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/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/ValueHandle.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Target/TargetLowering.h"
81 /// RegSortData - This class holds data which is used to order reuse candidates.
84 /// UsedByIndices - This represents the set of LSRUse indices which reference
85 /// a particular register.
86 SmallBitVector UsedByIndices;
90 void print(raw_ostream &OS) const;
96 void RegSortData::print(raw_ostream &OS) const {
97 OS << "[NumUses=" << UsedByIndices.count() << ']';
100 void RegSortData::dump() const {
101 print(errs()); errs() << '\n';
106 /// RegUseTracker - Map register candidates to information about how they are
108 class RegUseTracker {
109 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
111 RegUsesTy RegUsesMap;
112 SmallVector<const SCEV *, 16> RegSequence;
115 void CountRegister(const SCEV *Reg, size_t LUIdx);
116 void DropRegister(const SCEV *Reg, size_t LUIdx);
117 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
119 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
121 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
125 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
126 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
127 iterator begin() { return RegSequence.begin(); }
128 iterator end() { return RegSequence.end(); }
129 const_iterator begin() const { return RegSequence.begin(); }
130 const_iterator end() const { return RegSequence.end(); }
136 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
137 std::pair<RegUsesTy::iterator, bool> Pair =
138 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
139 RegSortData &RSD = Pair.first->second;
141 RegSequence.push_back(Reg);
142 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
143 RSD.UsedByIndices.set(LUIdx);
147 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
148 RegUsesTy::iterator It = RegUsesMap.find(Reg);
149 assert(It != RegUsesMap.end());
150 RegSortData &RSD = It->second;
151 assert(RSD.UsedByIndices.size() > LUIdx);
152 RSD.UsedByIndices.reset(LUIdx);
156 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
157 assert(LUIdx <= LastLUIdx);
159 // Update RegUses. The data structure is not optimized for this purpose;
160 // we must iterate through it and update each of the bit vectors.
161 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
163 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
164 if (LUIdx < UsedByIndices.size())
165 UsedByIndices[LUIdx] =
166 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
167 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
172 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
173 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
174 if (I == RegUsesMap.end())
176 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
177 int i = UsedByIndices.find_first();
178 if (i == -1) return false;
179 if ((size_t)i != LUIdx) return true;
180 return UsedByIndices.find_next(i) != -1;
183 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
184 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
185 assert(I != RegUsesMap.end() && "Unknown register!");
186 return I->second.UsedByIndices;
189 void RegUseTracker::clear() {
196 /// Formula - This class holds information that describes a formula for
197 /// computing satisfying a use. It may include broken-out immediates and scaled
200 /// AM - This is used to represent complex addressing, as well as other kinds
201 /// of interesting uses.
202 TargetLowering::AddrMode AM;
204 /// BaseRegs - The list of "base" registers for this use. When this is
205 /// non-empty, AM.HasBaseReg should be set to true.
206 SmallVector<const SCEV *, 2> BaseRegs;
208 /// ScaledReg - The 'scaled' register for this use. This should be non-null
209 /// when AM.Scale is not zero.
210 const SCEV *ScaledReg;
212 Formula() : ScaledReg(0) {}
214 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
216 unsigned getNumRegs() const;
217 const Type *getType() const;
219 void DeleteBaseReg(const SCEV *&S);
221 bool referencesReg(const SCEV *S) const;
222 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
223 const RegUseTracker &RegUses) const;
225 void print(raw_ostream &OS) const;
231 /// DoInitialMatch - Recursion helper for InitialMatch.
232 static void DoInitialMatch(const SCEV *S, Loop *L,
233 SmallVectorImpl<const SCEV *> &Good,
234 SmallVectorImpl<const SCEV *> &Bad,
235 ScalarEvolution &SE) {
236 // Collect expressions which properly dominate the loop header.
237 if (SE.properlyDominates(S, L->getHeader())) {
242 // Look at add operands.
243 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
244 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
246 DoInitialMatch(*I, L, Good, Bad, SE);
250 // Look at addrec operands.
251 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
252 if (!AR->getStart()->isZero()) {
253 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
254 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
255 AR->getStepRecurrence(SE),
256 // FIXME: AR->getNoWrapFlags()
257 AR->getLoop(), SCEV::FlagAnyWrap),
262 // Handle a multiplication by -1 (negation) if it didn't fold.
263 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
264 if (Mul->getOperand(0)->isAllOnesValue()) {
265 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
266 const SCEV *NewMul = SE.getMulExpr(Ops);
268 SmallVector<const SCEV *, 4> MyGood;
269 SmallVector<const SCEV *, 4> MyBad;
270 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
271 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
272 SE.getEffectiveSCEVType(NewMul->getType())));
273 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
274 E = MyGood.end(); I != E; ++I)
275 Good.push_back(SE.getMulExpr(NegOne, *I));
276 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
277 E = MyBad.end(); I != E; ++I)
278 Bad.push_back(SE.getMulExpr(NegOne, *I));
282 // Ok, we can't do anything interesting. Just stuff the whole thing into a
283 // register and hope for the best.
287 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
288 /// attempting to keep all loop-invariant and loop-computable values in a
289 /// single base register.
290 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
291 SmallVector<const SCEV *, 4> Good;
292 SmallVector<const SCEV *, 4> Bad;
293 DoInitialMatch(S, L, Good, Bad, SE);
295 const SCEV *Sum = SE.getAddExpr(Good);
297 BaseRegs.push_back(Sum);
298 AM.HasBaseReg = true;
301 const SCEV *Sum = SE.getAddExpr(Bad);
303 BaseRegs.push_back(Sum);
304 AM.HasBaseReg = true;
308 /// getNumRegs - Return the total number of register operands used by this
309 /// formula. This does not include register uses implied by non-constant
311 unsigned Formula::getNumRegs() const {
312 return !!ScaledReg + BaseRegs.size();
315 /// getType - Return the type of this formula, if it has one, or null
316 /// otherwise. This type is meaningless except for the bit size.
317 const Type *Formula::getType() const {
318 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
319 ScaledReg ? ScaledReg->getType() :
320 AM.BaseGV ? AM.BaseGV->getType() :
324 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
325 void Formula::DeleteBaseReg(const SCEV *&S) {
326 if (&S != &BaseRegs.back())
327 std::swap(S, BaseRegs.back());
331 /// referencesReg - Test if this formula references the given register.
332 bool Formula::referencesReg(const SCEV *S) const {
333 return S == ScaledReg ||
334 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
337 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
338 /// which are used by uses other than the use with the given index.
339 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
340 const RegUseTracker &RegUses) const {
342 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
344 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
345 E = BaseRegs.end(); I != E; ++I)
346 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
351 void Formula::print(raw_ostream &OS) const {
354 if (!First) OS << " + "; else First = false;
355 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
357 if (AM.BaseOffs != 0) {
358 if (!First) OS << " + "; else First = false;
361 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
362 E = BaseRegs.end(); I != E; ++I) {
363 if (!First) OS << " + "; else First = false;
364 OS << "reg(" << **I << ')';
366 if (AM.HasBaseReg && BaseRegs.empty()) {
367 if (!First) OS << " + "; else First = false;
368 OS << "**error: HasBaseReg**";
369 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
370 if (!First) OS << " + "; else First = false;
371 OS << "**error: !HasBaseReg**";
374 if (!First) OS << " + "; else First = false;
375 OS << AM.Scale << "*reg(";
384 void Formula::dump() const {
385 print(errs()); errs() << '\n';
388 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
389 /// without changing its value.
390 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
392 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
393 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
396 /// isAddSExtable - Return true if the given add can be sign-extended
397 /// without changing its value.
398 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
400 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
401 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
404 /// isMulSExtable - Return true if the given mul can be sign-extended
405 /// without changing its value.
406 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
408 IntegerType::get(SE.getContext(),
409 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
410 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
413 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
414 /// and if the remainder is known to be zero, or null otherwise. If
415 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
416 /// to Y, ignoring that the multiplication may overflow, which is useful when
417 /// the result will be used in a context where the most significant bits are
419 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
421 bool IgnoreSignificantBits = false) {
422 // Handle the trivial case, which works for any SCEV type.
424 return SE.getConstant(LHS->getType(), 1);
426 // Handle a few RHS special cases.
427 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
429 const APInt &RA = RC->getValue()->getValue();
430 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
432 if (RA.isAllOnesValue())
433 return SE.getMulExpr(LHS, RC);
434 // Handle x /s 1 as x.
439 // Check for a division of a constant by a constant.
440 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
443 const APInt &LA = C->getValue()->getValue();
444 const APInt &RA = RC->getValue()->getValue();
445 if (LA.srem(RA) != 0)
447 return SE.getConstant(LA.sdiv(RA));
450 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
451 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
452 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
453 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
454 IgnoreSignificantBits);
456 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
457 IgnoreSignificantBits);
458 if (!Start) return 0;
459 // FlagNW is independent of the start value, step direction, and is
460 // preserved with smaller magnitude steps.
461 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
462 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
467 // Distribute the sdiv over add operands, if the add doesn't overflow.
468 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
469 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
470 SmallVector<const SCEV *, 8> Ops;
471 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
473 const SCEV *Op = getExactSDiv(*I, RHS, SE,
474 IgnoreSignificantBits);
478 return SE.getAddExpr(Ops);
483 // Check for a multiply operand that we can pull RHS out of.
484 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
485 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
486 SmallVector<const SCEV *, 4> Ops;
488 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
492 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
493 IgnoreSignificantBits)) {
499 return Found ? SE.getMulExpr(Ops) : 0;
504 // Otherwise we don't know.
508 /// ExtractImmediate - If S involves the addition of a constant integer value,
509 /// return that integer value, and mutate S to point to a new SCEV with that
511 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
512 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
513 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
514 S = SE.getConstant(C->getType(), 0);
515 return C->getValue()->getSExtValue();
517 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
518 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
519 int64_t Result = ExtractImmediate(NewOps.front(), SE);
521 S = SE.getAddExpr(NewOps);
523 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
524 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
525 int64_t Result = ExtractImmediate(NewOps.front(), SE);
527 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
528 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
535 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
536 /// return that symbol, and mutate S to point to a new SCEV with that
538 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
539 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
540 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
541 S = SE.getConstant(GV->getType(), 0);
544 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
545 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
546 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
548 S = SE.getAddExpr(NewOps);
550 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
551 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
552 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
554 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
555 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
562 /// isAddressUse - Returns true if the specified instruction is using the
563 /// specified value as an address.
564 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
565 bool isAddress = isa<LoadInst>(Inst);
566 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
567 if (SI->getOperand(1) == OperandVal)
569 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
570 // Addressing modes can also be folded into prefetches and a variety
572 switch (II->getIntrinsicID()) {
574 case Intrinsic::prefetch:
575 case Intrinsic::x86_sse2_loadu_dq:
576 case Intrinsic::x86_sse2_loadu_pd:
577 case Intrinsic::x86_sse_loadu_ps:
578 case Intrinsic::x86_sse_storeu_ps:
579 case Intrinsic::x86_sse2_storeu_pd:
580 case Intrinsic::x86_sse2_storeu_dq:
581 case Intrinsic::x86_sse2_storel_dq:
582 if (II->getArgOperand(0) == OperandVal)
590 /// getAccessType - Return the type of the memory being accessed.
591 static const Type *getAccessType(const Instruction *Inst) {
592 const Type *AccessTy = Inst->getType();
593 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
594 AccessTy = SI->getOperand(0)->getType();
595 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
596 // Addressing modes can also be folded into prefetches and a variety
598 switch (II->getIntrinsicID()) {
600 case Intrinsic::x86_sse_storeu_ps:
601 case Intrinsic::x86_sse2_storeu_pd:
602 case Intrinsic::x86_sse2_storeu_dq:
603 case Intrinsic::x86_sse2_storel_dq:
604 AccessTy = II->getArgOperand(0)->getType();
609 // All pointers have the same requirements, so canonicalize them to an
610 // arbitrary pointer type to minimize variation.
611 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
612 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
613 PTy->getAddressSpace());
618 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
619 /// specified set are trivially dead, delete them and see if this makes any of
620 /// their operands subsequently dead.
