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);
116 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
118 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
122 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
123 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
124 iterator begin() { return RegSequence.begin(); }
125 iterator end() { return RegSequence.end(); }
126 const_iterator begin() const { return RegSequence.begin(); }
127 const_iterator end() const { return RegSequence.end(); }
133 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
134 std::pair<RegUsesTy::iterator, bool> Pair =
135 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
136 RegSortData &RSD = Pair.first->second;
138 RegSequence.push_back(Reg);
139 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
140 RSD.UsedByIndices.set(LUIdx);
144 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
145 if (!RegUsesMap.count(Reg)) return false;
146 const SmallBitVector &UsedByIndices =
147 RegUsesMap.find(Reg)->second.UsedByIndices;
148 int i = UsedByIndices.find_first();
149 if (i == -1) return false;
150 if ((size_t)i != LUIdx) return true;
151 return UsedByIndices.find_next(i) != -1;
154 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
155 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
156 assert(I != RegUsesMap.end() && "Unknown register!");
157 return I->second.UsedByIndices;
160 void RegUseTracker::clear() {
167 /// Formula - This class holds information that describes a formula for
168 /// computing satisfying a use. It may include broken-out immediates and scaled
171 /// AM - This is used to represent complex addressing, as well as other kinds
172 /// of interesting uses.
173 TargetLowering::AddrMode AM;
175 /// BaseRegs - The list of "base" registers for this use. When this is
176 /// non-empty, AM.HasBaseReg should be set to true.
177 SmallVector<const SCEV *, 2> BaseRegs;
179 /// ScaledReg - The 'scaled' register for this use. This should be non-null
180 /// when AM.Scale is not zero.
181 const SCEV *ScaledReg;
183 Formula() : ScaledReg(0) {}
185 void InitialMatch(const SCEV *S, Loop *L,
186 ScalarEvolution &SE, DominatorTree &DT);
188 unsigned getNumRegs() const;
189 const Type *getType() const;
191 bool referencesReg(const SCEV *S) const;
192 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
193 const RegUseTracker &RegUses) const;
195 void print(raw_ostream &OS) const;
201 /// DoInitialMatch - Recursion helper for InitialMatch.
202 static void DoInitialMatch(const SCEV *S, Loop *L,
203 SmallVectorImpl<const SCEV *> &Good,
204 SmallVectorImpl<const SCEV *> &Bad,
205 ScalarEvolution &SE, DominatorTree &DT) {
206 // Collect expressions which properly dominate the loop header.
207 if (S->properlyDominates(L->getHeader(), &DT)) {
212 // Look at add operands.
213 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
214 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
216 DoInitialMatch(*I, L, Good, Bad, SE, DT);
220 // Look at addrec operands.
221 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
222 if (!AR->getStart()->isZero()) {
223 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
224 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
225 AR->getStepRecurrence(SE),
227 L, Good, Bad, SE, DT);
231 // Handle a multiplication by -1 (negation) if it didn't fold.
232 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
233 if (Mul->getOperand(0)->isAllOnesValue()) {
234 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
235 const SCEV *NewMul = SE.getMulExpr(Ops);
237 SmallVector<const SCEV *, 4> MyGood;
238 SmallVector<const SCEV *, 4> MyBad;
239 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
240 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
241 SE.getEffectiveSCEVType(NewMul->getType())));
242 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
243 E = MyGood.end(); I != E; ++I)
244 Good.push_back(SE.getMulExpr(NegOne, *I));
245 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
246 E = MyBad.end(); I != E; ++I)
247 Bad.push_back(SE.getMulExpr(NegOne, *I));
251 // Ok, we can't do anything interesting. Just stuff the whole thing into a
252 // register and hope for the best.
256 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
257 /// attempting to keep all loop-invariant and loop-computable values in a
258 /// single base register.
259 void Formula::InitialMatch(const SCEV *S, Loop *L,
260 ScalarEvolution &SE, DominatorTree &DT) {
261 SmallVector<const SCEV *, 4> Good;
262 SmallVector<const SCEV *, 4> Bad;
263 DoInitialMatch(S, L, Good, Bad, SE, DT);
265 const SCEV *Sum = SE.getAddExpr(Good);
267 BaseRegs.push_back(Sum);
268 AM.HasBaseReg = true;
271 const SCEV *Sum = SE.getAddExpr(Bad);
273 BaseRegs.push_back(Sum);
274 AM.HasBaseReg = true;
278 /// getNumRegs - Return the total number of register operands used by this
279 /// formula. This does not include register uses implied by non-constant
281 unsigned Formula::getNumRegs() const {
282 return !!ScaledReg + BaseRegs.size();
285 /// getType - Return the type of this formula, if it has one, or null
286 /// otherwise. This type is meaningless except for the bit size.
287 const Type *Formula::getType() const {
288 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
289 ScaledReg ? ScaledReg->getType() :
290 AM.BaseGV ? AM.BaseGV->getType() :
294 /// referencesReg - Test if this formula references the given register.
295 bool Formula::referencesReg(const SCEV *S) const {
296 return S == ScaledReg ||
297 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
300 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
301 /// which are used by uses other than the use with the given index.
302 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
303 const RegUseTracker &RegUses) const {
305 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
307 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
308 E = BaseRegs.end(); I != E; ++I)
309 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
314 void Formula::print(raw_ostream &OS) const {
317 if (!First) OS << " + "; else First = false;
318 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
320 if (AM.BaseOffs != 0) {
321 if (!First) OS << " + "; else First = false;
324 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
325 E = BaseRegs.end(); I != E; ++I) {
326 if (!First) OS << " + "; else First = false;
327 OS << "reg(" << **I << ')';
329 if (AM.HasBaseReg && BaseRegs.empty()) {
330 if (!First) OS << " + "; else First = false;
331 OS << "**error: HasBaseReg**";
332 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
333 if (!First) OS << " + "; else First = false;
334 OS << "**error: !HasBaseReg**";
337 if (!First) OS << " + "; else First = false;
338 OS << AM.Scale << "*reg(";
347 void Formula::dump() const {
348 print(errs()); errs() << '\n';
351 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
352 /// without changing its value.
353 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
355 IntegerType::get(SE.getContext(),
356 SE.getTypeSizeInBits(AR->getType()) + 1);
357 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
360 /// isAddSExtable - Return true if the given add can be sign-extended
361 /// without changing its value.
362 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
364 IntegerType::get(SE.getContext(),
365 SE.getTypeSizeInBits(A->getType()) + 1);
366 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
369 /// isMulSExtable - Return true if the given add can be sign-extended
370 /// without changing its value.
371 static bool isMulSExtable(const SCEVMulExpr *A, ScalarEvolution &SE) {
373 IntegerType::get(SE.getContext(),
374 SE.getTypeSizeInBits(A->getType()) + 1);
375 return isa<SCEVMulExpr>(SE.getSignExtendExpr(A, WideTy));
378 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
379 /// and if the remainder is known to be zero, or null otherwise. If
380 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
381 /// to Y, ignoring that the multiplication may overflow, which is useful when
382 /// the result will be used in a context where the most significant bits are
384 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
386 bool IgnoreSignificantBits = false) {
387 // Handle the trivial case, which works for any SCEV type.
389 return SE.getConstant(LHS->getType(), 1);
391 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
393 if (RHS->isAllOnesValue())
394 return SE.getMulExpr(LHS, RHS);
396 // Check for a division of a constant by a constant.
397 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
398 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
401 if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
403 return SE.getConstant(C->getValue()->getValue()
404 .sdiv(RC->getValue()->getValue()));
407 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
408 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
409 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
410 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
411 IgnoreSignificantBits);
412 if (!Start) return 0;
413 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
414 IgnoreSignificantBits);
416 return SE.getAddRecExpr(Start, Step, AR->getLoop());
420 // Distribute the sdiv over add operands, if the add doesn't overflow.
421 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
422 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
423 SmallVector<const SCEV *, 8> Ops;
424 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
426 const SCEV *Op = getExactSDiv(*I, RHS, SE,
427 IgnoreSignificantBits);
431 return SE.getAddExpr(Ops);
435 // Check for a multiply operand that we can pull RHS out of.
436 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
437 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
438 SmallVector<const SCEV *, 4> Ops;
440 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
443 if (const SCEV *Q = getExactSDiv(*I, RHS, SE,
444 IgnoreSignificantBits)) {
451 return Found ? SE.getMulExpr(Ops) : 0;
454 // Otherwise we don't know.
458 /// ExtractImmediate - If S involves the addition of a constant integer value,
459 /// return that integer value, and mutate S to point to a new SCEV with that
461 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
462 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
463 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
464 S = SE.getConstant(C->getType(), 0);
465 return C->getValue()->getSExtValue();
467 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
468 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
469 int64_t Result = ExtractImmediate(NewOps.front(), SE);
470 S = SE.getAddExpr(NewOps);
472 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
473 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
474 int64_t Result = ExtractImmediate(NewOps.front(), SE);
475 S = SE.getAddRecExpr(NewOps, AR->getLoop());
481 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
482 /// return that symbol, and mutate S to point to a new SCEV with that
484 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
485 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
486 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
487 S = SE.getConstant(GV->getType(), 0);
490 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
491 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
492 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
493 S = SE.getAddExpr(NewOps);
495 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
496 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
497 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
498 S = SE.getAddRecExpr(NewOps, AR->getLoop());
504 /// isAddressUse - Returns true if the specified instruction is using the
505 /// specified value as an address.
506 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
507 bool isAddress = isa<LoadInst>(Inst);
508 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
509 if (SI->getOperand(1) == OperandVal)
511 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
512 // Addressing modes can also be folded into prefetches and a variety
514 switch (II->getIntrinsicID()) {
516 case Intrinsic::prefetch:
517 case Intrinsic::x86_sse2_loadu_dq:
518 case Intrinsic::x86_sse2_loadu_pd:
519 case Intrinsic::x86_sse_loadu_ps:
520 case Intrinsic::x86_sse_storeu_ps:
521 case Intrinsic::x86_sse2_storeu_pd:
522 case Intrinsic::x86_sse2_storeu_dq:
523 case Intrinsic::x86_sse2_storel_dq:
524 if (II->getOperand(1) == OperandVal)
532 /// getAccessType - Return the type of the memory being accessed.
533 static const Type *getAccessType(const Instruction *Inst) {
534 const Type *AccessTy = Inst->getType();
535 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
536 AccessTy = SI->getOperand(0)->getType();
537 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
538 // Addressing modes can also be folded into prefetches and a variety
540 switch (II->getIntrinsicID()) {
542 case Intrinsic::x86_sse_storeu_ps:
543 case Intrinsic::x86_sse2_storeu_pd:
544 case Intrinsic::x86_sse2_storeu_dq:
545 case Intrinsic::x86_sse2_storel_dq:
546 AccessTy = II->getOperand(1)->getType();
551 // All pointers have the same requirements, so canonicalize them to an
552 // arbitrary pointer type to minimize variation.
553 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
554 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
555 PTy->getAddressSpace());
560 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
561 /// specified set are trivially dead, delete them and see if this makes any of
562 /// their operands subsequently dead.
