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 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
165 if (I == RegUsesMap.end())
167 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
168 int i = UsedByIndices.find_first();
169 if (i == -1) return false;
170 if ((size_t)i != LUIdx) return true;
171 return UsedByIndices.find_next(i) != -1;
174 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
175 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
176 assert(I != RegUsesMap.end() && "Unknown register!");
177 return I->second.UsedByIndices;
180 void RegUseTracker::clear() {
187 /// Formula - This class holds information that describes a formula for
188 /// computing satisfying a use. It may include broken-out immediates and scaled
191 /// AM - This is used to represent complex addressing, as well as other kinds
192 /// of interesting uses.
193 TargetLowering::AddrMode AM;
195 /// BaseRegs - The list of "base" registers for this use. When this is
196 /// non-empty, AM.HasBaseReg should be set to true.
197 SmallVector<const SCEV *, 2> BaseRegs;
199 /// ScaledReg - The 'scaled' register for this use. This should be non-null
200 /// when AM.Scale is not zero.
201 const SCEV *ScaledReg;
203 Formula() : ScaledReg(0) {}
205 void InitialMatch(const SCEV *S, Loop *L,
206 ScalarEvolution &SE, DominatorTree &DT);
208 unsigned getNumRegs() const;
209 const Type *getType() const;
211 void DeleteBaseReg(const SCEV *&S);
213 bool referencesReg(const SCEV *S) const;
214 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
215 const RegUseTracker &RegUses) const;
217 void print(raw_ostream &OS) const;
223 /// DoInitialMatch - Recursion helper for InitialMatch.
224 static void DoInitialMatch(const SCEV *S, Loop *L,
225 SmallVectorImpl<const SCEV *> &Good,
226 SmallVectorImpl<const SCEV *> &Bad,
227 ScalarEvolution &SE, DominatorTree &DT) {
228 // Collect expressions which properly dominate the loop header.
229 if (S->properlyDominates(L->getHeader(), &DT)) {
234 // Look at add operands.
235 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
236 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
238 DoInitialMatch(*I, L, Good, Bad, SE, DT);
242 // Look at addrec operands.
243 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
244 if (!AR->getStart()->isZero()) {
245 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
246 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
247 AR->getStepRecurrence(SE),
249 L, Good, Bad, SE, DT);
253 // Handle a multiplication by -1 (negation) if it didn't fold.
254 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
255 if (Mul->getOperand(0)->isAllOnesValue()) {
256 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
257 const SCEV *NewMul = SE.getMulExpr(Ops);
259 SmallVector<const SCEV *, 4> MyGood;
260 SmallVector<const SCEV *, 4> MyBad;
261 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
262 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
263 SE.getEffectiveSCEVType(NewMul->getType())));
264 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
265 E = MyGood.end(); I != E; ++I)
266 Good.push_back(SE.getMulExpr(NegOne, *I));
267 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
268 E = MyBad.end(); I != E; ++I)
269 Bad.push_back(SE.getMulExpr(NegOne, *I));
273 // Ok, we can't do anything interesting. Just stuff the whole thing into a
274 // register and hope for the best.
278 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
279 /// attempting to keep all loop-invariant and loop-computable values in a
280 /// single base register.
281 void Formula::InitialMatch(const SCEV *S, Loop *L,
282 ScalarEvolution &SE, DominatorTree &DT) {
283 SmallVector<const SCEV *, 4> Good;
284 SmallVector<const SCEV *, 4> Bad;
285 DoInitialMatch(S, L, Good, Bad, SE, DT);
287 const SCEV *Sum = SE.getAddExpr(Good);
289 BaseRegs.push_back(Sum);
290 AM.HasBaseReg = true;
293 const SCEV *Sum = SE.getAddExpr(Bad);
295 BaseRegs.push_back(Sum);
296 AM.HasBaseReg = true;
300 /// getNumRegs - Return the total number of register operands used by this
301 /// formula. This does not include register uses implied by non-constant
303 unsigned Formula::getNumRegs() const {
304 return !!ScaledReg + BaseRegs.size();
307 /// getType - Return the type of this formula, if it has one, or null
308 /// otherwise. This type is meaningless except for the bit size.
309 const Type *Formula::getType() const {
310 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
311 ScaledReg ? ScaledReg->getType() :
312 AM.BaseGV ? AM.BaseGV->getType() :
316 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
317 void Formula::DeleteBaseReg(const SCEV *&S) {
318 if (&S != &BaseRegs.back())
319 std::swap(S, BaseRegs.back());
323 /// referencesReg - Test if this formula references the given register.
324 bool Formula::referencesReg(const SCEV *S) const {
325 return S == ScaledReg ||
326 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
329 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
330 /// which are used by uses other than the use with the given index.
331 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
332 const RegUseTracker &RegUses) const {
334 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
336 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
337 E = BaseRegs.end(); I != E; ++I)
338 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
343 void Formula::print(raw_ostream &OS) const {
346 if (!First) OS << " + "; else First = false;
347 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
349 if (AM.BaseOffs != 0) {
350 if (!First) OS << " + "; else First = false;
353 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
354 E = BaseRegs.end(); I != E; ++I) {
355 if (!First) OS << " + "; else First = false;
356 OS << "reg(" << **I << ')';
358 if (AM.HasBaseReg && BaseRegs.empty()) {
359 if (!First) OS << " + "; else First = false;
360 OS << "**error: HasBaseReg**";
361 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
362 if (!First) OS << " + "; else First = false;
363 OS << "**error: !HasBaseReg**";
366 if (!First) OS << " + "; else First = false;
367 OS << AM.Scale << "*reg(";
376 void Formula::dump() const {
377 print(errs()); errs() << '\n';
380 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
381 /// without changing its value.
382 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
384 IntegerType::get(SE.getContext(), 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(), SE.getTypeSizeInBits(A->getType()) + 1);
393 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
396 /// isMulSExtable - Return true if the given mul can be sign-extended
397 /// without changing its value.
398 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
400 IntegerType::get(SE.getContext(),
401 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
402 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
405 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
406 /// and if the remainder is known to be zero, or null otherwise. If
407 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
408 /// to Y, ignoring that the multiplication may overflow, which is useful when
409 /// the result will be used in a context where the most significant bits are
411 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
413 bool IgnoreSignificantBits = false) {
414 // Handle the trivial case, which works for any SCEV type.
416 return SE.getConstant(LHS->getType(), 1);
418 // Handle a few RHS special cases.
419 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
421 const APInt &RA = RC->getValue()->getValue();
422 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
424 if (RA.isAllOnesValue())
425 return SE.getMulExpr(LHS, RC);
426 // Handle x /s 1 as x.
431 // Check for a division of a constant by a constant.
432 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
435 const APInt &LA = C->getValue()->getValue();
436 const APInt &RA = RC->getValue()->getValue();
437 if (LA.srem(RA) != 0)
439 return SE.getConstant(LA.sdiv(RA));
442 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
443 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
444 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
445 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
446 IgnoreSignificantBits);
448 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
449 IgnoreSignificantBits);
450 if (!Start) return 0;
451 return SE.getAddRecExpr(Start, Step, AR->getLoop());
456 // Distribute the sdiv over add operands, if the add doesn't overflow.
457 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
458 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
459 SmallVector<const SCEV *, 8> Ops;
460 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
462 const SCEV *Op = getExactSDiv(*I, RHS, SE,
463 IgnoreSignificantBits);
467 return SE.getAddExpr(Ops);
472 // Check for a multiply operand that we can pull RHS out of.
473 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
474 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
475 SmallVector<const SCEV *, 4> Ops;
477 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
481 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
482 IgnoreSignificantBits)) {
488 return Found ? SE.getMulExpr(Ops) : 0;
493 // Otherwise we don't know.
497 /// ExtractImmediate - If S involves the addition of a constant integer value,
498 /// return that integer value, and mutate S to point to a new SCEV with that
500 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
501 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
502 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
503 S = SE.getConstant(C->getType(), 0);
504 return C->getValue()->getSExtValue();
506 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
507 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
508 int64_t Result = ExtractImmediate(NewOps.front(), SE);
510 S = SE.getAddExpr(NewOps);
512 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
513 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
514 int64_t Result = ExtractImmediate(NewOps.front(), SE);
516 S = SE.getAddRecExpr(NewOps, AR->getLoop());
522 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
523 /// return that symbol, and mutate S to point to a new SCEV with that
525 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
526 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
527 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
528 S = SE.getConstant(GV->getType(), 0);
531 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
532 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
533 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
535 S = SE.getAddExpr(NewOps);
537 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
538 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
539 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
541 S = SE.getAddRecExpr(NewOps, AR->getLoop());
547 /// isAddressUse - Returns true if the specified instruction is using the
548 /// specified value as an address.
549 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
550 bool isAddress = isa<LoadInst>(Inst);
551 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
552 if (SI->getOperand(1) == OperandVal)
554 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
555 // Addressing modes can also be folded into prefetches and a variety
557 switch (II->getIntrinsicID()) {
559 case Intrinsic::prefetch:
560 case Intrinsic::x86_sse2_loadu_dq:
561 case Intrinsic::x86_sse2_loadu_pd:
562 case Intrinsic::x86_sse_loadu_ps:
563 case Intrinsic::x86_sse_storeu_ps:
564 case Intrinsic::x86_sse2_storeu_pd:
565 case Intrinsic::x86_sse2_storeu_dq:
566 case Intrinsic::x86_sse2_storel_dq:
567 if (II->getArgOperand(0) == OperandVal)
575 /// getAccessType - Return the type of the memory being accessed.
576 static const Type *getAccessType(const Instruction *Inst) {
577 const Type *AccessTy = Inst->getType();
578 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
579 AccessTy = SI->getOperand(0)->getType();
580 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
581 // Addressing modes can also be folded into prefetches and a variety
583 switch (II->getIntrinsicID()) {
585 case Intrinsic::x86_sse_storeu_ps:
586 case Intrinsic::x86_sse2_storeu_pd:
587 case Intrinsic::x86_sse2_storeu_dq:
588 case Intrinsic::x86_sse2_storel_dq:
589 AccessTy = II->getArgOperand(0)->getType();
594 // All pointers have the same requirements, so canonicalize them to an
595 // arbitrary pointer type to minimize variation.
596 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
597 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
598 PTy->getAddressSpace());
603 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
604 /// specified set are trivially dead, delete them and see if this makes any of
605 /// their operands subsequently dead.
607 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
608 bool Changed = false;
610 while (!DeadInsts.empty()) {
611 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
613 if (I == 0 || !isInstructionTriviallyDead(I))
616 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
617 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
620 DeadInsts.push_back(U);
623 I->eraseFromParent();
632 /// Cost - This class is used to measure and compare candidate formulae.
634 /// TODO: Some of these could be merged. Also, a lexical ordering
635 /// isn't always optimal.
639 unsigned NumBaseAdds;
645 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
648 unsigned getNumRegs() const { return NumRegs; }
650 bool operator<(const Cost &Other) const;
654 void RateFormula(const Formula &F,
655 SmallPtrSet<const SCEV *, 16> &Regs,
656 const DenseSet<const SCEV *> &VisitedRegs,
658 const SmallVectorImpl<int64_t> &Offsets,
659 ScalarEvolution &SE, DominatorTree &DT);
661 void print(raw_ostream &OS) const;
665 void RateRegister(const SCEV *Reg,
666 SmallPtrSet<const SCEV *, 16> &Regs,
668 ScalarEvolution &SE, DominatorTree &DT);
669 void RatePrimaryRegister(const SCEV *Reg,
670 SmallPtrSet<const SCEV *, 16> &Regs,
672 ScalarEvolution &SE, DominatorTree &DT);
677 /// RateRegister - Tally up interesting quantities from the given register.