622 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
623 bool Changed = false;
625 while (!DeadInsts.empty()) {
626 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
628 if (I == 0 || !isInstructionTriviallyDead(I))
631 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
632 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
635 DeadInsts.push_back(U);
638 I->eraseFromParent();
647 /// Cost - This class is used to measure and compare candidate formulae.
649 /// TODO: Some of these could be merged. Also, a lexical ordering
650 /// isn't always optimal.
654 unsigned NumBaseAdds;
660 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
663 bool operator<(const Cost &Other) const;
667 void RateFormula(const Formula &F,
668 SmallPtrSet<const SCEV *, 16> &Regs,
669 const DenseSet<const SCEV *> &VisitedRegs,
671 const SmallVectorImpl<int64_t> &Offsets,
672 ScalarEvolution &SE, DominatorTree &DT);
674 void print(raw_ostream &OS) const;
678 void RateRegister(const SCEV *Reg,
679 SmallPtrSet<const SCEV *, 16> &Regs,
681 ScalarEvolution &SE, DominatorTree &DT);
682 void RatePrimaryRegister(const SCEV *Reg,
683 SmallPtrSet<const SCEV *, 16> &Regs,
685 ScalarEvolution &SE, DominatorTree &DT);
690 /// RateRegister - Tally up interesting quantities from the given register.
691 void Cost::RateRegister(const SCEV *Reg,
692 SmallPtrSet<const SCEV *, 16> &Regs,
694 ScalarEvolution &SE, DominatorTree &DT) {
695 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
696 if (AR->getLoop() == L)
697 AddRecCost += 1; /// TODO: This should be a function of the stride.
699 // If this is an addrec for a loop that's already been visited by LSR,
700 // don't second-guess its addrec phi nodes. LSR isn't currently smart
701 // enough to reason about more than one loop at a time. Consider these
702 // registers free and leave them alone.
703 else if (L->contains(AR->getLoop()) ||
704 (!AR->getLoop()->contains(L) &&
705 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
706 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
707 PHINode *PN = dyn_cast<PHINode>(I); ++I)
708 if (SE.isSCEVable(PN->getType()) &&
709 (SE.getEffectiveSCEVType(PN->getType()) ==
710 SE.getEffectiveSCEVType(AR->getType())) &&
711 SE.getSCEV(PN) == AR)
714 // If this isn't one of the addrecs that the loop already has, it
715 // would require a costly new phi and add. TODO: This isn't
716 // precisely modeled right now.
718 if (!Regs.count(AR->getStart()))
719 RateRegister(AR->getStart(), Regs, L, SE, DT);
722 // Add the step value register, if it needs one.
723 // TODO: The non-affine case isn't precisely modeled here.
724 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
725 if (!Regs.count(AR->getStart()))
726 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
730 // Rough heuristic; favor registers which don't require extra setup
731 // instructions in the preheader.
732 if (!isa<SCEVUnknown>(Reg) &&
733 !isa<SCEVConstant>(Reg) &&
734 !(isa<SCEVAddRecExpr>(Reg) &&
735 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
736 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
739 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
740 SE.hasComputableLoopEvolution(Reg, L);
743 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
745 void Cost::RatePrimaryRegister(const SCEV *Reg,
746 SmallPtrSet<const SCEV *, 16> &Regs,
748 ScalarEvolution &SE, DominatorTree &DT) {
749 if (Regs.insert(Reg))
750 RateRegister(Reg, Regs, L, SE, DT);
753 void Cost::RateFormula(const Formula &F,
754 SmallPtrSet<const SCEV *, 16> &Regs,
755 const DenseSet<const SCEV *> &VisitedRegs,
757 const SmallVectorImpl<int64_t> &Offsets,
758 ScalarEvolution &SE, DominatorTree &DT) {
759 // Tally up the registers.
760 if (const SCEV *ScaledReg = F.ScaledReg) {
761 if (VisitedRegs.count(ScaledReg)) {
765 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
767 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
768 E = F.BaseRegs.end(); I != E; ++I) {
769 const SCEV *BaseReg = *I;
770 if (VisitedRegs.count(BaseReg)) {
774 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
777 if (F.BaseRegs.size() > 1)
778 NumBaseAdds += F.BaseRegs.size() - 1;
780 // Tally up the non-zero immediates.
781 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
782 E = Offsets.end(); I != E; ++I) {
783 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
785 ImmCost += 64; // Handle symbolic values conservatively.
786 // TODO: This should probably be the pointer size.
787 else if (Offset != 0)
788 ImmCost += APInt(64, Offset, true).getMinSignedBits();
792 /// Loose - Set this cost to a loosing value.
802 /// operator< - Choose the lower cost.
803 bool Cost::operator<(const Cost &Other) const {
804 if (NumRegs != Other.NumRegs)
805 return NumRegs < Other.NumRegs;
806 if (AddRecCost != Other.AddRecCost)
807 return AddRecCost < Other.AddRecCost;
808 if (NumIVMuls != Other.NumIVMuls)
809 return NumIVMuls < Other.NumIVMuls;
810 if (NumBaseAdds != Other.NumBaseAdds)
811 return NumBaseAdds < Other.NumBaseAdds;
812 if (ImmCost != Other.ImmCost)
813 return ImmCost < Other.ImmCost;
814 if (SetupCost != Other.SetupCost)
815 return SetupCost < Other.SetupCost;
819 void Cost::print(raw_ostream &OS) const {
820 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
822 OS << ", with addrec cost " << AddRecCost;
824 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
825 if (NumBaseAdds != 0)
826 OS << ", plus " << NumBaseAdds << " base add"
827 << (NumBaseAdds == 1 ? "" : "s");
829 OS << ", plus " << ImmCost << " imm cost";
831 OS << ", plus " << SetupCost << " setup cost";
834 void Cost::dump() const {
835 print(errs()); errs() << '\n';
840 /// LSRFixup - An operand value in an instruction which is to be replaced
841 /// with some equivalent, possibly strength-reduced, replacement.
843 /// UserInst - The instruction which will be updated.
844 Instruction *UserInst;
846 /// OperandValToReplace - The operand of the instruction which will
847 /// be replaced. The operand may be used more than once; every instance
848 /// will be replaced.
849 Value *OperandValToReplace;
851 /// PostIncLoops - If this user is to use the post-incremented value of an
852 /// induction variable, this variable is non-null and holds the loop
853 /// associated with the induction variable.
854 PostIncLoopSet PostIncLoops;
856 /// LUIdx - The index of the LSRUse describing the expression which
857 /// this fixup needs, minus an offset (below).
860 /// Offset - A constant offset to be added to the LSRUse expression.
861 /// This allows multiple fixups to share the same LSRUse with different
862 /// offsets, for example in an unrolled loop.
865 bool isUseFullyOutsideLoop(const Loop *L) const;
869 void print(raw_ostream &OS) const;
876 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
878 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
879 /// value outside of the given loop.
880 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
881 // PHI nodes use their value in their incoming blocks.
882 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
883 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
884 if (PN->getIncomingValue(i) == OperandValToReplace &&
885 L->contains(PN->getIncomingBlock(i)))
890 return !L->contains(UserInst);
893 void LSRFixup::print(raw_ostream &OS) const {
895 // Store is common and interesting enough to be worth special-casing.
896 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
898 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
899 } else if (UserInst->getType()->isVoidTy())
900 OS << UserInst->getOpcodeName();
902 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
904 OS << ", OperandValToReplace=";
905 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
907 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
908 E = PostIncLoops.end(); I != E; ++I) {
909 OS << ", PostIncLoop=";
910 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
913 if (LUIdx != ~size_t(0))
914 OS << ", LUIdx=" << LUIdx;
917 OS << ", Offset=" << Offset;
920 void LSRFixup::dump() const {
921 print(errs()); errs() << '\n';
926 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
927 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
928 struct UniquifierDenseMapInfo {
929 static SmallVector<const SCEV *, 2> getEmptyKey() {
930 SmallVector<const SCEV *, 2> V;
931 V.push_back(reinterpret_cast<const SCEV *>(-1));
935 static SmallVector<const SCEV *, 2> getTombstoneKey() {
936 SmallVector<const SCEV *, 2> V;
937 V.push_back(reinterpret_cast<const SCEV *>(-2));
941 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
943 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
944 E = V.end(); I != E; ++I)
945 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
949 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
950 const SmallVector<const SCEV *, 2> &RHS) {
955 /// LSRUse - This class holds the state that LSR keeps for each use in
956 /// IVUsers, as well as uses invented by LSR itself. It includes information
957 /// about what kinds of things can be folded into the user, information about
958 /// the user itself, and information about how the use may be satisfied.
959 /// TODO: Represent multiple users of the same expression in common?
961 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
964 /// KindType - An enum for a kind of use, indicating what types of
965 /// scaled and immediate operands it might support.
967 Basic, ///< A normal use, with no folding.
968 Special, ///< A special case of basic, allowing -1 scales.
969 Address, ///< An address use; folding according to TargetLowering
970 ICmpZero ///< An equality icmp with both operands folded into one.
971 // TODO: Add a generic icmp too?
975 const Type *AccessTy;
977 SmallVector<int64_t, 8> Offsets;
981 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
982 /// LSRUse are outside of the loop, in which case some special-case heuristics
984 bool AllFixupsOutsideLoop;
986 /// WidestFixupType - This records the widest use type for any fixup using
987 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
988 /// max fixup widths to be equivalent, because the narrower one may be relying
989 /// on the implicit truncation to truncate away bogus bits.
990 const Type *WidestFixupType;
992 /// Formulae - A list of ways to build a value that can satisfy this user.
993 /// After the list is populated, one of these is selected heuristically and
994 /// used to formulate a replacement for OperandValToReplace in UserInst.
995 SmallVector<Formula, 12> Formulae;
997 /// Regs - The set of register candidates used by all formulae in this LSRUse.
998 SmallPtrSet<const SCEV *, 4> Regs;
1000 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
1001 MinOffset(INT64_MAX),
1002 MaxOffset(INT64_MIN),
1003 AllFixupsOutsideLoop(true),
1004 WidestFixupType(0) {}
1006 bool HasFormulaWithSameRegs(const Formula &F) const;
1007 bool InsertFormula(const Formula &F);
1008 void DeleteFormula(Formula &F);
1009 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1011 void print(raw_ostream &OS) const;
1017 /// HasFormula - Test whether this use as a formula which has the same
1018 /// registers as the given formula.
1019 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1020 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1021 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1022 // Unstable sort by host order ok, because this is only used for uniquifying.
1023 std::sort(Key.begin(), Key.end());
1024 return Uniquifier.count(Key);
1027 /// InsertFormula - If the given formula has not yet been inserted, add it to
1028 /// the list, and return true. Return false otherwise.
1029 bool LSRUse::InsertFormula(const Formula &F) {
1030 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1031 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1032 // Unstable sort by host order ok, because this is only used for uniquifying.
1033 std::sort(Key.begin(), Key.end());
1035 if (!Uniquifier.insert(Key).second)
1038 // Using a register to hold the value of 0 is not profitable.
1039 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1040 "Zero allocated in a scaled register!");
1042 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1043 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1044 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1047 // Add the formula to the list.
1048 Formulae.push_back(F);
1050 // Record registers now being used by this use.
1051 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1052 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1057 /// DeleteFormula - Remove the given formula from this use's list.
1058 void LSRUse::DeleteFormula(Formula &F) {
1059 if (&F != &Formulae.back())
1060 std::swap(F, Formulae.back());
1061 Formulae.pop_back();
1062 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1065 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1066 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1067 // Now that we've filtered out some formulae, recompute the Regs set.
1068 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1070 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1071 E = Formulae.end(); I != E; ++I) {
1072 const Formula &F = *I;
1073 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1074 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1077 // Update the RegTracker.
