1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/ADT/SmallBitVector.h"
69 #include "llvm/ADT/SetVector.h"
70 #include "llvm/ADT/DenseSet.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/ValueHandle.h"
73 #include "llvm/Support/raw_ostream.h"
74 #include "llvm/Target/TargetLowering.h"
80 /// RegSortData - This class holds data which is used to order reuse candidates.
83 /// UsedByIndices - This represents the set of LSRUse indices which reference
84 /// a particular register.
85 SmallBitVector UsedByIndices;
89 void print(raw_ostream &OS) const;
95 void RegSortData::print(raw_ostream &OS) const {
96 OS << "[NumUses=" << UsedByIndices.count() << ']';
99 void RegSortData::dump() const {
100 print(errs()); errs() << '\n';
105 /// RegUseTracker - Map register candidates to information about how they are
107 class RegUseTracker {
108 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
110 RegUsesTy RegUsesMap;
111 SmallVector<const SCEV *, 16> RegSequence;
114 void CountRegister(const SCEV *Reg, size_t LUIdx);
115 void DropRegister(const SCEV *Reg, size_t LUIdx);
116 void DropUse(size_t LUIdx);
118 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
120 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
124 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
125 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
126 iterator begin() { return RegSequence.begin(); }
127 iterator end() { return RegSequence.end(); }
128 const_iterator begin() const { return RegSequence.begin(); }
129 const_iterator end() const { return RegSequence.end(); }
135 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
136 std::pair<RegUsesTy::iterator, bool> Pair =
137 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
138 RegSortData &RSD = Pair.first->second;
140 RegSequence.push_back(Reg);
141 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
142 RSD.UsedByIndices.set(LUIdx);
146 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
147 RegUsesTy::iterator It = RegUsesMap.find(Reg);
148 assert(It != RegUsesMap.end());
149 RegSortData &RSD = It->second;
150 assert(RSD.UsedByIndices.size() > LUIdx);
151 RSD.UsedByIndices.reset(LUIdx);
155 RegUseTracker::DropUse(size_t LUIdx) {
156 // Remove the use index from every register's use list.
157 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
159 I->second.UsedByIndices.reset(LUIdx);
163 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
164 if (!RegUsesMap.count(Reg)) return false;
165 const SmallBitVector &UsedByIndices =
166 RegUsesMap.find(Reg)->second.UsedByIndices;
167 int i = UsedByIndices.find_first();
168 if (i == -1) return false;
169 if ((size_t)i != LUIdx) return true;
170 return UsedByIndices.find_next(i) != -1;
173 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
174 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
175 assert(I != RegUsesMap.end() && "Unknown register!");
176 return I->second.UsedByIndices;
179 void RegUseTracker::clear() {
186 /// Formula - This class holds information that describes a formula for
187 /// computing satisfying a use. It may include broken-out immediates and scaled
190 /// AM - This is used to represent complex addressing, as well as other kinds
191 /// of interesting uses.
192 TargetLowering::AddrMode AM;
194 /// BaseRegs - The list of "base" registers for this use. When this is
195 /// non-empty, AM.HasBaseReg should be set to true.
196 SmallVector<const SCEV *, 2> BaseRegs;
198 /// ScaledReg - The 'scaled' register for this use. This should be non-null
199 /// when AM.Scale is not zero.
200 const SCEV *ScaledReg;
202 Formula() : ScaledReg(0) {}
204 void InitialMatch(const SCEV *S, Loop *L,
205 ScalarEvolution &SE, DominatorTree &DT);
207 unsigned getNumRegs() const;
208 const Type *getType() const;
210 void DeleteBaseReg(const SCEV *&S);
212 bool referencesReg(const SCEV *S) const;
213 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
214 const RegUseTracker &RegUses) const;
216 void print(raw_ostream &OS) const;
222 /// DoInitialMatch - Recursion helper for InitialMatch.
223 static void DoInitialMatch(const SCEV *S, Loop *L,
224 SmallVectorImpl<const SCEV *> &Good,
225 SmallVectorImpl<const SCEV *> &Bad,
226 ScalarEvolution &SE, DominatorTree &DT) {
227 // Collect expressions which properly dominate the loop header.
228 if (S->properlyDominates(L->getHeader(), &DT)) {
233 // Look at add operands.
234 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
235 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
237 DoInitialMatch(*I, L, Good, Bad, SE, DT);
241 // Look at addrec operands.
242 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
243 if (!AR->getStart()->isZero()) {
244 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
245 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
246 AR->getStepRecurrence(SE),
248 L, Good, Bad, SE, DT);
252 // Handle a multiplication by -1 (negation) if it didn't fold.
253 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
254 if (Mul->getOperand(0)->isAllOnesValue()) {
255 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
256 const SCEV *NewMul = SE.getMulExpr(Ops);
258 SmallVector<const SCEV *, 4> MyGood;
259 SmallVector<const SCEV *, 4> MyBad;
260 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
261 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
262 SE.getEffectiveSCEVType(NewMul->getType())));
263 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
264 E = MyGood.end(); I != E; ++I)
265 Good.push_back(SE.getMulExpr(NegOne, *I));
266 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
267 E = MyBad.end(); I != E; ++I)
268 Bad.push_back(SE.getMulExpr(NegOne, *I));
272 // Ok, we can't do anything interesting. Just stuff the whole thing into a
273 // register and hope for the best.
277 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
278 /// attempting to keep all loop-invariant and loop-computable values in a
279 /// single base register.
280 void Formula::InitialMatch(const SCEV *S, Loop *L,
281 ScalarEvolution &SE, DominatorTree &DT) {
282 SmallVector<const SCEV *, 4> Good;
283 SmallVector<const SCEV *, 4> Bad;
284 DoInitialMatch(S, L, Good, Bad, SE, DT);
286 const SCEV *Sum = SE.getAddExpr(Good);
288 BaseRegs.push_back(Sum);
289 AM.HasBaseReg = true;
292 const SCEV *Sum = SE.getAddExpr(Bad);
294 BaseRegs.push_back(Sum);
295 AM.HasBaseReg = true;
299 /// getNumRegs - Return the total number of register operands used by this
300 /// formula. This does not include register uses implied by non-constant
302 unsigned Formula::getNumRegs() const {
303 return !!ScaledReg + BaseRegs.size();
306 /// getType - Return the type of this formula, if it has one, or null
307 /// otherwise. This type is meaningless except for the bit size.
308 const Type *Formula::getType() const {
309 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
310 ScaledReg ? ScaledReg->getType() :
311 AM.BaseGV ? AM.BaseGV->getType() :
315 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
316 void Formula::DeleteBaseReg(const SCEV *&S) {
317 if (&S != &BaseRegs.back())
318 std::swap(S, BaseRegs.back());
322 /// referencesReg - Test if this formula references the given register.
323 bool Formula::referencesReg(const SCEV *S) const {
324 return S == ScaledReg ||
325 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
328 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
329 /// which are used by uses other than the use with the given index.
330 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
331 const RegUseTracker &RegUses) const {
333 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
335 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
336 E = BaseRegs.end(); I != E; ++I)
337 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
342 void Formula::print(raw_ostream &OS) const {
345 if (!First) OS << " + "; else First = false;
346 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
348 if (AM.BaseOffs != 0) {
349 if (!First) OS << " + "; else First = false;
352 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
353 E = BaseRegs.end(); I != E; ++I) {
354 if (!First) OS << " + "; else First = false;
355 OS << "reg(" << **I << ')';
357 if (AM.HasBaseReg && BaseRegs.empty()) {
358 if (!First) OS << " + "; else First = false;
359 OS << "**error: HasBaseReg**";
360 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
361 if (!First) OS << " + "; else First = false;
362 OS << "**error: !HasBaseReg**";
365 if (!First) OS << " + "; else First = false;
366 OS << AM.Scale << "*reg(";
375 void Formula::dump() const {
376 print(errs()); errs() << '\n';
379 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
380 /// without changing its value.
381 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
383 IntegerType::get(SE.getContext(),
384 SE.getTypeSizeInBits(AR->getType()) + 1);
385 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
388 /// isAddSExtable - Return true if the given add can be sign-extended
389 /// without changing its value.
390 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
392 IntegerType::get(SE.getContext(),
393 SE.getTypeSizeInBits(A->getType()) + 1);
394 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
397 /// isMulSExtable - Return true if the given add can be sign-extended
398 /// without changing its value.
399 static bool isMulSExtable(const SCEVMulExpr *A, ScalarEvolution &SE) {
401 IntegerType::get(SE.getContext(),
402 SE.getTypeSizeInBits(A->getType()) + 1);
403 return isa<SCEVMulExpr>(SE.getSignExtendExpr(A, WideTy));
406 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
407 /// and if the remainder is known to be zero, or null otherwise. If
408 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
409 /// to Y, ignoring that the multiplication may overflow, which is useful when
410 /// the result will be used in a context where the most significant bits are
412 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
414 bool IgnoreSignificantBits = false) {
415 // Handle the trivial case, which works for any SCEV type.
417 return SE.getConstant(LHS->getType(), 1);
419 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
421 if (RHS->isAllOnesValue())
422 return SE.getMulExpr(LHS, RHS);
424 // Check for a division of a constant by a constant.
425 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
426 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
429 if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
431 return SE.getConstant(C->getValue()->getValue()
432 .sdiv(RC->getValue()->getValue()));
435 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
436 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
437 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
438 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
439 IgnoreSignificantBits);
440 if (!Start) return 0;
441 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
442 IgnoreSignificantBits);
444 return SE.getAddRecExpr(Start, Step, AR->getLoop());
448 // Distribute the sdiv over add operands, if the add doesn't overflow.
449 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
450 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
451 SmallVector<const SCEV *, 8> Ops;
452 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
454 const SCEV *Op = getExactSDiv(*I, RHS, SE,
455 IgnoreSignificantBits);
459 return SE.getAddExpr(Ops);
463 // Check for a multiply operand that we can pull RHS out of.
464 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
465 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
466 SmallVector<const SCEV *, 4> Ops;
468 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
472 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
473 IgnoreSignificantBits)) {
479 return Found ? SE.getMulExpr(Ops) : 0;
482 // Otherwise we don't know.
486 /// ExtractImmediate - If S involves the addition of a constant integer value,
487 /// return that integer value, and mutate S to point to a new SCEV with that
489 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
490 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
491 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
492 S = SE.getConstant(C->getType(), 0);
493 return C->getValue()->getSExtValue();
495 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
496 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
497 int64_t Result = ExtractImmediate(NewOps.front(), SE);
498 S = SE.getAddExpr(NewOps);
500 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
501 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
502 int64_t Result = ExtractImmediate(NewOps.front(), SE);
503 S = SE.getAddRecExpr(NewOps, AR->getLoop());
509 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
510 /// return that symbol, and mutate S to point to a new SCEV with that
512 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
513 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
514 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
515 S = SE.getConstant(GV->getType(), 0);
518 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
519 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
520 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
521 S = SE.getAddExpr(NewOps);
523 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
524 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
525 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
526 S = SE.getAddRecExpr(NewOps, AR->getLoop());
532 /// isAddressUse - Returns true if the specified instruction is using the
533 /// specified value as an address.
534 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
535 bool isAddress = isa<LoadInst>(Inst);
536 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
537 if (SI->getOperand(1) == OperandVal)
539 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
540 // Addressing modes can also be folded into prefetches and a variety
542 switch (II->getIntrinsicID()) {
544 case Intrinsic::prefetch:
545 case Intrinsic::x86_sse2_loadu_dq:
546 case Intrinsic::x86_sse2_loadu_pd:
547 case Intrinsic::x86_sse_loadu_ps:
548 case Intrinsic::x86_sse_storeu_ps:
549 case Intrinsic::x86_sse2_storeu_pd:
550 case Intrinsic::x86_sse2_storeu_dq:
551 case Intrinsic::x86_sse2_storel_dq:
552 if (II->getOperand(1) == OperandVal)
560 /// getAccessType - Return the type of the memory being accessed.
