop?
2129 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2130 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2131 SmallPtrSet<const SCEV *, 8> Inserted;
2133 while (!Worklist.empty()) {
2134 const SCEV *S = Worklist.pop_back_val();
2136 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2137 Worklist.append(N->op_begin(), N->op_end());
2138 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2139 Worklist.push_back(C->getOperand());
2140 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2141 Worklist.push_back(D->getLHS());
2142 Worklist.push_back(D->getRHS());
2143 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2144 if (!Inserted.insert(U)) continue;
2145 const Value *V = U->getValue();
2146 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2147 // Look for instructions defined outside the loop.
2148 if (L->contains(Inst)) continue;
2149 } else if (isa<UndefValue>(V))
2150 // Undef doesn't have a live range, so it doesn't matter.
2152 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2154 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2155 // Ignore non-instructions.
2158 // Ignore instructions in other functions (as can happen with
2160 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2162 // Ignore instructions not dominated by the loop.
2163 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2164 UserInst->getParent() :
2165 cast<PHINode>(UserInst)->getIncomingBlock(
2166 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2167 if (!DT.dominates(L->getHeader(), UseBB))
2169 // Ignore uses which are part of other SCEV expressions, to avoid
2170 // analyzing them multiple times.
2171 if (SE.isSCEVable(UserInst->getType())) {
2172 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2173 // If the user is a no-op, look through to its uses.
2174 if (!isa<SCEVUnknown>(UserS))
2178 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2182 // Ignore icmp instructions which are already being analyzed.
2183 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2184 unsigned OtherIdx = !UI.getOperandNo();
2185 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2186 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2190 LSRFixup &LF = getNewFixup();
2191 LF.UserInst = const_cast<Instruction *>(UserInst);
2192 LF.OperandValToReplace = UI.getUse();
2193 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2195 LF.Offset = P.second;
2196 LSRUse &LU = Uses[LF.LUIdx];
2197 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2198 InsertSupplementalFormula(U, LU, LF.LUIdx);
2199 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2206 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2207 /// separate registers. If C is non-null, multiply each subexpression by C.
2208 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2209 SmallVectorImpl<const SCEV *> &Ops,
2210 SmallVectorImpl<const SCEV *> &UninterestingOps,
2212 ScalarEvolution &SE) {
2213 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2214 // Break out add operands.
2215 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2217 CollectSubexprs(*I, C, Ops, UninterestingOps, L, SE);
2219 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2220 // Split a non-zero base out of an addrec.
2221 if (!AR->getStart()->isZero()) {
2222 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2223 AR->getStepRecurrence(SE),
2225 C, Ops, UninterestingOps, L, SE);
2226 CollectSubexprs(AR->getStart(), C, Ops, UninterestingOps, L, SE);
2229 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2230 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2231 if (Mul->getNumOperands() == 2)
2232 if (const SCEVConstant *Op0 =
2233 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2234 CollectSubexprs(Mul->getOperand(1),
2235 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2236 Ops, UninterestingOps, L, SE);
2241 // Otherwise use the value itself. Loop-variant "unknown" values are
2242 // uninteresting; we won't be able to do anything meaningful with them.
2243 if (!C && isa<SCEVUnknown>(S) && !S->isLoopInvariant(L))
2244 UninterestingOps.push_back(S);
2246 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2249 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2251 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2254 // Arbitrarily cap recursion to protect compile time.
2255 if (Depth >= 3) return;
2257 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2258 const SCEV *BaseReg = Base.BaseRegs[i];
2260 SmallVector<const SCEV *, 8> AddOps, UninterestingAddOps;
2261 CollectSubexprs(BaseReg, 0, AddOps, UninterestingAddOps, L, SE);
2263 // Add any uninteresting values as one register, as we won't be able to
2264 // form any interesting reassociation opportunities with them. They'll
2265 // just have to be added inside the loop no matter what we do.
2266 if (!UninterestingAddOps.empty())
2267 AddOps.push_back(SE.getAddExpr(UninterestingAddOps));
2269 if (AddOps.size() == 1) continue;
2271 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2272 JE = AddOps.end(); J != JE; ++J) {
2273 // Don't pull a constant into a register if the constant could be folded
2274 // into an immediate field.
2275 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2276 Base.getNumRegs() > 1,
2277 LU.Kind, LU.AccessTy, TLI, SE))
2280 // Collect all operands except *J.
2281 SmallVector<const SCEV *, 8> InnerAddOps
2282 ( ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2284 (next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2286 // Don't leave just a constant behind in a register if the constant could
2287 // be folded into an immediate field.
2288 if (InnerAddOps.size() == 1 &&
2289 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2290 Base.getNumRegs() > 1,
2291 LU.Kind, LU.AccessTy, TLI, SE))
2294 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2295 if (InnerSum->isZero())
2298 F.BaseRegs[i] = InnerSum;
2299 F.BaseRegs.push_back(*J);
2300 if (InsertFormula(LU, LUIdx, F))
2301 // If that formula hadn't been seen before, recurse to find more like
2303 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2308 /// GenerateCombinations - Generate a formula consisting of all of the
2309 /// loop-dominating registers added into a single register.
2310 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2312 // This method is only interesting on a plurality of registers.