1078 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1079 E = OldRegs.end(); I != E; ++I)
1080 if (!Regs.count(*I))
1081 RegUses.DropRegister(*I, LUIdx);
1084 void LSRUse::print(raw_ostream &OS) const {
1085 OS << "LSR Use: Kind=";
1087 case Basic: OS << "Basic"; break;
1088 case Special: OS << "Special"; break;
1089 case ICmpZero: OS << "ICmpZero"; break;
1091 OS << "Address of ";
1092 if (AccessTy->isPointerTy())
1093 OS << "pointer"; // the full pointer type could be really verbose
1098 OS << ", Offsets={";
1099 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1100 E = Offsets.end(); I != E; ++I) {
1102 if (llvm::next(I) != E)
1107 if (AllFixupsOutsideLoop)
1108 OS << ", all-fixups-outside-loop";
1110 if (WidestFixupType)
1111 OS << ", widest fixup type: " << *WidestFixupType;
1114 void LSRUse::dump() const {
1115 print(errs()); errs() << '\n';
1118 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1119 /// be completely folded into the user instruction at isel time. This includes
1120 /// address-mode folding and special icmp tricks.
1121 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1122 LSRUse::KindType Kind, const Type *AccessTy,
1123 const TargetLowering *TLI) {
1125 case LSRUse::Address:
1126 // If we have low-level target information, ask the target if it can
1127 // completely fold this address.
1128 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1130 // Otherwise, just guess that reg+reg addressing is legal.
1131 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1133 case LSRUse::ICmpZero:
1134 // There's not even a target hook for querying whether it would be legal to
1135 // fold a GV into an ICmp.
1139 // ICmp only has two operands; don't allow more than two non-trivial parts.
1140 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1143 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1144 // putting the scaled register in the other operand of the icmp.
1145 if (AM.Scale != 0 && AM.Scale != -1)
1148 // If we have low-level target information, ask the target if it can fold an
1149 // integer immediate on an icmp.
1150 if (AM.BaseOffs != 0) {
1151 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1158 // Only handle single-register values.
1159 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1161 case LSRUse::Special:
1162 // Only handle -1 scales, or no scale.
1163 return AM.Scale == 0 || AM.Scale == -1;
1169 static bool isLegalUse(TargetLowering::AddrMode AM,
1170 int64_t MinOffset, int64_t MaxOffset,
1171 LSRUse::KindType Kind, const Type *AccessTy,
1172 const TargetLowering *TLI) {
1173 // Check for overflow.
1174 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1177 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1178 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1179 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1180 // Check for overflow.
1181 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1184 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1185 return isLegalUse(AM, Kind, AccessTy, TLI);
1190 static bool isAlwaysFoldable(int64_t BaseOffs,
1191 GlobalValue *BaseGV,
1193 LSRUse::KindType Kind, const Type *AccessTy,
1194 const TargetLowering *TLI) {
1195 // Fast-path: zero is always foldable.
1196 if (BaseOffs == 0 && !BaseGV) return true;
1198 // Conservatively, create an address with an immediate and a
1199 // base and a scale.
1200 TargetLowering::AddrMode AM;
1201 AM.BaseOffs = BaseOffs;
1203 AM.HasBaseReg = HasBaseReg;
1204 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1206 // Canonicalize a scale of 1 to a base register if the formula doesn't
1207 // already have a base register.
1208 if (!AM.HasBaseReg && AM.Scale == 1) {
1210 AM.HasBaseReg = true;
1213 return isLegalUse(AM, Kind, AccessTy, TLI);
1216 static bool isAlwaysFoldable(const SCEV *S,
1217 int64_t MinOffset, int64_t MaxOffset,
1219 LSRUse::KindType Kind, const Type *AccessTy,
1220 const TargetLowering *TLI,
1221 ScalarEvolution &SE) {
1222 // Fast-path: zero is always foldable.
1223 if (S->isZero()) return true;
1225 // Conservatively, create an address with an immediate and a
1226 // base and a scale.
1227 int64_t BaseOffs = ExtractImmediate(S, SE);
1228 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1230 // If there's anything else involved, it's not foldable.
1231 if (!S->isZero()) return false;
1233 // Fast-path: zero is always foldable.
1234 if (BaseOffs == 0 && !BaseGV) return true;
1236 // Conservatively, create an address with an immediate and a
1237 // base and a scale.
1238 TargetLowering::AddrMode AM;
1239 AM.BaseOffs = BaseOffs;
1241 AM.HasBaseReg = HasBaseReg;
1242 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1244 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1249 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1250 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1251 struct UseMapDenseMapInfo {
1252 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1253 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1256 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1257 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1261 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1262 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1263 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1267 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1268 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1273 /// LSRInstance - This class holds state for the main loop strength reduction
1277 ScalarEvolution &SE;
1280 const TargetLowering *const TLI;
1284 /// IVIncInsertPos - This is the insert position that the current loop's
1285 /// induction variable increment should be placed. In simple loops, this is
1286 /// the latch block's terminator. But in more complicated cases, this is a
1287 /// position which will dominate all the in-loop post-increment users.
1288 Instruction *IVIncInsertPos;
1290 /// Factors - Interesting factors between use strides.
1291 SmallSetVector<int64_t, 8> Factors;
1293 /// Types - Interesting use types, to facilitate truncation reuse.
1294 SmallSetVector<const Type *, 4> Types;
1296 /// Fixups - The list of operands which are to be replaced.
1297 SmallVector<LSRFixup, 16> Fixups;
1299 /// Uses - The list of interesting uses.
1300 SmallVector<LSRUse, 16> Uses;
1302 /// RegUses - Track which uses use which register candidates.
1303 RegUseTracker RegUses;
1305 void OptimizeShadowIV();
1306 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1307 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1308 void OptimizeLoopTermCond();
1310 void CollectInterestingTypesAndFactors();
1311 void CollectFixupsAndInitialFormulae();
1313 LSRFixup &getNewFixup() {
1314 Fixups.push_back(LSRFixup());
1315 return Fixups.back();
1318 // Support for sharing of LSRUses between LSRFixups.
1319 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1321 UseMapDenseMapInfo> UseMapTy;
1324 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1325 LSRUse::KindType Kind, const Type *AccessTy);
1327 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1328 LSRUse::KindType Kind,
1329 const Type *AccessTy);
1331 void DeleteUse(LSRUse &LU, size_t LUIdx);
1333 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1336 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1337 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1338 void CountRegisters(const Formula &F, size_t LUIdx);
1339 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1341 void CollectLoopInvariantFixupsAndFormulae();
1343 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1344 unsigned Depth = 0);
1345 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1346 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1347 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1348 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1349 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1350 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1351 void GenerateCrossUseConstantOffsets();
1352 void GenerateAllReuseFormulae();
1354 void FilterOutUndesirableDedicatedRegisters();
1356 size_t EstimateSearchSpaceComplexity() const;
1357 void NarrowSearchSpaceByDetectingSupersets();
1358 void NarrowSearchSpaceByCollapsingUnrolledCode();
1359 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1360 void NarrowSearchSpaceByPickingWinnerRegs();
1361 void NarrowSearchSpaceUsingHeuristics();
1363 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1365 SmallVectorImpl<const Formula *> &Workspace,
1366 const Cost &CurCost,
1367 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1368 DenseSet<const SCEV *> &VisitedRegs) const;
1369 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1371 BasicBlock::iterator
1372 HoistInsertPosition(BasicBlock::iterator IP,
1373 const SmallVectorImpl<Instruction *> &Inputs) const;
1374 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1376 const LSRUse &LU) const;
1378 Value *Expand(const LSRFixup &LF,
1380 BasicBlock::iterator IP,
1381 SCEVExpander &Rewriter,
1382 SmallVectorImpl<WeakVH> &DeadInsts) const;
1383 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1385 SCEVExpander &Rewriter,
1386 SmallVectorImpl<WeakVH> &DeadInsts,
1388 void Rewrite(const LSRFixup &LF,
1390 SCEVExpander &Rewriter,
1391 SmallVectorImpl<WeakVH> &DeadInsts,
1393 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1396 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1398 bool getChanged() const { return Changed; }
1400 void print_factors_and_types(raw_ostream &OS) const;
1401 void print_fixups(raw_ostream &OS) const;
1402 void print_uses(raw_ostream &OS) const;
1403 void print(raw_ostream &OS) const;
1409 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1410 /// inside the loop then try to eliminate the cast operation.
1411 void LSRInstance::OptimizeShadowIV() {
1412 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1413 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1416 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1417 UI != E; /* empty */) {
1418 IVUsers::const_iterator CandidateUI = UI;
1420 Instruction *ShadowUse = CandidateUI->getUser();
1421 const Type *DestTy = NULL;
1423 /* If shadow use is a int->float cast then insert a second IV
1424 to eliminate this cast.
1426 for (unsigned i = 0; i < n; ++i)
1432 for (unsigned i = 0; i < n; ++i, ++d)
1435 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1436 DestTy = UCast->getDestTy();
1437 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1438 DestTy = SCast->getDestTy();
1439 if (!DestTy) continue;
1442 // If target does not support DestTy natively then do not apply
1443 // this transformation.
1444 EVT DVT = TLI->getValueType(DestTy);
1445 if (!TLI->isTypeLegal(DVT)) continue;
1448 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1450 if (PH->getNumIncomingValues() != 2) continue;
1452 const Type *SrcTy = PH->getType();
1453 int Mantissa = DestTy->getFPMantissaWidth();
1454 if (Mantissa == -1) continue;
1455 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1458 unsigned Entry, Latch;
1459 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1467 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1468 if (!Init) continue;
1469 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1471 BinaryOperator *Incr =
1472 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1473 if (!Incr) continue;
1474 if (Incr->getOpcode() != Instruction::Add
1475 && Incr->getOpcode() != Instruction::Sub)
1478 /* Initialize new IV, double d = 0.0 in above example. */
1479 ConstantInt *C = NULL;
1480 if (Incr->getOperand(0) == PH)
1481 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1482 else if (Incr->getOperand(1) == PH)
1483 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1489 // Ignore negative constants, as the code below doesn't handle them
1490 // correctly. TODO: Remove this restriction.
1491 if (!C->getValue().isStrictlyPositive()) continue;
1493 /* Add new PHINode. */
1494 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1496 /* create new increment. '++d' in above example. */
1497 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1498 BinaryOperator *NewIncr =
1499 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1500 Instruction::FAdd : Instruction::FSub,
1501 NewPH, CFP, "IV.S.next.", Incr);
1503 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1504 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1506 /* Remove cast operation */
1507 ShadowUse->replaceAllUsesWith(NewPH);
1508 ShadowUse->eraseFromParent();
1514 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1515 /// set the IV user and stride information and return true, otherwise return
1517 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1518 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1519 if (UI->getUser() == Cond) {
1520 // NOTE: we could handle setcc instructions with multiple uses here, but
1521 // InstCombine does it as well for simple uses, it's not clear that it
1522 // occurs enough in real life to handle.
1529 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1530 /// a max computation.
1532 /// This is a narrow solution to a specific, but acute, problem. For loops
1538 /// } while (++i < n);
1540 /// the trip count isn't just 'n', because 'n' might not be positive. And
1541 /// unfortunately this can come up even for loops where the user didn't use
1542 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1543 /// will commonly be lowered like this:
1549 /// } while (++i < n);
1552 /// and then it's possible for subsequent optimization to obscure the if
1553 /// test in such a way that indvars can't find it.
1555 /// When indvars can't find the if test in loops like this, it creates a
1556 /// max expression, which allows it to give the loop a canonical
1557 /// induction variable:
1560 /// max = n < 1 ? 1 : n;
1563 /// } while (++i != max);
1565 /// Canonical induction variables are necessary because the loop passes
1566 /// are designed around them. The most obvious example of this is the
1567 /// LoopInfo analysis, which doesn't remember trip count values. It
1568 /// expects to be able to rediscover the trip count each time it is
1569 /// needed, and it does this using a simple analysis that only succeeds if
1570 /// the loop has a canonical induction variable.
1572 /// However, when it comes time to generate code, the maximum operation
1573 /// can be quite costly, especially if it's inside of an outer loop.
1575 /// This function solves this problem by detecting this type of loop and
1576 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1577 /// the instructions for the maximum computation.
1579 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1580 // Check that the loop matches the pattern we're looking for.