678 void Cost::RateRegister(const SCEV *Reg,
679 SmallPtrSet<const SCEV *, 16> &Regs,
681 ScalarEvolution &SE, DominatorTree &DT) {
682 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
683 if (AR->getLoop() == L)
684 AddRecCost += 1; /// TODO: This should be a function of the stride.
686 // If this is an addrec for a loop that's already been visited by LSR,
687 // don't second-guess its addrec phi nodes. LSR isn't currently smart
688 // enough to reason about more than one loop at a time. Consider these
689 // registers free and leave them alone.
690 else if (L->contains(AR->getLoop()) ||
691 (!AR->getLoop()->contains(L) &&
692 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
693 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
694 PHINode *PN = dyn_cast<PHINode>(I); ++I)
695 if (SE.isSCEVable(PN->getType()) &&
696 (SE.getEffectiveSCEVType(PN->getType()) ==
697 SE.getEffectiveSCEVType(AR->getType())) &&
698 SE.getSCEV(PN) == AR)
701 // If this isn't one of the addrecs that the loop already has, it
702 // would require a costly new phi and add. TODO: This isn't
703 // precisely modeled right now.
705 if (!Regs.count(AR->getStart()))
706 RateRegister(AR->getStart(), Regs, L, SE, DT);
709 // Add the step value register, if it needs one.
710 // TODO: The non-affine case isn't precisely modeled here.
711 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
712 if (!Regs.count(AR->getStart()))
713 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
717 // Rough heuristic; favor registers which don't require extra setup
718 // instructions in the preheader.
719 if (!isa<SCEVUnknown>(Reg) &&
720 !isa<SCEVConstant>(Reg) &&
721 !(isa<SCEVAddRecExpr>(Reg) &&
722 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
723 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
727 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
729 void Cost::RatePrimaryRegister(const SCEV *Reg,
730 SmallPtrSet<const SCEV *, 16> &Regs,
732 ScalarEvolution &SE, DominatorTree &DT) {
733 if (Regs.insert(Reg))
734 RateRegister(Reg, Regs, L, SE, DT);
737 void Cost::RateFormula(const Formula &F,
738 SmallPtrSet<const SCEV *, 16> &Regs,
739 const DenseSet<const SCEV *> &VisitedRegs,
741 const SmallVectorImpl<int64_t> &Offsets,
742 ScalarEvolution &SE, DominatorTree &DT) {
743 // Tally up the registers.
744 if (const SCEV *ScaledReg = F.ScaledReg) {
745 if (VisitedRegs.count(ScaledReg)) {
749 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
751 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
752 E = F.BaseRegs.end(); I != E; ++I) {
753 const SCEV *BaseReg = *I;
754 if (VisitedRegs.count(BaseReg)) {
758 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
760 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
761 BaseReg->hasComputableLoopEvolution(L);
764 if (F.BaseRegs.size() > 1)
765 NumBaseAdds += F.BaseRegs.size() - 1;
767 // Tally up the non-zero immediates.
768 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
769 E = Offsets.end(); I != E; ++I) {
770 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
772 ImmCost += 64; // Handle symbolic values conservatively.
773 // TODO: This should probably be the pointer size.
774 else if (Offset != 0)
775 ImmCost += APInt(64, Offset, true).getMinSignedBits();
779 /// Loose - Set this cost to a loosing value.
789 /// operator< - Choose the lower cost.
790 bool Cost::operator<(const Cost &Other) const {
791 if (NumRegs != Other.NumRegs)
792 return NumRegs < Other.NumRegs;
793 if (AddRecCost != Other.AddRecCost)
794 return AddRecCost < Other.AddRecCost;
795 if (NumIVMuls != Other.NumIVMuls)
796 return NumIVMuls < Other.NumIVMuls;
797 if (NumBaseAdds != Other.NumBaseAdds)
798 return NumBaseAdds < Other.NumBaseAdds;
799 if (ImmCost != Other.ImmCost)
800 return ImmCost < Other.ImmCost;
801 if (SetupCost != Other.SetupCost)
802 return SetupCost < Other.SetupCost;
806 void Cost::print(raw_ostream &OS) const {
807 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
809 OS << ", with addrec cost " << AddRecCost;
811 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
812 if (NumBaseAdds != 0)
813 OS << ", plus " << NumBaseAdds << " base add"
814 << (NumBaseAdds == 1 ? "" : "s");
816 OS << ", plus " << ImmCost << " imm cost";
818 OS << ", plus " << SetupCost << " setup cost";
821 void Cost::dump() const {
822 print(errs()); errs() << '\n';
827 /// LSRFixup - An operand value in an instruction which is to be replaced
828 /// with some equivalent, possibly strength-reduced, replacement.
830 /// UserInst - The instruction which will be updated.
831 Instruction *UserInst;
833 /// OperandValToReplace - The operand of the instruction which will
834 /// be replaced. The operand may be used more than once; every instance
835 /// will be replaced.
836 Value *OperandValToReplace;
838 /// PostIncLoops - If this user is to use the post-incremented value of an
839 /// induction variable, this variable is non-null and holds the loop
840 /// associated with the induction variable.
841 PostIncLoopSet PostIncLoops;
843 /// LUIdx - The index of the LSRUse describing the expression which
844 /// this fixup needs, minus an offset (below).
847 /// Offset - A constant offset to be added to the LSRUse expression.
848 /// This allows multiple fixups to share the same LSRUse with different
849 /// offsets, for example in an unrolled loop.
852 bool isUseFullyOutsideLoop(const Loop *L) const;
856 void print(raw_ostream &OS) const;
863 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
865 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
866 /// value outside of the given loop.
867 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
868 // PHI nodes use their value in their incoming blocks.
869 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
870 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
871 if (PN->getIncomingValue(i) == OperandValToReplace &&
872 L->contains(PN->getIncomingBlock(i)))
877 return !L->contains(UserInst);
880 void LSRFixup::print(raw_ostream &OS) const {
882 // Store is common and interesting enough to be worth special-casing.
883 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
885 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
886 } else if (UserInst->getType()->isVoidTy())
887 OS << UserInst->getOpcodeName();
889 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
891 OS << ", OperandValToReplace=";
892 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
894 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
895 E = PostIncLoops.end(); I != E; ++I) {
896 OS << ", PostIncLoop=";
897 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
900 if (LUIdx != ~size_t(0))
901 OS << ", LUIdx=" << LUIdx;
904 OS << ", Offset=" << Offset;
907 void LSRFixup::dump() const {
908 print(errs()); errs() << '\n';
913 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
914 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
915 struct UniquifierDenseMapInfo {
916 static SmallVector<const SCEV *, 2> getEmptyKey() {
917 SmallVector<const SCEV *, 2> V;
918 V.push_back(reinterpret_cast<const SCEV *>(-1));
922 static SmallVector<const SCEV *, 2> getTombstoneKey() {
923 SmallVector<const SCEV *, 2> V;
924 V.push_back(reinterpret_cast<const SCEV *>(-2));
928 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
930 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
931 E = V.end(); I != E; ++I)
932 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
936 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
937 const SmallVector<const SCEV *, 2> &RHS) {
942 /// LSRUse - This class holds the state that LSR keeps for each use in
943 /// IVUsers, as well as uses invented by LSR itself. It includes information
944 /// about what kinds of things can be folded into the user, information about
945 /// the user itself, and information about how the use may be satisfied.
946 /// TODO: Represent multiple users of the same expression in common?
948 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
951 /// KindType - An enum for a kind of use, indicating what types of
952 /// scaled and immediate operands it might support.
954 Basic, ///< A normal use, with no folding.
955 Special, ///< A special case of basic, allowing -1 scales.
956 Address, ///< An address use; folding according to TargetLowering
957 ICmpZero ///< An equality icmp with both operands folded into one.
958 // TODO: Add a generic icmp too?
962 const Type *AccessTy;
964 SmallVector<int64_t, 8> Offsets;
968 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
969 /// LSRUse are outside of the loop, in which case some special-case heuristics
971 bool AllFixupsOutsideLoop;
973 /// WidestFixupType - This records the widest use type for any fixup using
974 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
975 /// max fixup widths to be equivalent, because the narrower one may be relying
976 /// on the implicit truncation to truncate away bogus bits.
977 const Type *WidestFixupType;
979 /// Formulae - A list of ways to build a value that can satisfy this user.
980 /// After the list is populated, one of these is selected heuristically and
981 /// used to formulate a replacement for OperandValToReplace in UserInst.
982 SmallVector<Formula, 12> Formulae;
984 /// Regs - The set of register candidates used by all formulae in this LSRUse.
985 SmallPtrSet<const SCEV *, 4> Regs;
987 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
988 MinOffset(INT64_MAX),
989 MaxOffset(INT64_MIN),
990 AllFixupsOutsideLoop(true),
991 WidestFixupType(0) {}
993 bool HasFormulaWithSameRegs(const Formula &F) const;
994 bool InsertFormula(const Formula &F);
995 void DeleteFormula(Formula &F);
996 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
998 void print(raw_ostream &OS) const;
1004 /// HasFormula - Test whether this use as a formula which has the same
1005 /// registers as the given formula.
1006 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1007 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1008 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1009 // Unstable sort by host order ok, because this is only used for uniquifying.
1010 std::sort(Key.begin(), Key.end());
1011 return Uniquifier.count(Key);
1014 /// InsertFormula - If the given formula has not yet been inserted, add it to
1015 /// the list, and return true. Return false otherwise.
1016 bool LSRUse::InsertFormula(const Formula &F) {
1017 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1018 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1019 // Unstable sort by host order ok, because this is only used for uniquifying.
1020 std::sort(Key.begin(), Key.end());
1022 if (!Uniquifier.insert(Key).second)
1025 // Using a register to hold the value of 0 is not profitable.
1026 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1027 "Zero allocated in a scaled register!");
1029 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1030 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1031 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1034 // Add the formula to the list.
1035 Formulae.push_back(F);
1037 // Record registers now being used by this use.
1038 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1039 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1044 /// DeleteFormula - Remove the given formula from this use's list.
1045 void LSRUse::DeleteFormula(Formula &F) {
1046 if (&F != &Formulae.back())
1047 std::swap(F, Formulae.back());
1048 Formulae.pop_back();
1049 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1052 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1053 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1054 // Now that we've filtered out some formulae, recompute the Regs set.
1055 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1057 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1058 E = Formulae.end(); I != E; ++I) {
1059 const Formula &F = *I;
1060 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1061 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1064 // Update the RegTracker.
1065 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1066 E = OldRegs.end(); I != E; ++I)
1067 if (!Regs.count(*I))
1068 RegUses.DropRegister(*I, LUIdx);
1071 void LSRUse::print(raw_ostream &OS) const {
1072 OS << "LSR Use: Kind=";
1074 case Basic: OS << "Basic"; break;
1075 case Special: OS << "Special"; break;
1076 case ICmpZero: OS << "ICmpZero"; break;
1078 OS << "Address of ";
1079 if (AccessTy->isPointerTy())
1080 OS << "pointer"; // the full pointer type could be really verbose
1085 OS << ", Offsets={";
1086 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1087 E = Offsets.end(); I != E; ++I) {
1089 if (llvm::next(I) != E)
1094 if (AllFixupsOutsideLoop)
1095 OS << ", all-fixups-outside-loop";
1097 if (WidestFixupType)
1098 OS << ", widest fixup type: " << *WidestFixupType;
1101 void LSRUse::dump() const {
1102 print(errs()); errs() << '\n';
1105 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1106 /// be completely folded into the user instruction at isel time. This includes
1107 /// address-mode folding and special icmp tricks.