564 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
565 bool Changed = false;
567 while (!DeadInsts.empty()) {
568 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
570 if (I == 0 || !isInstructionTriviallyDead(I))
573 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
574 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
577 DeadInsts.push_back(U);
580 I->eraseFromParent();
589 /// Cost - This class is used to measure and compare candidate formulae.
591 /// TODO: Some of these could be merged. Also, a lexical ordering
592 /// isn't always optimal.
596 unsigned NumBaseAdds;
602 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
605 unsigned getNumRegs() const { return NumRegs; }
607 bool operator<(const Cost &Other) const;
611 void RateFormula(const Formula &F,
612 SmallPtrSet<const SCEV *, 16> &Regs,
613 const DenseSet<const SCEV *> &VisitedRegs,
615 const SmallVectorImpl<int64_t> &Offsets,
616 ScalarEvolution &SE, DominatorTree &DT);
618 void print(raw_ostream &OS) const;
622 void RateRegister(const SCEV *Reg,
623 SmallPtrSet<const SCEV *, 16> &Regs,
625 ScalarEvolution &SE, DominatorTree &DT);
626 void RatePrimaryRegister(const SCEV *Reg,
627 SmallPtrSet<const SCEV *, 16> &Regs,
629 ScalarEvolution &SE, DominatorTree &DT);
634 /// RateRegister - Tally up interesting quantities from the given register.
635 void Cost::RateRegister(const SCEV *Reg,
636 SmallPtrSet<const SCEV *, 16> &Regs,
638 ScalarEvolution &SE, DominatorTree &DT) {
639 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
640 if (AR->getLoop() == L)
641 AddRecCost += 1; /// TODO: This should be a function of the stride.
643 // If this is an addrec for a loop that's already been visited by LSR,
644 // don't second-guess its addrec phi nodes. LSR isn't currently smart
645 // enough to reason about more than one loop at a time. Consider these
646 // registers free and leave them alone.
647 else if (L->contains(AR->getLoop()) ||
648 (!AR->getLoop()->contains(L) &&
649 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
650 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
651 PHINode *PN = dyn_cast<PHINode>(I); ++I)
652 if (SE.isSCEVable(PN->getType()) &&
653 (SE.getEffectiveSCEVType(PN->getType()) ==
654 SE.getEffectiveSCEVType(AR->getType())) &&
655 SE.getSCEV(PN) == AR)
658 // If this isn't one of the addrecs that the loop already has, it
659 // would require a costly new phi and add. TODO: This isn't
660 // precisely modeled right now.
662 if (!Regs.count(AR->getStart()))
663 RateRegister(AR->getStart(), Regs, L, SE, DT);
666 // Add the step value register, if it needs one.
667 // TODO: The non-affine case isn't precisely modeled here.
668 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
669 if (!Regs.count(AR->getStart()))
670 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
674 // Rough heuristic; favor registers which don't require extra setup
675 // instructions in the preheader.
676 if (!isa<SCEVUnknown>(Reg) &&
677 !isa<SCEVConstant>(Reg) &&
678 !(isa<SCEVAddRecExpr>(Reg) &&
679 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
680 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
684 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
686 void Cost::RatePrimaryRegister(const SCEV *Reg,
687 SmallPtrSet<const SCEV *, 16> &Regs,
689 ScalarEvolution &SE, DominatorTree &DT) {
690 if (Regs.insert(Reg))
691 RateRegister(Reg, Regs, L, SE, DT);
694 void Cost::RateFormula(const Formula &F,
695 SmallPtrSet<const SCEV *, 16> &Regs,
696 const DenseSet<const SCEV *> &VisitedRegs,
698 const SmallVectorImpl<int64_t> &Offsets,
699 ScalarEvolution &SE, DominatorTree &DT) {
700 // Tally up the registers.
701 if (const SCEV *ScaledReg = F.ScaledReg) {
702 if (VisitedRegs.count(ScaledReg)) {
706 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
708 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
709 E = F.BaseRegs.end(); I != E; ++I) {
710 const SCEV *BaseReg = *I;
711 if (VisitedRegs.count(BaseReg)) {
715 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
717 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
718 BaseReg->hasComputableLoopEvolution(L);
721 if (F.BaseRegs.size() > 1)
722 NumBaseAdds += F.BaseRegs.size() - 1;
724 // Tally up the non-zero immediates.
725 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
726 E = Offsets.end(); I != E; ++I) {
727 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
729 ImmCost += 64; // Handle symbolic values conservatively.
730 // TODO: This should probably be the pointer size.
731 else if (Offset != 0)
732 ImmCost += APInt(64, Offset, true).getMinSignedBits();
736 /// Loose - Set this cost to a loosing value.
746 /// operator< - Choose the lower cost.
747 bool Cost::operator<(const Cost &Other) const {
748 if (NumRegs != Other.NumRegs)
749 return NumRegs < Other.NumRegs;
750 if (AddRecCost != Other.AddRecCost)
751 return AddRecCost < Other.AddRecCost;
752 if (NumIVMuls != Other.NumIVMuls)
753 return NumIVMuls < Other.NumIVMuls;
754 if (NumBaseAdds != Other.NumBaseAdds)
755 return NumBaseAdds < Other.NumBaseAdds;
756 if (ImmCost != Other.ImmCost)
757 return ImmCost < Other.ImmCost;
758 if (SetupCost != Other.SetupCost)
759 return SetupCost < Other.SetupCost;
763 void Cost::print(raw_ostream &OS) const {
764 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
766 OS << ", with addrec cost " << AddRecCost;
768 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
769 if (NumBaseAdds != 0)
770 OS << ", plus " << NumBaseAdds << " base add"
771 << (NumBaseAdds == 1 ? "" : "s");
773 OS << ", plus " << ImmCost << " imm cost";
775 OS << ", plus " << SetupCost << " setup cost";
778 void Cost::dump() const {
779 print(errs()); errs() << '\n';
784 /// LSRFixup - An operand value in an instruction which is to be replaced
785 /// with some equivalent, possibly strength-reduced, replacement.
787 /// UserInst - The instruction which will be updated.
788 Instruction *UserInst;
790 /// OperandValToReplace - The operand of the instruction which will
791 /// be replaced. The operand may be used more than once; every instance
792 /// will be replaced.
793 Value *OperandValToReplace;
795 /// PostIncLoops - If this user is to use the post-incremented value of an
796 /// induction variable, this variable is non-null and holds the loop
797 /// associated with the induction variable.
798 PostIncLoopSet PostIncLoops;
800 /// LUIdx - The index of the LSRUse describing the expression which
801 /// this fixup needs, minus an offset (below).
804 /// Offset - A constant offset to be added to the LSRUse expression.
805 /// This allows multiple fixups to share the same LSRUse with different
806 /// offsets, for example in an unrolled loop.
809 bool isUseFullyOutsideLoop(const Loop *L) const;
813 void print(raw_ostream &OS) const;
820 : UserInst(0), OperandValToReplace(0),
821 LUIdx(~size_t(0)), Offset(0) {}
823 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
824 /// value outside of the given loop.
825 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
826 // PHI nodes use their value in their incoming blocks.
827 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
828 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
829 if (PN->getIncomingValue(i) == OperandValToReplace &&
830 L->contains(PN->getIncomingBlock(i)))
835 return !L->contains(UserInst);
838 void LSRFixup::print(raw_ostream &OS) const {
840 // Store is common and interesting enough to be worth special-casing.
841 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
843 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
844 } else if (UserInst->getType()->isVoidTy())
845 OS << UserInst->getOpcodeName();
847 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
849 OS << ", OperandValToReplace=";
850 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
852 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
853 E = PostIncLoops.end(); I != E; ++I) {
854 OS << ", PostIncLoop=";
855 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
858 if (LUIdx != ~size_t(0))
859 OS << ", LUIdx=" << LUIdx;
862 OS << ", Offset=" << Offset;
865 void LSRFixup::dump() const {
866 print(errs()); errs() << '\n';
871 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
872 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
873 struct UniquifierDenseMapInfo {
874 static SmallVector<const SCEV *, 2> getEmptyKey() {
875 SmallVector<const SCEV *, 2> V;
876 V.push_back(reinterpret_cast<const SCEV *>(-1));
880 static SmallVector<const SCEV *, 2> getTombstoneKey() {
881 SmallVector<const SCEV *, 2> V;
882 V.push_back(reinterpret_cast<const SCEV *>(-2));
886 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
888 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
889 E = V.end(); I != E; ++I)
890 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
894 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
895 const SmallVector<const SCEV *, 2> &RHS) {
900 /// LSRUse - This class holds the state that LSR keeps for each use in
901 /// IVUsers, as well as uses invented by LSR itself. It includes information
902 /// about what kinds of things can be folded into the user, information about
903 /// the user itself, and information about how the use may be satisfied.
904 /// TODO: Represent multiple users of the same expression in common?
906 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
909 /// KindType - An enum for a kind of use, indicating what types of
910 /// scaled and immediate operands it might support.
912 Basic, ///< A normal use, with no folding.
913 Special, ///< A special case of basic, allowing -1 scales.
914 Address, ///< An address use; folding according to TargetLowering
915 ICmpZero ///< An equality icmp with both operands folded into one.
916 // TODO: Add a generic icmp too?
920 const Type *AccessTy;
922 SmallVector<int64_t, 8> Offsets;
926 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
927 /// LSRUse are outside of the loop, in which case some special-case heuristics
929 bool AllFixupsOutsideLoop;
931 /// Formulae - A list of ways to build a value that can satisfy this user.
932 /// After the list is populated, one of these is selected heuristically and
933 /// used to formulate a replacement for OperandValToReplace in UserInst.
934 SmallVector<Formula, 12> Formulae;
936 /// Regs - The set of register candidates used by all formulae in this LSRUse.
937 SmallPtrSet<const SCEV *, 4> Regs;
939 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
940 MinOffset(INT64_MAX),
941 MaxOffset(INT64_MIN),
942 AllFixupsOutsideLoop(true) {}
944 bool InsertFormula(const Formula &F);
945 void DeleteFormula(Formula &F);
949 void print(raw_ostream &OS) const;
953 /// InsertFormula - If the given formula has not yet been inserted, add it to
954 /// the list, and return true. Return false otherwise.
955 bool LSRUse::InsertFormula(const Formula &F) {
956 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
957 if (F.ScaledReg) Key.push_back(F.ScaledReg);
958 // Unstable sort by host order ok, because this is only used for uniquifying.
959 std::sort(Key.begin(), Key.end());
961 if (!Uniquifier.insert(Key).second)
964 // Using a register to hold the value of 0 is not profitable.
965 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
966 "Zero allocated in a scaled register!");
968 for (SmallVectorImpl<const SCEV *>::const_iterator I =
969 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
970 assert(!(*I)->isZero() && "Zero allocated in a base register!");
973 // Add the formula to the list.
974 Formulae.push_back(F);
976 // Record registers now being used by this use.