1581 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1582 Cond->getPredicate() != CmpInst::ICMP_NE)
1585 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1586 if (!Sel || !Sel->hasOneUse()) return Cond;
1588 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1589 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1591 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1593 // Add one to the backedge-taken count to get the trip count.
1594 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1595 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1597 // Check for a max calculation that matches the pattern. There's no check
1598 // for ICMP_ULE here because the comparison would be with zero, which
1599 // isn't interesting.
1600 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1601 const SCEVNAryExpr *Max = 0;
1602 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1603 Pred = ICmpInst::ICMP_SLE;
1605 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1606 Pred = ICmpInst::ICMP_SLT;
1608 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1609 Pred = ICmpInst::ICMP_ULT;
1616 // To handle a max with more than two operands, this optimization would
1617 // require additional checking and setup.
1618 if (Max->getNumOperands() != 2)
1621 const SCEV *MaxLHS = Max->getOperand(0);
1622 const SCEV *MaxRHS = Max->getOperand(1);
1624 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1625 // for a comparison with 1. For <= and >=, a comparison with zero.
1627 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1630 // Check the relevant induction variable for conformance to
1632 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1633 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1634 if (!AR || !AR->isAffine() ||
1635 AR->getStart() != One ||
1636 AR->getStepRecurrence(SE) != One)
1639 assert(AR->getLoop() == L &&
1640 "Loop condition operand is an addrec in a different loop!");
1642 // Check the right operand of the select, and remember it, as it will
1643 // be used in the new comparison instruction.
1645 if (ICmpInst::isTrueWhenEqual(Pred)) {
1646 // Look for n+1, and grab n.
1647 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1648 if (isa<ConstantInt>(BO->getOperand(1)) &&
1649 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1650 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1651 NewRHS = BO->getOperand(0);
1652 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1653 if (isa<ConstantInt>(BO->getOperand(1)) &&
1654 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1655 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1656 NewRHS = BO->getOperand(0);
1659 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1660 NewRHS = Sel->getOperand(1);
1661 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1662 NewRHS = Sel->getOperand(2);
1663 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1664 NewRHS = SU->getValue();
1666 // Max doesn't match expected pattern.
1669 // Determine the new comparison opcode. It may be signed or unsigned,
1670 // and the original comparison may be either equality or inequality.
1671 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1672 Pred = CmpInst::getInversePredicate(Pred);
1674 // Ok, everything looks ok to change the condition into an SLT or SGE and
1675 // delete the max calculation.
1677 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1679 // Delete the max calculation instructions.
1680 Cond->replaceAllUsesWith(NewCond);
1681 CondUse->setUser(NewCond);
1682 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1683 Cond->eraseFromParent();
1684 Sel->eraseFromParent();
1685 if (Cmp->use_empty())
1686 Cmp->eraseFromParent();
1690 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1691 /// postinc iv when possible.
1693 LSRInstance::OptimizeLoopTermCond() {
1694 SmallPtrSet<Instruction *, 4> PostIncs;
1696 BasicBlock *LatchBlock = L->getLoopLatch();
1697 SmallVector<BasicBlock*, 8> ExitingBlocks;
1698 L->getExitingBlocks(ExitingBlocks);
1700 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1701 BasicBlock *ExitingBlock = ExitingBlocks[i];
1703 // Get the terminating condition for the loop if possible. If we
1704 // can, we want to change it to use a post-incremented version of its
1705 // induction variable, to allow coalescing the live ranges for the IV into
1706 // one register value.
1708 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1711 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1712 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1715 // Search IVUsesByStride to find Cond's IVUse if there is one.
1716 IVStrideUse *CondUse = 0;
1717 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1718 if (!FindIVUserForCond(Cond, CondUse))
1721 // If the trip count is computed in terms of a max (due to ScalarEvolution
1722 // being unable to find a sufficient guard, for example), change the loop
1723 // comparison to use SLT or ULT instead of NE.
1724 // One consequence of doing this now is that it disrupts the count-down
1725 // optimization. That's not always a bad thing though, because in such
1726 // cases it may still be worthwhile to avoid a max.
1727 Cond = OptimizeMax(Cond, CondUse);
1729 // If this exiting block dominates the latch block, it may also use
1730 // the post-inc value if it won't be shared with other uses.
1731 // Check for dominance.
1732 if (!DT.dominates(ExitingBlock, LatchBlock))
1735 // Conservatively avoid trying to use the post-inc value in non-latch
1736 // exits if there may be pre-inc users in intervening blocks.
1737 if (LatchBlock != ExitingBlock)
1738 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1739 // Test if the use is reachable from the exiting block. This dominator
1740 // query is a conservative approximation of reachability.
1741 if (&*UI != CondUse &&
1742 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1743 // Conservatively assume there may be reuse if the quotient of their
1744 // strides could be a legal scale.
1745 const SCEV *A = IU.getStride(*CondUse, L);
1746 const SCEV *B = IU.getStride(*UI, L);
1747 if (!A || !B) continue;
1748 if (SE.getTypeSizeInBits(A->getType()) !=
1749 SE.getTypeSizeInBits(B->getType())) {
1750 if (SE.getTypeSizeInBits(A->getType()) >
1751 SE.getTypeSizeInBits(B->getType()))
1752 B = SE.getSignExtendExpr(B, A->getType());
1754 A = SE.getSignExtendExpr(A, B->getType());
1756 if (const SCEVConstant *D =
1757 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1758 const ConstantInt *C = D->getValue();
1759 // Stride of one or negative one can have reuse with non-addresses.
1760 if (C->isOne() || C->isAllOnesValue())
1761 goto decline_post_inc;
1762 // Avoid weird situations.
1763 if (C->getValue().getMinSignedBits() >= 64 ||
1764 C->getValue().isMinSignedValue())
1765 goto decline_post_inc;
1766 // Without TLI, assume that any stride might be valid, and so any
1767 // use might be shared.
1769 goto decline_post_inc;
1770 // Check for possible scaled-address reuse.
1771 const Type *AccessTy = getAccessType(UI->getUser());
1772 TargetLowering::AddrMode AM;
1773 AM.Scale = C->getSExtValue();
1774 if (TLI->isLegalAddressingMode(AM, AccessTy))
1775 goto decline_post_inc;
1776 AM.Scale = -AM.Scale;
1777 if (TLI->isLegalAddressingMode(AM, AccessTy))
1778 goto decline_post_inc;
1782 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1785 // It's possible for the setcc instruction to be anywhere in the loop, and
1786 // possible for it to have multiple users. If it is not immediately before
1787 // the exiting block branch, move it.
1788 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1789 if (Cond->hasOneUse()) {
1790 Cond->moveBefore(TermBr);
1792 // Clone the terminating condition and insert into the loopend.
1793 ICmpInst *OldCond = Cond;
1794 Cond = cast<ICmpInst>(Cond->clone());
1795 Cond->setName(L->getHeader()->getName() + ".termcond");
1796 ExitingBlock->getInstList().insert(TermBr, Cond);
1798 // Clone the IVUse, as the old use still exists!
1799 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1800 TermBr->replaceUsesOfWith(OldCond, Cond);
1804 // If we get to here, we know that we can transform the setcc instruction to
1805 // use the post-incremented version of the IV, allowing us to coalesce the
1806 // live ranges for the IV correctly.
1807 CondUse->transformToPostInc(L);
1810 PostIncs.insert(Cond);
1814 // Determine an insertion point for the loop induction variable increment. It
1815 // must dominate all the post-inc comparisons we just set up, and it must
1816 // dominate the loop latch edge.
1817 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1818 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1819 E = PostIncs.end(); I != E; ++I) {
1821 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1823 if (BB == (*I)->getParent())
1824 IVIncInsertPos = *I;
1825 else if (BB != IVIncInsertPos->getParent())
1826 IVIncInsertPos = BB->getTerminator();
1830 /// reconcileNewOffset - Determine if the given use can accomodate a fixup
1831 /// at the given offset and other details. If so, update the use and
1834 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1835 LSRUse::KindType Kind, const Type *AccessTy) {
1836 int64_t NewMinOffset = LU.MinOffset;
1837 int64_t NewMaxOffset = LU.MaxOffset;
1838 const Type *NewAccessTy = AccessTy;
1840 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1841 // something conservative, however this can pessimize in the case that one of
1842 // the uses will have all its uses outside the loop, for example.
1843 if (LU.Kind != Kind)
1845 // Conservatively assume HasBaseReg is true for now.
1846 if (NewOffset < LU.MinOffset) {
1847 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1848 Kind, AccessTy, TLI))
1850 NewMinOffset = NewOffset;
1851 } else if (NewOffset > LU.MaxOffset) {
1852 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1853 Kind, AccessTy, TLI))
1855 NewMaxOffset = NewOffset;
1857 // Check for a mismatched access type, and fall back conservatively as needed.
1858 // TODO: Be less conservative when the type is similar and can use the same
1859 // addressing modes.
1860 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1861 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1864 LU.MinOffset = NewMinOffset;
1865 LU.MaxOffset = NewMaxOffset;
1866 LU.AccessTy = NewAccessTy;
1867 if (NewOffset != LU.Offsets.back())
1868 LU.Offsets.push_back(NewOffset);
1872 /// getUse - Return an LSRUse index and an offset value for a fixup which
1873 /// needs the given expression, with the given kind and optional access type.
1874 /// Either reuse an existing use or create a new one, as needed.
1875 std::pair<size_t, int64_t>
1876 LSRInstance::getUse(const SCEV *&Expr,
1877 LSRUse::KindType Kind, const Type *AccessTy) {
1878 const SCEV *Copy = Expr;
1879 int64_t Offset = ExtractImmediate(Expr, SE);
1881 // Basic uses can't accept any offset, for example.
1882 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1887 std::pair<UseMapTy::iterator, bool> P =
1888 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1890 // A use already existed with this base.
1891 size_t LUIdx = P.first->second;
1892 LSRUse &LU = Uses[LUIdx];
1893 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1895 return std::make_pair(LUIdx, Offset);
1898 // Create a new use.
1899 size_t LUIdx = Uses.size();
1900 P.first->second = LUIdx;
1901 Uses.push_back(LSRUse(Kind, AccessTy));
1902 LSRUse &LU = Uses[LUIdx];
1904 // We don't need to track redundant offsets, but we don't need to go out
1905 // of our way here to avoid them.
1906 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1907 LU.Offsets.push_back(Offset);
1909 LU.MinOffset = Offset;
1910 LU.MaxOffset = Offset;
1911 return std::make_pair(LUIdx, Offset);
1914 /// DeleteUse - Delete the given use from the Uses list.
1915 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1916 if (&LU != &Uses.back())
1917 std::swap(LU, Uses.back());
1921 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1924 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1925 /// a formula that has the same registers as the given formula.
1927 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1928 const LSRUse &OrigLU) {
1929 // Search all uses for the formula. This could be more clever.
1930 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1931 LSRUse &LU = Uses[LUIdx];
1932 // Check whether this use is close enough to OrigLU, to see whether it's
1933 // worthwhile looking through its formulae.
1934 // Ignore ICmpZero uses because they may contain formulae generated by
1935 // GenerateICmpZeroScales, in which case adding fixup offsets may
1937 if (&LU != &OrigLU &&
1938 LU.Kind != LSRUse::ICmpZero &&
1939 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1940 LU.WidestFixupType == OrigLU.WidestFixupType &&
1941 LU.HasFormulaWithSameRegs(OrigF)) {
1942 // Scan through this use's formulae.
1943 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1944 E = LU.Formulae.end(); I != E; ++I) {
1945 const Formula &F = *I;
1946 // Check to see if this formula has the same registers and symbols
1948 if (F.BaseRegs == OrigF.BaseRegs &&
1949 F.ScaledReg == OrigF.ScaledReg &&
1950 F.AM.BaseGV == OrigF.AM.BaseGV &&
1951 F.AM.Scale == OrigF.AM.Scale) {
1952 if (F.AM.BaseOffs == 0)
1954 // This is the formula where all the registers and symbols matched;
1955 // there aren't going to be any others. Since we declined it, we
1956 // can skip the rest of the formulae and procede to the next LSRUse.
1963 // Nothing looked good.