2313 if (Base.BaseRegs.size() <= 1) return;
2317 SmallVector<const SCEV *, 4> Ops;
2318 for (SmallVectorImpl<const SCEV *>::const_iterator
2319 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2320 const SCEV *BaseReg = *I;
2321 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2322 !BaseReg->hasComputableLoopEvolution(L))
2323 Ops.push_back(BaseReg);
2325 F.BaseRegs.push_back(BaseReg);
2327 if (Ops.size() > 1) {
2328 const SCEV *Sum = SE.getAddExpr(Ops);
2329 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2330 // opportunity to fold something. For now, just ignore such cases
2331 // rather than proceed with zero in a register.
2332 if (!Sum->isZero()) {
2333 F.BaseRegs.push_back(Sum);
2334 (void)InsertFormula(LU, LUIdx, F);
2339 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2340 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2342 // We can't add a symbolic offset if the address already contains one.
2343 if (Base.AM.BaseGV) return;
2345 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2346 const SCEV *G = Base.BaseRegs[i];
2347 GlobalValue *GV = ExtractSymbol(G, SE);
2348 if (G->isZero() || !GV)
2352 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2353 LU.Kind, LU.AccessTy, TLI))
2356 (void)InsertFormula(LU, LUIdx, F);
2360 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2361 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2363 // TODO: For now, just add the min and max offset, because it usually isn't
2364 // worthwhile looking at everything inbetween.
2365 SmallVector<int64_t, 4> Worklist;
2366 Worklist.push_back(LU.MinOffset);
2367 if (LU.MaxOffset != LU.MinOffset)
2368 Worklist.push_back(LU.MaxOffset);
2370 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2371 const SCEV *G = Base.BaseRegs[i];
2373 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2374 E = Worklist.end(); I != E; ++I) {
2376 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2377 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2378 LU.Kind, LU.AccessTy, TLI)) {
2379 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2381 (void)InsertFormula(LU, LUIdx, F);
2385 int64_t Imm = ExtractImmediate(G, SE);
2386 if (G->isZero() || Imm == 0)
2389 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2390 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2391 LU.Kind, LU.AccessTy, TLI))
2394 (void)InsertFormula(LU, LUIdx, F);
2398 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2399 /// the comparison. For example, x == y -> x*c == y*c.
2400 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2402 if (LU.Kind != LSRUse::ICmpZero) return;
2404 // Determine the integer type for the base formula.
2405 const Type *IntTy = Base.getType();
2407 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2409 // Don't do this if there is more than one offset.
2410 if (LU.MinOffset != LU.MaxOffset) return;
2412 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2414 // Check each interesting stride.
2415 for (SmallSetVector<int64_t, 8>::const_iterator
2416 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2417 int64_t Factor = *I;
2419 // Check that the multiplication doesn't overflow.
2420 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2422 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2423 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2426 // Check that multiplying with the use offset doesn't overflow.
2427 int64_t Offset = LU.MinOffset;
2428 if (Offset == INT64_MIN && Factor == -1)
2430 Offset = (uint64_t)Offset * Factor;
2431 if (Offset / Factor != LU.MinOffset)
2435 F.AM.BaseOffs = NewBaseOffs;
2437 // Check that this scale is legal.
2438 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2441 // Compensate for the use having MinOffset built into it.
2442 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2444 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2446 // Check that multiplying with each base register doesn't overflow.
2447 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2448 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2449 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2453 // Check that multiplying with the scaled register doesn't overflow.
2455 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2456 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2460 // If we make it here and it's legal, add it.
2461 (void)InsertFormula(LU, LUIdx, F);
2466 /// GenerateScales - Generate stride factor reuse formulae by making use of
2467 /// scaled-offset address modes, for example.
2468 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2469 // Determine the integer type for the base formula.
2470 const Type *IntTy = Base.getType();
2473 // If this Formula already has a scaled register, we can't add another one.
2474 if (Base.AM.Scale != 0) return;
2476 // Check each interesting stride.
2477 for (SmallSetVector<int64_t, 8>::const_iterator
2478 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2479 int64_t Factor = *I;
2481 Base.AM.Scale = Factor;
2482 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2483 // Check whether this scale is going to be legal.
2484 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2485 LU.Kind, LU.AccessTy, TLI)) {
2486 // As a special-case, handle special out-of-loop Basic users specially.
2487 // TODO: Reconsider this special case.
2488 if (LU.Kind == LSRUse::Basic &&
2489 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2490 LSRUse::Special, LU.AccessTy, TLI) &&
2491 LU.AllFixupsOutsideLoop)
2492 LU.Kind = LSRUse::Special;
2496 // For an ICmpZero, negating a solitary base register won't lead to
2498 if (LU.Kind == LSRUse::ICmpZero &&
2499 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2501 // For each addrec base reg, apply the scale, if possible.
2502 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2503 if (const SCEVAddRecExpr *AR =
2504 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2505 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2506 if (FactorS->isZero())
2508 // Divide out the factor, ignoring high bits, since we'll be
2509 // scaling the value back up in the end.
2510 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2511 // TODO: This could be optimized to avoid all the copying.
2513 F.ScaledReg = Quotient;
2514 F.DeleteBaseReg(F.BaseRegs[i]);
2515 (void)InsertFormula(LU, LUIdx, F);
2521 /// GenerateTruncates - Generate reuse formulae from different IV types.