977 if (F.ScaledReg) Regs.insert(F.ScaledReg);
978 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
983 /// DeleteFormula - Remove the given formula from this use's list.
984 void LSRUse::DeleteFormula(Formula &F) {
985 std::swap(F, Formulae.back());
989 void LSRUse::print(raw_ostream &OS) const {
990 OS << "LSR Use: Kind=";
992 case Basic: OS << "Basic"; break;
993 case Special: OS << "Special"; break;
994 case ICmpZero: OS << "ICmpZero"; break;
997 if (AccessTy->isPointerTy())
998 OS << "pointer"; // the full pointer type could be really verbose
1003 OS << ", Offsets={";
1004 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1005 E = Offsets.end(); I != E; ++I) {
1012 if (AllFixupsOutsideLoop)
1013 OS << ", all-fixups-outside-loop";
1016 void LSRUse::dump() const {
1017 print(errs()); errs() << '\n';
1020 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1021 /// be completely folded into the user instruction at isel time. This includes
1022 /// address-mode folding and special icmp tricks.
1023 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1024 LSRUse::KindType Kind, const Type *AccessTy,
1025 const TargetLowering *TLI) {
1027 case LSRUse::Address:
1028 // If we have low-level target information, ask the target if it can
1029 // completely fold this address.
1030 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1032 // Otherwise, just guess that reg+reg addressing is legal.
1033 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1035 case LSRUse::ICmpZero:
1036 // There's not even a target hook for querying whether it would be legal to
1037 // fold a GV into an ICmp.
1041 // ICmp only has two operands; don't allow more than two non-trivial parts.
1042 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1045 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1046 // putting the scaled register in the other operand of the icmp.
1047 if (AM.Scale != 0 && AM.Scale != -1)
1050 // If we have low-level target information, ask the target if it can fold an
1051 // integer immediate on an icmp.
1052 if (AM.BaseOffs != 0) {
1053 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1060 // Only handle single-register values.
1061 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1063 case LSRUse::Special:
1064 // Only handle -1 scales, or no scale.
1065 return AM.Scale == 0 || AM.Scale == -1;
1071 static bool isLegalUse(TargetLowering::AddrMode AM,
1072 int64_t MinOffset, int64_t MaxOffset,
1073 LSRUse::KindType Kind, const Type *AccessTy,
1074 const TargetLowering *TLI) {
1075 // Check for overflow.
1076 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1079 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1080 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1081 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1082 // Check for overflow.
1083 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1086 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1087 return isLegalUse(AM, Kind, AccessTy, TLI);
1092 static bool isAlwaysFoldable(int64_t BaseOffs,
1093 GlobalValue *BaseGV,
1095 LSRUse::KindType Kind, const Type *AccessTy,
1096 const TargetLowering *TLI) {
1097 // Fast-path: zero is always foldable.
1098 if (BaseOffs == 0 && !BaseGV) return true;
1100 // Conservatively, create an address with an immediate and a
1101 // base and a scale.
1102 TargetLowering::AddrMode AM;
1103 AM.BaseOffs = BaseOffs;
1105 AM.HasBaseReg = HasBaseReg;
1106 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1108 return isLegalUse(AM, Kind, AccessTy, TLI);
1111 static bool isAlwaysFoldable(const SCEV *S,
1112 int64_t MinOffset, int64_t MaxOffset,
1114 LSRUse::KindType Kind, const Type *AccessTy,
1115 const TargetLowering *TLI,
1116 ScalarEvolution &SE) {
1117 // Fast-path: zero is always foldable.
1118 if (S->isZero()) return true;
1120 // Conservatively, create an address with an immediate and a
1121 // base and a scale.
1122 int64_t BaseOffs = ExtractImmediate(S, SE);
1123 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1125 // If there's anything else involved, it's not foldable.
1126 if (!S->isZero()) return false;
1128 // Fast-path: zero is always foldable.
1129 if (BaseOffs == 0 && !BaseGV) return true;
1131 // Conservatively, create an address with an immediate and a
1132 // base and a scale.
1133 TargetLowering::AddrMode AM;
1134 AM.BaseOffs = BaseOffs;
1136 AM.HasBaseReg = HasBaseReg;
1137 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1139 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1142 /// FormulaSorter - This class implements an ordering for formulae which sorts
1143 /// the by their standalone cost.
1144 class FormulaSorter {
1145 /// These two sets are kept empty, so that we compute standalone costs.
1146 DenseSet<const SCEV *> VisitedRegs;
1147 SmallPtrSet<const SCEV *, 16> Regs;
1150 ScalarEvolution &SE;
1154 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1155 : L(l), LU(&lu), SE(se), DT(dt) {}
1157 bool operator()(const Formula &A, const Formula &B) {
1159 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1162 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1164 return CostA < CostB;
1168 /// LSRInstance - This class holds state for the main loop strength reduction
1172 ScalarEvolution &SE;
1175 const TargetLowering *const TLI;
1179 /// IVIncInsertPos - This is the insert position that the current loop's
1180 /// induction variable increment should be placed. In simple loops, this is
1181 /// the latch block's terminator. But in more complicated cases, this is a
1182 /// position which will dominate all the in-loop post-increment users.
1183 Instruction *IVIncInsertPos;
1185 /// Factors - Interesting factors between use strides.
1186 SmallSetVector<int64_t, 8> Factors;
1188 /// Types - Interesting use types, to facilitate truncation reuse.
1189 SmallSetVector<const Type *, 4> Types;
1191 /// Fixups - The list of operands which are to be replaced.
1192 SmallVector<LSRFixup, 16> Fixups;
1194 /// Uses - The list of interesting uses.
1195 SmallVector<LSRUse, 16> Uses;
1197 /// RegUses - Track which uses use which register candidates.
1198 RegUseTracker RegUses;
1200 void OptimizeShadowIV();
1201 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1202 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1203 bool OptimizeLoopTermCond();
1205 void CollectInterestingTypesAndFactors();
1206 void CollectFixupsAndInitialFormulae();
1208 LSRFixup &getNewFixup() {
1209 Fixups.push_back(LSRFixup());
1210 return Fixups.back();
1213 // Support for sharing of LSRUses between LSRFixups.
1214 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1217 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1218 LSRUse::KindType Kind, const Type *AccessTy);
1220 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1221 LSRUse::KindType Kind,
1222 const Type *AccessTy);
1225 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1226 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1227 void CountRegisters(const Formula &F, size_t LUIdx);
1228 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1230 void CollectLoopInvariantFixupsAndFormulae();
1232 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1233 unsigned Depth = 0);
1234 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1235 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1236 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1237 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1238 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1239 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1240 void GenerateCrossUseConstantOffsets();
1241 void GenerateAllReuseFormulae();
1243 void FilterOutUndesirableDedicatedRegisters();
1245 size_t EstimateSearchSpaceComplexity() const;
1246 void NarrowSearchSpaceUsingHeuristics();
1248 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1250 SmallVectorImpl<const Formula *> &Workspace,
1251 const Cost &CurCost,
1252 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1253 DenseSet<const SCEV *> &VisitedRegs) const;
1254 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1256 BasicBlock::iterator
1257 HoistInsertPosition(BasicBlock::iterator IP,
1258 const SmallVectorImpl<Instruction *> &Inputs) const;
1259 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1261 const LSRUse &LU) const;
1263 Value *Expand(const LSRFixup &LF,
1265 BasicBlock::iterator IP,
1266 SCEVExpander &Rewriter,
1267 SmallVectorImpl<WeakVH> &DeadInsts) const;
1268 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1270 SCEVExpander &Rewriter,
1271 SmallVectorImpl<WeakVH> &DeadInsts,
1273 void Rewrite(const LSRFixup &LF,
1275 SCEVExpander &Rewriter,
1276 SmallVectorImpl<WeakVH> &DeadInsts,
1278 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1281 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1283 bool getChanged() const { return Changed; }
1285 void print_factors_and_types(raw_ostream &OS) const;
1286 void print_fixups(raw_ostream &OS) const;
1287 void print_uses(raw_ostream &OS) const;
1288 void print(raw_ostream &OS) const;
1294 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1295 /// inside the loop then try to eliminate the cast operation.
1296 void LSRInstance::OptimizeShadowIV() {
1297 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1298 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1301 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1302 UI != E; /* empty */) {
1303 IVUsers::const_iterator CandidateUI = UI;
1305 Instruction *ShadowUse = CandidateUI->getUser();
1306 const Type *DestTy = NULL;
1308 /* If shadow use is a int->float cast then insert a second IV
1309 to eliminate this cast.
1311 for (unsigned i = 0; i < n; ++i)
1317 for (unsigned i = 0; i < n; ++i, ++d)
1320 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1321 DestTy = UCast->getDestTy();
1322 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1323 DestTy = SCast->getDestTy();
1324 if (!DestTy) continue;
1327 // If target does not support DestTy natively then do not apply
1328 // this transformation.
1329 EVT DVT = TLI->getValueType(DestTy);
1330 if (!TLI->isTypeLegal(DVT)) continue;
1333 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1335 if (PH->getNumIncomingValues() != 2) continue;
1337 const Type *SrcTy = PH->getType();
1338 int Mantissa = DestTy->getFPMantissaWidth();
1339 if (Mantissa == -1) continue;
1340 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1343 unsigned Entry, Latch;
1344 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1352 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1353 if (!Init) continue;
1354 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1356 BinaryOperator *Incr =
1357 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1358 if (!Incr) continue;
1359 if (Incr->getOpcode() != Instruction::Add
1360 && Incr->getOpcode() != Instruction::Sub)
1363 /* Initialize new IV, double d = 0.0 in above example. */
1364 ConstantInt *C = NULL;
1365 if (Incr->getOperand(0) == PH)
1366 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1367 else if (Incr->getOperand(1) == PH)
1368 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1374 // Ignore negative constants, as the code below doesn't handle them
1375 // correctly. TODO: Remove this restriction.
1376 if (!C->getValue().isStrictlyPositive()) continue;
1378 /* Add new PHINode. */
1379 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1381 /* create new increment. '++d' in above example. */
1382 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1383 BinaryOperator *NewIncr =
1384 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1385 Instruction::FAdd : Instruction::FSub,
1386 NewPH, CFP, "IV.S.next.", Incr);
1388 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1389 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1391 /* Remove cast operation */
1392 ShadowUse->replaceAllUsesWith(NewPH);
1393 ShadowUse->eraseFromParent();
1398 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1399 /// set the IV user and stride information and return true, otherwise return
1401 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond,
1402 IVStrideUse *&CondUse) {
1403 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1404 if (UI->getUser() == Cond) {
1405 // NOTE: we could handle setcc instructions with multiple uses here, but
1406 // InstCombine does it as well for simple uses, it's not clear that it
1407 // occurs enough in real life to handle.
1414 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1415 /// a max computation.
1417 /// This is a narrow solution to a specific, but acute, problem. For loops
1423 /// } while (++i < n);
1425 /// the trip count isn't just 'n', because 'n' might not be positive. And
1426 /// unfortunately this can come up even for loops where the user didn't use
1427 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1428 /// will commonly be lowered like this:
1434 /// } while (++i < n);
1437 /// and then it's possible for subsequent optimization to obscure the if
1438 /// test in such a way that indvars can't find it.