1108 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1109 LSRUse::KindType Kind, const Type *AccessTy,
1110 const TargetLowering *TLI) {
1112 case LSRUse::Address:
1113 // If we have low-level target information, ask the target if it can
1114 // completely fold this address.
1115 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1117 // Otherwise, just guess that reg+reg addressing is legal.
1118 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1120 case LSRUse::ICmpZero:
1121 // There's not even a target hook for querying whether it would be legal to
1122 // fold a GV into an ICmp.
1126 // ICmp only has two operands; don't allow more than two non-trivial parts.
1127 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1130 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1131 // putting the scaled register in the other operand of the icmp.
1132 if (AM.Scale != 0 && AM.Scale != -1)
1135 // If we have low-level target information, ask the target if it can fold an
1136 // integer immediate on an icmp.
1137 if (AM.BaseOffs != 0) {
1138 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1145 // Only handle single-register values.
1146 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1148 case LSRUse::Special:
1149 // Only handle -1 scales, or no scale.
1150 return AM.Scale == 0 || AM.Scale == -1;
1156 static bool isLegalUse(TargetLowering::AddrMode AM,
1157 int64_t MinOffset, int64_t MaxOffset,
1158 LSRUse::KindType Kind, const Type *AccessTy,
1159 const TargetLowering *TLI) {
1160 // Check for overflow.
1161 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1164 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1165 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1166 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1167 // Check for overflow.
1168 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1171 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1172 return isLegalUse(AM, Kind, AccessTy, TLI);
1177 static bool isAlwaysFoldable(int64_t BaseOffs,
1178 GlobalValue *BaseGV,
1180 LSRUse::KindType Kind, const Type *AccessTy,
1181 const TargetLowering *TLI) {
1182 // Fast-path: zero is always foldable.
1183 if (BaseOffs == 0 && !BaseGV) return true;
1185 // Conservatively, create an address with an immediate and a
1186 // base and a scale.
1187 TargetLowering::AddrMode AM;
1188 AM.BaseOffs = BaseOffs;
1190 AM.HasBaseReg = HasBaseReg;
1191 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1193 // Canonicalize a scale of 1 to a base register if the formula doesn't
1194 // already have a base register.
1195 if (!AM.HasBaseReg && AM.Scale == 1) {
1197 AM.HasBaseReg = true;
1200 return isLegalUse(AM, Kind, AccessTy, TLI);
1203 static bool isAlwaysFoldable(const SCEV *S,
1204 int64_t MinOffset, int64_t MaxOffset,
1206 LSRUse::KindType Kind, const Type *AccessTy,
1207 const TargetLowering *TLI,
1208 ScalarEvolution &SE) {
1209 // Fast-path: zero is always foldable.
1210 if (S->isZero()) return true;
1212 // Conservatively, create an address with an immediate and a
1213 // base and a scale.
1214 int64_t BaseOffs = ExtractImmediate(S, SE);
1215 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1217 // If there's anything else involved, it's not foldable.
1218 if (!S->isZero()) return false;
1220 // Fast-path: zero is always foldable.
1221 if (BaseOffs == 0 && !BaseGV) return true;
1223 // Conservatively, create an address with an immediate and a
1224 // base and a scale.
1225 TargetLowering::AddrMode AM;
1226 AM.BaseOffs = BaseOffs;
1228 AM.HasBaseReg = HasBaseReg;
1229 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1231 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1236 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1237 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1238 struct UseMapDenseMapInfo {
1239 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1240 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1243 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1244 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1248 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1249 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1250 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1254 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1255 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1260 /// FormulaSorter - This class implements an ordering for formulae which sorts
1261 /// the by their standalone cost.
1262 class FormulaSorter {
1263 /// These two sets are kept empty, so that we compute standalone costs.
1264 DenseSet<const SCEV *> VisitedRegs;
1265 SmallPtrSet<const SCEV *, 16> Regs;
1268 ScalarEvolution &SE;
1272 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1273 : L(l), LU(&lu), SE(se), DT(dt) {}
1275 bool operator()(const Formula &A, const Formula &B) {
1277 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1280 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1282 return CostA < CostB;
1286 /// LSRInstance - This class holds state for the main loop strength reduction
1290 ScalarEvolution &SE;
1293 const TargetLowering *const TLI;
1297 /// IVIncInsertPos - This is the insert position that the current loop's
1298 /// induction variable increment should be placed. In simple loops, this is
1299 /// the latch block's terminator. But in more complicated cases, this is a
1300 /// position which will dominate all the in-loop post-increment users.
1301 Instruction *IVIncInsertPos;
1303 /// Factors - Interesting factors between use strides.
1304 SmallSetVector<int64_t, 8> Factors;
1306 /// Types - Interesting use types, to facilitate truncation reuse.
1307 SmallSetVector<const Type *, 4> Types;
1309 /// Fixups - The list of operands which are to be replaced.
1310 SmallVector<LSRFixup, 16> Fixups;
1312 /// Uses - The list of interesting uses.
1313 SmallVector<LSRUse, 16> Uses;
1315 /// RegUses - Track which uses use which register candidates.
1316 RegUseTracker RegUses;
1318 void OptimizeShadowIV();
1319 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1320 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1321 void OptimizeLoopTermCond();
1323 void CollectInterestingTypesAndFactors();
1324 void CollectFixupsAndInitialFormulae();
1326 LSRFixup &getNewFixup() {
1327 Fixups.push_back(LSRFixup());
1328 return Fixups.back();
1331 // Support for sharing of LSRUses between LSRFixups.
1332 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1334 UseMapDenseMapInfo> UseMapTy;
1337 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1338 LSRUse::KindType Kind, const Type *AccessTy);
1340 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1341 LSRUse::KindType Kind,
1342 const Type *AccessTy);
1344 void DeleteUse(LSRUse &LU);
1346 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1349 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1350 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1351 void CountRegisters(const Formula &F, size_t LUIdx);
1352 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1354 void CollectLoopInvariantFixupsAndFormulae();
1356 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1357 unsigned Depth = 0);
1358 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1359 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1360 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1361 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1362 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1363 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1364 void GenerateCrossUseConstantOffsets();
1365 void GenerateAllReuseFormulae();
1367 void FilterOutUndesirableDedicatedRegisters();
1369 size_t EstimateSearchSpaceComplexity() const;
1370 void NarrowSearchSpaceUsingHeuristics();
1372 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1374 SmallVectorImpl<const Formula *> &Workspace,
1375 const Cost &CurCost,
1376 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1377 DenseSet<const SCEV *> &VisitedRegs) const;
1378 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1380 BasicBlock::iterator
1381 HoistInsertPosition(BasicBlock::iterator IP,
1382 const SmallVectorImpl<Instruction *> &Inputs) const;
1383 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1385 const LSRUse &LU) const;
1387 Value *Expand(const LSRFixup &LF,
1389 BasicBlock::iterator IP,
1390 SCEVExpander &Rewriter,
1391 SmallVectorImpl<WeakVH> &DeadInsts) const;
1392 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1394 SCEVExpander &Rewriter,
1395 SmallVectorImpl<WeakVH> &DeadInsts,
1397 void Rewrite(const LSRFixup &LF,
1399 SCEVExpander &Rewriter,
1400 SmallVectorImpl<WeakVH> &DeadInsts,
1402 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1405 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1407 bool getChanged() const { return Changed; }
1409 void print_factors_and_types(raw_ostream &OS) const;
1410 void print_fixups(raw_ostream &OS) const;
1411 void print_uses(raw_ostream &OS) const;
1412 void print(raw_ostream &OS) const;
1418 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1419 /// inside the loop then try to eliminate the cast operation.
1420 void LSRInstance::OptimizeShadowIV() {
1421 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1422 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1425 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1426 UI != E; /* empty */) {
1427 IVUsers::const_iterator CandidateUI = UI;
1429 Instruction *ShadowUse = CandidateUI->getUser();
1430 const Type *DestTy = NULL;
1432 /* If shadow use is a int->float cast then insert a second IV
1433 to eliminate this cast.
1435 for (unsigned i = 0; i < n; ++i)
1441 for (unsigned i = 0; i < n; ++i, ++d)
1444 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1445 DestTy = UCast->getDestTy();
1446 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1447 DestTy = SCast->getDestTy();
1448 if (!DestTy) continue;
1451 // If target does not support DestTy natively then do not apply
1452 // this transformation.
1453 EVT DVT = TLI->getValueType(DestTy);
1454 if (!TLI->isTypeLegal(DVT)) continue;
1457 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1459 if (PH->getNumIncomingValues() != 2) continue;
1461 const Type *SrcTy = PH->getType();
1462 int Mantissa = DestTy->getFPMantissaWidth();
1463 if (Mantissa == -1) continue;
1464 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1467 unsigned Entry, Latch;
1468 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1476 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1477 if (!Init) continue;
1478 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1480 BinaryOperator *Incr =
1481 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1482 if (!Incr) continue;
1483 if (Incr->getOpcode() != Instruction::Add
1484 && Incr->getOpcode() != Instruction::Sub)
1487 /* Initialize new IV, double d = 0.0 in above example. */
1488 ConstantInt *C = NULL;
1489 if (Incr->getOperand(0) == PH)
1490 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1491 else if (Incr->getOperand(1) == PH)
1492 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1498 // Ignore negative constants, as the code below doesn't handle them
1499 // correctly. TODO: Remove this restriction.
1500 if (!C->getValue().isStrictlyPositive()) continue;
1502 /* Add new PHINode. */
1503 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1505 /* create new increment. '++d' in above example. */
1506 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1507 BinaryOperator *NewIncr =
1508 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1509 Instruction::FAdd : Instruction::FSub,
1510 NewPH, CFP, "IV.S.next.", Incr);
1512 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1513 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1515 /* Remove cast operation */
1516 ShadowUse->replaceAllUsesWith(NewPH);
1517 ShadowUse->eraseFromParent();
1523 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1524 /// set the IV user and stride information and return true, otherwise return
1526 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1527 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1528 if (UI->getUser() == Cond) {
1529 // NOTE: we could handle setcc instructions with multiple uses here, but
1530 // InstCombine does it as well for simple uses, it's not clear that it
1531 // occurs enough in real life to handle.
1538 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1539 /// a max computation.
1541 /// This is a narrow solution to a specific, but acute, problem. For loops
1547 /// } while (++i < n);
1549 /// the trip count isn't just 'n', because 'n' might not be positive. And
1550 /// unfortunately this can come up even for loops where the user didn't use
1551 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1552 /// will commonly be lowered like this:
1558 /// } while (++i < n);
1561 /// and then it's possible for subsequent optimization to obscure the if
1562 /// test in such a way that indvars can't find it.
1564 /// When indvars can't find the if test in loops like this, it creates a
1565 /// max expression, which allows it to give the loop a canonical
1566 /// induction variable:
1569 /// max = n < 1 ? 1 : n;
1572 /// } while (++i != max);
1574 /// Canonical induction variables are necessary because the loop passes
1575 /// are designed around them. The most obvious example of this is the
1576 /// LoopInfo analysis, which doesn't remember trip count values. It
1577 /// expects to be able to rediscover the trip count each time it is
1578 /// needed, and it does this using a simple analysis that only succeeds if
1579 /// the loop has a canonical induction variable.
1581 /// However, when it comes time to generate code, the maximum operation
1582 /// can be quite costly, especially if it's inside of an outer loop.
1584 /// This function solves this problem by detecting this type of loop and
1585 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1586 /// the instructions for the maximum computation.
1588 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1589 // Check that the loop matches the pattern we're looking for.
1590 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1591 Cond->getPredicate() != CmpInst::ICMP_NE)
1594 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1595 if (!Sel || !Sel->hasOneUse()) return Cond;
1597 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1598 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1600 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1602 // Add one to the backedge-taken count to get the trip count.