2522 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2523 // This requires TargetLowering to tell us which truncates are free.
2526 // Don't bother truncating symbolic values.
2527 if (Base.AM.BaseGV) return;
2529 // Determine the integer type for the base formula.
2530 const Type *DstTy = Base.getType();
2532 DstTy = SE.getEffectiveSCEVType(DstTy);
2534 for (SmallSetVector<const Type *, 4>::const_iterator
2535 I = Types.begin(), E = Types.end(); I != E; ++I) {
2536 const Type *SrcTy = *I;
2537 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2540 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2541 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2542 JE = F.BaseRegs.end(); J != JE; ++J)
2543 *J = SE.getAnyExtendExpr(*J, SrcTy);
2545 // TODO: This assumes we've done basic processing on all uses and
2546 // have an idea what the register usage is.
2547 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2550 (void)InsertFormula(LU, LUIdx, F);
2557 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2558 /// defer modifications so that the search phase doesn't have to worry about
2559 /// the data structures moving underneath it.
2563 const SCEV *OrigReg;
2565 WorkItem(size_t LI, int64_t I, const SCEV *R)
2566 : LUIdx(LI), Imm(I), OrigReg(R) {}
2568 void print(raw_ostream &OS) const;
2574 void WorkItem::print(raw_ostream &OS) const {
2575 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2576 << " , add offset " << Imm;
2579 void WorkItem::dump() const {
2580 print(errs()); errs() << '\n';
2583 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2584 /// distance apart and try to form reuse opportunities between them.
2585 void LSRInstance::GenerateCrossUseConstantOffsets() {
2586 // Group the registers by their value without any added constant offset.
2587 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2588 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2590 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2591 SmallVector<const SCEV *, 8> Sequence;
2592 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2594 const SCEV *Reg = *I;
2595 int64_t Imm = ExtractImmediate(Reg, SE);
2596 std::pair<RegMapTy::iterator, bool> Pair =
2597 Map.insert(std::make_pair(Reg, ImmMapTy()));
2599 Sequence.push_back(Reg);
2600 Pair.first->second.insert(std::make_pair(Imm, *I));
2601 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2604 // Now examine each set of registers with the same base value. Build up
2605 // a list of work to do and do the work in a separate step so that we're
2606 // not adding formulae and register counts while we're searching.
2607 SmallVector<WorkItem, 32> WorkItems;
2608 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2609 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2610 E = Sequence.end(); I != E; ++I) {
2611 const SCEV *Reg = *I;
2612 const ImmMapTy &Imms = Map.find(Reg)->second;
2614 // It's not worthwhile looking for reuse if there's only one offset.
2615 if (Imms.size() == 1)
2618 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2619 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2621 dbgs() << ' ' << J->first;
2624 // Examine each offset.
2625 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2627 const SCEV *OrigReg = J->second;
2629 int64_t JImm = J->first;
2630 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2632 if (!isa<SCEVConstant>(OrigReg) &&
2633 UsedByIndicesMap[Reg].count() == 1) {
2634 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2638 // Conservatively examine offsets between this orig reg a few selected
2640 ImmMapTy::const_iterator OtherImms[] = {
2641 Imms.begin(), prior(Imms.end()),
2642 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2644 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2645 ImmMapTy::const_iterator M = OtherImms[i];
2646 if (M == J || M == JE) continue;
2648 // Compute the difference between the two.
2649 int64_t Imm = (uint64_t)JImm - M->first;
2650 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2651 LUIdx = UsedByIndices.find_next(LUIdx))
2652 // Make a memo of this use, offset, and register tuple.
2653 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2654 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2661 UsedByIndicesMap.clear();
2662 UniqueItems.clear();
2664 // Now iterate through the worklist and add new formulae.
2665 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2666 E = WorkItems.end(); I != E; ++I) {
2667 const WorkItem &WI = *I;
2668 size_t LUIdx = WI.LUIdx;
2669 LSRUse &LU = Uses[LUIdx];
2670 int64_t Imm = WI.Imm;
2671 const SCEV *OrigReg = WI.OrigReg;
2673 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2674 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2675 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2677 // TODO: Use a more targeted data structure.
2678 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2679 const Formula &F = LU.Formulae[L];
2680 // Use the immediate in the scaled register.
2681 if (F.ScaledReg == OrigReg) {
2682 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2683 Imm * (uint64_t)F.AM.Scale;
2684 // Don't create 50 + reg(-50).
2685 if (F.referencesReg(SE.getSCEV(
2686 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2689 NewF.AM.BaseOffs = Offs;
2690 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2691 LU.Kind, LU.AccessTy, TLI))
2693 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2695 // If the new scale is a constant in a register, and adding the constant
2696 // value to the immediate would produce a value closer to zero than the
2697 // immediate itself, then the formula isn't worthwhile.
2698 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2699 if (C->getValue()->getValue().isNegative() !=
2700 (NewF.AM.BaseOffs < 0) &&
2701 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2702 .ule(abs64(NewF.AM.BaseOffs)))
2706 (void)InsertFormula(LU, LUIdx, NewF);
2708 // Use the immediate in a base register.
2709 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2710 const SCEV *BaseReg = F.BaseRegs[N];
2711 if (BaseReg != OrigReg)
2714 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2715 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2716 LU.Kind, LU.AccessTy, TLI))
2718 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2720 // If the new formula has a constant in a register, and adding the
2721 // constant value to the immediate would produce a value closer to
2722 // zero than the immediate itself, then the formula isn't worthwhile.