561 static const Type *getAccessType(const Instruction *Inst) {
562 const Type *AccessTy = Inst->getType();
563 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
564 AccessTy = SI->getOperand(0)->getType();
565 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
566 // Addressing modes can also be folded into prefetches and a variety
568 switch (II->getIntrinsicID()) {
570 case Intrinsic::x86_sse_storeu_ps:
571 case Intrinsic::x86_sse2_storeu_pd:
572 case Intrinsic::x86_sse2_storeu_dq:
573 case Intrinsic::x86_sse2_storel_dq:
574 AccessTy = II->getOperand(1)->getType();
579 // All pointers have the same requirements, so canonicalize them to an
580 // arbitrary pointer type to minimize variation.
581 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
582 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
583 PTy->getAddressSpace());
588 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
589 /// specified set are trivially dead, delete them and see if this makes any of
590 /// their operands subsequently dead.
592 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
593 bool Changed = false;
595 while (!DeadInsts.empty()) {
596 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
598 if (I == 0 || !isInstructionTriviallyDead(I))
601 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
602 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
605 DeadInsts.push_back(U);
608 I->eraseFromParent();
617 /// Cost - This class is used to measure and compare candidate formulae.
619 /// TODO: Some of these could be merged. Also, a lexical ordering
620 /// isn't always optimal.
624 unsigned NumBaseAdds;
630 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
633 unsigned getNumRegs() const { return NumRegs; }
635 bool operator<(const Cost &Other) const;
639 void RateFormula(const Formula &F,
640 SmallPtrSet<const SCEV *, 16> &Regs,
641 const DenseSet<const SCEV *> &VisitedRegs,
643 const SmallVectorImpl<int64_t> &Offsets,
644 ScalarEvolution &SE, DominatorTree &DT);
646 void print(raw_ostream &OS) const;
650 void RateRegister(const SCEV *Reg,
651 SmallPtrSet<const SCEV *, 16> &Regs,
653 ScalarEvolution &SE, DominatorTree &DT);
654 void RatePrimaryRegister(const SCEV *Reg,
655 SmallPtrSet<const SCEV *, 16> &Regs,
657 ScalarEvolution &SE, DominatorTree &DT);
662 /// RateRegister - Tally up interesting quantities from the given register.
663 void Cost::RateRegister(const SCEV *Reg,
664 SmallPtrSet<const SCEV *, 16> &Regs,
666 ScalarEvolution &SE, DominatorTree &DT) {
667 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
668 if (AR->getLoop() == L)
669 AddRecCost += 1; /// TODO: This should be a function of the stride.
671 // If this is an addrec for a loop that's already been visited by LSR,
672 // don't second-guess its addrec phi nodes. LSR isn't currently smart
673 // enough to reason about more than one loop at a time. Consider these
674 // registers free and leave them alone.
675 else if (L->contains(AR->getLoop()) ||
676 (!AR->getLoop()->contains(L) &&
677 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
678 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
679 PHINode *PN = dyn_cast<PHINode>(I); ++I)
680 if (SE.isSCEVable(PN->getType()) &&
681 (SE.getEffectiveSCEVType(PN->getType()) ==
682 SE.getEffectiveSCEVType(AR->getType())) &&
683 SE.getSCEV(PN) == AR)
686 // If this isn't one of the addrecs that the loop already has, it
687 // would require a costly new phi and add. TODO: This isn't
688 // precisely modeled right now.
690 if (!Regs.count(AR->getStart()))
691 RateRegister(AR->getStart(), Regs, L, SE, DT);
694 // Add the step value register, if it needs one.
695 // TODO: The non-affine case isn't precisely modeled here.
696 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
697 if (!Regs.count(AR->getStart()))
698 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
702 // Rough heuristic; favor registers which don't require extra setup
703 // instructions in the preheader.
704 if (!isa<SCEVUnknown>(Reg) &&
705 !isa<SCEVConstant>(Reg) &&
706 !(isa<SCEVAddRecExpr>(Reg) &&
707 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
708 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
712 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
714 void Cost::RatePrimaryRegister(const SCEV *Reg,
715 SmallPtrSet<const SCEV *, 16> &Regs,
717 ScalarEvolution &SE, DominatorTree &DT) {
718 if (Regs.insert(Reg))
719 RateRegister(Reg, Regs, L, SE, DT);
722 void Cost::RateFormula(const Formula &F,
723 SmallPtrSet<const SCEV *, 16> &Regs,
724 const DenseSet<const SCEV *> &VisitedRegs,
726 const SmallVectorImpl<int64_t> &Offsets,
727 ScalarEvolution &SE, DominatorTree &DT) {
728 // Tally up the registers.
729 if (const SCEV *ScaledReg = F.ScaledReg) {
730 if (VisitedRegs.count(ScaledReg)) {
734 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
736 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
737 E = F.BaseRegs.end(); I != E; ++I) {
738 const SCEV *BaseReg = *I;
739 if (VisitedRegs.count(BaseReg)) {
743 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
745 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
746 BaseReg->hasComputableLoopEvolution(L);
749 if (F.BaseRegs.size() > 1)
750 NumBaseAdds += F.BaseRegs.size() - 1;
752 // Tally up the non-zero immediates.
753 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
754 E = Offsets.end(); I != E; ++I) {
755 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
757 ImmCost += 64; // Handle symbolic values conservatively.
758 // TODO: This should probably be the pointer size.
759 else if (Offset != 0)
760 ImmCost += APInt(64, Offset, true).getMinSignedBits();
764 /// Loose - Set this cost to a loosing value.
774 /// operator< - Choose the lower cost.
775 bool Cost::operator<(const Cost &Other) const {
776 if (NumRegs != Other.NumRegs)
777 return NumRegs < Other.NumRegs;
778 if (AddRecCost != Other.AddRecCost)
779 return AddRecCost < Other.AddRecCost;
780 if (NumIVMuls != Other.NumIVMuls)
781 return NumIVMuls < Other.NumIVMuls;
782 if (NumBaseAdds != Other.NumBaseAdds)
783 return NumBaseAdds < Other.NumBaseAdds;
784 if (ImmCost != Other.ImmCost)
785 return ImmCost < Other.ImmCost;
786 if (SetupCost != Other.SetupCost)
787 return SetupCost < Other.SetupCost;
791 void Cost::print(raw_ostream &OS) const {
792 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
794 OS << ", with addrec cost " << AddRecCost;
796 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
797 if (NumBaseAdds != 0)
798 OS << ", plus " << NumBaseAdds << " base add"
799 << (NumBaseAdds == 1 ? "" : "s");
801 OS << ", plus " << ImmCost << " imm cost";
803 OS << ", plus " << SetupCost << " setup cost";
806 void Cost::dump() const {
807 print(errs()); errs() << '\n';
812 /// LSRFixup - An operand value in an instruction which is to be replaced
813 /// with some equivalent, possibly strength-reduced, replacement.
815 /// UserInst - The instruction which will be updated.
816 Instruction *UserInst;
818 /// OperandValToReplace - The operand of the instruction which will
819 /// be replaced. The operand may be used more than once; every instance
820 /// will be replaced.
821 Value *OperandValToReplace;
823 /// PostIncLoops - If this user is to use the post-incremented value of an
824 /// induction variable, this variable is non-null and holds the loop
825 /// associated with the induction variable.
826 PostIncLoopSet PostIncLoops;
828 /// LUIdx - The index of the LSRUse describing the expression which
829 /// this fixup needs, minus an offset (below).
832 /// Offset - A constant offset to be added to the LSRUse expression.
833 /// This allows multiple fixups to share the same LSRUse with different
834 /// offsets, for example in an unrolled loop.
837 bool isUseFullyOutsideLoop(const Loop *L) const;
841 void print(raw_ostream &OS) const;
848 : UserInst(0), OperandValToReplace(0),
849 LUIdx(~size_t(0)), Offset(0) {}
851 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
852 /// value outside of the given loop.
853 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
854 // PHI nodes use their value in their incoming blocks.
855 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
856 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
857 if (PN->getIncomingValue(i) == OperandValToReplace &&
858 L->contains(PN->getIncomingBlock(i)))
863 return !L->contains(UserInst);
866 void LSRFixup::print(raw_ostream &OS) const {
868 // Store is common and interesting enough to be worth special-casing.
869 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
871 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
872 } else if (UserInst->getType()->isVoidTy())
873 OS << UserInst->getOpcodeName();
875 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
877 OS << ", OperandValToReplace=";
878 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
880 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
881 E = PostIncLoops.end(); I != E; ++I) {
882 OS << ", PostIncLoop=";
883 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
886 if (LUIdx != ~size_t(0))
887 OS << ", LUIdx=" << LUIdx;
890 OS << ", Offset=" << Offset;
893 void LSRFixup::dump() const {
894 print(errs()); errs() << '\n';
899 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
900 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
901 struct UniquifierDenseMapInfo {
902 static SmallVector<const SCEV *, 2> getEmptyKey() {
903 SmallVector<const SCEV *, 2> V;
904 V.push_back(reinterpret_cast<const SCEV *>(-1));
908 static SmallVector<const SCEV *, 2> getTombstoneKey() {
909 SmallVector<const SCEV *, 2> V;
910 V.push_back(reinterpret_cast<const SCEV *>(-2));
914 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
916 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
917 E = V.end(); I != E; ++I)
918 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
922 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
923 const SmallVector<const SCEV *, 2> &RHS) {
928 /// LSRUse - This class holds the state that LSR keeps for each use in
929 /// IVUsers, as well as uses invented by LSR itself. It includes information
930 /// about what kinds of things can be folded into the user, information about
931 /// the user itself, and information about how the use may be satisfied.
932 /// TODO: Represent multiple users of the same expression in common?
934 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
937 /// KindType - An enum for a kind of use, indicating what types of
938 /// scaled and immediate operands it might support.
940 Basic, ///< A normal use, with no folding.
941 Special, ///< A special case of basic, allowing -1 scales.
942 Address, ///< An address use; folding according to TargetLowering
943 ICmpZero ///< An equality icmp with both operands folded into one.
944 // TODO: Add a generic icmp too?
948 const Type *AccessTy;
950 SmallVector<int64_t, 8> Offsets;
954 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
955 /// LSRUse are outside of the loop, in which case some special-case heuristics
957 bool AllFixupsOutsideLoop;
959 /// Formulae - A list of ways to build a value that can satisfy this user.
960 /// After the list is populated, one of these is selected heuristically and
961 /// used to formulate a replacement for OperandValToReplace in UserInst.
962 SmallVector<Formula, 12> Formulae;
964 /// Regs - The set of register candidates used by all formulae in this LSRUse.
965 SmallPtrSet<const SCEV *, 4> Regs;
967 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
968 MinOffset(INT64_MAX),
969 MaxOffset(INT64_MIN),
970 AllFixupsOutsideLoop(true) {}
972 bool HasFormulaWithSameRegs(const Formula &F) const;
973 bool InsertFormula(const Formula &F);
974 void DeleteFormula(Formula &F);
975 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
979 void print(raw_ostream &OS) const;
983 /// HasFormula - Test whether this use as a formula which has the same
984 /// registers as the given formula.
985 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
986 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
987 if (F.ScaledReg) Key.push_back(F.ScaledReg);
988 // Unstable sort by host order ok, because this is only used for uniquifying.
989 std::sort(Key.begin(), Key.end());
990 return Uniquifier.count(Key);
993 /// InsertFormula - If the given formula has not yet been inserted, add it to
994 /// the list, and return true. Return false otherwise.