1603 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1604 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1606 // Check for a max calculation that matches the pattern. There's no check
1607 // for ICMP_ULE here because the comparison would be with zero, which
1608 // isn't interesting.
1609 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1610 const SCEVNAryExpr *Max = 0;
1611 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1612 Pred = ICmpInst::ICMP_SLE;
1614 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1615 Pred = ICmpInst::ICMP_SLT;
1617 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1618 Pred = ICmpInst::ICMP_ULT;
1625 // To handle a max with more than two operands, this optimization would
1626 // require additional checking and setup.
1627 if (Max->getNumOperands() != 2)
1630 const SCEV *MaxLHS = Max->getOperand(0);
1631 const SCEV *MaxRHS = Max->getOperand(1);
1633 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1634 // for a comparison with 1. For <= and >=, a comparison with zero.
1636 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1639 // Check the relevant induction variable for conformance to
1641 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1642 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1643 if (!AR || !AR->isAffine() ||
1644 AR->getStart() != One ||
1645 AR->getStepRecurrence(SE) != One)
1648 assert(AR->getLoop() == L &&
1649 "Loop condition operand is an addrec in a different loop!");
1651 // Check the right operand of the select, and remember it, as it will
1652 // be used in the new comparison instruction.
1654 if (ICmpInst::isTrueWhenEqual(Pred)) {
1655 // Look for n+1, and grab n.
1656 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1657 if (isa<ConstantInt>(BO->getOperand(1)) &&
1658 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1659 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1660 NewRHS = BO->getOperand(0);
1661 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1662 if (isa<ConstantInt>(BO->getOperand(1)) &&
1663 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1664 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1665 NewRHS = BO->getOperand(0);
1668 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1669 NewRHS = Sel->getOperand(1);
1670 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1671 NewRHS = Sel->getOperand(2);
1672 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1673 NewRHS = SU->getValue();
1675 // Max doesn't match expected pattern.
1678 // Determine the new comparison opcode. It may be signed or unsigned,
1679 // and the original comparison may be either equality or inequality.
1680 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1681 Pred = CmpInst::getInversePredicate(Pred);
1683 // Ok, everything looks ok to change the condition into an SLT or SGE and
1684 // delete the max calculation.
1686 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1688 // Delete the max calculation instructions.
1689 Cond->replaceAllUsesWith(NewCond);
1690 CondUse->setUser(NewCond);
1691 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1692 Cond->eraseFromParent();
1693 Sel->eraseFromParent();
1694 if (Cmp->use_empty())
1695 Cmp->eraseFromParent();
1699 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1700 /// postinc iv when possible.
1702 LSRInstance::OptimizeLoopTermCond() {
1703 SmallPtrSet<Instruction *, 4> PostIncs;
1705 BasicBlock *LatchBlock = L->getLoopLatch();
1706 SmallVector<BasicBlock*, 8> ExitingBlocks;
1707 L->getExitingBlocks(ExitingBlocks);
1709 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1710 BasicBlock *ExitingBlock = ExitingBlocks[i];
1712 // Get the terminating condition for the loop if possible. If we
1713 // can, we want to change it to use a post-incremented version of its
1714 // induction variable, to allow coalescing the live ranges for the IV into
1715 // one register value.
1717 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1720 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1721 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1724 // Search IVUsesByStride to find Cond's IVUse if there is one.
1725 IVStrideUse *CondUse = 0;
1726 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1727 if (!FindIVUserForCond(Cond, CondUse))
1730 // If the trip count is computed in terms of a max (due to ScalarEvolution
1731 // being unable to find a sufficient guard, for example), change the loop
1732 // comparison to use SLT or ULT instead of NE.
1733 // One consequence of doing this now is that it disrupts the count-down
1734 // optimization. That's not always a bad thing though, because in such
1735 // cases it may still be worthwhile to avoid a max.
1736 Cond = OptimizeMax(Cond, CondUse);
1738 // If this exiting block dominates the latch block, it may also use
1739 // the post-inc value if it won't be shared with other uses.
1740 // Check for dominance.
1741 if (!DT.dominates(ExitingBlock, LatchBlock))
1744 // Conservatively avoid trying to use the post-inc value in non-latch
1745 // exits if there may be pre-inc users in intervening blocks.
1746 if (LatchBlock != ExitingBlock)
1747 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1748 // Test if the use is reachable from the exiting block. This dominator
1749 // query is a conservative approximation of reachability.
1750 if (&*UI != CondUse &&
1751 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1752 // Conservatively assume there may be reuse if the quotient of their
1753 // strides could be a legal scale.
1754 const SCEV *A = IU.getStride(*CondUse, L);
1755 const SCEV *B = IU.getStride(*UI, L);
1756 if (!A || !B) continue;
1757 if (SE.getTypeSizeInBits(A->getType()) !=
1758 SE.getTypeSizeInBits(B->getType())) {
1759 if (SE.getTypeSizeInBits(A->getType()) >
1760 SE.getTypeSizeInBits(B->getType()))
1761 B = SE.getSignExtendExpr(B, A->getType());
1763 A = SE.getSignExtendExpr(A, B->getType());
1765 if (const SCEVConstant *D =
1766 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1767 const ConstantInt *C = D->getValue();
1768 // Stride of one or negative one can have reuse with non-addresses.
1769 if (C->isOne() || C->isAllOnesValue())
1770 goto decline_post_inc;
1771 // Avoid weird situations.
1772 if (C->getValue().getMinSignedBits() >= 64 ||
1773 C->getValue().isMinSignedValue())
1774 goto decline_post_inc;
1775 // Without TLI, assume that any stride might be valid, and so any
1776 // use might be shared.
1778 goto decline_post_inc;
1779 // Check for possible scaled-address reuse.
1780 const Type *AccessTy = getAccessType(UI->getUser());
1781 TargetLowering::AddrMode AM;
1782 AM.Scale = C->getSExtValue();
1783 if (TLI->isLegalAddressingMode(AM, AccessTy))
1784 goto decline_post_inc;
1785 AM.Scale = -AM.Scale;
1786 if (TLI->isLegalAddressingMode(AM, AccessTy))
1787 goto decline_post_inc;
1791 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1794 // It's possible for the setcc instruction to be anywhere in the loop, and
1795 // possible for it to have multiple users. If it is not immediately before
1796 // the exiting block branch, move it.
1797 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1798 if (Cond->hasOneUse()) {
1799 Cond->moveBefore(TermBr);
1801 // Clone the terminating condition and insert into the loopend.
1802 ICmpInst *OldCond = Cond;
1803 Cond = cast<ICmpInst>(Cond->clone());
1804 Cond->setName(L->getHeader()->getName() + ".termcond");
1805 ExitingBlock->getInstList().insert(TermBr, Cond);
1807 // Clone the IVUse, as the old use still exists!
1808 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1809 TermBr->replaceUsesOfWith(OldCond, Cond);
1813 // If we get to here, we know that we can transform the setcc instruction to
1814 // use the post-incremented version of the IV, allowing us to coalesce the
1815 // live ranges for the IV correctly.
1816 CondUse->transformToPostInc(L);
1819 PostIncs.insert(Cond);
1823 // Determine an insertion point for the loop induction variable increment. It
1824 // must dominate all the post-inc comparisons we just set up, and it must
1825 // dominate the loop latch edge.
1826 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1827 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1828 E = PostIncs.end(); I != E; ++I) {
1830 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1832 if (BB == (*I)->getParent())
1833 IVIncInsertPos = *I;
1834 else if (BB != IVIncInsertPos->getParent())
1835 IVIncInsertPos = BB->getTerminator();
1839 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1840 /// at the given offset and other details. If so, update the use and
1843 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1844 LSRUse::KindType Kind, const Type *AccessTy) {
1845 int64_t NewMinOffset = LU.MinOffset;
1846 int64_t NewMaxOffset = LU.MaxOffset;
1847 const Type *NewAccessTy = AccessTy;
1849 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1850 // something conservative, however this can pessimize in the case that one of
1851 // the uses will have all its uses outside the loop, for example.
1852 if (LU.Kind != Kind)
1854 // Conservatively assume HasBaseReg is true for now.
1855 if (NewOffset < LU.MinOffset) {
1856 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1857 Kind, AccessTy, TLI))
1859 NewMinOffset = NewOffset;
1860 } else if (NewOffset > LU.MaxOffset) {
1861 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1862 Kind, AccessTy, TLI))
1864 NewMaxOffset = NewOffset;
1866 // Check for a mismatched access type, and fall back conservatively as needed.
1867 // TODO: Be less conservative when the type is similar and can use the same
1868 // addressing modes.
1869 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1870 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1873 LU.MinOffset = NewMinOffset;
1874 LU.MaxOffset = NewMaxOffset;
1875 LU.AccessTy = NewAccessTy;
1876 if (NewOffset != LU.Offsets.back())
1877 LU.Offsets.push_back(NewOffset);
1881 /// getUse - Return an LSRUse index and an offset value for a fixup which
1882 /// needs the given expression, with the given kind and optional access type.
1883 /// Either reuse an existing use or create a new one, as needed.
1884 std::pair<size_t, int64_t>
1885 LSRInstance::getUse(const SCEV *&Expr,
1886 LSRUse::KindType Kind, const Type *AccessTy) {
1887 const SCEV *Copy = Expr;
1888 int64_t Offset = ExtractImmediate(Expr, SE);
1890 // Basic uses can't accept any offset, for example.
1891 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1896 std::pair<UseMapTy::iterator, bool> P =
1897 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1899 // A use already existed with this base.
1900 size_t LUIdx = P.first->second;
1901 LSRUse &LU = Uses[LUIdx];
1902 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1904 return std::make_pair(LUIdx, Offset);
1907 // Create a new use.
1908 size_t LUIdx = Uses.size();
1909 P.first->second = LUIdx;
1910 Uses.push_back(LSRUse(Kind, AccessTy));
1911 LSRUse &LU = Uses[LUIdx];
1913 // We don't need to track redundant offsets, but we don't need to go out
1914 // of our way here to avoid them.
1915 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1916 LU.Offsets.push_back(Offset);
1918 LU.MinOffset = Offset;
1919 LU.MaxOffset = Offset;
1920 return std::make_pair(LUIdx, Offset);
1923 /// DeleteUse - Delete the given use from the Uses list.
1924 void LSRInstance::DeleteUse(LSRUse &LU) {
1925 if (&LU != &Uses.back())
1926 std::swap(LU, Uses.back());
1930 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1931 /// a formula that has the same registers as the given formula.
1933 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1934 const LSRUse &OrigLU) {
1935 // Search all uses for the formula. This could be more clever. Ignore
1936 // ICmpZero uses because they may contain formulae generated by
1937 // GenerateICmpZeroScales, in which case adding fixup offsets may
1939 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1940 LSRUse &LU = Uses[LUIdx];
1941 if (&LU != &OrigLU &&
1942 LU.Kind != LSRUse::ICmpZero &&
1943 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1944 LU.WidestFixupType == OrigLU.WidestFixupType &&
1945 LU.HasFormulaWithSameRegs(OrigF)) {
1946 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1947 E = LU.Formulae.end(); I != E; ++I) {
1948 const Formula &F = *I;
1949 if (F.BaseRegs == OrigF.BaseRegs &&
1950 F.ScaledReg == OrigF.ScaledReg &&
1951 F.AM.BaseGV == OrigF.AM.BaseGV &&
1952 F.AM.Scale == OrigF.AM.Scale &&
1954 if (F.AM.BaseOffs == 0)
1965 void LSRInstance::CollectInterestingTypesAndFactors() {
1966 SmallSetVector<const SCEV *, 4> Strides;
1968 // Collect interesting types and strides.
