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());
454 // Distribute the sdiv over add operands, if the add doesn't overflow.
455 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
456 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
457 SmallVector<const SCEV *, 8> Ops;
458 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
460 const SCEV *Op = getExactSDiv(*I, RHS, SE,
461 IgnoreSignificantBits);
465 return SE.getAddExpr(Ops);
469 // Check for a multiply operand that we can pull RHS out of.
470 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
471 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
472 SmallVector<const SCEV *, 4> Ops;
474 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
478 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
479 IgnoreSignificantBits)) {
485 return Found ? SE.getMulExpr(Ops) : 0;
488 // Otherwise we don't know.
492 /// ExtractImmediate - If S involves the addition of a constant integer value,
493 /// return that integer value, and mutate S to point to a new SCEV with that
495 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
496 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
497 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
498 S = SE.getConstant(C->getType(), 0);
499 return C->getValue()->getSExtValue();
501 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
502 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
503 int64_t Result = ExtractImmediate(NewOps.front(), SE);
504 S = SE.getAddExpr(NewOps);
506 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
507 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
508 int64_t Result = ExtractImmediate(NewOps.front(), SE);
509 S = SE.getAddRecExpr(NewOps, AR->getLoop());
515 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
516 /// return that symbol, and mutate S to point to a new SCEV with that
518 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
519 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
520 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
521 S = SE.getConstant(GV->getType(), 0);
524 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
525 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
526 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
527 S = SE.getAddExpr(NewOps);
529 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
530 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
531 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
532 S = SE.getAddRecExpr(NewOps, AR->getLoop());
538 /// isAddressUse - Returns true if the specified instruction is using the
539 /// specified value as an address.
540 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
541 bool isAddress = isa<LoadInst>(Inst);
542 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
543 if (SI->getOperand(1) == OperandVal)
545 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
546 // Addressing modes can also be folded into prefetches and a variety
548 switch (II->getIntrinsicID()) {
550 case Intrinsic::prefetch:
551 case Intrinsic::x86_sse2_loadu_dq:
552 case Intrinsic::x86_sse2_loadu_pd:
553 case Intrinsic::x86_sse_loadu_ps:
554 case Intrinsic::x86_sse_storeu_ps:
555 case Intrinsic::x86_sse2_storeu_pd:
556 case Intrinsic::x86_sse2_storeu_dq:
557 case Intrinsic::x86_sse2_storel_dq:
558 if (II->getOperand(1) == OperandVal)
566 /// getAccessType - Return the type of the memory being accessed.
567 static const Type *getAccessType(const Instruction *Inst) {
568 const Type *AccessTy = Inst->getType();
569 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
570 AccessTy = SI->getOperand(0)->getType();
571 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
572 // Addressing modes can also be folded into prefetches and a variety
574 switch (II->getIntrinsicID()) {
576 case Intrinsic::x86_sse_storeu_ps:
577 case Intrinsic::x86_sse2_storeu_pd:
578 case Intrinsic::x86_sse2_storeu_dq:
579 case Intrinsic::x86_sse2_storel_dq:
580 AccessTy = II->getOperand(1)->getType();
585 // All pointers have the same requirements, so canonicalize them to an
586 // arbitrary pointer type to minimize variation.
587 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
588 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
589 PTy->getAddressSpace());
594 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
595 /// specified set are trivially dead, delete them and see if this makes any of
596 /// their operands subsequently dead.
598 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
599 bool Changed = false;
601 while (!DeadInsts.empty()) {
602 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
604 if (I == 0 || !isInstructionTriviallyDead(I))
607 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
608 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
611 DeadInsts.push_back(U);
614 I->eraseFromParent();
623 /// Cost - This class is used to measure and compare candidate formulae.
625 /// TODO: Some of these could be merged. Also, a lexical ordering
626 /// isn't always optimal.
630 unsigned NumBaseAdds;
636 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
639 unsigned getNumRegs() const { return NumRegs; }
641 bool operator<(const Cost &Other) const;
645 void RateFormula(const Formula &F,
646 SmallPtrSet<const SCEV *, 16> &Regs,
647 const DenseSet<const SCEV *> &VisitedRegs,
649 const SmallVectorImpl<int64_t> &Offsets,
650 ScalarEvolution &SE, DominatorTree &DT);
652 void print(raw_ostream &OS) const;
656 void RateRegister(const SCEV *Reg,
657 SmallPtrSet<const SCEV *, 16> &Regs,
659 ScalarEvolution &SE, DominatorTree &DT);
660 void RatePrimaryRegister(const SCEV *Reg,
661 SmallPtrSet<const SCEV *, 16> &Regs,
663 ScalarEvolution &SE, DominatorTree &DT);
668 /// RateRegister - Tally up interesting quantities from the given register.
669 void Cost::RateRegister(const SCEV *Reg,
670 SmallPtrSet<const SCEV *, 16> &Regs,
672 ScalarEvolution &SE, DominatorTree &DT) {
673 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
674 if (AR->getLoop() == L)
675 AddRecCost += 1; /// TODO: This should be a function of the stride.
677 // If this is an addrec for a loop that's already been visited by LSR,
678 // don't second-guess its addrec phi nodes. LSR isn't currently smart
679 // enough to reason about more than one loop at a time. Consider these
680 // registers free and leave them alone.
681 else if (L->contains(AR->getLoop()) ||
682 (!AR->getLoop()->contains(L) &&
683 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
684 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
685 PHINode *PN = dyn_cast<PHINode>(I); ++I)
686 if (SE.isSCEVable(PN->getType()) &&
687 (SE.getEffectiveSCEVType(PN->getType()) ==
688 SE.getEffectiveSCEVType(AR->getType())) &&
689 SE.getSCEV(PN) == AR)
692 // If this isn't one of the addrecs that the loop already has, it
693 // would require a costly new phi and add. TODO: This isn't
694 // precisely modeled right now.
696 if (!Regs.count(AR->getStart()))
697 RateRegister(AR->getStart(), Regs, L, SE, DT);
700 // Add the step value register, if it needs one.
701 // TODO: The non-affine case isn't precisely modeled here.
702 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
703 if (!Regs.count(AR->getStart()))
704 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
708 // Rough heuristic; favor registers which don't require extra setup
709 // instructions in the preheader.
710 if (!isa<SCEVUnknown>(Reg) &&
711 !isa<SCEVConstant>(Reg) &&
712 !(isa<SCEVAddRecExpr>(Reg) &&
713 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
714 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
718 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
720 void Cost::RatePrimaryRegister(const SCEV *Reg,
721 SmallPtrSet<const SCEV *, 16> &Regs,
723 ScalarEvolution &SE, DominatorTree &DT) {
724 if (Regs.insert(Reg))
725 RateRegister(Reg, Regs, L, SE, DT);
728 void Cost::RateFormula(const Formula &F,
729 SmallPtrSet<const SCEV *, 16> &Regs,
730 const DenseSet<const SCEV *> &VisitedRegs,
732 const SmallVectorImpl<int64_t> &Offsets,
733 ScalarEvolution &SE, DominatorTree &DT) {
734 // Tally up the registers.
735 if (const SCEV *ScaledReg = F.ScaledReg) {
736 if (VisitedRegs.count(ScaledReg)) {
740 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
742 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
743 E = F.BaseRegs.end(); I != E; ++I) {
744 const SCEV *BaseReg = *I;
745 if (VisitedRegs.count(BaseReg)) {
749 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
751 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
752 BaseReg->hasComputableLoopEvolution(L);
755 if (F.BaseRegs.size() > 1)
756 NumBaseAdds += F.BaseRegs.size() - 1;
758 // Tally up the non-zero immediates.
759 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
760 E = Offsets.end(); I != E; ++I) {
761 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
763 ImmCost += 64; // Handle symbolic values conservatively.
764 // TODO: This should probably be the pointer size.
765 else if (Offset != 0)
766 ImmCost += APInt(64, Offset, true).getMinSignedBits();
770 /// Loose - Set this cost to a loosing value.
780 /// operator< - Choose the lower cost.
781 bool Cost::operator<(const Cost &Other) const {
782 if (NumRegs != Other.NumRegs)
783 return NumRegs < Other.NumRegs;
784 if (AddRecCost != Other.AddRecCost)
785 return AddRecCost < Other.AddRecCost;
786 if (NumIVMuls != Other.NumIVMuls)
787 return NumIVMuls < Other.NumIVMuls;
788 if (NumBaseAdds != Other.NumBaseAdds)
789 return NumBaseAdds < Other.NumBaseAdds;
790 if (ImmCost != Other.ImmCost)
791 return ImmCost < Other.ImmCost;
792 if (SetupCost != Other.SetupCost)
793 return SetupCost < Other.SetupCost;
797 void Cost::print(raw_ostream &OS) const {
798 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
800 OS << ", with addrec cost " << AddRecCost;
802 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
803 if (NumBaseAdds != 0)
804 OS << ", plus " << NumBaseAdds << " base add"
805 << (NumBaseAdds == 1 ? "" : "s");
807 OS << ", plus " << ImmCost << " imm cost";
809 OS << ", plus " << SetupCost << " setup cost";
812 void Cost::dump() const {
813 print(errs()); errs() << '\n';
818 /// LSRFixup - An operand value in an instruction which is to be replaced
819 /// with some equivalent, possibly strength-reduced, replacement.
821 /// UserInst - The instruction which will be updated.
822 Instruction *UserInst;
824 /// OperandValToReplace - The operand of the instruction which will
825 /// be replaced. The operand may be used more than once; every instance
826 /// will be replaced.
827 Value *OperandValToReplace;
829 /// PostIncLoops - If this user is to use the post-incremented value of an
830 /// induction variable, this variable is non-null and holds the loop
831 /// associated with the induction variable.
832 PostIncLoopSet PostIncLoops;
834 /// LUIdx - The index of the LSRUse describing the expression which
835 /// this fixup needs, minus an offset (below).
838 /// Offset - A constant offset to be added to the LSRUse expression.
839 /// This allows multiple fixups to share the same LSRUse with different
840 /// offsets, for example in an unrolled loop.
843 bool isUseFullyOutsideLoop(const Loop *L) const;
847 void print(raw_ostream &OS) const;
854 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
856 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
857 /// value outside of the given loop.
858 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
859 // PHI nodes use their value in their incoming blocks.
860 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
861 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
862 if (PN->getIncomingValue(i) == OperandValToReplace &&
863 L->contains(PN->getIncomingBlock(i)))
868 return !L->contains(UserInst);
871 void LSRFixup::print(raw_ostream &OS) const {
873 // Store is common and interesting enough to be worth special-casing.