2723 for (SmallVectorImpl<const SCEV *>::const_iterator
2724 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2726 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2727 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2728 abs64(NewF.AM.BaseOffs)) &&
2729 (C->getValue()->getValue() +
2730 NewF.AM.BaseOffs).countTrailingZeros() >=
2731 CountTrailingZeros_64(NewF.AM.BaseOffs))
2735 (void)InsertFormula(LU, LUIdx, NewF);
2744 /// GenerateAllReuseFormulae - Generate formulae for each use.
2746 LSRInstance::GenerateAllReuseFormulae() {
2747 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2748 // queries are more precise.
2749 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2750 LSRUse &LU = Uses[LUIdx];
2751 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2752 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2753 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2754 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2756 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2757 LSRUse &LU = Uses[LUIdx];
2758 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2759 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2760 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2761 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2762 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2763 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2764 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2765 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2767 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2768 LSRUse &LU = Uses[LUIdx];
2769 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2770 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2773 GenerateCrossUseConstantOffsets();
2776 /// If their are multiple formulae with the same set of registers used
2777 /// by other uses, pick the best one and delete the others.
2778 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2780 bool ChangedFormulae = false;
2783 // Collect the best formula for each unique set of shared registers. This
2784 // is reset for each use.
2785 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2787 BestFormulaeTy BestFormulae;
2789 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2790 LSRUse &LU = Uses[LUIdx];
2791 FormulaSorter Sorter(L, LU, SE, DT);
2792 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2795 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2796 FIdx != NumForms; ++FIdx) {
2797 Formula &F = LU.Formulae[FIdx];
2799 SmallVector<const SCEV *, 2> Key;
2800 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2801 JE = F.BaseRegs.end(); J != JE; ++J) {
2802 const SCEV *Reg = *J;
2803 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2807 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2808 Key.push_back(F.ScaledReg);
2809 // Unstable sort by host order ok, because this is only used for
2811 std::sort(Key.begin(), Key.end());
2813 std::pair<BestFormulaeTy::const_iterator, bool> P =
2814 BestFormulae.insert(std::make_pair(Key, FIdx));
2816 Formula &Best = LU.Formulae[P.first->second];
2817 if (Sorter.operator()(F, Best))
2819 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2821 " in favor of formula "; Best.print(dbgs());
2824 ChangedFormulae = true;
2826 LU.DeleteFormula(F);
2834 // Now that we've filtered out some formulae, recompute the Regs set.
2836 LU.RecomputeRegs(LUIdx, RegUses);
2838 // Reset this to prepare for the next use.
2839 BestFormulae.clear();
2842 DEBUG(if (ChangedFormulae) {
2844 "After filtering out undesirable candidates:\n";
2849 // This is a rough guess that seems to work fairly well.
2850 static const size_t ComplexityLimit = UINT16_MAX;
2852 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2853 /// solutions the solver might have to consider. It almost never considers
2854 /// this many solutions because it prune the search space, but the pruning
2855 /// isn't always sufficient.
2856 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2858 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2859 E = Uses.end(); I != E; ++I) {
2860 size_t FSize = I->Formulae.size();
2861 if (FSize >= ComplexityLimit) {
2862 Power = ComplexityLimit;
2866 if (Power >= ComplexityLimit)
2872 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2873 /// formulae to choose from, use some rough heuristics to prune down the number
2874 /// of formulae. This keeps the main solver from taking an extraordinary amount
2875 /// of time in some worst-case scenarios.
2876 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2877 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2878 DEBUG(dbgs() << "The search space is too complex.\n");
2880 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2881 "which use a superset of registers used by other "
2884 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2885 LSRUse &LU = Uses[LUIdx];
2887 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2888 Formula &F = LU.Formulae[i];
2889 // Look for a formula with a constant or GV in a register. If the use
2890 // also has a formula with that same value in an immediate field,
2891 // delete the one that uses a register.
2892 for (SmallVectorImpl<const SCEV *>::const_iterator
2893 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2894 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2896 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2897 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2898 (I - F.BaseRegs.begin()));
2899 if (LU.HasFormulaWithSameRegs(NewF)) {
2900 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2901 LU.DeleteFormula(F);
2907 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2908 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2911 NewF.AM.BaseGV = GV;
2912 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2913 (I - F.BaseRegs.begin()));
2914 if (LU.HasFormulaWithSameRegs(NewF)) {
2915 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2917 LU.DeleteFormula(F);
2928 LU.RecomputeRegs(LUIdx, RegUses);
2931 DEBUG(dbgs() << "After pre-selection:\n";
2932 print_uses(dbgs()));
2935 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2936 DEBUG(dbgs() << "The search space is too complex.\n");
2938 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2939 "separated by a constant offset will use the same "
2942 // This is especially useful for unrolled loops.
2944 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2945 LSRUse &LU = Uses[LUIdx];
2946 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2947 E = LU.Formulae.end(); I != E; ++I) {
2948 const Formula &F = *I;
2949 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2950 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2951 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2952 /*HasBaseReg=*/false,
2953 LU.Kind, LU.AccessTy)) {
2954 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2957 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2959 // Delete formulae from the new use which are no longer legal.