1969 SmallVector<const SCEV *, 4> Worklist;
1970 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1971 const SCEV *Expr = IU.getExpr(*UI);
1973 // Collect interesting types.
1974 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1976 // Add strides for mentioned loops.
1977 Worklist.push_back(Expr);
1979 const SCEV *S = Worklist.pop_back_val();
1980 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1981 Strides.insert(AR->getStepRecurrence(SE));
1982 Worklist.push_back(AR->getStart());
1983 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1984 Worklist.append(Add->op_begin(), Add->op_end());
1986 } while (!Worklist.empty());
1989 // Compute interesting factors from the set of interesting strides.
1990 for (SmallSetVector<const SCEV *, 4>::const_iterator
1991 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1992 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1993 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
1994 const SCEV *OldStride = *I;
1995 const SCEV *NewStride = *NewStrideIter;
1997 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1998 SE.getTypeSizeInBits(NewStride->getType())) {
1999 if (SE.getTypeSizeInBits(OldStride->getType()) >
2000 SE.getTypeSizeInBits(NewStride->getType()))
2001 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2003 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2005 if (const SCEVConstant *Factor =
2006 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2008 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2009 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2010 } else if (const SCEVConstant *Factor =
2011 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2014 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2015 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2019 // If all uses use the same type, don't bother looking for truncation-based
2021 if (Types.size() == 1)
2024 DEBUG(print_factors_and_types(dbgs()));
2027 void LSRInstance::CollectFixupsAndInitialFormulae() {
2028 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2030 LSRFixup &LF = getNewFixup();
2031 LF.UserInst = UI->getUser();
2032 LF.OperandValToReplace = UI->getOperandValToReplace();
2033 LF.PostIncLoops = UI->getPostIncLoops();
2035 LSRUse::KindType Kind = LSRUse::Basic;
2036 const Type *AccessTy = 0;
2037 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2038 Kind = LSRUse::Address;
2039 AccessTy = getAccessType(LF.UserInst);
2042 const SCEV *S = IU.getExpr(*UI);
2044 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2045 // (N - i == 0), and this allows (N - i) to be the expression that we work
2046 // with rather than just N or i, so we can consider the register
2047 // requirements for both N and i at the same time. Limiting this code to
2048 // equality icmps is not a problem because all interesting loops use
2049 // equality icmps, thanks to IndVarSimplify.
2050 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2051 if (CI->isEquality()) {
2052 // Swap the operands if needed to put the OperandValToReplace on the
2053 // left, for consistency.
2054 Value *NV = CI->getOperand(1);
2055 if (NV == LF.OperandValToReplace) {
2056 CI->setOperand(1, CI->getOperand(0));
2057 CI->setOperand(0, NV);
2058 NV = CI->getOperand(1);
2062 // x == y --> x - y == 0
2063 const SCEV *N = SE.getSCEV(NV);
2064 if (N->isLoopInvariant(L)) {
2065 Kind = LSRUse::ICmpZero;
2066 S = SE.getMinusSCEV(N, S);
2069 // -1 and the negations of all interesting strides (except the negation
2070 // of -1) are now also interesting.
2071 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2072 if (Factors[i] != -1)
2073 Factors.insert(-(uint64_t)Factors[i]);
2077 // Set up the initial formula for this use.
2078 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2080 LF.Offset = P.second;
2081 LSRUse &LU = Uses[LF.LUIdx];
2082 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2083 if (!LU.WidestFixupType ||
2084 SE.getTypeSizeInBits(LU.WidestFixupType) <
2085 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2086 LU.WidestFixupType = LF.OperandValToReplace->getType();
2088 // If this is the first use of this LSRUse, give it a formula.
2089 if (LU.Formulae.empty()) {
2090 InsertInitialFormula(S, LU, LF.LUIdx);
2091 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2095 DEBUG(print_fixups(dbgs()));
2098 /// InsertInitialFormula - Insert a formula for the given expression into
2099 /// the given use, separating out loop-variant portions from loop-invariant
2100 /// and loop-computable portions.
2102 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2104 F.InitialMatch(S, L, SE, DT);
2105 bool Inserted = InsertFormula(LU, LUIdx, F);
2106 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2109 /// InsertSupplementalFormula - Insert a simple single-register formula for
2110 /// the given expression into the given use.
2112 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2113 LSRUse &LU, size_t LUIdx) {
2115 F.BaseRegs.push_back(S);
2116 F.AM.HasBaseReg = true;
2117 bool Inserted = InsertFormula(LU, LUIdx, F);
2118 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2121 /// CountRegisters - Note which registers are used by the given formula,
2122 /// updating RegUses.
2123 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2125 RegUses.CountRegister(F.ScaledReg, LUIdx);
2126 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2127 E = F.BaseRegs.end(); I != E; ++I)
2128 RegUses.CountRegister(*I, LUIdx);
2131 /// InsertFormula - If the given formula has not yet been inserted, add it to
2132 /// the list, and return true. Return false otherwise.
2133 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2134 if (!LU.InsertFormula(F))
2137 CountRegisters(F, LUIdx);
2141 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2142 /// loop-invariant values which we're tracking. These other uses will pin these
2143 /// values in registers, making them less profitable for elimination.
2144 /// TODO: This currently misses non-constant addrec step registers.
2145 /// TODO: Should this give more weight to users inside the loop?
2147 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2148 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2149 SmallPtrSet<const SCEV *, 8> Inserted;
2151 while (!Worklist.empty()) {
2152 const SCEV *S = Worklist.pop_back_val();
2154 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2155 Worklist.append(N->op_begin(), N->op_end());
2156 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2157 Worklist.push_back(C->getOperand());
2158 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2159 Worklist.push_back(D->getLHS());
2160 Worklist.push_back(D->getRHS());
2161 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2162 if (!Inserted.insert(U)) continue;
2163 const Value *V = U->getValue();
2164 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2165 // Look for instructions defined outside the loop.
2166 if (L->contains(Inst)) continue;
2167 } else if (isa<UndefValue>(V))
2168 // Undef doesn't have a live range, so it doesn't matter.
2170 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2172 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2173 // Ignore non-instructions.
2176 // Ignore instructions in other functions (as can happen with
2178 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2180 // Ignore instructions not dominated by the loop.
2181 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2182 UserInst->getParent() :
2183 cast<PHINode>(UserInst)->getIncomingBlock(
2184 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2185 if (!DT.dominates(L->getHeader(), UseBB))
2187 // Ignore uses which are part of other SCEV expressions, to avoid
2188 // analyzing them multiple times.
2189 if (SE.isSCEVable(UserInst->getType())) {
2190 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2191 // If the user is a no-op, look through to its uses.
2192 if (!isa<SCEVUnknown>(UserS))
2196 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2200 // Ignore icmp instructions which are already being analyzed.
2201 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2202 unsigned OtherIdx = !UI.getOperandNo();
2203 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2204 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
2208 LSRFixup &LF = getNewFixup();
2209 LF.UserInst = const_cast<Instruction *>(UserInst);
2210 LF.OperandValToReplace = UI.getUse();
2211 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2213 LF.Offset = P.second;
2214 LSRUse &LU = Uses[LF.LUIdx];
2215 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2216 if (!LU.WidestFixupType ||
2217 SE.getTypeSizeInBits(LU.WidestFixupType) <
2218 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2219 LU.WidestFixupType = LF.OperandValToReplace->getType();
2220 InsertSupplementalFormula(U, LU, LF.LUIdx);
2221 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2228 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2229 /// separate registers. If C is non-null, multiply each subexpression by C.
2230 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2231 SmallVectorImpl<const SCEV *> &Ops,
2233 ScalarEvolution &SE) {
2234 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2235 // Break out add operands.
2236 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2238 CollectSubexprs(*I, C, Ops, L, SE);
2240 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2241 // Split a non-zero base out of an addrec.
2242 if (!AR->getStart()->isZero()) {
2243 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2244 AR->getStepRecurrence(SE),
2247 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2250 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2251 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2252 if (Mul->getNumOperands() == 2)
2253 if (const SCEVConstant *Op0 =
2254 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2255 CollectSubexprs(Mul->getOperand(1),
2256 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2262 // Otherwise use the value itself, optionally with a scale applied.
2263 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2266 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2268 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2271 // Arbitrarily cap recursion to protect compile time.
2272 if (Depth >= 3) return;
2274 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2275 const SCEV *BaseReg = Base.BaseRegs[i];
2277 SmallVector<const SCEV *, 8> AddOps;
2278 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2280 if (AddOps.size() == 1) continue;
2282 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2283 JE = AddOps.end(); J != JE; ++J) {
2285 // Loop-variant "unknown" values are uninteresting; we won't be able to
2286 // do anything meaningful with them.
2287 if (isa<SCEVUnknown>(*J) && !(*J)->isLoopInvariant(L))
2290 // Don't pull a constant into a register if the constant could be folded
2291 // into an immediate field.
2292 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2293 Base.getNumRegs() > 1,
2294 LU.Kind, LU.AccessTy, TLI, SE))
2297 // Collect all operands except *J.
2298 SmallVector<const SCEV *, 8> InnerAddOps
2299 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2301 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2303 // Don't leave just a constant behind in a register if the constant could
2304 // be folded into an immediate field.
2305 if (InnerAddOps.size() == 1 &&
2306 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2307 Base.getNumRegs() > 1,
2308 LU.Kind, LU.AccessTy, TLI, SE))
2311 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2312 if (InnerSum->isZero())
2315 F.BaseRegs[i] = InnerSum;
2316 F.BaseRegs.push_back(*J);
2317 if (InsertFormula(LU, LUIdx, F))
2318 // If that formula hadn't been seen before, recurse to find more like
2320 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2325 /// GenerateCombinations - Generate a formula consisting of all of the
2326 /// loop-dominating registers added into a single register.
2327 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2329 // This method is only interesting on a plurality of registers.
2330 if (Base.BaseRegs.size() <= 1) return;
2334 SmallVector<const SCEV *, 4> Ops;
2335 for (SmallVectorImpl<const SCEV *>::const_iterator
2336 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2337 const SCEV *BaseReg = *I;
2338 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2339 !BaseReg->hasComputableLoopEvolution(L))
2340 Ops.push_back(BaseReg);
2342 F.BaseRegs.push_back(BaseReg);
2344 if (Ops.size() > 1) {
2345 const SCEV *Sum = SE.getAddExpr(Ops);
2346 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2347 // opportunity to fold something. For now, just ignore such cases
2348 // rather than proceed with zero in a register.
2349 if (!Sum->isZero()) {
2350 F.BaseRegs.push_back(Sum);
2351 (void)InsertFormula(LU, LUIdx, F);
2356 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2357 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2359 // We can't add a symbolic offset if the address already contains one.
2360 if (Base.AM.BaseGV) return;
2362 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2363 const SCEV *G = Base.BaseRegs[i];
2364 GlobalValue *GV = ExtractSymbol(G, SE);
2365 if (G->isZero() || !GV)
2369 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2370 LU.Kind, LU.AccessTy, TLI))
2373 (void)InsertFormula(LU, LUIdx, F);
2377 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2378 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2380 // TODO: For now, just add the min and max offset, because it usually isn't
2381 // worthwhile looking at everything inbetween.
2382 SmallVector<int64_t, 2> Worklist;
2383 Worklist.push_back(LU.MinOffset);
2384 if (LU.MaxOffset != LU.MinOffset)
2385 Worklist.push_back(LU.MaxOffset);
2387 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2388 const SCEV *G = Base.BaseRegs[i];
2390 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2391 E = Worklist.end(); I != E; ++I) {
2393 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2394 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2395 LU.Kind, LU.AccessTy, TLI)) {
2396 // Add the offset to the base register.