995 bool LSRUse::InsertFormula(const Formula &F) {
996 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
997 if (F.ScaledReg) Key.push_back(F.ScaledReg);
998 // Unstable sort by host order ok, because this is only used for uniquifying.
999 std::sort(Key.begin(), Key.end());
1001 if (!Uniquifier.insert(Key).second)
1004 // Using a register to hold the value of 0 is not profitable.
1005 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1006 "Zero allocated in a scaled register!");
1008 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1009 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1010 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1013 // Add the formula to the list.
1014 Formulae.push_back(F);
1016 // Record registers now being used by this use.
1017 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1018 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1023 /// DeleteFormula - Remove the given formula from this use's list.
1024 void LSRUse::DeleteFormula(Formula &F) {
1025 if (&F != &Formulae.back())
1026 std::swap(F, Formulae.back());
1027 Formulae.pop_back();
1028 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1031 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1032 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1033 // Now that we've filtered out some formulae, recompute the Regs set.
1034 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1036 for (size_t FIdx = 0, NumForms = Formulae.size(); FIdx != NumForms; ++FIdx) {
1037 Formula &F = Formulae[FIdx];
1038 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1039 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1042 // Update the RegTracker.
1043 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1044 E = OldRegs.end(); I != E; ++I)
1045 if (!Regs.count(*I))
1046 RegUses.DropRegister(*I, LUIdx);
1049 void LSRUse::print(raw_ostream &OS) const {
1050 OS << "LSR Use: Kind=";
1052 case Basic: OS << "Basic"; break;
1053 case Special: OS << "Special"; break;
1054 case ICmpZero: OS << "ICmpZero"; break;
1056 OS << "Address of ";
1057 if (AccessTy->isPointerTy())
1058 OS << "pointer"; // the full pointer type could be really verbose
1063 OS << ", Offsets={";
1064 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1065 E = Offsets.end(); I != E; ++I) {
1072 if (AllFixupsOutsideLoop)
1073 OS << ", all-fixups-outside-loop";
1076 void LSRUse::dump() const {
1077 print(errs()); errs() << '\n';
1080 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1081 /// be completely folded into the user instruction at isel time. This includes
1082 /// address-mode folding and special icmp tricks.
1083 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1084 LSRUse::KindType Kind, const Type *AccessTy,
1085 const TargetLowering *TLI) {
1087 case LSRUse::Address:
1088 // If we have low-level target information, ask the target if it can
1089 // completely fold this address.
1090 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1092 // Otherwise, just guess that reg+reg addressing is legal.
1093 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1095 case LSRUse::ICmpZero:
1096 // There's not even a target hook for querying whether it would be legal to
1097 // fold a GV into an ICmp.
1101 // ICmp only has two operands; don't allow more than two non-trivial parts.
1102 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1105 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1106 // putting the scaled register in the other operand of the icmp.
1107 if (AM.Scale != 0 && AM.Scale != -1)
1110 // If we have low-level target information, ask the target if it can fold an
1111 // integer immediate on an icmp.
1112 if (AM.BaseOffs != 0) {
1113 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1120 // Only handle single-register values.
1121 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1123 case LSRUse::Special:
1124 // Only handle -1 scales, or no scale.
1125 return AM.Scale == 0 || AM.Scale == -1;
1131 static bool isLegalUse(TargetLowering::AddrMode AM,
1132 int64_t MinOffset, int64_t MaxOffset,
1133 LSRUse::KindType Kind, const Type *AccessTy,
1134 const TargetLowering *TLI) {
1135 // Check for overflow.
1136 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1139 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1140 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1141 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1142 // Check for overflow.
1143 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1146 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1147 return isLegalUse(AM, Kind, AccessTy, TLI);
1152 static bool isAlwaysFoldable(int64_t BaseOffs,
1153 GlobalValue *BaseGV,
1155 LSRUse::KindType Kind, const Type *AccessTy,
1156 const TargetLowering *TLI) {
1157 // Fast-path: zero is always foldable.
1158 if (BaseOffs == 0 && !BaseGV) return true;
1160 // Conservatively, create an address with an immediate and a
1161 // base and a scale.
1162 TargetLowering::AddrMode AM;
1163 AM.BaseOffs = BaseOffs;
1165 AM.HasBaseReg = HasBaseReg;
1166 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1168 // Canonicalize a scale of 1 to a base register if the formula doesn't
1169 // already have a base register.
1170 if (!AM.HasBaseReg && AM.Scale == 1) {
1172 AM.HasBaseReg = true;
1175 return isLegalUse(AM, Kind, AccessTy, TLI);
1178 static bool isAlwaysFoldable(const SCEV *S,
1179 int64_t MinOffset, int64_t MaxOffset,
1181 LSRUse::KindType Kind, const Type *AccessTy,
1182 const TargetLowering *TLI,
1183 ScalarEvolution &SE) {
1184 // Fast-path: zero is always foldable.
1185 if (S->isZero()) return true;
1187 // Conservatively, create an address with an immediate and a
1188 // base and a scale.
1189 int64_t BaseOffs = ExtractImmediate(S, SE);
1190 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1192 // If there's anything else involved, it's not foldable.
1193 if (!S->isZero()) return false;
1195 // Fast-path: zero is always foldable.
1196 if (BaseOffs == 0 && !BaseGV) return true;
1198 // Conservatively, create an address with an immediate and a
1199 // base and a scale.
1200 TargetLowering::AddrMode AM;
1201 AM.BaseOffs = BaseOffs;
1203 AM.HasBaseReg = HasBaseReg;
1204 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1206 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1209 /// FormulaSorter - This class implements an ordering for formulae which sorts
1210 /// the by their standalone cost.
1211 class FormulaSorter {
1212 /// These two sets are kept empty, so that we compute standalone costs.
1213 DenseSet<const SCEV *> VisitedRegs;
1214 SmallPtrSet<const SCEV *, 16> Regs;
1217 ScalarEvolution &SE;
1221 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1222 : L(l), LU(&lu), SE(se), DT(dt) {}
1224 bool operator()(const Formula &A, const Formula &B) {
1226 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1229 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1231 return CostA < CostB;
1235 /// LSRInstance - This class holds state for the main loop strength reduction
1239 ScalarEvolution &SE;
1242 const TargetLowering *const TLI;
1246 /// IVIncInsertPos - This is the insert position that the current loop's
1247 /// induction variable increment should be placed. In simple loops, this is
1248 /// the latch block's terminator. But in more complicated cases, this is a
1249 /// position which will dominate all the in-loop post-increment users.
1250 Instruction *IVIncInsertPos;
1252 /// Factors - Interesting factors between use strides.
1253 SmallSetVector<int64_t, 8> Factors;
1255 /// Types - Interesting use types, to facilitate truncation reuse.
1256 SmallSetVector<const Type *, 4> Types;
1258 /// Fixups - The list of operands which are to be replaced.
1259 SmallVector<LSRFixup, 16> Fixups;
1261 /// Uses - The list of interesting uses.
1262 SmallVector<LSRUse, 16> Uses;
1264 /// RegUses - Track which uses use which register candidates.
1265 RegUseTracker RegUses;
1267 void OptimizeShadowIV();
1268 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1269 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1270 bool OptimizeLoopTermCond();
1272 void CollectInterestingTypesAndFactors();
1273 void CollectFixupsAndInitialFormulae();
1275 LSRFixup &getNewFixup() {
1276 Fixups.push_back(LSRFixup());
1277 return Fixups.back();
1280 // Support for sharing of LSRUses between LSRFixups.
1281 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1284 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1286 LSRUse::KindType Kind, const Type *AccessTy);
1288 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1289 LSRUse::KindType Kind,
1290 const Type *AccessTy);
1292 void DeleteUse(LSRUse &LU);
1294 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1297 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1298 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1299 void CountRegisters(const Formula &F, size_t LUIdx);
1300 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1302 void CollectLoopInvariantFixupsAndFormulae();
1304 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1305 unsigned Depth = 0);
1306 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1307 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1308 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1309 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1310 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1311 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1312 void GenerateCrossUseConstantOffsets();
1313 void GenerateAllReuseFormulae();
1315 void FilterOutUndesirableDedicatedRegisters();
1317 size_t EstimateSearchSpaceComplexity() const;
1318 void NarrowSearchSpaceUsingHeuristics();
1320 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1322 SmallVectorImpl<const Formula *> &Workspace,
1323 const Cost &CurCost,
1324 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1325 DenseSet<const SCEV *> &VisitedRegs) const;
1326 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1328 BasicBlock::iterator
1329 HoistInsertPosition(BasicBlock::iterator IP,
1330 const SmallVectorImpl<Instruction *> &Inputs) const;
1331 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1333 const LSRUse &LU) const;
1335 Value *Expand(const LSRFixup &LF,
1337 BasicBlock::iterator IP,
1338 SCEVExpander &Rewriter,
1339 SmallVectorImpl<WeakVH> &DeadInsts) const;
1340 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1342 SCEVExpander &Rewriter,
1343 SmallVectorImpl<WeakVH> &DeadInsts,
1345 void Rewrite(const LSRFixup &LF,
1347 SCEVExpander &Rewriter,
1348 SmallVectorImpl<WeakVH> &DeadInsts,
1350 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1353 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1355 bool getChanged() const { return Changed; }
1357 void print_factors_and_types(raw_ostream &OS) const;
1358 void print_fixups(raw_ostream &OS) const;
1359 void print_uses(raw_ostream &OS) const;
1360 void print(raw_ostream &OS) const;
1366 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1367 /// inside the loop then try to eliminate the cast operation.
1368 void LSRInstance::OptimizeShadowIV() {
1369 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1370 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1373 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1374 UI != E; /* empty */) {
1375 IVUsers::const_iterator CandidateUI = UI;
1377 Instruction *ShadowUse = CandidateUI->getUser();
1378 const Type *DestTy = NULL;
1380 /* If shadow use is a int->float cast then insert a second IV
1381 to eliminate this cast.
1383 for (unsigned i = 0; i < n; ++i)
1389 for (unsigned i = 0; i < n; ++i, ++d)
1392 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1393 DestTy = UCast->getDestTy();
1394 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1395 DestTy = SCast->getDestTy();
1396 if (!DestTy) continue;
1399 // If target does not support DestTy natively then do not apply
1400 // this transformation.
1401 EVT DVT = TLI->getValueType(DestTy);
1402 if (!TLI->isTypeLegal(DVT)) continue;
1405 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1407 if (PH->getNumIncomingValues() != 2) continue;
1409 const Type *SrcTy = PH->getType();
1410 int Mantissa = DestTy->getFPMantissaWidth();
1411 if (Mantissa == -1) continue;
1412 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1415 unsigned Entry, Latch;
1416 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1424 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1425 if (!Init) continue;
1426 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1428 BinaryOperator *Incr =
1429 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1430 if (!Incr) continue;
1431 if (Incr->getOpcode() != Instruction::Add
1432 && Incr->getOpcode() != Instruction::Sub)
1435 /* Initialize new IV, double d = 0.0 in above example. */
1436 ConstantInt *C = NULL;
1437 if (Incr->getOperand(0) == PH)
1438 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1439 else if (Incr->getOperand(1) == PH)
1440 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1446 // Ignore negative constants, as the code below doesn't handle them
1447 // correctly. TODO: Remove this restriction.