2961 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2962 Formula &F = LUThatHas->Formulae[i];
2963 if (!isLegalUse(F.AM,
2964 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2965 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2966 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2968 LUThatHas->DeleteFormula(F);
2975 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
2977 // Update the relocs to reference the new use.
2978 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
2979 E = Fixups.end(); I != E; ++I) {
2980 LSRFixup &Fixup = *I;
2981 if (Fixup.LUIdx == LUIdx) {
2982 Fixup.LUIdx = LUThatHas - &Uses.front();
2983 Fixup.Offset += F.AM.BaseOffs;
2984 DEBUG(errs() << "New fixup has offset "
2985 << Fixup.Offset << '\n');
2987 if (Fixup.LUIdx == NumUses-1)
2988 Fixup.LUIdx = LUIdx;
2991 // Delete the old use.
3002 DEBUG(dbgs() << "After pre-selection:\n";
3003 print_uses(dbgs()));
3006 // With all other options exhausted, loop until the system is simple
3007 // enough to handle.
3008 SmallPtrSet<const SCEV *, 4> Taken;
3009 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3010 // Ok, we have too many of formulae on our hands to conveniently handle.
3011 // Use a rough heuristic to thin out the list.
3012 DEBUG(dbgs() << "The search space is too complex.\n");
3014 // Pick the register which is used by the most LSRUses, which is likely
3015 // to be a good reuse register candidate.
3016 const SCEV *Best = 0;
3017 unsigned BestNum = 0;
3018 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3020 const SCEV *Reg = *I;
3021 if (Taken.count(Reg))
3026 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3027 if (Count > BestNum) {
3034 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3035 << " will yield profitable reuse.\n");
3038 // In any use with formulae which references this register, delete formulae
3039 // which don't reference it.
3040 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3041 LSRUse &LU = Uses[LUIdx];
3042 if (!LU.Regs.count(Best)) continue;
3045 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3046 Formula &F = LU.Formulae[i];
3047 if (!F.referencesReg(Best)) {
3048 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3049 LU.DeleteFormula(F);
3053 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3059 LU.RecomputeRegs(LUIdx, RegUses);
3062 DEBUG(dbgs() << "After pre-selection:\n";
3063 print_uses(dbgs()));
3067 /// SolveRecurse - This is the recursive solver.
3068 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3070 SmallVectorImpl<const Formula *> &Workspace,
3071 const Cost &CurCost,
3072 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3073 DenseSet<const SCEV *> &VisitedRegs) const {
3076 // - use more aggressive filtering
3077 // - sort the formula so that the most profitable solutions are found first
3078 // - sort the uses too
3080 // - don't compute a cost, and then compare. compare while computing a cost
3082 // - track register sets with SmallBitVector
3084 const LSRUse &LU = Uses[Workspace.size()];
3086 // If this use references any register that's already a part of the
3087 // in-progress solution, consider it a requirement that a formula must
3088 // reference that register in order to be considered. This prunes out
3089 // unprofitable searching.
3090 SmallSetVector<const SCEV *, 4> ReqRegs;
3091 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3092 E = CurRegs.end(); I != E; ++I)
3093 if (LU.Regs.count(*I))
3096 bool AnySatisfiedReqRegs = false;
3097 SmallPtrSet<const SCEV *, 16> NewRegs;
3100 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3101 E = LU.Formulae.end(); I != E; ++I) {
3102 const Formula &F = *I;
3104 // Ignore formulae which do not use any of the required registers.
3105 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3106 JE = ReqRegs.end(); J != JE; ++J) {
3107 const SCEV *Reg = *J;
3108 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3109 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3113 AnySatisfiedReqRegs = true;
3115 // Evaluate the cost of the current formula. If it's already worse than
3116 // the current best, prune the search at that point.
3119 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3120 if (NewCost < SolutionCost) {
3121 Workspace.push_back(&F);
3122 if (Workspace.size() != Uses.size()) {
3123 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3124 NewRegs, VisitedRegs);
3125 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3126 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3128 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3129 dbgs() << ". Regs:";
3130 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3131 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3132 dbgs() << ' ' << **I;
3135 SolutionCost = NewCost;
3136 Solution = Workspace;
3138 Workspace.pop_back();
3143 // If none of the formulae had all of the required registers, relax the
3144 // constraint so that we don't exclude all formulae.
3145 if (!AnySatisfiedReqRegs) {
3146 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3152 /// Solve - Choose one formula from each use. Return the results in the given
3153 /// Solution vector.
3154 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3155 SmallVector<const Formula *, 8> Workspace;
3157 SolutionCost.Loose();
3159 SmallPtrSet<const SCEV *, 16> CurRegs;
3160 DenseSet<const SCEV *> VisitedRegs;
3161 Workspace.reserve(Uses.size());
3163 // SolveRecurse does all the work.
3164 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3165 CurRegs, VisitedRegs);
3167 // Ok, we've now made all our decisions.
3168 DEBUG(dbgs() << "\n"
3169 "The chosen solution requires "; SolutionCost.print(dbgs());
3171 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3173 Uses[i].print(dbgs());
3176 Solution[i]->print(dbgs());
3180 assert(Solution.size() == Uses.size() && "Malformed solution!");
3183 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3184 /// the dominator tree far as we can go while still being dominated by the
3185 /// input positions. This helps canonicalize the insert position, which
3186 /// encourages sharing.