2397 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2398 // If it cancelled out, drop the base register, otherwise update it.
2399 if (NewG->isZero()) {
2400 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2401 F.BaseRegs.pop_back();
2403 F.BaseRegs[i] = NewG;
2405 (void)InsertFormula(LU, LUIdx, F);
2409 int64_t Imm = ExtractImmediate(G, SE);
2410 if (G->isZero() || Imm == 0)
2413 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2414 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2415 LU.Kind, LU.AccessTy, TLI))
2418 (void)InsertFormula(LU, LUIdx, F);
2422 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2423 /// the comparison. For example, x == y -> x*c == y*c.
2424 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2426 if (LU.Kind != LSRUse::ICmpZero) return;
2428 // Determine the integer type for the base formula.
2429 const Type *IntTy = Base.getType();
2431 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2433 // Don't do this if there is more than one offset.
2434 if (LU.MinOffset != LU.MaxOffset) return;
2436 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2438 // Check each interesting stride.
2439 for (SmallSetVector<int64_t, 8>::const_iterator
2440 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2441 int64_t Factor = *I;
2443 // Check that the multiplication doesn't overflow.
2444 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2446 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2447 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2450 // Check that multiplying with the use offset doesn't overflow.
2451 int64_t Offset = LU.MinOffset;
2452 if (Offset == INT64_MIN && Factor == -1)
2454 Offset = (uint64_t)Offset * Factor;
2455 if (Offset / Factor != LU.MinOffset)
2459 F.AM.BaseOffs = NewBaseOffs;
2461 // Check that this scale is legal.
2462 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2465 // Compensate for the use having MinOffset built into it.
2466 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2468 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2470 // Check that multiplying with each base register doesn't overflow.
2471 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2472 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2473 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2477 // Check that multiplying with the scaled register doesn't overflow.
2479 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2480 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2484 // If we make it here and it's legal, add it.
2485 (void)InsertFormula(LU, LUIdx, F);
2490 /// GenerateScales - Generate stride factor reuse formulae by making use of
2491 /// scaled-offset address modes, for example.
2492 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2493 // Determine the integer type for the base formula.
2494 const Type *IntTy = Base.getType();
2497 // If this Formula already has a scaled register, we can't add another one.
2498 if (Base.AM.Scale != 0) return;
2500 // Check each interesting stride.
2501 for (SmallSetVector<int64_t, 8>::const_iterator
2502 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2503 int64_t Factor = *I;
2505 Base.AM.Scale = Factor;
2506 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2507 // Check whether this scale is going to be legal.
2508 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2509 LU.Kind, LU.AccessTy, TLI)) {
2510 // As a special-case, handle special out-of-loop Basic users specially.
2511 // TODO: Reconsider this special case.
2512 if (LU.Kind == LSRUse::Basic &&
2513 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2514 LSRUse::Special, LU.AccessTy, TLI) &&
2515 LU.AllFixupsOutsideLoop)
2516 LU.Kind = LSRUse::Special;
2520 // For an ICmpZero, negating a solitary base register won't lead to
2522 if (LU.Kind == LSRUse::ICmpZero &&
2523 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2525 // For each addrec base reg, apply the scale, if possible.
2526 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2527 if (const SCEVAddRecExpr *AR =
2528 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2529 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2530 if (FactorS->isZero())
2532 // Divide out the factor, ignoring high bits, since we'll be
2533 // scaling the value back up in the end.
2534 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2535 // TODO: This could be optimized to avoid all the copying.
2537 F.ScaledReg = Quotient;
2538 F.DeleteBaseReg(F.BaseRegs[i]);
2539 (void)InsertFormula(LU, LUIdx, F);
2545 /// GenerateTruncates - Generate reuse formulae from different IV types.
2546 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2547 // This requires TargetLowering to tell us which truncates are free.
2550 // Don't bother truncating symbolic values.
2551 if (Base.AM.BaseGV) return;
2553 // Determine the integer type for the base formula.
2554 const Type *DstTy = Base.getType();
2556 DstTy = SE.getEffectiveSCEVType(DstTy);
2558 for (SmallSetVector<const Type *, 4>::const_iterator
2559 I = Types.begin(), E = Types.end(); I != E; ++I) {
2560 const Type *SrcTy = *I;
2561 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2564 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2565 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2566 JE = F.BaseRegs.end(); J != JE; ++J)
2567 *J = SE.getAnyExtendExpr(*J, SrcTy);
2569 // TODO: This assumes we've done basic processing on all uses and
2570 // have an idea what the register usage is.
2571 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2574 (void)InsertFormula(LU, LUIdx, F);
2581 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2582 /// defer modifications so that the search phase doesn't have to worry about
2583 /// the data structures moving underneath it.
2587 const SCEV *OrigReg;
2589 WorkItem(size_t LI, int64_t I, const SCEV *R)
2590 : LUIdx(LI), Imm(I), OrigReg(R) {}
2592 void print(raw_ostream &OS) const;
2598 void WorkItem::print(raw_ostream &OS) const {
2599 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2600 << " , add offset " << Imm;
2603 void WorkItem::dump() const {
2604 print(errs()); errs() << '\n';
2607 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2608 /// distance apart and try to form reuse opportunities between them.
2609 void LSRInstance::GenerateCrossUseConstantOffsets() {
2610 // Group the registers by their value without any added constant offset.
2611 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2612 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2614 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2615 SmallVector<const SCEV *, 8> Sequence;
2616 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2618 const SCEV *Reg = *I;
2619 int64_t Imm = ExtractImmediate(Reg, SE);
2620 std::pair<RegMapTy::iterator, bool> Pair =
2621 Map.insert(std::make_pair(Reg, ImmMapTy()));
2623 Sequence.push_back(Reg);
2624 Pair.first->second.insert(std::make_pair(Imm, *I));
2625 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2628 // Now examine each set of registers with the same base value. Build up
2629 // a list of work to do and do the work in a separate step so that we're
2630 // not adding formulae and register counts while we're searching.
2631 SmallVector<WorkItem, 32> WorkItems;
2632 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2633 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2634 E = Sequence.end(); I != E; ++I) {
2635 const SCEV *Reg = *I;
2636 const ImmMapTy &Imms = Map.find(Reg)->second;
2638 // It's not worthwhile looking for reuse if there's only one offset.
2639 if (Imms.size() == 1)
2642 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2643 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2645 dbgs() << ' ' << J->first;
2648 // Examine each offset.
2649 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2651 const SCEV *OrigReg = J->second;
2653 int64_t JImm = J->first;
2654 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2656 if (!isa<SCEVConstant>(OrigReg) &&
2657 UsedByIndicesMap[Reg].count() == 1) {
2658 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2662 // Conservatively examine offsets between this orig reg a few selected
2664 ImmMapTy::const_iterator OtherImms[] = {
2665 Imms.begin(), prior(Imms.end()),
2666 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2668 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2669 ImmMapTy::const_iterator M = OtherImms[i];
2670 if (M == J || M == JE) continue;
2672 // Compute the difference between the two.
2673 int64_t Imm = (uint64_t)JImm - M->first;
2674 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2675 LUIdx = UsedByIndices.find_next(LUIdx))
2676 // Make a memo of this use, offset, and register tuple.
2677 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2678 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2685 UsedByIndicesMap.clear();
2686 UniqueItems.clear();
2688 // Now iterate through the worklist and add new formulae.
2689 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2690 E = WorkItems.end(); I != E; ++I) {
2691 const WorkItem &WI = *I;
2692 size_t LUIdx = WI.LUIdx;
2693 LSRUse &LU = Uses[LUIdx];
2694 int64_t Imm = WI.Imm;
2695 const SCEV *OrigReg = WI.OrigReg;
2697 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2698 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2699 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2701 // TODO: Use a more targeted data structure.
2702 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2703 const Formula &F = LU.Formulae[L];
2704 // Use the immediate in the scaled register.
2705 if (F.ScaledReg == OrigReg) {
2706 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2707 Imm * (uint64_t)F.AM.Scale;
2708 // Don't create 50 + reg(-50).
2709 if (F.referencesReg(SE.getSCEV(
2710 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2713 NewF.AM.BaseOffs = Offs;
2714 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2715 LU.Kind, LU.AccessTy, TLI))
2717 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2719 // If the new scale is a constant in a register, and adding the constant
2720 // value to the immediate would produce a value closer to zero than the
2721 // immediate itself, then the formula isn't worthwhile.
2722 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2723 if (C->getValue()->getValue().isNegative() !=
2724 (NewF.AM.BaseOffs < 0) &&
2725 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2726 .ule(abs64(NewF.AM.BaseOffs)))
2730 (void)InsertFormula(LU, LUIdx, NewF);
2732 // Use the immediate in a base register.
2733 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2734 const SCEV *BaseReg = F.BaseRegs[N];
2735 if (BaseReg != OrigReg)
2738 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2739 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2740 LU.Kind, LU.AccessTy, TLI))
2742 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2744 // If the new formula has a constant in a register, and adding the
2745 // constant value to the immediate would produce a value closer to
2746 // zero than the immediate itself, then the formula isn't worthwhile.
2747 for (SmallVectorImpl<const SCEV *>::const_iterator
2748 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2750 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2751 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2752 abs64(NewF.AM.BaseOffs)) &&
2753 (C->getValue()->getValue() +
2754 NewF.AM.BaseOffs).countTrailingZeros() >=
2755 CountTrailingZeros_64(NewF.AM.BaseOffs))
2759 (void)InsertFormula(LU, LUIdx, NewF);
2768 /// GenerateAllReuseFormulae - Generate formulae for each use.
2770 LSRInstance::GenerateAllReuseFormulae() {
2771 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2772 // queries are more precise.
2773 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2774 LSRUse &LU = Uses[LUIdx];
2775 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2776 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2777 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2778 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2780 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2781 LSRUse &LU = Uses[LUIdx];
2782 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2783 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2784 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2785 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2786 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2787 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2788 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2789 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2791 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2792 LSRUse &LU = Uses[LUIdx];
2793 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2794 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2797 GenerateCrossUseConstantOffsets();
2799 DEBUG(dbgs() << "\n"
2800 "After generating reuse formulae:\n";
2801 print_uses(dbgs()));
2804 /// If their are multiple formulae with the same set of registers used
2805 /// by other uses, pick the best one and delete the others.
2806 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2808 bool ChangedFormulae = false;
2811 // Collect the best formula for each unique set of shared registers. This
2812 // is reset for each use.
2813 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2815 BestFormulaeTy BestFormulae;
2817 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2818 LSRUse &LU = Uses[LUIdx];
2819 FormulaSorter Sorter(L, LU, SE, DT);
2820 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2823 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2824 FIdx != NumForms; ++FIdx) {
2825 Formula &F = LU.Formulae[FIdx];
2827 SmallVector<const SCEV *, 2> Key;
2828 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2829 JE = F.BaseRegs.end(); J != JE; ++J) {
2830 const SCEV *Reg = *J;
2831 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2835 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2836 Key.push_back(F.ScaledReg);
2837 // Unstable sort by host order ok, because this is only used for
2839 std::sort(Key.begin(), Key.end());
2841 std::pair<BestFormulaeTy::const_iterator, bool> P =
2842 BestFormulae.insert(std::make_pair(Key, FIdx));
2844 Formula &Best = LU.Formulae[P.first->second];
2845 if (Sorter.operator()(F, Best))
2847 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2849 " in favor of formula "; Best.print(dbgs());
2852 ChangedFormulae = true;
2854 LU.DeleteFormula(F);
2862 // Now that we've filtered out some formulae, recompute the Regs set.
2864 LU.RecomputeRegs(LUIdx, RegUses);
2866 // Reset this to prepare for the next use.