1448 if (!C->getValue().isStrictlyPositive()) continue;
1450 /* Add new PHINode. */
1451 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1453 /* create new increment. '++d' in above example. */
1454 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1455 BinaryOperator *NewIncr =
1456 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1457 Instruction::FAdd : Instruction::FSub,
1458 NewPH, CFP, "IV.S.next.", Incr);
1460 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1461 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1463 /* Remove cast operation */
1464 ShadowUse->replaceAllUsesWith(NewPH);
1465 ShadowUse->eraseFromParent();
1470 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1471 /// set the IV user and stride information and return true, otherwise return
1473 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond,
1474 IVStrideUse *&CondUse) {
1475 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1476 if (UI->getUser() == Cond) {
1477 // NOTE: we could handle setcc instructions with multiple uses here, but
1478 // InstCombine does it as well for simple uses, it's not clear that it
1479 // occurs enough in real life to handle.
1486 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1487 /// a max computation.
1489 /// This is a narrow solution to a specific, but acute, problem. For loops
1495 /// } while (++i < n);
1497 /// the trip count isn't just 'n', because 'n' might not be positive. And
1498 /// unfortunately this can come up even for loops where the user didn't use
1499 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1500 /// will commonly be lowered like this:
1506 /// } while (++i < n);
1509 /// and then it's possible for subsequent optimization to obscure the if
1510 /// test in such a way that indvars can't find it.
1512 /// When indvars can't find the if test in loops like this, it creates a
1513 /// max expression, which allows it to give the loop a canonical
1514 /// induction variable:
1517 /// max = n < 1 ? 1 : n;
1520 /// } while (++i != max);
1522 /// Canonical induction variables are necessary because the loop passes
1523 /// are designed around them. The most obvious example of this is the
1524 /// LoopInfo analysis, which doesn't remember trip count values. It
1525 /// expects to be able to rediscover the trip count each time it is
1526 /// needed, and it does this using a simple analysis that only succeeds if
1527 /// the loop has a canonical induction variable.
1529 /// However, when it comes time to generate code, the maximum operation
1530 /// can be quite costly, especially if it's inside of an outer loop.
1532 /// This function solves this problem by detecting this type of loop and
1533 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1534 /// the instructions for the maximum computation.
1536 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1537 // Check that the loop matches the pattern we're looking for.
1538 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1539 Cond->getPredicate() != CmpInst::ICMP_NE)
1542 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1543 if (!Sel || !Sel->hasOneUse()) return Cond;
1545 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1546 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1548 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1550 // Add one to the backedge-taken count to get the trip count.
1551 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1552 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1554 // Check for a max calculation that matches the pattern. There's no check
1555 // for ICMP_ULE here because the comparison would be with zero, which
1556 // isn't interesting.
1557 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1558 const SCEVNAryExpr *Max = 0;
1559 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1560 Pred = ICmpInst::ICMP_SLE;
1562 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1563 Pred = ICmpInst::ICMP_SLT;
1565 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1566 Pred = ICmpInst::ICMP_ULT;
1573 // To handle a max with more than two operands, this optimization would
1574 // require additional checking and setup.
1575 if (Max->getNumOperands() != 2)
1578 const SCEV *MaxLHS = Max->getOperand(0);
1579 const SCEV *MaxRHS = Max->getOperand(1);
1581 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1582 // for a comparison with 1. For <= and >=, a comparison with zero.
1584 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1587 // Check the relevant induction variable for conformance to
1589 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1590 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1591 if (!AR || !AR->isAffine() ||
1592 AR->getStart() != One ||
1593 AR->getStepRecurrence(SE) != One)
1596 assert(AR->getLoop() == L &&
1597 "Loop condition operand is an addrec in a different loop!");
1599 // Check the right operand of the select, and remember it, as it will
1600 // be used in the new comparison instruction.
1602 if (ICmpInst::isTrueWhenEqual(Pred)) {
1603 // Look for n+1, and grab n.
1604 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1605 if (isa<ConstantInt>(BO->getOperand(1)) &&
1606 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1607 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1608 NewRHS = BO->getOperand(0);
1609 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1610 if (isa<ConstantInt>(BO->getOperand(1)) &&
1611 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1612 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1613 NewRHS = BO->getOperand(0);
1616 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1617 NewRHS = Sel->getOperand(1);
1618 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1619 NewRHS = Sel->getOperand(2);
1621 llvm_unreachable("Max doesn't match expected pattern!");
1623 // Determine the new comparison opcode. It may be signed or unsigned,
1624 // and the original comparison may be either equality or inequality.
1625 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1626 Pred = CmpInst::getInversePredicate(Pred);
1628 // Ok, everything looks ok to change the condition into an SLT or SGE and
1629 // delete the max calculation.
1631 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1633 // Delete the max calculation instructions.
1634 Cond->replaceAllUsesWith(NewCond);
1635 CondUse->setUser(NewCond);
1636 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1637 Cond->eraseFromParent();
1638 Sel->eraseFromParent();
1639 if (Cmp->use_empty())
1640 Cmp->eraseFromParent();
1644 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1645 /// postinc iv when possible.
1647 LSRInstance::OptimizeLoopTermCond() {
1648 SmallPtrSet<Instruction *, 4> PostIncs;
1650 BasicBlock *LatchBlock = L->getLoopLatch();
1651 SmallVector<BasicBlock*, 8> ExitingBlocks;
1652 L->getExitingBlocks(ExitingBlocks);
1654 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1655 BasicBlock *ExitingBlock = ExitingBlocks[i];
1657 // Get the terminating condition for the loop if possible. If we
1658 // can, we want to change it to use a post-incremented version of its
1659 // induction variable, to allow coalescing the live ranges for the IV into
1660 // one register value.
1662 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1665 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1666 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1669 // Search IVUsesByStride to find Cond's IVUse if there is one.
1670 IVStrideUse *CondUse = 0;
1671 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1672 if (!FindIVUserForCond(Cond, CondUse))
1675 // If the trip count is computed in terms of a max (due to ScalarEvolution
1676 // being unable to find a sufficient guard, for example), change the loop
1677 // comparison to use SLT or ULT instead of NE.
1678 // One consequence of doing this now is that it disrupts the count-down
1679 // optimization. That's not always a bad thing though, because in such
1680 // cases it may still be worthwhile to avoid a max.
1681 Cond = OptimizeMax(Cond, CondUse);
1683 // If this exiting block dominates the latch block, it may also use
1684 // the post-inc value if it won't be shared with other uses.
1685 // Check for dominance.
1686 if (!DT.dominates(ExitingBlock, LatchBlock))
1689 // Conservatively avoid trying to use the post-inc value in non-latch
1690 // exits if there may be pre-inc users in intervening blocks.
1691 if (LatchBlock != ExitingBlock)
1692 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1693 // Test if the use is reachable from the exiting block. This dominator
1694 // query is a conservative approximation of reachability.
1695 if (&*UI != CondUse &&
1696 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1697 // Conservatively assume there may be reuse if the quotient of their
1698 // strides could be a legal scale.
1699 const SCEV *A = IU.getStride(*CondUse, L);
1700 const SCEV *B = IU.getStride(*UI, L);
1701 if (!A || !B) continue;
1702 if (SE.getTypeSizeInBits(A->getType()) !=
1703 SE.getTypeSizeInBits(B->getType())) {
1704 if (SE.getTypeSizeInBits(A->getType()) >
1705 SE.getTypeSizeInBits(B->getType()))
1706 B = SE.getSignExtendExpr(B, A->getType());
1708 A = SE.getSignExtendExpr(A, B->getType());
1710 if (const SCEVConstant *D =
1711 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1712 // Stride of one or negative one can have reuse with non-addresses.
1713 if (D->getValue()->isOne() ||
1714 D->getValue()->isAllOnesValue())
1715 goto decline_post_inc;
1716 // Avoid weird situations.
1717 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1718 D->getValue()->getValue().isMinSignedValue())
1719 goto decline_post_inc;
1720 // Without TLI, assume that any stride might be valid, and so any
1721 // use might be shared.
1723 goto decline_post_inc;
1724 // Check for possible scaled-address reuse.
1725 const Type *AccessTy = getAccessType(UI->getUser());
1726 TargetLowering::AddrMode AM;
1727 AM.Scale = D->getValue()->getSExtValue();
1728 if (TLI->isLegalAddressingMode(AM, AccessTy))
1729 goto decline_post_inc;
1730 AM.Scale = -AM.Scale;
1731 if (TLI->isLegalAddressingMode(AM, AccessTy))
1732 goto decline_post_inc;
1736 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1739 // It's possible for the setcc instruction to be anywhere in the loop, and
1740 // possible for it to have multiple users. If it is not immediately before
1741 // the exiting block branch, move it.
1742 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1743 if (Cond->hasOneUse()) {
1744 Cond->moveBefore(TermBr);
1746 // Clone the terminating condition and insert into the loopend.
1747 ICmpInst *OldCond = Cond;
1748 Cond = cast<ICmpInst>(Cond->clone());
1749 Cond->setName(L->getHeader()->getName() + ".termcond");
1750 ExitingBlock->getInstList().insert(TermBr, Cond);
1752 // Clone the IVUse, as the old use still exists!
1753 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1754 TermBr->replaceUsesOfWith(OldCond, Cond);
1758 // If we get to here, we know that we can transform the setcc instruction to
1759 // use the post-incremented version of the IV, allowing us to coalesce the
1760 // live ranges for the IV correctly.
1761 CondUse->transformToPostInc(L);
1764 PostIncs.insert(Cond);
1768 // Determine an insertion point for the loop induction variable increment. It
1769 // must dominate all the post-inc comparisons we just set up, and it must
1770 // dominate the loop latch edge.
1771 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1772 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1773 E = PostIncs.end(); I != E; ++I) {
1775 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1777 if (BB == (*I)->getParent())
1778 IVIncInsertPos = *I;
1779 else if (BB != IVIncInsertPos->getParent())
1780 IVIncInsertPos = BB->getTerminator();
1787 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1789 LSRUse::KindType Kind, const Type *AccessTy) {
1790 int64_t NewMinOffset = LU.MinOffset;
1791 int64_t NewMaxOffset = LU.MaxOffset;
1792 const Type *NewAccessTy = AccessTy;
1794 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1795 // something conservative, however this can pessimize in the case that one of
1796 // the uses will have all its uses outside the loop, for example.
1797 if (LU.Kind != Kind)
1799 // Conservatively assume HasBaseReg is true for now.
1800 if (NewOffset < LU.MinOffset) {
1801 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1802 Kind, AccessTy, TLI))
1804 NewMinOffset = NewOffset;
1805 } else if (NewOffset > LU.MaxOffset) {
1806 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1807 Kind, AccessTy, TLI))
1809 NewMaxOffset = NewOffset;
1811 // Check for a mismatched access type, and fall back conservatively as needed.
1812 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1813 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1816 LU.MinOffset = NewMinOffset;
1817 LU.MaxOffset = NewMaxOffset;
1818 LU.AccessTy = NewAccessTy;
1819 if (NewOffset != LU.Offsets.back())
1820 LU.Offsets.push_back(NewOffset);
1824 /// getUse - Return an LSRUse index and an offset value for a fixup which
1825 /// needs the given expression, with the given kind and optional access type.
1826 /// Either reuse an existing use or create a new one, as needed.
1827 std::pair<size_t, int64_t>
1828 LSRInstance::getUse(const SCEV *&Expr,
1829 LSRUse::KindType Kind, const Type *AccessTy) {
1830 const SCEV *Copy = Expr;
1831 int64_t Offset = ExtractImmediate(Expr, SE);
1833 // Basic uses can't accept any offset, for example.
1834 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1839 std::pair<UseMapTy::iterator, bool> P =
1840 UseMap.insert(std::make_pair(Expr, 0));
1842 // A use already existed with this base.
1843 size_t LUIdx = P.first->second;
1844 LSRUse &LU = Uses[LUIdx];
1845 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1847 return std::make_pair(LUIdx, Offset);
1850 // Create a new use.
1851 size_t LUIdx = Uses.size();
1852 P.first->second = LUIdx;
1853 Uses.push_back(LSRUse(Kind, AccessTy));
1854 LSRUse &LU = Uses[LUIdx];
1856 // We don't need to track redundant offsets, but we don't need to go out
1857 // of our way here to avoid them.