3187 BasicBlock::iterator
3188 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3189 const SmallVectorImpl<Instruction *> &Inputs)
3192 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3193 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3196 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3197 if (!Rung) return IP;
3198 Rung = Rung->getIDom();
3199 if (!Rung) return IP;
3200 IDom = Rung->getBlock();
3202 // Don't climb into a loop though.
3203 const Loop *IDomLoop = LI.getLoopFor(IDom);
3204 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3205 if (IDomDepth <= IPLoopDepth &&
3206 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3210 bool AllDominate = true;
3211 Instruction *BetterPos = 0;
3212 Instruction *Tentative = IDom->getTerminator();
3213 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3214 E = Inputs.end(); I != E; ++I) {
3215 Instruction *Inst = *I;
3216 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3217 AllDominate = false;
3220 // Attempt to find an insert position in the middle of the block,
3221 // instead of at the end, so that it can be used for other expansions.
3222 if (IDom == Inst->getParent() &&
3223 (!BetterPos || DT.dominates(BetterPos, Inst)))
3224 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3237 /// AdjustInsertPositionForExpand - Determine an input position which will be
3238 /// dominated by the operands and which will dominate the result.
3239 BasicBlock::iterator
3240 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3242 const LSRUse &LU) const {
3243 // Collect some instructions which must be dominated by the
3244 // expanding replacement. These must be dominated by any operands that
3245 // will be required in the expansion.
3246 SmallVector<Instruction *, 4> Inputs;
3247 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3248 Inputs.push_back(I);
3249 if (LU.Kind == LSRUse::ICmpZero)
3250 if (Instruction *I =
3251 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3252 Inputs.push_back(I);
3253 if (LF.PostIncLoops.count(L)) {
3254 if (LF.isUseFullyOutsideLoop(L))
3255 Inputs.push_back(L->getLoopLatch()->getTerminator());
3257 Inputs.push_back(IVIncInsertPos);
3259 // The expansion must also be dominated by the increment positions of any
3260 // loops it for which it is using post-inc mode.
3261 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3262 E = LF.PostIncLoops.end(); I != E; ++I) {
3263 const Loop *PIL = *I;
3264 if (PIL == L) continue;
3266 // Be dominated by the loop exit.
3267 SmallVector<BasicBlock *, 4> ExitingBlocks;
3268 PIL->getExitingBlocks(ExitingBlocks);
3269 if (!ExitingBlocks.empty()) {
3270 BasicBlock *BB = ExitingBlocks[0];
3271 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3272 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3273 Inputs.push_back(BB->getTerminator());
3277 // Then, climb up the immediate dominator tree as far as we can go while
3278 // still being dominated by the input positions.
3279 IP = HoistInsertPosition(IP, Inputs);
3281 // Don't insert instructions before PHI nodes.
3282 while (isa<PHINode>(IP)) ++IP;
3284 // Ignore debug intrinsics.
3285 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3290 /// Expand - Emit instructions for the leading candidate expression for this
3291 /// LSRUse (this is called "expanding").
3292 Value *LSRInstance::Expand(const LSRFixup &LF,
3294 BasicBlock::iterator IP,
3295 SCEVExpander &Rewriter,
3296 SmallVectorImpl<WeakVH> &DeadInsts) const {
3297 const LSRUse &LU = Uses[LF.LUIdx];
3299 // Determine an input position which will be dominated by the operands and
3300 // which will dominate the result.
3301 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3303 // Inform the Rewriter if we have a post-increment use, so that it can
3304 // perform an advantageous expansion.
3305 Rewriter.setPostInc(LF.PostIncLoops);
3307 // This is the type that the user actually needs.
3308 const Type *OpTy = LF.OperandValToReplace->getType();
3309 // This will be the type that we'll initially expand to.
3310 const Type *Ty = F.getType();
3312 // No type known; just expand directly to the ultimate type.
3314 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3315 // Expand directly to the ultimate type if it's the right size.
3317 // This is the type to do integer arithmetic in.
3318 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3320 // Build up a list of operands to add together to form the full base.
3321 SmallVector<const SCEV *, 8> Ops;
3323 // Expand the BaseRegs portion.
3324 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3325 E = F.BaseRegs.end(); I != E; ++I) {
3326 const SCEV *Reg = *I;
3327 assert(!Reg->isZero() && "Zero allocated in a base register!");
3329 // If we're expanding for a post-inc user, make the post-inc adjustment.
3330 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3331 Reg = TransformForPostIncUse(Denormalize, Reg,
3332 LF.UserInst, LF.OperandValToReplace,
3335 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3338 // Flush the operand list to suppress SCEVExpander hoisting.
3340 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3342 Ops.push_back(SE.getUnknown(FullV));
3345 // Expand the ScaledReg portion.
3346 Value *ICmpScaledV = 0;
3347 if (F.AM.Scale != 0) {
3348 const SCEV *ScaledS = F.ScaledReg;
3350 // If we're expanding for a post-inc user, make the post-inc adjustment.