1440 /// When indvars can't find the if test in loops like this, it creates a
1441 /// max expression, which allows it to give the loop a canonical
1442 /// induction variable:
1445 /// max = n < 1 ? 1 : n;
1448 /// } while (++i != max);
1450 /// Canonical induction variables are necessary because the loop passes
1451 /// are designed around them. The most obvious example of this is the
1452 /// LoopInfo analysis, which doesn't remember trip count values. It
1453 /// expects to be able to rediscover the trip count each time it is
1454 /// needed, and it does this using a simple analysis that only succeeds if
1455 /// the loop has a canonical induction variable.
1457 /// However, when it comes time to generate code, the maximum operation
1458 /// can be quite costly, especially if it's inside of an outer loop.
1460 /// This function solves this problem by detecting this type of loop and
1461 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1462 /// the instructions for the maximum computation.
1464 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1465 // Check that the loop matches the pattern we're looking for.
1466 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1467 Cond->getPredicate() != CmpInst::ICMP_NE)
1470 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1471 if (!Sel || !Sel->hasOneUse()) return Cond;
1473 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1474 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1476 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1478 // Add one to the backedge-taken count to get the trip count.
1479 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1480 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1482 // Check for a max calculation that matches the pattern. There's no check
1483 // for ICMP_ULE here because the comparison would be with zero, which
1484 // isn't interesting.
1485 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1486 const SCEVNAryExpr *Max = 0;
1487 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1488 Pred = ICmpInst::ICMP_SLE;
1490 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1491 Pred = ICmpInst::ICMP_SLT;
1493 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1494 Pred = ICmpInst::ICMP_ULT;
1501 // To handle a max with more than two operands, this optimization would
1502 // require additional checking and setup.
1503 if (Max->getNumOperands() != 2)
1506 const SCEV *MaxLHS = Max->getOperand(0);
1507 const SCEV *MaxRHS = Max->getOperand(1);
1509 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1510 // for a comparison with 1. For <= and >=, a comparison with zero.
1512 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1515 // Check the relevant induction variable for conformance to
1517 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1518 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1519 if (!AR || !AR->isAffine() ||
1520 AR->getStart() != One ||
1521 AR->getStepRecurrence(SE) != One)
1524 assert(AR->getLoop() == L &&
1525 "Loop condition operand is an addrec in a different loop!");
1527 // Check the right operand of the select, and remember it, as it will
1528 // be used in the new comparison instruction.
1530 if (ICmpInst::isTrueWhenEqual(Pred)) {
1531 // Look for n+1, and grab n.
1532 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1533 if (isa<ConstantInt>(BO->getOperand(1)) &&
1534 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1535 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1536 NewRHS = BO->getOperand(0);
1537 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1538 if (isa<ConstantInt>(BO->getOperand(1)) &&
1539 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1540 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1541 NewRHS = BO->getOperand(0);
1544 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1545 NewRHS = Sel->getOperand(1);
1546 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1547 NewRHS = Sel->getOperand(2);
1549 llvm_unreachable("Max doesn't match expected pattern!");
1551 // Determine the new comparison opcode. It may be signed or unsigned,
1552 // and the original comparison may be either equality or inequality.
1553 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1554 Pred = CmpInst::getInversePredicate(Pred);
1556 // Ok, everything looks ok to change the condition into an SLT or SGE and
1557 // delete the max calculation.
1559 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1561 // Delete the max calculation instructions.
1562 Cond->replaceAllUsesWith(NewCond);
1563 CondUse->setUser(NewCond);
1564 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1565 Cond->eraseFromParent();
1566 Sel->eraseFromParent();
1567 if (Cmp->use_empty())
1568 Cmp->eraseFromParent();
1572 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1573 /// postinc iv when possible.
1575 LSRInstance::OptimizeLoopTermCond() {
1576 SmallPtrSet<Instruction *, 4> PostIncs;
1578 BasicBlock *LatchBlock = L->getLoopLatch();
1579 SmallVector<BasicBlock*, 8> ExitingBlocks;
1580 L->getExitingBlocks(ExitingBlocks);
1582 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1583 BasicBlock *ExitingBlock = ExitingBlocks[i];
1585 // Get the terminating condition for the loop if possible. If we
1586 // can, we want to change it to use a post-incremented version of its
1587 // induction variable, to allow coalescing the live ranges for the IV into
1588 // one register value.
1590 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1593 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1594 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1597 // Search IVUsesByStride to find Cond's IVUse if there is one.
1598 IVStrideUse *CondUse = 0;
1599 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1600 if (!FindIVUserForCond(Cond, CondUse))
1603 // If the trip count is computed in terms of a max (due to ScalarEvolution
1604 // being unable to find a sufficient guard, for example), change the loop
1605 // comparison to use SLT or ULT instead of NE.
1606 // One consequence of doing this now is that it disrupts the count-down
1607 // optimization. That's not always a bad thing though, because in such
1608 // cases it may still be worthwhile to avoid a max.
1609 Cond = OptimizeMax(Cond, CondUse);
1611 // If this exiting block dominates the latch block, it may also use
1612 // the post-inc value if it won't be shared with other uses.
1613 // Check for dominance.
1614 if (!DT.dominates(ExitingBlock, LatchBlock))
1617 // Conservatively avoid trying to use the post-inc value in non-latch
1618 // exits if there may be pre-inc users in intervening blocks.
1619 if (LatchBlock != ExitingBlock)
1620 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1621 // Test if the use is reachable from the exiting block. This dominator
1622 // query is a conservative approximation of reachability.
1623 if (&*UI != CondUse &&
1624 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1625 // Conservatively assume there may be reuse if the quotient of their
1626 // strides could be a legal scale.
1627 const SCEV *A = IU.getStride(*CondUse, L);
1628 const SCEV *B = IU.getStride(*UI, L);
1629 if (!A || !B) continue;
1630 if (SE.getTypeSizeInBits(A->getType()) !=
1631 SE.getTypeSizeInBits(B->getType())) {
1632 if (SE.getTypeSizeInBits(A->getType()) >
1633 SE.getTypeSizeInBits(B->getType()))
1634 B = SE.getSignExtendExpr(B, A->getType());
1636 A = SE.getSignExtendExpr(A, B->getType());
1638 if (const SCEVConstant *D =
1639 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1640 // Stride of one or negative one can have reuse with non-addresses.
1641 if (D->getValue()->isOne() ||
1642 D->getValue()->isAllOnesValue())
1643 goto decline_post_inc;
1644 // Avoid weird situations.
1645 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1646 D->getValue()->getValue().isMinSignedValue())
1647 goto decline_post_inc;
1648 // Without TLI, assume that any stride might be valid, and so any
1649 // use might be shared.
1651 goto decline_post_inc;
1652 // Check for possible scaled-address reuse.
1653 const Type *AccessTy = getAccessType(UI->getUser());
1654 TargetLowering::AddrMode AM;
1655 AM.Scale = D->getValue()->getSExtValue();
1656 if (TLI->isLegalAddressingMode(AM, AccessTy))
1657 goto decline_post_inc;
1658 AM.Scale = -AM.Scale;
1659 if (TLI->isLegalAddressingMode(AM, AccessTy))
1660 goto decline_post_inc;
1664 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1667 // It's possible for the setcc instruction to be anywhere in the loop, and
1668 // possible for it to have multiple users. If it is not immediately before
1669 // the exiting block branch, move it.
1670 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1671 if (Cond->hasOneUse()) {
1672 Cond->moveBefore(TermBr);
1674 // Clone the terminating condition and insert into the loopend.
1675 ICmpInst *OldCond = Cond;
1676 Cond = cast<ICmpInst>(Cond->clone());
1677 Cond->setName(L->getHeader()->getName() + ".termcond");
1678 ExitingBlock->getInstList().insert(TermBr, Cond);
1680 // Clone the IVUse, as the old use still exists!
1681 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1682 TermBr->replaceUsesOfWith(OldCond, Cond);
1686 // If we get to here, we know that we can transform the setcc instruction to
1687 // use the post-incremented version of the IV, allowing us to coalesce the
1688 // live ranges for the IV correctly.
1689 CondUse->transformToPostInc(L);
1692 PostIncs.insert(Cond);
1696 // Determine an insertion point for the loop induction variable increment. It
1697 // must dominate all the post-inc comparisons we just set up, and it must
1698 // dominate the loop latch edge.
1699 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1700 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1701 E = PostIncs.end(); I != E; ++I) {
1703 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1705 if (BB == (*I)->getParent())
1706 IVIncInsertPos = *I;
1707 else if (BB != IVIncInsertPos->getParent())
1708 IVIncInsertPos = BB->getTerminator();
1715 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1716 LSRUse::KindType Kind, const Type *AccessTy) {
1717 int64_t NewMinOffset = LU.MinOffset;
1718 int64_t NewMaxOffset = LU.MaxOffset;
1719 const Type *NewAccessTy = AccessTy;
1721 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1722 // something conservative, however this can pessimize in the case that one of
1723 // the uses will have all its uses outside the loop, for example.
1724 if (LU.Kind != Kind)
1726 // Conservatively assume HasBaseReg is true for now.
1727 if (NewOffset < LU.MinOffset) {
1728 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true,
1729 Kind, AccessTy, TLI))
1731 NewMinOffset = NewOffset;
1732 } else if (NewOffset > LU.MaxOffset) {
1733 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true,
1734 Kind, AccessTy, TLI))
1736 NewMaxOffset = NewOffset;
1738 // Check for a mismatched access type, and fall back conservatively as needed.
1739 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1740 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1743 LU.MinOffset = NewMinOffset;
1744 LU.MaxOffset = NewMaxOffset;
1745 LU.AccessTy = NewAccessTy;
1746 if (NewOffset != LU.Offsets.back())
1747 LU.Offsets.push_back(NewOffset);
1751 /// getUse - Return an LSRUse index and an offset value for a fixup which
1752 /// needs the given expression, with the given kind and optional access type.
1753 /// Either reuse an existing use or create a new one, as needed.
1754 std::pair<size_t, int64_t>
1755 LSRInstance::getUse(const SCEV *&Expr,
1756 LSRUse::KindType Kind, const Type *AccessTy) {
1757 const SCEV *Copy = Expr;
1758 int64_t Offset = ExtractImmediate(Expr, SE);
1760 // Basic uses can't accept any offset, for example.
1761 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1766 std::pair<UseMapTy::iterator, bool> P =
1767 UseMap.insert(std::make_pair(Expr, 0));
1769 // A use already existed with this base.
1770 size_t LUIdx = P.first->second;
1771 LSRUse &LU = Uses[LUIdx];
1772 if (reconcileNewOffset(LU, Offset, Kind, AccessTy))
1774 return std::make_pair(LUIdx, Offset);
1777 // Create a new use.