2867 BestFormulae.clear();
2870 DEBUG(if (ChangedFormulae) {
2872 "After filtering out undesirable candidates:\n";
2877 // This is a rough guess that seems to work fairly well.
2878 static const size_t ComplexityLimit = UINT16_MAX;
2880 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2881 /// solutions the solver might have to consider. It almost never considers
2882 /// this many solutions because it prune the search space, but the pruning
2883 /// isn't always sufficient.
2884 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2886 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2887 E = Uses.end(); I != E; ++I) {
2888 size_t FSize = I->Formulae.size();
2889 if (FSize >= ComplexityLimit) {
2890 Power = ComplexityLimit;
2894 if (Power >= ComplexityLimit)
2900 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2901 /// formulae to choose from, use some rough heuristics to prune down the number
2902 /// of formulae. This keeps the main solver from taking an extraordinary amount
2903 /// of time in some worst-case scenarios.
2904 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2905 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2906 DEBUG(dbgs() << "The search space is too complex.\n");
2908 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2909 "which use a superset of registers used by other "
2912 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2913 LSRUse &LU = Uses[LUIdx];
2915 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2916 Formula &F = LU.Formulae[i];
2917 // Look for a formula with a constant or GV in a register. If the use
2918 // also has a formula with that same value in an immediate field,
2919 // delete the one that uses a register.
2920 for (SmallVectorImpl<const SCEV *>::const_iterator
2921 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2922 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2924 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2925 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2926 (I - F.BaseRegs.begin()));
2927 if (LU.HasFormulaWithSameRegs(NewF)) {
2928 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2929 LU.DeleteFormula(F);
2935 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2936 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2939 NewF.AM.BaseGV = GV;
2940 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2941 (I - F.BaseRegs.begin()));
2942 if (LU.HasFormulaWithSameRegs(NewF)) {
2943 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2945 LU.DeleteFormula(F);
2956 LU.RecomputeRegs(LUIdx, RegUses);
2959 DEBUG(dbgs() << "After pre-selection:\n";
2960 print_uses(dbgs()));
2963 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2964 DEBUG(dbgs() << "The search space is too complex.\n");
2966 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2967 "separated by a constant offset will use the same "
2970 // This is especially useful for unrolled loops.
2972 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2973 LSRUse &LU = Uses[LUIdx];
2974 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2975 E = LU.Formulae.end(); I != E; ++I) {
2976 const Formula &F = *I;
2977 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2978 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2979 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2980 /*HasBaseReg=*/false,
2981 LU.Kind, LU.AccessTy)) {
2982 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
2985 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
2987 // Delete formulae from the new use which are no longer legal.
2989 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
2990 Formula &F = LUThatHas->Formulae[i];
2991 if (!isLegalUse(F.AM,
2992 LUThatHas->MinOffset, LUThatHas->MaxOffset,
2993 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
2994 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2996 LUThatHas->DeleteFormula(F);
3003 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3005 // Update the relocs to reference the new use.
3006 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3007 E = Fixups.end(); I != E; ++I) {
3008 LSRFixup &Fixup = *I;
3009 if (Fixup.LUIdx == LUIdx) {
3010 Fixup.LUIdx = LUThatHas - &Uses.front();
3011 Fixup.Offset += F.AM.BaseOffs;
3012 DEBUG(dbgs() << "New fixup has offset "
3013 << Fixup.Offset << '\n');
3015 if (Fixup.LUIdx == NumUses-1)
3016 Fixup.LUIdx = LUIdx;
3019 // Delete the old use.
3030 DEBUG(dbgs() << "After pre-selection:\n";
3031 print_uses(dbgs()));
3034 // With all other options exhausted, loop until the system is simple
3035 // enough to handle.
3036 SmallPtrSet<const SCEV *, 4> Taken;
3037 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3038 // Ok, we have too many of formulae on our hands to conveniently handle.
3039 // Use a rough heuristic to thin out the list.
3040 DEBUG(dbgs() << "The search space is too complex.\n");
3042 // Pick the register which is used by the most LSRUses, which is likely
3043 // to be a good reuse register candidate.
3044 const SCEV *Best = 0;
3045 unsigned BestNum = 0;
3046 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3048 const SCEV *Reg = *I;
3049 if (Taken.count(Reg))
3054 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3055 if (Count > BestNum) {
3062 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3063 << " will yield profitable reuse.\n");
3066 // In any use with formulae which references this register, delete formulae
3067 // which don't reference it.
3068 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3069 LSRUse &LU = Uses[LUIdx];
3070 if (!LU.Regs.count(Best)) continue;
3073 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3074 Formula &F = LU.Formulae[i];
3075 if (!F.referencesReg(Best)) {
3076 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3077 LU.DeleteFormula(F);
3081 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3087 LU.RecomputeRegs(LUIdx, RegUses);
3090 DEBUG(dbgs() << "After pre-selection:\n";
3091 print_uses(dbgs()));
3095 /// SolveRecurse - This is the recursive solver.
3096 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3098 SmallVectorImpl<const Formula *> &Workspace,
3099 const Cost &CurCost,
3100 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3101 DenseSet<const SCEV *> &VisitedRegs) const {
3104 // - use more aggressive filtering
3105 // - sort the formula so that the most profitable solutions are found first
3106 // - sort the uses too
3108 // - don't compute a cost, and then compare. compare while computing a cost
3110 // - track register sets with SmallBitVector
3112 const LSRUse &LU = Uses[Workspace.size()];
3114 // If this use references any register that's already a part of the
3115 // in-progress solution, consider it a requirement that a formula must
3116 // reference that register in order to be considered. This prunes out
3117 // unprofitable searching.
3118 SmallSetVector<const SCEV *, 4> ReqRegs;
3119 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3120 E = CurRegs.end(); I != E; ++I)
3121 if (LU.Regs.count(*I))
3124 bool AnySatisfiedReqRegs = false;
3125 SmallPtrSet<const SCEV *, 16> NewRegs;
3128 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3129 E = LU.Formulae.end(); I != E; ++I) {
3130 const Formula &F = *I;
3132 // Ignore formulae which do not use any of the required registers.
3133 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3134 JE = ReqRegs.end(); J != JE; ++J) {
3135 const SCEV *Reg = *J;
3136 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3137 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3141 AnySatisfiedReqRegs = true;
3143 // Evaluate the cost of the current formula. If it's already worse than
3144 // the current best, prune the search at that point.
3147 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3148 if (NewCost < SolutionCost) {
3149 Workspace.push_back(&F);
3150 if (Workspace.size() != Uses.size()) {
3151 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3152 NewRegs, VisitedRegs);
3153 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3154 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3156 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3157 dbgs() << ". Regs:";
3158 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3159 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3160 dbgs() << ' ' << **I;
3163 SolutionCost = NewCost;
3164 Solution = Workspace;
3166 Workspace.pop_back();
3171 // If none of the formulae had all of the required registers, relax the
3172 // constraint so that we don't exclude all formulae.
3173 if (!AnySatisfiedReqRegs) {
3174 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3180 /// Solve - Choose one formula from each use. Return the results in the given
3181 /// Solution vector.
3182 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3183 SmallVector<const Formula *, 8> Workspace;
3185 SolutionCost.Loose();
3187 SmallPtrSet<const SCEV *, 16> CurRegs;
3188 DenseSet<const SCEV *> VisitedRegs;
3189 Workspace.reserve(Uses.size());
3191 // SolveRecurse does all the work.
3192 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3193 CurRegs, VisitedRegs);
3195 // Ok, we've now made all our decisions.
3196 DEBUG(dbgs() << "\n"
3197 "The chosen solution requires "; SolutionCost.print(dbgs());
3199 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3201 Uses[i].print(dbgs());
3204 Solution[i]->print(dbgs());
3208 assert(Solution.size() == Uses.size() && "Malformed solution!");
3211 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3212 /// the dominator tree far as we can go while still being dominated by the
3213 /// input positions. This helps canonicalize the insert position, which
3214 /// encourages sharing.
3215 BasicBlock::iterator
3216 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3217 const SmallVectorImpl<Instruction *> &Inputs)
3220 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3221 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3224 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3225 if (!Rung) return IP;
3226 Rung = Rung->getIDom();
3227 if (!Rung) return IP;
3228 IDom = Rung->getBlock();
3230 // Don't climb into a loop though.
3231 const Loop *IDomLoop = LI.getLoopFor(IDom);
3232 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3233 if (IDomDepth <= IPLoopDepth &&
3234 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3238 bool AllDominate = true;
3239 Instruction *BetterPos = 0;
3240 Instruction *Tentative = IDom->getTerminator();
3241 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3242 E = Inputs.end(); I != E; ++I) {
3243 Instruction *Inst = *I;
3244 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3245 AllDominate = false;
3248 // Attempt to find an insert position in the middle of the block,
3249 // instead of at the end, so that it can be used for other expansions.
3250 if (IDom == Inst->getParent() &&
3251 (!BetterPos || DT.dominates(BetterPos, Inst)))
3252 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3265 /// AdjustInsertPositionForExpand - Determine an input position which will be
3266 /// dominated by the operands and which will dominate the result.
3267 BasicBlock::iterator
3268 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3270 const LSRUse &LU) const {
3271 // Collect some instructions which must be dominated by the
3272 // expanding replacement. These must be dominated by any operands that
3273 // will be required in the expansion.
3274 SmallVector<Instruction *, 4> Inputs;
3275 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3276 Inputs.push_back(I);
3277 if (LU.Kind == LSRUse::ICmpZero)
3278 if (Instruction *I =
3279 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3280 Inputs.push_back(I);
3281 if (LF.PostIncLoops.count(L)) {
3282 if (LF.isUseFullyOutsideLoop(L))
3283 Inputs.push_back(L->getLoopLatch()->getTerminator());
3285 Inputs.push_back(IVIncInsertPos);
3287 // The expansion must also be dominated by the increment positions of any
3288 // loops it for which it is using post-inc mode.
3289 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3290 E = LF.PostIncLoops.end(); I != E; ++I) {
3291 const Loop *PIL = *I;
3292 if (PIL == L) continue;
3294 // Be dominated by the loop exit.
3295 SmallVector<BasicBlock *, 4> ExitingBlocks;
3296 PIL->getExitingBlocks(ExitingBlocks);
3297 if (!ExitingBlocks.empty()) {
3298 BasicBlock *BB = ExitingBlocks[0];
3299 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3300 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3301 Inputs.push_back(BB->getTerminator());
3305 // Then, climb up the immediate dominator tree as far as we can go while
3306 // still being dominated by the input positions.
3307 IP = HoistInsertPosition(IP, Inputs);
3309 // Don't insert instructions before PHI nodes.
3310 while (isa<PHINode>(IP)) ++IP;
3312 // Ignore debug intrinsics.
3313 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3318 /// Expand - Emit instructions for the leading candidate expression for this
3319 /// LSRUse (this is called "expanding").
3320 Value *LSRInstance::Expand(const LSRFixup &LF,
3322 BasicBlock::iterator IP,
3323 SCEVExpander &Rewriter,
3324 SmallVectorImpl<WeakVH> &DeadInsts) const {
3325 const LSRUse &LU = Uses[LF.LUIdx];
3327 // Determine an input position which will be dominated by the operands and
3328 // which will dominate the result.
3329 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3331 // Inform the Rewriter if we have a post-increment use, so that it can
3332 // perform an advantageous expansion.
3333 Rewriter.setPostInc(LF.PostIncLoops);
3335 // This is the type that the user actually needs.
3336 const Type *OpTy = LF.OperandValToReplace->getType();
3337 // This will be the type that we'll initially expand to.