3351 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3352 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3353 LF.UserInst, LF.OperandValToReplace,
3356 if (LU.Kind == LSRUse::ICmpZero) {
3357 // An interesting way of "folding" with an icmp is to use a negated
3358 // scale, which we'll implement by inserting it into the other operand
3360 assert(F.AM.Scale == -1 &&
3361 "The only scale supported by ICmpZero uses is -1!");
3362 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3364 // Otherwise just expand the scaled register and an explicit scale,
3365 // which is expected to be matched as part of the address.
3366 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3367 ScaledS = SE.getMulExpr(ScaledS,
3368 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3369 Ops.push_back(ScaledS);
3371 // Flush the operand list to suppress SCEVExpander hoisting.
3372 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3374 Ops.push_back(SE.getUnknown(FullV));
3378 // Expand the GV portion.
3380 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3382 // Flush the operand list to suppress SCEVExpander hoisting.
3383 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3385 Ops.push_back(SE.getUnknown(FullV));
3388 // Expand the immediate portion.
3389 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3391 if (LU.Kind == LSRUse::ICmpZero) {
3392 // The other interesting way of "folding" with an ICmpZero is to use a
3393 // negated immediate.
3395 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3397 Ops.push_back(SE.getUnknown(ICmpScaledV));
3398 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3401 // Just add the immediate values. These again are expected to be matched
3402 // as part of the address.
3403 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3407 // Emit instructions summing all the operands.
3408 const SCEV *FullS = Ops.empty() ?
3409 SE.getConstant(IntTy, 0) :
3411 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3413 // We're done expanding now, so reset the rewriter.
3414 Rewriter.clearPostInc();
3416 // An ICmpZero Formula represents an ICmp which we're handling as a
3417 // comparison against zero. Now that we've expanded an expression for that
3418 // form, update the ICmp's other operand.
3419 if (LU.Kind == LSRUse::ICmpZero) {
3420 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3421 DeadInsts.push_back(CI->getOperand(1));
3422 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3423 "a scale at the same time!");
3424 if (F.AM.Scale == -1) {
3425 if (ICmpScaledV->getType() != OpTy) {
3427 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3429 ICmpScaledV, OpTy, "tmp", CI);
3432 CI->setOperand(1, ICmpScaledV);
3434 assert(F.AM.Scale == 0 &&
3435 "ICmp does not support folding a global value and "
3436 "a scale at the same time!");
3437 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3439 if (C->getType() != OpTy)
3440 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3444 CI->setOperand(1, C);
3451 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3452 /// of their operands effectively happens in their predecessor blocks, so the
3453 /// expression may need to be expanded in multiple places.
3454 void LSRInstance::RewriteForPHI(PHINode *PN,
3457 SCEVExpander &Rewriter,
3458 SmallVectorImpl<WeakVH> &DeadInsts,
3460 DenseMap<BasicBlock *, Value *> Inserted;
3461 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3462 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3463 BasicBlock *BB = PN->getIncomingBlock(i);
3465 // If this is a critical edge, split the edge so that we do not insert
3466 // the code on all predecessor/successor paths. We do this unless this
3467 // is the canonical backedge for this loop, which complicates post-inc
3469 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3470 !isa<IndirectBrInst>(BB->getTerminator()) &&
3471 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3472 // Split the critical edge.
3473 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3475 // If PN is outside of the loop and BB is in the loop, we want to
3476 // move the block to be immediately before the PHI block, not
3477 // immediately after BB.
3478 if (L->contains(BB) && !L->contains(PN))
3479 NewBB->moveBefore(PN->getParent());
3481 // Splitting the edge can reduce the number of PHI entries we have.
3482 e = PN->getNumIncomingValues();
3484 i = PN->getBasicBlockIndex(BB);
3487 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3488 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3490 PN->setIncomingValue(i, Pair.first->second);
3492 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3494 // If this is reuse-by-noop-cast, insert the noop cast.
3495 const Type *OpTy = LF.OperandValToReplace->getType();
3496 if (FullV->getType() != OpTy)
3498 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3500 FullV, LF.OperandValToReplace->getType(),
3501 "tmp", BB->getTerminator());
3503 PN->setIncomingValue(i, FullV);
3504 Pair.first->second = FullV;
3509 /// Rewrite - Emit instructions for the leading candidate expression for this
3510 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3511 /// the newly expanded value.
3512 void LSRInstance::Rewrite(const LSRFixup &LF,
3514 SCEVExpander &Rewriter,
3515 SmallVectorImpl<WeakVH> &DeadInsts,
3517 // First, find an insertion point that dominates UserInst. For PHI nodes,
3518 // find the nearest block which dominates all the relevant uses.
3519 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3520 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3522 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3524 // If this is reuse-by-noop-cast, insert the noop cast.
3525 const Type *OpTy = LF.OperandValToReplace->getType();
3526 if (FullV->getType() != OpTy) {
3528 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3529 FullV, OpTy, "tmp", LF.UserInst);
3533 // Update the user. ICmpZero is handled specially here (for now) because
3534 // Expand may have updated one of the operands of the icmp already, and
3535 // its new value may happen to be equal to LF.OperandValToReplace, in
3536 // which case doing replaceUsesOfWith leads to replacing both operands
3537 // with the same value. TODO: Reorganize this.
3538 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3539 LF.UserInst->setOperand(0, FullV);
3541 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3544 DeadInsts.push_back(LF.OperandValToReplace);
3547 /// ImplementSolution - Rewrite all the fixup locations with new values,
3548 /// following the chosen solution.