1778 size_t LUIdx = Uses.size();
1779 P.first->second = LUIdx;
1780 Uses.push_back(LSRUse(Kind, AccessTy));
1781 LSRUse &LU = Uses[LUIdx];
1783 // We don't need to track redundant offsets, but we don't need to go out
1784 // of our way here to avoid them.
1785 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1786 LU.Offsets.push_back(Offset);
1788 LU.MinOffset = Offset;
1789 LU.MaxOffset = Offset;
1790 return std::make_pair(LUIdx, Offset);
1793 void LSRInstance::CollectInterestingTypesAndFactors() {
1794 SmallSetVector<const SCEV *, 4> Strides;
1796 // Collect interesting types and strides.
1797 SmallVector<const SCEV *, 4> Worklist;
1798 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1799 const SCEV *Expr = IU.getExpr(*UI);
1801 // Collect interesting types.
1802 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1804 // Add strides for mentioned loops.
1805 Worklist.push_back(Expr);
1807 const SCEV *S = Worklist.pop_back_val();
1808 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1809 Strides.insert(AR->getStepRecurrence(SE));
1810 Worklist.push_back(AR->getStart());
1811 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1812 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1814 } while (!Worklist.empty());
1817 // Compute interesting factors from the set of interesting strides.
1818 for (SmallSetVector<const SCEV *, 4>::const_iterator
1819 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1820 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1821 next(I); NewStrideIter != E; ++NewStrideIter) {
1822 const SCEV *OldStride = *I;
1823 const SCEV *NewStride = *NewStrideIter;
1825 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1826 SE.getTypeSizeInBits(NewStride->getType())) {
1827 if (SE.getTypeSizeInBits(OldStride->getType()) >
1828 SE.getTypeSizeInBits(NewStride->getType()))
1829 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1831 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1833 if (const SCEVConstant *Factor =
1834 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1836 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1837 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1838 } else if (const SCEVConstant *Factor =
1839 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1842 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1843 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1847 // If all uses use the same type, don't bother looking for truncation-based
1849 if (Types.size() == 1)
1852 DEBUG(print_factors_and_types(dbgs()));
1855 void LSRInstance::CollectFixupsAndInitialFormulae() {
1856 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1858 LSRFixup &LF = getNewFixup();
1859 LF.UserInst = UI->getUser();
1860 LF.OperandValToReplace = UI->getOperandValToReplace();
1861 LF.PostIncLoops = UI->getPostIncLoops();
1863 LSRUse::KindType Kind = LSRUse::Basic;
1864 const Type *AccessTy = 0;
1865 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1866 Kind = LSRUse::Address;
1867 AccessTy = getAccessType(LF.UserInst);
1870 const SCEV *S = IU.getExpr(*UI);
1872 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1873 // (N - i == 0), and this allows (N - i) to be the expression that we work
1874 // with rather than just N or i, so we can consider the register
1875 // requirements for both N and i at the same time. Limiting this code to
1876 // equality icmps is not a problem because all interesting loops use
1877 // equality icmps, thanks to IndVarSimplify.
1878 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1879 if (CI->isEquality()) {
1880 // Swap the operands if needed to put the OperandValToReplace on the
1881 // left, for consistency.
1882 Value *NV = CI->getOperand(1);
1883 if (NV == LF.OperandValToReplace) {
1884 CI->setOperand(1, CI->getOperand(0));
1885 CI->setOperand(0, NV);
1888 // x == y --> x - y == 0
1889 const SCEV *N = SE.getSCEV(NV);
1890 if (N->isLoopInvariant(L)) {
1891 Kind = LSRUse::ICmpZero;
1892 S = SE.getMinusSCEV(N, S);
1895 // -1 and the negations of all interesting strides (except the negation
1896 // of -1) are now also interesting.
1897 for (size_t i = 0, e = Factors.size(); i != e; ++i)
1898 if (Factors[i] != -1)
1899 Factors.insert(-(uint64_t)Factors[i]);
1903 // Set up the initial formula for this use.
1904 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
1906 LF.Offset = P.second;
1907 LSRUse &LU = Uses[LF.LUIdx];
1908 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
1910 // If this is the first use of this LSRUse, give it a formula.
1911 if (LU.Formulae.empty()) {
1912 InsertInitialFormula(S, LU, LF.LUIdx);
1913 CountRegisters(LU.Formulae.back(), LF.LUIdx);
1917 DEBUG(print_fixups(dbgs()));
1921 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
1923 F.InitialMatch(S, L, SE, DT);
1924 bool Inserted = InsertFormula(LU, LUIdx, F);
1925 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
1929 LSRInstance::InsertSupplementalFormula(const SCEV *S,
1930 LSRUse &LU, size_t LUIdx) {
1932 F.BaseRegs.push_back(S);
1933 F.AM.HasBaseReg = true;
1934 bool Inserted = InsertFormula(LU, LUIdx, F);
1935 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
1938 /// CountRegisters - Note which registers are used by the given formula,
1939 /// updating RegUses.
1940 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
1942 RegUses.CountRegister(F.ScaledReg, LUIdx);
1943 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1944 E = F.BaseRegs.end(); I != E; ++I)
1945 RegUses.CountRegister(*I, LUIdx);
1948 /// InsertFormula - If the given formula has not yet been inserted, add it to
1949 /// the list, and return true. Return false otherwise.
1950 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
1951 if (!LU.InsertFormula(F))
1954 CountRegisters(F, LUIdx);
1958 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
1959 /// loop-invariant values which we're tracking. These other uses will pin these
1960 /// values in registers, making them less profitable for elimination.
1961 /// TODO: This currently misses non-constant addrec step registers.
1962 /// TODO: Should this give more weight to users inside the loop?
1964 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
1965 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
1966 SmallPtrSet<const SCEV *, 8> Inserted;
1968 while (!Worklist.empty()) {
1969 const SCEV *S = Worklist.pop_back_val();
1971 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
1972 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
1973 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1974 Worklist.push_back(C->getOperand());
1975 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1976 Worklist.push_back(D->getLHS());
1977 Worklist.push_back(D->getRHS());
1978 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
1979 if (!Inserted.insert(U)) continue;
1980 const Value *V = U->getValue();
1981 if (const Instruction *Inst = dyn_cast<Instruction>(V))
1982 if (L->contains(Inst)) continue;
1983 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1985 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
1986 // Ignore non-instructions.
1989 // Ignore instructions in other functions (as can happen with
1991 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
1993 // Ignore instructions not dominated by the loop.
1994 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
1995 UserInst->getParent() :
1996 cast<PHINode>(UserInst)->getIncomingBlock(
1997 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
1998 if (!DT.dominates(L->getHeader(), UseBB))
2000 // Ignore uses which are part of other SCEV expressions, to avoid
2001 // analyzing them multiple times.
2002 if (SE.isSCEVable(UserInst->getType())) {
2003 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2004 // If the user is a no-op, look through to its uses.
2005 if (!isa<SCEVUnknown>(UserS))
2009 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2013 // Ignore icmp instructions which are already being analyzed.
2014 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2015 unsigned OtherIdx = !UI.getOperandNo();
2016 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2017 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2021 LSRFixup &LF = getNewFixup();
2022 LF.UserInst = const_cast<Instruction *>(UserInst);
2023 LF.OperandValToReplace = UI.getUse();
2024 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2026 LF.Offset = P.second;
2027 LSRUse &LU = Uses[LF.LUIdx];
2028 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2029 InsertSupplementalFormula(U, LU, LF.LUIdx);
2030 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2037 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2038 /// separate registers. If C is non-null, multiply each subexpression by C.
2039 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2040 SmallVectorImpl<const SCEV *> &Ops,
2041 ScalarEvolution &SE) {
2042 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2043 // Break out add operands.
2044 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2046 CollectSubexprs(*I, C, Ops, SE);
2048 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2049 // Split a non-zero base out of an addrec.
2050 if (!AR->getStart()->isZero()) {
2051 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2052 AR->getStepRecurrence(SE),
2053 AR->getLoop()), C, Ops, SE);
2054 CollectSubexprs(AR->getStart(), C, Ops, SE);
2057 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2058 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2059 if (Mul->getNumOperands() == 2)
2060 if (const SCEVConstant *Op0 =
2061 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2062 CollectSubexprs(Mul->getOperand(1),
2063 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2069 // Otherwise use the value itself.
2070 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2073 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2075 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2078 // Arbitrarily cap recursion to protect compile time.
2079 if (Depth >= 3) return;
2081 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2082 const SCEV *BaseReg = Base.BaseRegs[i];
2084 SmallVector<const SCEV *, 8> AddOps;
2085 CollectSubexprs(BaseReg, 0, AddOps, SE);
2086 if (AddOps.size() == 1) continue;
2088 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2089 JE = AddOps.end(); J != JE; ++J) {
2090 // Don't pull a constant into a register if the constant could be folded
2091 // into an immediate field.
2092 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2093 Base.getNumRegs() > 1,
2094 LU.Kind, LU.AccessTy, TLI, SE))
2097 // Collect all operands except *J.
2098 SmallVector<const SCEV *, 8> InnerAddOps;
2099 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2100 KE = AddOps.end(); K != KE; ++K)
2102 InnerAddOps.push_back(*K);
2104 // Don't leave just a constant behind in a register if the constant could
2105 // be folded into an immediate field.
2106 if (InnerAddOps.size() == 1 &&
2107 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2108 Base.getNumRegs() > 1,
2109 LU.Kind, LU.AccessTy, TLI, SE))
2112 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2113 if (InnerSum->isZero())
2116 F.BaseRegs[i] = InnerSum;
2117 F.BaseRegs.push_back(*J);
2118 if (InsertFormula(LU, LUIdx, F))
2119 // If that formula hadn't been seen before, recurse to find more like
2121 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2126 /// GenerateCombinations - Generate a formula consisting of all of the
2127 /// loop-dominating registers added into a single register.
2128 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2130 // This method is only interesting on a plurality of registers.
2131 if (Base.BaseRegs.size() <= 1) return;
2135 SmallVector<const SCEV *, 4> Ops;
2136 for (SmallVectorImpl<const SCEV *>::const_iterator
2137 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2138 const SCEV *BaseReg = *I;
2139 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2140 !BaseReg->hasComputableLoopEvolution(L))
2141 Ops.push_back(BaseReg);
2143 F.BaseRegs.push_back(BaseReg);
2145 if (Ops.size() > 1) {
2146 const SCEV *Sum = SE.getAddExpr(Ops);
2147 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2148 // opportunity to fold something. For now, just ignore such cases
2149 // rather than proceed with zero in a register.
2150 if (!Sum->isZero()) {
2151 F.BaseRegs.push_back(Sum);
2152 (void)InsertFormula(LU, LUIdx, F);
2157 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2158 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2160 // We can't add a symbolic offset if the address already contains one.
2161 if (Base.AM.BaseGV) return;
2163 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2164 const SCEV *G = Base.BaseRegs[i];
2165 GlobalValue *GV = ExtractSymbol(G, SE);
2166 if (G->isZero() || !GV)
2170 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2171 LU.Kind, LU.AccessTy, TLI))
2174 (void)InsertFormula(LU, LUIdx, F);
2178 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2179 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2181 // TODO: For now, just add the min and max offset, because it usually isn't
2182 // worthwhile looking at everything inbetween.