3338 const Type *Ty = F.getType();
3340 // No type known; just expand directly to the ultimate type.
3342 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3343 // Expand directly to the ultimate type if it's the right size.
3345 // This is the type to do integer arithmetic in.
3346 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3348 // Build up a list of operands to add together to form the full base.
3349 SmallVector<const SCEV *, 8> Ops;
3351 // Expand the BaseRegs portion.
3352 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3353 E = F.BaseRegs.end(); I != E; ++I) {
3354 const SCEV *Reg = *I;
3355 assert(!Reg->isZero() && "Zero allocated in a base register!");
3357 // If we're expanding for a post-inc user, make the post-inc adjustment.
3358 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3359 Reg = TransformForPostIncUse(Denormalize, Reg,
3360 LF.UserInst, LF.OperandValToReplace,
3363 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3366 // Flush the operand list to suppress SCEVExpander hoisting.
3368 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3370 Ops.push_back(SE.getUnknown(FullV));
3373 // Expand the ScaledReg portion.
3374 Value *ICmpScaledV = 0;
3375 if (F.AM.Scale != 0) {
3376 const SCEV *ScaledS = F.ScaledReg;
3378 // If we're expanding for a post-inc user, make the post-inc adjustment.
3379 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3380 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3381 LF.UserInst, LF.OperandValToReplace,
3384 if (LU.Kind == LSRUse::ICmpZero) {
3385 // An interesting way of "folding" with an icmp is to use a negated
3386 // scale, which we'll implement by inserting it into the other operand
3388 assert(F.AM.Scale == -1 &&
3389 "The only scale supported by ICmpZero uses is -1!");
3390 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3392 // Otherwise just expand the scaled register and an explicit scale,
3393 // which is expected to be matched as part of the address.
3394 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3395 ScaledS = SE.getMulExpr(ScaledS,
3396 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3397 Ops.push_back(ScaledS);
3399 // Flush the operand list to suppress SCEVExpander hoisting.
3400 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3402 Ops.push_back(SE.getUnknown(FullV));
3406 // Expand the GV portion.
3408 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3410 // Flush the operand list to suppress SCEVExpander hoisting.
3411 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3413 Ops.push_back(SE.getUnknown(FullV));
3416 // Expand the immediate portion.
3417 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3419 if (LU.Kind == LSRUse::ICmpZero) {
3420 // The other interesting way of "folding" with an ICmpZero is to use a
3421 // negated immediate.
3423 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3425 Ops.push_back(SE.getUnknown(ICmpScaledV));
3426 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3429 // Just add the immediate values. These again are expected to be matched
3430 // as part of the address.
3431 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3435 // Emit instructions summing all the operands.
3436 const SCEV *FullS = Ops.empty() ?
3437 SE.getConstant(IntTy, 0) :
3439 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3441 // We're done expanding now, so reset the rewriter.
3442 Rewriter.clearPostInc();
3444 // An ICmpZero Formula represents an ICmp which we're handling as a
3445 // comparison against zero. Now that we've expanded an expression for that
3446 // form, update the ICmp's other operand.
3447 if (LU.Kind == LSRUse::ICmpZero) {
3448 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3449 DeadInsts.push_back(CI->getOperand(1));
3450 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3451 "a scale at the same time!");
3452 if (F.AM.Scale == -1) {
3453 if (ICmpScaledV->getType() != OpTy) {
3455 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3457 ICmpScaledV, OpTy, "tmp", CI);
3460 CI->setOperand(1, ICmpScaledV);
3462 assert(F.AM.Scale == 0 &&
3463 "ICmp does not support folding a global value and "
3464 "a scale at the same time!");
3465 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3467 if (C->getType() != OpTy)
3468 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3472 CI->setOperand(1, C);
3479 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3480 /// of their operands effectively happens in their predecessor blocks, so the
3481 /// expression may need to be expanded in multiple places.
3482 void LSRInstance::RewriteForPHI(PHINode *PN,
3485 SCEVExpander &Rewriter,
3486 SmallVectorImpl<WeakVH> &DeadInsts,
3488 DenseMap<BasicBlock *, Value *> Inserted;
3489 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3490 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3491 BasicBlock *BB = PN->getIncomingBlock(i);
3493 // If this is a critical edge, split the edge so that we do not insert
3494 // the code on all predecessor/successor paths. We do this unless this
3495 // is the canonical backedge for this loop, which complicates post-inc
3497 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3498 !isa<IndirectBrInst>(BB->getTerminator()) &&
3499 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3500 // Split the critical edge.
3501 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3503 // If PN is outside of the loop and BB is in the loop, we want to
3504 // move the block to be immediately before the PHI block, not
3505 // immediately after BB.
3506 if (L->contains(BB) && !L->contains(PN))
3507 NewBB->moveBefore(PN->getParent());
3509 // Splitting the edge can reduce the number of PHI entries we have.
3510 e = PN->getNumIncomingValues();
3512 i = PN->getBasicBlockIndex(BB);
3515 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3516 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3518 PN->setIncomingValue(i, Pair.first->second);
3520 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3522 // If this is reuse-by-noop-cast, insert the noop cast.
3523 const Type *OpTy = LF.OperandValToReplace->getType();
3524 if (FullV->getType() != OpTy)
3526 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3528 FullV, LF.OperandValToReplace->getType(),
3529 "tmp", BB->getTerminator());
3531 PN->setIncomingValue(i, FullV);
3532 Pair.first->second = FullV;
3537 /// Rewrite - Emit instructions for the leading candidate expression for this
3538 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3539 /// the newly expanded value.
3540 void LSRInstance::Rewrite(const LSRFixup &LF,
3542 SCEVExpander &Rewriter,
3543 SmallVectorImpl<WeakVH> &DeadInsts,
3545 // First, find an insertion point that dominates UserInst. For PHI nodes,
3546 // find the nearest block which dominates all the relevant uses.
3547 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3548 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3550 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3552 // If this is reuse-by-noop-cast, insert the noop cast.
3553 const Type *OpTy = LF.OperandValToReplace->getType();
3554 if (FullV->getType() != OpTy) {
3556 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3557 FullV, OpTy, "tmp", LF.UserInst);
3561 // Update the user. ICmpZero is handled specially here (for now) because
3562 // Expand may have updated one of the operands of the icmp already, and
3563 // its new value may happen to be equal to LF.OperandValToReplace, in
3564 // which case doing replaceUsesOfWith leads to replacing both operands
3565 // with the same value. TODO: Reorganize this.
3566 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3567 LF.UserInst->setOperand(0, FullV);
3569 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3572 DeadInsts.push_back(LF.OperandValToReplace);
3575 /// ImplementSolution - Rewrite all the fixup locations with new values,
3576 /// following the chosen solution.
3578 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3580 // Keep track of instructions we may have made dead, so that
3581 // we can remove them after we are done working.
3582 SmallVector<WeakVH, 16> DeadInsts;
3584 SCEVExpander Rewriter(SE);
3585 Rewriter.disableCanonicalMode();
3586 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3588 // Expand the new value definitions and update the users.
3589 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3590 E = Fixups.end(); I != E; ++I) {
3591 const LSRFixup &Fixup = *I;
3593 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3598 // Clean up after ourselves. This must be done before deleting any
3602 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3605 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3606 : IU(P->getAnalysis<IVUsers>()),
3607 SE(P->getAnalysis<ScalarEvolution>()),
3608 DT(P->getAnalysis<DominatorTree>()),
3609 LI(P->getAnalysis<LoopInfo>()),
3610 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3612 // If LoopSimplify form is not available, stay out of trouble.
3613 if (!L->isLoopSimplifyForm()) return;
3615 // If there's no interesting work to be done, bail early.
3616 if (IU.empty()) return;
3618 DEBUG(dbgs() << "\nLSR on loop ";
3619 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3622 // First, perform some low-level loop optimizations.
3624 OptimizeLoopTermCond();
3626 // Start collecting data and preparing for the solver.
3627 CollectInterestingTypesAndFactors();
3628 CollectFixupsAndInitialFormulae();
3629 CollectLoopInvariantFixupsAndFormulae();
3631 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3632 print_uses(dbgs()));
3634 // Now use the reuse data to generate a bunch of interesting ways
3635 // to formulate the values needed for the uses.
3636 GenerateAllReuseFormulae();
3638 FilterOutUndesirableDedicatedRegisters();
3639 NarrowSearchSpaceUsingHeuristics();
3641 SmallVector<const Formula *, 8> Solution;
3644 // Release memory that is no longer needed.
3650 // Formulae should be legal.
3651 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3652 E = Uses.end(); I != E; ++I) {
3653 const LSRUse &LU = *I;
3654 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3655 JE = LU.Formulae.end(); J != JE; ++J)
3656 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3657 LU.Kind, LU.AccessTy, TLI) &&
3658 "Illegal formula generated!");
3662 // Now that we've decided what we want, make it so.
3663 ImplementSolution(Solution, P);
3666 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3667 if (Factors.empty() && Types.empty()) return;
3669 OS << "LSR has identified the following interesting factors and types: ";
3672 for (SmallSetVector<int64_t, 8>::const_iterator
3673 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3674 if (!First) OS << ", ";
3679 for (SmallSetVector<const Type *, 4>::const_iterator
3680 I = Types.begin(), E = Types.end(); I != E; ++I) {
3681 if (!First) OS << ", ";
3683 OS << '(' << **I << ')';
3688 void LSRInstance::print_fixups(raw_ostream &OS) const {
3689 OS << "LSR is examining the following fixup sites:\n";
3690 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3691 E = Fixups.end(); I != E; ++I) {
3698 void LSRInstance::print_uses(raw_ostream &OS) const {
3699 OS << "LSR is examining the following uses:\n";
3700 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3701 E = Uses.end(); I != E; ++I) {
3702 const LSRUse &LU = *I;
3706 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3707 JE = LU.Formulae.end(); J != JE; ++J) {
3715 void LSRInstance::print(raw_ostream &OS) const {
3716 print_factors_and_types(OS);
3721 void LSRInstance::dump() const {
3722 print(errs()); errs() << '\n';
3727 class LoopStrengthReduce : public LoopPass {
3728 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3729 /// transformation profitability.
3730 const TargetLowering *const TLI;
3733 static char ID; // Pass ID, replacement for typeid
3734 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3737 bool runOnLoop(Loop *L, LPPassManager &LPM);
3738 void getAnalysisUsage(AnalysisUsage &AU) const;
3743 char LoopStrengthReduce::ID = 0;
3744 INITIALIZE_PASS(LoopStrengthReduce, "loop-reduce",
3745 "Loop Strength Reduction", false, false);
3747 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3748 return new LoopStrengthReduce(TLI);
3751 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3752 : LoopPass(ID), TLI(tli) {}
3754 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3755 // We split critical edges, so we change the CFG. However, we do update
3756 // many analyses if they are around.
3757 AU.addPreservedID(LoopSimplifyID);
3758 AU.addPreserved("domfrontier");
3760 AU.addRequired<LoopInfo>();
3761 AU.addPreserved<LoopInfo>();
3762 AU.addRequiredID(LoopSimplifyID);
3763 AU.addRequired<DominatorTree>();
3764 AU.addPreserved<DominatorTree>();
3765 AU.addRequired<ScalarEvolution>();
3766 AU.addPreserved<ScalarEvolution>();
3767 AU.addRequired<IVUsers>();
3768 AU.addPreserved<IVUsers>();
3771 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3772 bool Changed = false;
3774 // Run the main LSR transformation.
3775 Changed |= LSRInstance(TLI, L, this).getChanged();
3777 // At this point, it is worth checking to see if any recurrence PHIs are also
3778 // dead, so that we can remove them as well.
3779 Changed |= DeleteDeadPHIs(L->getHeader());