874 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
876 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
877 } else if (UserInst->getType()->isVoidTy())
878 OS << UserInst->getOpcodeName();
880 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
882 OS << ", OperandValToReplace=";
883 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
885 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
886 E = PostIncLoops.end(); I != E; ++I) {
887 OS << ", PostIncLoop=";
888 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
891 if (LUIdx != ~size_t(0))
892 OS << ", LUIdx=" << LUIdx;
895 OS << ", Offset=" << Offset;
898 void LSRFixup::dump() const {
899 print(errs()); errs() << '\n';
904 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
905 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
906 struct UniquifierDenseMapInfo {
907 static SmallVector<const SCEV *, 2> getEmptyKey() {
908 SmallVector<const SCEV *, 2> V;
909 V.push_back(reinterpret_cast<const SCEV *>(-1));
913 static SmallVector<const SCEV *, 2> getTombstoneKey() {
914 SmallVector<const SCEV *, 2> V;
915 V.push_back(reinterpret_cast<const SCEV *>(-2));
919 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
921 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
922 E = V.end(); I != E; ++I)
923 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
927 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
928 const SmallVector<const SCEV *, 2> &RHS) {
933 /// LSRUse - This class holds the state that LSR keeps for each use in
934 /// IVUsers, as well as uses invented by LSR itself. It includes information
935 /// about what kinds of things can be folded into the user, information about
936 /// the user itself, and information about how the use may be satisfied.
937 /// TODO: Represent multiple users of the same expression in common?
939 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
942 /// KindType - An enum for a kind of use, indicating what types of
943 /// scaled and immediate operands it might support.
945 Basic, ///< A normal use, with no folding.
946 Special, ///< A special case of basic, allowing -1 scales.
947 Address, ///< An address use; folding according to TargetLowering
948 ICmpZero ///< An equality icmp with both operands folded into one.
949 // TODO: Add a generic icmp too?
953 const Type *AccessTy;
955 SmallVector<int64_t, 8> Offsets;
959 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
960 /// LSRUse are outside of the loop, in which case some special-case heuristics
962 bool AllFixupsOutsideLoop;
964 /// Formulae - A list of ways to build a value that can satisfy this user.
965 /// After the list is populated, one of these is selected heuristically and
966 /// used to formulate a replacement for OperandValToReplace in UserInst.
967 SmallVector<Formula, 12> Formulae;
969 /// Regs - The set of register candidates used by all formulae in this LSRUse.
970 SmallPtrSet<const SCEV *, 4> Regs;
972 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
973 MinOffset(INT64_MAX),
974 MaxOffset(INT64_MIN),
975 AllFixupsOutsideLoop(true) {}
977 bool HasFormulaWithSameRegs(const Formula &F) const;
978 bool InsertFormula(const Formula &F);
979 void DeleteFormula(Formula &F);
980 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
984 void print(raw_ostream &OS) const;
990 /// HasFormula - Test whether this use as a formula which has the same
991 /// registers as the given formula.
992 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
993 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
994 if (F.ScaledReg) Key.push_back(F.ScaledReg);
995 // Unstable sort by host order ok, because this is only used for uniquifying.
996 std::sort(Key.begin(), Key.end());
997 return Uniquifier.count(Key);
1000 /// InsertFormula - If the given formula has not yet been inserted, add it to
1001 /// the list, and return true. Return false otherwise.
1002 bool LSRUse::InsertFormula(const Formula &F) {
1003 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1004 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1005 // Unstable sort by host order ok, because this is only used for uniquifying.
1006 std::sort(Key.begin(), Key.end());
1008 if (!Uniquifier.insert(Key).second)
1011 // Using a register to hold the value of 0 is not profitable.
1012 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1013 "Zero allocated in a scaled register!");
1015 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1016 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1017 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1020 // Add the formula to the list.
1021 Formulae.push_back(F);
1023 // Record registers now being used by this use.
1024 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1025 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1030 /// DeleteFormula - Remove the given formula from this use's list.
1031 void LSRUse::DeleteFormula(Formula &F) {
1032 if (&F != &Formulae.back())
1033 std::swap(F, Formulae.back());
1034 Formulae.pop_back();
1035 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1038 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1039 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1040 // Now that we've filtered out some formulae, recompute the Regs set.
1041 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1043 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1044 E = Formulae.end(); I != E; ++I) {
1045 const Formula &F = *I;
1046 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1047 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1050 // Update the RegTracker.
1051 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1052 E = OldRegs.end(); I != E; ++I)
1053 if (!Regs.count(*I))
1054 RegUses.DropRegister(*I, LUIdx);
1057 void LSRUse::print(raw_ostream &OS) const {
1058 OS << "LSR Use: Kind=";
1060 case Basic: OS << "Basic"; break;
1061 case Special: OS << "Special"; break;
1062 case ICmpZero: OS << "ICmpZero"; break;
1064 OS << "Address of ";
1065 if (AccessTy->isPointerTy())
1066 OS << "pointer"; // the full pointer type could be really verbose
1071 OS << ", Offsets={";
1072 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1073 E = Offsets.end(); I != E; ++I) {
1080 if (AllFixupsOutsideLoop)
1081 OS << ", all-fixups-outside-loop";
1084 void LSRUse::dump() const {
1085 print(errs()); errs() << '\n';
1088 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1089 /// be completely folded into the user instruction at isel time. This includes
1090 /// address-mode folding and special icmp tricks.
1091 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1092 LSRUse::KindType Kind, const Type *AccessTy,
1093 const TargetLowering *TLI) {
1095 case LSRUse::Address:
1096 // If we have low-level target information, ask the target if it can
1097 // completely fold this address.
1098 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1100 // Otherwise, just guess that reg+reg addressing is legal.
1101 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1103 case LSRUse::ICmpZero:
1104 // There's not even a target hook for querying whether it would be legal to
1105 // fold a GV into an ICmp.
1109 // ICmp only has two operands; don't allow more than two non-trivial parts.
1110 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1113 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1114 // putting the scaled register in the other operand of the icmp.
1115 if (AM.Scale != 0 && AM.Scale != -1)
1118 // If we have low-level target information, ask the target if it can fold an
1119 // integer immediate on an icmp.
1120 if (AM.BaseOffs != 0) {
1121 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1128 // Only handle single-register values.
1129 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1131 case LSRUse::Special:
1132 // Only handle -1 scales, or no scale.
1133 return AM.Scale == 0 || AM.Scale == -1;
1139 static bool isLegalUse(TargetLowering::AddrMode AM,
1140 int64_t MinOffset, int64_t MaxOffset,
1141 LSRUse::KindType Kind, const Type *AccessTy,
1142 const TargetLowering *TLI) {
1143 // Check for overflow.
1144 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1147 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1148 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1149 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1150 // Check for overflow.
1151 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1154 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1155 return isLegalUse(AM, Kind, AccessTy, TLI);
1160 static bool isAlwaysFoldable(int64_t BaseOffs,
1161 GlobalValue *BaseGV,
1163 LSRUse::KindType Kind, const Type *AccessTy,
1164 const TargetLowering *TLI) {
1165 // Fast-path: zero is always foldable.
1166 if (BaseOffs == 0 && !BaseGV) return true;
1168 // Conservatively, create an address with an immediate and a
1169 // base and a scale.
1170 TargetLowering::AddrMode AM;
1171 AM.BaseOffs = BaseOffs;
1173 AM.HasBaseReg = HasBaseReg;
1174 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1176 // Canonicalize a scale of 1 to a base register if the formula doesn't
1177 // already have a base register.
1178 if (!AM.HasBaseReg && AM.Scale == 1) {
1180 AM.HasBaseReg = true;
1183 return isLegalUse(AM, Kind, AccessTy, TLI);
1186 static bool isAlwaysFoldable(const SCEV *S,
1187 int64_t MinOffset, int64_t MaxOffset,
1189 LSRUse::KindType Kind, const Type *AccessTy,
1190 const TargetLowering *TLI,
1191 ScalarEvolution &SE) {
1192 // Fast-path: zero is always foldable.
1193 if (S->isZero()) return true;
1195 // Conservatively, create an address with an immediate and a
1196 // base and a scale.
1197 int64_t BaseOffs = ExtractImmediate(S, SE);
1198 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1200 // If there's anything else involved, it's not foldable.
1201 if (!S->isZero()) return false;
1203 // Fast-path: zero is always foldable.
1204 if (BaseOffs == 0 && !BaseGV) return true;
1206 // Conservatively, create an address with an immediate and a
1207 // base and a scale.
1208 TargetLowering::AddrMode AM;
1209 AM.BaseOffs = BaseOffs;
1211 AM.HasBaseReg = HasBaseReg;
1212 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1214 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1219 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1220 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1221 struct UseMapDenseMapInfo {
1222 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1223 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1226 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1227 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1231 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1232 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1233 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1237 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1238 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1243 /// FormulaSorter - This class implements an ordering for formulae which sorts
1244 /// the by their standalone cost.
1245 class FormulaSorter {
1246 /// These two sets are kept empty, so that we compute standalone costs.
1247 DenseSet<const SCEV *> VisitedRegs;
1248 SmallPtrSet<const SCEV *, 16> Regs;
1251 ScalarEvolution &SE;
1255 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1256 : L(l), LU(&lu), SE(se), DT(dt) {}
1258 bool operator()(const Formula &A, const Formula &B) {
1260 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1263 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1265 return CostA < CostB;
1269 /// LSRInstance - This class holds state for the main loop strength reduction
1273 ScalarEvolution &SE;
1276 const TargetLowering *const TLI;
1280 /// IVIncInsertPos - This is the insert position that the current loop's
1281 /// induction variable increment should be placed. In simple loops, this is
1282 /// the latch block's terminator. But in more complicated cases, this is a
1283 /// position which will dominate all the in-loop post-increment users.
1284 Instruction *IVIncInsertPos;
1286 /// Factors - Interesting factors between use strides.
1287 SmallSetVector<int64_t, 8> Factors;
1289 /// Types - Interesting use types, to facilitate truncation reuse.
1290 SmallSetVector<const Type *, 4> Types;
1292 /// Fixups - The list of operands which are to be replaced.
1293 SmallVector<LSRFixup, 16> Fixups;
1295 /// Uses - The list of interesting uses.
1296 SmallVector<LSRUse, 16> Uses;
1298 /// RegUses - Track which uses use which register candidates.
1299 RegUseTracker RegUses;
1301 void OptimizeShadowIV();
1302 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1303 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1304 void OptimizeLoopTermCond();
1306 void CollectInterestingTypesAndFactors();
1307 void CollectFixupsAndInitialFormulae();
1309 LSRFixup &getNewFixup() {
1310 Fixups.push_back(LSRFixup());
1311 return Fixups.back();
1314 // Support for sharing of LSRUses between LSRFixups.