1967 void LSRInstance::CollectInterestingTypesAndFactors() {
1968 SmallSetVector<const SCEV *, 4> Strides;
1970 // Collect interesting types and strides.
1971 SmallVector<const SCEV *, 4> Worklist;
1972 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1973 const SCEV *Expr = IU.getExpr(*UI);
1975 // Collect interesting types.
1976 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1978 // Add strides for mentioned loops.
1979 Worklist.push_back(Expr);
1981 const SCEV *S = Worklist.pop_back_val();
1982 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1983 Strides.insert(AR->getStepRecurrence(SE));
1984 Worklist.push_back(AR->getStart());
1985 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1986 Worklist.append(Add->op_begin(), Add->op_end());
1988 } while (!Worklist.empty());
1991 // Compute interesting factors from the set of interesting strides.
1992 for (SmallSetVector<const SCEV *, 4>::const_iterator
1993 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1994 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1995 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
1996 const SCEV *OldStride = *I;
1997 const SCEV *NewStride = *NewStrideIter;
1999 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2000 SE.getTypeSizeInBits(NewStride->getType())) {
2001 if (SE.getTypeSizeInBits(OldStride->getType()) >
2002 SE.getTypeSizeInBits(NewStride->getType()))
2003 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2005 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2007 if (const SCEVConstant *Factor =
2008 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2010 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2011 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2012 } else if (const SCEVConstant *Factor =
2013 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2016 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2017 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2021 // If all uses use the same type, don't bother looking for truncation-based
2023 if (Types.size() == 1)
2026 DEBUG(print_factors_and_types(dbgs()));
2029 void LSRInstance::CollectFixupsAndInitialFormulae() {
2030 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2032 LSRFixup &LF = getNewFixup();
2033 LF.UserInst = UI->getUser();
2034 LF.OperandValToReplace = UI->getOperandValToReplace();
2035 LF.PostIncLoops = UI->getPostIncLoops();
2037 LSRUse::KindType Kind = LSRUse::Basic;
2038 const Type *AccessTy = 0;
2039 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2040 Kind = LSRUse::Address;
2041 AccessTy = getAccessType(LF.UserInst);
2044 const SCEV *S = IU.getExpr(*UI);
2046 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2047 // (N - i == 0), and this allows (N - i) to be the expression that we work
2048 // with rather than just N or i, so we can consider the register
2049 // requirements for both N and i at the same time. Limiting this code to
2050 // equality icmps is not a problem because all interesting loops use
2051 // equality icmps, thanks to IndVarSimplify.
2052 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2053 if (CI->isEquality()) {
2054 // Swap the operands if needed to put the OperandValToReplace on the
2055 // left, for consistency.
2056 Value *NV = CI->getOperand(1);
2057 if (NV == LF.OperandValToReplace) {
2058 CI->setOperand(1, CI->getOperand(0));
2059 CI->setOperand(0, NV);
2060 NV = CI->getOperand(1);
2064 // x == y --> x - y == 0
2065 const SCEV *N = SE.getSCEV(NV);
2066 if (SE.isLoopInvariant(N, L)) {
2067 Kind = LSRUse::ICmpZero;
2068 S = SE.getMinusSCEV(N, S);
2071 // -1 and the negations of all interesting strides (except the negation
2072 // of -1) are now also interesting.
2073 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2074 if (Factors[i] != -1)
2075 Factors.insert(-(uint64_t)Factors[i]);
2079 // Set up the initial formula for this use.
2080 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2082 LF.Offset = P.second;
2083 LSRUse &LU = Uses[LF.LUIdx];
2084 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2085 if (!LU.WidestFixupType ||
2086 SE.getTypeSizeInBits(LU.WidestFixupType) <
2087 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2088 LU.WidestFixupType = LF.OperandValToReplace->getType();
2090 // If this is the first use of this LSRUse, give it a formula.
2091 if (LU.Formulae.empty()) {
2092 InsertInitialFormula(S, LU, LF.LUIdx);
2093 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2097 DEBUG(print_fixups(dbgs()));
2100 /// InsertInitialFormula - Insert a formula for the given expression into
2101 /// the given use, separating out loop-variant portions from loop-invariant
2102 /// and loop-computable portions.
2104 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2106 F.InitialMatch(S, L, SE);
2107 bool Inserted = InsertFormula(LU, LUIdx, F);
2108 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2111 /// InsertSupplementalFormula - Insert a simple single-register formula for
2112 /// the given expression into the given use.
2114 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2115 LSRUse &LU, size_t LUIdx) {
2117 F.BaseRegs.push_back(S);
2118 F.AM.HasBaseReg = true;
2119 bool Inserted = InsertFormula(LU, LUIdx, F);
2120 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2123 /// CountRegisters - Note which registers are used by the given formula,
2124 /// updating RegUses.
2125 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2127 RegUses.CountRegister(F.ScaledReg, LUIdx);
2128 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2129 E = F.BaseRegs.end(); I != E; ++I)
2130 RegUses.CountRegister(*I, LUIdx);
2133 /// InsertFormula - If the given formula has not yet been inserted, add it to
2134 /// the list, and return true. Return false otherwise.
2135 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2136 if (!LU.InsertFormula(F))
2139 CountRegisters(F, LUIdx);
2143 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2144 /// loop-invariant values which we're tracking. These other uses will pin these
2145 /// values in registers, making them less profitable for elimination.
2146 /// TODO: This currently misses non-constant addrec step registers.
2147 /// TODO: Should this give more weight to users inside the loop?
2149 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2150 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2151 SmallPtrSet<const SCEV *, 8> Inserted;
2153 while (!Worklist.empty()) {
2154 const SCEV *S = Worklist.pop_back_val();
2156 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2157 Worklist.append(N->op_begin(), N->op_end());
2158 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2159 Worklist.push_back(C->getOperand());
2160 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2161 Worklist.push_back(D->getLHS());
2162 Worklist.push_back(D->getRHS());
2163 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2164 if (!Inserted.insert(U)) continue;
2165 const Value *V = U->getValue();
2166 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2167 // Look for instructions defined outside the loop.
2168 if (L->contains(Inst)) continue;
2169 } else if (isa<UndefValue>(V))
2170 // Undef doesn't have a live range, so it doesn't matter.
2172 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2174 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2175 // Ignore non-instructions.
2178 // Ignore instructions in other functions (as can happen with
2180 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2182 // Ignore instructions not dominated by the loop.
2183 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2184 UserInst->getParent() :
2185 cast<PHINode>(UserInst)->getIncomingBlock(
2186 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2187 if (!DT.dominates(L->getHeader(), UseBB))
2189 // Ignore uses which are part of other SCEV expressions, to avoid
2190 // analyzing them multiple times.
2191 if (SE.isSCEVable(UserInst->getType())) {
2192 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2193 // If the user is a no-op, look through to its uses.
2194 if (!isa<SCEVUnknown>(UserS))
2198 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2202 // Ignore icmp instructions which are already being analyzed.
2203 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2204 unsigned OtherIdx = !UI.getOperandNo();
2205 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2206 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2210 LSRFixup &LF = getNewFixup();
2211 LF.UserInst = const_cast<Instruction *>(UserInst);
2212 LF.OperandValToReplace = UI.getUse();
2213 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2215 LF.Offset = P.second;
2216 LSRUse &LU = Uses[LF.LUIdx];
2217 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2218 if (!LU.WidestFixupType ||
2219 SE.getTypeSizeInBits(LU.WidestFixupType) <
2220 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2221 LU.WidestFixupType = LF.OperandValToReplace->getType();
2222 InsertSupplementalFormula(U, LU, LF.LUIdx);
2223 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2230 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2231 /// separate registers. If C is non-null, multiply each subexpression by C.
2232 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2233 SmallVectorImpl<const SCEV *> &Ops,
2235 ScalarEvolution &SE) {
2236 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2237 // Break out add operands.
2238 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2240 CollectSubexprs(*I, C, Ops, L, SE);
2242 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2243 // Split a non-zero base out of an addrec.
2244 if (!AR->getStart()->isZero()) {
2245 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2246 AR->getStepRecurrence(SE),
2248 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2251 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2254 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2255 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2256 if (Mul->getNumOperands() == 2)
2257 if (const SCEVConstant *Op0 =
2258 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2259 CollectSubexprs(Mul->getOperand(1),
2260 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2266 // Otherwise use the value itself, optionally with a scale applied.
2267 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2270 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2272 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2275 // Arbitrarily cap recursion to protect compile time.
2276 if (Depth >= 3) return;
2278 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2279 const SCEV *BaseReg = Base.BaseRegs[i];
2281 SmallVector<const SCEV *, 8> AddOps;
2282 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2284 if (AddOps.size() == 1) continue;
2286 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2287 JE = AddOps.end(); J != JE; ++J) {
2289 // Loop-variant "unknown" values are uninteresting; we won't be able to
2290 // do anything meaningful with them.
2291 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2294 // Don't pull a constant into a register if the constant could be folded
2295 // into an immediate field.
2296 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2297 Base.getNumRegs() > 1,
2298 LU.Kind, LU.AccessTy, TLI, SE))
2301 // Collect all operands except *J.
2302 SmallVector<const SCEV *, 8> InnerAddOps
2303 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2305 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2307 // Don't leave just a constant behind in a register if the constant could
2308 // be folded into an immediate field.
2309 if (InnerAddOps.size() == 1 &&
2310 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2311 Base.getNumRegs() > 1,
2312 LU.Kind, LU.AccessTy, TLI, SE))
2315 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2316 if (InnerSum->isZero())
2319 F.BaseRegs[i] = InnerSum;
2320 F.BaseRegs.push_back(*J);
2321 if (InsertFormula(LU, LUIdx, F))
2322 // If that formula hadn't been seen before, recurse to find more like
2324 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2329 /// GenerateCombinations - Generate a formula consisting of all of the
2330 /// loop-dominating registers added into a single register.
2331 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2333 // This method is only interesting on a plurality of registers.
2334 if (Base.BaseRegs.size() <= 1) return;
2338 SmallVector<const SCEV *, 4> Ops;
2339 for (SmallVectorImpl<const SCEV *>::const_iterator
2340 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2341 const SCEV *BaseReg = *I;
2342 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2343 !SE.hasComputableLoopEvolution(BaseReg, L))
2344 Ops.push_back(BaseReg);
2346 F.BaseRegs.push_back(BaseReg);
2348 if (Ops.size() > 1) {
2349 const SCEV *Sum = SE.getAddExpr(Ops);
2350 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2351 // opportunity to fold something. For now, just ignore such cases
2352 // rather than proceed with zero in a register.
2353 if (!Sum->isZero()) {
2354 F.BaseRegs.push_back(Sum);
2355 (void)InsertFormula(LU, LUIdx, F);
2360 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2361 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2363 // We can't add a symbolic offset if the address already contains one.
2364 if (Base.AM.BaseGV) return;
2366 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2367 const SCEV *G = Base.BaseRegs[i];
2368 GlobalValue *GV = ExtractSymbol(G, SE);
2369 if (G->isZero() || !GV)
2373 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2374 LU.Kind, LU.AccessTy, TLI))
2377 (void)InsertFormula(LU, LUIdx, F);
2381 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2382 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2384 // TODO: For now, just add the min and max offset, because it usually isn't
2385 // worthwhile looking at everything inbetween.
2386 SmallVector<int64_t, 2> Worklist;
2387 Worklist.push_back(LU.MinOffset);
2388 if (LU.MaxOffset != LU.MinOffset)
2389 Worklist.push_back(LU.MaxOffset);
2391 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2392 const SCEV *G = Base.BaseRegs[i];
2394 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2395 E = Worklist.end(); I != E; ++I) {
2397 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2398 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2399 LU.Kind, LU.AccessTy, TLI)) {
2400 // Add the offset to the base register.
2401 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2402 // If it cancelled out, drop the base register, otherwise update it.
2403 if (NewG->isZero()) {
2404 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2405 F.BaseRegs.pop_back();
2407 F.BaseRegs[i] = NewG;
2409 (void)InsertFormula(LU, LUIdx, F);
2413 int64_t Imm = ExtractImmediate(G, SE);
2414 if (G->isZero() || Imm == 0)
2417 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2418 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2419 LU.Kind, LU.AccessTy, TLI))
2422 (void)InsertFormula(LU, LUIdx, F);
2426 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2427 /// the comparison. For example, x == y -> x*c == y*c.