1858 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1859 LU.Offsets.push_back(Offset);
1861 LU.MinOffset = Offset;
1862 LU.MaxOffset = Offset;
1863 return std::make_pair(LUIdx, Offset);
1866 /// DeleteUse - Delete the given use from the Uses list.
1867 void LSRInstance::DeleteUse(LSRUse &LU) {
1868 if (&LU != &Uses.back())
1869 std::swap(LU, Uses.back());
1873 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1874 /// a formula that has the same registers as the given formula.
1876 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1877 const LSRUse &OrigLU) {
1878 // Search all uses for the formula. This could be more clever. Ignore
1879 // ICmpZero uses because they may contain formulae generated by
1880 // GenerateICmpZeroScales, in which case adding fixup offsets may
1882 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1883 LSRUse &LU = Uses[LUIdx];
1884 if (&LU != &OrigLU &&
1885 LU.Kind != LSRUse::ICmpZero &&
1886 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1887 LU.HasFormulaWithSameRegs(OrigF)) {
1888 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
1889 FIdx != NumForms; ++FIdx) {
1890 Formula &F = LU.Formulae[FIdx];
1891 if (F.BaseRegs == OrigF.BaseRegs &&
1892 F.ScaledReg == OrigF.ScaledReg &&
1893 F.AM.BaseGV == OrigF.AM.BaseGV &&
1894 F.AM.Scale == OrigF.AM.Scale &&
1896 if (F.AM.BaseOffs == 0)
1907 void LSRInstance::CollectInterestingTypesAndFactors() {
1908 SmallSetVector<const SCEV *, 4> Strides;
1910 // Collect interesting types and strides.
1911 SmallVector<const SCEV *, 4> Worklist;
1912 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1913 const SCEV *Expr = IU.getExpr(*UI);
1915 // Collect interesting types.
1916 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1918 // Add strides for mentioned loops.
1919 Worklist.push_back(Expr);
1921 const SCEV *S = Worklist.pop_back_val();
1922 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1923 Strides.insert(AR->getStepRecurrence(SE));
1924 Worklist.push_back(AR->getStart());
1925 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1926 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1928 } while (!Worklist.empty());
1931 // Compute interesting factors from the set of interesting strides.
1932 for (SmallSetVector<const SCEV *, 4>::const_iterator
1933 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1934 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1935 next(I); NewStrideIter != E; ++NewStrideIter) {
1936 const SCEV *OldStride = *I;
1937 const SCEV *NewStride = *NewStrideIter;
1939 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1940 SE.getTypeSizeInBits(NewStride->getType())) {
1941 if (SE.getTypeSizeInBits(OldStride->getType()) >
1942 SE.getTypeSizeInBits(NewStride->getType()))
1943 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1945 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1947 if (const SCEVConstant *Factor =
1948 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1950 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1951 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1952 } else if (const SCEVConstant *Factor =
1953 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1956 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1957 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1961 // If all uses use the same type, don't bother looking for truncation-based
1963 if (Types.size() == 1)
1966 DEBUG(print_factors_and_types(dbgs()));
1969 void LSRInstance::CollectFixupsAndInitialFormulae() {
1970 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1972 LSRFixup &LF = getNewFixup();
1973 LF.UserInst = UI->getUser();
1974 LF.OperandValToReplace = UI->getOperandValToReplace();
1975 LF.PostIncLoops = UI->getPostIncLoops();
1977 LSRUse::KindType Kind = LSRUse::Basic;
1978 const Type *AccessTy = 0;
1979 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1980 Kind = LSRUse::Address;
1981 AccessTy = getAccessType(LF.UserInst);
1984 const SCEV *S = IU.getExpr(*UI);
1986 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1987 // (N - i == 0), and this allows (N - i) to be the expression that we work
1988 // with rather than just N or i, so we can consider the register
1989 // requirements for both N and i at the same time. Limiting this code to
1990 // equality icmps is not a problem because all interesting loops use
1991 // equality icmps, thanks to IndVarSimplify.
1992 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1993 if (CI->isEquality()) {
1994 // Swap the operands if needed to put the OperandValToReplace on the
1995 // left, for consistency.
1996 Value *NV = CI->getOperand(1);
1997 if (NV == LF.OperandValToReplace) {
1998 CI->setOperand(1, CI->getOperand(0));
1999 CI->setOperand(0, NV);
2003 // x == y --> x - y == 0
2004 const SCEV *N = SE.getSCEV(NV);
2005 if (N->isLoopInvariant(L)) {
2006 Kind = LSRUse::ICmpZero;
2007 S = SE.getMinusSCEV(N, S);
2010 // -1 and the negations of all interesting strides (except the negation
2011 // of -1) are now also interesting.
2012 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2013 if (Factors[i] != -1)
2014 Factors.insert(-(uint64_t)Factors[i]);
2018 // Set up the initial formula for this use.
2019 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2021 LF.Offset = P.second;
2022 LSRUse &LU = Uses[LF.LUIdx];
2023 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2025 // If this is the first use of this LSRUse, give it a formula.
2026 if (LU.Formulae.empty()) {
2027 InsertInitialFormula(S, LU, LF.LUIdx);
2028 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2032 DEBUG(print_fixups(dbgs()));
2036 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2038 F.InitialMatch(S, L, SE, DT);
2039 bool Inserted = InsertFormula(LU, LUIdx, F);
2040 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2044 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2045 LSRUse &LU, size_t LUIdx) {
2047 F.BaseRegs.push_back(S);
2048 F.AM.HasBaseReg = true;
2049 bool Inserted = InsertFormula(LU, LUIdx, F);
2050 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2053 /// CountRegisters - Note which registers are used by the given formula,
2054 /// updating RegUses.
2055 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2057 RegUses.CountRegister(F.ScaledReg, LUIdx);
2058 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2059 E = F.BaseRegs.end(); I != E; ++I)
2060 RegUses.CountRegister(*I, LUIdx);
2063 /// InsertFormula - If the given formula has not yet been inserted, add it to
2064 /// the list, and return true. Return false otherwise.
2065 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2066 if (!LU.InsertFormula(F))
2069 CountRegisters(F, LUIdx);
2073 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2074 /// loop-invariant values which we're tracking. These other uses will pin these
2075 /// values in registers, making them less profitable for elimination.
2076 /// TODO: This currently misses non-constant addrec step registers.
2077 /// TODO: Should this give more weight to users inside the loop?
2079 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2080 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2081 SmallPtrSet<const SCEV *, 8> Inserted;
2083 while (!Worklist.empty()) {
2084 const SCEV *S = Worklist.pop_back_val();
2086 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2087 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
2088 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2089 Worklist.push_back(C->getOperand());
2090 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2091 Worklist.push_back(D->getLHS());
2092 Worklist.push_back(D->getRHS());
2093 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2094 if (!Inserted.insert(U)) continue;
2095 const Value *V = U->getValue();
2096 if (const Instruction *Inst = dyn_cast<Instruction>(V))
2097 if (L->contains(Inst)) continue;
2098 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2100 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2101 // Ignore non-instructions.
2104 // Ignore instructions in other functions (as can happen with
2106 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2108 // Ignore instructions not dominated by the loop.
2109 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2110 UserInst->getParent() :
2111 cast<PHINode>(UserInst)->getIncomingBlock(
2112 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2113 if (!DT.dominates(L->getHeader(), UseBB))
2115 // Ignore uses which are part of other SCEV expressions, to avoid
2116 // analyzing them multiple times.
2117 if (SE.isSCEVable(UserInst->getType())) {
2118 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2119 // If the user is a no-op, look through to its uses.
2120 if (!isa<SCEVUnknown>(UserS))
2124 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2128 // Ignore icmp instructions which are already being analyzed.
2129 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2130 unsigned OtherIdx = !UI.getOperandNo();
2131 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2132 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2136 LSRFixup &LF = getNewFixup();
2137 LF.UserInst = const_cast<Instruction *>(UserInst);
2138 LF.OperandValToReplace = UI.getUse();
2139 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2141 LF.Offset = P.second;
2142 LSRUse &LU = Uses[LF.LUIdx];
2143 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2144 InsertSupplementalFormula(U, LU, LF.LUIdx);
2145 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2152 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2153 /// separate registers. If C is non-null, multiply each subexpression by C.
2154 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2155 SmallVectorImpl<const SCEV *> &Ops,
2156 ScalarEvolution &SE) {
2157 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2158 // Break out add operands.
2159 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2161 CollectSubexprs(*I, C, Ops, SE);
2163 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2164 // Split a non-zero base out of an addrec.
2165 if (!AR->getStart()->isZero()) {
2166 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2167 AR->getStepRecurrence(SE),
2168 AR->getLoop()), C, Ops, SE);
2169 CollectSubexprs(AR->getStart(), C, Ops, SE);
2172 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2173 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2174 if (Mul->getNumOperands() == 2)
2175 if (const SCEVConstant *Op0 =
2176 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2177 CollectSubexprs(Mul->getOperand(1),
2178 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2184 // Otherwise use the value itself.
2185 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2188 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2190 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2193 // Arbitrarily cap recursion to protect compile time.
2194 if (Depth >= 3) return;
2196 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2197 const SCEV *BaseReg = Base.BaseRegs[i];
2199 SmallVector<const SCEV *, 8> AddOps;
2200 CollectSubexprs(BaseReg, 0, AddOps, SE);
2201 if (AddOps.size() == 1) continue;
2203 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2204 JE = AddOps.end(); J != JE; ++J) {
2205 // Don't pull a constant into a register if the constant could be folded
2206 // into an immediate field.
2207 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2208 Base.getNumRegs() > 1,
2209 LU.Kind, LU.AccessTy, TLI, SE))
2212 // Collect all operands except *J.
2213 SmallVector<const SCEV *, 8> InnerAddOps;
2214 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2215 KE = AddOps.end(); K != KE; ++K)
2217 InnerAddOps.push_back(*K);
2219 // Don't leave just a constant behind in a register if the constant could
2220 // be folded into an immediate field.
2221 if (InnerAddOps.size() == 1 &&
2222 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2223 Base.getNumRegs() > 1,
2224 LU.Kind, LU.AccessTy, TLI, SE))
2227 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2228 if (InnerSum->isZero())
2231 F.BaseRegs[i] = InnerSum;
2232 F.BaseRegs.push_back(*J);
2233 if (InsertFormula(LU, LUIdx, F))
2234 // If that formula hadn't been seen before, recurse to find more like
2236 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2241 /// GenerateCombinations - Generate a formula consisting of all of the
2242 /// loop-dominating registers added into a single register.
2243 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2245 // This method is only interesting on a plurality of registers.
2246 if (Base.BaseRegs.size() <= 1) return;
2250 SmallVector<const SCEV *, 4> Ops;
2251 for (SmallVectorImpl<const SCEV *>::const_iterator
2252 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2253 const SCEV *BaseReg = *I;
2254 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2255 !BaseReg->hasComputableLoopEvolution(L))
2256 Ops.push_back(BaseReg);
2258 F.BaseRegs.push_back(BaseReg);
2260 if (Ops.size() > 1) {
2261 const SCEV *Sum = SE.getAddExpr(Ops);
2262 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2263 // opportunity to fold something. For now, just ignore such cases
2264 // rather than proceed with zero in a register.
2265 if (!Sum->isZero()) {
2266 F.BaseRegs.push_back(Sum);
2267 (void)InsertFormula(LU, LUIdx, F);
2272 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2273 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2275 // We can't add a symbolic offset if the address already contains one.