1315 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1317 UseMapDenseMapInfo> UseMapTy;
1320 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1321 LSRUse::KindType Kind, const Type *AccessTy);
1323 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1324 LSRUse::KindType Kind,
1325 const Type *AccessTy);
1327 void DeleteUse(LSRUse &LU);
1329 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1332 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1333 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1334 void CountRegisters(const Formula &F, size_t LUIdx);
1335 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1337 void CollectLoopInvariantFixupsAndFormulae();
1339 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1340 unsigned Depth = 0);
1341 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1342 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1343 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1344 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1345 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1346 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1347 void GenerateCrossUseConstantOffsets();
1348 void GenerateAllReuseFormulae();
1350 void FilterOutUndesirableDedicatedRegisters();
1352 size_t EstimateSearchSpaceComplexity() const;
1353 void NarrowSearchSpaceUsingHeuristics();
1355 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1357 SmallVectorImpl<const Formula *> &Workspace,
1358 const Cost &CurCost,
1359 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1360 DenseSet<const SCEV *> &VisitedRegs) const;
1361 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1363 BasicBlock::iterator
1364 HoistInsertPosition(BasicBlock::iterator IP,
1365 const SmallVectorImpl<Instruction *> &Inputs) const;
1366 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1368 const LSRUse &LU) const;
1370 Value *Expand(const LSRFixup &LF,
1372 BasicBlock::iterator IP,
1373 SCEVExpander &Rewriter,
1374 SmallVectorImpl<WeakVH> &DeadInsts) const;
1375 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1377 SCEVExpander &Rewriter,
1378 SmallVectorImpl<WeakVH> &DeadInsts,
1380 void Rewrite(const LSRFixup &LF,
1382 SCEVExpander &Rewriter,
1383 SmallVectorImpl<WeakVH> &DeadInsts,
1385 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1388 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1390 bool getChanged() const { return Changed; }
1392 void print_factors_and_types(raw_ostream &OS) const;
1393 void print_fixups(raw_ostream &OS) const;
1394 void print_uses(raw_ostream &OS) const;
1395 void print(raw_ostream &OS) const;
1401 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1402 /// inside the loop then try to eliminate the cast operation.
1403 void LSRInstance::OptimizeShadowIV() {
1404 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1405 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1408 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1409 UI != E; /* empty */) {
1410 IVUsers::const_iterator CandidateUI = UI;
1412 Instruction *ShadowUse = CandidateUI->getUser();
1413 const Type *DestTy = NULL;
1415 /* If shadow use is a int->float cast then insert a second IV
1416 to eliminate this cast.
1418 for (unsigned i = 0; i < n; ++i)
1424 for (unsigned i = 0; i < n; ++i, ++d)
1427 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1428 DestTy = UCast->getDestTy();
1429 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1430 DestTy = SCast->getDestTy();
1431 if (!DestTy) continue;
1434 // If target does not support DestTy natively then do not apply
1435 // this transformation.
1436 EVT DVT = TLI->getValueType(DestTy);
1437 if (!TLI->isTypeLegal(DVT)) continue;
1440 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1442 if (PH->getNumIncomingValues() != 2) continue;
1444 const Type *SrcTy = PH->getType();
1445 int Mantissa = DestTy->getFPMantissaWidth();
1446 if (Mantissa == -1) continue;
1447 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1450 unsigned Entry, Latch;
1451 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1459 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1460 if (!Init) continue;
1461 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1463 BinaryOperator *Incr =
1464 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1465 if (!Incr) continue;
1466 if (Incr->getOpcode() != Instruction::Add
1467 && Incr->getOpcode() != Instruction::Sub)
1470 /* Initialize new IV, double d = 0.0 in above example. */
1471 ConstantInt *C = NULL;
1472 if (Incr->getOperand(0) == PH)
1473 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1474 else if (Incr->getOperand(1) == PH)
1475 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1481 // Ignore negative constants, as the code below doesn't handle them
1482 // correctly. TODO: Remove this restriction.
1483 if (!C->getValue().isStrictlyPositive()) continue;
1485 /* Add new PHINode. */
1486 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1488 /* create new increment. '++d' in above example. */
1489 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1490 BinaryOperator *NewIncr =
1491 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1492 Instruction::FAdd : Instruction::FSub,
1493 NewPH, CFP, "IV.S.next.", Incr);
1495 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1496 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1498 /* Remove cast operation */
1499 ShadowUse->replaceAllUsesWith(NewPH);
1500 ShadowUse->eraseFromParent();
1506 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1507 /// set the IV user and stride information and return true, otherwise return
1509 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1510 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1511 if (UI->getUser() == Cond) {
1512 // NOTE: we could handle setcc instructions with multiple uses here, but
1513 // InstCombine does it as well for simple uses, it's not clear that it
1514 // occurs enough in real life to handle.
1521 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1522 /// a max computation.
1524 /// This is a narrow solution to a specific, but acute, problem. For loops
1530 /// } while (++i < n);
1532 /// the trip count isn't just 'n', because 'n' might not be positive. And
1533 /// unfortunately this can come up even for loops where the user didn't use
1534 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1535 /// will commonly be lowered like this:
1541 /// } while (++i < n);
1544 /// and then it's possible for subsequent optimization to obscure the if
1545 /// test in such a way that indvars can't find it.
1547 /// When indvars can't find the if test in loops like this, it creates a
1548 /// max expression, which allows it to give the loop a canonical
1549 /// induction variable:
1552 /// max = n < 1 ? 1 : n;
1555 /// } while (++i != max);
1557 /// Canonical induction variables are necessary because the loop passes
1558 /// are designed around them. The most obvious example of this is the
1559 /// LoopInfo analysis, which doesn't remember trip count values. It
1560 /// expects to be able to rediscover the trip count each time it is
1561 /// needed, and it does this using a simple analysis that only succeeds if
1562 /// the loop has a canonical induction variable.
1564 /// However, when it comes time to generate code, the maximum operation
1565 /// can be quite costly, especially if it's inside of an outer loop.
1567 /// This function solves this problem by detecting this type of loop and
1568 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1569 /// the instructions for the maximum computation.
1571 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1572 // Check that the loop matches the pattern we're looking for.
1573 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1574 Cond->getPredicate() != CmpInst::ICMP_NE)
1577 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1578 if (!Sel || !Sel->hasOneUse()) return Cond;
1580 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1581 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1583 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1585 // Add one to the backedge-taken count to get the trip count.
1586 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1587 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1589 // Check for a max calculation that matches the pattern. There's no check
1590 // for ICMP_ULE here because the comparison would be with zero, which
1591 // isn't interesting.
1592 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1593 const SCEVNAryExpr *Max = 0;
1594 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1595 Pred = ICmpInst::ICMP_SLE;
1597 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1598 Pred = ICmpInst::ICMP_SLT;
1600 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1601 Pred = ICmpInst::ICMP_ULT;
1608 // To handle a max with more than two operands, this optimization would
1609 // require additional checking and setup.
1610 if (Max->getNumOperands() != 2)
1613 const SCEV *MaxLHS = Max->getOperand(0);
1614 const SCEV *MaxRHS = Max->getOperand(1);
1616 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1617 // for a comparison with 1. For <= and >=, a comparison with zero.
1619 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1622 // Check the relevant induction variable for conformance to
1624 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1625 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1626 if (!AR || !AR->isAffine() ||
1627 AR->getStart() != One ||
1628 AR->getStepRecurrence(SE) != One)
1631 assert(AR->getLoop() == L &&
1632 "Loop condition operand is an addrec in a different loop!");
1634 // Check the right operand of the select, and remember it, as it will
1635 // be used in the new comparison instruction.
1637 if (ICmpInst::isTrueWhenEqual(Pred)) {
1638 // Look for n+1, and grab n.
1639 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1640 if (isa<ConstantInt>(BO->getOperand(1)) &&
1641 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1642 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1643 NewRHS = BO->getOperand(0);
1644 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1645 if (isa<ConstantInt>(BO->getOperand(1)) &&
1646 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1647 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1648 NewRHS = BO->getOperand(0);
1651 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1652 NewRHS = Sel->getOperand(1);
1653 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1654 NewRHS = Sel->getOperand(2);
1655 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1656 NewRHS = SU->getValue();
1658 // Max doesn't match expected pattern.
1661 // Determine the new comparison opcode. It may be signed or unsigned,
1662 // and the original comparison may be either equality or inequality.
1663 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1664 Pred = CmpInst::getInversePredicate(Pred);
1666 // Ok, everything looks ok to change the condition into an SLT or SGE and
1667 // delete the max calculation.
1669 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1671 // Delete the max calculation instructions.
1672 Cond->replaceAllUsesWith(NewCond);
1673 CondUse->setUser(NewCond);
1674 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1675 Cond->eraseFromParent();
1676 Sel->eraseFromParent();
1677 if (Cmp->use_empty())
1678 Cmp->eraseFromParent();
1682 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1683 /// postinc iv when possible.
1685 LSRInstance::OptimizeLoopTermCond() {
1686 SmallPtrSet<Instruction *, 4> PostIncs;
1688 BasicBlock *LatchBlock = L->getLoopLatch();
1689 SmallVector<BasicBlock*, 8> ExitingBlocks;
1690 L->getExitingBlocks(ExitingBlocks);
1692 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1693 BasicBlock *ExitingBlock = ExitingBlocks[i];
1695 // Get the terminating condition for the loop if possible. If we
1696 // can, we want to change it to use a post-incremented version of its
1697 // induction variable, to allow coalescing the live ranges for the IV into
1698 // one register value.
1700 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1703 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1704 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1707 // Search IVUsesByStride to find Cond's IVUse if there is one.
1708 IVStrideUse *CondUse = 0;
1709 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1710 if (!FindIVUserForCond(Cond, CondUse))
1713 // If the trip count is computed in terms of a max (due to ScalarEvolution
1714 // being unable to find a sufficient guard, for example), change the loop
1715 // comparison to use SLT or ULT instead of NE.
1716 // One consequence of doing this now is that it disrupts the count-down
1717 // optimization. That's not always a bad thing though, because in such
1718 // cases it may still be worthwhile to avoid a max.
1719 Cond = OptimizeMax(Cond, CondUse);
1721 // If this exiting block dominates the latch block, it may also use
1722 // the post-inc value if it won't be shared with other uses.
1723 // Check for dominance.
1724 if (!DT.dominates(ExitingBlock, LatchBlock))
1727 // Conservatively avoid trying to use the post-inc value in non-latch
1728 // exits if there may be pre-inc users in intervening blocks.
1729 if (LatchBlock != ExitingBlock)
1730 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1731 // Test if the use is reachable from the exiting block. This dominator
1732 // query is a conservative approximation of reachability.