2428 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2430 if (LU.Kind != LSRUse::ICmpZero) return;
2432 // Determine the integer type for the base formula.
2433 const Type *IntTy = Base.getType();
2435 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2437 // Don't do this if there is more than one offset.
2438 if (LU.MinOffset != LU.MaxOffset) return;
2440 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2442 // Check each interesting stride.
2443 for (SmallSetVector<int64_t, 8>::const_iterator
2444 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2445 int64_t Factor = *I;
2447 // Check that the multiplication doesn't overflow.
2448 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2450 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2451 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2454 // Check that multiplying with the use offset doesn't overflow.
2455 int64_t Offset = LU.MinOffset;
2456 if (Offset == INT64_MIN && Factor == -1)
2458 Offset = (uint64_t)Offset * Factor;
2459 if (Offset / Factor != LU.MinOffset)
2463 F.AM.BaseOffs = NewBaseOffs;
2465 // Check that this scale is legal.
2466 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2469 // Compensate for the use having MinOffset built into it.
2470 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2472 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2474 // Check that multiplying with each base register doesn't overflow.
2475 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2476 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2477 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2481 // Check that multiplying with the scaled register doesn't overflow.
2483 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2484 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2488 // If we make it here and it's legal, add it.
2489 (void)InsertFormula(LU, LUIdx, F);
2494 /// GenerateScales - Generate stride factor reuse formulae by making use of
2495 /// scaled-offset address modes, for example.
2496 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2497 // Determine the integer type for the base formula.
2498 const Type *IntTy = Base.getType();
2501 // If this Formula already has a scaled register, we can't add another one.
2502 if (Base.AM.Scale != 0) return;
2504 // Check each interesting stride.
2505 for (SmallSetVector<int64_t, 8>::const_iterator
2506 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2507 int64_t Factor = *I;
2509 Base.AM.Scale = Factor;
2510 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2511 // Check whether this scale is going to be legal.
2512 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2513 LU.Kind, LU.AccessTy, TLI)) {
2514 // As a special-case, handle special out-of-loop Basic users specially.
2515 // TODO: Reconsider this special case.
2516 if (LU.Kind == LSRUse::Basic &&
2517 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2518 LSRUse::Special, LU.AccessTy, TLI) &&
2519 LU.AllFixupsOutsideLoop)
2520 LU.Kind = LSRUse::Special;
2524 // For an ICmpZero, negating a solitary base register won't lead to
2526 if (LU.Kind == LSRUse::ICmpZero &&
2527 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2529 // For each addrec base reg, apply the scale, if possible.
2530 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2531 if (const SCEVAddRecExpr *AR =
2532 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2533 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2534 if (FactorS->isZero())
2536 // Divide out the factor, ignoring high bits, since we'll be
2537 // scaling the value back up in the end.
2538 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2539 // TODO: This could be optimized to avoid all the copying.
2541 F.ScaledReg = Quotient;
2542 F.DeleteBaseReg(F.BaseRegs[i]);
2543 (void)InsertFormula(LU, LUIdx, F);
2549 /// GenerateTruncates - Generate reuse formulae from different IV types.
2550 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2551 // This requires TargetLowering to tell us which truncates are free.
2554 // Don't bother truncating symbolic values.
2555 if (Base.AM.BaseGV) return;
2557 // Determine the integer type for the base formula.
2558 const Type *DstTy = Base.getType();
2560 DstTy = SE.getEffectiveSCEVType(DstTy);
2562 for (SmallSetVector<const Type *, 4>::const_iterator
2563 I = Types.begin(), E = Types.end(); I != E; ++I) {
2564 const Type *SrcTy = *I;
2565 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2568 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2569 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2570 JE = F.BaseRegs.end(); J != JE; ++J)
2571 *J = SE.getAnyExtendExpr(*J, SrcTy);
2573 // TODO: This assumes we've done basic processing on all uses and
2574 // have an idea what the register usage is.
2575 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2578 (void)InsertFormula(LU, LUIdx, F);
2585 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2586 /// defer modifications so that the search phase doesn't have to worry about
2587 /// the data structures moving underneath it.
2591 const SCEV *OrigReg;
2593 WorkItem(size_t LI, int64_t I, const SCEV *R)
2594 : LUIdx(LI), Imm(I), OrigReg(R) {}
2596 void print(raw_ostream &OS) const;
2602 void WorkItem::print(raw_ostream &OS) const {
2603 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2604 << " , add offset " << Imm;
2607 void WorkItem::dump() const {
2608 print(errs()); errs() << '\n';
2611 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2612 /// distance apart and try to form reuse opportunities between them.
2613 void LSRInstance::GenerateCrossUseConstantOffsets() {
2614 // Group the registers by their value without any added constant offset.
2615 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2616 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2618 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2619 SmallVector<const SCEV *, 8> Sequence;
2620 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2622 const SCEV *Reg = *I;
2623 int64_t Imm = ExtractImmediate(Reg, SE);
2624 std::pair<RegMapTy::iterator, bool> Pair =
2625 Map.insert(std::make_pair(Reg, ImmMapTy()));
2627 Sequence.push_back(Reg);
2628 Pair.first->second.insert(std::make_pair(Imm, *I));
2629 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2632 // Now examine each set of registers with the same base value. Build up
2633 // a list of work to do and do the work in a separate step so that we're
2634 // not adding formulae and register counts while we're searching.
2635 SmallVector<WorkItem, 32> WorkItems;
2636 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2637 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2638 E = Sequence.end(); I != E; ++I) {
2639 const SCEV *Reg = *I;
2640 const ImmMapTy &Imms = Map.find(Reg)->second;
2642 // It's not worthwhile looking for reuse if there's only one offset.
2643 if (Imms.size() == 1)
2646 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2647 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2649 dbgs() << ' ' << J->first;
2652 // Examine each offset.
2653 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2655 const SCEV *OrigReg = J->second;
2657 int64_t JImm = J->first;
2658 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2660 if (!isa<SCEVConstant>(OrigReg) &&
2661 UsedByIndicesMap[Reg].count() == 1) {
2662 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2666 // Conservatively examine offsets between this orig reg a few selected
2668 ImmMapTy::const_iterator OtherImms[] = {
2669 Imms.begin(), prior(Imms.end()),
2670 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2672 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2673 ImmMapTy::const_iterator M = OtherImms[i];
2674 if (M == J || M == JE) continue;
2676 // Compute the difference between the two.
2677 int64_t Imm = (uint64_t)JImm - M->first;
2678 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2679 LUIdx = UsedByIndices.find_next(LUIdx))
2680 // Make a memo of this use, offset, and register tuple.
2681 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2682 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2689 UsedByIndicesMap.clear();
2690 UniqueItems.clear();
2692 // Now iterate through the worklist and add new formulae.
2693 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2694 E = WorkItems.end(); I != E; ++I) {
2695 const WorkItem &WI = *I;
2696 size_t LUIdx = WI.LUIdx;
2697 LSRUse &LU = Uses[LUIdx];
2698 int64_t Imm = WI.Imm;
2699 const SCEV *OrigReg = WI.OrigReg;
2701 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2702 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2703 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2705 // TODO: Use a more targeted data structure.
2706 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2707 const Formula &F = LU.Formulae[L];
2708 // Use the immediate in the scaled register.
2709 if (F.ScaledReg == OrigReg) {
2710 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2711 Imm * (uint64_t)F.AM.Scale;
2712 // Don't create 50 + reg(-50).
2713 if (F.referencesReg(SE.getSCEV(
2714 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2717 NewF.AM.BaseOffs = Offs;
2718 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2719 LU.Kind, LU.AccessTy, TLI))
2721 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2723 // If the new scale is a constant in a register, and adding the constant
2724 // value to the immediate would produce a value closer to zero than the
2725 // immediate itself, then the formula isn't worthwhile.
2726 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2727 if (C->getValue()->getValue().isNegative() !=
2728 (NewF.AM.BaseOffs < 0) &&
2729 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2730 .ule(abs64(NewF.AM.BaseOffs)))
2734 (void)InsertFormula(LU, LUIdx, NewF);
2736 // Use the immediate in a base register.
2737 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2738 const SCEV *BaseReg = F.BaseRegs[N];
2739 if (BaseReg != OrigReg)
2742 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2743 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2744 LU.Kind, LU.AccessTy, TLI))
2746 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2748 // If the new formula has a constant in a register, and adding the
2749 // constant value to the immediate would produce a value closer to
2750 // zero than the immediate itself, then the formula isn't worthwhile.
2751 for (SmallVectorImpl<const SCEV *>::const_iterator
2752 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2754 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2755 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2756 abs64(NewF.AM.BaseOffs)) &&
2757 (C->getValue()->getValue() +
2758 NewF.AM.BaseOffs).countTrailingZeros() >=
2759 CountTrailingZeros_64(NewF.AM.BaseOffs))
2763 (void)InsertFormula(LU, LUIdx, NewF);
2772 /// GenerateAllReuseFormulae - Generate formulae for each use.
2774 LSRInstance::GenerateAllReuseFormulae() {
2775 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2776 // queries are more precise.
2777 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2778 LSRUse &LU = Uses[LUIdx];
2779 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2780 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2781 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2782 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2784 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2785 LSRUse &LU = Uses[LUIdx];
2786 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2787 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2788 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2789 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2790 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2791 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2792 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2793 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2795 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2796 LSRUse &LU = Uses[LUIdx];
2797 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2798 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2801 GenerateCrossUseConstantOffsets();
2803 DEBUG(dbgs() << "\n"
2804 "After generating reuse formulae:\n";
2805 print_uses(dbgs()));
2808 /// If there are multiple formulae with the same set of registers used
2809 /// by other uses, pick the best one and delete the others.
2810 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2811 DenseSet<const SCEV *> VisitedRegs;
2812 SmallPtrSet<const SCEV *, 16> Regs;
2814 bool ChangedFormulae = false;
2817 // Collect the best formula for each unique set of shared registers. This
2818 // is reset for each use.
2819 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2821 BestFormulaeTy BestFormulae;
2823 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2824 LSRUse &LU = Uses[LUIdx];
2825 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2828 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2829 FIdx != NumForms; ++FIdx) {
2830 Formula &F = LU.Formulae[FIdx];
2832 SmallVector<const SCEV *, 2> Key;
2833 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2834 JE = F.BaseRegs.end(); J != JE; ++J) {
2835 const SCEV *Reg = *J;
2836 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2840 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2841 Key.push_back(F.ScaledReg);
2842 // Unstable sort by host order ok, because this is only used for
2844 std::sort(Key.begin(), Key.end());
2846 std::pair<BestFormulaeTy::const_iterator, bool> P =
2847 BestFormulae.insert(std::make_pair(Key, FIdx));
2849 Formula &Best = LU.Formulae[P.first->second];
2852 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2855 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2857 if (CostF < CostBest)
2859 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2861 " in favor of formula "; Best.print(dbgs());
2864 ChangedFormulae = true;
2866 LU.DeleteFormula(F);
2874 // Now that we've filtered out some formulae, recompute the Regs set.
2876 LU.RecomputeRegs(LUIdx, RegUses);
2878 // Reset this to prepare for the next use.
2879 BestFormulae.clear();
2882 DEBUG(if (ChangedFormulae) {
2884 "After filtering out undesirable candidates:\n";
2889 // This is a rough guess that seems to work fairly well.
2890 static const size_t ComplexityLimit = UINT16_MAX;
2892 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2893 /// solutions the solver might have to consider. It almost never considers
2894 /// this many solutions because it prune the search space, but the pruning
2895 /// isn't always sufficient.
2896 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2898 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2899 E = Uses.end(); I != E; ++I) {
2900 size_t FSize = I->Formulae.size();
2901 if (FSize >= ComplexityLimit) {
2902 Power = ComplexityLimit;
2906 if (Power >= ComplexityLimit)
2912 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2913 /// of the registers of another formula, it won't help reduce register
2914 /// pressure (though it may not necessarily hurt register pressure); remove
2915 /// it to simplify the system.