3550 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3552 // Keep track of instructions we may have made dead, so that
3553 // we can remove them after we are done working.
3554 SmallVector<WeakVH, 16> DeadInsts;
3556 SCEVExpander Rewriter(SE);
3557 Rewriter.disableCanonicalMode();
3558 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3560 // Expand the new value definitions and update the users.
3561 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3562 E = Fixups.end(); I != E; ++I) {
3563 const LSRFixup &Fixup = *I;
3565 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3570 // Clean up after ourselves. This must be done before deleting any
3574 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3577 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3578 : IU(P->getAnalysis<IVUsers>()),
3579 SE(P->getAnalysis<ScalarEvolution>()),
3580 DT(P->getAnalysis<DominatorTree>()),
3581 LI(P->getAnalysis<LoopInfo>()),
3582 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3584 // If LoopSimplify form is not available, stay out of trouble.
3585 if (!L->isLoopSimplifyForm()) return;
3587 // If there's no interesting work to be done, bail early.
3588 if (IU.empty()) return;
3590 DEBUG(dbgs() << "\nLSR on loop ";
3591 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3594 // First, perform some low-level loop optimizations.
3596 OptimizeLoopTermCond();
3598 // Start collecting data and preparing for the solver.
3599 CollectInterestingTypesAndFactors();
3600 CollectFixupsAndInitialFormulae();
3601 CollectLoopInvariantFixupsAndFormulae();
3603 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3604 print_uses(dbgs()));
3606 // Now use the reuse data to generate a bunch of interesting ways
3607 // to formulate the values needed for the uses.
3608 GenerateAllReuseFormulae();
3610 DEBUG(dbgs() << "\n"
3611 "After generating reuse formulae:\n";
3612 print_uses(dbgs()));
3614 FilterOutUndesirableDedicatedRegisters();
3615 NarrowSearchSpaceUsingHeuristics();
3617 SmallVector<const Formula *, 8> Solution;
3620 // Release memory that is no longer needed.
3626 // Formulae should be legal.
3627 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3628 E = Uses.end(); I != E; ++I) {
3629 const LSRUse &LU = *I;
3630 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3631 JE = LU.Formulae.end(); J != JE; ++J)
3632 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3633 LU.Kind, LU.AccessTy, TLI) &&
3634 "Illegal formula generated!");
3638 // Now that we've decided what we want, make it so.
3639 ImplementSolution(Solution, P);
3642 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3643 if (Factors.empty() && Types.empty()) return;
3645 OS << "LSR has identified the following interesting factors and types: ";
3648 for (SmallSetVector<int64_t, 8>::const_iterator
3649 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3650 if (!First) OS << ", ";
3655 for (SmallSetVector<const Type *, 4>::const_iterator
3656 I = Types.begin(), E = Types.end(); I != E; ++I) {
3657 if (!First) OS << ", ";
3659 OS << '(' << **I << ')';
3664 void LSRInstance::print_fixups(raw_ostream &OS) const {
3665 OS << "LSR is examining the following fixup sites:\n";
3666 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3667 E = Fixups.end(); I != E; ++I) {
3674 void LSRInstance::print_uses(raw_ostream &OS) const {
3675 OS << "LSR is examining the following uses:\n";
3676 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3677 E = Uses.end(); I != E; ++I) {
3678 const LSRUse &LU = *I;
3682 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3683 JE = LU.Formulae.end(); J != JE; ++J) {
3691 void LSRInstance::print(raw_ostream &OS) const {
3692 print_factors_and_types(OS);
3697 void LSRInstance::dump() const {
3698 print(errs()); errs() << '\n';
3703 class LoopStrengthReduce : public LoopPass {
3704 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3705 /// transformation profitability.
3706 const TargetLowering *const TLI;
3709 static char ID; // Pass ID, replacement for typeid
3710 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3713 bool runOnLoop(Loop *L, LPPassManager &LPM);
3714 void getAnalysisUsage(AnalysisUsage &AU) const;
3719 char LoopStrengthReduce::ID = 0;
3720 static RegisterPass<LoopStrengthReduce>
3721 X("loop-reduce", "Loop Strength Reduction");
3723 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3724 return new LoopStrengthReduce(TLI);
3727 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3728 : LoopPass(&ID), TLI(tli) {}
3730 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3731 // We split critical edges, so we change the CFG. However, we do update
3732 // many analyses if they are around.
3733 AU.addPreservedID(LoopSimplifyID);
3734 AU.addPreserved("domfrontier");
3736 AU.addRequired<LoopInfo>();
3737 AU.addPreserved<LoopInfo>();
3738 AU.addRequiredID(LoopSimplifyID);
3739 AU.addRequired<DominatorTree>();
3740 AU.addPreserved<DominatorTree>();
3741 AU.addRequired<ScalarEvolution>();
3742 AU.addPreserved<ScalarEvolution>();
3743 AU.addRequired<IVUsers>();
3744 AU.addPreserved<IVUsers>();
3747 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3748 bool Changed = false;
3750 // Run the main LSR transformation.
3751 Changed |= LSRInstance(TLI, L, this).getChanged();
3753 // At this point, it is worth checking to see if any recurrence PHIs are also
3754 // dead, so that we can remove them as well.
3755 Changed |= DeleteDeadPHIs(L->getHeader());