1733 if (&*UI != CondUse &&
1734 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1735 // Conservatively assume there may be reuse if the quotient of their
1736 // strides could be a legal scale.
1737 const SCEV *A = IU.getStride(*CondUse, L);
1738 const SCEV *B = IU.getStride(*UI, L);
1739 if (!A || !B) continue;
1740 if (SE.getTypeSizeInBits(A->getType()) !=
1741 SE.getTypeSizeInBits(B->getType())) {
1742 if (SE.getTypeSizeInBits(A->getType()) >
1743 SE.getTypeSizeInBits(B->getType()))
1744 B = SE.getSignExtendExpr(B, A->getType());
1746 A = SE.getSignExtendExpr(A, B->getType());
1748 if (const SCEVConstant *D =
1749 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1750 const ConstantInt *C = D->getValue();
1751 // Stride of one or negative one can have reuse with non-addresses.
1752 if (C->isOne() || C->isAllOnesValue())
1753 goto decline_post_inc;
1754 // Avoid weird situations.
1755 if (C->getValue().getMinSignedBits() >= 64 ||
1756 C->getValue().isMinSignedValue())
1757 goto decline_post_inc;
1758 // Without TLI, assume that any stride might be valid, and so any
1759 // use might be shared.
1761 goto decline_post_inc;
1762 // Check for possible scaled-address reuse.
1763 const Type *AccessTy = getAccessType(UI->getUser());
1764 TargetLowering::AddrMode AM;
1765 AM.Scale = C->getSExtValue();
1766 if (TLI->isLegalAddressingMode(AM, AccessTy))
1767 goto decline_post_inc;
1768 AM.Scale = -AM.Scale;
1769 if (TLI->isLegalAddressingMode(AM, AccessTy))
1770 goto decline_post_inc;
1774 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1777 // It's possible for the setcc instruction to be anywhere in the loop, and
1778 // possible for it to have multiple users. If it is not immediately before
1779 // the exiting block branch, move it.
1780 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1781 if (Cond->hasOneUse()) {
1782 Cond->moveBefore(TermBr);
1784 // Clone the terminating condition and insert into the loopend.
1785 ICmpInst *OldCond = Cond;
1786 Cond = cast<ICmpInst>(Cond->clone());
1787 Cond->setName(L->getHeader()->getName() + ".termcond");
1788 ExitingBlock->getInstList().insert(TermBr, Cond);
1790 // Clone the IVUse, as the old use still exists!
1791 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1792 TermBr->replaceUsesOfWith(OldCond, Cond);
1796 // If we get to here, we know that we can transform the setcc instruction to
1797 // use the post-incremented version of the IV, allowing us to coalesce the
1798 // live ranges for the IV correctly.
1799 CondUse->transformToPostInc(L);
1802 PostIncs.insert(Cond);
1806 // Determine an insertion point for the loop induction variable increment. It
1807 // must dominate all the post-inc comparisons we just set up, and it must
1808 // dominate the loop latch edge.
1809 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1810 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1811 E = PostIncs.end(); I != E; ++I) {
1813 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1815 if (BB == (*I)->getParent())
1816 IVIncInsertPos = *I;
1817 else if (BB != IVIncInsertPos->getParent())
1818 IVIncInsertPos = BB->getTerminator();
1822 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1823 /// at the given offset and other details. If so, update the use and
1826 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1827 LSRUse::KindType Kind, const Type *AccessTy) {
1828 int64_t NewMinOffset = LU.MinOffset;
1829 int64_t NewMaxOffset = LU.MaxOffset;
1830 const Type *NewAccessTy = AccessTy;
1832 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1833 // something conservative, however this can pessimize in the case that one of
1834 // the uses will have all its uses outside the loop, for example.
1835 if (LU.Kind != Kind)
1837 // Conservatively assume HasBaseReg is true for now.
1838 if (NewOffset < LU.MinOffset) {
1839 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1840 Kind, AccessTy, TLI))
1842 NewMinOffset = NewOffset;
1843 } else if (NewOffset > LU.MaxOffset) {
1844 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1845 Kind, AccessTy, TLI))
1847 NewMaxOffset = NewOffset;
1849 // Check for a mismatched access type, and fall back conservatively as needed.
1850 // TODO: Be less conservative when the type is similar and can use the same
1851 // addressing modes.
1852 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1853 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1856 LU.MinOffset = NewMinOffset;
1857 LU.MaxOffset = NewMaxOffset;
1858 LU.AccessTy = NewAccessTy;
1859 if (NewOffset != LU.Offsets.back())
1860 LU.Offsets.push_back(NewOffset);
1864 /// getUse - Return an LSRUse index and an offset value for a fixup which
1865 /// needs the given expression, with the given kind and optional access type.
1866 /// Either reuse an existing use or create a new one, as needed.
1867 std::pair<size_t, int64_t>
1868 LSRInstance::getUse(const SCEV *&Expr,
1869 LSRUse::KindType Kind, const Type *AccessTy) {
1870 const SCEV *Copy = Expr;
1871 int64_t Offset = ExtractImmediate(Expr, SE);
1873 // Basic uses can't accept any offset, for example.
1874 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1879 std::pair<UseMapTy::iterator, bool> P =
1880 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1882 // A use already existed with this base.
1883 size_t LUIdx = P.first->second;
1884 LSRUse &LU = Uses[LUIdx];
1885 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1887 return std::make_pair(LUIdx, Offset);
1890 // Create a new use.
1891 size_t LUIdx = Uses.size();
1892 P.first->second = LUIdx;
1893 Uses.push_back(LSRUse(Kind, AccessTy));
1894 LSRUse &LU = Uses[LUIdx];
1896 // We don't need to track redundant offsets, but we don't need to go out
1897 // of our way here to avoid them.
1898 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1899 LU.Offsets.push_back(Offset);
1901 LU.MinOffset = Offset;
1902 LU.MaxOffset = Offset;
1903 return std::make_pair(LUIdx, Offset);
1906 /// DeleteUse - Delete the given use from the Uses list.
1907 void LSRInstance::DeleteUse(LSRUse &LU) {
1908 if (&LU != &Uses.back())
1909 std::swap(LU, Uses.back());
1913 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1914 /// a formula that has the same registers as the given formula.
1916 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1917 const LSRUse &OrigLU) {
1918 // Search all uses for the formula. This could be more clever. Ignore
1919 // ICmpZero uses because they may contain formulae generated by
1920 // GenerateICmpZeroScales, in which case adding fixup offsets may
1922 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1923 LSRUse &LU = Uses[LUIdx];
1924 if (&LU != &OrigLU &&
1925 LU.Kind != LSRUse::ICmpZero &&
1926 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1927 LU.HasFormulaWithSameRegs(OrigF)) {
1928 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1929 E = LU.Formulae.end(); I != E; ++I) {
1930 const Formula &F = *I;
1931 if (F.BaseRegs == OrigF.BaseRegs &&
1932 F.ScaledReg == OrigF.ScaledReg &&
1933 F.AM.BaseGV == OrigF.AM.BaseGV &&
1934 F.AM.Scale == OrigF.AM.Scale &&
1936 if (F.AM.BaseOffs == 0)
1947 void LSRInstance::CollectInterestingTypesAndFactors() {
1948 SmallSetVector<const SCEV *, 4> Strides;
1950 // Collect interesting types and strides.
1951 SmallVector<const SCEV *, 4> Worklist;
1952 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1953 const SCEV *Expr = IU.getExpr(*UI);
1955 // Collect interesting types.
1956 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1958 // Add strides for mentioned loops.
1959 Worklist.push_back(Expr);
1961 const SCEV *S = Worklist.pop_back_val();
1962 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1963 Strides.insert(AR->getStepRecurrence(SE));
1964 Worklist.push_back(AR->getStart());
1965 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1966 Worklist.append(Add->op_begin(), Add->op_end());
1968 } while (!Worklist.empty());
1971 // Compute interesting factors from the set of interesting strides.
1972 for (SmallSetVector<const SCEV *, 4>::const_iterator
1973 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1974 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1975 next(I); NewStrideIter != E; ++NewStrideIter) {
1976 const SCEV *OldStride = *I;
1977 const SCEV *NewStride = *NewStrideIter;
1979 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1980 SE.getTypeSizeInBits(NewStride->getType())) {
1981 if (SE.getTypeSizeInBits(OldStride->getType()) >
1982 SE.getTypeSizeInBits(NewStride->getType()))
1983 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1985 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1987 if (const SCEVConstant *Factor =
1988 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1990 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1991 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1992 } else if (const SCEVConstant *Factor =
1993 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1996 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1997 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2001 // If all uses use the same type, don't bother looking for truncation-based
2003 if (Types.size() == 1)
2006 DEBUG(print_factors_and_types(dbgs()));
2009 void LSRInstance::CollectFixupsAndInitialFormulae() {
2010 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2012 LSRFixup &LF = getNewFixup();
2013 LF.UserInst = UI->getUser();
2014 LF.OperandValToReplace = UI->getOperandValToReplace();
2015 LF.PostIncLoops = UI->getPostIncLoops();
2017 LSRUse::KindType Kind = LSRUse::Basic;
2018 const Type *AccessTy = 0;
2019 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2020 Kind = LSRUse::Address;
2021 AccessTy = getAccessType(LF.UserInst);
2024 const SCEV *S = IU.getExpr(*UI);
2026 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2027 // (N - i == 0), and this allows (N - i) to be the expression that we work
2028 // with rather than just N or i, so we can consider the register
2029 // requirements for both N and i at the same time. Limiting this code to
2030 // equality icmps is not a problem because all interesting loops use
2031 // equality icmps, thanks to IndVarSimplify.
2032 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2033 if (CI->isEquality()) {
2034 // Swap the operands if needed to put the OperandValToReplace on the
2035 // left, for consistency.
2036 Value *NV = CI->getOperand(1);
2037 if (NV == LF.OperandValToReplace) {
2038 CI->setOperand(1, CI->getOperand(0));
2039 CI->setOperand(0, NV);
2040 NV = CI->getOperand(1);
2044 // x == y --> x - y == 0
2045 const SCEV *N = SE.getSCEV(NV);
2046 if (N->isLoopInvariant(L)) {
2047 Kind = LSRUse::ICmpZero;
2048 S = SE.getMinusSCEV(N, S);
2051 // -1 and the negations of all interesting strides (except the negation
2052 // of -1) are now also interesting.
2053 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2054 if (Factors[i] != -1)
2055 Factors.insert(-(uint64_t)Factors[i]);
2059 // Set up the initial formula for this use.