2916 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2917 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2918 DEBUG(dbgs() << "The search space is too complex.\n");
2920 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2921 "which use a superset of registers used by other "
2924 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2925 LSRUse &LU = Uses[LUIdx];
2927 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2928 Formula &F = LU.Formulae[i];
2929 // Look for a formula with a constant or GV in a register. If the use
2930 // also has a formula with that same value in an immediate field,
2931 // delete the one that uses a register.
2932 for (SmallVectorImpl<const SCEV *>::const_iterator
2933 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2934 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2936 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2937 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2938 (I - F.BaseRegs.begin()));
2939 if (LU.HasFormulaWithSameRegs(NewF)) {
2940 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2941 LU.DeleteFormula(F);
2947 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2948 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2951 NewF.AM.BaseGV = GV;
2952 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2953 (I - F.BaseRegs.begin()));
2954 if (LU.HasFormulaWithSameRegs(NewF)) {
2955 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2957 LU.DeleteFormula(F);
2968 LU.RecomputeRegs(LUIdx, RegUses);
2971 DEBUG(dbgs() << "After pre-selection:\n";
2972 print_uses(dbgs()));
2976 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
2977 /// for expressions like A, A+1, A+2, etc., allocate a single register for
2979 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
2980 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2981 DEBUG(dbgs() << "The search space is too complex.\n");
2983 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
2984 "separated by a constant offset will use the same "
2987 // This is especially useful for unrolled loops.
2989 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2990 LSRUse &LU = Uses[LUIdx];
2991 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2992 E = LU.Formulae.end(); I != E; ++I) {
2993 const Formula &F = *I;
2994 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
2995 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2996 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
2997 /*HasBaseReg=*/false,
2998 LU.Kind, LU.AccessTy)) {
2999 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3002 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3004 // Update the relocs to reference the new use.
3005 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3006 E = Fixups.end(); I != E; ++I) {
3007 LSRFixup &Fixup = *I;
3008 if (Fixup.LUIdx == LUIdx) {
3009 Fixup.LUIdx = LUThatHas - &Uses.front();
3010 Fixup.Offset += F.AM.BaseOffs;
3011 // Add the new offset to LUThatHas' offset list.
3012 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3013 LUThatHas->Offsets.push_back(Fixup.Offset);
3014 if (Fixup.Offset > LUThatHas->MaxOffset)
3015 LUThatHas->MaxOffset = Fixup.Offset;
3016 if (Fixup.Offset < LUThatHas->MinOffset)
3017 LUThatHas->MinOffset = Fixup.Offset;
3019 DEBUG(dbgs() << "New fixup has offset "
3020 << Fixup.Offset << '\n');
3022 if (Fixup.LUIdx == NumUses-1)
3023 Fixup.LUIdx = LUIdx;
3026 // Delete formulae from the new use which are no longer legal.
3028 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3029 Formula &F = LUThatHas->Formulae[i];
3030 if (!isLegalUse(F.AM,
3031 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3032 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3033 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3035 LUThatHas->DeleteFormula(F);
3042 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3044 // Delete the old use.
3045 DeleteUse(LU, LUIdx);
3055 DEBUG(dbgs() << "After pre-selection:\n";
3056 print_uses(dbgs()));
3060 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3061 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3062 /// we've done more filtering, as it may be able to find more formulae to
3064 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3065 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3066 DEBUG(dbgs() << "The search space is too complex.\n");
3068 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3069 "undesirable dedicated registers.\n");
3071 FilterOutUndesirableDedicatedRegisters();
3073 DEBUG(dbgs() << "After pre-selection:\n";
3074 print_uses(dbgs()));
3078 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3079 /// to be profitable, and then in any use which has any reference to that
3080 /// register, delete all formulae which do not reference that register.
3081 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3082 // With all other options exhausted, loop until the system is simple
3083 // enough to handle.
3084 SmallPtrSet<const SCEV *, 4> Taken;
3085 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3086 // Ok, we have too many of formulae on our hands to conveniently handle.
3087 // Use a rough heuristic to thin out the list.
3088 DEBUG(dbgs() << "The search space is too complex.\n");
3090 // Pick the register which is used by the most LSRUses, which is likely
3091 // to be a good reuse register candidate.
3092 const SCEV *Best = 0;
3093 unsigned BestNum = 0;
3094 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3096 const SCEV *Reg = *I;
3097 if (Taken.count(Reg))
3102 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3103 if (Count > BestNum) {
3110 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3111 << " will yield profitable reuse.\n");
3114 // In any use with formulae which references this register, delete formulae
3115 // which don't reference it.
3116 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3117 LSRUse &LU = Uses[LUIdx];
3118 if (!LU.Regs.count(Best)) continue;
3121 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3122 Formula &F = LU.Formulae[i];
3123 if (!F.referencesReg(Best)) {
3124 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3125 LU.DeleteFormula(F);
3129 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3135 LU.RecomputeRegs(LUIdx, RegUses);
3138 DEBUG(dbgs() << "After pre-selection:\n";
3139 print_uses(dbgs()));
3143 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3144 /// formulae to choose from, use some rough heuristics to prune down the number
3145 /// of formulae. This keeps the main solver from taking an extraordinary amount
3146 /// of time in some worst-case scenarios.
3147 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3148 NarrowSearchSpaceByDetectingSupersets();
3149 NarrowSearchSpaceByCollapsingUnrolledCode();
3150 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3151 NarrowSearchSpaceByPickingWinnerRegs();
3154 /// SolveRecurse - This is the recursive solver.
3155 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3157 SmallVectorImpl<const Formula *> &Workspace,
3158 const Cost &CurCost,
3159 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3160 DenseSet<const SCEV *> &VisitedRegs) const {
3163 // - use more aggressive filtering
3164 // - sort the formula so that the most profitable solutions are found first
3165 // - sort the uses too
3167 // - don't compute a cost, and then compare. compare while computing a cost
3169 // - track register sets with SmallBitVector
3171 const LSRUse &LU = Uses[Workspace.size()];
3173 // If this use references any register that's already a part of the
3174 // in-progress solution, consider it a requirement that a formula must
3175 // reference that register in order to be considered. This prunes out
3176 // unprofitable searching.
3177 SmallSetVector<const SCEV *, 4> ReqRegs;
3178 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3179 E = CurRegs.end(); I != E; ++I)
3180 if (LU.Regs.count(*I))
3183 bool AnySatisfiedReqRegs = false;
3184 SmallPtrSet<const SCEV *, 16> NewRegs;
3187 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3188 E = LU.Formulae.end(); I != E; ++I) {
3189 const Formula &F = *I;
3191 // Ignore formulae which do not use any of the required registers.
3192 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3193 JE = ReqRegs.end(); J != JE; ++J) {
3194 const SCEV *Reg = *J;
3195 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3196 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3200 AnySatisfiedReqRegs = true;
3202 // Evaluate the cost of the current formula. If it's already worse than
3203 // the current best, prune the search at that point.
3206 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3207 if (NewCost < SolutionCost) {
3208 Workspace.push_back(&F);
3209 if (Workspace.size() != Uses.size()) {
3210 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3211 NewRegs, VisitedRegs);
3212 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3213 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3215 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3216 dbgs() << ". Regs:";
3217 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3218 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3219 dbgs() << ' ' << **I;
3222 SolutionCost = NewCost;
3223 Solution = Workspace;
3225 Workspace.pop_back();
3230 // If none of the formulae had all of the required registers, relax the
3231 // constraint so that we don't exclude all formulae.
3232 if (!AnySatisfiedReqRegs) {
3233 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3239 /// Solve - Choose one formula from each use. Return the results in the given
3240 /// Solution vector.
3241 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3242 SmallVector<const Formula *, 8> Workspace;
3244 SolutionCost.Loose();
3246 SmallPtrSet<const SCEV *, 16> CurRegs;
3247 DenseSet<const SCEV *> VisitedRegs;
3248 Workspace.reserve(Uses.size());
3250 // SolveRecurse does all the work.
3251 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3252 CurRegs, VisitedRegs);
3254 // Ok, we've now made all our decisions.
3255 DEBUG(dbgs() << "\n"
3256 "The chosen solution requires "; SolutionCost.print(dbgs());
3258 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3260 Uses[i].print(dbgs());
3263 Solution[i]->print(dbgs());
3267 assert(Solution.size() == Uses.size() && "Malformed solution!");
3270 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3271 /// the dominator tree far as we can go while still being dominated by the
3272 /// input positions. This helps canonicalize the insert position, which
3273 /// encourages sharing.
3274 BasicBlock::iterator
3275 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3276 const SmallVectorImpl<Instruction *> &Inputs)
3279 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3280 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3283 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3284 if (!Rung) return IP;
3285 Rung = Rung->getIDom();
3286 if (!Rung) return IP;
3287 IDom = Rung->getBlock();
3289 // Don't climb into a loop though.
3290 const Loop *IDomLoop = LI.getLoopFor(IDom);
3291 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3292 if (IDomDepth <= IPLoopDepth &&
3293 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3297 bool AllDominate = true;
3298 Instruction *BetterPos = 0;
3299 Instruction *Tentative = IDom->getTerminator();
3300 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3301 E = Inputs.end(); I != E; ++I) {
3302 Instruction *Inst = *I;
3303 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3304 AllDominate = false;
3307 // Attempt to find an insert position in the middle of the block,
3308 // instead of at the end, so that it can be used for other expansions.
3309 if (IDom == Inst->getParent() &&
3310 (!BetterPos || DT.dominates(BetterPos, Inst)))
3311 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3324 /// AdjustInsertPositionForExpand - Determine an input position which will be
3325 /// dominated by the operands and which will dominate the result.
3326 BasicBlock::iterator
3327 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3329 const LSRUse &LU) const {
3330 // Collect some instructions which must be dominated by the
3331 // expanding replacement. These must be dominated by any operands that
3332 // will be required in the expansion.
3333 SmallVector<Instruction *, 4> Inputs;
3334 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3335 Inputs.push_back(I);
3336 if (LU.Kind == LSRUse::ICmpZero)
3337 if (Instruction *I =
3338 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3339 Inputs.push_back(I);
3340 if (LF.PostIncLoops.count(L)) {
3341 if (LF.isUseFullyOutsideLoop(L))
3342 Inputs.push_back(L->getLoopLatch()->getTerminator());
3344 Inputs.push_back(IVIncInsertPos);
3346 // The expansion must also be dominated by the increment positions of any
3347 // loops it for which it is using post-inc mode.
3348 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3349 E = LF.PostIncLoops.end(); I != E; ++I) {
3350 const Loop *PIL = *I;
3351 if (PIL == L) continue;
3353 // Be dominated by the loop exit.
3354 SmallVector<BasicBlock *, 4> ExitingBlocks;
3355 PIL->getExitingBlocks(ExitingBlocks);
3356 if (!ExitingBlocks.empty()) {
3357 BasicBlock *BB = ExitingBlocks[0];
3358 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3359 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3360 Inputs.push_back(BB->getTerminator());
3364 // Then, climb up the immediate dominator tree as far as we can go while
3365 // still being dominated by the input positions.
3366 IP = HoistInsertPosition(IP, Inputs);
3368 // Don't insert instructions before PHI nodes.
3369 while (isa<PHINode>(IP)) ++IP;
3371 // Ignore debug intrinsics.
3372 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3377 /// Expand - Emit instructions for the leading candidate expression for this
3378 /// LSRUse (this is called "expanding").
3379 Value *LSRInstance::Expand(const LSRFixup &LF,
3381 BasicBlock::iterator IP,
3382 SCEVExpander &Rewriter,
3383 SmallVectorImpl<WeakVH> &DeadInsts) const {
3384 const LSRUse &LU = Uses[LF.LUIdx];
3386 // Determine an input position which will be dominated by the operands and
3387 // which will dominate the result.
3388 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3390 // Inform the Rewriter if we have a post-increment use, so that it can
3391 // perform an advantageous expansion.
3392 Rewriter.setPostInc(LF.PostIncLoops);
3394 // This is the type that the user actually needs.