2183 SmallVector<int64_t, 4> Worklist;
2184 Worklist.push_back(LU.MinOffset);
2185 if (LU.MaxOffset != LU.MinOffset)
2186 Worklist.push_back(LU.MaxOffset);
2188 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2189 const SCEV *G = Base.BaseRegs[i];
2191 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2192 E = Worklist.end(); I != E; ++I) {
2194 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2195 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2196 LU.Kind, LU.AccessTy, TLI)) {
2197 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2199 (void)InsertFormula(LU, LUIdx, F);
2203 int64_t Imm = ExtractImmediate(G, SE);
2204 if (G->isZero() || Imm == 0)
2207 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2208 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2209 LU.Kind, LU.AccessTy, TLI))
2212 (void)InsertFormula(LU, LUIdx, F);
2216 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2217 /// the comparison. For example, x == y -> x*c == y*c.
2218 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2220 if (LU.Kind != LSRUse::ICmpZero) return;
2222 // Determine the integer type for the base formula.
2223 const Type *IntTy = Base.getType();
2225 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2227 // Don't do this if there is more than one offset.
2228 if (LU.MinOffset != LU.MaxOffset) return;
2230 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2232 // Check each interesting stride.
2233 for (SmallSetVector<int64_t, 8>::const_iterator
2234 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2235 int64_t Factor = *I;
2238 // Check that the multiplication doesn't overflow.
2239 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2241 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2242 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2245 // Check that multiplying with the use offset doesn't overflow.
2246 int64_t Offset = LU.MinOffset;
2247 if (Offset == INT64_MIN && Factor == -1)
2249 Offset = (uint64_t)Offset * Factor;
2250 if (Offset / Factor != LU.MinOffset)
2253 // Check that this scale is legal.
2254 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2257 // Compensate for the use having MinOffset built into it.
2258 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2260 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2262 // Check that multiplying with each base register doesn't overflow.
2263 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2264 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2265 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2269 // Check that multiplying with the scaled register doesn't overflow.
2271 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2272 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2276 // If we make it here and it's legal, add it.
2277 (void)InsertFormula(LU, LUIdx, F);
2282 /// GenerateScales - Generate stride factor reuse formulae by making use of
2283 /// scaled-offset address modes, for example.
2284 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
2286 // Determine the integer type for the base formula.
2287 const Type *IntTy = Base.getType();
2290 // If this Formula already has a scaled register, we can't add another one.
2291 if (Base.AM.Scale != 0) return;
2293 // Check each interesting stride.
2294 for (SmallSetVector<int64_t, 8>::const_iterator
2295 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2296 int64_t Factor = *I;
2298 Base.AM.Scale = Factor;
2299 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2300 // Check whether this scale is going to be legal.
2301 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2302 LU.Kind, LU.AccessTy, TLI)) {
2303 // As a special-case, handle special out-of-loop Basic users specially.
2304 // TODO: Reconsider this special case.
2305 if (LU.Kind == LSRUse::Basic &&
2306 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2307 LSRUse::Special, LU.AccessTy, TLI) &&
2308 LU.AllFixupsOutsideLoop)
2309 LU.Kind = LSRUse::Special;
2313 // For an ICmpZero, negating a solitary base register won't lead to
2315 if (LU.Kind == LSRUse::ICmpZero &&
2316 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2318 // For each addrec base reg, apply the scale, if possible.
2319 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2320 if (const SCEVAddRecExpr *AR =
2321 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2322 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2323 if (FactorS->isZero())
2325 // Divide out the factor, ignoring high bits, since we'll be
2326 // scaling the value back up in the end.
2327 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2328 // TODO: This could be optimized to avoid all the copying.
2330 F.ScaledReg = Quotient;
2331 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2332 F.BaseRegs.pop_back();
2333 (void)InsertFormula(LU, LUIdx, F);
2339 /// GenerateTruncates - Generate reuse formulae from different IV types.
2340 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
2342 // This requires TargetLowering to tell us which truncates are free.
2345 // Don't bother truncating symbolic values.
2346 if (Base.AM.BaseGV) return;
2348 // Determine the integer type for the base formula.
2349 const Type *DstTy = Base.getType();
2351 DstTy = SE.getEffectiveSCEVType(DstTy);
2353 for (SmallSetVector<const Type *, 4>::const_iterator
2354 I = Types.begin(), E = Types.end(); I != E; ++I) {
2355 const Type *SrcTy = *I;
2356 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2359 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2360 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2361 JE = F.BaseRegs.end(); J != JE; ++J)
2362 *J = SE.getAnyExtendExpr(*J, SrcTy);
2364 // TODO: This assumes we've done basic processing on all uses and
2365 // have an idea what the register usage is.
2366 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2369 (void)InsertFormula(LU, LUIdx, F);
2376 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2377 /// defer modifications so that the search phase doesn't have to worry about
2378 /// the data structures moving underneath it.
2382 const SCEV *OrigReg;
2384 WorkItem(size_t LI, int64_t I, const SCEV *R)
2385 : LUIdx(LI), Imm(I), OrigReg(R) {}
2387 void print(raw_ostream &OS) const;
2393 void WorkItem::print(raw_ostream &OS) const {
2394 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2395 << " , add offset " << Imm;
2398 void WorkItem::dump() const {
2399 print(errs()); errs() << '\n';
2402 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2403 /// distance apart and try to form reuse opportunities between them.
2404 void LSRInstance::GenerateCrossUseConstantOffsets() {
2405 // Group the registers by their value without any added constant offset.
2406 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2407 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2409 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2410 SmallVector<const SCEV *, 8> Sequence;
2411 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2413 const SCEV *Reg = *I;
2414 int64_t Imm = ExtractImmediate(Reg, SE);
2415 std::pair<RegMapTy::iterator, bool> Pair =
2416 Map.insert(std::make_pair(Reg, ImmMapTy()));
2418 Sequence.push_back(Reg);
2419 Pair.first->second.insert(std::make_pair(Imm, *I));
2420 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2423 // Now examine each set of registers with the same base value. Build up
2424 // a list of work to do and do the work in a separate step so that we're
2425 // not adding formulae and register counts while we're searching.
2426 SmallVector<WorkItem, 32> WorkItems;
2427 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2428 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2429 E = Sequence.end(); I != E; ++I) {
2430 const SCEV *Reg = *I;
2431 const ImmMapTy &Imms = Map.find(Reg)->second;
2433 // It's not worthwhile looking for reuse if there's only one offset.
2434 if (Imms.size() == 1)
2437 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2438 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2440 dbgs() << ' ' << J->first;
2443 // Examine each offset.
2444 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2446 const SCEV *OrigReg = J->second;
2448 int64_t JImm = J->first;
2449 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2451 if (!isa<SCEVConstant>(OrigReg) &&
2452 UsedByIndicesMap[Reg].count() == 1) {
2453 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2457 // Conservatively examine offsets between this orig reg a few selected
2459 ImmMapTy::const_iterator OtherImms[] = {
2460 Imms.begin(), prior(Imms.end()),
2461 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2463 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2464 ImmMapTy::const_iterator M = OtherImms[i];
2465 if (M == J || M == JE) continue;
2467 // Compute the difference between the two.
2468 int64_t Imm = (uint64_t)JImm - M->first;
2469 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2470 LUIdx = UsedByIndices.find_next(LUIdx))
2471 // Make a memo of this use, offset, and register tuple.
2472 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2473 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2480 UsedByIndicesMap.clear();
2481 UniqueItems.clear();
2483 // Now iterate through the worklist and add new formulae.
2484 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2485 E = WorkItems.end(); I != E; ++I) {
2486 const WorkItem &WI = *I;
2487 size_t LUIdx = WI.LUIdx;
2488 LSRUse &LU = Uses[LUIdx];
2489 int64_t Imm = WI.Imm;
2490 const SCEV *OrigReg = WI.OrigReg;
2492 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2493 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2494 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2496 // TODO: Use a more targeted data structure.
2497 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2498 Formula F = LU.Formulae[L];
2499 // Use the immediate in the scaled register.
2500 if (F.ScaledReg == OrigReg) {
2501 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2502 Imm * (uint64_t)F.AM.Scale;
2503 // Don't create 50 + reg(-50).
2504 if (F.referencesReg(SE.getSCEV(
2505 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2508 NewF.AM.BaseOffs = Offs;
2509 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2510 LU.Kind, LU.AccessTy, TLI))
2512 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2514 // If the new scale is a constant in a register, and adding the constant
2515 // value to the immediate would produce a value closer to zero than the
2516 // immediate itself, then the formula isn't worthwhile.
2517 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2518 if (C->getValue()->getValue().isNegative() !=
2519 (NewF.AM.BaseOffs < 0) &&
2520 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2521 .ule(abs64(NewF.AM.BaseOffs)))
2525 (void)InsertFormula(LU, LUIdx, NewF);
2527 // Use the immediate in a base register.
2528 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2529 const SCEV *BaseReg = F.BaseRegs[N];
2530 if (BaseReg != OrigReg)
2533 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2534 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2535 LU.Kind, LU.AccessTy, TLI))
2537 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2539 // If the new formula has a constant in a register, and adding the
2540 // constant value to the immediate would produce a value closer to
2541 // zero than the immediate itself, then the formula isn't worthwhile.
2542 for (SmallVectorImpl<const SCEV *>::const_iterator
2543 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2545 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2546 if (C->getValue()->getValue().isNegative() !=
2547 (NewF.AM.BaseOffs < 0) &&
2548 C->getValue()->getValue().abs()
2549 .ule(abs64(NewF.AM.BaseOffs)))
2553 (void)InsertFormula(LU, LUIdx, NewF);
2562 /// GenerateAllReuseFormulae - Generate formulae for each use.
2564 LSRInstance::GenerateAllReuseFormulae() {
2565 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2566 // queries are more precise.
2567 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2568 LSRUse &LU = Uses[LUIdx];
2569 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2570 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2571 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2572 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2574 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2575 LSRUse &LU = Uses[LUIdx];
2576 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2577 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2578 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2579 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2580 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2581 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2582 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2583 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2585 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2586 LSRUse &LU = Uses[LUIdx];
2587 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2588 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2591 GenerateCrossUseConstantOffsets();
2594 /// If their are multiple formulae with the same set of registers used
2595 /// by other uses, pick the best one and delete the others.
2596 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2598 bool Changed = false;
2601 // Collect the best formula for each unique set of shared registers. This
2602 // is reset for each use.