2060 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2062 LF.Offset = P.second;
2063 LSRUse &LU = Uses[LF.LUIdx];
2064 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2066 // If this is the first use of this LSRUse, give it a formula.
2067 if (LU.Formulae.empty()) {
2068 InsertInitialFormula(S, LU, LF.LUIdx);
2069 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2073 DEBUG(print_fixups(dbgs()));
2076 /// InsertInitialFormula - Insert a formula for the given expression into
2077 /// the given use, separating out loop-variant portions from loop-invariant
2078 /// and loop-computable portions.
2080 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2082 F.InitialMatch(S, L, SE, DT);
2083 bool Inserted = InsertFormula(LU, LUIdx, F);
2084 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2087 /// InsertSupplementalFormula - Insert a simple single-register formula for
2088 /// the given expression into the given use.
2090 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2091 LSRUse &LU, size_t LUIdx) {
2093 F.BaseRegs.push_back(S);
2094 F.AM.HasBaseReg = true;
2095 bool Inserted = InsertFormula(LU, LUIdx, F);
2096 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2099 /// CountRegisters - Note which registers are used by the given formula,
2100 /// updating RegUses.
2101 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2103 RegUses.CountRegister(F.ScaledReg, LUIdx);
2104 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2105 E = F.BaseRegs.end(); I != E; ++I)
2106 RegUses.CountRegister(*I, LUIdx);
2109 /// InsertFormula - If the given formula has not yet been inserted, add it to
2110 /// the list, and return true. Return false otherwise.
2111 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2112 if (!LU.InsertFormula(F))
2115 CountRegisters(F, LUIdx);
2119 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2120 /// loop-invariant values which we're tracking. These other uses will pin these
2121 /// values in registers, making them less profitable for elimination.
2122 /// TODO: This currently misses non-constant addrec step registers.
2123 /// TODO: Should this give more weight to users inside the loop?
2125 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2126 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2127 SmallPtrSet<const SCEV *, 8> Inserted;
2129 while (!Worklist.empty()) {
2130 const SCEV *S = Worklist.pop_back_val();
2132 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2133 Worklist.append(N->op_begin(), N->op_end());
2134 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2135 Worklist.push_back(C->getOperand());
2136 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2137 Worklist.push_back(D->getLHS());
2138 Worklist.push_back(D->getRHS());
2139 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2140 if (!Inserted.insert(U)) continue;
2141 const Value *V = U->getValue();
2142 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2143 // Look for instructions defined outside the loop.
2144 if (L->contains(Inst)) continue;
2145 } else if (isa<UndefValue>(V))
2146 // Undef doesn't have a live range, so it doesn't matter.
2148 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2150 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2151 // Ignore non-instructions.
2154 // Ignore instructions in other functions (as can happen with
2156 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2158 // Ignore instructions not dominated by the loop.
2159 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2160 UserInst->getParent() :
2161 cast<PHINode>(UserInst)->getIncomingBlock(
2162 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2163 if (!DT.dominates(L->getHeader(), UseBB))
2165 // Ignore uses which are part of other SCEV expressions, to avoid
2166 // analyzing them multiple times.
2167 if (SE.isSCEVable(UserInst->getType())) {
2168 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2169 // If the user is a no-op, look through to its uses.
2170 if (!isa<SCEVUnknown>(UserS))
2174 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2178 // Ignore icmp instructions which are already being analyzed.
2179 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2180 unsigned OtherIdx = !UI.getOperandNo();
2181 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2182 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2186 LSRFixup &LF = getNewFixup();
2187 LF.UserInst = const_cast<Instruction *>(UserInst);
2188 LF.OperandValToReplace = UI.getUse();
2189 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2191 LF.Offset = P.second;
2192 LSRUse &LU = Uses[LF.LUIdx];
2193 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2194 InsertSupplementalFormula(U, LU, LF.LUIdx);
2195 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2202 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2203 /// separate registers. If C is non-null, multiply each subexpression by C.
2204 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2205 SmallVectorImpl<const SCEV *> &Ops,
2206 ScalarEvolution &SE) {
2207 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2208 // Break out add operands.
2209 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2211 CollectSubexprs(*I, C, Ops, SE);
2213 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2214 // Split a non-zero base out of an addrec.
2215 if (!AR->getStart()->isZero()) {
2216 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2217 AR->getStepRecurrence(SE),
2218 AR->getLoop()), C, Ops, SE);
2219 CollectSubexprs(AR->getStart(), C, Ops, SE);
2222 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2223 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2224 if (Mul->getNumOperands() == 2)
2225 if (const SCEVConstant *Op0 =
2226 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2227 CollectSubexprs(Mul->getOperand(1),
2228 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2234 // Otherwise use the value itself.
2235 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2238 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2240 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2243 // Arbitrarily cap recursion to protect compile time.
2244 if (Depth >= 3) return;
2246 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2247 const SCEV *BaseReg = Base.BaseRegs[i];
2249 SmallVector<const SCEV *, 8> AddOps;
2250 CollectSubexprs(BaseReg, 0, AddOps, SE);
2251 if (AddOps.size() == 1) continue;
2253 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2254 JE = AddOps.end(); J != JE; ++J) {
2255 // Don't pull a constant into a register if the constant could be folded
2256 // into an immediate field.
2257 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2258 Base.getNumRegs() > 1,
2259 LU.Kind, LU.AccessTy, TLI, SE))
2262 // Collect all operands except *J.
2263 SmallVector<const SCEV *, 8> InnerAddOps
2264 ( ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2266 (next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2268 // Don't leave just a constant behind in a register if the constant could
2269 // be folded into an immediate field.
2270 if (InnerAddOps.size() == 1 &&
2271 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2272 Base.getNumRegs() > 1,
2273 LU.Kind, LU.AccessTy, TLI, SE))
2276 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2277 if (InnerSum->isZero())
2280 F.BaseRegs[i] = InnerSum;
2281 F.BaseRegs.push_back(*J);
2282 if (InsertFormula(LU, LUIdx, F))
2283 // If that formula hadn't been seen before, recurse to find more like
2285 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2290 /// GenerateCombinations - Generate a formula consisting of all of the
2291 /// loop-dominating registers added into a single register.
2292 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2294 // This method is only interesting on a plurality of registers.
2295 if (Base.BaseRegs.size() <= 1) return;
2299 SmallVector<const SCEV *, 4> Ops;
2300 for (SmallVectorImpl<const SCEV *>::const_iterator
2301 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2302 const SCEV *BaseReg = *I;
2303 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2304 !BaseReg->hasComputableLoopEvolution(L))
2305 Ops.push_back(BaseReg);
2307 F.BaseRegs.push_back(BaseReg);
2309 if (Ops.size() > 1) {
2310 const SCEV *Sum = SE.getAddExpr(Ops);
2311 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2312 // opportunity to fold something. For now, just ignore such cases
2313 // rather than proceed with zero in a register.
2314 if (!Sum->isZero()) {
2315 F.BaseRegs.push_back(Sum);
2316 (void)InsertFormula(LU, LUIdx, F);
2321 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2322 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2324 // We can't add a symbolic offset if the address already contains one.
2325 if (Base.AM.BaseGV) return;
2327 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2328 const SCEV *G = Base.BaseRegs[i];
2329 GlobalValue *GV = ExtractSymbol(G, SE);
2330 if (G->isZero() || !GV)
2334 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2335 LU.Kind, LU.AccessTy, TLI))
2338 (void)InsertFormula(LU, LUIdx, F);
2342 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2343 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2345 // TODO: For now, just add the min and max offset, because it usually isn't
2346 // worthwhile looking at everything inbetween.
2347 SmallVector<int64_t, 4> Worklist;
2348 Worklist.push_back(LU.MinOffset);
2349 if (LU.MaxOffset != LU.MinOffset)
2350 Worklist.push_back(LU.MaxOffset);
2352 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2353 const SCEV *G = Base.BaseRegs[i];
2355 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2356 E = Worklist.end(); I != E; ++I) {
2358 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2359 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2360 LU.Kind, LU.AccessTy, TLI)) {
2361 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2363 (void)InsertFormula(LU, LUIdx, F);
2367 int64_t Imm = ExtractImmediate(G, SE);
2368 if (G->isZero() || Imm == 0)
2371 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2372 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2373 LU.Kind, LU.AccessTy, TLI))
2376 (void)InsertFormula(LU, LUIdx, F);
2380 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2381 /// the comparison. For example, x == y -> x*c == y*c.
2382 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2384 if (LU.Kind != LSRUse::ICmpZero) return;
2386 // Determine the integer type for the base formula.
2387 const Type *IntTy = Base.getType();
2389 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2391 // Don't do this if there is more than one offset.
2392 if (LU.MinOffset != LU.MaxOffset) return;
2394 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2396 // Check each interesting stride.
2397 for (SmallSetVector<int64_t, 8>::const_iterator
2398 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2399 int64_t Factor = *I;
2402 // Check that the multiplication doesn't overflow.
2403 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2405 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2406 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2409 // Check that multiplying with the use offset doesn't overflow.
2410 int64_t Offset = LU.MinOffset;
2411 if (Offset == INT64_MIN && Factor == -1)
2413 Offset = (uint64_t)Offset * Factor;
2414 if (Offset / Factor != LU.MinOffset)
2417 // Check that this scale is legal.
2418 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2421 // Compensate for the use having MinOffset built into it.
2422 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2424 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2426 // Check that multiplying with each base register doesn't overflow.
2427 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2428 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2429 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2433 // Check that multiplying with the scaled register doesn't overflow.
2435 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2436 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2440 // If we make it here and it's legal, add it.
2441 (void)InsertFormula(LU, LUIdx, F);
2446 /// GenerateScales - Generate stride factor reuse formulae by making use of
2447 /// scaled-offset address modes, for example.
2448 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2449 // Determine the integer type for the base formula.
2450 const Type *IntTy = Base.getType();
2453 // If this Formula already has a scaled register, we can't add another one.
2454 if (Base.AM.Scale != 0) return;
2456 // Check each interesting stride.
2457 for (SmallSetVector<int64_t, 8>::const_iterator
2458 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2459 int64_t Factor = *I;
2461 Base.AM.Scale = Factor;
2462 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2463 // Check whether this scale is going to be legal.
2464 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2465 LU.Kind, LU.AccessTy, TLI)) {
2466 // As a special-case, handle special out-of-loop Basic users specially.
2467 // TODO: Reconsider this special case.
2468 if (LU.Kind == LSRUse::Basic &&
2469 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2470 LSRUse::Special, LU.AccessTy, TLI) &&
2471 LU.AllFixupsOutsideLoop)
2472 LU.Kind = LSRUse::Special;
2476 // For an ICmpZero, negating a solitary base register won't lead to
2478 if (LU.Kind == LSRUse::ICmpZero &&
2479 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2481 // For each addrec base reg, apply the scale, if possible.