3395 const Type *OpTy = LF.OperandValToReplace->getType();
3396 // This will be the type that we'll initially expand to.
3397 const Type *Ty = F.getType();
3399 // No type known; just expand directly to the ultimate type.
3401 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3402 // Expand directly to the ultimate type if it's the right size.
3404 // This is the type to do integer arithmetic in.
3405 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3407 // Build up a list of operands to add together to form the full base.
3408 SmallVector<const SCEV *, 8> Ops;
3410 // Expand the BaseRegs portion.
3411 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3412 E = F.BaseRegs.end(); I != E; ++I) {
3413 const SCEV *Reg = *I;
3414 assert(!Reg->isZero() && "Zero allocated in a base register!");
3416 // If we're expanding for a post-inc user, make the post-inc adjustment.
3417 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3418 Reg = TransformForPostIncUse(Denormalize, Reg,
3419 LF.UserInst, LF.OperandValToReplace,
3422 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3425 // Flush the operand list to suppress SCEVExpander hoisting.
3427 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3429 Ops.push_back(SE.getUnknown(FullV));
3432 // Expand the ScaledReg portion.
3433 Value *ICmpScaledV = 0;
3434 if (F.AM.Scale != 0) {
3435 const SCEV *ScaledS = F.ScaledReg;
3437 // If we're expanding for a post-inc user, make the post-inc adjustment.
3438 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3439 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3440 LF.UserInst, LF.OperandValToReplace,
3443 if (LU.Kind == LSRUse::ICmpZero) {
3444 // An interesting way of "folding" with an icmp is to use a negated
3445 // scale, which we'll implement by inserting it into the other operand
3447 assert(F.AM.Scale == -1 &&
3448 "The only scale supported by ICmpZero uses is -1!");
3449 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3451 // Otherwise just expand the scaled register and an explicit scale,
3452 // which is expected to be matched as part of the address.
3453 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3454 ScaledS = SE.getMulExpr(ScaledS,
3455 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3456 Ops.push_back(ScaledS);
3458 // Flush the operand list to suppress SCEVExpander hoisting.
3459 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3461 Ops.push_back(SE.getUnknown(FullV));
3465 // Expand the GV portion.
3467 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3469 // Flush the operand list to suppress SCEVExpander hoisting.
3470 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3472 Ops.push_back(SE.getUnknown(FullV));
3475 // Expand the immediate portion.
3476 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3478 if (LU.Kind == LSRUse::ICmpZero) {
3479 // The other interesting way of "folding" with an ICmpZero is to use a
3480 // negated immediate.
3482 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3484 Ops.push_back(SE.getUnknown(ICmpScaledV));
3485 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3488 // Just add the immediate values. These again are expected to be matched
3489 // as part of the address.
3490 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3494 // Emit instructions summing all the operands.
3495 const SCEV *FullS = Ops.empty() ?
3496 SE.getConstant(IntTy, 0) :
3498 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3500 // We're done expanding now, so reset the rewriter.
3501 Rewriter.clearPostInc();
3503 // An ICmpZero Formula represents an ICmp which we're handling as a
3504 // comparison against zero. Now that we've expanded an expression for that
3505 // form, update the ICmp's other operand.
3506 if (LU.Kind == LSRUse::ICmpZero) {
3507 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3508 DeadInsts.push_back(CI->getOperand(1));
3509 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3510 "a scale at the same time!");
3511 if (F.AM.Scale == -1) {
3512 if (ICmpScaledV->getType() != OpTy) {
3514 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3516 ICmpScaledV, OpTy, "tmp", CI);
3519 CI->setOperand(1, ICmpScaledV);
3521 assert(F.AM.Scale == 0 &&
3522 "ICmp does not support folding a global value and "
3523 "a scale at the same time!");
3524 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3526 if (C->getType() != OpTy)
3527 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3531 CI->setOperand(1, C);
3538 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3539 /// of their operands effectively happens in their predecessor blocks, so the
3540 /// expression may need to be expanded in multiple places.
3541 void LSRInstance::RewriteForPHI(PHINode *PN,
3544 SCEVExpander &Rewriter,
3545 SmallVectorImpl<WeakVH> &DeadInsts,
3547 DenseMap<BasicBlock *, Value *> Inserted;
3548 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3549 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3550 BasicBlock *BB = PN->getIncomingBlock(i);
3552 // If this is a critical edge, split the edge so that we do not insert
3553 // the code on all predecessor/successor paths. We do this unless this
3554 // is the canonical backedge for this loop, which complicates post-inc
3556 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3557 !isa<IndirectBrInst>(BB->getTerminator())) {
3558 Loop *PNLoop = LI.getLoopFor(PN->getParent());
3559 if (!PNLoop || PN->getParent() != PNLoop->getHeader()) {
3560 // Split the critical edge.
3561 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3563 // If PN is outside of the loop and BB is in the loop, we want to
3564 // move the block to be immediately before the PHI block, not
3565 // immediately after BB.
3566 if (L->contains(BB) && !L->contains(PN))
3567 NewBB->moveBefore(PN->getParent());
3569 // Splitting the edge can reduce the number of PHI entries we have.
3570 e = PN->getNumIncomingValues();
3572 i = PN->getBasicBlockIndex(BB);
3576 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3577 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3579 PN->setIncomingValue(i, Pair.first->second);
3581 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3583 // If this is reuse-by-noop-cast, insert the noop cast.
3584 const Type *OpTy = LF.OperandValToReplace->getType();
3585 if (FullV->getType() != OpTy)
3587 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3589 FullV, LF.OperandValToReplace->getType(),
3590 "tmp", BB->getTerminator());
3592 PN->setIncomingValue(i, FullV);
3593 Pair.first->second = FullV;
3598 /// Rewrite - Emit instructions for the leading candidate expression for this
3599 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3600 /// the newly expanded value.
3601 void LSRInstance::Rewrite(const LSRFixup &LF,
3603 SCEVExpander &Rewriter,
3604 SmallVectorImpl<WeakVH> &DeadInsts,
3606 // First, find an insertion point that dominates UserInst. For PHI nodes,
3607 // find the nearest block which dominates all the relevant uses.
3608 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3609 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3611 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3613 // If this is reuse-by-noop-cast, insert the noop cast.
3614 const Type *OpTy = LF.OperandValToReplace->getType();
3615 if (FullV->getType() != OpTy) {
3617 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3618 FullV, OpTy, "tmp", LF.UserInst);
3622 // Update the user. ICmpZero is handled specially here (for now) because
3623 // Expand may have updated one of the operands of the icmp already, and
3624 // its new value may happen to be equal to LF.OperandValToReplace, in
3625 // which case doing replaceUsesOfWith leads to replacing both operands
3626 // with the same value. TODO: Reorganize this.
3627 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3628 LF.UserInst->setOperand(0, FullV);
3630 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3633 DeadInsts.push_back(LF.OperandValToReplace);
3636 /// ImplementSolution - Rewrite all the fixup locations with new values,
3637 /// following the chosen solution.
3639 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3641 // Keep track of instructions we may have made dead, so that
3642 // we can remove them after we are done working.
3643 SmallVector<WeakVH, 16> DeadInsts;
3645 SCEVExpander Rewriter(SE);
3646 Rewriter.disableCanonicalMode();
3647 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3649 // Expand the new value definitions and update the users.
3650 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3651 E = Fixups.end(); I != E; ++I) {
3652 const LSRFixup &Fixup = *I;
3654 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3659 // Clean up after ourselves. This must be done before deleting any
3663 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3666 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3667 : IU(P->getAnalysis<IVUsers>()),
3668 SE(P->getAnalysis<ScalarEvolution>()),
3669 DT(P->getAnalysis<DominatorTree>()),
3670 LI(P->getAnalysis<LoopInfo>()),
3671 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3673 // If LoopSimplify form is not available, stay out of trouble.
3674 if (!L->isLoopSimplifyForm()) return;
3676 // If there's no interesting work to be done, bail early.
3677 if (IU.empty()) return;
3679 DEBUG(dbgs() << "\nLSR on loop ";
3680 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3683 // First, perform some low-level loop optimizations.
3685 OptimizeLoopTermCond();
3687 // Start collecting data and preparing for the solver.
3688 CollectInterestingTypesAndFactors();
3689 CollectFixupsAndInitialFormulae();
3690 CollectLoopInvariantFixupsAndFormulae();
3692 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3693 print_uses(dbgs()));
3695 // Now use the reuse data to generate a bunch of interesting ways
3696 // to formulate the values needed for the uses.
3697 GenerateAllReuseFormulae();
3699 FilterOutUndesirableDedicatedRegisters();
3700 NarrowSearchSpaceUsingHeuristics();
3702 SmallVector<const Formula *, 8> Solution;
3705 // Release memory that is no longer needed.
3711 // Formulae should be legal.
3712 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3713 E = Uses.end(); I != E; ++I) {
3714 const LSRUse &LU = *I;
3715 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3716 JE = LU.Formulae.end(); J != JE; ++J)
3717 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3718 LU.Kind, LU.AccessTy, TLI) &&
3719 "Illegal formula generated!");
3723 // Now that we've decided what we want, make it so.
3724 ImplementSolution(Solution, P);
3727 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3728 if (Factors.empty() && Types.empty()) return;
3730 OS << "LSR has identified the following interesting factors and types: ";
3733 for (SmallSetVector<int64_t, 8>::const_iterator
3734 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3735 if (!First) OS << ", ";
3740 for (SmallSetVector<const Type *, 4>::const_iterator
3741 I = Types.begin(), E = Types.end(); I != E; ++I) {
3742 if (!First) OS << ", ";
3744 OS << '(' << **I << ')';
3749 void LSRInstance::print_fixups(raw_ostream &OS) const {
3750 OS << "LSR is examining the following fixup sites:\n";
3751 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3752 E = Fixups.end(); I != E; ++I) {
3759 void LSRInstance::print_uses(raw_ostream &OS) const {
3760 OS << "LSR is examining the following uses:\n";
3761 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3762 E = Uses.end(); I != E; ++I) {
3763 const LSRUse &LU = *I;
3767 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3768 JE = LU.Formulae.end(); J != JE; ++J) {
3776 void LSRInstance::print(raw_ostream &OS) const {
3777 print_factors_and_types(OS);
3782 void LSRInstance::dump() const {
3783 print(errs()); errs() << '\n';
3788 class LoopStrengthReduce : public LoopPass {
3789 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3790 /// transformation profitability.
3791 const TargetLowering *const TLI;
3794 static char ID; // Pass ID, replacement for typeid
3795 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3798 bool runOnLoop(Loop *L, LPPassManager &LPM);
3799 void getAnalysisUsage(AnalysisUsage &AU) const;
3804 char LoopStrengthReduce::ID = 0;
3805 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3806 "Loop Strength Reduction", false, false)
3807 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3808 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3809 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3810 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3811 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3812 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3813 "Loop Strength Reduction", false, false)
3816 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3817 return new LoopStrengthReduce(TLI);
3820 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3821 : LoopPass(ID), TLI(tli) {
3822 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3825 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3826 // We split critical edges, so we change the CFG. However, we do update
3827 // many analyses if they are around.
3828 AU.addPreservedID(LoopSimplifyID);
3830 AU.addRequired<LoopInfo>();
3831 AU.addPreserved<LoopInfo>();
3832 AU.addRequiredID(LoopSimplifyID);
3833 AU.addRequired<DominatorTree>();
3834 AU.addPreserved<DominatorTree>();
3835 AU.addRequired<ScalarEvolution>();
3836 AU.addPreserved<ScalarEvolution>();
3837 // Requiring LoopSimplify a second time here prevents IVUsers from running
3838 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3839 AU.addRequiredID(LoopSimplifyID);
3840 AU.addRequired<IVUsers>();
3841 AU.addPreserved<IVUsers>();
3844 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3845 bool Changed = false;
3847 // Run the main LSR transformation.
3848 Changed |= LSRInstance(TLI, L, this).getChanged();
3850 // At this point, it is worth checking to see if any recurrence PHIs are also
3851 // dead, so that we can remove them as well.
3852 Changed |= DeleteDeadPHIs(L->getHeader());