2276 if (Base.AM.BaseGV) return;
2278 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2279 const SCEV *G = Base.BaseRegs[i];
2280 GlobalValue *GV = ExtractSymbol(G, SE);
2281 if (G->isZero() || !GV)
2285 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2286 LU.Kind, LU.AccessTy, TLI))
2289 (void)InsertFormula(LU, LUIdx, F);
2293 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2294 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2296 // TODO: For now, just add the min and max offset, because it usually isn't
2297 // worthwhile looking at everything inbetween.
2298 SmallVector<int64_t, 4> Worklist;
2299 Worklist.push_back(LU.MinOffset);
2300 if (LU.MaxOffset != LU.MinOffset)
2301 Worklist.push_back(LU.MaxOffset);
2303 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2304 const SCEV *G = Base.BaseRegs[i];
2306 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2307 E = Worklist.end(); I != E; ++I) {
2309 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2310 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2311 LU.Kind, LU.AccessTy, TLI)) {
2312 F.BaseRegs[i] = SE.getAddExpr(G, SE.getConstant(G->getType(), *I));
2314 (void)InsertFormula(LU, LUIdx, F);
2318 int64_t Imm = ExtractImmediate(G, SE);
2319 if (G->isZero() || Imm == 0)
2322 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2323 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2324 LU.Kind, LU.AccessTy, TLI))
2327 (void)InsertFormula(LU, LUIdx, F);
2331 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2332 /// the comparison. For example, x == y -> x*c == y*c.
2333 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2335 if (LU.Kind != LSRUse::ICmpZero) return;
2337 // Determine the integer type for the base formula.
2338 const Type *IntTy = Base.getType();
2340 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2342 // Don't do this if there is more than one offset.
2343 if (LU.MinOffset != LU.MaxOffset) return;
2345 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2347 // Check each interesting stride.
2348 for (SmallSetVector<int64_t, 8>::const_iterator
2349 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2350 int64_t Factor = *I;
2353 // Check that the multiplication doesn't overflow.
2354 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2356 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2357 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2360 // Check that multiplying with the use offset doesn't overflow.
2361 int64_t Offset = LU.MinOffset;
2362 if (Offset == INT64_MIN && Factor == -1)
2364 Offset = (uint64_t)Offset * Factor;
2365 if (Offset / Factor != LU.MinOffset)
2368 // Check that this scale is legal.
2369 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2372 // Compensate for the use having MinOffset built into it.
2373 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2375 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2377 // Check that multiplying with each base register doesn't overflow.
2378 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2379 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2380 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2384 // Check that multiplying with the scaled register doesn't overflow.
2386 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2387 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2391 // If we make it here and it's legal, add it.
2392 (void)InsertFormula(LU, LUIdx, F);
2397 /// GenerateScales - Generate stride factor reuse formulae by making use of
2398 /// scaled-offset address modes, for example.
2399 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
2401 // Determine the integer type for the base formula.
2402 const Type *IntTy = Base.getType();
2405 // If this Formula already has a scaled register, we can't add another one.
2406 if (Base.AM.Scale != 0) return;
2408 // Check each interesting stride.
2409 for (SmallSetVector<int64_t, 8>::const_iterator
2410 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2411 int64_t Factor = *I;
2413 Base.AM.Scale = Factor;
2414 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2415 // Check whether this scale is going to be legal.
2416 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2417 LU.Kind, LU.AccessTy, TLI)) {
2418 // As a special-case, handle special out-of-loop Basic users specially.
2419 // TODO: Reconsider this special case.
2420 if (LU.Kind == LSRUse::Basic &&
2421 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2422 LSRUse::Special, LU.AccessTy, TLI) &&
2423 LU.AllFixupsOutsideLoop)
2424 LU.Kind = LSRUse::Special;
2428 // For an ICmpZero, negating a solitary base register won't lead to
2430 if (LU.Kind == LSRUse::ICmpZero &&
2431 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2433 // For each addrec base reg, apply the scale, if possible.
2434 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2435 if (const SCEVAddRecExpr *AR =
2436 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2437 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2438 if (FactorS->isZero())
2440 // Divide out the factor, ignoring high bits, since we'll be
2441 // scaling the value back up in the end.
2442 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2443 // TODO: This could be optimized to avoid all the copying.
2445 F.ScaledReg = Quotient;
2446 F.DeleteBaseReg(F.BaseRegs[i]);
2447 (void)InsertFormula(LU, LUIdx, F);
2453 /// GenerateTruncates - Generate reuse formulae from different IV types.
2454 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
2456 // This requires TargetLowering to tell us which truncates are free.
2459 // Don't bother truncating symbolic values.
2460 if (Base.AM.BaseGV) return;
2462 // Determine the integer type for the base formula.
2463 const Type *DstTy = Base.getType();
2465 DstTy = SE.getEffectiveSCEVType(DstTy);
2467 for (SmallSetVector<const Type *, 4>::const_iterator
2468 I = Types.begin(), E = Types.end(); I != E; ++I) {
2469 const Type *SrcTy = *I;
2470 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2473 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2474 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2475 JE = F.BaseRegs.end(); J != JE; ++J)
2476 *J = SE.getAnyExtendExpr(*J, SrcTy);
2478 // TODO: This assumes we've done basic processing on all uses and
2479 // have an idea what the register usage is.
2480 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2483 (void)InsertFormula(LU, LUIdx, F);
2490 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2491 /// defer modifications so that the search phase doesn't have to worry about
2492 /// the data structures moving underneath it.
2496 const SCEV *OrigReg;
2498 WorkItem(size_t LI, int64_t I, const SCEV *R)
2499 : LUIdx(LI), Imm(I), OrigReg(R) {}
2501 void print(raw_ostream &OS) const;
2507 void WorkItem::print(raw_ostream &OS) const {
2508 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2509 << " , add offset " << Imm;
2512 void WorkItem::dump() const {
2513 print(errs()); errs() << '\n';
2516 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2517 /// distance apart and try to form reuse opportunities between them.
2518 void LSRInstance::GenerateCrossUseConstantOffsets() {
2519 // Group the registers by their value without any added constant offset.
2520 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2521 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2523 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2524 SmallVector<const SCEV *, 8> Sequence;
2525 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2527 const SCEV *Reg = *I;
2528 int64_t Imm = ExtractImmediate(Reg, SE);
2529 std::pair<RegMapTy::iterator, bool> Pair =
2530 Map.insert(std::make_pair(Reg, ImmMapTy()));
2532 Sequence.push_back(Reg);
2533 Pair.first->second.insert(std::make_pair(Imm, *I));
2534 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2537 // Now examine each set of registers with the same base value. Build up
2538 // a list of work to do and do the work in a separate step so that we're
2539 // not adding formulae and register counts while we're searching.
2540 SmallVector<WorkItem, 32> WorkItems;
2541 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2542 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2543 E = Sequence.end(); I != E; ++I) {
2544 const SCEV *Reg = *I;
2545 const ImmMapTy &Imms = Map.find(Reg)->second;
2547 // It's not worthwhile looking for reuse if there's only one offset.
2548 if (Imms.size() == 1)
2551 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2552 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2554 dbgs() << ' ' << J->first;
2557 // Examine each offset.
2558 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2560 const SCEV *OrigReg = J->second;
2562 int64_t JImm = J->first;
2563 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2565 if (!isa<SCEVConstant>(OrigReg) &&
2566 UsedByIndicesMap[Reg].count() == 1) {
2567 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2571 // Conservatively examine offsets between this orig reg a few selected
2573 ImmMapTy::const_iterator OtherImms[] = {
2574 Imms.begin(), prior(Imms.end()),
2575 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2577 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2578 ImmMapTy::const_iterator M = OtherImms[i];
2579 if (M == J || M == JE) continue;
2581 // Compute the difference between the two.
2582 int64_t Imm = (uint64_t)JImm - M->first;
2583 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2584 LUIdx = UsedByIndices.find_next(LUIdx))
2585 // Make a memo of this use, offset, and register tuple.
2586 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2587 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2594 UsedByIndicesMap.clear();
2595 UniqueItems.clear();
2597 // Now iterate through the worklist and add new formulae.
2598 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2599 E = WorkItems.end(); I != E; ++I) {
2600 const WorkItem &WI = *I;
2601 size_t LUIdx = WI.LUIdx;
2602 LSRUse &LU = Uses[LUIdx];
2603 int64_t Imm = WI.Imm;
2604 const SCEV *OrigReg = WI.OrigReg;
2606 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2607 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2608 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2610 // TODO: Use a more targeted data structure.
2611 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2612 Formula F = LU.Formulae[L];
2613 // Use the immediate in the scaled register.
2614 if (F.ScaledReg == OrigReg) {
2615 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2616 Imm * (uint64_t)F.AM.Scale;
2617 // Don't create 50 + reg(-50).
2618 if (F.referencesReg(SE.getSCEV(
2619 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2622 NewF.AM.BaseOffs = Offs;
2623 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2624 LU.Kind, LU.AccessTy, TLI))
2626 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2628 // If the new scale is a constant in a register, and adding the constant
2629 // value to the immediate would produce a value closer to zero than the
2630 // immediate itself, then the formula isn't worthwhile.
2631 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2632 if (C->getValue()->getValue().isNegative() !=
2633 (NewF.AM.BaseOffs < 0) &&
2634 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2635 .ule(abs64(NewF.AM.BaseOffs)))
2639 (void)InsertFormula(LU, LUIdx, NewF);
2641 // Use the immediate in a base register.
2642 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2643 const SCEV *BaseReg = F.BaseRegs[N];
2644 if (BaseReg != OrigReg)
2647 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2648 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2649 LU.Kind, LU.AccessTy, TLI))
2651 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2653 // If the new formula has a constant in a register, and adding the
2654 // constant value to the immediate would produce a value closer to
2655 // zero than the immediate itself, then the formula isn't worthwhile.
2656 for (SmallVectorImpl<const SCEV *>::const_iterator
2657 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2659 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2660 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2661 abs64(NewF.AM.BaseOffs)) &&
2662 (C->getValue()->getValue() +
2663 NewF.AM.BaseOffs).countTrailingZeros() >=
2664 CountTrailingZeros_64(NewF.AM.BaseOffs))
2668 (void)InsertFormula(LU, LUIdx, NewF);
2677 /// GenerateAllReuseFormulae - Generate formulae for each use.
2679 LSRInstance::GenerateAllReuseFormulae() {
2680 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2681 // queries are more precise.
2682 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2683 LSRUse &LU = Uses[LUIdx];
2684 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2685 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2686 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2687 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2689 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2690 LSRUse &LU = Uses[LUIdx];
2691 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2692 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2693 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2694 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2695 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2696 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2697 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2698 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2700 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2701 LSRUse &LU = Uses[LUIdx];
2702 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2703 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2706 GenerateCrossUseConstantOffsets();
2709 /// If their are multiple formulae with the same set of registers used
2710 /// by other uses, pick the best one and delete the others.
2711 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2713 bool Changed = false;
2716 // Collect the best formula for each unique set of shared registers. This
2717 // is reset for each use.
2718 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2720 BestFormulaeTy BestFormulae;
2722 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2723 LSRUse &LU = Uses[LUIdx];
2724 FormulaSorter Sorter(L, LU, SE, DT);
2725 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << "\n");
2728 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2729 FIdx != NumForms; ++FIdx) {
2730 Formula &F = LU.Formulae[FIdx];
2732 SmallVector<const SCEV *, 2> Key;
2733 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2734 JE = F.BaseRegs.end(); J != JE; ++J) {
2735 const SCEV *Reg = *J;
2736 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2740 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2741 Key.push_back(F.ScaledReg);
2742 // Unstable sort by host order ok, because this is only used for
2744 std::sort(Key.begin(), Key.end());
2746 std::pair<BestFormulaeTy::const_iterator, bool> P =
2747 BestFormulae.insert(std::make_pair(Key, FIdx));
2749 Formula &Best = LU.Formulae[P.first->second];
2750 if (Sorter.operator()(F, Best))
2752 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2754 " in favor of formula "; Best.print(dbgs());
2759 LU.DeleteFormula(F);
2767 // Now that we've filtered out some formulae, recompute the Regs set.