2482 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2483 if (const SCEVAddRecExpr *AR =
2484 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2485 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2486 if (FactorS->isZero())
2488 // Divide out the factor, ignoring high bits, since we'll be
2489 // scaling the value back up in the end.
2490 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2491 // TODO: This could be optimized to avoid all the copying.
2493 F.ScaledReg = Quotient;
2494 F.DeleteBaseReg(F.BaseRegs[i]);
2495 (void)InsertFormula(LU, LUIdx, F);
2501 /// GenerateTruncates - Generate reuse formulae from different IV types.
2502 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2503 // This requires TargetLowering to tell us which truncates are free.
2506 // Don't bother truncating symbolic values.
2507 if (Base.AM.BaseGV) return;
2509 // Determine the integer type for the base formula.
2510 const Type *DstTy = Base.getType();
2512 DstTy = SE.getEffectiveSCEVType(DstTy);
2514 for (SmallSetVector<const Type *, 4>::const_iterator
2515 I = Types.begin(), E = Types.end(); I != E; ++I) {
2516 const Type *SrcTy = *I;
2517 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2520 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2521 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2522 JE = F.BaseRegs.end(); J != JE; ++J)
2523 *J = SE.getAnyExtendExpr(*J, SrcTy);
2525 // TODO: This assumes we've done basic processing on all uses and
2526 // have an idea what the register usage is.
2527 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2530 (void)InsertFormula(LU, LUIdx, F);
2537 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2538 /// defer modifications so that the search phase doesn't have to worry about
2539 /// the data structures moving underneath it.
2543 const SCEV *OrigReg;
2545 WorkItem(size_t LI, int64_t I, const SCEV *R)
2546 : LUIdx(LI), Imm(I), OrigReg(R) {}
2548 void print(raw_ostream &OS) const;
2554 void WorkItem::print(raw_ostream &OS) const {
2555 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2556 << " , add offset " << Imm;
2559 void WorkItem::dump() const {
2560 print(errs()); errs() << '\n';
2563 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2564 /// distance apart and try to form reuse opportunities between them.
2565 void LSRInstance::GenerateCrossUseConstantOffsets() {
2566 // Group the registers by their value without any added constant offset.
2567 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2568 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2570 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2571 SmallVector<const SCEV *, 8> Sequence;
2572 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2574 const SCEV *Reg = *I;
2575 int64_t Imm = ExtractImmediate(Reg, SE);
2576 std::pair<RegMapTy::iterator, bool> Pair =
2577 Map.insert(std::make_pair(Reg, ImmMapTy()));
2579 Sequence.push_back(Reg);
2580 Pair.first->second.insert(std::make_pair(Imm, *I));
2581 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2584 // Now examine each set of registers with the same base value. Build up
2585 // a list of work to do and do the work in a separate step so that we're
2586 // not adding formulae and register counts while we're searching.
2587 SmallVector<WorkItem, 32> WorkItems;
2588 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2589 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2590 E = Sequence.end(); I != E; ++I) {
2591 const SCEV *Reg = *I;
2592 const ImmMapTy &Imms = Map.find(Reg)->second;
2594 // It's not worthwhile looking for reuse if there's only one offset.
2595 if (Imms.size() == 1)
2598 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2599 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2601 dbgs() << ' ' << J->first;
2604 // Examine each offset.
2605 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2607 const SCEV *OrigReg = J->second;
2609 int64_t JImm = J->first;
2610 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2612 if (!isa<SCEVConstant>(OrigReg) &&
2613 UsedByIndicesMap[Reg].count() == 1) {
2614 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2618 // Conservatively examine offsets between this orig reg a few selected
2620 ImmMapTy::const_iterator OtherImms[] = {
2621 Imms.begin(), prior(Imms.end()),
2622 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2624 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2625 ImmMapTy::const_iterator M = OtherImms[i];
2626 if (M == J || M == JE) continue;
2628 // Compute the difference between the two.
2629 int64_t Imm = (uint64_t)JImm - M->first;
2630 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2631 LUIdx = UsedByIndices.find_next(LUIdx))
2632 // Make a memo of this use, offset, and register tuple.
2633 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2634 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2641 UsedByIndicesMap.clear();
2642 UniqueItems.clear();
2644 // Now iterate through the worklist and add new formulae.
2645 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2646 E = WorkItems.end(); I != E; ++I) {
2647 const WorkItem &WI = *I;
2648 size_t LUIdx = WI.LUIdx;
2649 LSRUse &LU = Uses[LUIdx];
2650 int64_t Imm = WI.Imm;
2651 const SCEV *OrigReg = WI.OrigReg;
2653 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2654 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2655 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2657 // TODO: Use a more targeted data structure.
2658 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2659 const Formula &F = LU.Formulae[L];
2660 // Use the immediate in the scaled register.
2661 if (F.ScaledReg == OrigReg) {
2662 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2663 Imm * (uint64_t)F.AM.Scale;
2664 // Don't create 50 + reg(-50).
2665 if (F.referencesReg(SE.getSCEV(
2666 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2669 NewF.AM.BaseOffs = Offs;
2670 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2671 LU.Kind, LU.AccessTy, TLI))
2673 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2675 // If the new scale is a constant in a register, and adding the constant
2676 // value to the immediate would produce a value closer to zero than the
2677 // immediate itself, then the formula isn't worthwhile.
2678 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2679 if (C->getValue()->getValue().isNegative() !=
2680 (NewF.AM.BaseOffs < 0) &&
2681 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2682 .ule(abs64(NewF.AM.BaseOffs)))
2686 (void)InsertFormula(LU, LUIdx, NewF);
2688 // Use the immediate in a base register.
2689 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2690 const SCEV *BaseReg = F.BaseRegs[N];
2691 if (BaseReg != OrigReg)
2694 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2695 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2696 LU.Kind, LU.AccessTy, TLI))
2698 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2700 // If the new formula has a constant in a register, and adding the
2701 // constant value to the immediate would produce a value closer to
2702 // zero than the immediate itself, then the formula isn't worthwhile.
2703 for (SmallVectorImpl<const SCEV *>::const_iterator
2704 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2706 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2707 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2708 abs64(NewF.AM.BaseOffs)) &&
2709 (C->getValue()->getValue() +
2710 NewF.AM.BaseOffs).countTrailingZeros() >=
2711 CountTrailingZeros_64(NewF.AM.BaseOffs))
2715 (void)InsertFormula(LU, LUIdx, NewF);
2724 /// GenerateAllReuseFormulae - Generate formulae for each use.
2726 LSRInstance::GenerateAllReuseFormulae() {
2727 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2728 // queries are more precise.
2729 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2730 LSRUse &LU = Uses[LUIdx];
2731 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2732 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2733 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2734 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2736 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2737 LSRUse &LU = Uses[LUIdx];
2738 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2739 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2740 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2741 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2742 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2743 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2744 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2745 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2747 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2748 LSRUse &LU = Uses[LUIdx];
2749 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2750 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2753 GenerateCrossUseConstantOffsets();
2756 /// If their are multiple formulae with the same set of registers used
2757 /// by other uses, pick the best one and delete the others.
2758 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2760 bool ChangedFormulae = false;
2763 // Collect the best formula for each unique set of shared registers. This
2764 // is reset for each use.
2765 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2767 BestFormulaeTy BestFormulae;
2769 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2770 LSRUse &LU = Uses[LUIdx];
2771 FormulaSorter Sorter(L, LU, SE, DT);
2772 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2775 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2776 FIdx != NumForms; ++FIdx) {
2777 Formula &F = LU.Formulae[FIdx];
2779 SmallVector<const SCEV *, 2> Key;
2780 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2781 JE = F.BaseRegs.end(); J != JE; ++J) {
2782 const SCEV *Reg = *J;
2783 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2787 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2788 Key.push_back(F.ScaledReg);
2789 // Unstable sort by host order ok, because this is only used for
2791 std::sort(Key.begin(), Key.end());
2793 std::pair<BestFormulaeTy::const_iterator, bool> P =
2794 BestFormulae.insert(std::make_pair(Key, FIdx));
2796 Formula &Best = LU.Formulae[P.first->second];
2797 if (Sorter.operator()(F, Best))
2799 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2801 " in favor of formula "; Best.print(dbgs());
2804 ChangedFormulae = true;
2806 LU.DeleteFormula(F);
2814 // Now that we've filtered out some formulae, recompute the Regs set.
2816 LU.RecomputeRegs(LUIdx, RegUses);
2818 // Reset this to prepare for the next use.
2819 BestFormulae.clear();
2822 DEBUG(if (ChangedFormulae) {
2824 "After filtering out undesirable candidates:\n";
2829 // This is a rough guess that seems to work fairly well.
2830 static const size_t ComplexityLimit = UINT16_MAX;
2832 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2833 /// solutions the solver might have to consider. It almost never considers
2834 /// this many solutions because it prune the search space, but the pruning
2835 /// isn't always sufficient.
2836 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2838 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2839 E = Uses.end(); I != E; ++I) {
2840 size_t FSize = I->Formulae.size();
2841 if (FSize >= ComplexityLimit) {
2842 Power = ComplexityLimit;
2846 if (Power >= ComplexityLimit)
2852 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2853 /// formulae to choose from, use some rough heuristics to prune down the number
2854 /// of formulae. This keeps the main solver from taking an extraordinary amount
2855 /// of time in some worst-case scenarios.
2856 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2857 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2858 DEBUG(dbgs() << "The search space is too complex.\n");
2860 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2861 "which use a superset of registers used by other "
2864 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2865 LSRUse &LU = Uses[LUIdx];
2867 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2868 Formula &F = LU.Formulae[i];
2869 // Look for a formula with a constant or GV in a register. If the use
2870 // also has a formula with that same value in an immediate field,
2871 // delete the one that uses a register.
2872 for (SmallVectorImpl<const SCEV *>::const_iterator
2873 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2874 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2876 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2877 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2878 (I - F.BaseRegs.begin()));
2879 if (LU.HasFormulaWithSameRegs(NewF)) {
2880 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2881 LU.DeleteFormula(F);
2887 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2888 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2891 NewF.AM.BaseGV = GV;
2892 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2893 (I - F.BaseRegs.begin()));
2894 if (LU.HasFormulaWithSameRegs(NewF)) {
2895 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2897 LU.DeleteFormula(F);
2908 LU.RecomputeRegs(LUIdx, RegUses);
2911 DEBUG(dbgs() << "After pre-selection:\n";
2912 print_uses(dbgs()));
2915 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2916 DEBUG(dbgs() << "The search space is too complex.\n");
2918 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2919 "separated by a constant offset will use the same "
2922 // This is especially useful for unrolled loops.