2603 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2605 BestFormulaeTy BestFormulae;
2607 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2608 LSRUse &LU = Uses[LUIdx];
2609 FormulaSorter Sorter(L, LU, SE, DT);
2610 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << "\n");
2612 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2613 FIdx != NumForms; ++FIdx) {
2614 Formula &F = LU.Formulae[FIdx];
2616 SmallVector<const SCEV *, 2> Key;
2617 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2618 JE = F.BaseRegs.end(); J != JE; ++J) {
2619 const SCEV *Reg = *J;
2620 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2624 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2625 Key.push_back(F.ScaledReg);
2626 // Unstable sort by host order ok, because this is only used for
2628 std::sort(Key.begin(), Key.end());
2630 std::pair<BestFormulaeTy::const_iterator, bool> P =
2631 BestFormulae.insert(std::make_pair(Key, FIdx));
2633 Formula &Best = LU.Formulae[P.first->second];
2634 if (Sorter.operator()(F, Best))
2636 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2638 " in favor of formula "; Best.print(dbgs());
2643 LU.DeleteFormula(F);
2650 // Now that we've filtered out some formulae, recompute the Regs set.
2652 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2653 FIdx != NumForms; ++FIdx) {
2654 Formula &F = LU.Formulae[FIdx];
2655 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2656 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2659 // Reset this to prepare for the next use.
2660 BestFormulae.clear();
2663 DEBUG(if (Changed) {
2665 "After filtering out undesirable candidates:\n";
2670 // This is a rough guess that seems to work fairly well.
2671 static const size_t ComplexityLimit = UINT16_MAX;
2673 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2674 /// solutions the solver might have to consider. It almost never considers
2675 /// this many solutions because it prune the search space, but the pruning
2676 /// isn't always sufficient.
2677 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2679 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2680 E = Uses.end(); I != E; ++I) {
2681 size_t FSize = I->Formulae.size();
2682 if (FSize >= ComplexityLimit) {
2683 Power = ComplexityLimit;
2687 if (Power >= ComplexityLimit)
2693 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2694 /// formulae to choose from, use some rough heuristics to prune down the number
2695 /// of formulae. This keeps the main solver from taking an extraordinary amount
2696 /// of time in some worst-case scenarios.
2697 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2698 SmallPtrSet<const SCEV *, 4> Taken;
2699 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2700 // Ok, we have too many of formulae on our hands to conveniently handle.
2701 // Use a rough heuristic to thin out the list.
2702 DEBUG(dbgs() << "The search space is too complex.\n");
2704 // Pick the register which is used by the most LSRUses, which is likely
2705 // to be a good reuse register candidate.
2706 const SCEV *Best = 0;
2707 unsigned BestNum = 0;
2708 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2710 const SCEV *Reg = *I;
2711 if (Taken.count(Reg))
2716 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2717 if (Count > BestNum) {
2724 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2725 << " will yield profitable reuse.\n");
2728 // In any use with formulae which references this register, delete formulae
2729 // which don't reference it.
2730 for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(),
2731 E = Uses.end(); I != E; ++I) {
2733 if (!LU.Regs.count(Best)) continue;
2735 // Clear out the set of used regs; it will be recomputed.
2738 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2739 Formula &F = LU.Formulae[i];
2740 if (!F.referencesReg(Best)) {
2741 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2742 LU.DeleteFormula(F);
2745 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
2749 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2750 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2754 DEBUG(dbgs() << "After pre-selection:\n";
2755 print_uses(dbgs()));
2759 /// SolveRecurse - This is the recursive solver.
2760 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2762 SmallVectorImpl<const Formula *> &Workspace,
2763 const Cost &CurCost,
2764 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2765 DenseSet<const SCEV *> &VisitedRegs) const {
2768 // - use more aggressive filtering
2769 // - sort the formula so that the most profitable solutions are found first
2770 // - sort the uses too
2772 // - don't compute a cost, and then compare. compare while computing a cost
2774 // - track register sets with SmallBitVector
2776 const LSRUse &LU = Uses[Workspace.size()];
2778 // If this use references any register that's already a part of the
2779 // in-progress solution, consider it a requirement that a formula must
2780 // reference that register in order to be considered. This prunes out
2781 // unprofitable searching.
2782 SmallSetVector<const SCEV *, 4> ReqRegs;
2783 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
2784 E = CurRegs.end(); I != E; ++I)
2785 if (LU.Regs.count(*I))
2788 bool AnySatisfiedReqRegs = false;
2789 SmallPtrSet<const SCEV *, 16> NewRegs;
2792 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2793 E = LU.Formulae.end(); I != E; ++I) {
2794 const Formula &F = *I;
2796 // Ignore formulae which do not use any of the required registers.
2797 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
2798 JE = ReqRegs.end(); J != JE; ++J) {
2799 const SCEV *Reg = *J;
2800 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
2801 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
2805 AnySatisfiedReqRegs = true;
2807 // Evaluate the cost of the current formula. If it's already worse than
2808 // the current best, prune the search at that point.
2811 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
2812 if (NewCost < SolutionCost) {
2813 Workspace.push_back(&F);
2814 if (Workspace.size() != Uses.size()) {
2815 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
2816 NewRegs, VisitedRegs);
2817 if (F.getNumRegs() == 1 && Workspace.size() == 1)
2818 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
2820 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
2821 dbgs() << ". Regs:";
2822 for (SmallPtrSet<const SCEV *, 16>::const_iterator
2823 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
2824 dbgs() << ' ' << **I;
2827 SolutionCost = NewCost;
2828 Solution = Workspace;
2830 Workspace.pop_back();
2835 // If none of the formulae had all of the required registers, relax the
2836 // constraint so that we don't exclude all formulae.
2837 if (!AnySatisfiedReqRegs) {
2838 assert(!ReqRegs.empty() && "Solver failed even without required registers");
2844 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
2845 SmallVector<const Formula *, 8> Workspace;
2847 SolutionCost.Loose();
2849 SmallPtrSet<const SCEV *, 16> CurRegs;
2850 DenseSet<const SCEV *> VisitedRegs;
2851 Workspace.reserve(Uses.size());
2853 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
2854 CurRegs, VisitedRegs);
2856 // Ok, we've now made all our decisions.
2857 DEBUG(dbgs() << "\n"
2858 "The chosen solution requires "; SolutionCost.print(dbgs());
2860 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
2862 Uses[i].print(dbgs());
2865 Solution[i]->print(dbgs());
2870 /// getImmediateDominator - A handy utility for the specific DominatorTree
2871 /// query that we need here.
2873 static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
2874 DomTreeNode *Node = DT.getNode(BB);
2875 if (!Node) return 0;
2876 Node = Node->getIDom();
2877 if (!Node) return 0;
2878 return Node->getBlock();
2881 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
2882 /// the dominator tree far as we can go while still being dominated by the
2883 /// input positions. This helps canonicalize the insert position, which
2884 /// encourages sharing.
2885 BasicBlock::iterator
2886 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
2887 const SmallVectorImpl<Instruction *> &Inputs)
2890 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
2891 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
2894 for (BasicBlock *Rung = IP->getParent(); ; Rung = IDom) {
2895 IDom = getImmediateDominator(Rung, DT);
2896 if (!IDom) return IP;
2898 // Don't climb into a loop though.
2899 const Loop *IDomLoop = LI.getLoopFor(IDom);
2900 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
2901 if (IDomDepth <= IPLoopDepth &&
2902 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
2906 bool AllDominate = true;
2907 Instruction *BetterPos = 0;
2908 Instruction *Tentative = IDom->getTerminator();
2909 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
2910 E = Inputs.end(); I != E; ++I) {
2911 Instruction *Inst = *I;
2912 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
2913 AllDominate = false;
2916 // Attempt to find an insert position in the middle of the block,
2917 // instead of at the end, so that it can be used for other expansions.
2918 if (IDom == Inst->getParent() &&
2919 (!BetterPos || DT.dominates(BetterPos, Inst)))
2920 BetterPos = llvm::next(BasicBlock::iterator(Inst));
2933 /// AdjustInsertPositionForExpand - Determine an input position which will be
2934 /// dominated by the operands and which will dominate the result.
2935 BasicBlock::iterator
2936 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2938 const LSRUse &LU) const {
2939 // Collect some instructions which must be dominated by the
2940 // expanding replacement. These must be dominated by any operands that
2941 // will be required in the expansion.
2942 SmallVector<Instruction *, 4> Inputs;
2943 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
2944 Inputs.push_back(I);
2945 if (LU.Kind == LSRUse::ICmpZero)
2946 if (Instruction *I =
2947 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
2948 Inputs.push_back(I);
2949 if (LF.PostIncLoops.count(L)) {
2950 if (LF.isUseFullyOutsideLoop(L))
2951 Inputs.push_back(L->getLoopLatch()->getTerminator());
2953 Inputs.push_back(IVIncInsertPos);
2955 // The expansion must also be dominated by the increment positions of any
2956 // loops it for which it is using post-inc mode.
2957 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
2958 E = LF.PostIncLoops.end(); I != E; ++I) {
2959 const Loop *PIL = *I;
2960 if (PIL == L) continue;
2962 // Be dominated by the loop exit.
2963 SmallVector<BasicBlock *, 4> ExitingBlocks;
2964 PIL->getExitingBlocks(ExitingBlocks);
2965 if (!ExitingBlocks.empty()) {
2966 BasicBlock *BB = ExitingBlocks[0];
2967 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
2968 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
2969 Inputs.push_back(BB->getTerminator());
2973 // Then, climb up the immediate dominator tree as far as we can go while
2974 // still being dominated by the input positions.
2975 IP = HoistInsertPosition(IP, Inputs);
2977 // Don't insert instructions before PHI nodes.
2978 while (isa<PHINode>(IP)) ++IP;
2980 // Ignore debug intrinsics.
2981 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
2986 Value *LSRInstance::Expand(const LSRFixup &LF,
2988 BasicBlock::iterator IP,
2989 SCEVExpander &Rewriter,
2990 SmallVectorImpl<WeakVH> &DeadInsts) const {
2991 const LSRUse &LU = Uses[LF.LUIdx];
2993 // Determine an input position which will be dominated by the operands and
2994 // which will dominate the result.
2995 IP = AdjustInsertPositionForExpand(IP, LF, LU);
2997 // Inform the Rewriter if we have a post-increment use, so that it can
2998 // perform an advantageous expansion.
2999 Rewriter.setPostInc(LF.PostIncLoops);
3001 // This is the type that the user actually needs.
3002 const Type *OpTy = LF.OperandValToReplace->getType();
3003 // This will be the type that we'll initially expand to.
3004 const Type *Ty = F.getType();
3006 // No type known; just expand directly to the ultimate type.
3008 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3009 // Expand directly to the ultimate type if it's the right size.
3011 // This is the type to do integer arithmetic in.
3012 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3014 // Build up a list of operands to add together to form the full base.
3015 SmallVector<const SCEV *, 8> Ops;
3017 // Expand the BaseRegs portion.
3018 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3019 E = F.BaseRegs.end(); I != E; ++I) {
3020 const SCEV *Reg = *I;
3021 assert(!Reg->isZero() && "Zero allocated in a base register!");
3023 // If we're expanding for a post-inc user, make the post-inc adjustment.
3024 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3025 Reg = TransformForPostIncUse(Denormalize, Reg,
3026 LF.UserInst, LF.OperandValToReplace,
3029 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3032 // Flush the operand list to suppress SCEVExpander hoisting.