2769 LU.RecomputeRegs(LUIdx, RegUses);
2771 // Reset this to prepare for the next use.
2772 BestFormulae.clear();
2775 DEBUG(if (Changed) {
2777 "After filtering out undesirable candidates:\n";
2782 // This is a rough guess that seems to work fairly well.
2783 static const size_t ComplexityLimit = UINT16_MAX;
2785 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2786 /// solutions the solver might have to consider. It almost never considers
2787 /// this many solutions because it prune the search space, but the pruning
2788 /// isn't always sufficient.
2789 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2791 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2792 E = Uses.end(); I != E; ++I) {
2793 size_t FSize = I->Formulae.size();
2794 if (FSize >= ComplexityLimit) {
2795 Power = ComplexityLimit;
2799 if (Power >= ComplexityLimit)
2805 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2806 /// formulae to choose from, use some rough heuristics to prune down the number
2807 /// of formulae. This keeps the main solver from taking an extraordinary amount
2808 /// of time in some worst-case scenarios.
2809 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2810 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2811 DEBUG(dbgs() << "The search space is too complex.\n");
2813 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2814 "which use a superset of registers used by other "
2817 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2818 LSRUse &LU = Uses[LUIdx];
2820 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2821 Formula &F = LU.Formulae[i];
2822 for (SmallVectorImpl<const SCEV *>::const_iterator
2823 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2824 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2826 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2827 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2828 (I - F.BaseRegs.begin()));
2829 if (LU.HasFormulaWithSameRegs(NewF)) {
2830 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2831 LU.DeleteFormula(F);
2837 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2838 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2841 NewF.AM.BaseGV = GV;
2842 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2843 (I - F.BaseRegs.begin()));
2844 if (LU.HasFormulaWithSameRegs(NewF)) {
2845 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2847 LU.DeleteFormula(F);
2858 LU.RecomputeRegs(LUIdx, RegUses);
2861 DEBUG(dbgs() << "After pre-selection:\n";
2862 print_uses(dbgs()));
2865 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2866 DEBUG(dbgs() << "The search space is too complex.\n");
2868 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2869 "separated by a constant offset will use the same "
2872 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2873 LSRUse &LU = Uses[LUIdx];
2874 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2875 FIdx != NumForms; ++FIdx) {
2876 Formula &F = LU.Formulae[FIdx];
2877 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2878 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2879 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2880 /*HasBaseReg=*/false,
2881 LU.Kind, LU.AccessTy)) {
2882 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2885 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2887 // Delete formulae from the new use which are no longer legal.
2889 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2890 Formula &F = LUThatHas->Formulae[i];
2891 if (!isLegalUse(F.AM,
2892 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2893 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2894 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2896 LUThatHas->DeleteFormula(F);
2903 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
2905 // Update the relocs to reference the new use.
2906 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
2907 if (Fixups[i].LUIdx == LUIdx) {
2908 Fixups[i].LUIdx = LUThatHas - &Uses.front();
2909 Fixups[i].Offset += F.AM.BaseOffs;
2910 DEBUG(errs() << "New fixup has offset "
2911 << Fixups[i].Offset << "\n");
2913 if (Fixups[i].LUIdx == NumUses-1)
2914 Fixups[i].LUIdx = LUIdx;
2917 // Delete the old use.
2928 DEBUG(dbgs() << "After pre-selection:\n";
2929 print_uses(dbgs()));
2932 SmallPtrSet<const SCEV *, 4> Taken;
2933 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2934 // Ok, we have too many of formulae on our hands to conveniently handle.
2935 // Use a rough heuristic to thin out the list.
2936 DEBUG(dbgs() << "The search space is too complex.\n");
2938 // Pick the register which is used by the most LSRUses, which is likely
2939 // to be a good reuse register candidate.
2940 const SCEV *Best = 0;
2941 unsigned BestNum = 0;
2942 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2944 const SCEV *Reg = *I;
2945 if (Taken.count(Reg))
2950 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2951 if (Count > BestNum) {
2958 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2959 << " will yield profitable reuse.\n");
2962 // In any use with formulae which references this register, delete formulae
2963 // which don't reference it.
2964 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2965 LSRUse &LU = Uses[LUIdx];
2966 if (!LU.Regs.count(Best)) continue;
2969 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2970 Formula &F = LU.Formulae[i];
2971 if (!F.referencesReg(Best)) {
2972 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2973 LU.DeleteFormula(F);
2977 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
2983 LU.RecomputeRegs(LUIdx, RegUses);
2986 DEBUG(dbgs() << "After pre-selection:\n";
2987 print_uses(dbgs()));
2991 /// SolveRecurse - This is the recursive solver.
2992 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2994 SmallVectorImpl<const Formula *> &Workspace,
2995 const Cost &CurCost,
2996 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2997 DenseSet<const SCEV *> &VisitedRegs) const {
3000 // - use more aggressive filtering
3001 // - sort the formula so that the most profitable solutions are found first
3002 // - sort the uses too
3004 // - don't compute a cost, and then compare. compare while computing a cost
3006 // - track register sets with SmallBitVector
3008 const LSRUse &LU = Uses[Workspace.size()];
3010 // If this use references any register that's already a part of the
3011 // in-progress solution, consider it a requirement that a formula must
3012 // reference that register in order to be considered. This prunes out
3013 // unprofitable searching.
3014 SmallSetVector<const SCEV *, 4> ReqRegs;
3015 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3016 E = CurRegs.end(); I != E; ++I)
3017 if (LU.Regs.count(*I))
3020 bool AnySatisfiedReqRegs = false;
3021 SmallPtrSet<const SCEV *, 16> NewRegs;
3024 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3025 E = LU.Formulae.end(); I != E; ++I) {
3026 const Formula &F = *I;
3028 // Ignore formulae which do not use any of the required registers.
3029 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3030 JE = ReqRegs.end(); J != JE; ++J) {
3031 const SCEV *Reg = *J;
3032 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3033 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3037 AnySatisfiedReqRegs = true;
3039 // Evaluate the cost of the current formula. If it's already worse than
3040 // the current best, prune the search at that point.
3043 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3044 if (NewCost < SolutionCost) {
3045 Workspace.push_back(&F);
3046 if (Workspace.size() != Uses.size()) {
3047 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3048 NewRegs, VisitedRegs);
3049 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3050 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3052 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3053 dbgs() << ". Regs:";
3054 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3055 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3056 dbgs() << ' ' << **I;
3059 SolutionCost = NewCost;
3060 Solution = Workspace;
3062 Workspace.pop_back();
3067 // If none of the formulae had all of the required registers, relax the
3068 // constraint so that we don't exclude all formulae.
3069 if (!AnySatisfiedReqRegs) {
3070 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3076 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3077 SmallVector<const Formula *, 8> Workspace;
3079 SolutionCost.Loose();
3081 SmallPtrSet<const SCEV *, 16> CurRegs;
3082 DenseSet<const SCEV *> VisitedRegs;
3083 Workspace.reserve(Uses.size());
3085 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3086 CurRegs, VisitedRegs);
3088 // Ok, we've now made all our decisions.
3089 DEBUG(dbgs() << "\n"
3090 "The chosen solution requires "; SolutionCost.print(dbgs());
3092 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3094 Uses[i].print(dbgs());
3097 Solution[i]->print(dbgs());
3102 /// getImmediateDominator - A handy utility for the specific DominatorTree
3103 /// query that we need here.
3105 static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
3106 DomTreeNode *Node = DT.getNode(BB);
3107 if (!Node) return 0;
3108 Node = Node->getIDom();
3109 if (!Node) return 0;
3110 return Node->getBlock();
3113 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3114 /// the dominator tree far as we can go while still being dominated by the
3115 /// input positions. This helps canonicalize the insert position, which
3116 /// encourages sharing.
3117 BasicBlock::iterator
3118 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3119 const SmallVectorImpl<Instruction *> &Inputs)
3122 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3123 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3126 for (BasicBlock *Rung = IP->getParent(); ; Rung = IDom) {
3127 IDom = getImmediateDominator(Rung, DT);
3128 if (!IDom) return IP;
3130 // Don't climb into a loop though.
3131 const Loop *IDomLoop = LI.getLoopFor(IDom);
3132 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3133 if (IDomDepth <= IPLoopDepth &&
3134 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3138 bool AllDominate = true;
3139 Instruction *BetterPos = 0;
3140 Instruction *Tentative = IDom->getTerminator();
3141 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3142 E = Inputs.end(); I != E; ++I) {
3143 Instruction *Inst = *I;
3144 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3145 AllDominate = false;
3148 // Attempt to find an insert position in the middle of the block,
3149 // instead of at the end, so that it can be used for other expansions.
3150 if (IDom == Inst->getParent() &&
3151 (!BetterPos || DT.dominates(BetterPos, Inst)))
3152 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3165 /// AdjustInsertPositionForExpand - Determine an input position which will be
3166 /// dominated by the operands and which will dominate the result.
3167 BasicBlock::iterator
3168 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3170 const LSRUse &LU) const {
3171 // Collect some instructions which must be dominated by the
3172 // expanding replacement. These must be dominated by any operands that
3173 // will be required in the expansion.
3174 SmallVector<Instruction *, 4> Inputs;
3175 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3176 Inputs.push_back(I);
3177 if (LU.Kind == LSRUse::ICmpZero)
3178 if (Instruction *I =
3179 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3180 Inputs.push_back(I);
3181 if (LF.PostIncLoops.count(L)) {
3182 if (LF.isUseFullyOutsideLoop(L))
3183 Inputs.push_back(L->getLoopLatch()->getTerminator());
3185 Inputs.push_back(IVIncInsertPos);
3187 // The expansion must also be dominated by the increment positions of any
3188 // loops it for which it is using post-inc mode.
3189 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3190 E = LF.PostIncLoops.end(); I != E; ++I) {
3191 const Loop *PIL = *I;
3192 if (PIL == L) continue;
3194 // Be dominated by the loop exit.
3195 SmallVector<BasicBlock *, 4> ExitingBlocks;
3196 PIL->getExitingBlocks(ExitingBlocks);
3197 if (!ExitingBlocks.empty()) {
3198 BasicBlock *BB = ExitingBlocks[0];
3199 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3200 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3201 Inputs.push_back(BB->getTerminator());
3205 // Then, climb up the immediate dominator tree as far as we can go while
3206 // still being dominated by the input positions.
3207 IP = HoistInsertPosition(IP, Inputs);
3209 // Don't insert instructions before PHI nodes.
3210 while (isa<PHINode>(IP)) ++IP;
3212 // Ignore debug intrinsics.
3213 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3218 Value *LSRInstance::Expand(const LSRFixup &LF,
3220 BasicBlock::iterator IP,
3221 SCEVExpander &Rewriter,
3222 SmallVectorImpl<WeakVH> &DeadInsts) const {
3223 const LSRUse &LU = Uses[LF.LUIdx];
3225 // Determine an input position which will be dominated by the operands and
3226 // which will dominate the result.
3227 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3229 // Inform the Rewriter if we have a post-increment use, so that it can
3230 // perform an advantageous expansion.
3231 Rewriter.setPostInc(LF.PostIncLoops);
3233 // This is the type that the user actually needs.
3234 const Type *OpTy = LF.OperandValToReplace->getType();
3235 // This will be the type that we'll initially expand to.
3236 const Type *Ty = F.getType();
3238 // No type known; just expand directly to the ultimate type.