2924 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2925 LSRUse &LU = Uses[LUIdx];
2926 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2927 E = LU.Formulae.end(); I != E; ++I) {
2928 const Formula &F = *I;
2929 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2930 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2931 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2932 /*HasBaseReg=*/false,
2933 LU.Kind, LU.AccessTy)) {
2934 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2937 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2939 // Delete formulae from the new use which are no longer legal.
2941 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2942 Formula &F = LUThatHas->Formulae[i];
2943 if (!isLegalUse(F.AM,
2944 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2945 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2946 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2948 LUThatHas->DeleteFormula(F);
2955 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
2957 // Update the relocs to reference the new use.
2958 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
2959 E = Fixups.end(); I != E; ++I) {
2960 LSRFixup &Fixup = *I;
2961 if (Fixup.LUIdx == LUIdx) {
2962 Fixup.LUIdx = LUThatHas - &Uses.front();
2963 Fixup.Offset += F.AM.BaseOffs;
2964 DEBUG(errs() << "New fixup has offset "
2965 << Fixup.Offset << '\n');
2967 if (Fixup.LUIdx == NumUses-1)
2968 Fixup.LUIdx = LUIdx;
2971 // Delete the old use.
2982 DEBUG(dbgs() << "After pre-selection:\n";
2983 print_uses(dbgs()));
2986 // With all other options exhausted, loop until the system is simple
2987 // enough to handle.
2988 SmallPtrSet<const SCEV *, 4> Taken;
2989 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2990 // Ok, we have too many of formulae on our hands to conveniently handle.
2991 // Use a rough heuristic to thin out the list.
2992 DEBUG(dbgs() << "The search space is too complex.\n");
2994 // Pick the register which is used by the most LSRUses, which is likely
2995 // to be a good reuse register candidate.
2996 const SCEV *Best = 0;
2997 unsigned BestNum = 0;
2998 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3000 const SCEV *Reg = *I;
3001 if (Taken.count(Reg))
3006 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3007 if (Count > BestNum) {
3014 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3015 << " will yield profitable reuse.\n");
3018 // In any use with formulae which references this register, delete formulae
3019 // which don't reference it.
3020 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3021 LSRUse &LU = Uses[LUIdx];
3022 if (!LU.Regs.count(Best)) continue;
3025 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3026 Formula &F = LU.Formulae[i];
3027 if (!F.referencesReg(Best)) {
3028 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3029 LU.DeleteFormula(F);
3033 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3039 LU.RecomputeRegs(LUIdx, RegUses);
3042 DEBUG(dbgs() << "After pre-selection:\n";
3043 print_uses(dbgs()));
3047 /// SolveRecurse - This is the recursive solver.
3048 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3050 SmallVectorImpl<const Formula *> &Workspace,
3051 const Cost &CurCost,
3052 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3053 DenseSet<const SCEV *> &VisitedRegs) const {
3056 // - use more aggressive filtering
3057 // - sort the formula so that the most profitable solutions are found first
3058 // - sort the uses too
3060 // - don't compute a cost, and then compare. compare while computing a cost
3062 // - track register sets with SmallBitVector
3064 const LSRUse &LU = Uses[Workspace.size()];
3066 // If this use references any register that's already a part of the
3067 // in-progress solution, consider it a requirement that a formula must
3068 // reference that register in order to be considered. This prunes out
3069 // unprofitable searching.
3070 SmallSetVector<const SCEV *, 4> ReqRegs;
3071 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3072 E = CurRegs.end(); I != E; ++I)
3073 if (LU.Regs.count(*I))
3076 bool AnySatisfiedReqRegs = false;
3077 SmallPtrSet<const SCEV *, 16> NewRegs;
3080 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3081 E = LU.Formulae.end(); I != E; ++I) {
3082 const Formula &F = *I;
3084 // Ignore formulae which do not use any of the required registers.
3085 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3086 JE = ReqRegs.end(); J != JE; ++J) {
3087 const SCEV *Reg = *J;
3088 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3089 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3093 AnySatisfiedReqRegs = true;
3095 // Evaluate the cost of the current formula. If it's already worse than
3096 // the current best, prune the search at that point.
3099 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3100 if (NewCost < SolutionCost) {
3101 Workspace.push_back(&F);
3102 if (Workspace.size() != Uses.size()) {
3103 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3104 NewRegs, VisitedRegs);
3105 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3106 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3108 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3109 dbgs() << ". Regs:";
3110 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3111 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3112 dbgs() << ' ' << **I;
3115 SolutionCost = NewCost;
3116 Solution = Workspace;
3118 Workspace.pop_back();
3123 // If none of the formulae had all of the required registers, relax the
3124 // constraint so that we don't exclude all formulae.
3125 if (!AnySatisfiedReqRegs) {
3126 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3132 /// Solve - Choose one formula from each use. Return the results in the given
3133 /// Solution vector.
3134 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3135 SmallVector<const Formula *, 8> Workspace;
3137 SolutionCost.Loose();
3139 SmallPtrSet<const SCEV *, 16> CurRegs;
3140 DenseSet<const SCEV *> VisitedRegs;
3141 Workspace.reserve(Uses.size());
3143 // SolveRecurse does all the work.
3144 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3145 CurRegs, VisitedRegs);
3147 // Ok, we've now made all our decisions.
3148 DEBUG(dbgs() << "\n"
3149 "The chosen solution requires "; SolutionCost.print(dbgs());
3151 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3153 Uses[i].print(dbgs());
3156 Solution[i]->print(dbgs());
3160 assert(Solution.size() == Uses.size() && "Malformed solution!");
3163 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3164 /// the dominator tree far as we can go while still being dominated by the
3165 /// input positions. This helps canonicalize the insert position, which
3166 /// encourages sharing.
3167 BasicBlock::iterator
3168 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3169 const SmallVectorImpl<Instruction *> &Inputs)
3172 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3173 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3176 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3177 if (!Rung) return IP;
3178 Rung = Rung->getIDom();
3179 if (!Rung) return IP;
3180 IDom = Rung->getBlock();
3182 // Don't climb into a loop though.
3183 const Loop *IDomLoop = LI.getLoopFor(IDom);
3184 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3185 if (IDomDepth <= IPLoopDepth &&
3186 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3190 bool AllDominate = true;
3191 Instruction *BetterPos = 0;
3192 Instruction *Tentative = IDom->getTerminator();
3193 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3194 E = Inputs.end(); I != E; ++I) {
3195 Instruction *Inst = *I;
3196 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3197 AllDominate = false;
3200 // Attempt to find an insert position in the middle of the block,
3201 // instead of at the end, so that it can be used for other expansions.
3202 if (IDom == Inst->getParent() &&
3203 (!BetterPos || DT.dominates(BetterPos, Inst)))
3204 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3217 /// AdjustInsertPositionForExpand - Determine an input position which will be
3218 /// dominated by the operands and which will dominate the result.
3219 BasicBlock::iterator
3220 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3222 const LSRUse &LU) const {
3223 // Collect some instructions which must be dominated by the
3224 // expanding replacement. These must be dominated by any operands that
3225 // will be required in the expansion.
3226 SmallVector<Instruction *, 4> Inputs;
3227 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3228 Inputs.push_back(I);
3229 if (LU.Kind == LSRUse::ICmpZero)
3230 if (Instruction *I =
3231 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3232 Inputs.push_back(I);
3233 if (LF.PostIncLoops.count(L)) {
3234 if (LF.isUseFullyOutsideLoop(L))
3235 Inputs.push_back(L->getLoopLatch()->getTerminator());
3237 Inputs.push_back(IVIncInsertPos);
3239 // The expansion must also be dominated by the increment positions of any
3240 // loops it for which it is using post-inc mode.
3241 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3242 E = LF.PostIncLoops.end(); I != E; ++I) {
3243 const Loop *PIL = *I;
3244 if (PIL == L) continue;
3246 // Be dominated by the loop exit.
3247 SmallVector<BasicBlock *, 4> ExitingBlocks;
3248 PIL->getExitingBlocks(ExitingBlocks);
3249 if (!ExitingBlocks.empty()) {
3250 BasicBlock *BB = ExitingBlocks[0];
3251 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3252 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3253 Inputs.push_back(BB->getTerminator());
3257 // Then, climb up the immediate dominator tree as far as we can go while
3258 // still being dominated by the input positions.
3259 IP = HoistInsertPosition(IP, Inputs);
3261 // Don't insert instructions before PHI nodes.
3262 while (isa<PHINode>(IP)) ++IP;
3264 // Ignore debug intrinsics.
3265 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3270 /// Expand - Emit instructions for the leading candidate expression for this
3271 /// LSRUse (this is called "expanding").
3272 Value *LSRInstance::Expand(const LSRFixup &LF,
3274 BasicBlock::iterator IP,
3275 SCEVExpander &Rewriter,
3276 SmallVectorImpl<WeakVH> &DeadInsts) const {
3277 const LSRUse &LU = Uses[LF.LUIdx];
3279 // Determine an input position which will be dominated by the operands and
3280 // which will dominate the result.
3281 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3283 // Inform the Rewriter if we have a post-increment use, so that it can
3284 // perform an advantageous expansion.
3285 Rewriter.setPostInc(LF.PostIncLoops);
3287 // This is the type that the user actually needs.
3288 const Type *OpTy = LF.OperandValToReplace->getType();
3289 // This will be the type that we'll initially expand to.
3290 const Type *Ty = F.getType();
3292 // No type known; just expand directly to the ultimate type.
3294 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3295 // Expand directly to the ultimate type if it's the right size.
3297 // This is the type to do integer arithmetic in.
3298 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3300 // Build up a list of operands to add together to form the full base.
3301 SmallVector<const SCEV *, 8> Ops;
3303 // Expand the BaseRegs portion.
3304 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3305 E = F.BaseRegs.end(); I != E; ++I) {
3306 const SCEV *Reg = *I;
3307 assert(!Reg->isZero() && "Zero allocated in a base register!");
3309 // If we're expanding for a post-inc user, make the post-inc adjustment.
3310 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3311 Reg = TransformForPostIncUse(Denormalize, Reg,
3312 LF.UserInst, LF.OperandValToReplace,
3315 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3318 // Flush the operand list to suppress SCEVExpander hoisting.
3320 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3322 Ops.push_back(SE.getUnknown(FullV));
3325 // Expand the ScaledReg portion.
3326 Value *ICmpScaledV = 0;
3327 if (F.AM.Scale != 0) {
3328 const SCEV *ScaledS = F.ScaledReg;
3330 // If we're expanding for a post-inc user, make the post-inc adjustment.