3034 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3036 Ops.push_back(SE.getUnknown(FullV));
3039 // Expand the ScaledReg portion.
3040 Value *ICmpScaledV = 0;
3041 if (F.AM.Scale != 0) {
3042 const SCEV *ScaledS = F.ScaledReg;
3044 // If we're expanding for a post-inc user, make the post-inc adjustment.
3045 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3046 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3047 LF.UserInst, LF.OperandValToReplace,
3050 if (LU.Kind == LSRUse::ICmpZero) {
3051 // An interesting way of "folding" with an icmp is to use a negated
3052 // scale, which we'll implement by inserting it into the other operand
3054 assert(F.AM.Scale == -1 &&
3055 "The only scale supported by ICmpZero uses is -1!");
3056 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3058 // Otherwise just expand the scaled register and an explicit scale,
3059 // which is expected to be matched as part of the address.
3060 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3061 ScaledS = SE.getMulExpr(ScaledS,
3062 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3063 Ops.push_back(ScaledS);
3065 // Flush the operand list to suppress SCEVExpander hoisting.
3066 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3068 Ops.push_back(SE.getUnknown(FullV));
3072 // Expand the GV portion.
3074 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3076 // Flush the operand list to suppress SCEVExpander hoisting.
3077 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3079 Ops.push_back(SE.getUnknown(FullV));
3082 // Expand the immediate portion.
3083 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3085 if (LU.Kind == LSRUse::ICmpZero) {
3086 // The other interesting way of "folding" with an ICmpZero is to use a
3087 // negated immediate.
3089 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3091 Ops.push_back(SE.getUnknown(ICmpScaledV));
3092 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3095 // Just add the immediate values. These again are expected to be matched
3096 // as part of the address.
3097 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3101 // Emit instructions summing all the operands.
3102 const SCEV *FullS = Ops.empty() ?
3103 SE.getConstant(IntTy, 0) :
3105 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3107 // We're done expanding now, so reset the rewriter.
3108 Rewriter.clearPostInc();
3110 // An ICmpZero Formula represents an ICmp which we're handling as a
3111 // comparison against zero. Now that we've expanded an expression for that
3112 // form, update the ICmp's other operand.
3113 if (LU.Kind == LSRUse::ICmpZero) {
3114 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3115 DeadInsts.push_back(CI->getOperand(1));
3116 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3117 "a scale at the same time!");
3118 if (F.AM.Scale == -1) {
3119 if (ICmpScaledV->getType() != OpTy) {
3121 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3123 ICmpScaledV, OpTy, "tmp", CI);
3126 CI->setOperand(1, ICmpScaledV);
3128 assert(F.AM.Scale == 0 &&
3129 "ICmp does not support folding a global value and "
3130 "a scale at the same time!");
3131 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3133 if (C->getType() != OpTy)
3134 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3138 CI->setOperand(1, C);
3145 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3146 /// of their operands effectively happens in their predecessor blocks, so the
3147 /// expression may need to be expanded in multiple places.
3148 void LSRInstance::RewriteForPHI(PHINode *PN,
3151 SCEVExpander &Rewriter,
3152 SmallVectorImpl<WeakVH> &DeadInsts,
3154 DenseMap<BasicBlock *, Value *> Inserted;
3155 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3156 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3157 BasicBlock *BB = PN->getIncomingBlock(i);
3159 // If this is a critical edge, split the edge so that we do not insert
3160 // the code on all predecessor/successor paths. We do this unless this
3161 // is the canonical backedge for this loop, which complicates post-inc
3163 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3164 !isa<IndirectBrInst>(BB->getTerminator()) &&
3165 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3166 // Split the critical edge.
3167 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3169 // If PN is outside of the loop and BB is in the loop, we want to
3170 // move the block to be immediately before the PHI block, not
3171 // immediately after BB.
3172 if (L->contains(BB) && !L->contains(PN))
3173 NewBB->moveBefore(PN->getParent());
3175 // Splitting the edge can reduce the number of PHI entries we have.
3176 e = PN->getNumIncomingValues();
3178 i = PN->getBasicBlockIndex(BB);
3181 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3182 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3184 PN->setIncomingValue(i, Pair.first->second);
3186 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3188 // If this is reuse-by-noop-cast, insert the noop cast.
3189 const Type *OpTy = LF.OperandValToReplace->getType();
3190 if (FullV->getType() != OpTy)
3192 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3194 FullV, LF.OperandValToReplace->getType(),
3195 "tmp", BB->getTerminator());
3197 PN->setIncomingValue(i, FullV);
3198 Pair.first->second = FullV;
3203 /// Rewrite - Emit instructions for the leading candidate expression for this
3204 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3205 /// the newly expanded value.
3206 void LSRInstance::Rewrite(const LSRFixup &LF,
3208 SCEVExpander &Rewriter,
3209 SmallVectorImpl<WeakVH> &DeadInsts,
3211 // First, find an insertion point that dominates UserInst. For PHI nodes,
3212 // find the nearest block which dominates all the relevant uses.
3213 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3214 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3216 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3218 // If this is reuse-by-noop-cast, insert the noop cast.
3219 const Type *OpTy = LF.OperandValToReplace->getType();
3220 if (FullV->getType() != OpTy) {
3222 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3223 FullV, OpTy, "tmp", LF.UserInst);
3227 // Update the user. ICmpZero is handled specially here (for now) because
3228 // Expand may have updated one of the operands of the icmp already, and
3229 // its new value may happen to be equal to LF.OperandValToReplace, in
3230 // which case doing replaceUsesOfWith leads to replacing both operands
3231 // with the same value. TODO: Reorganize this.
3232 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3233 LF.UserInst->setOperand(0, FullV);
3235 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3238 DeadInsts.push_back(LF.OperandValToReplace);
3242 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3244 // Keep track of instructions we may have made dead, so that
3245 // we can remove them after we are done working.
3246 SmallVector<WeakVH, 16> DeadInsts;
3248 SCEVExpander Rewriter(SE);
3249 Rewriter.disableCanonicalMode();
3250 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3252 // Expand the new value definitions and update the users.
3253 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3254 size_t LUIdx = Fixups[i].LUIdx;
3256 Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
3261 // Clean up after ourselves. This must be done before deleting any
3265 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3268 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3269 : IU(P->getAnalysis<IVUsers>()),
3270 SE(P->getAnalysis<ScalarEvolution>()),
3271 DT(P->getAnalysis<DominatorTree>()),
3272 LI(P->getAnalysis<LoopInfo>()),
3273 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3275 // If LoopSimplify form is not available, stay out of trouble.
3276 if (!L->isLoopSimplifyForm()) return;
3278 // If there's no interesting work to be done, bail early.
3279 if (IU.empty()) return;
3281 DEBUG(dbgs() << "\nLSR on loop ";
3282 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3285 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3286 /// inside the loop then try to eliminate the cast operation.
3289 // Change loop terminating condition to use the postinc iv when possible.
3290 Changed |= OptimizeLoopTermCond();
3292 CollectInterestingTypesAndFactors();
3293 CollectFixupsAndInitialFormulae();
3294 CollectLoopInvariantFixupsAndFormulae();
3296 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3297 print_uses(dbgs()));
3299 // Now use the reuse data to generate a bunch of interesting ways
3300 // to formulate the values needed for the uses.
3301 GenerateAllReuseFormulae();
3303 DEBUG(dbgs() << "\n"
3304 "After generating reuse formulae:\n";
3305 print_uses(dbgs()));
3307 FilterOutUndesirableDedicatedRegisters();
3308 NarrowSearchSpaceUsingHeuristics();
3310 SmallVector<const Formula *, 8> Solution;
3312 assert(Solution.size() == Uses.size() && "Malformed solution!");
3314 // Release memory that is no longer needed.
3320 // Formulae should be legal.
3321 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3322 E = Uses.end(); I != E; ++I) {
3323 const LSRUse &LU = *I;
3324 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3325 JE = LU.Formulae.end(); J != JE; ++J)
3326 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3327 LU.Kind, LU.AccessTy, TLI) &&
3328 "Illegal formula generated!");
3332 // Now that we've decided what we want, make it so.
3333 ImplementSolution(Solution, P);
3336 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3337 if (Factors.empty() && Types.empty()) return;
3339 OS << "LSR has identified the following interesting factors and types: ";
3342 for (SmallSetVector<int64_t, 8>::const_iterator
3343 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3344 if (!First) OS << ", ";
3349 for (SmallSetVector<const Type *, 4>::const_iterator
3350 I = Types.begin(), E = Types.end(); I != E; ++I) {
3351 if (!First) OS << ", ";
3353 OS << '(' << **I << ')';
3358 void LSRInstance::print_fixups(raw_ostream &OS) const {
3359 OS << "LSR is examining the following fixup sites:\n";
3360 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3361 E = Fixups.end(); I != E; ++I) {
3362 const LSRFixup &LF = *I;
3369 void LSRInstance::print_uses(raw_ostream &OS) const {
3370 OS << "LSR is examining the following uses:\n";
3371 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3372 E = Uses.end(); I != E; ++I) {
3373 const LSRUse &LU = *I;
3377 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3378 JE = LU.Formulae.end(); J != JE; ++J) {
3386 void LSRInstance::print(raw_ostream &OS) const {
3387 print_factors_and_types(OS);
3392 void LSRInstance::dump() const {
3393 print(errs()); errs() << '\n';
3398 class LoopStrengthReduce : public LoopPass {
3399 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3400 /// transformation profitability.
3401 const TargetLowering *const TLI;
3404 static char ID; // Pass ID, replacement for typeid
3405 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3408 bool runOnLoop(Loop *L, LPPassManager &LPM);
3409 void getAnalysisUsage(AnalysisUsage &AU) const;
3414 char LoopStrengthReduce::ID = 0;
3415 static RegisterPass<LoopStrengthReduce>
3416 X("loop-reduce", "Loop Strength Reduction");
3418 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3419 return new LoopStrengthReduce(TLI);
3422 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3423 : LoopPass(&ID), TLI(tli) {}
3425 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3426 // We split critical edges, so we change the CFG. However, we do update
3427 // many analyses if they are around.
3428 AU.addPreservedID(LoopSimplifyID);
3429 AU.addPreserved("domfrontier");
3431 AU.addRequired<LoopInfo>();
3432 AU.addPreserved<LoopInfo>();
3433 AU.addRequiredID(LoopSimplifyID);
3434 AU.addRequired<DominatorTree>();
3435 AU.addPreserved<DominatorTree>();
3436 AU.addRequired<ScalarEvolution>();
3437 AU.addPreserved<ScalarEvolution>();
3438 AU.addRequired<IVUsers>();
3439 AU.addPreserved<IVUsers>();
3442 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3443 bool Changed = false;
3445 // Run the main LSR transformation.
3446 Changed |= LSRInstance(TLI, L, this).getChanged();
3448 // At this point, it is worth checking to see if any recurrence PHIs are also
3449 // dead, so that we can remove them as well.
3450 Changed |= DeleteDeadPHIs(L->getHeader());