3240 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3241 // Expand directly to the ultimate type if it's the right size.
3243 // This is the type to do integer arithmetic in.
3244 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3246 // Build up a list of operands to add together to form the full base.
3247 SmallVector<const SCEV *, 8> Ops;
3249 // Expand the BaseRegs portion.
3250 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3251 E = F.BaseRegs.end(); I != E; ++I) {
3252 const SCEV *Reg = *I;
3253 assert(!Reg->isZero() && "Zero allocated in a base register!");
3255 // If we're expanding for a post-inc user, make the post-inc adjustment.
3256 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3257 Reg = TransformForPostIncUse(Denormalize, Reg,
3258 LF.UserInst, LF.OperandValToReplace,
3261 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3264 // Flush the operand list to suppress SCEVExpander hoisting.
3266 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3268 Ops.push_back(SE.getUnknown(FullV));
3271 // Expand the ScaledReg portion.
3272 Value *ICmpScaledV = 0;
3273 if (F.AM.Scale != 0) {
3274 const SCEV *ScaledS = F.ScaledReg;
3276 // If we're expanding for a post-inc user, make the post-inc adjustment.
3277 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3278 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3279 LF.UserInst, LF.OperandValToReplace,
3282 if (LU.Kind == LSRUse::ICmpZero) {
3283 // An interesting way of "folding" with an icmp is to use a negated
3284 // scale, which we'll implement by inserting it into the other operand
3286 assert(F.AM.Scale == -1 &&
3287 "The only scale supported by ICmpZero uses is -1!");
3288 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3290 // Otherwise just expand the scaled register and an explicit scale,
3291 // which is expected to be matched as part of the address.
3292 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3293 ScaledS = SE.getMulExpr(ScaledS,
3294 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3295 Ops.push_back(ScaledS);
3297 // Flush the operand list to suppress SCEVExpander hoisting.
3298 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3300 Ops.push_back(SE.getUnknown(FullV));
3304 // Expand the GV portion.
3306 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3308 // Flush the operand list to suppress SCEVExpander hoisting.
3309 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3311 Ops.push_back(SE.getUnknown(FullV));
3314 // Expand the immediate portion.
3315 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3317 if (LU.Kind == LSRUse::ICmpZero) {
3318 // The other interesting way of "folding" with an ICmpZero is to use a
3319 // negated immediate.
3321 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3323 Ops.push_back(SE.getUnknown(ICmpScaledV));
3324 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3327 // Just add the immediate values. These again are expected to be matched
3328 // as part of the address.
3329 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3333 // Emit instructions summing all the operands.
3334 const SCEV *FullS = Ops.empty() ?
3335 SE.getConstant(IntTy, 0) :
3337 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3339 // We're done expanding now, so reset the rewriter.
3340 Rewriter.clearPostInc();
3342 // An ICmpZero Formula represents an ICmp which we're handling as a
3343 // comparison against zero. Now that we've expanded an expression for that
3344 // form, update the ICmp's other operand.
3345 if (LU.Kind == LSRUse::ICmpZero) {
3346 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3347 DeadInsts.push_back(CI->getOperand(1));
3348 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3349 "a scale at the same time!");
3350 if (F.AM.Scale == -1) {
3351 if (ICmpScaledV->getType() != OpTy) {
3353 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3355 ICmpScaledV, OpTy, "tmp", CI);
3358 CI->setOperand(1, ICmpScaledV);
3360 assert(F.AM.Scale == 0 &&
3361 "ICmp does not support folding a global value and "
3362 "a scale at the same time!");
3363 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3365 if (C->getType() != OpTy)
3366 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3370 CI->setOperand(1, C);
3377 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3378 /// of their operands effectively happens in their predecessor blocks, so the
3379 /// expression may need to be expanded in multiple places.
3380 void LSRInstance::RewriteForPHI(PHINode *PN,
3383 SCEVExpander &Rewriter,
3384 SmallVectorImpl<WeakVH> &DeadInsts,
3386 DenseMap<BasicBlock *, Value *> Inserted;
3387 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3388 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3389 BasicBlock *BB = PN->getIncomingBlock(i);
3391 // If this is a critical edge, split the edge so that we do not insert
3392 // the code on all predecessor/successor paths. We do this unless this
3393 // is the canonical backedge for this loop, which complicates post-inc
3395 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3396 !isa<IndirectBrInst>(BB->getTerminator()) &&
3397 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3398 // Split the critical edge.
3399 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3401 // If PN is outside of the loop and BB is in the loop, we want to
3402 // move the block to be immediately before the PHI block, not
3403 // immediately after BB.
3404 if (L->contains(BB) && !L->contains(PN))
3405 NewBB->moveBefore(PN->getParent());
3407 // Splitting the edge can reduce the number of PHI entries we have.
3408 e = PN->getNumIncomingValues();
3410 i = PN->getBasicBlockIndex(BB);
3413 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3414 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3416 PN->setIncomingValue(i, Pair.first->second);
3418 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3420 // If this is reuse-by-noop-cast, insert the noop cast.
3421 const Type *OpTy = LF.OperandValToReplace->getType();
3422 if (FullV->getType() != OpTy)
3424 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3426 FullV, LF.OperandValToReplace->getType(),
3427 "tmp", BB->getTerminator());
3429 PN->setIncomingValue(i, FullV);
3430 Pair.first->second = FullV;
3435 /// Rewrite - Emit instructions for the leading candidate expression for this
3436 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3437 /// the newly expanded value.
3438 void LSRInstance::Rewrite(const LSRFixup &LF,
3440 SCEVExpander &Rewriter,
3441 SmallVectorImpl<WeakVH> &DeadInsts,
3443 // First, find an insertion point that dominates UserInst. For PHI nodes,
3444 // find the nearest block which dominates all the relevant uses.
3445 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3446 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3448 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3450 // If this is reuse-by-noop-cast, insert the noop cast.
3451 const Type *OpTy = LF.OperandValToReplace->getType();
3452 if (FullV->getType() != OpTy) {
3454 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3455 FullV, OpTy, "tmp", LF.UserInst);
3459 // Update the user. ICmpZero is handled specially here (for now) because
3460 // Expand may have updated one of the operands of the icmp already, and
3461 // its new value may happen to be equal to LF.OperandValToReplace, in
3462 // which case doing replaceUsesOfWith leads to replacing both operands
3463 // with the same value. TODO: Reorganize this.
3464 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3465 LF.UserInst->setOperand(0, FullV);
3467 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3470 DeadInsts.push_back(LF.OperandValToReplace);
3474 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3476 // Keep track of instructions we may have made dead, so that
3477 // we can remove them after we are done working.
3478 SmallVector<WeakVH, 16> DeadInsts;
3480 SCEVExpander Rewriter(SE);
3481 Rewriter.disableCanonicalMode();
3482 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3484 // Expand the new value definitions and update the users.
3485 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3486 size_t LUIdx = Fixups[i].LUIdx;
3488 Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
3493 // Clean up after ourselves. This must be done before deleting any
3497 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3500 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3501 : IU(P->getAnalysis<IVUsers>()),
3502 SE(P->getAnalysis<ScalarEvolution>()),
3503 DT(P->getAnalysis<DominatorTree>()),
3504 LI(P->getAnalysis<LoopInfo>()),
3505 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3507 // If LoopSimplify form is not available, stay out of trouble.
3508 if (!L->isLoopSimplifyForm()) return;
3510 // If there's no interesting work to be done, bail early.
3511 if (IU.empty()) return;
3513 DEBUG(dbgs() << "\nLSR on loop ";
3514 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3517 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3518 /// inside the loop then try to eliminate the cast operation.
3521 // Change loop terminating condition to use the postinc iv when possible.
3522 Changed |= OptimizeLoopTermCond();
3524 CollectInterestingTypesAndFactors();
3525 CollectFixupsAndInitialFormulae();
3526 CollectLoopInvariantFixupsAndFormulae();
3528 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3529 print_uses(dbgs()));
3531 // Now use the reuse data to generate a bunch of interesting ways
3532 // to formulate the values needed for the uses.
3533 GenerateAllReuseFormulae();
3535 DEBUG(dbgs() << "\n"
3536 "After generating reuse formulae:\n";
3537 print_uses(dbgs()));
3539 FilterOutUndesirableDedicatedRegisters();
3540 NarrowSearchSpaceUsingHeuristics();
3542 SmallVector<const Formula *, 8> Solution;
3544 assert(Solution.size() == Uses.size() && "Malformed solution!");
3546 // Release memory that is no longer needed.
3552 // Formulae should be legal.
3553 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3554 E = Uses.end(); I != E; ++I) {
3555 const LSRUse &LU = *I;
3556 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3557 JE = LU.Formulae.end(); J != JE; ++J)
3558 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3559 LU.Kind, LU.AccessTy, TLI) &&
3560 "Illegal formula generated!");
3564 // Now that we've decided what we want, make it so.
3565 ImplementSolution(Solution, P);
3568 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3569 if (Factors.empty() && Types.empty()) return;
3571 OS << "LSR has identified the following interesting factors and types: ";
3574 for (SmallSetVector<int64_t, 8>::const_iterator
3575 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3576 if (!First) OS << ", ";
3581 for (SmallSetVector<const Type *, 4>::const_iterator
3582 I = Types.begin(), E = Types.end(); I != E; ++I) {
3583 if (!First) OS << ", ";
3585 OS << '(' << **I << ')';
3590 void LSRInstance::print_fixups(raw_ostream &OS) const {
3591 OS << "LSR is examining the following fixup sites:\n";
3592 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3593 E = Fixups.end(); I != E; ++I) {
3594 const LSRFixup &LF = *I;
3601 void LSRInstance::print_uses(raw_ostream &OS) const {
3602 OS << "LSR is examining the following uses:\n";
3603 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3604 E = Uses.end(); I != E; ++I) {
3605 const LSRUse &LU = *I;
3609 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3610 JE = LU.Formulae.end(); J != JE; ++J) {
3618 void LSRInstance::print(raw_ostream &OS) const {
3619 print_factors_and_types(OS);
3624 void LSRInstance::dump() const {
3625 print(errs()); errs() << '\n';
3630 class LoopStrengthReduce : public LoopPass {
3631 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3632 /// transformation profitability.
3633 const TargetLowering *const TLI;
3636 static char ID; // Pass ID, replacement for typeid
3637 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3640 bool runOnLoop(Loop *L, LPPassManager &LPM);
3641 void getAnalysisUsage(AnalysisUsage &AU) const;
3646 char LoopStrengthReduce::ID = 0;
3647 static RegisterPass<LoopStrengthReduce>
3648 X("loop-reduce", "Loop Strength Reduction");
3650 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3651 return new LoopStrengthReduce(TLI);
3654 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3655 : LoopPass(&ID), TLI(tli) {}
3657 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3658 // We split critical edges, so we change the CFG. However, we do update
3659 // many analyses if they are around.
3660 AU.addPreservedID(LoopSimplifyID);
3661 AU.addPreserved("domfrontier");
3663 AU.addRequired<LoopInfo>();
3664 AU.addPreserved<LoopInfo>();
3665 AU.addRequiredID(LoopSimplifyID);
3666 AU.addRequired<DominatorTree>();
3667 AU.addPreserved<DominatorTree>();
3668 AU.addRequired<ScalarEvolution>();
3669 AU.addPreserved<ScalarEvolution>();
3670 AU.addRequired<IVUsers>();
3671 AU.addPreserved<IVUsers>();
3674 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3675 bool Changed = false;
3677 // Run the main LSR transformation.
3678 Changed |= LSRInstance(TLI, L, this).getChanged();
3680 // At this point, it is worth checking to see if any recurrence PHIs are also
3681 // dead, so that we can remove them as well.
3682 Changed |= DeleteDeadPHIs(L->getHeader());