3331 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3332 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3333 LF.UserInst, LF.OperandValToReplace,
3336 if (LU.Kind == LSRUse::ICmpZero) {
3337 // An interesting way of "folding" with an icmp is to use a negated
3338 // scale, which we'll implement by inserting it into the other operand
3340 assert(F.AM.Scale == -1 &&
3341 "The only scale supported by ICmpZero uses is -1!");
3342 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3344 // Otherwise just expand the scaled register and an explicit scale,
3345 // which is expected to be matched as part of the address.
3346 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3347 ScaledS = SE.getMulExpr(ScaledS,
3348 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3349 Ops.push_back(ScaledS);
3351 // Flush the operand list to suppress SCEVExpander hoisting.
3352 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3354 Ops.push_back(SE.getUnknown(FullV));
3358 // Expand the GV portion.
3360 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3362 // Flush the operand list to suppress SCEVExpander hoisting.
3363 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3365 Ops.push_back(SE.getUnknown(FullV));
3368 // Expand the immediate portion.
3369 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3371 if (LU.Kind == LSRUse::ICmpZero) {
3372 // The other interesting way of "folding" with an ICmpZero is to use a
3373 // negated immediate.
3375 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3377 Ops.push_back(SE.getUnknown(ICmpScaledV));
3378 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3381 // Just add the immediate values. These again are expected to be matched
3382 // as part of the address.
3383 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3387 // Emit instructions summing all the operands.
3388 const SCEV *FullS = Ops.empty() ?
3389 SE.getConstant(IntTy, 0) :
3391 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3393 // We're done expanding now, so reset the rewriter.
3394 Rewriter.clearPostInc();
3396 // An ICmpZero Formula represents an ICmp which we're handling as a
3397 // comparison against zero. Now that we've expanded an expression for that
3398 // form, update the ICmp's other operand.
3399 if (LU.Kind == LSRUse::ICmpZero) {
3400 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3401 DeadInsts.push_back(CI->getOperand(1));
3402 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3403 "a scale at the same time!");
3404 if (F.AM.Scale == -1) {
3405 if (ICmpScaledV->getType() != OpTy) {
3407 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3409 ICmpScaledV, OpTy, "tmp", CI);
3412 CI->setOperand(1, ICmpScaledV);
3414 assert(F.AM.Scale == 0 &&
3415 "ICmp does not support folding a global value and "
3416 "a scale at the same time!");
3417 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3419 if (C->getType() != OpTy)
3420 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3424 CI->setOperand(1, C);
3431 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3432 /// of their operands effectively happens in their predecessor blocks, so the
3433 /// expression may need to be expanded in multiple places.
3434 void LSRInstance::RewriteForPHI(PHINode *PN,
3437 SCEVExpander &Rewriter,
3438 SmallVectorImpl<WeakVH> &DeadInsts,
3440 DenseMap<BasicBlock *, Value *> Inserted;
3441 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3442 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3443 BasicBlock *BB = PN->getIncomingBlock(i);
3445 // If this is a critical edge, split the edge so that we do not insert
3446 // the code on all predecessor/successor paths. We do this unless this
3447 // is the canonical backedge for this loop, which complicates post-inc
3449 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3450 !isa<IndirectBrInst>(BB->getTerminator()) &&
3451 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3452 // Split the critical edge.
3453 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3455 // If PN is outside of the loop and BB is in the loop, we want to
3456 // move the block to be immediately before the PHI block, not
3457 // immediately after BB.
3458 if (L->contains(BB) && !L->contains(PN))
3459 NewBB->moveBefore(PN->getParent());
3461 // Splitting the edge can reduce the number of PHI entries we have.
3462 e = PN->getNumIncomingValues();
3464 i = PN->getBasicBlockIndex(BB);
3467 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3468 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3470 PN->setIncomingValue(i, Pair.first->second);
3472 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3474 // If this is reuse-by-noop-cast, insert the noop cast.
3475 const Type *OpTy = LF.OperandValToReplace->getType();
3476 if (FullV->getType() != OpTy)
3478 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3480 FullV, LF.OperandValToReplace->getType(),
3481 "tmp", BB->getTerminator());
3483 PN->setIncomingValue(i, FullV);
3484 Pair.first->second = FullV;
3489 /// Rewrite - Emit instructions for the leading candidate expression for this
3490 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3491 /// the newly expanded value.
3492 void LSRInstance::Rewrite(const LSRFixup &LF,
3494 SCEVExpander &Rewriter,
3495 SmallVectorImpl<WeakVH> &DeadInsts,
3497 // First, find an insertion point that dominates UserInst. For PHI nodes,
3498 // find the nearest block which dominates all the relevant uses.
3499 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3500 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3502 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3504 // If this is reuse-by-noop-cast, insert the noop cast.
3505 const Type *OpTy = LF.OperandValToReplace->getType();
3506 if (FullV->getType() != OpTy) {
3508 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3509 FullV, OpTy, "tmp", LF.UserInst);
3513 // Update the user. ICmpZero is handled specially here (for now) because
3514 // Expand may have updated one of the operands of the icmp already, and
3515 // its new value may happen to be equal to LF.OperandValToReplace, in
3516 // which case doing replaceUsesOfWith leads to replacing both operands
3517 // with the same value. TODO: Reorganize this.
3518 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3519 LF.UserInst->setOperand(0, FullV);
3521 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3524 DeadInsts.push_back(LF.OperandValToReplace);
3527 /// ImplementSolution - Rewrite all the fixup locations with new values,
3528 /// following the chosen solution.
3530 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3532 // Keep track of instructions we may have made dead, so that
3533 // we can remove them after we are done working.
3534 SmallVector<WeakVH, 16> DeadInsts;
3536 SCEVExpander Rewriter(SE);
3537 Rewriter.disableCanonicalMode();
3538 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3540 // Expand the new value definitions and update the users.
3541 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3542 E = Fixups.end(); I != E; ++I) {
3543 const LSRFixup &Fixup = *I;
3545 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3550 // Clean up after ourselves. This must be done before deleting any
3554 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3557 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3558 : IU(P->getAnalysis<IVUsers>()),
3559 SE(P->getAnalysis<ScalarEvolution>()),
3560 DT(P->getAnalysis<DominatorTree>()),
3561 LI(P->getAnalysis<LoopInfo>()),
3562 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3564 // If LoopSimplify form is not available, stay out of trouble.
3565 if (!L->isLoopSimplifyForm()) return;
3567 // If there's no interesting work to be done, bail early.
3568 if (IU.empty()) return;
3570 DEBUG(dbgs() << "\nLSR on loop ";
3571 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3574 // First, perform some low-level loop optimizations.
3576 OptimizeLoopTermCond();
3578 // Start collecting data and preparing for the solver.
3579 CollectInterestingTypesAndFactors();
3580 CollectFixupsAndInitialFormulae();
3581 CollectLoopInvariantFixupsAndFormulae();
3583 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3584 print_uses(dbgs()));
3586 // Now use the reuse data to generate a bunch of interesting ways
3587 // to formulate the values needed for the uses.
3588 GenerateAllReuseFormulae();
3590 DEBUG(dbgs() << "\n"
3591 "After generating reuse formulae:\n";
3592 print_uses(dbgs()));
3594 FilterOutUndesirableDedicatedRegisters();
3595 NarrowSearchSpaceUsingHeuristics();
3597 SmallVector<const Formula *, 8> Solution;
3600 // Release memory that is no longer needed.
3606 // Formulae should be legal.
3607 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3608 E = Uses.end(); I != E; ++I) {
3609 const LSRUse &LU = *I;
3610 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3611 JE = LU.Formulae.end(); J != JE; ++J)
3612 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3613 LU.Kind, LU.AccessTy, TLI) &&
3614 "Illegal formula generated!");
3618 // Now that we've decided what we want, make it so.
3619 ImplementSolution(Solution, P);
3622 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3623 if (Factors.empty() && Types.empty()) return;
3625 OS << "LSR has identified the following interesting factors and types: ";
3628 for (SmallSetVector<int64_t, 8>::const_iterator
3629 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3630 if (!First) OS << ", ";
3635 for (SmallSetVector<const Type *, 4>::const_iterator
3636 I = Types.begin(), E = Types.end(); I != E; ++I) {
3637 if (!First) OS << ", ";
3639 OS << '(' << **I << ')';
3644 void LSRInstance::print_fixups(raw_ostream &OS) const {
3645 OS << "LSR is examining the following fixup sites:\n";
3646 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3647 E = Fixups.end(); I != E; ++I) {
3654 void LSRInstance::print_uses(raw_ostream &OS) const {
3655 OS << "LSR is examining the following uses:\n";
3656 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3657 E = Uses.end(); I != E; ++I) {
3658 const LSRUse &LU = *I;
3662 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3663 JE = LU.Formulae.end(); J != JE; ++J) {
3671 void LSRInstance::print(raw_ostream &OS) const {
3672 print_factors_and_types(OS);
3677 void LSRInstance::dump() const {
3678 print(errs()); errs() << '\n';
3683 class LoopStrengthReduce : public LoopPass {
3684 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3685 /// transformation profitability.
3686 const TargetLowering *const TLI;
3689 static char ID; // Pass ID, replacement for typeid
3690 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3693 bool runOnLoop(Loop *L, LPPassManager &LPM);
3694 void getAnalysisUsage(AnalysisUsage &AU) const;
3699 char LoopStrengthReduce::ID = 0;
3700 static RegisterPass<LoopStrengthReduce>
3701 X("loop-reduce", "Loop Strength Reduction");
3703 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3704 return new LoopStrengthReduce(TLI);
3707 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3708 : LoopPass(&ID), TLI(tli) {}
3710 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3711 // We split critical edges, so we change the CFG. However, we do update
3712 // many analyses if they are around.
3713 AU.addPreservedID(LoopSimplifyID);
3714 AU.addPreserved("domfrontier");
3716 AU.addRequired<LoopInfo>();
3717 AU.addPreserved<LoopInfo>();
3718 AU.addRequiredID(LoopSimplifyID);
3719 AU.addRequired<DominatorTree>();
3720 AU.addPreserved<DominatorTree>();
3721 AU.addRequired<ScalarEvolution>();
3722 AU.addPreserved<ScalarEvolution>();
3723 AU.addRequired<IVUsers>();
3724 AU.addPreserved<IVUsers>();
3727 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3728 bool Changed = false;
3730 // Run the main LSR transformation.
3731 Changed |= LSRInstance(TLI, L, this).getChanged();
3733 // At this point, it is worth checking to see if any recurrence PHIs are also
3734 // dead, so that we can remove them as well.
3735 Changed |= DeleteDeadPHIs(L->getHeader());